representation methods data

• Business Essentials
• Leadership & Management
• Credential of Leadership, Impact, and Management in Business (CLIMB)
• Entrepreneurship & Innovation
• Digital Transformation
• Finance & Accounting
• Business in Society
• For Organizations
• Support Portal
• Media Coverage
• Founding Donors
• Leadership Team

• Harvard Business School →
• HBS Online →
• Business Insights →

Business Insights

Harvard Business School Online's Business Insights Blog provides the career insights you need to achieve your goals and gain confidence in your business skills.

• Career Development
• Communication
• Decision-Making
• Earning Your MBA
• Negotiation
• News & Events
• Productivity
• Staff Spotlight
• Student Profiles
• Work-Life Balance
• AI Essentials for Business
• Alternative Investments
• Business Analytics
• Business Strategy
• Business and Climate Change
• Design Thinking and Innovation
• Digital Marketing Strategy
• Disruptive Strategy
• Economics for Managers
• Entrepreneurship Essentials
• Financial Accounting
• Global Business
• Launching Tech Ventures
• Leadership Principles
• Leadership, Ethics, and Corporate Accountability
• Leading with Finance
• Management Essentials
• Negotiation Mastery
• Organizational Leadership
• Power and Influence for Positive Impact
• Strategy Execution
• Sustainable Business Strategy
• Sustainable Investing
• Winning with Digital Platforms

17 Data Visualization Techniques All Professionals Should Know

• 17 Sep 2019

There’s a growing demand for business analytics and data expertise in the workforce. But you don’t need to be a professional analyst to benefit from data-related skills.

Becoming skilled at common data visualization techniques can help you reap the rewards of data-driven decision-making , including increased confidence and potential cost savings. Learning how to effectively visualize data could be the first step toward using data analytics and data science to your advantage to add value to your organization.

Several data visualization techniques can help you become more effective in your role. Here are 17 essential data visualization techniques all professionals should know, as well as tips to help you effectively present your data.

Access your free e-book today.

What Is Data Visualization?

Data visualization is the process of creating graphical representations of information. This process helps the presenter communicate data in a way that’s easy for the viewer to interpret and draw conclusions.

There are many different techniques and tools you can leverage to visualize data, so you want to know which ones to use and when. Here are some of the most important data visualization techniques all professionals should know.

Data Visualization Techniques

The type of data visualization technique you leverage will vary based on the type of data you’re working with, in addition to the story you’re telling with your data .

Here are some important data visualization techniques to know:

• Gantt Chart
• Box and Whisker Plot
• Waterfall Chart
• Scatter Plot
• Pictogram Chart
• Highlight Table
• Bullet Graph
• Choropleth Map
• Network Diagram
• Correlation Matrices

1. Pie Chart

Pie charts are one of the most common and basic data visualization techniques, used across a wide range of applications. Pie charts are ideal for illustrating proportions, or part-to-whole comparisons.

Because pie charts are relatively simple and easy to read, they’re best suited for audiences who might be unfamiliar with the information or are only interested in the key takeaways. For viewers who require a more thorough explanation of the data, pie charts fall short in their ability to display complex information.

2. Bar Chart

The classic bar chart , or bar graph, is another common and easy-to-use method of data visualization. In this type of visualization, one axis of the chart shows the categories being compared, and the other, a measured value. The length of the bar indicates how each group measures according to the value.

One drawback is that labeling and clarity can become problematic when there are too many categories included. Like pie charts, they can also be too simple for more complex data sets.

3. Histogram

Unlike bar charts, histograms illustrate the distribution of data over a continuous interval or defined period. These visualizations are helpful in identifying where values are concentrated, as well as where there are gaps or unusual values.

Histograms are especially useful for showing the frequency of a particular occurrence. For instance, if you’d like to show how many clicks your website received each day over the last week, you can use a histogram. From this visualization, you can quickly determine which days your website saw the greatest and fewest number of clicks.

4. Gantt Chart

Gantt charts are particularly common in project management, as they’re useful in illustrating a project timeline or progression of tasks. In this type of chart, tasks to be performed are listed on the vertical axis and time intervals on the horizontal axis. Horizontal bars in the body of the chart represent the duration of each activity.

Utilizing Gantt charts to display timelines can be incredibly helpful, and enable team members to keep track of every aspect of a project. Even if you’re not a project management professional, familiarizing yourself with Gantt charts can help you stay organized.

5. Heat Map

A heat map is a type of visualization used to show differences in data through variations in color. These charts use color to communicate values in a way that makes it easy for the viewer to quickly identify trends. Having a clear legend is necessary in order for a user to successfully read and interpret a heatmap.

There are many possible applications of heat maps. For example, if you want to analyze which time of day a retail store makes the most sales, you can use a heat map that shows the day of the week on the vertical axis and time of day on the horizontal axis. Then, by shading in the matrix with colors that correspond to the number of sales at each time of day, you can identify trends in the data that allow you to determine the exact times your store experiences the most sales.

6. A Box and Whisker Plot

A box and whisker plot , or box plot, provides a visual summary of data through its quartiles. First, a box is drawn from the first quartile to the third of the data set. A line within the box represents the median. “Whiskers,” or lines, are then drawn extending from the box to the minimum (lower extreme) and maximum (upper extreme). Outliers are represented by individual points that are in-line with the whiskers.

This type of chart is helpful in quickly identifying whether or not the data is symmetrical or skewed, as well as providing a visual summary of the data set that can be easily interpreted.

7. Waterfall Chart

A waterfall chart is a visual representation that illustrates how a value changes as it’s influenced by different factors, such as time. The main goal of this chart is to show the viewer how a value has grown or declined over a defined period. For example, waterfall charts are popular for showing spending or earnings over time.

8. Area Chart

An area chart , or area graph, is a variation on a basic line graph in which the area underneath the line is shaded to represent the total value of each data point. When several data series must be compared on the same graph, stacked area charts are used.

This method of data visualization is useful for showing changes in one or more quantities over time, as well as showing how each quantity combines to make up the whole. Stacked area charts are effective in showing part-to-whole comparisons.

9. Scatter Plot

Another technique commonly used to display data is a scatter plot . A scatter plot displays data for two variables as represented by points plotted against the horizontal and vertical axis. This type of data visualization is useful in illustrating the relationships that exist between variables and can be used to identify trends or correlations in data.

Scatter plots are most effective for fairly large data sets, since it’s often easier to identify trends when there are more data points present. Additionally, the closer the data points are grouped together, the stronger the correlation or trend tends to be.

10. Pictogram Chart

Pictogram charts , or pictograph charts, are particularly useful for presenting simple data in a more visual and engaging way. These charts use icons to visualize data, with each icon representing a different value or category. For example, data about time might be represented by icons of clocks or watches. Each icon can correspond to either a single unit or a set number of units (for example, each icon represents 100 units).

In addition to making the data more engaging, pictogram charts are helpful in situations where language or cultural differences might be a barrier to the audience’s understanding of the data.

11. Timeline

Timelines are the most effective way to visualize a sequence of events in chronological order. They’re typically linear, with key events outlined along the axis. Timelines are used to communicate time-related information and display historical data.

Timelines allow you to highlight the most important events that occurred, or need to occur in the future, and make it easy for the viewer to identify any patterns appearing within the selected time period. While timelines are often relatively simple linear visualizations, they can be made more visually appealing by adding images, colors, fonts, and decorative shapes.

12. Highlight Table

A highlight table is a more engaging alternative to traditional tables. By highlighting cells in the table with color, you can make it easier for viewers to quickly spot trends and patterns in the data. These visualizations are useful for comparing categorical data.

Depending on the data visualization tool you’re using, you may be able to add conditional formatting rules to the table that automatically color cells that meet specified conditions. For instance, when using a highlight table to visualize a company’s sales data, you may color cells red if the sales data is below the goal, or green if sales were above the goal. Unlike a heat map, the colors in a highlight table are discrete and represent a single meaning or value.

13. Bullet Graph

A bullet graph is a variation of a bar graph that can act as an alternative to dashboard gauges to represent performance data. The main use for a bullet graph is to inform the viewer of how a business is performing in comparison to benchmarks that are in place for key business metrics.

In a bullet graph, the darker horizontal bar in the middle of the chart represents the actual value, while the vertical line represents a comparative value, or target. If the horizontal bar passes the vertical line, the target for that metric has been surpassed. Additionally, the segmented colored sections behind the horizontal bar represent range scores, such as “poor,” “fair,” or “good.”

14. Choropleth Maps

A choropleth map uses color, shading, and other patterns to visualize numerical values across geographic regions. These visualizations use a progression of color (or shading) on a spectrum to distinguish high values from low.

Choropleth maps allow viewers to see how a variable changes from one region to the next. A potential downside to this type of visualization is that the exact numerical values aren’t easily accessible because the colors represent a range of values. Some data visualization tools, however, allow you to add interactivity to your map so the exact values are accessible.

15. Word Cloud

A word cloud , or tag cloud, is a visual representation of text data in which the size of the word is proportional to its frequency. The more often a specific word appears in a dataset, the larger it appears in the visualization. In addition to size, words often appear bolder or follow a specific color scheme depending on their frequency.

Word clouds are often used on websites and blogs to identify significant keywords and compare differences in textual data between two sources. They are also useful when analyzing qualitative datasets, such as the specific words consumers used to describe a product.

16. Network Diagram

Network diagrams are a type of data visualization that represent relationships between qualitative data points. These visualizations are composed of nodes and links, also called edges. Nodes are singular data points that are connected to other nodes through edges, which show the relationship between multiple nodes.

There are many use cases for network diagrams, including depicting social networks, highlighting the relationships between employees at an organization, or visualizing product sales across geographic regions.

17. Correlation Matrix

A correlation matrix is a table that shows correlation coefficients between variables. Each cell represents the relationship between two variables, and a color scale is used to communicate whether the variables are correlated and to what extent.

Correlation matrices are useful to summarize and find patterns in large data sets. In business, a correlation matrix might be used to analyze how different data points about a specific product might be related, such as price, advertising spend, launch date, etc.

Other Data Visualization Options

While the examples listed above are some of the most commonly used techniques, there are many other ways you can visualize data to become a more effective communicator. Some other data visualization options include:

• Bubble clouds
• Circle views
• Dendrograms
• Dot distribution maps
• Open-high-low-close charts
• Polar areas
• Radial trees
• Ring Charts
• Sankey diagram
• Span charts
• Streamgraphs
• Wedge stack graphs
• Violin plots

Tips For Creating Effective Visualizations

Creating effective data visualizations requires more than just knowing how to choose the best technique for your needs. There are several considerations you should take into account to maximize your effectiveness when it comes to presenting data.

Related : What to Keep in Mind When Creating Data Visualizations in Excel

One of the most important steps is to evaluate your audience. For example, if you’re presenting financial data to a team that works in an unrelated department, you’ll want to choose a fairly simple illustration. On the other hand, if you’re presenting financial data to a team of finance experts, it’s likely you can safely include more complex information.

Another helpful tip is to avoid unnecessary distractions. Although visual elements like animation can be a great way to add interest, they can also distract from the key points the illustration is trying to convey and hinder the viewer’s ability to quickly understand the information.

Finally, be mindful of the colors you utilize, as well as your overall design. While it’s important that your graphs or charts are visually appealing, there are more practical reasons you might choose one color palette over another. For instance, using low contrast colors can make it difficult for your audience to discern differences between data points. Using colors that are too bold, however, can make the illustration overwhelming or distracting for the viewer.

Related : Bad Data Visualization: 5 Examples of Misleading Data

Visuals to Interpret and Share Information

No matter your role or title within an organization, data visualization is a skill that’s important for all professionals. Being able to effectively present complex data through easy-to-understand visual representations is invaluable when it comes to communicating information with members both inside and outside your business.

There’s no shortage in how data visualization can be applied in the real world. Data is playing an increasingly important role in the marketplace today, and data literacy is the first step in understanding how analytics can be used in business.

Are you interested in improving your analytical skills? Learn more about Business Analytics , our eight-week online course that can help you use data to generate insights and tackle business decisions.

This post was updated on January 20, 2022. It was originally published on September 17, 2019.

About the Author

• School Guide
• Mathematics
• Number System and Arithmetic
• Trigonometry
• Probability
• Mensuration
• Maths Formulas
• Class 8 Maths Notes
• Class 9 Maths Notes
• Class 10 Maths Notes
• Class 11 Maths Notes
• Class 12 Maths Notes

What are the different ways of Data Representation?

• What are the Different Kinds of Data Scientists?
• Different types of Coding Schemes to represent data
• Graphical Representation of Data
• What are the types of statistics?
• Textual Presentation of Data: Meaning, Suitability, and Drawbacks
• Diagrammatic and Graphic Presentation of Data
• Different Types of Data in Data Mining
• Tabular Presentation of Data: Meaning, Objectives, Features and Merits
• What is a Dataset: Types, Features, and Examples
• Data Manipulation: Definition, Examples, and Uses
• Collection and Presentation of Data
• What is Data Organization?
• Different forms of data representation in today's world
• Difference Between Presentation and Representation
• What are the Basic Data Types in PHP ?
• Processing of Raw Data to Tidy Data in R
• Graph and its representations
• Difference between Information and Data
• Data Preprocessing and Its Types
• What are Numbers?
• Propositional Logic
• Pigeonhole Principle
• L U Decomposition
• Set Theory - Definition, Types of Sets, Symbols & Examples
• Predicates and Quantifiers
• Binary to Decimal Converter
• Number System in Maths
• Puzzle 26 | (Know Average Salary without Disclosing Individual Salaries)

The process of collecting the data and analyzing that data in large quantity is known as statistics. It is a branch of mathematics trading with the collection, analysis, interpretation, and presentation of numeral facts and figures.

It is a numerical statement that helps us to collect and analyze the data in large quantity the statistics are based on two of its concepts:

• Statistical Data 
• Statistical Science

Statistics must be expressed numerically and should be collected systematically.

Data Representation

The word data refers to constituting people, things, events, ideas. It can be a title, an integer, or anycast.  After collecting data the investigator has to condense them in tabular form to study their salient features. Such an arrangement is known as the presentation of data.

It refers to the process of condensing the collected data in a tabular form or graphically. This arrangement of data is known as Data Representation.

The row can be placed in different orders like it can be presented in ascending orders, descending order, or can be presented in alphabetical order. 

Example: Let the marks obtained by 10 students of class V in a class test, out of 50 according to their roll numbers, be: 39, 44, 49, 40, 22, 10, 45, 38, 15, 50 The data in the given form is known as raw data. The above given data can be placed in the serial order as shown below: Roll No. Marks 1 39 2 44 3 49 4 40 5 22 6 10 7 45 8 38 9 14 10 50 Now, if you want to analyse the standard of achievement of the students. If you arrange them in ascending or descending order, it will give you a better picture. Ascending order: 10, 15, 22, 38, 39, 40, 44. 45, 49, 50 Descending order: 50, 49, 45, 44, 40, 39, 38, 22, 15, 10 When the row is placed in ascending or descending order is known as arrayed data.

Types of Graphical Data Representation

Bar chart helps us to represent the collected data visually. The collected data can be visualized horizontally or vertically in a bar chart like amounts and frequency. It can be grouped or single. It helps us in comparing different items. By looking at all the bars, it is easy to say which types in a group of data influence the other.

Now let us understand bar chart by taking this example  Let the marks obtained by 5 students of class V in a class test, out of 10 according to their names, be: 7,8,4,9,6 The data in the given form is known as raw data. The above given data can be placed in the bar chart as shown below: Name Marks Akshay 7 Maya 8 Dhanvi 4 Jaslen 9 Muskan 6

A histogram is the graphical representation of data. It is similar to the appearance of a bar graph but there is a lot of difference between histogram and bar graph because a bar graph helps to measure the frequency of categorical data. A categorical data means it is based on two or more categories like gender, months, etc. Whereas histogram is used for quantitative data.

For example:

The graph which uses lines and points to present the change in time is known as a line graph. Line graphs can be based on the number of animals left on earth, the increasing population of the world day by day, or the increasing or decreasing the number of bitcoins day by day, etc. The line graphs tell us about the changes occurring across the world over time. In a  line graph, we can tell about two or more types of changes occurring around the world.

For Example:

Pie chart is a type of graph that involves a structural graphic representation of numerical proportion. It can be replaced in most cases by other plots like a bar chart, box plot, dot plot, etc. As per the research, it is shown that it is difficult to compare the different sections of a given pie chart, or if it is to compare data across different pie charts.

Frequency Distribution Table

A frequency distribution table is a chart that helps us to summarise the value and the frequency of the chart. This frequency distribution table has two columns, The first column consist of the list of the various outcome in the data, While the second column list the frequency of each outcome of the data. By putting this kind of data into a table it helps us to make it easier to understand and analyze the data. 

For Example: To create a frequency distribution table, we would first need to list all the outcomes in the data. In this example, the results are 0 runs, 1 run, 2 runs, and 3 runs. We would list these numerals in numerical ranking in the foremost queue. Subsequently, we ought to calculate how many times per result happened. They scored 0 runs in the 1st, 4th, 7th, and 8th innings, 1 run in the 2nd, 5th, and the 9th innings, 2 runs in the 6th inning, and 3 runs in the 3rd inning. We set the frequency of each result in the double queue. You can notice that the table is a vastly more useful method to show this data.  Baseball Team Runs Per Inning Number of Runs Frequency           0       4           1        3            2        1            3        1

Sample Questions

Question 1: Considering the school fee submission of 10 students of class 10th is given below:

In order to draw the bar graph for the data above, we prepare the frequency table as given below. Fee submission No. of Students Paid   6 Not paid    4 Now we have to represent the data by using the bar graph. It can be drawn by following the steps given below: Step 1: firstly we have to draw the two axis of the graph X-axis and the Y-axis. The varieties of the data must be put on the X-axis (the horizontal line) and the frequencies of the data must be put on the Y-axis (the vertical line) of the graph. Step 2: After drawing both the axis now we have to give the numeric scale to the Y-axis (the vertical line) of the graph It should be started from zero and ends up with the highest value of the data. Step 3: After the decision of the range at the Y-axis now we have to give it a suitable difference of the numeric scale. Like it can be 0,1,2,3…….or 0,10,20,30 either we can give it a numeric scale like 0,20,40,60… Step 4: Now on the X-axis we have to label it appropriately. Step 5: Now we have to draw the bars according to the data but we have to keep in mind that all the bars should be of the same length and there should be the same distance between each graph

Question 2: Watch the subsequent pie chart that denotes the money spent by Megha at the funfair. The suggested colour indicates the quantity paid for each variety. The total value of the data is 15 and the amount paid on each variety is diagnosed as follows:

Chocolates – 3

Wafers – 3

Toys – 2

Rides – 7

To convert this into pie chart percentage, we apply the formula:  (Frequency/Total Frequency) × 100 Let us convert the above data into a percentage: Amount paid on rides: (7/15) × 100 = 47% Amount paid on toys: (2/15) × 100 = 13% Amount paid on wafers: (3/15) × 100 = 20% Amount paid on chocolates: (3/15) × 100 = 20 %

Question 3: The line graph given below shows how Devdas’s height changes as he grows.

Given below is a line graph showing the height changes in Devdas’s as he grows. Observe the graph and answer the questions below.

(i) What was the height of  Devdas’s at 8 years? Answer: 65 inches (ii) What was the height of  Devdas’s at 6 years? Answer:  50 inches (iii) What was the height of  Devdas’s at 2 years? Answer: 35 inches (iv) How much has  Devdas’s grown from 2 to 8 years? Answer: 30 inches (v) When was  Devdas’s 35 inches tall? Answer: 2 years.

Please Login to comment...

Similar reads.

• School Learning

Improve your Coding Skills with Practice

What kind of Experience do you want to share?

An official website of the United States government

The .gov means it's official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you're on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings
• Browse Titles

NCBI Bookshelf. A service of the National Library of Medicine, National Institutes of Health.

National Research Council; Division of Behavioral and Social Sciences and Education; Commission on Behavioral and Social Sciences and Education; Committee on Basic Research in the Behavioral and Social Sciences; Gerstein DR, Luce RD, Smelser NJ, et al., editors. The Behavioral and Social Sciences: Achievements and Opportunities. Washington (DC): National Academies Press (US); 1988.

The Behavioral and Social Sciences: Achievements and Opportunities.

• Hardcopy Version at National Academies Press

5 Methods of Data Collection, Representation, and Analysis

This chapter concerns research on collecting, representing, and analyzing the data that underlie behavioral and social sciences knowledge. Such research, methodological in character, includes ethnographic and historical approaches, scaling, axiomatic measurement, and statistics, with its important relatives, econometrics and psychometrics. The field can be described as including the self-conscious study of how scientists draw inferences and reach conclusions from observations. Since statistics is the largest and most prominent of methodological approaches and is used by researchers in virtually every discipline, statistical work draws the lion’s share of this chapter’s attention.

Problems of interpreting data arise whenever inherent variation or measurement fluctuations create challenges to understand data or to judge whether observed relationships are significant, durable, or general. Some examples: Is a sharp monthly (or yearly) increase in the rate of juvenile delinquency (or unemployment) in a particular area a matter for alarm, an ordinary periodic or random fluctuation, or the result of a change or quirk in reporting method? Do the temporal patterns seen in such repeated observations reflect a direct causal mechanism, a complex of indirect ones, or just imperfections in the data? Is a decrease in auto injuries an effect of a new seat-belt law? Are the disagreements among people describing some aspect of a subculture too great to draw valid inferences about that aspect of the culture?

Such issues of inference are often closely connected to substantive theory and specific data, and to some extent it is difficult and perhaps misleading to treat methods of data collection, representation, and analysis separately. This report does so, as do all sciences to some extent, because the methods developed often are far more general than the specific problems that originally gave rise to them. There is much transfer of new ideas from one substantive field to another—and to and from fields outside the behavioral and social sciences. Some of the classical methods of statistics arose in studies of astronomical observations, biological variability, and human diversity. The major growth of the classical methods occurred in the twentieth century, greatly stimulated by problems in agriculture and genetics. Some methods for uncovering geometric structures in data, such as multidimensional scaling and factor analysis, originated in research on psychological problems, but have been applied in many other sciences. Some time-series methods were developed originally to deal with economic data, but they are equally applicable to many other kinds of data.

• In economics: large-scale models of the U.S. economy; effects of taxation, money supply, and other government fiscal and monetary policies; theories of duopoly, oligopoly, and rational expectations; economic effects of slavery.
• In psychology: test calibration; the formation of subjective probabilities, their revision in the light of new information, and their use in decision making; psychiatric epidemiology and mental health program evaluation.
• In sociology and other fields: victimization and crime rates; effects of incarceration and sentencing policies; deployment of police and fire-fighting forces; discrimination, antitrust, and regulatory court cases; social networks; population growth and forecasting; and voting behavior.

Even such an abridged listing makes clear that improvements in methodology are valuable across the spectrum of empirical research in the behavioral and social sciences as well as in application to policy questions. Clearly, methodological research serves many different purposes, and there is a need to develop different approaches to serve those different purposes, including exploratory data analysis, scientific inference about hypotheses and population parameters, individual decision making, forecasting what will happen in the event or absence of intervention, and assessing causality from both randomized experiments and observational data.

This discussion of methodological research is divided into three areas: design, representation, and analysis. The efficient design of investigations must take place before data are collected because it involves how much, what kind of, and how data are to be collected. What type of study is feasible: experimental, sample survey, field observation, or other? What variables should be measured, controlled, and randomized? How extensive a subject pool or observational period is appropriate? How can study resources be allocated most effectively among various sites, instruments, and subsamples?

The construction of useful representations of the data involves deciding what kind of formal structure best expresses the underlying qualitative and quantitative concepts that are being used in a given study. For example, cost of living is a simple concept to quantify if it applies to a single individual with unchanging tastes in stable markets (that is, markets offering the same array of goods from year to year at varying prices), but as a national aggregate for millions of households and constantly changing consumer product markets, the cost of living is not easy to specify clearly or measure reliably. Statisticians, economists, sociologists, and other experts have long struggled to make the cost of living a precise yet practicable concept that is also efficient to measure, and they must continually modify it to reflect changing circumstances.

Data analysis covers the final step of characterizing and interpreting research findings: Can estimates of the relations between variables be made? Can some conclusion be drawn about correlation, cause and effect, or trends over time? How uncertain are the estimates and conclusions and can that uncertainty be reduced by analyzing the data in a different way? Can computers be used to display complex results graphically for quicker or better understanding or to suggest different ways of proceeding?

Advances in analysis, data representation, and research design feed into and reinforce one another in the course of actual scientific work. The intersections between methodological improvements and empirical advances are an important aspect of the multidisciplinary thrust of progress in the behavioral and social sciences.

• Designs for Data Collection

Four broad kinds of research designs are used in the behavioral and social sciences: experimental, survey, comparative, and ethnographic.

Experimental designs, in either the laboratory or field settings, systematically manipulate a few variables while others that may affect the outcome are held constant, randomized, or otherwise controlled. The purpose of randomized experiments is to ensure that only one or a few variables can systematically affect the results, so that causes can be attributed. Survey designs include the collection and analysis of data from censuses, sample surveys, and longitudinal studies and the examination of various relationships among the observed phenomena. Randomization plays a different role here than in experimental designs: it is used to select members of a sample so that the sample is as representative of the whole population as possible. Comparative designs involve the retrieval of evidence that is recorded in the flow of current or past events in different times or places and the interpretation and analysis of this evidence. Ethnographic designs, also known as participant-observation designs, involve a researcher in intensive and direct contact with a group, community, or population being studied, through participation, observation, and extended interviewing.

Experimental Designs

Laboratory experiments.

Laboratory experiments underlie most of the work reported in Chapter 1 , significant parts of Chapter 2 , and some of the newest lines of research in Chapter 3 . Laboratory experiments extend and adapt classical methods of design first developed, for the most part, in the physical and life sciences and agricultural research. Their main feature is the systematic and independent manipulation of a few variables and the strict control or randomization of all other variables that might affect the phenomenon under study. For example, some studies of animal motivation involve the systematic manipulation of amounts of food and feeding schedules while other factors that may also affect motivation, such as body weight, deprivation, and so on, are held constant. New designs are currently coming into play largely because of new analytic and computational methods (discussed below, in “Advances in Statistical Inference and Analysis”).

Two examples of empirically important issues that demonstrate the need for broadening classical experimental approaches are open-ended responses and lack of independence of successive experimental trials. The first concerns the design of research protocols that do not require the strict segregation of the events of an experiment into well-defined trials, but permit a subject to respond at will. These methods are needed when what is of interest is how the respondent chooses to allocate behavior in real time and across continuously available alternatives. Such empirical methods have long been used, but they can generate very subtle and difficult problems in experimental design and subsequent analysis. As theories of allocative behavior of all sorts become more sophisticated and precise, the experimental requirements become more demanding, so the need to better understand and solve this range of design issues is an outstanding challenge to methodological ingenuity.

The second issue arises in repeated-trial designs when the behavior on successive trials, even if it does not exhibit a secular trend (such as a learning curve), is markedly influenced by what has happened in the preceding trial or trials. The more naturalistic the experiment and the more sensitive the meas urements taken, the more likely it is that such effects will occur. But such sequential dependencies in observations cause a number of important conceptual and technical problems in summarizing the data and in testing analytical models, which are not yet completely understood. In the absence of clear solutions, such effects are sometimes ignored by investigators, simplifying the data analysis but leaving residues of skepticism about the reliability and significance of the experimental results. With continuing development of sensitive measures in repeated-trial designs, there is a growing need for more advanced concepts and methods for dealing with experimental results that may be influenced by sequential dependencies.

Randomized Field Experiments

The state of the art in randomized field experiments, in which different policies or procedures are tested in controlled trials under real conditions, has advanced dramatically over the past two decades. Problems that were once considered major methodological obstacles—such as implementing randomized field assignment to treatment and control groups and protecting the randomization procedure from corruption—have been largely overcome. While state-of-the-art standards are not achieved in every field experiment, the commitment to reaching them is rising steadily, not only among researchers but also among customer agencies and sponsors.

The health insurance experiment described in Chapter 2 is an example of a major randomized field experiment that has had and will continue to have important policy reverberations in the design of health care financing. Field experiments with the negative income tax (guaranteed minimum income) conducted in the 1970s were significant in policy debates, even before their completion, and provided the most solid evidence available on how tax-based income support programs and marginal tax rates can affect the work incentives and family structures of the poor. Important field experiments have also been carried out on alternative strategies for the prevention of delinquency and other criminal behavior, reform of court procedures, rehabilitative programs in mental health, family planning, and special educational programs, among other areas.

In planning field experiments, much hinges on the definition and design of the experimental cells, the particular combinations needed of treatment and control conditions for each set of demographic or other client sample characteristics, including specification of the minimum number of cases needed in each cell to test for the presence of effects. Considerations of statistical power, client availability, and the theoretical structure of the inquiry enter into such specifications. Current important methodological thresholds are to find better ways of predicting recruitment and attrition patterns in the sample, of designing experiments that will be statistically robust in the face of problematic sample recruitment or excessive attrition, and of ensuring appropriate acquisition and analysis of data on the attrition component of the sample.

Also of major significance are improvements in integrating detailed process and outcome measurements in field experiments. To conduct research on program effects under field conditions requires continual monitoring to determine exactly what is being done—the process—how it corresponds to what was projected at the outset. Relatively unintrusive, inexpensive, and effective implementation measures are of great interest. There is, in parallel, a growing emphasis on designing experiments to evaluate distinct program components in contrast to summary measures of net program effects.

Finally, there is an important opportunity now for further theoretical work to model organizational processes in social settings and to design and select outcome variables that, in the relatively short time of most field experiments, can predict longer-term effects: For example, in job-training programs, what are the effects on the community (role models, morale, referral networks) or on individual skills, motives, or knowledge levels that are likely to translate into sustained changes in career paths and income levels?

Survey Designs

Many people have opinions about how societal mores, economic conditions, and social programs shape lives and encourage or discourage various kinds of behavior. People generalize from their own cases, and from the groups to which they belong, about such matters as how much it costs to raise a child, the extent to which unemployment contributes to divorce, and so on. In fact, however, effects vary so much from one group to another that homespun generalizations are of little use. Fortunately, behavioral and social scientists have been able to bridge the gaps between personal perspectives and collective realities by means of survey research. In particular, governmental information systems include volumes of extremely valuable survey data, and the facility of modern computers to store, disseminate, and analyze such data has significantly improved empirical tests and led to new understandings of social processes.

Within this category of research designs, two major types are distinguished: repeated cross-sectional surveys and longitudinal panel surveys. In addition, and cross-cutting these types, there is a major effort under way to improve and refine the quality of survey data by investigating features of human memory and of question formation that affect survey response.

Repeated cross-sectional designs can either attempt to measure an entire population—as does the oldest U.S. example, the national decennial census—or they can rest on samples drawn from a population. The general principle is to take independent samples at two or more times, measuring the variables of interest, such as income levels, housing plans, or opinions about public affairs, in the same way. The General Social Survey, collected by the National Opinion Research Center with National Science Foundation support, is a repeated cross sectional data base that was begun in 1972. One methodological question of particular salience in such data is how to adjust for nonresponses and “don’t know” responses. Another is how to deal with self-selection bias. For example, to compare the earnings of women and men in the labor force, it would be mistaken to first assume that the two samples of labor-force participants are randomly selected from the larger populations of men and women; instead, one has to consider and incorporate in the analysis the factors that determine who is in the labor force.

In longitudinal panels, a sample is drawn at one point in time and the relevant variables are measured at this and subsequent times for the same people. In more complex versions, some fraction of each panel may be replaced or added to periodically, such as expanding the sample to include households formed by the children of the original sample. An example of panel data developed in this way is the Panel Study of Income Dynamics (PSID), conducted by the University of Michigan since 1968 (discussed in Chapter 3 ).

Comparing the fertility or income of different people in different circumstances at the same time to find correlations always leaves a large proportion of the variability unexplained, but common sense suggests that much of the unexplained variability is actually explicable. There are systematic reasons for individual outcomes in each person’s past achievements, in parental models, upbringing, and earlier sequences of experiences. Unfortunately, asking people about the past is not particularly helpful: people remake their views of the past to rationalize the present and so retrospective data are often of uncertain validity. In contrast, generation-long longitudinal data allow readings on the sequence of past circumstances uncolored by later outcomes. Such data are uniquely useful for studying the causes and consequences of naturally occurring decisions and transitions. Thus, as longitudinal studies continue, quantitative analysis is becoming feasible about such questions as: How are the decisions of individuals affected by parental experience? Which aspects of early decisions constrain later opportunities? And how does detailed background experience leave its imprint? Studies like the two-decade-long PSID are bringing within grasp a complete generational cycle of detailed data on fertility, work life, household structure, and income.

Advances in Longitudinal Designs

Large-scale longitudinal data collection projects are uniquely valuable as vehicles for testing and improving survey research methodology. In ways that lie beyond the scope of a cross-sectional survey, longitudinal studies can sometimes be designed—without significant detriment to their substantive interests—to facilitate the evaluation and upgrading of data quality; the analysis of relative costs and effectiveness of alternative techniques of inquiry; and the standardization or coordination of solutions to problems of method, concept, and measurement across different research domains.

Some areas of methodological improvement include discoveries about the impact of interview mode on response (mail, telephone, face-to-face); the effects of nonresponse on the representativeness of a sample (due to respondents’ refusal or interviewers’ failure to contact); the effects on behavior of continued participation over time in a sample survey; the value of alternative methods of adjusting for nonresponse and incomplete observations (such as imputation of missing data, variable case weighting); the impact on response of specifying different recall periods, varying the intervals between interviews, or changing the length of interviews; and the comparison and calibration of results obtained by longitudinal surveys, randomized field experiments, laboratory studies, onetime surveys, and administrative records.

It should be especially noted that incorporating improvements in methodology and data quality has been and will no doubt continue to be crucial to the growing success of longitudinal studies. Panel designs are intrinsically more vulnerable than other designs to statistical biases due to cumulative item non-response, sample attrition, time-in-sample effects, and error margins in repeated measures, all of which may produce exaggerated estimates of change. Over time, a panel that was initially representative may become much less representative of a population, not only because of attrition in the sample, but also because of changes in immigration patterns, age structure, and the like. Longitudinal studies are also subject to changes in scientific and societal contexts that may create uncontrolled drifts over time in the meaning of nominally stable questions or concepts as well as in the underlying behavior. Also, a natural tendency to expand over time the range of topics and thus the interview lengths, which increases the burdens on respondents, may lead to deterioration of data quality or relevance. Careful methodological research to understand and overcome these problems has been done, and continued work as a component of new longitudinal studies is certain to advance the overall state of the art.

Longitudinal studies are sometimes pressed for evidence they are not designed to produce: for example, in important public policy questions concerning the impact of government programs in such areas as health promotion, disease prevention, or criminal justice. By using research designs that combine field experiments (with randomized assignment to program and control conditions) and longitudinal surveys, one can capitalize on the strongest merits of each: the experimental component provides stronger evidence for casual statements that are critical for evaluating programs and for illuminating some fundamental theories; the longitudinal component helps in the estimation of long-term program effects and their attenuation. Coupling experiments to ongoing longitudinal studies is not often feasible, given the multiple constraints of not disrupting the survey, developing all the complicated arrangements that go into a large-scale field experiment, and having the populations of interest overlap in useful ways. Yet opportunities to join field experiments to surveys are of great importance. Coupled studies can produce vital knowledge about the empirical conditions under which the results of longitudinal surveys turn out to be similar to—or divergent from—those produced by randomized field experiments. A pattern of divergence and similarity has begun to emerge in coupled studies; additional cases are needed to understand why some naturally occurring social processes and longitudinal design features seem to approximate formal random allocation and others do not. The methodological implications of such new knowledge go well beyond program evaluation and survey research. These findings bear directly on the confidence scientists—and others—can have in conclusions from observational studies of complex behavioral and social processes, particularly ones that cannot be controlled or simulated within the confines of a laboratory environment.

Memory and the Framing of Questions

A very important opportunity to improve survey methods lies in the reduction of nonsampling error due to questionnaire context, phrasing of questions, and, generally, the semantic and social-psychological aspects of surveys. Survey data are particularly affected by the fallibility of human memory and the sensitivity of respondents to the framework in which a question is asked. This sensitivity is especially strong for certain types of attitudinal and opinion questions. Efforts are now being made to bring survey specialists into closer contact with researchers working on memory function, knowledge representation, and language in order to uncover and reduce this kind of error.

Memory for events is often inaccurate, biased toward what respondents believe to be true—or should be true—about the world. In many cases in which data are based on recollection, improvements can be achieved by shifting to techniques of structured interviewing and calibrated forms of memory elicitation, such as specifying recent, brief time periods (for example, in the last seven days) within which respondents recall certain types of events with acceptable accuracy.

• “Taking things altogether, how would you describe your marriage? Would you say that your marriage is very happy, pretty happy, or not too happy?”
• “Taken altogether how would you say things are these days—would you say you are very happy, pretty happy, or not too happy?”

Presenting this sequence in both directions on different forms showed that the order affected answers to the general happiness question but did not change the marital happiness question: responses to the specific issue swayed subsequent responses to the general one, but not vice versa. The explanations for and implications of such order effects on the many kinds of questions and sequences that can be used are not simple matters. Further experimentation on the design of survey instruments promises not only to improve the accuracy and reliability of survey research, but also to advance understanding of how people think about and evaluate their behavior from day to day.

Comparative Designs

Both experiments and surveys involve interventions or questions by the scientist, who then records and analyzes the responses. In contrast, many bodies of social and behavioral data of considerable value are originally derived from records or collections that have accumulated for various nonscientific reasons, quite often administrative in nature, in firms, churches, military organizations, and governments at all levels. Data of this kind can sometimes be subjected to careful scrutiny, summary, and inquiry by historians and social scientists, and statistical methods have increasingly been used to develop and evaluate inferences drawn from such data. Some of the main comparative approaches are cross-national aggregate comparisons, selective comparison of a limited number of cases, and historical case studies.

Among the more striking problems facing the scientist using such data are the vast differences in what has been recorded by different agencies whose behavior is being compared (this is especially true for parallel agencies in different nations), the highly unrepresentative or idiosyncratic sampling that can occur in the collection of such data, and the selective preservation and destruction of records. Means to overcome these problems form a substantial methodological research agenda in comparative research. An example of the method of cross-national aggregative comparisons is found in investigations by political scientists and sociologists of the factors that underlie differences in the vitality of institutions of political democracy in different societies. Some investigators have stressed the existence of a large middle class, others the level of education of a population, and still others the development of systems of mass communication. In cross-national aggregate comparisons, a large number of nations are arrayed according to some measures of political democracy and then attempts are made to ascertain the strength of correlations between these and the other variables. In this line of analysis it is possible to use a variety of statistical cluster and regression techniques to isolate and assess the possible impact of certain variables on the institutions under study. While this kind of research is cross-sectional in character, statements about historical processes are often invoked to explain the correlations.

More limited selective comparisons, applied by many of the classic theorists, involve asking similar kinds of questions but over a smaller range of societies. Why did democracy develop in such different ways in America, France, and England? Why did northeastern Europe develop rational bourgeois capitalism, in contrast to the Mediterranean and Asian nations? Modern scholars have turned their attention to explaining, for example, differences among types of fascism between the two World Wars, and similarities and differences among modern state welfare systems, using these comparisons to unravel the salient causes. The questions asked in these instances are inevitably historical ones.

Historical case studies involve only one nation or region, and so they may not be geographically comparative. However, insofar as they involve tracing the transformation of a society’s major institutions and the role of its main shaping events, they involve a comparison of different periods of a nation’s or a region’s history. The goal of such comparisons is to give a systematic account of the relevant differences. Sometimes, particularly with respect to the ancient societies, the historical record is very sparse, and the methods of history and archaeology mesh in the reconstruction of complex social arrangements and patterns of change on the basis of few fragments.

Like all research designs, comparative ones have distinctive vulnerabilities and advantages: One of the main advantages of using comparative designs is that they greatly expand the range of data, as well as the amount of variation in those data, for study. Consequently, they allow for more encompassing explanations and theories that can relate highly divergent outcomes to one another in the same framework. They also contribute to reducing any cultural biases or tendencies toward parochialism among scientists studying common human phenomena.

One main vulnerability in such designs arises from the problem of achieving comparability. Because comparative study involves studying societies and other units that are dissimilar from one another, the phenomena under study usually occur in very different contexts—so different that in some cases what is called an event in one society cannot really be regarded as the same type of event in another. For example, a vote in a Western democracy is different from a vote in an Eastern bloc country, and a voluntary vote in the United States means something different from a compulsory vote in Australia. These circumstances make for interpretive difficulties in comparing aggregate rates of voter turnout in different countries.

The problem of achieving comparability appears in historical analysis as well. For example, changes in laws and enforcement and recording procedures over time change the definition of what is and what is not a crime, and for that reason it is difficult to compare the crime rates over time. Comparative researchers struggle with this problem continually, working to fashion equivalent measures; some have suggested the use of different measures (voting, letters to the editor, street demonstration) in different societies for common variables (political participation), to try to take contextual factors into account and to achieve truer comparability.

A second vulnerability is controlling variation. Traditional experiments make conscious and elaborate efforts to control the variation of some factors and thereby assess the causal significance of others. In surveys as well as experiments, statistical methods are used to control sources of variation and assess suspected causal significance. In comparative and historical designs, this kind of control is often difficult to attain because the sources of variation are many and the number of cases few. Scientists have made efforts to approximate such control in these cases of “many variables, small N.” One is the method of paired comparisons. If an investigator isolates 15 American cities in which racial violence has been recurrent in the past 30 years, for example, it is helpful to match them with 15 cities of similar population size, geographical region, and size of minorities—such characteristics are controls—and then search for systematic differences between the two sets of cities. Another method is to select, for comparative purposes, a sample of societies that resemble one another in certain critical ways, such as size, common language, and common level of development, thus attempting to hold these factors roughly constant, and then seeking explanations among other factors in which the sampled societies differ from one another.

Ethnographic Designs

Traditionally identified with anthropology, ethnographic research designs are playing increasingly significant roles in most of the behavioral and social sciences. The core of this methodology is participant-observation, in which a researcher spends an extended period of time with the group under study, ideally mastering the local language, dialect, or special vocabulary, and participating in as many activities of the group as possible. This kind of participant-observation is normally coupled with extensive open-ended interviewing, in which people are asked to explain in depth the rules, norms, practices, and beliefs through which (from their point of view) they conduct their lives. A principal aim of ethnographic study is to discover the premises on which those rules, norms, practices, and beliefs are built.

The use of ethnographic designs by anthropologists has contributed significantly to the building of knowledge about social and cultural variation. And while these designs continue to center on certain long-standing features—extensive face-to-face experience in the community, linguistic competence, participation, and open-ended interviewing—there are newer trends in ethnographic work. One major trend concerns its scale. Ethnographic methods were originally developed largely for studying small-scale groupings known variously as village, folk, primitive, preliterate, or simple societies. Over the decades, these methods have increasingly been applied to the study of small groups and networks within modern (urban, industrial, complex) society, including the contemporary United States. The typical subjects of ethnographic study in modern society are small groups or relatively small social networks, such as outpatient clinics, medical schools, religious cults and churches, ethnically distinctive urban neighborhoods, corporate offices and factories, and government bureaus and legislatures.

As anthropologists moved into the study of modern societies, researchers in other disciplines—particularly sociology, psychology, and political science—began using ethnographic methods to enrich and focus their own insights and findings. At the same time, studies of large-scale structures and processes have been aided by the use of ethnographic methods, since most large-scale changes work their way into the fabric of community, neighborhood, and family, affecting the daily lives of people. Ethnographers have studied, for example, the impact of new industry and new forms of labor in “backward” regions; the impact of state-level birth control policies on ethnic groups; and the impact on residents in a region of building a dam or establishing a nuclear waste dump. Ethnographic methods have also been used to study a number of social processes that lend themselves to its particular techniques of observation and interview—processes such as the formation of class and racial identities, bureaucratic behavior, legislative coalitions and outcomes, and the formation and shifting of consumer tastes.

Advances in structured interviewing (see above) have proven especially powerful in the study of culture. Techniques for understanding kinship systems, concepts of disease, color terminologies, ethnobotany, and ethnozoology have been radically transformed and strengthened by coupling new interviewing methods with modem measurement and scaling techniques (see below). These techniques have made possible more precise comparisons among cultures and identification of the most competent and expert persons within a culture. The next step is to extend these methods to study the ways in which networks of propositions (such as boys like sports, girls like babies) are organized to form belief systems. Much evidence suggests that people typically represent the world around them by means of relatively complex cognitive models that involve interlocking propositions. The techniques of scaling have been used to develop models of how people categorize objects, and they have great potential for further development, to analyze data pertaining to cultural propositions.

Ideological Systems

Perhaps the most fruitful area for the application of ethnographic methods in recent years has been the systematic study of ideologies in modern society. Earlier studies of ideology were in small-scale societies that were rather homogeneous. In these studies researchers could report on a single culture, a uniform system of beliefs and values for the society as a whole. Modern societies are much more diverse both in origins and number of subcultures, related to different regions, communities, occupations, or ethnic groups. Yet these subcultures and ideologies share certain underlying assumptions or at least must find some accommodation with the dominant value and belief systems in the society.

The challenge is to incorporate this greater complexity of structure and process into systematic descriptions and interpretations. One line of work carried out by researchers has tried to track the ways in which ideologies are created, transmitted, and shared among large populations that have traditionally lacked the social mobility and communications technologies of the West. This work has concentrated on large-scale civilizations such as China, India, and Central America. Gradually, the focus has generalized into a concern with the relationship between the great traditions—the central lines of cosmopolitan Confucian, Hindu, or Mayan culture, including aesthetic standards, irrigation technologies, medical systems, cosmologies and calendars, legal codes, poetic genres, and religious doctrines and rites—and the little traditions, those identified with rural, peasant communities. How are the ideological doctrines and cultural values of the urban elites, the great traditions, transmitted to local communities? How are the little traditions, the ideas from the more isolated, less literate, and politically weaker groups in society, transmitted to the elites?

India and southern Asia have been fruitful areas for ethnographic research on these questions. The great Hindu tradition was present in virtually all local contexts through the presence of high-caste individuals in every community. It operated as a pervasive standard of value for all members of society, even in the face of strong little traditions. The situation is surprisingly akin to that of modern, industrialized societies. The central research questions are the degree and the nature of penetration of dominant ideology, even in groups that appear marginal and subordinate and have no strong interest in sharing the dominant value system. In this connection the lowest and poorest occupational caste—the untouchables—serves as an ultimate test of the power of ideology and cultural beliefs to unify complex hierarchical social systems.

Historical Reconstruction

Another current trend in ethnographic methods is its convergence with archival methods. One joining point is the application of descriptive and interpretative procedures used by ethnographers to reconstruct the cultures that created historical documents, diaries, and other records, to interview history, so to speak. For example, a revealing study showed how the Inquisition in the Italian countryside between the 1570s and 1640s gradually worked subtle changes in an ancient fertility cult in peasant communities; the peasant beliefs and rituals assimilated many elements of witchcraft after learning them from their persecutors. A good deal of social history—particularly that of the family—has drawn on discoveries made in the ethnographic study of primitive societies. As described in Chapter 4 , this particular line of inquiry rests on a marriage of ethnographic, archival, and demographic approaches.

Other lines of ethnographic work have focused on the historical dimensions of nonliterate societies. A strikingly successful example in this kind of effort is a study of head-hunting. By combining an interpretation of local oral tradition with the fragmentary observations that were made by outside observers (such as missionaries, traders, colonial officials), historical fluctuations in the rate and significance of head-hunting were shown to be partly in response to such international forces as the great depression and World War II. Researchers are also investigating the ways in which various groups in contemporary societies invent versions of traditions that may or may not reflect the actual history of the group. This process has been observed among elites seeking political and cultural legitimation and among hard-pressed minorities (for example, the Basque in Spain, the Welsh in Great Britain) seeking roots and political mobilization in a larger society.

Ethnography is a powerful method to record, describe, and interpret the system of meanings held by groups and to discover how those meanings affect the lives of group members. It is a method well adapted to the study of situations in which people interact with one another and the researcher can interact with them as well, so that information about meanings can be evoked and observed. Ethnography is especially suited to exploration and elucidation of unsuspected connections; ideally, it is used in combination with other methods—experimental, survey, or comparative—to establish with precision the relative strengths and weaknesses of such connections. By the same token, experimental, survey, and comparative methods frequently yield connections, the meaning of which is unknown; ethnographic methods are a valuable way to determine them.

• Models for Representing Phenomena

The objective of any science is to uncover the structure and dynamics of the phenomena that are its subject, as they are exhibited in the data. Scientists continuously try to describe possible structures and ask whether the data can, with allowance for errors of measurement, be described adequately in terms of them. Over a long time, various families of structures have recurred throughout many fields of science; these structures have become objects of study in their own right, principally by statisticians, other methodological specialists, applied mathematicians, and philosophers of logic and science. Methods have evolved to evaluate the adequacy of particular structures to account for particular types of data. In the interest of clarity we discuss these structures in this section and the analytical methods used for estimation and evaluation of them in the next section, although in practice they are closely intertwined.

A good deal of mathematical and statistical modeling attempts to describe the relations, both structural and dynamic, that hold among variables that are presumed to be representable by numbers. Such models are applicable in the behavioral and social sciences only to the extent that appropriate numerical measurement can be devised for the relevant variables. In many studies the phenomena in question and the raw data obtained are not intrinsically numerical, but qualitative, such as ethnic group identifications. The identifying numbers used to code such questionnaire categories for computers are no more than labels, which could just as well be letters or colors. One key question is whether there is some natural way to move from the qualitative aspects of such data to a structural representation that involves one of the well-understood numerical or geometric models or whether such an attempt would be inherently inappropriate for the data in question. The decision as to whether or not particular empirical data can be represented in particular numerical or more complex structures is seldom simple, and strong intuitive biases or a priori assumptions about what can and cannot be done may be misleading.

Recent decades have seen rapid and extensive development and application of analytical methods attuned to the nature and complexity of social science data. Examples of nonnumerical modeling are increasing. Moreover, the widespread availability of powerful computers is probably leading to a qualitative revolution, it is affecting not only the ability to compute numerical solutions to numerical models, but also to work out the consequences of all sorts of structures that do not involve numbers at all. The following discussion gives some indication of the richness of past progress and of future prospects although it is by necessity far from exhaustive.

In describing some of the areas of new and continuing research, we have organized this section on the basis of whether the representations are fundamentally probabilistic or not. A further useful distinction is between representations of data that are highly discrete or categorical in nature (such as whether a person is male or female) and those that are continuous in nature (such as a person’s height). Of course, there are intermediate cases involving both types of variables, such as color stimuli that are characterized by discrete hues (red, green) and a continuous luminance measure. Probabilistic models lead very naturally to questions of estimation and statistical evaluation of the correspondence between data and model. Those that are not probabilistic involve additional problems of dealing with and representing sources of variability that are not explicitly modeled. At the present time, scientists understand some aspects of structure, such as geometries, and some aspects of randomness, as embodied in probability models, but do not yet adequately understand how to put the two together in a single unified model. Table 5-1 outlines the way we have organized this discussion and shows where the examples in this section lie.

A Classification of Structural Models.

Probability Models

Some behavioral and social sciences variables appear to be more or less continuous, for example, utility of goods, loudness of sounds, or risk associated with uncertain alternatives. Many other variables, however, are inherently categorical, often with only two or a few values possible: for example, whether a person is in or out of school, employed or not employed, identifies with a major political party or political ideology. And some variables, such as moral attitudes, are typically measured in research with survey questions that allow only categorical responses. Much of the early probability theory was formulated only for continuous variables; its use with categorical variables was not really justified, and in some cases it may have been misleading. Recently, very significant advances have been made in how to deal explicitly with categorical variables. This section first describes several contemporary approaches to models involving categorical variables, followed by ones involving continuous representations.

Log-Linear Models for Categorical Variables

Many recent models for analyzing categorical data of the kind usually displayed as counts (cell frequencies) in multidimensional contingency tables are subsumed under the general heading of log-linear models, that is, linear models in the natural logarithms of the expected counts in each cell in the table. These recently developed forms of statistical analysis allow one to partition variability due to various sources in the distribution of categorical attributes, and to isolate the effects of particular variables or combinations of them.

Present log-linear models were first developed and used by statisticians and sociologists and then found extensive application in other social and behavioral sciences disciplines. When applied, for instance, to the analysis of social mobility, such models separate factors of occupational supply and demand from other factors that impede or propel movement up and down the social hierarchy. With such models, for example, researchers discovered the surprising fact that occupational mobility patterns are strikingly similar in many nations of the world (even among disparate nations like the United States and most of the Eastern European socialist countries), and from one time period to another, once allowance is made for differences in the distributions of occupations. The log-linear and related kinds of models have also made it possible to identify and analyze systematic differences in mobility among nations and across time. As another example of applications, psychologists and others have used log-linear models to analyze attitudes and their determinants and to link attitudes to behavior. These methods have also diffused to and been used extensively in the medical and biological sciences.

Regression Models for Categorical Variables

Models that permit one variable to be explained or predicted by means of others, called regression models, are the workhorses of much applied statistics; this is especially true when the dependent (explained) variable is continuous. For a two-valued dependent variable, such as alive or dead, models and approximate theory and computational methods for one explanatory variable were developed in biometry about 50 years ago. Computer programs able to handle many explanatory variables, continuous or categorical, are readily available today. Even now, however, the accuracy of the approximate theory on given data is an open question.

Using classical utility theory, economists have developed discrete choice models that turn out to be somewhat related to the log-linear and categorical regression models. Models for limited dependent variables, especially those that cannot take on values above or below a certain level (such as weeks unemployed, number of children, and years of schooling) have been used profitably in economics and in some other areas. For example, censored normal variables (called tobits in economics), in which observed values outside certain limits are simply counted, have been used in studying decisions to go on in school. It will require further research and development to incorporate information about limited ranges of variables fully into the main multivariate methodologies. In addition, with respect to the assumptions about distribution and functional form conventionally made in discrete response models, some new methods are now being developed that show promise of yielding reliable inferences without making unrealistic assumptions; further research in this area promises significant progress.

One problem arises from the fact that many of the categorical variables collected by the major data bases are ordered. For example, attitude surveys frequently use a 3-, 5-, or 7-point scale (from high to low) without specifying numerical intervals between levels. Social class and educational levels are often described by ordered categories. Ignoring order information, which many traditional statistical methods do, may be inefficient or inappropriate, but replacing the categories by successive integers or other arbitrary scores may distort the results. (For additional approaches to this question, see sections below on ordered structures.) Regression-like analysis of ordinal categorical variables is quite well developed, but their multivariate analysis needs further research. New log-bilinear models have been proposed, but to date they deal specifically with only two or three categorical variables. Additional research extending the new models, improving computational algorithms, and integrating the models with work on scaling promise to lead to valuable new knowledge.

Models for Event Histories

Event-history studies yield the sequence of events that respondents to a survey sample experience over a period of time; for example, the timing of marriage, childbearing, or labor force participation. Event-history data can be used to study educational progress, demographic processes (migration, fertility, and mortality), mergers of firms, labor market behavior, and even riots, strikes, and revolutions. As interest in such data has grown, many researchers have turned to models that pertain to changes in probabilities over time to describe when and how individuals move among a set of qualitative states.

Much of the progress in models for event-history data builds on recent developments in statistics and biostatistics for life-time, failure-time, and hazard models. Such models permit the analysis of qualitative transitions in a population whose members are undergoing partially random organic deterioration, mechanical wear, or other risks over time. With the increased complexity of event-history data that are now being collected, and the extension of event-history data bases over very long periods of time, new problems arise that cannot be effectively handled by older types of analysis. Among the problems are repeated transitions, such as between unemployment and employment or marriage and divorce; more than one time variable (such as biological age, calendar time, duration in a stage, and time exposed to some specified condition); latent variables (variables that are explicitly modeled even though not observed); gaps in the data; sample attrition that is not randomly distributed over the categories; and respondent difficulties in recalling the exact timing of events.

Models for Multiple-Item Measurement

For a variety of reasons, researchers typically use multiple measures (or multiple indicators) to represent theoretical concepts. Sociologists, for example, often rely on two or more variables (such as occupation and education) to measure an individual’s socioeconomic position; educational psychologists ordinarily measure a student’s ability with multiple test items. Despite the fact that the basic observations are categorical, in a number of applications this is interpreted as a partitioning of something continuous. For example, in test theory one thinks of the measures of both item difficulty and respondent ability as continuous variables, possibly multidimensional in character.

Classical test theory and newer item-response theories in psychometrics deal with the extraction of information from multiple measures. Testing, which is a major source of data in education and other areas, results in millions of test items stored in archives each year for purposes ranging from college admissions to job-training programs for industry. One goal of research on such test data is to be able to make comparisons among persons or groups even when different test items are used. Although the information collected from each respondent is intentionally incomplete in order to keep the tests short and simple, item-response techniques permit researchers to reconstitute the fragments into an accurate picture of overall group proficiencies. These new methods provide a better theoretical handle on individual differences, and they are expected to be extremely important in developing and using tests. For example, they have been used in attempts to equate different forms of a test given in successive waves during a year, a procedure made necessary in large-scale testing programs by legislation requiring disclosure of test-scoring keys at the time results are given.

An example of the use of item-response theory in a significant research effort is the National Assessment of Educational Progress (NAEP). The goal of this project is to provide accurate, nationally representative information on the average (rather than individual) proficiency of American children in a wide variety of academic subjects as they progress through elementary and secondary school. This approach is an improvement over the use of trend data on university entrance exams, because NAEP estimates of academic achievements (by broad characteristics such as age, grade, region, ethnic background, and so on) are not distorted by the self-selected character of those students who seek admission to college, graduate, and professional programs.

Item-response theory also forms the basis of many new psychometric instruments, known as computerized adaptive testing, currently being implemented by the U.S. military services and under additional development in many testing organizations. In adaptive tests, a computer program selects items for each examinee based upon the examinee’s success with previous items. Generally, each person gets a slightly different set of items and the equivalence of scale scores is established by using item-response theory. Adaptive testing can greatly reduce the number of items needed to achieve a given level of measurement accuracy.

Nonlinear, Nonadditive Models

Virtually all statistical models now in use impose a linearity or additivity assumption of some kind, sometimes after a nonlinear transformation of variables. Imposing these forms on relationships that do not, in fact, possess them may well result in false descriptions and spurious effects. Unwary users, especially of computer software packages, can easily be misled. But more realistic nonlinear and nonadditive multivariate models are becoming available. Extensive use with empirical data is likely to force many changes and enhancements in such models and stimulate quite different approaches to nonlinear multivariate analysis in the next decade.

Geometric and Algebraic Models

Geometric and algebraic models attempt to describe underlying structural relations among variables. In some cases they are part of a probabilistic approach, such as the algebraic models underlying regression or the geometric representations of correlations between items in a technique called factor analysis. In other cases, geometric and algebraic models are developed without explicitly modeling the element of randomness or uncertainty that is always present in the data. Although this latter approach to behavioral and social sciences problems has been less researched than the probabilistic one, there are some advantages in developing the structural aspects independent of the statistical ones. We begin the discussion with some inherently geometric representations and then turn to numerical representations for ordered data.

Although geometry is a huge mathematical topic, little of it seems directly applicable to the kinds of data encountered in the behavioral and social sciences. A major reason is that the primitive concepts normally used in geometry—points, lines, coincidence—do not correspond naturally to the kinds of qualitative observations usually obtained in behavioral and social sciences contexts. Nevertheless, since geometric representations are used to reduce bodies of data, there is a real need to develop a deeper understanding of when such representations of social or psychological data make sense. Moreover, there is a practical need to understand why geometric computer algorithms, such as those of multidimensional scaling, work as well as they apparently do. A better understanding of the algorithms will increase the efficiency and appropriateness of their use, which becomes increasingly important with the widespread availability of scaling programs for microcomputers.

Over the past 50 years several kinds of well-understood scaling techniques have been developed and widely used to assist in the search for appropriate geometric representations of empirical data. The whole field of scaling is now entering a critical juncture in terms of unifying and synthesizing what earlier appeared to be disparate contributions. Within the past few years it has become apparent that several major methods of analysis, including some that are based on probabilistic assumptions, can be unified under the rubric of a single generalized mathematical structure. For example, it has recently been demonstrated that such diverse approaches as nonmetric multidimensional scaling, principal-components analysis, factor analysis, correspondence analysis, and log-linear analysis have more in common in terms of underlying mathematical structure than had earlier been realized.

Nonmetric multidimensional scaling is a method that begins with data about the ordering established by subjective similarity (or nearness) between pairs of stimuli. The idea is to embed the stimuli into a metric space (that is, a geometry with a measure of distance between points) in such a way that distances between points corresponding to stimuli exhibit the same ordering as do the data. This method has been successfully applied to phenomena that, on other grounds, are known to be describable in terms of a specific geometric structure; such applications were used to validate the procedures. Such validation was done, for example, with respect to the perception of colors, which are known to be describable in terms of a particular three-dimensional structure known as the Euclidean color coordinates. Similar applications have been made with Morse code symbols and spoken phonemes. The technique is now used in some biological and engineering applications, as well as in some of the social sciences, as a method of data exploration and simplification.

One question of interest is how to develop an axiomatic basis for various geometries using as a primitive concept an observable such as the subject’s ordering of the relative similarity of one pair of stimuli to another, which is the typical starting point of such scaling. The general task is to discover properties of the qualitative data sufficient to ensure that a mapping into the geometric structure exists and, ideally, to discover an algorithm for finding it. Some work of this general type has been carried out: for example, there is an elegant set of axioms based on laws of color matching that yields the three-dimensional vectorial representation of color space. But the more general problem of understanding the conditions under which the multidimensional scaling algorithms are suitable remains unsolved. In addition, work is needed on understanding more general, non-Euclidean spatial models.

Ordered Factorial Systems

One type of structure common throughout the sciences arises when an ordered dependent variable is affected by two or more ordered independent variables. This is the situation to which regression and analysis-of-variance models are often applied; it is also the structure underlying the familiar physical identities, in which physical units are expressed as products of the powers of other units (for example, energy has the unit of mass times the square of the unit of distance divided by the square of the unit of time).

There are many examples of these types of structures in the behavioral and social sciences. One example is the ordering of preference of commodity bundles—collections of various amounts of commodities—which may be revealed directly by expressions of preference or indirectly by choices among alternative sets of bundles. A related example is preferences among alternative courses of action that involve various outcomes with differing degrees of uncertainty; this is one of the more thoroughly investigated problems because of its potential importance in decision making. A psychological example is the trade-off between delay and amount of reward, yielding those combinations that are equally reinforcing. In a common, applied kind of problem, a subject is given descriptions of people in terms of several factors, for example, intelligence, creativity, diligence, and honesty, and is asked to rate them according to a criterion such as suitability for a particular job.

In all these cases and a myriad of others like them the question is whether the regularities of the data permit a numerical representation. Initially, three types of representations were studied quite fully: the dependent variable as a sum, a product, or a weighted average of the measures associated with the independent variables. The first two representations underlie some psychological and economic investigations, as well as a considerable portion of physical measurement and modeling in classical statistics. The third representation, averaging, has proved most useful in understanding preferences among uncertain outcomes and the amalgamation of verbally described traits, as well as some physical variables.

For each of these three cases—adding, multiplying, and averaging—researchers know what properties or axioms of order the data must satisfy for such a numerical representation to be appropriate. On the assumption that one or another of these representations exists, and using numerical ratings by subjects instead of ordering, a scaling technique called functional measurement (referring to the function that describes how the dependent variable relates to the independent ones) has been developed and applied in a number of domains. What remains problematic is how to encompass at the ordinal level the fact that some random error intrudes into nearly all observations and then to show how that randomness is represented at the numerical level; this continues to be an unresolved and challenging research issue.

During the past few years considerable progress has been made in understanding certain representations inherently different from those just discussed. The work has involved three related thrusts. The first is a scheme of classifying structures according to how uniquely their representation is constrained. The three classical numerical representations are known as ordinal, interval, and ratio scale types. For systems with continuous numerical representations and of scale type at least as rich as the ratio one, it has been shown that only one additional type can exist. A second thrust is to accept structural assumptions, like factorial ones, and to derive for each scale the possible functional relations among the independent variables. And the third thrust is to develop axioms for the properties of an order relation that leads to the possible representations. Much is now known about the possible nonadditive representations of both the multifactor case and the one where stimuli can be combined, such as combining sound intensities.

Closely related to this classification of structures is the question: What statements, formulated in terms of the measures arising in such representations, can be viewed as meaningful in the sense of corresponding to something empirical? Statements here refer to any scientific assertions, including statistical ones, formulated in terms of the measures of the variables and logical and mathematical connectives. These are statements for which asserting truth or falsity makes sense. In particular, statements that remain invariant under certain symmetries of structure have played an important role in classical geometry, dimensional analysis in physics, and in relating measurement and statistical models applied to the same phenomenon. In addition, these ideas have been used to construct models in more formally developed areas of the behavioral and social sciences, such as psychophysics. Current research has emphasized the communality of these historically independent developments and is attempting both to uncover systematic, philosophically sound arguments as to why invariance under symmetries is as important as it appears to be and to understand what to do when structures lack symmetry, as, for example, when variables have an inherent upper bound.

Many subjects do not seem to be correctly represented in terms of distances in continuous geometric space. Rather, in some cases, such as the relations among meanings of words—which is of great interest in the study of memory representations—a description in terms of tree-like, hierarchial structures appears to be more illuminating. This kind of description appears appropriate both because of the categorical nature of the judgments and the hierarchial, rather than trade-off, nature of the structure. Individual items are represented as the terminal nodes of the tree, and groupings by different degrees of similarity are shown as intermediate nodes, with the more general groupings occurring nearer the root of the tree. Clustering techniques, requiring considerable computational power, have been and are being developed. Some successful applications exist, but much more refinement is anticipated.

Network Models

Several other lines of advanced modeling have progressed in recent years, opening new possibilities for empirical specification and testing of a variety of theories. In social network data, relationships among units, rather than the units themselves, are the primary objects of study: friendships among persons, trade ties among nations, cocitation clusters among research scientists, interlocking among corporate boards of directors. Special models for social network data have been developed in the past decade, and they give, among other things, precise new measures of the strengths of relational ties among units. A major challenge in social network data at present is to handle the statistical dependence that arises when the units sampled are related in complex ways.

• Statistical Inference and Analysis

As was noted earlier, questions of design, representation, and analysis are intimately intertwined. Some issues of inference and analysis have been discussed above as related to specific data collection and modeling approaches. This section discusses some more general issues of statistical inference and advances in several current approaches to them.

Causal Inference

Behavioral and social scientists use statistical methods primarily to infer the effects of treatments, interventions, or policy factors. Previous chapters included many instances of causal knowledge gained this way. As noted above, the large experimental study of alternative health care financing discussed in Chapter 2 relied heavily on statistical principles and techniques, including randomization, in the design of the experiment and the analysis of the resulting data. Sophisticated designs were necessary in order to answer a variety of questions in a single large study without confusing the effects of one program difference (such as prepayment or fee for service) with the effects of another (such as different levels of deductible costs), or with effects of unobserved variables (such as genetic differences). Statistical techniques were also used to ascertain which results applied across the whole enrolled population and which were confined to certain subgroups (such as individuals with high blood pressure) and to translate utilization rates across different programs and types of patients into comparable overall dollar costs and health outcomes for alternative financing options.

A classical experiment, with systematic but randomly assigned variation of the variables of interest (or some reasonable approach to this), is usually considered the most rigorous basis from which to draw such inferences. But random samples or randomized experimental manipulations are not always feasible or ethically acceptable. Then, causal inferences must be drawn from observational studies, which, however well designed, are less able to ensure that the observed (or inferred) relationships among variables provide clear evidence on the underlying mechanisms of cause and effect.

Certain recurrent challenges have been identified in studying causal inference. One challenge arises from the selection of background variables to be measured, such as the sex, nativity, or parental religion of individuals in a comparative study of how education affects occupational success. The adequacy of classical methods of matching groups in background variables and adjusting for covariates needs further investigation. Statistical adjustment of biases linked to measured background variables is possible, but it can become complicated. Current work in adjustment for selectivity bias is aimed at weakening implausible assumptions, such as normality, when carrying out these adjustments. Even after adjustment has been made for the measured background variables, other, unmeasured variables are almost always still affecting the results (such as family transfers of wealth or reading habits). Analyses of how the conclusions might change if such unmeasured variables could be taken into account is essential in attempting to make causal inferences from an observational study, and systematic work on useful statistical models for such sensitivity analyses is just beginning.

The third important issue arises from the necessity for distinguishing among competing hypotheses when the explanatory variables are measured with different degrees of precision. Both the estimated size and significance of an effect are diminished when it has large measurement error, and the coefficients of other correlated variables are affected even when the other variables are measured perfectly. Similar results arise from conceptual errors, when one measures only proxies for a theoretical construct (such as years of education to represent amount of learning). In some cases, there are procedures for simultaneously or iteratively estimating both the precision of complex measures and their effect on a particular criterion.

Although complex models are often necessary to infer causes, once their output is available, it should be translated into understandable displays for evaluation. Results that depend on the accuracy of a multivariate model and the associated software need to be subjected to appropriate checks, including the evaluation of graphical displays, group comparisons, and other analyses.

New Statistical Techniques

Internal resampling.

One of the great contributions of twentieth-century statistics was to demonstrate how a properly drawn sample of sufficient size, even if it is only a tiny fraction of the population of interest, can yield very good estimates of most population characteristics. When enough is known at the outset about the characteristic in question—for example, that its distribution is roughly normal—inference from the sample data to the population as a whole is straightforward, and one can easily compute measures of the certainty of inference, a common example being the 95 percent confidence interval around an estimate. But population shapes are sometimes unknown or uncertain, and so inference procedures cannot be so simple. Furthermore, more often than not, it is difficult to assess even the degree of uncertainty associated with complex data and with the statistics needed to unravel complex social and behavioral phenomena.

Internal resampling methods attempt to assess this uncertainty by generating a number of simulated data sets similar to the one actually observed. The definition of similar is crucial, and many methods that exploit different types of similarity have been devised. These methods provide researchers the freedom to choose scientifically appropriate procedures and to replace procedures that are valid under assumed distributional shapes with ones that are not so restricted. Flexible and imaginative computer simulation is the key to these methods. For a simple random sample, the “bootstrap” method repeatedly resamples the obtained data (with replacement) to generate a distribution of possible data sets. The distribution of any estimator can thereby be simulated and measures of the certainty of inference be derived. The “jackknife” method repeatedly omits a fraction of the data and in this way generates a distribution of possible data sets that can also be used to estimate variability. These methods can also be used to remove or reduce bias. For example, the ratio-estimator, a statistic that is commonly used in analyzing sample surveys and censuses, is known to be biased, and the jackknife method can usually remedy this defect. The methods have been extended to other situations and types of analysis, such as multiple regression.

There are indications that under relatively general conditions, these methods, and others related to them, allow more accurate estimates of the uncertainty of inferences than do the traditional ones that are based on assumed (usually, normal) distributions when that distributional assumption is unwarranted. For complex samples, such internal resampling or subsampling facilitates estimating the sampling variances of complex statistics.

An older and simpler, but equally important, idea is to use one independent subsample in searching the data to develop a model and at least one separate subsample for estimating and testing a selected model. Otherwise, it is next to impossible to make allowances for the excessively close fitting of the model that occurs as a result of the creative search for the exact characteristics of the sample data—characteristics that are to some degree random and will not predict well to other samples.

Robust Techniques

Many technical assumptions underlie the analysis of data. Some, like the assumption that each item in a sample is drawn independently of other items, can be weakened when the data are sufficiently structured to admit simple alternative models, such as serial correlation. Usually, these models require that a few parameters be estimated. Assumptions about shapes of distributions, normality being the most common, have proved to be particularly important, and considerable progress has been made in dealing with the consequences of different assumptions.

More recently, robust techniques have been designed that permit sharp, valid discriminations among possible values of parameters of central tendency for a wide variety of alternative distributions by reducing the weight given to occasional extreme deviations. It turns out that by giving up, say, 10 percent of the discrimination that could be provided under the rather unrealistic assumption of normality, one can greatly improve performance in more realistic situations, especially when unusually large deviations are relatively common.

These valuable modifications of classical statistical techniques have been extended to multiple regression, in which procedures of iterative reweighting can now offer relatively good performance for a variety of underlying distributional shapes. They should be extended to more general schemes of analysis.

In some contexts—notably the most classical uses of analysis of variance—the use of adequate robust techniques should help to bring conventional statistical practice closer to the best standards that experts can now achieve.

Many Interrelated Parameters

In trying to give a more accurate representation of the real world than is possible with simple models, researchers sometimes use models with many parameters, all of which must be estimated from the data. Classical principles of estimation, such as straightforward maximum-likelihood, do not yield reliable estimates unless either the number of observations is much larger than the number of parameters to be estimated or special designs are used in conjunction with strong assumptions. Bayesian methods do not draw a distinction between fixed and random parameters, and so may be especially appropriate for such problems.

A variety of statistical methods have recently been developed that can be interpreted as treating many of the parameters as or similar to random quantities, even if they are regarded as representing fixed quantities to be estimated. Theory and practice demonstrate that such methods can improve the simpler fixed-parameter methods from which they evolved, especially when the number of observations is not large relative to the number of parameters. Successful applications include college and graduate school admissions, where quality of previous school is treated as a random parameter when the data are insufficient to separately estimate it well. Efforts to create appropriate models using this general approach for small-area estimation and undercount adjustment in the census are important potential applications.

Missing Data

In data analysis, serious problems can arise when certain kinds of (quantitative or qualitative) information is partially or wholly missing. Various approaches to dealing with these problems have been or are being developed. One of the methods developed recently for dealing with certain aspects of missing data is called multiple imputation: each missing value in a data set is replaced by several values representing a range of possibilities, with statistical dependence among missing values reflected by linkage among their replacements. It is currently being used to handle a major problem of incompatibility between the 1980 and previous Bureau of Census public-use tapes with respect to occupation codes. The extension of these techniques to address such problems as nonresponse to income questions in the Current Population Survey has been examined in exploratory applications with great promise.

Computer Packages and Expert Systems

The development of high-speed computing and data handling has fundamentally changed statistical analysis. Methodologies for all kinds of situations are rapidly being developed and made available for use in computer packages that may be incorporated into interactive expert systems. This computing capability offers the hope that much data analyses will be more carefully and more effectively done than previously and that better strategies for data analysis will move from the practice of expert statisticians, some of whom may not have tried to articulate their own strategies, to both wide discussion and general use.

But powerful tools can be hazardous, as witnessed by occasional dire misuses of existing statistical packages. Until recently the only strategies available were to train more expert methodologists or to train substantive scientists in more methodology, but without the updating of their training it tends to become outmoded. Now there is the opportunity to capture in expert systems the current best methodological advice and practice. If that opportunity is exploited, standard methodological training of social scientists will shift to emphasizing strategies in using good expert systems—including understanding the nature and importance of the comments it provides—rather than in how to patch together something on one’s own. With expert systems, almost all behavioral and social scientists should become able to conduct any of the more common styles of data analysis more effectively and with more confidence than all but the most expert do today. However, the difficulties in developing expert systems that work as hoped for should not be underestimated. Human experts cannot readily explicate all of the complex cognitive network that constitutes an important part of their knowledge. As a result, the first attempts at expert systems were not especially successful (as discussed in Chapter 1 ). Additional work is expected to overcome these limitations, but it is not clear how long it will take.

Exploratory Analysis and Graphic Presentation

The formal focus of much statistics research in the middle half of the twentieth century was on procedures to confirm or reject precise, a priori hypotheses developed in advance of collecting data—that is, procedures to determine statistical significance. There was relatively little systematic work on realistically rich strategies for the applied researcher to use when attacking real-world problems with their multiplicity of objectives and sources of evidence. More recently, a species of quantitative detective work, called exploratory data analysis, has received increasing attention. In this approach, the researcher seeks out possible quantitative relations that may be present in the data. The techniques are flexible and include an important component of graphic representations. While current techniques have evolved for single responses in situations of modest complexity, extensions to multiple responses and to single responses in more complex situations are now possible.

Graphic and tabular presentation is a research domain in active renaissance, stemming in part from suggestions for new kinds of graphics made possible by computer capabilities, for example, hanging histograms and easily assimilated representations of numerical vectors. Research on data presentation has been carried out by statisticians, psychologists, cartographers, and other specialists, and attempts are now being made to incorporate findings and concepts from linguistics, industrial and publishing design, aesthetics, and classification studies in library science. Another influence has been the rapidly increasing availability of powerful computational hardware and software, now available even on desktop computers. These ideas and capabilities are leading to an increasing number of behavioral experiments with substantial statistical input. Nonetheless, criteria of good graphic and tabular practice are still too much matters of tradition and dogma, without adequate empirical evidence or theoretical coherence. To broaden the respective research outlooks and vigorously develop such evidence and coherence, extended collaborations between statistical and mathematical specialists and other scientists are needed, a major objective being to understand better the visual and cognitive processes (see Chapter 1 ) relevant to effective use of graphic or tabular approaches.

Combining Evidence

Combining evidence from separate sources is a recurrent scientific task, and formal statistical methods for doing so go back 30 years or more. These methods include the theory and practice of combining tests of individual hypotheses, sequential design and analysis of experiments, comparisons of laboratories, and Bayesian and likelihood paradigms.

There is now growing interest in more ambitious analytical syntheses, which are often called meta-analyses. One stimulus has been the appearance of syntheses explicitly combining all existing investigations in particular fields, such as prison parole policy, classroom size in primary schools, cooperative studies of therapeutic treatments for coronary heart disease, early childhood education interventions, and weather modification experiments. In such fields, a serious approach to even the simplest question—how to put together separate estimates of effect size from separate investigations—leads quickly to difficult and interesting issues. One issue involves the lack of independence among the available studies, due, for example, to the effect of influential teachers on the research projects of their students. Another issue is selection bias, because only some of the studies carried out, usually those with “significant” findings, are available and because the literature search may not find out all relevant studies that are available. In addition, experts agree, although informally, that the quality of studies from different laboratories and facilities differ appreciably and that such information probably should be taken into account. Inevitably, the studies to be included used different designs and concepts and controlled or measured different variables, making it difficult to know how to combine them.

Rich, informal syntheses, allowing for individual appraisal, may be better than catch-all formal modeling, but the literature on formal meta-analytic models is growing and may be an important area of discovery in the next decade, relevant both to statistical analysis per se and to improved syntheses in the behavioral and social and other sciences.

• Opportunities and Needs

This chapter has cited a number of methodological topics associated with behavioral and social sciences research that appear to be particularly active and promising at the present time. As throughout the report, they constitute illustrative examples of what the committee believes to be important areas of research in the coming decade. In this section we describe recommendations for an additional $16 million annually to facilitate both the development of methodologically oriented research and, equally important, its communication throughout the research community.

Methodological studies, including early computer implementations, have for the most part been carried out by individual investigators with small teams of colleagues or students. Occasionally, such research has been associated with quite large substantive projects, and some of the current developments of computer packages, graphics, and expert systems clearly require large, organized efforts, which often lie at the boundary between grant-supported work and commercial development. As such research is often a key to understanding complex bodies of behavioral and social sciences data, it is vital to the health of these sciences that research support continue on methods relevant to problems of modeling, statistical analysis, representation, and related aspects of behavioral and social sciences data. Researchers and funding agencies should also be especially sympathetic to the inclusion of such basic methodological work in large experimental and longitudinal studies. Additional funding for work in this area, both in terms of individual research grants on methodological issues and in terms of augmentation of large projects to include additional methodological aspects, should be provided largely in the form of investigator-initiated project grants.

Ethnographic and comparative studies also typically rely on project grants to individuals and small groups of investigators. While this type of support should continue, provision should also be made to facilitate the execution of studies using these methods by research teams and to provide appropriate methodological training through the mechanisms outlined below.

Overall, we recommend an increase of $4 million in the level of investigator-initiated grant support for methodological work. An additional $1 million should be devoted to a program of centers for methodological research.

Many of the new methods and models described in the chapter, if and when adopted to any large extent, will demand substantially greater amounts of research devoted to appropriate analysis and computer implementation. New user interfaces and numerical algorithms will need to be designed and new computer programs written. And even when generally available methods (such as maximum-likelihood) are applicable, model application still requires skillful development in particular contexts. Many of the familiar general methods that are applied in the statistical analysis of data are known to provide good approximations when sample sizes are sufficiently large, but their accuracy varies with the specific model and data used. To estimate the accuracy requires extensive numerical exploration. Investigating the sensitivity of results to the assumptions of the models is important and requires still more creative, thoughtful research. It takes substantial efforts of these kinds to bring any new model on line, and the need becomes increasingly important and difficult as statistical models move toward greater realism, usefulness, complexity, and availability in computer form. More complexity in turn will increase the demand for computational power. Although most of this demand can be satisfied by increasingly powerful desktop computers, some access to mainframe and even supercomputers will be needed in selected cases. We recommend an additional $4 million annually to cover the growth in computational demands for model development and testing.

Interaction and cooperation between the developers and the users of statistical and mathematical methods need continual stimulation—both ways. Efforts should be made to teach new methods to a wider variety of potential users than is now the case. Several ways appear effective for methodologists to communicate to empirical scientists: running summer training programs for graduate students, faculty, and other researchers; encouraging graduate students, perhaps through degree requirements, to make greater use of the statistical, mathematical, and methodological resources at their own or affiliated universities; associating statistical and mathematical research specialists with large-scale data collection projects; and developing statistical packages that incorporate expert systems in applying the methods.

Methodologists, in turn, need to become more familiar with the problems actually faced by empirical scientists in the laboratory and especially in the field. Several ways appear useful for communication in this direction: encouraging graduate students in methodological specialties, perhaps through degree requirements, to work directly on empirical research; creating postdoctoral fellowships aimed at integrating such specialists into ongoing data collection projects; and providing for large data collection projects to engage relevant methodological specialists. In addition, research on and development of statistical packages and expert systems should be encouraged to involve the multidisciplinary collaboration of experts with experience in statistical, computer, and cognitive sciences.

A final point has to do with the promise held out by bringing different research methods to bear on the same problems. As our discussions of research methods in this and other chapters have emphasized, different methods have different powers and limitations, and each is designed especially to elucidate one or more particular facets of a subject. An important type of interdisciplinary work is the collaboration of specialists in different research methodologies on a substantive issue, examples of which have been noted throughout this report. If more such research were conducted cooperatively, the power of each method pursued separately would be increased. To encourage such multidisciplinary work, we recommend increased support for fellowships, research workshops, and training institutes.

Funding for fellowships, both pre-and postdoctoral, should be aimed at giving methodologists experience with substantive problems and at upgrading the methodological capabilities of substantive scientists. Such targeted fellowship support should be increased by $4 million annually, of which $3 million should be for predoctoral fellowships emphasizing the enrichment of methodological concentrations. The new support needed for research workshops is estimated to be $1 million annually. And new support needed for various kinds of advanced training institutes aimed at rapidly diffusing new methodological findings among substantive scientists is estimated to be $2 million annually.

• Cite this Page National Research Council; Division of Behavioral and Social Sciences and Education; Commission on Behavioral and Social Sciences and Education; Committee on Basic Research in the Behavioral and Social Sciences; Gerstein DR, Luce RD, Smelser NJ, et al., editors. The Behavioral and Social Sciences: Achievements and Opportunities. Washington (DC): National Academies Press (US); 1988. 5, Methods of Data Collection, Representation, and Analysis.
• PDF version of this title (16M)

In this Page

Other titles in this collection.

• The National Academies Collection: Reports funded by National Institutes of Health

Recent Activity

• Methods of Data Collection, Representation, and Analysis - The Behavioral and So... Methods of Data Collection, Representation, and Analysis - The Behavioral and Social Sciences: Achievements and Opportunities

Your browsing activity is empty.

Activity recording is turned off.

Turn recording back on

Connect with NLM

National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894

Web Policies FOIA HHS Vulnerability Disclosure

Help Accessibility Careers

• school Campus Bookshelves
• menu_book Bookshelves
• perm_media Learning Objects
• login Login
• how_to_reg Request Instructor Account
• hub Instructor Commons
• Download Page (PDF)
• Download Full Book (PDF)
• Periodic Table
• Physics Constants
• Scientific Calculator
• Reference & Cite
• Tools expand_more
• Readability

selected template will load here

This action is not available.

2.1: Types of Data Representation

• Last updated
• Save as PDF
• Page ID 5696

Two common types of graphic displays are bar charts and histograms. Both bar charts and histograms use vertical or horizontal bars to represent the number of data points in each category or interval. The main difference graphically is that in a  bar chart  there are spaces between the bars and in a  histogram  there are not spaces between the bars. Why does this subtle difference exist and what does it imply about graphic displays in general?

Displaying Data

It is often easier for people to interpret relative sizes of data when that data is displayed graphically. Note that a  categorical variable  is a variable that can take on one of a limited number of values and a  quantitative variable  is a variable that takes on numerical values that represent a measurable quantity. Examples of categorical variables are tv stations, the state someone lives in, and eye color while examples of quantitative variables are the height of students or the population of a city. There are a few common ways of displaying data graphically that you should be familiar with. 

A  pie chart  shows the relative proportions of data in different categories.  Pie charts  are excellent ways of displaying categorical data with easily separable groups. The following pie chart shows six categories labeled A−F.  The size of each pie slice is determined by the central angle. Since there are 360 o  in a circle, the size of the central angle θ A  of category A can be found by:

CK-12 Foundation -  https://www.flickr.com/photos/slgc/16173880801  - CCSA

A  bar chart  displays frequencies of categories of data. The bar chart below has 5 categories, and shows the TV channel preferences for 53 adults. The horizontal axis could have also been labeled News, Sports, Local News, Comedy, Action Movies. The reason why the bars are separated by spaces is to emphasize the fact that they are categories and not continuous numbers. For example, just because you split your time between channel 8 and channel 44 does not mean on average you watch channel 26. Categories can be numbers so you need to be very careful.

CK-12 Foundation -  https://www.flickr.com/photos/slgc/16173880801  - CCSA

A  histogram  displays frequencies of quantitative data that has been sorted into intervals. The following is a histogram that shows the heights of a class of 53 students. Notice the largest category is 56-60 inches with 18 people.

A  boxplot  (also known as a  box and whiskers plot ) is another way to display quantitative data. It displays the five 5 number summary (minimum, Q1,  median , Q3, maximum). The box can either be vertically or horizontally displayed depending on the labeling of the axis. The box does not need to be perfectly symmetrical because it represents data that might not be perfectly symmetrical.

Earlier, you were asked about the difference between histograms and bar charts. The reason for the space in bar charts but no space in histograms is bar charts graph categorical variables while histograms graph quantitative variables. It would be extremely improper to forget the space with bar charts because you would run the risk of implying a spectrum from one side of the chart to the other. Note that in the bar chart where TV stations where shown, the station numbers were not listed horizontally in order by size. This was to emphasize the fact that the stations were categories.

Create a boxplot of the following numbers in your calculator.

8.5, 10.9, 9.1, 7.5, 7.2, 6, 2.3, 5.5

Enter the data into L1 by going into the Stat menu.

CK-12 Foundation - CCSA

Then turn the statplot on and choose boxplot.

Use Zoomstat to automatically center the window on the boxplot.

Create a pie chart to represent the preferences of 43 hungry students.

• Other – 5
• Burritos – 7
• Burgers – 9
• Pizza – 22

Create a bar chart representing the preference for sports of a group of 23 people.

• Football – 12
• Baseball – 10
• Basketball – 8
• Hockey – 3

Create a histogram for the income distribution of 200 million people.

• Below $50,000 is 100 million people
• Between $50,000 and $100,000 is 50 million people
• Between $100,000 and $150,000 is 40 million people
• Above $150,000 is 10 million people

1. What types of graphs show categorical data?

2. What types of graphs show quantitative data?

A math class of 30 students had the following grades:

3. Create a bar chart for this data.

4. Create a pie chart for this data.

5. Which graph do you think makes a better visual representation of the data?

A set of 20 exam scores is 67, 94, 88, 76, 85, 93, 55, 87, 80, 81, 80, 61, 90, 84, 75, 93, 75, 68, 100, 98

6. Create a histogram for this data. Use your best judgment to decide what the intervals should be.

7. Find the  five number summary  for this data.

8. Use the  five number summary  to create a boxplot for this data.

9. Describe the data shown in the boxplot below.

10. Describe the data shown in the histogram below.

A math class of 30 students has the following eye colors:

11. Create a bar chart for this data.

12. Create a pie chart for this data.

13. Which graph do you think makes a better visual representation of the data?

14. Suppose you have data that shows the breakdown of registered republicans by state. What types of graphs could you use to display this data?

15. From which types of graphs could you obtain information about the spread of the data? Note that spread is a measure of how spread out all of the data is.

Review (Answers)

To see the Review answers, open this  PDF file  and look for section 15.4. 

Additional Resources

PLIX: Play, Learn, Interact, eXplore - Baby Due Date Histogram

Practice: Types of Data Representation

Real World: Prepare for Impact

The Behavioral and Social Sciences: Achievements and Opportunities (1988)

Chapter: 5. methods of data collection, representation, and anlysis.

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

l - Methods of Data Collection, Representation Analysis , and

SMethods of Data Collection. Representation, and This chapter concerns research on collecting, representing, and analyzing the data that underlie behavioral and social sciences knowledge. Such research, methodological in character, includes ethnographic and historical approaches, scaling, axiomatic measurement, and statistics, with its important relatives, econometrics and psychometrics. The field can be described as including the self-conscious study of how scientists draw inferences and reach conclusions from observations. Since statistics is the largest and most prominent of meth- odological approaches and is used by researchers in virtually every discipline, statistical work draws the lion's share of this chapter's attention. Problems of interpreting data arise whenever inherent variation or measure- ment fluctuations create challenges to understand data or to judge whether observed relationships are significant, durable, or general. Some examples: Is a sharp monthly (or yearly) increase in the rate of juvenile delinquency (or unemployment) in a particular area a matter for alarm, an ordinary periodic or random fluctuation, or the result of a change or quirk in reporting method? Do the temporal patterns seen in such repeated observations reflect a direct causal mechanism, a complex of indirect ones, or just imperfections in the Analysis 167

168 / The Behavioral and Social Sciences data? Is a decrease in auto injuries an effect of a new seat-belt law? Are the disagreements among people describing some aspect of a subculture too great to draw valid inferences about that aspect of the culture? Such issues of inference are often closely connected to substantive theory and specific data, and to some extent it is difficult and perhaps misleading to treat methods of data collection, representation, and analysis separately. This report does so, as do all sciences to some extent, because the methods developed often are far more general than the specific problems that originally gave rise to them. There is much transfer of new ideas from one substantive field to another—and to and from fields outside the behavioral and social sciences. Some of the classical methods of statistics arose in studies of astronomical observations, biological variability, and human diversity. The major growth of the classical methods occurred in the twentieth century, greatly stimulated by problems in agriculture and genetics. Some methods for uncovering geometric structures in data, such as multidimensional scaling and factor analysis, orig- inated in research on psychological problems, but have been applied in many other sciences. Some time-series methods were developed originally to deal with economic data, but they are equally applicable to many other kinds of data. Within the behavioral and social sciences, statistical methods have been developed in and have contributed to an enormous variety of research, includ- ing: · In economics: large-scale models of the U.S. economy; effects of taxa- tion, money supply, and other government fiscal and monetary policies; theories of duopoly, oligopoly, and rational expectations; economic effects of slavery. · In psychology: test calibration; the formation of subjective probabilities, their revision in the light of new information, and their use in decision making; psychiatric epidemiology and mental health program evaluation. · In sociology and other fields: victimization and crime rates; effects of incarceration and sentencing policies; deployment of police and fire-fight- ing forces; discrimination, antitrust, and regulatory court cases; social net- works; population growth and forecasting; and voting behavior. Even such an abridged listing makes clear that improvements in method- ology are valuable across the spectrum of empirical research in the behavioral and social sciences as well as in application to policy questions. Clearly, meth- odological research serves many different purposes, and there is a need to develop different approaches to serve those different purposes, including ex- ploratory data analysis, scientific inference about hypotheses and population parameters, individual decision making, forecasting what will happen in the event or absence of intervention, and assessing causality from both randomized experiments and observational data.

Methods of Data Collection, Representation, and Analysis / 169 This discussion of methodological research is divided into three areas: de- sign, representation, and analysis. The efficient design of investigations must take place before data are collected because it involves how much, what kind of, and how data are to be collected. What type of study is feasible: experi- mental, sample survey, field observation, or other? What variables should be measured, controlled, and randomized? How extensive a subject pool or ob- servational period is appropriate? How can study resources be allocated most effectively among various sites, instruments, and subsamples? The construction of useful representations of the data involves deciding what kind of formal structure best expresses the underlying qualitative and quanti- tative concepts that are being used in a given study. For example, cost of living is a simple concept to quantify if it applies to a single individual with unchang- ing tastes in stable markets (that is, markets offering the same array of goods from year to year at varying prices), but as a national aggregate for millions of households and constantly changing consumer product markets, the cost of living is not easy to specify clearly or measure reliably. Statisticians, economists, sociologists, and other experts have long struggled to make the cost of living a precise yet practicable concept that is also efficient to measure, and they must continually modify it to reflect changing circumstances. Data analysis covers the final step of characterizing and interpreting research findings: Can estimates of the relations between variables be made? Can some conclusion be drawn about correlation, cause and effect, or trends over time? How uncertain are the estimates and conclusions and can that uncertainty be reduced by analyzing the data in a different way? Can computers be used to display complex results graphically for quicker or better understanding or to suggest different ways of proceeding? Advances in analysis, data representation, and research design feed into and reinforce one another in the course of actual scientific work. The intersections between methodological improvements and empirical advances are an impor- tant aspect of the multidisciplinary thrust of progress in the behavioral and . socla. . sciences. DESIGNS FOR DATA COLLECTION Four broad kinds of research designs are used in the behavioral and social sciences: experimental, survey, comparative, and ethnographic. Experimental designs, in either the laboratory or field settings, systematically manipulate a few variables while others that may affect the outcome are held constant, randomized, or otherwise controlled. The purpose of randomized experiments is to ensure that only one or a few variables can systematically affect the results, so that causes can be attributed. Survey designs include the collection and analysis of data from censuses, sample surveys, and longitudinal studies and the examination of various relationships among the observed phe-

170 / The Behavioral and Social Sciences nomena. Randomization plays a different role here than in experimental de- signs: it is used to select members of a sample so that the sample is as repre- sentative of the whole population as possible. Comparative designs involve the retrieval of evidence that is recorded in the flow of current or past events in different times or places and the interpretation and analysis of this evidence. Ethnographic designs, also known as participant-observation designs, involve a researcher in intensive and direct contact with a group, community, or pop- ulation being studied, through participation, observation, and extended inter- vlewlng. Experimental Designs Laboratory Experiments Laboratory experiments underlie most of the work reported in Chapter 1, significant parts of Chapter 2, and some of the newest lines of research in Chapter 3. Laboratory experiments extend and adapt classical methods of de- sign first developed, for the most part, in the physical and life sciences and agricultural research. Their main feature is the systematic and independent manipulation of a few variables and the strict control or randomization of all other variables that might affect the phenomenon under study. For example, some studies of animal motivation involve the systematic manipulation of amounts of food and feeding schedules while other factors that may also affect motiva- tion, such as body weight, deprivation, and so on, are held constant. New designs are currently coming into play largely because of new analytic and computational methods (discussed below, in "Advances in Statistical Inference and Analysis". Two examples of empirically important issues that demonstrate the need for broadening classical experimental approaches are open-ended responses and lack of independence of successive experimental trials. The first concerns the design of research protocols that do not require the strict segregation of the events of an experiment into well-defined trials, but permit a subject to respond at will. These methods are needed when what is of interest is how the respond- ent chooses to allocate behavior in real time and across continuously available alternatives. Such empirical methods have long been used, but they can gen- erate very subtle and difficult problems in experimental design and subsequent analysis. As theories of allocative behavior of all sorts become more sophisti- cated and precise, the experimental requirements become more demanding, so the need to better understand and solve this range of design issues is an outstanding challenge to methodological ingenuity. The second issue arises in repeated-trial designs when the behavior on suc- cessive trials, even if it does not exhibit a secular trend (such as a learning curve), is markedly influenced by what has happened in the preceding trial or trials. The more naturalistic the experiment and the more sensitive the meas-

Methods of Data Collection, Representation, and Analysis / 171 urements taken, the more likely it is that such effects will occur. But such sequential dependencies in observations cause a number of important concep- tual and technical problems in summarizing the data and in testing analytical models, which are not yet completely understood. In the absence of clear solutions, such effects are sometimes ignored by investigators, simplifying the data analysis but leaving residues of skepticism about the reliability and sig- nificance of the experimental results. With continuing development of sensitive measures in repeated-trial designs, there is a growing need for more advanced concepts and methods for dealing with experimental results that may be influ- enced by sequential dependencies. Randomized Field Experiments The state of the art in randomized field experiments, in which different policies or procedures are tested in controlled trials under real conditions, has advanced dramatically over the past two decades. Problems that were once considered major methodological obstacles such as implementing random- ized field assignment to treatment and control groups and protecting the ran- domization procedure from corruption have been largely overcome. While state-of-the-art standards are not achieved in every field experiment, the com- mitment to reaching them is rising steadily, not only among researchers but also among customer agencies and sponsors. The health insurance experiment described in Chapter 2 is an example of a major randomized field experiment that has had and will continue to have important policy reverberations in the design of health care financing. Field experiments with the negative income tax (guaranteed minimum income) con- ducted in the 1970s were significant in policy debates, even before their com- pletion, and provided the most solid evidence available on how tax-based income support programs and marginal tax rates can affect the work incentives and family structures of the poor. Important field experiments have also been carried out on alternative strategies for the prevention of delinquency and other criminal behavior, reform of court procedures, rehabilitative programs in men- tal health, family planning, and special educational programs, among other areas. In planning field experiments, much hinges on the definition and design of the experimental cells, the particular combinations needed of treatment and control conditions for each set of demographic or other client sample charac- teristics, including specification of the minimum number of cases needed in each cell to test for the presence of effects. Considerations of statistical power, client availability, and the theoretical structure of the inquiry enter into such specifications. Current important methodological thresholds are to find better ways of predicting recruitment and attrition patterns in the sample, of designing experiments that will be statistically robust in the face of problematic sample

172 / The Behavioral and Social Sciences recruitment or excessive attrition, and of ensuring appropriate acquisition and analysis of data on the attrition component of the sample. Also of major significance are improvements in integrating detailed process and outcome measurements in field experiments. To conduct research on pro- gram effects under held conditions requires continual monitoring to determine exactly what is being done—the process how it corresponds to what was projected at the outset. Relatively unintrusive, inexpensive, and effective im- plementation measures are of great interest. There is, in parallel, a growing emphasis on designing experiments to evaluate distinct program components in contrast to summary measures of net program effects. Finally, there is an important opportunity now for further theoretical work to model organizational processes in social settings and to design and select outcome variables that, in the relatively short time of most field experiments, can predict longer-term effects: For example, in job-training programs, what are the effects on the community (role models, morale, referral networks) or on individual skills, motives, or knowledge levels that are likely to translate into sustained changes in career paths and income levels? Survey Designs Many people have opinions about how societal mores, economic conditions, and social programs shape lives and encourage or discourage various kinds of behavior. People generalize from their own cases, and from the groups to which they belong, about such matters as how much it costs to raise a child, the extent to which unemployment contributes to divorce, and so on. In fact, however, effects vary so much from one group to another that homespun generalizations are of little use. Fortunately, behavioral and social scientists have been able to bridge the gaps between personal perspectives and collective realities by means of survey research. In particular, governmental information systems include volumes of extremely valuable survey data, and the facility of modern com- puters to store, disseminate, and analyze such data has significantly improved empirical tests and led to new understandings of social processes. Within this category of research designs, two major types are distinguished: repeated cross-sectional surveys and longitudinal panel surveys. In addition, and cross-cutting these types, there is a major effort under way to improve and refine the quality of survey data by investigating features of human memory and of question formation that affect survey response. Repeated cross-sectional designs can either attempt to measure an entire population as does the oldest U.S. example, the national decennial census or they can rest on samples drawn from a population. The general principle is to take independent samples at two or more times, measuring the variables of interest, such as income levels, housing plans, or opinions about public affairs, in the same way. The General Social Survey, collected by the National Opinion Research Center with National Science Foundation support, is a repeated cross-

Methods of Data Collection, Representation, and Analysis / 173 sectional data base that was begun in 1972. One methodological question of particular salience in such data is how to adjust for nonresponses and "don't know" responses. Another is how to deal with self-selection bias. For example, to compare the earnings of women and men in the labor force, it would be mistaken to first assume that the two samples of labor-force participants are randomly selected from the larger populations of men and women; instead, one has to consider and incorporate in the analysis the factors that determine who is in the labor force. In longitudinal panels, a sample is drawn at one point in time and the relevant variables are measured at this and subsequent times for the same people. In more complex versions, some fraction of each panel may be replaced or added to periodically, such as expanding the sample to include households formed by the children of the original sample. An example of panel data developed in this way is the Panel Study of Income Dynamics (PSID), conducted by the University of Michigan since 1968 (discussed in Chapter 35. Comparing the fertility or income of different people in different circum- stances at the same time to kind correlations always leaves a large proportion of the variability unexplained, but common sense suggests that much of the unexplained variability is actually explicable. There are systematic reasons for individual outcomes in each person's past achievements, in parental models, upbringing, and earlier sequences of experiences. Unfortunately, asking people about the past is not particularly helpful: people remake their views of the past to rationalize the present and so retrospective data are often of uncertain va- lidity. In contrast, generation-long longitudinal data allow readings on the sequence of past circumstances uncolored by later outcomes. Such data are uniquely useful for studying the causes and consequences of naturally occur- ring decisions and transitions. Thus, as longitudinal studies continue, quant,i- tative analysis is becoming feasible about such questions as: How are the de- cisions of individuals affected by parental experience? Which aspects of early decisions constrain later opportunities? And how does detailed background experience leave its imprint? Studies like the two-decade-long PSID are bring- ing within grasp a complete generational cycle of detailed data on fertility, work life, household structure, and income. Advances in Longitudinal Designs Large-scale longitudinal data collection projects are uniquely valuable as vehicles for testing and improving survey research methodology. In ways that lie beyond the scope of a cross-sectional survey, longitudinal studies can some- times be designed without significant detriment to their substantive inter- ests to facilitate the evaluation and upgrading of data quality; the analysis of relative costs and effectiveness of alternative techniques of inquiry; and the standardization or coordination of solutions to problems of method, concept, and measurement across different research domains.

174 / The Behavioral and Social Sciences Some areas of methodological improvement include discoveries about the impact of interview mode on response (mail, telephone, face-to-face); the effects of nonresponse on the representativeness of a sample (due to respondents' refusal or interviewers' failure to contact); the effects on behavior of continued participation over time in a sample survey; the value of alternative methods of adjusting for nonresponse and incomplete observations (such as imputation of missing data, variable case weighting); the impact on response of specifying different recall periods, varying the intervals between interviews, or changing the length of interviews; and the comparison and calibration of results obtained by longitudinal surveys, randomized field experiments, laboratory studies, one- time surveys, and administrative records. It should be especially noted that incorporating improvements in method- ology and data quality has been and will no doubt continue to be crucial to the growing success of longitudinal studies. Panel designs are intrinsically more vulnerable than other designs to statistical biases due to cumulative item non- response, sample attrition, time-in-sample effects, and error margins in re- peated measures, all of which may produce exaggerated estimates of change. Over time, a panel that was initially representative may become much less representative of a population, not only because of attrition in the sample, but also because of changes in immigration patterns, age structure, and the like. Longitudinal studies are also subject to changes in scientific and societal con- texts that may create uncontrolled drifts over time in the meaning of nominally stable questions or concepts as well as in the underlying behavior. Also, a natural tendency to expand over time the range of topics and thus the interview lengths, which increases the burdens on respondents, may lead to deterioration of data quality or relevance. Careful methodological research to understand and overcome these problems has been done, and continued work as a com- ponent of new longitudinal studies is certain to advance the overall state of the art. Longitudinal studies are sometimes pressed for evidence they are not de- signed to produce: for example, in important public policy questions concern- ing the impact of government programs in such areas as health promotion, disease prevention, or criminal justice. By using research designs that combine field experiments (with randomized assignment to program and control con- ditions) and longitudinal surveys, one can capitalize on the strongest merits of each: the experimental component provides stronger evidence for casual state- ments that are critical for evaluating programs and for illuminating some fun- damental theories; the longitudinal component helps in the estimation of long- term program effects and their attenuation. Coupling experiments to ongoing longitudinal studies is not often feasible, given the multiple constraints of not disrupting the survey, developing all the complicated arrangements that go into a large-scale field experiment, and having the populations of interest over- lap in useful ways. Yet opportunities to join field experiments to surveys are

Methods of Data Collection, Representation, and Analysis / 175 of great importance. Coupled studies can produce vital knowledge about the empirical conditions under which the results of longitudinal surveys turn out to be similar to—or divergent from those produced by randomized field experiments. A pattern of divergence and similarity has begun to emerge in coupled studies; additional cases are needed to understand why some naturally occurring social processes and longitudinal design features seem to approxi- mate formal random allocation and others do not. The methodological impli- cations of such new knowledge go well beyond program evaluation and survey research. These findings bear directly on the confidence scientists and oth- ers can have in conclusions from observational studies of complex behavioral and social processes, particularly ones that cannot be controlled or simulated within the confines of a laboratory environment. Memory and the Framing of questions A very important opportunity to improve survey methods lies in the reduc- tion of nonsampling error due to questionnaire context, phrasing of questions, and, generally, the semantic and social-psychological aspects of surveys. Survey data are particularly affected by the fallibility of human memory and the sen- sitivity of respondents to the framework in which a question is asked. This sensitivity is especially strong for certain types of attitudinal and opinion ques- tions. Efforts are now being made to bring survey specialists into closer contact with researchers working on memory function, knowledge representation, and language in order to uncover and reduce this kind of error. Memory for events is often inaccurate, biased toward what respondents believe to be true or should be true—about the world. In many cases in which data are based on recollection, improvements can be achieved by shifting to techniques of structured interviewing and calibrated forms of memory elic- itation, such as specifying recent, brief time periods (for example, in the last seven days) within which respondents recall certain types of events with ac- ceptable accuracy. Experiments on individual decision making show that the way a question is framed predictably alters the responses. Analysts of survey data find that some small changes in the wording of certain kinds of questions can produce large differences in the answers, although other wording changes have little effect. Even simply changing the order in which some questions are presented can produce large differences, although for other questions the order of presenta- tion does not matter. For example, the following questions were among those asked in one wave of the General Social Survey: · "Taking things altogether, how would you describe your marriage? Would you say that your marriage is very happy, pretty happy, or not too happy?" · "Taken altogether how would you say things are these days—would you say you are very happy, pretty happy, or not too happy?"

176 / The Behavioral and Social Sciences Presenting this sequence in both directions on different forms showed that the order affected answers to the general happiness question but did not change the marital happiness question: responses to the specific issue swayed subse- quent responses to the general one, but not vice versa. The explanations for and implications of such order effects on the many kinds of questions and sequences that can be used are not simple matters. Further experimentation on the design of survey instruments promises not only to improve the accuracy and reliability of survey research, but also to advance understanding of how people think about and evaluate their behavior from day to day. Comparative Designs Both experiments and surveys involve interventions or questions by the scientist, who then records and analyzes the responses. In contrast, many bodies of social and behavioral data of considerable value are originally derived from records or collections that have accumulated for various nonscientific reasons, quite often administrative in nature, in firms, churches, military or- ganizations, and governments at all levels. Data of this kind can sometimes be subjected to careful scrutiny, summary, and inquiry by historians and social scientists, and statistical methods have increasingly been used to develop and evaluate inferences drawn from such data. Some of the main comparative approaches are'. cross-national aggregate comparisons, selective comparison of a limited number of cases, and historical case studies. Among the more striking problems facing the scientist using such data are the vast differences in what has been recorded by different agencies whose behavior is being compared (this is especially true for parallel agencies in different nations), the highly unrepresentative or idiosyncratic sampling that can occur in the collection of such data, and the selective preservation and destruction of records. Means to overcome these problems form a substantial methodological research agenda in comparative research. An example of the method of cross-national aggregative comparisons is found in investigations by political scientists and sociologists of the factors that underlie differences in the vitality of institutions of political democracy in different societies. Some investigators have stressed the existence of a large middle class, others the level of education of a population, and still others the development of systems of mass communication. In cross-national aggregate comparisons, a large number of nations are arrayed according to some measures of political democracy and then attempts are made to ascertain the strength of correlations between these and the other variables. In this line of analysis it is possible to use a variety of statistical cluster and regression techniques to isolate and assess the possible impact of certain variables on the institutions under study. While this kind of research is cross-sectional in character, statements about historical processes are often invoked to explain the correlations.

Methods of Data Collection, Representation, and Analysis / 177 More limited selective comparisons, applied by many of the classic theorists, involve asking similar kinds of questions but over a smaller range of societies. Why did democracy develop in such different ways in America, France, and England? Why did northeastern Europe develop rational bourgeois capitalism, in contrast to the Mediterranean and Asian nations? Modern scholars have turned their attention to explaining, for example, differences among types of fascism between the two World Wars, and similarities and differences among modern state welfare systems, using these comparisons to unravel the salient causes. The questions asked in these instances are inevitably historical ones. Historical case studies involve only one nation or region, and so they may not be geographically comparative. However, insofar as they involve tracing the transformation of a society's major institutions and the role of its main shaping events, they involve a comparison of different periods of a nation's or a region's history. The goal of such comparisons is to give a systematic account of the relevant differences. Sometimes, particularly with respect to the ancient societies, the historical record is very sparse, and the methods of history and archaeology mesh in the reconstruction of complex social arrangements and patterns of change on the basis of few fragments. Like all research designs, comparative ones have distinctive vulnerabilities and advantages: One of the main advantages of using comparative designs is that they greatly expand the range of data, as well as the amount of variation in those data, for study. Consequently, they allow for more encompassing explanations and theories that can relate highly divergent outcomes to one another in the same framework. They also contribute to reducing any cultural biases or tendencies toward parochialism among scientists studying common human phenomena. One main vulnerability in such designs arises from the problem of achieving comparability. Because comparative study involves studying societies and other units that are dissimilar from one another, the phenomena under study usually occur in very different contexts—so different that in some cases what is called an event in one society cannot really be regarded as the same type of event in another. For example, a vote in a Western democracy is different from a vote in an Eastern bloc country, and a voluntary vote in the United States means something different from a compulsory vote in Australia. These circumstances make for interpretive difficulties in comparing aggregate rates of voter turnout in different countries. The problem of achieving comparability appears in historical analysis as well. For example, changes in laws and enforcement and recording procedures over time change the definition of what is and what is not a crime, and for that reason it is difficult to compare the crime rates over time. Comparative re- searchers struggle with this problem continually, working to fashion equivalent measures; some have suggested the use of different measures (voting, letters to the editor, street demonstration) in different societies for common variables

178 / The Behavioral and Social Sciences (political participation), to try to take contextual factors into account and to achieve truer comparability. A second vulnerability is controlling variation. Traditional experiments make conscious and elaborate efforts to control the variation of some factors and thereby assess the causal significance of others. In surveys as well as experi- ments, statistical methods are used to control sources of variation and assess suspected causal significance. In comparative and historical designs, this kind of control is often difficult to attain because the sources of variation are many and the number of cases few. Scientists have made efforts to approximate such control in these cases of "many variables, small N." One is the method of paired comparisons. If an investigator isolates 15 American cities in which racial violence has been recurrent in the past 30 years, for example, it is helpful to match them with IS cities of similar population size, geographical region, and size of minorities- such characteristics are controls—and then search for sys- tematic differences between the two sets of cities. Another method is to select, for comparative purposes, a sample of societies that resemble one another in certain critical ways, such as size, common language, and common level of development, thus attempting to hold these factors roughly constant, and then seeking explanations among other factors in which the sampled societies differ from one another. Ethnographic Designs Traditionally identified with anthropology, ethnographic research designs are playing increasingly significant roles in most of the behavioral and social sciences. The core of this methodology is participant-observation, in which a researcher spends an extended period of time with the group under study, ideally mastering the local language, dialect, or special vocabulary, and partic- ipating in as many activities of the group as possible. This kind of participant- observation is normally coupled with extensive open-ended interviewing, in which people are asked to explain in depth the rules, norms, practices, and beliefs through which (from their point of view) they conduct their lives. A principal aim of ethnographic study is to discover the premises on which those rules, norms, practices, and beliefs are built. The use of ethnographic designs by anthropologists has contributed signif- icantly to the building of knowledge about social and cultural variation. And while these designs continue to center on certain long-standing features— extensive face-to-face experience in- the community, linguistic competence, participation, and open-ended interviewing- there are newer trends in eth- nographic work. One major trend concerns its scale. Ethnographic methods were originally developed largely for studying small-scale groupings known variously as village, folk, primitive, preliterate, or simple societies. Over the decades, these methods have increasingly been applied to the study of small

Methods of Data Collection, Representation, and Analysis / 179 groups and networks within modern (urban, industrial, complex) society, in- cluding the contemporary United States. The typical subjects of ethnographic study in modern society are small groups or relatively small social networks, such as outpatient clinics, medical schools, religious cults and churches, ethn- ically distinctive urban neighborhoods, corporate offices and factories, and government bureaus and legislatures. As anthropologists moved into the study of modern societies, researchers in other disciplines particularly sociology, psychology, and political science- began using ethnographic methods to enrich and focus their own insights and findings. At the same time, studies of large-scale structures and processes have been aided by the use of ethnographic methods, since most large-scale changes work their way into the fabric of community, neighborhood, and family, af- fecting the daily lives of people. Ethnographers have studied, for example, the impact of new industry and new forms of labor in "backward" regions; the impact of state-level birth control policies on ethnic groups; and the impact on residents in a region of building a dam or establishing a nuclear waste dump. Ethnographic methods have also been used to study a number of social pro- cesses that lend themselves to its particular techniques of observation and interview—processes such as the formation of class and racial identities, bu- reaucratic behavior, legislative coalitions and outcomes, and the formation and shifting of consumer tastes. Advances in structured interviewing (see above) have proven especially pow- erful in the study of culture. Techniques for understanding kinship systems, concepts of disease, color terminologies, ethnobotany, and ethnozoology have been radically transformed and strengthened by coupling new interviewing methods with modem measurement and scaling techniques (see below). These techniques have made possible more precise comparisons among cultures and identification of the most competent and expert persons within a culture. The next step is to extend these methods to study the ways in which networks of propositions (such as boys like sports, girls like babies) are organized to form belief systems. Much evidence suggests that people typically represent the world around them by means of relatively complex cognitive models that in- volve interlocking propositions. The techniques of scaling have been used to develop models of how people categorize objects, and they have great potential for further development, to analyze data pertaining to cultural propositions. Ideological Systems Perhaps the most fruitful area for the application of ethnographic methods in recent years has been the systematic study of ideologies in modern society. Earlier studies of ideology were in small-scale societies that were rather ho- mogeneous. In these studies researchers could report on a single culture, a uniform system of beliefs and values for the society as a whole. Modern societies are much more diverse both in origins and number of subcultures, related to

180 / The Behavioral and Social Sciences different regions, communities, occupations, or ethnic groups. Yet these sub- cultures and ideologies share certain underlying assumptions or at least must find some accommodation with the dominant value and belief systems in the society. The challenge is to incorporate this greater complexity of structure and process into systematic descriptions and interpretations. One line of work carried out by researchers has tried to track the ways in which ideologies are created, transmitted, and shared among large populations that have tradition- ally lacked the social mobility and communications technologies of the West. This work has concentrated on large-scale civilizations such as China, India, and Central America. Gradually, the focus has generalized into a concern with the relationship between the great traditions—the central lines of cosmopolitan Confucian, Hindu, or Mayan culture, including aesthetic standards, irrigation technologies, medical systems, cosmologies and calendars, legal codes, poetic genres, and religious doctrines and rites and the little traditions, those iden- tified with rural, peasant communities. How are the ideological doctrines and cultural values of the urban elites, the great traditions, transmitted to local communities? How are the little traditions, the ideas from the more isolated, less literate, and politically weaker groups in society, transmitted to the elites? India and southern Asia have been fruitful areas for ethnographic research on these questions. The great Hindu tradition was present in virtually all local contexts through the presence of high-caste individuals in every community. It operated as a pervasive standard of value for all members of society, even in the face of strong little traditions. The situation is surprisingly akin to that of modern, industrialized societies. The central research questions are the degree and the nature of penetration of dominant ideology, even in groups that appear marginal and subordinate and have no strong interest in sharing the dominant value system. In this connection the lowest and poorest occupational caste— the untouchables- serves as an ultimate test of the power of ideology and cultural beliefs to unify complex hierarchical social systems. Historical Reconstruction Another current trend in ethnographic methods is its convergence with archival methods. One joining point is the application of descriptive and in- terpretative procedures used by ethnographers to reconstruct the cultures that created historical documents, diaries, and other records, to interview history, so to speak. For example, a revealing study showed how the Inquisition in the Italian countryside between the 1570s and 1640s gradually worked subtle changes in an ancient fertility cult in peasant communities; the peasant beliefs and rituals assimilated many elements of witchcraft after learning them from their persecutors. A good deal of social history particularly that of the fam- ily has drawn on discoveries made in the ethnographic study of primitive societies. As described in Chapter 4, this particular line of inquiry rests on a marriage of ethnographic, archival, and demographic approaches.

Methods of Data Collection, Representation, and Analysis / 181 Other lines of ethnographic work have focused on the historical dimensions of nonliterate societies. A strikingly successful example in this kind of effort is a study of head-hunting. By combining an interpretation of local oral tradition with the fragmentary observations that were made by outside observers (such as missionaries, traders, colonial officials), historical fluctuations in the rate and significance of head-hunting were shown to be partly in response to such international forces as the great depression and World War II. Researchers are also investigating the ways in which various groups in contemporary societies invent versions of traditions that may or may not reflect the actual history of the group. This process has been observed among elites seeking political and cultural legitimation and among hard-pressed minorities (for example, the Basque in Spain, the Welsh in Great Britain) seeking roots and political mo- . .1 . . . olllzatlon in a arger society. Ethnography is a powerful method to record, describe, and interpret the system of meanings held by groups and to discover how those meanings affect the lives of group members. It is a method well adapted to the study of situations in which people interact with one another and the researcher can interact with them as well, so that information about meanings can be evoked and observed. Ethnography is especially suited to exploration and elucidation of unsuspected connections; ideally, it is used in combination with other methods—experi- mental, survey, or comparative to establish with precision the relative strengths and weaknesses of such connections. By the same token, experimental, survey, and comparative methods frequently yield connections, the meaning of which is unknown; ethnographic methods are a valuable way to determine them. MODELS FOR REPRESENTING PHENOMENA The objective of any science is to uncover the structure and dynamics of the phenomena that are its subject, as they are exhibited in the data. Scientists continuously try to describe possible structures and ask whether the data can, with allowance for errors of measurement, be described adequately in terms of them. Over a long time, various families of structures have recurred throughout many fields of science; these structures have become objects of study in their own right, principally by statisticians, other methodological specialists, applied mathematicians, and philosophers of logic and science. Methods have evolved to evaluate the adequacy of particular structures to account for particular types of data. In the interest of clarity we discuss these structures in this section and the analytical methods used for estimation and evaluation of them in the next section, although in practice they are closely intertwined. A good deal of mathematical and statistical modeling attempts to describe the relations, both structural and dynamic, that hold among variables that are presumed to be representable by numbers. Such models are applicable in the behavioral and social sciences only to the extent that appropriate numerical

182 / The Behavioral and Social Sciences measurement can be devised for the relevant variables. In many studies the phenomena in question and the raw data obtained are not intrinsically nu- merical, but qualitative, such as ethnic group identifications. The identifying numbers used to code such questionnaire categories for computers are no more than labels, which could just as well be letters or colors. One key question is whether there is some natural way to move from the qualitative aspects of such data to a structural representation that involves one of the well-understood numerical or geometric models or whether such an attempt would be inherently inappropriate for the data in question. The decision as to whether or not particular empirical data can be represented in particular numerical or more complex structures is seldom simple, and strong intuitive biases or a priori assumptions about what can and cannot be done may be misleading. Recent decades have seen rapid and extensive development and application of analytical methods attuned to the nature and complexity of social science data. Examples of nonnumerical modeling are increasing. Moreover, the wide- spread availability of powerful computers is probably leading to a qualitative revolution, it is affecting not only the ability to compute numerical solutions to numerical models, but also to work out the consequences of all sorts of structures that do not involve numbers at all. The following discussion gives some indication of the richness of past progress and of future prospects al- though it is by necessity far from exhaustive. In describing some of the areas of new and continuing research, we have organized this section on the basis of whether the representations are funda- mentally probabilistic or not. A further useful distinction is between represen- tations of data that are highly discrete or categorical in nature (such as whether a person is male or female) and those that are continuous in nature (such as a person's height). Of course, there are intermediate cases involving both types of variables, such as color stimuli that are characterized by discrete hues (red, green) and a continuous luminance measure. Probabilistic models lead very naturally to questions of estimation and statistical evaluation of the correspon- dence between data and model. Those that are not probabilistic involve addi- tional problems of dealing with and representing sources of variability that are not explicitly modeled. At the present time, scientists understand some aspects of structure, such as geometries, and some aspects of randomness, as embodied in probability models, but do not yet adequately understand how to put the two together in a single unibed model. Table 5-1 outlines the way we have organized this discussion and shows where the examples in this section lie. Probability Models Some behavioral and social sciences variables appear to be more or less continuous, for example, utility of goods, loudness of sounds, or risk associated with uncertain alternatives. Many other variables, however, are inherently cat-

Methods of Data Collection, Representation, and Analysis / 183 TABLE S- 1 A Classification of Structural Models Nature of the Variables Nature of the Representation Categorical Continuous Probabilistic Log-linear and Multi-item related models measurement Event histories Nonlinear, nonadditive models Geometric and Clustering Scaling algebraic Network models Ordered factorial systems egorical, often with only two or a few values possible: for example, whether a person is in or out of school, employed or not employed, identifies with a major political party or political ideology. And some variables, such as moral attitudes, are typically measured in research with survey questions that allow only categorical responses. Much of the early probability theory was formulated only for continuous variables; its use with categorical variables was not really justified, and in some cases it may have been misleading. Recently, very sig- nificant advances have been made in how to deal explicitly with categorical variables. This section first describes several contemporary approaches to models involving categorical variables, followed by ones involving continuous repre- sentations. Log-Linear Models for Categorical Variables Many recent models for analyzing categorical data of the kind usually dis- played as counts (cell frequencies) in multidimensional contingency tables are subsumed under the general heading of log-linear models, that is, linear models in the natural logarithms of the expected counts in each cell in the table. These recently developed forms of statistical analysis allow one to partition variability due to various sources in the distribution of categorical attributes, and to isolate the effects of particular variables or combinations of them. Present log-linear models were first developed and used by statisticians and sociologists and then found extensive application in other social and behavioral sciences disciplines. When applied, for instance, to the analysis of social mo- bility, such models separate factors of occupational supply and demand from other factors that impede or propel movement up and down the social hier- archy. With such models, for example, researchers discovered the surprising fact that occupational mobility patterns are strikingly similar in many nations of the world (even among disparate nations like the United States and most of the Eastem European socialist countries), and from one time period to another, once allowance is made for differences in the distributions of occupations. The

184 / The Behavioral and Social Sciences log-linear and related kinds of models have also made it possible to identify and analyze systematic differences in mobility among nations and across time. As another example of applications, psychologists and others have used log- linear models to analyze attitudes and their determinants and to link attitudes to behavior. These methods have also diffused to and been used extensively in the medical and biological sciences. Regression Modelsfor Categorical Variables Models that permit one variable to be explained or predicted by means of others, called regression models, are the workhorses of much applied statistics; this is especially true when the dependent (explained) variable is continuous. For a two-valued dependent variable, such as alive or dead, models and ap- proximate theory and computational methods for one explanatory variable were developed in biometry about 50 years ago. Computer programs able to handle many explanatory variables, continuous or categorical, are readily avail- able today. Even now, however, the accuracy of the approximate theory on . given c .ata IS an open question. Using classical utility theory, economists have developed discrete choice models that turn out to be somewhat related to the log-linear and categorical regression models. Models for limited dependent variables, especially those that cannot take on values above or below a certain level (such as weeks unemployed, number of children, and years of schooling) have been used profitably in economics and in some other areas. For example, censored normal variables (called tobits in economics), in which observed values outside certain limits are simply counted, have been used in studying decisions to go on in school. It will require further research and development to incorporate infor- mation about limited ranges of variables fully into the main multivariate meth- odologies. In addition, with respect to the assumptions about distribution and functional form conventionally made in discrete response models, some new methods are now being developed that show promise of yielding reliable in- ferences without making unrealistic assumptions; further research in this area . ~ promises slgnl~cant progress. One problem arises from the fact that many of the categorical variables collected by the major data bases are ordered. For example, attitude surveys frequently use a 3-, 5-, or 7-point scale (from high to low) without specifying numerical intervals between levels. Social class and educational levels are often described by ordered categories. Ignoring order information, which many tra- ditional statistical methods do, may be inefficient or inappropriate, but replac- ing the categories by successive integers or other arbitrary scores may distort the results. (For additional approaches to this question, see sections below on ordered structures.) Regression-like analysis of ordinal categorical variables is quite well developed, but their multivariate analysis needs further research. New log-bilinear models have been proposed, but to date they deal specifically

Methods of Data Collection, Representation, and Analysis / 18S with only two or three categorical variables. Additional research extending the new models, improving computational algorithms, and integrating the models with work on scaling promise to lead to valuable new knowledge. Models for Event Histories Event-history studies yield the sequence of events that respondents to a survey sample experience over a period of time; for example, the timing of marriage, childbearing, or labor force participation. Event-history data can be used to study educational progress, demographic processes (migration, fertility, and mortality), mergers of firms, labor market behavior? and even riots, strikes, and revolutions. As interest in such data has grown, many researchers have turned to models that pertain to changes in probabilities over time to describe when and how individuals move among a set of qualitative states. Much of the progress in models for event-history data builds on recent developments in statistics and biostatistics for life-time, failure-time, and haz- ard models. Such models permit the analysis of qualitative transitions in a population whose members are undergoing partially random organic deterio- ration, mechanical wear, or other risks over time. With the increased com- plexity of event-history data that are now being collected, and the extension of event-history data bases over very long periods of time, new problems arise that cannot be effectively handled by older types of analysis. Among the prob- lems are repeated transitions, such as between unemployment and employment or marriage and divorce; more than one time variable (such as biological age, calendar time, duration in a stage, and time exposed to some specified con- dition); latent variables (variables that are explicitly modeled even though not observed); gaps in the data; sample attrition that is not randomly distributed over the categories; and respondent difficulties in recalling the exact timing of events. Models for Multiple-Item Measurement For a variety of reasons, researchers typically use multiple measures (or multiple indicators) to represent theoretical concepts. Sociologists, for example, often rely on two or more variables (such as occupation and education) to measure an individual's socioeconomic position; educational psychologists or- dinarily measure a student's ability with multiple test items. Despite the fact that the basic observations are categorical, in a number of applications this is interpreted as a partitioning of something continuous. For example, in test theory one thinks of the measures of both item difficulty and respondent ability as continuous variables, possibly multidimensional in character. Classical test theory and newer item-response theories in psychometrics deal with the extraction of information from multiple measures. Testing, which is a major source of data in education and other areas, results in millions of test

186 / The Behavioral and Social Sciences items stored in archives each year for purposes ranging from college admissions to job-training programs for industry. One goal of research on such test data is to be able to make comparisons among persons or groups even when different test items are used. Although the information collected from each respondent is intentionally incomplete in order to keep the tests short and simple, item- response techniques permit researchers to reconstitute the fragments into an accurate picture of overall group proficiencies. These new methods provide a better theoretical handle on individual differences, and they are expected to be extremely important in developing and using tests. For example, they have been used in attempts to equate different forms of a test given in successive waves during a year, a procedure made necessary in large-scale testing programs by legislation requiring disclosure of test-scoring keys at the time results are given. An example of the use of item-response theory in a significant research effort is the National Assessment of Educational Progress (NAEP). The goal of this project is to provide accurate, nationally representative information on the average (rather than individual) proficiency of American children in a wide variety of academic subjects as they progress through elementary and secondary school. This approach is an improvement over the use of trend data on uni- versity entrance exams, because NAEP estimates of academic achievements (by broad characteristics such as age, grade, region, ethnic background, and so on) are not distorted by the self-selected character of those students who seek admission to college, graduate, and professional programs. Item-response theory also forms the basis of many new psychometric in- struments, known as computerized adaptive testing, currently being imple- mented by the U.S. military services and under additional development in many testing organizations. In adaptive tests, a computer program selects items for each examiner based upon the examinee's success with previous items. Gen- erally, each person gets a slightly different set of items and the equivalence of scale scores is established by using item-response theory. Adaptive testing can greatly reduce the number of items needed to achieve a given level of meas- urement accuracy. Nonlinear, Nonadditive Models Virtually all statistical models now in use impose a linearity or additivity assumption of some kind, sometimes after a nonlinear transformation of var- iables. Imposing these forms on relationships that do not, in fact, possess them may well result in false descriptions and spurious effects. Unwary users, es- pecially of computer software packages, can easily be misled. But more realistic nonlinear and nonadditive multivariate models are becoming available. Exten- sive use with empirical data is likely to force many changes and enhancements in such models and stimulate quite different approaches to nonlinear multi- variate analysis in the next decade.

Methods of Data Collection, Representation, and Analysis / 187 Geometric and Algebraic Models Geometric and algebraic models attempt to describe underlying structural relations among variables. In some cases they are part of a probabilistic ap- proach, such as the algebraic models underlying regression or the geometric representations of correlations between items in a technique called factor anal- ysis. In other cases, geometric and algebraic models are developed without explicitly modeling the element of randomness or uncertainty that is always present in the data. Although this latter approach to behavioral and social sciences problems has been less researched than the probabilistic one, there are some advantages in developing the structural aspects independent of the statistical ones. We begin the discussion with some inherently geometric rep- resentations and then turn to numerical representations for ordered data. Although geometry is a huge mathematical topic, little of it seems directly applicable to the kinds of data encountered in the behavioral and social sci- ences. A major reason is that the primitive concepts normally used in geome- try points, lines, coincidence—do not correspond naturally to the kinds of qualitative observations usually obtained in behavioral and social sciences con- texts. Nevertheless, since geometric representations are used to reduce bodies of data, there is a real need to develop a deeper understanding of when such representations of social or psychological data make sense. Moreover, there is a practical need to understand why geometric computer algorithms, such as those of multidimensional scaling, work as well as they apparently do. A better understanding of the algorithms will increase the efficiency and appropriate- ness of their use, which becomes increasingly important with the widespread availability of scaling programs for microcomputers. Scaling Over the past 50 years several kinds of well-understood scaling techniques have been developed and widely used to assist in the search for appropriate geometric representations of empirical data. The whole field of scaling is now entering a critical juncture in terms of unifying and synthesizing what earlier appeared to be disparate contributions. Within the past few years it has become apparent that several major methods of analysis, including some that are based on probabilistic assumptions, can be unified under the rubric of a single gen- eralized mathematical structure. For example, it has recently been demon- strated that such diverse approaches as nonmetric multidimensional scaling, principal-components analysis, factor analysis, correspondence analysis, and log-linear analysis have more in common in terms of underlying mathematical structure than had earlier been realized. Nonmetric multidimensional scaling is a method that begins with data about the ordering established by subjective similarity (or nearness) between pairs of stimuli. The idea is to embed the stimuli into a metric space (that is, a geometry

188 / The Behavioral and Social Sciences with a measure of distance between points) in such a way that distances between points corresponding to stimuli exhibit the same ordering as do the data. This method has been successfully applied to phenomena that, on other grounds, are known to be describable in terms of a specific geometric structure; such applications were used to validate the procedures. Such validation was done, for example, with respect to the perception of colors, which are known to be describable in terms of a particular three-dimensional structure known as the Euclidean color coordinates. Similar applications have been made with Morse code symbols and spoken phonemes. The technique is now used in some biological and engineering applications, as well as in some of the social sciences, as a method of data exploration and simplification. One question of interest is how to develop an axiomatic basis for various geometries using as a primitive concept an observable such as the subject's ordering of the relative similarity of one pair of stimuli to another, which is the typical starting point of such scaling. The general task is to discover properties of the qualitative data sufficient to ensure that a mapping into the geometric structure exists and, ideally, to discover an algorithm for finding it. Some work of this general type has been carried out: for example, there is an elegant set of axioms based on laws of color matching that yields the three-dimensional vectorial representation of color space. But the more general problem of un- derstanding the conditions under which the multidimensional scaling algo- rithms are suitable remains unsolved. In addition, work is needed on under- standing more general, non-Euclidean spatial models. Ordered Factorial Systems One type of structure common throughout the sciences arises when an ordered dependent variable is affected by two or more ordered independent variables. This is the situation to which regression and analysis-of-variance models are often applied; it is also the structure underlying the familiar physical identities, in which physical units are expressed as products of the powers of other units (for example, energy has the unit of mass times the square of the unit of distance divided by the square of the unit of time). There are many examples of these types of structures in the behavioral and social sciences. One example is the ordering of preference of commodity bun- dles collections of various amounts of commodities which may be revealed directly by expressions of preference or indirectly by choices among alternative sets of bundles. A related example is preferences among alternative courses of action that involve various outcomes with differing degrees of uncertainty; this is one of the more thoroughly investigated problems because of its potential importance in decision making. A psychological example is the trade-off be- tween delay and amount of reward, yielding those combinations that are equally reinforcing. In a common, applied kind of problem, a subject is given descrip- tions of people in terms of several factors, for example, intelligence, creativity,

Methods of Data Collection, Representation, and Analysis / 189 diligence, and honesty, and is asked to rate them according to a criterion such as suitability for a particular job. In all these cases and a myriad of others like them the question is whether the regularities of the data permit a numerical representation. Initially, three types of representations were studied quite fully: the dependent variable as a sum, a product, or a weighted average of the measures associated with the independent variables. The first two representations underlie some psycholog- ical and economic investigations, as well as a considerable portion of physical measurement and modeling in classical statistics. The third representation, averaging, has proved most useful in understanding preferences among un- certain outcomes and the amalgamation of verbally described traits, as well as some physical variables. For each of these three cases adding, multiplying, and averaging re- searchers know what properties or axioms of order the data must satisfy for such a numerical representation to be appropriate. On the assumption that one or another of these representations exists, and using numerical ratings by sub- jects instead of ordering, a scaling technique called functional measurement (referring to the function that describes how the dependent variable relates to the independent ones) has been developed and applied in a number of domains. What remains problematic is how to encompass at the ordinal level the fact that some random error intrudes into nearly all observations and then to show how that randomness is represented at the numerical level; this continues to be an unresolved and challenging research issue. During the past few years considerable progress has been made in under- standing certain representations inherently different from those just discussed. The work has involved three related thrusts. The first is a scheme of classifying structures according to how uniquely their representation is constrained. The three classical numerical representations are known as ordinal, interval, and ratio scale types. For systems with continuous numerical representations and of scale type at least as rich as the ratio one, it has been shown that only one additional type can exist. A second thrust is to accept structural assumptions, like factorial ones, and to derive for each scale the possible functional relations among the independent variables. And the third thrust is to develop axioms for the properties of an order relation that leads to the possible representations. Much is now known about the possible nonadditive representations of both the muItifactor case and the one where stimuli can be combined, such as . . . . . . . com fining sounc . intensities. Closely related to this classification of structures is the question: What state- ments, formulated in terms of the measures arising in such representations, can be viewed as meaningful in the sense of corresponding to something em- pirical? Statements here refer to any scientific assertions, including statistical ones, formulated in terms of the measures of the variables and logical and mathematical connectives. These are statements for which asserting truth or /

190 / The Behavioral and Social Sciences falsity makes sense. In particular, statements that remain invariant under certain symmetries of structure have played an important role in classical geometry, dimensional analysis in physics, and in relating measurement and statistical models applied to the same phenomenon. In addition, these ideas have been used to construct models in more formally developed areas of the behavioral and social sciences, such as psychophysics. Current research has emphasized the communality of these historically independent developments and is at- tempting both to uncover systematic, philosophically sound arguments as to why invariance under symmetries is as important as it appears to be and to understand what to do when structures lack symmetry, as, for example, when variables have an inherent upper bound. Clustering Many subjects do not seem to be correctly represented in terms of distances in continuous geometric space. Rather, in some cases, such as the relations among meanings of words which is of great interest in the study of memory representations a description in terms of tree-like, hierarchial structures ap- pears to be more illuminating. This kind of description appears appropriate both because of the categorical nature of the judgments and the hierarchial, rather than trade-off, nature of the structure. Individual items are represented as the terminal nodes of the tree, and groupings by different degrees of similarity are shown as intermediate nodes, with the more general groupings occurring nearer the root of the tree. Clustering techniques, requiring considerable com- putational power, have been and are being developed. Some successful appli- cations exist, but much more refinement is anticipated. Network Models Several other lines of advanced modeling have progressed in recent years, opening new possibilities for empirical specification and testing of a variety of theories. In social network data, relationships among units, rather than the units themselves, are the primary objects of study: friendships among persons, trade ties among nations, cocitation clusters among research scientists, inter- locking among corporate boards of directors. Special models for social network data have been developed in the past decade, and they give, among other things, precise new measures of the strengths of relational ties among units. A major challenge in social network data at present is to handle the statistical depend- ence that arises when the units sampled are related in complex ways. STATISTICAL INFERENCE AND ANALYSIS As was noted earlier, questions of design, representation, and analysis are intimately intertwined. Some issues of inference and analysis have been dis-

Methods of Data Collection, Representation, and Analysis / 191 cussed above as related to specific data collection and modeling approaches. This section discusses some more general issues of statistical inference and advances in several current approaches to them. Causal Inference Behavioral and social scientists use statistical methods primarily to infer the effects of treatments, interventions, or policy factors. Previous chapters in- cluded many instances of causal knowledge gained this way. As noted above, the large experimental study of alternative health care financing discussed in Chapter 2 relied heavily on statistical principles and techniques, including randomization, in the design of the experiment and the analysis of the resulting data. Sophisticated designs were necessary in order to answer a variety of questions in a single large study without confusing the effects of one program difference (such as prepayment or fee for service) with the effects of another (such as different levels of deductible costs), or with effects of unobserved variables (such as genetic differences). Statistical techniques were also used to ascertain which results applied across the whole enrolled population and which were confined to certain subgroups (such as individuals with high blood pres- sure) and to translate utilization rates across different programs and types of patients into comparable overall dollar costs and health outcomes for alternative financing options. A classical experiment, with systematic but randomly assigned variation of the variables of interest (or some reasonable approach to this), is usually con- sidered the most rigorous basis from which to draw such inferences. But ran- dom samples or randomized experimental manipulations are not always fea- sible or ethically acceptable. Then, causal inferences must be drawn from observational studies, which, however well designed, are less able to ensure that the observed (or inferred) relationships among variables provide clear evidence on the underlying mechanisms of cause and effect. Certain recurrent challenges have been identified in studying causal infer- ence. One challenge arises from the selection of background variables to be measured, such as the sex, nativity, or parental religion of individuals in a comparative study of how education affects occupational success. The adequacy of classical methods of matching groups in background variables and adjusting for covariates needs further investigation. Statistical adjustment of biases linked to measured background variables is possible, but it can become complicated. Current work in adjustment for selectivity bias is aimed at weakening implau- sible assumptions, such as normality, when carrying out these adjustments. Even after adjustment has been made for the measured background variables, other, unmeasured variables are almost always still affecting the results (such as family transfers of wealth or reading habits). Analyses of how the conclusions might change if such unmeasured variables could be taken into account is

192 / The Behavioral and Social Sciences essential in attempting to make causal inferences from an observational study, and systematic work on useful statistical models for such sensitivity analyses is just beginning. The third important issue arises from the necessity for distinguishing among competing hypotheses when the explanatory variables are measured with dif- ferent degrees of precision. Both the estimated size and significance of an effect are diminished when it has large measurement error, and the coefficients of other correlated variables are affected even when the other variables are meas- ured perfectly. Similar results arise from conceptual errors, when one measures only proxies for a theoretical construct (such as years of education to represent amount of learning). In some cases, there are procedures for simultaneously or iteratively estimating both the precision of complex measures and their effect . . On a particu tar criterion. Although complex models are often necessary to infer causes, once their output is available, it should be translated into understandable displays for evaluation Results that depend on the accuracy of a multivariate model and the associated software need to be subjected to appropriate checks, including the evaluation of graphical displays, group comparisons, and other analyses. New Statistical Techniques Internal Resampling One of the great contributions of twentieth-century statistics was to dem- onstrate how a properly drawn sample of sufficient size, even if it is only a tiny fraction of the population of interest, can yield very good estimates of most population characteristics. When enough is known at the outset about the characteristic in question for example, that its distribution is roughly nor- mal inference from the sample data to the population as a whole is straight- forward, and one can easily compute measures of the certainty of inference, a common example being the 9S percent confidence interval around an estimate. But population shapes are sometimes unknown or uncertain, and so inference procedures cannot be so simple. Furthermore, more often than not, it is difficult to assess even the degree of uncertainty associated with complex data and with the statistics needed to unravel complex social and behavioral phenomena. Internal resampling methods attempt to assess this uncertainty by generating a number of simulated data sets similar to the one actually observed. The definition of similar is crucial, and many methods that exploit different types of similarity have been devised. These methods provide researchers the freedom to choose scientifically appropriate procedures and to replace procedures that are valid under assumed distributional shapes with ones that are not so re- stricted. Flexible and imaginative computer simulation is. the key to these methods. For a simple random sample, the "bootstrap" method repeatedly resamples the obtained data (with replacement) to generate a distribution of

Methods of Data Collection, Representation, and Analysis / 193 possible data sets. The distribution of any estimator can thereby be simulated and measures of the certainty of inference be derived. The "jackknife" method repeatedly omits a fraction of the data and in this way generates a distribution of possible data sets that can also be used to estimate variability. These methods can also be used to remove or reduce bias. For example, the ratio-estimator, a statistic that is commonly used in analyzing sample surveys and censuses, is known to be biased, and the jackknife method can usually remedy this defect. The methods have been extended to other situations and types of analysis, such as multiple regression. There are indications that under relatively general conditions, these methods, and others related to them, allow more accurate estimates of the uncertainty of inferences than do the traditional ones that are based on assumed (usually, normal) distributions when that distributional assumption is unwarranted. For complex samples, such internal resampling or subsampling facilitates estimat- ing the sampling variances of complex statistics. An older and simpler, but equally important, idea is to use one independent subsample in searching the data to develop a model and at least one separate subsample for estimating and testing a selected model. Otherwise, it is next to impossible to make allowances for the excessively close fitting of the model that occurs as a result of the creative search for the exact characteristics of the sample data characteristics that are to some degree random and will not predict well to other samples. Robust Techniques Many technical assumptions underlie the analysis of data. Some, like the assumption that each item in a sample is drawn independently of other items, can be weakened when the data are sufficiently structured to admit simple alternative models, such as serial correlation. Usually, these models require that a few parameters be estimated. Assumptions about shapes of distributions, normality being the most common, have proved to be particularly important, and considerable progress has been made in dealing with the consequences of different assumptions. More recently, robust techniques have been designed that permit sharp, valid discriminations among possible values of parameters of central tendency for a wide variety of alternative distributions by reducing the weight given to oc- casional extreme deviations. It turns out that by giving up, say, 10 percent of the discrimination that could be provided under the rather unrealistic as- sumption of normality, one can greatly improve performance in more realistic situations, especially when unusually large deviations are relatively common. These valuable modifications of classical statistical techniques have been extended to multiple regression, in which procedures of iterative reweighting can now offer relatively good performance for a variety of underlying distri- butional shapes. They should be extended to more general schemes of analysis.

194 / The Behavioral and Social Sciences In some contexts notably the most classical uses of analysis of variance the use of adequate robust techniques should help to bring conventional statistical practice closer to the best standards that experts can now achieve. Many Interrelated Parameters In trying to give a more accurate representation of the real world than is possible with simple models, researchers sometimes use models with many parameters, all of which must be estimated from the data. Classical principles of estimation, such as straightforward maximum-likelihood, do not yield re- liable estimates unless either the number of observations is much larger than the number of parameters to be estimated or special designs are used in con- junction with strong assumptions. Bayesian methods do not draw a distinction between fixed and random parameters, and so may be especially appropriate for such problems. A variety of statistical methods have recently been developed that can be interpreted as treating many of the parameters as or similar to random quan- tities, even if they are regarded as representing fixed quantities to be estimated. Theory and practice demonstrate that such methods can improve the simpler fixed-parameter methods from which they evolved, especially when the num- ber of observations is not large relative to the number of parameters. Successful applications include college and graduate school admissions, where quality of previous school is treated as a random parameter when the data are insufficient to separately estimate it well. Efforts to create appropriate models using this general approach for small-area estimation and underc.ount adjustment in the census are important potential applications. Missing Data In data analysis, serious problems can arise when certain kinds of (quanti- tative or qualitative) information is partially or wholly missing. Various ap- proaches to dealing with these problems have been or are being developed. One of the methods developed recently for dealing with certain aspects of missing data is called multiple imputation: each missing value in a data set is replaced by several values representing a range of possibilities, with statistical dependence among missing values reflected by linkage among their replace- ments. It is currently being used to handle a major problem of incompatibility between the 1980 and previous Bureau of Census public-use tapes with respect to occupation codes. The extension of these techniques to address such prob- lems as nonresponse to income questions in the Current Population Survey has been examined in exploratory applications with great promise. Computing Computer Packages and Expert Systems The development of high-speed computing and data handling has funda- mentally changed statistical analysis. Methodologies for all kinds of situations

Methods of Data Collection, Representation, and Analysis / l9S are rapidly being developed and made available for use in computer packages that may be incorporated into interactive expert systems. This computing ca- pability offers the hope that much data analyses will be more carefully and more effectively done than previously and that better strategies for data analysis will move from the practice of expert statisticians, some of whom may not have tried to articulate their own strategies, to both wide discussion and general use. But powerful tools can be hazardous, as witnessed by occasional dire misuses of existing statistical packages. Until recently the only strategies available were to train more expert methodologists or to train substantive scientists in more methodology, but without the updating of their training it tends to become outmoded. Now there is the opportunity to capture in expert systems the current best methodological advice and practice. If that opportunity is ex- ploited, standard methodological training of social scientists will shift to em- phasizing strategies in using good expert systems - including understanding the nature and importance of the comments it provides rather than in how to patch together something on one's own. With expert systems, almost all behavioral and social scientists should become able to conduct any of the more common styles of data analysis more effectively and with more confidence than all but the most expert do today. However, the difficulties in developing expert systems that work as hoped for should not be underestimated. Human experts cannot readily explicate all of the complex cognitive network that constitutes an important part of their knowledge. As a result, the first attempts at expert systems were not especially successful (as discussed in Chapter 1~. Additional work is expected to overcome these limitations, but it is not clear how long it will take. Exploratory Analysis and Graphic Presentation The formal focus of much statistics research in the middle half of the twen- tieth century was on procedures to confirm or reject precise, a priori hypotheses developed in advance of collecting data—that is, procedures to determine statistical significance. There was relatively little systematic work on realistically rich strategies for the applied researcher to use when attacking real-world problems with their multiplicity of objectives and sources of evidence. More recently, a species of quantitative detective work, called exploratory data anal- ysis, has received increasing attention. In this approach, the researcher seeks out possible quantitative relations that may be present in the data. The tech- niques are flexible and include an important component of graphic represen- tations. While current techniques have evolved for single responses in situa- tions of modest complexity, extensions to multiple responses and to single responses in more complex situations are now possible. Graphic and tabular presentation is a research domain in active renaissance, stemming in part from suggestions for new kinds of graphics made possible by computer capabilities, for example, hanging histograms and easily assimi- lated representations of numerical vectors. Research on data presentation has

196 / The Behavioral and Social Sciences been carried out by statisticians, psychologists, cartographers, and other spe- cialists, and attempts are now being made to incorporate findings and concepts from linguistics, industrial and publishing design, aesthetics, and classification studies in library science. Another influence has been the rapidly increasing availability of powerful computational hardware and software, now available even on desktop computers. These ideas and capabilities are leading to an increasing number of behavioral experiments with substantial statistical input. Nonetheless, criteria of good graphic and tabular practice are still too much matters of tradition and dogma, without adequate empirical evidence or theo- retical coherence. To broaden the respective research outlooks and vigorously develop such evidence and coherence, extended collaborations between statis- tical and mathematical specialists and other scientists are needed, a major objective being to understand better the visual and cognitive processes (see Chapter 1) relevant to effective use of graphic or tabular approaches. Combining Evidence Combining evidence from separate sources is a recurrent scientific task, and formal statistical methods for doing so go back 30 years or more. These methods include the theory and practice of combining tests of individual hypotheses, sequential design and analysis of experiments, comparisons of laboratories, and Bayesian and likelihood paradigms. There is now growing interest in more ambitious analytical syntheses, which are often called meta-analyses. One stimulus has been the appearance of syntheses explicitly combining all existing investigations in particular fields, such as prison parole policy, classroom size in primary schools, cooperative studies of ther- apeutic treatments for coronary heart disease, early childhood education in- terventions, and weather modification experiments. In such fields, a serious approach to even the simplest question how to put together separate esti- mates of effect size from separate investigations leads quickly to difficult and interesting issues. One issue involves the lack of independence among the available studies, due, for example, to the effect of influential teachers on the research projects of their students. Another issue is selection bias, because only some of the studies carried out, usually those with "significant" findings, are available and because the literature search may not find out all relevant studies that are available. In addition, experts agree, although informally, that the quality of studies from different laboratories and facilities differ appreciably and that such information probably should be taken into account. Inevitably, the studies to be included used different designs and concepts and controlled or measured different variables, making it difficult to know how to combine them. Rich, informal syntheses, allowing for individual appraisal, may be better than catch-all formal modeling, but the literature on formal meta-analytic models

Methods of Data Collection, Representation, and Analysis / 197 is growing and may be an important area of discovery in the next decade, relevant both to statistical analysis per se and to improved syntheses in the behavioral and social and other sciences. OPPORTUNITIES AND NEEDS This chapter has cited a number of methodological topics associated with 1 1 · ~ 1 . ~ . ~ oenav~ora~ and social sciences research that appear to be particularly active and promising at the present time. As throughout the report, they constitute illus- trative examples of what the committee believes to be important areas of re- search in the coming decade. In this section we describe recommendations for an additional $16 million annually to facilitate both the development of meth- odologically oriented research and, equally important, its communication throughout the research community. Methodological studies, including early computer implementations, have for the most part been carried out by individual investigators with small teams of colleagues or students. Occasionally, such research has been associated with quite large substantive projects, and some of the current developments of computer packages, graphics, and expert systems clearly require large, orga- nized efforts, which often lie at the boundary between grant-supported work and commercial development. As such research is often a key to understanding complex bodies of behavioral and social sciences data, it is vital to the health of these sciences that research support continue on methods relevant to prob- lems of modeling, statistical analysis, representation, and related aspects of behavioral and social sciences data. Researchers and funding agencies should also be especially sympathetic to the inclusion of such basic methodological work in large experimental and longitudinal studies. Additional funding for work in this area, both in terms of individual research grants on methodological issues and in terms of augmentation of large projects to include additional methodological aspects, should be provided largely in the form of investigator- initiated project grants. Ethnographic and comparative studies also typically rely on project grants to individuals and small groups of investigators. While this type of support should continue, provision should also be made to facilitate the execution of studies using these methods by research teams and to provide appropriate methodological training through the mechanisms outlined below. Overall, we recommend an increase of $4 million in the level of investigator- initiated grant support for methodological work. An additional $1 million should be devoted to a program of centers for methodological research. Many of the new methods and models described in the chapter, if and when adopted to any large extent, will demand substantially greater amounts of research devoted to appropriate analysis and computer implementation. New

198 / The Behavioral and Social Sciences user interfaces and numerical algorithms will need to be designed and new computer programs written. And even when generally available methods (such as maximum-likelihood) are applicable, model application still requires skillful development in particular contexts. Many of the familiar general methods that are applied in the statistical analysis of data are known to provide good ap- proximations when sample sizes are sufficiently large, but their accuracy varies with the specific model and data used. To estimate the accuracy requires ex- tensive numerical exploration. Investigating the sensitivity of results to the assumptions of the models is important and requires still more creative, thoughtful research. It takes substantial efforts of these kinds to bring any new model on line, and the need becomes increasingly important and difficult as statistical models move toward greater realism, usefulness, complexity, and availability in computer form. More complexity in turn will increase the demand for com- putational power. Although most of this demand can be satisfied by increas- ingly powerful desktop computers, some access to mainframe and even su- percomputers will be needed in selected cases. We recommend an additional $4 million annually to cover the growth in computational demands for model development and testing. Interaction and cooperation between the developers and the users of statis- tical and mathematical methods need continual stimulation both ways. Ef- forts should be made to teach new methods to a wider variety of potential users than is now the case. Several ways appear effective for methodologists to com- municate to empirical scientists: running summer training programs for grad- uate students, faculty, and other researchers; encouraging graduate students, perhaps through degree requirements, to make greater use of the statistical, mathematical, and methodological resources at their own or affiliated univer- sities; associating statistical and mathematical research specialists with large- scale data collection projects; and developing statistical packages that incor- porate expert systems in applying the methods. Methodologists, in turn, need to become more familiar with the problems actually faced by empirical scientists in the laboratory and especially in the field. Several ways appear useful for communication in this direction: encour- aging graduate students in methodological specialties, perhaps through degree requirements, to work directly on empirical research; creating postdoctoral fellowships aimed at integrating such specialists into ongoing data collection projects; and providing for large data collection projects to engage relevant methodological specialists. In addition, research on and development of sta- tistical packages and expert systems should be encouraged to involve the mul- tidisciplinary collaboration of experts with experience in statistical, computer, . . . anc ~ cognitive sciences. A final point has to do with the promise held out by bringing different research methods to bear on the same problems. As our discussions of research methods in this and other chapters have emphasized, different methods have

Methods of Data Collection, Representation, and Analysis / 199 different powers and limitations, and each is designed especially to elucidate one or more particular facets of a subject. An important type of interdisciplinary work is the collaboration of specialists in different research methodologies on a substantive issue, examples of which have been noted throughout this report. If more such research were conducted cooperatively, the power of each method pursued separately would be increased. To encourage such multidisciplinary work, we recommend increased support for fellowships, research workshops, anc . tramlug institutes. Funding for fellowships, both pre- and postdoctoral, should be aimed at giving methodologists experience with substantive problems and at upgrading the methodological capabilities of substantive scientists. Such targeted fellow- ship support should be increased by $4 million annually, of which $3 million should be for predoctoral fellowships emphasizing the enrichment of meth- odological concentrations. The new support needed for research workshops is estimated to be $1 million annually. And new support needed for various kinds of advanced training institutes aimed at rapidly diffusing new methodological findings among substantive scientists is estimated to be $2 million annually.

This volume explores the scientific frontiers and leading edges of research across the fields of anthropology, economics, political science, psychology, sociology, history, business, education, geography, law, and psychiatry, as well as the newer, more specialized areas of artificial intelligence, child development, cognitive science, communications, demography, linguistics, and management and decision science. It includes recommendations concerning new resources, facilities, and programs that may be needed over the next several years to ensure rapid progress and provide a high level of returns to basic research.

READ FREE ONLINE

Welcome to OpenBook!

You're looking at OpenBook, NAP.edu's online reading room since 1999. Based on feedback from you, our users, we've made some improvements that make it easier than ever to read thousands of publications on our website.

Do you want to take a quick tour of the OpenBook's features?

Show this book's table of contents , where you can jump to any chapter by name.

...or use these buttons to go back to the previous chapter or skip to the next one.

Jump up to the previous page or down to the next one. Also, you can type in a page number and press Enter to go directly to that page in the book.

To search the entire text of this book, type in your search term here and press Enter .

Share a link to this book page on your preferred social network or via email.

View our suggested citation for this chapter.

Ready to take your reading offline? Click here to buy this book in print or download it as a free PDF, if available.

Get Email Updates

Do you enjoy reading reports from the Academies online for free ? Sign up for email notifications and we'll let you know about new publications in your areas of interest when they're released.

• Data representation

Bytes of memory

• Abstract machine

Unsigned integer representation

Signed integer representation, pointer representation, array representation, compiler layout, array access performance, collection representation.

• Consequences of size and alignment rules

Uninitialized objects

Pointer arithmetic, undefined behavior.

• Computer arithmetic

Arena allocation

This course is about learning how computers work, from the perspective of systems software: what makes programs work fast or slow, and how properties of the machines we program impact the programs we write. We want to communicate ideas, tools, and an experimental approach.

The course divides into six units:

• Assembly & machine programming
• Storage & caching
• Kernel programming
• Process management
• Concurrency

The first unit, data representation , is all about how different forms of data can be represented in terms the computer can understand.

Computer memory is kind of like a Lite Brite.

A Lite Brite is big black backlit pegboard coupled with a supply of colored pegs, in a limited set of colors. You can plug in the pegs to make all kinds of designs. A computer’s memory is like a vast pegboard where each slot holds one of 256 different colors. The colors are numbered 0 through 255, so each slot holds one byte . (A byte is a number between 0 and 255, inclusive.)

A slot of computer memory is identified by its address . On a computer with M bytes of memory, and therefore M slots, you can think of the address as a number between 0 and M −1. My laptop has 16 gibibytes of memory, so M = 16×2 30 = 2 34 = 17,179,869,184 = 0x4'0000'0000 —a very large number!

The problem of data representation is the problem of representing all the concepts we might want to use in programming—integers, fractions, real numbers, sets, pictures, texts, buildings, animal species, relationships—using the limited medium of addresses and bytes.

Powers of ten and powers of two. Digital computers love the number two and all powers of two. The electronics of digital computers are based on the bit , the smallest unit of storage, which a base-two digit: either 0 or 1. More complicated objects are represented by collections of bits. This choice has many scale and error-correction advantages. It also refracts upwards to larger choices, and even into terminology. Memory chips, for example, have capacities based on large powers of two, such as 2 30 bytes. Since 2 10 = 1024 is pretty close to 1,000, 2 20 = 1,048,576 is pretty close to a million, and 2 30 = 1,073,741,824 is pretty close to a billion, it’s common to refer to 2 30 bytes of memory as “a giga byte,” even though that actually means 10 9 = 1,000,000,000 bytes. But for greater precision, there are terms that explicitly signal the use of powers of two. 2 30 is a gibibyte : the “-bi-” component means “binary.”

Virtual memory. Modern computers actually abstract their memory spaces using a technique called virtual memory . The lowest-level kind of address, called a physical address , really does take on values between 0 and M −1. However, even on a 16GiB machine like my laptop, the addresses we see in programs can take on values like 0x7ffe'ea2c'aa67 that are much larger than M −1 = 0x3'ffff'ffff . The addresses used in programs are called virtual addresses . They’re incredibly useful for protection: since different running programs have logically independent address spaces, it’s much less likely that a bug in one program will crash the whole machine. We’ll learn about virtual memory in much more depth in the kernel unit ; the distinction between virtual and physical addresses is not as critical for data representation.

Most programming languages prevent their users from directly accessing memory. But not C and C++! These languages let you access any byte of memory with a valid address. This is powerful; it is also very dangerous. But it lets us get a hands-on view of how computers really work.

C++ programs accomplish their work by constructing, examining, and modifying objects . An object is a region of data storage that contains a value, such as the integer 12. (The standard specifically says “a region of data storage in the execution environment, the contents of which can represent values”.) Memory is called “memory” because it remembers object values.

In this unit, we often use functions called hexdump to examine memory. These functions are defined in hexdump.cc . hexdump_object(x) prints out the bytes of memory that comprise an object named x , while hexdump(ptr, size) prints out the size bytes of memory starting at a pointer ptr .

For example, in datarep1/add.cc , we might use hexdump_object to examine the memory used to represent some integers:

This display reports that a , b , and c are each four bytes long; that a , b , and c are located at different, nonoverlapping addresses (the long hex number in the first column); and shows us how the numbers 1, 2, and 3 are represented in terms of bytes. (More on that later.)

The compiler, hardware, and standard together define how objects of different types map to bytes. Each object uses a contiguous range of addresses (and thus bytes), and objects never overlap (objects that are active simultaneously are always stored in distinct address ranges).

Since C and C++ are designed to help software interface with hardware devices, their standards are transparent about how objects are stored. A C++ program can ask how big an object is using the sizeof keyword. sizeof(T) returns the number of bytes in the representation of an object of type T , and sizeof(x) returns the size of object x . The result of sizeof is a value of type size_t , which is an unsigned integer type large enough to hold any representable size. On 64-bit architectures, such as x86-64 (our focus in this course), size_t can hold numbers between 0 and 2 64 –1.

Qualitatively different objects may have the same data representation. For example, the following three objects have the same data representation on x86-64, which you can verify using hexdump :

In C and C++, you can’t reliably tell the type of an object by looking at the contents of its memory. That’s why tricks like our different addf*.cc functions work.

An object can have many names. For example, here, local and *ptr refer to the same object:

The different names for an object are sometimes called aliases .

There are five objects here:

• ch1 , a global variable
• ch2 , a constant (non-modifiable) global variable
• ch3 , a local variable
• ch4 , a local variable
• the anonymous storage allocated by new char and accessed by *ch4

Each object has a lifetime , which is called storage duration by the standard. There are three different kinds of lifetime.

• static lifetime: The object lasts as long as the program runs. ( ch1 , ch2 )
• automatic lifetime: The compiler allocates and destroys the object automatically as the program runs, based on the object’s scope (the region of the program in which it is meaningful). ( ch3 , ch4 )
• dynamic lifetime: The programmer allocates and destroys the object explicitly. ( *allocated_ch )

Objects with dynamic lifetime aren’t easy to use correctly. Dynamic lifetime causes many serious problems in C programs, including memory leaks, use-after-free, double-free, and so forth. Those serious problems cause undefined behavior and play a “disastrously central role” in “our ongoing computer security nightmare” . But dynamic lifetime is critically important. Only with dynamic lifetime can you construct an object whose size isn’t known at compile time, or construct an object that outlives the function that created it.

The compiler and operating system work together to put objects at different addresses. A program’s address space (which is the range of addresses accessible to a program) divides into regions called segments . Objects with different lifetimes are placed into different segments. The most important segments are:

• Code (also known as text or read-only data ). Contains instructions and constant global objects. Unmodifiable; static lifetime.
• Data . Contains non-constant global objects. Modifiable; static lifetime.
• Heap . Modifiable; dynamic lifetime.
• Stack . Modifiable; automatic lifetime.

The compiler decides on a segment for each object based on its lifetime. The final compiler phase, which is called the linker , then groups all the program’s objects by segment (so, for instance, global variables from different compiler runs are grouped together into a single segment). Finally, when a program runs, the operating system loads the segments into memory. (The stack and heap segments grow on demand.)

We can use a program to investigate where objects with different lifetimes are stored. (See cs61-lectures/datarep2/mexplore0.cc .) This shows address ranges like this:

Constant global data and global data have the same lifetime, but are stored in different segments. The operating system uses different segments so it can prevent the program from modifying constants. It marks the code segment, which contains functions (instructions) and constant global data, as read-only, and any attempt to modify code-segment memory causes a crash (a “Segmentation violation”).

An executable is normally at least as big as the static-lifetime data (the code and data segments together). Since all that data must be in memory for the entire lifetime of the program, it’s written to disk and then loaded by the OS before the program starts running. There is an exception, however: the “bss” segment is used to hold modifiable static-lifetime data with initial value zero. Such data is common, since all static-lifetime data is initialized to zero unless otherwise specified in the program text. Rather than storing a bunch of zeros in the object files and executable, the compiler and linker simply track the location and size of all zero-initialized global data. The operating system sets this memory to zero during the program load process. Clearing memory is faster than loading data from disk, so this optimization saves both time (the program loads faster) and space (the executable is smaller).

Abstract machine and hardware

Programming involves turning an idea into hardware instructions. This transformation happens in multiple steps, some you control and some controlled by other programs.

First you have an idea , like “I want to make a flappy bird iPhone game.” The computer can’t (yet) understand that idea. So you transform the idea into a program , written in some programming language . This process is called programming.

A C++ program actually runs on an abstract machine . The behavior of this machine is defined by the C++ standard , a technical document. This document is supposed to be so precisely written as to have an exact mathematical meaning, defining exactly how every C++ program behaves. But the document can’t run programs!

C++ programs run on hardware (mostly), and the hardware determines what behavior we see. Mapping abstract machine behavior to instructions on real hardware is the task of the C++ compiler (and the standard library and operating system). A C++ compiler is correct if and only if it translates each correct program to instructions that simulate the expected behavior of the abstract machine.

This same rough series of transformations happens for any programming language, although some languages use interpreters rather than compilers.

A bit is the fundamental unit of digital information: it’s either 0 or 1.

C++ manages memory in units of bytes —8 contiguous bits that together can represent numbers between 0 and 255. C’s unit for a byte is char : the abstract machine says a byte is stored in char . That means an unsigned char holds values in the inclusive range [0, 255].

The C++ standard actually doesn’t require that a byte hold 8 bits, and on some crazy machines from decades ago , bytes could hold nine bits! (!?)

But larger numbers, such as 258, don’t fit in a single byte. To represent such numbers, we must use multiple bytes. The abstract machine doesn’t specify exactly how this is done—it’s the compiler and hardware’s job to implement a choice. But modern computers always use place–value notation , just like in decimal numbers. In decimal, the number 258 is written with three digits, the meanings of which are determined both by the digit and by their place in the overall number:

\[ 258 = 2\times10^2 + 5\times10^1 + 8\times10^0 \]

The computer uses base 256 instead of base 10. Two adjacent bytes can represent numbers between 0 and \(255\times256+255 = 65535 = 2^{16}-1\) , inclusive. A number larger than this would take three or more bytes.

\[ 258 = 1\times256^1 + 2\times256^0 \]

On x86-64, the ones place, the least significant byte, is on the left, at the lowest address in the contiguous two-byte range used to represent the integer. This is the opposite of how decimal numbers are written: decimal numbers put the most significant digit on the left. The representation choice of putting the least-significant byte in the lowest address is called little-endian representation. x86-64 uses little-endian representation.

Some computers actually store multi-byte integers the other way, with the most significant byte stored in the lowest address; that’s called big-endian representation. The Internet’s fundamental protocols, such as IP and TCP, also use big-endian order for multi-byte integers, so big-endian is also called “network” byte order.

The C++ standard defines five fundamental unsigned integer types, along with relationships among their sizes. Here they are, along with their actual sizes and ranges on x86-64:

Other architectures and operating systems implement different ranges for these types. For instance, on IA32 machines like Intel’s Pentium (the 32-bit processors that predated x86-64), sizeof(long) was 4, not 8.

Note that all values of a smaller unsigned integer type can fit in any larger unsigned integer type. When a value of a larger unsigned integer type is placed in a smaller unsigned integer object, however, not every value fits; for instance, the unsigned short value 258 doesn’t fit in an unsigned char x . When this occurs, the C++ abstract machine requires that the smaller object’s value equals the least -significant bits of the larger value (so x will equal 2).

In addition to these types, whose sizes can vary, C++ has integer types whose sizes are fixed. uint8_t , uint16_t , uint32_t , and uint64_t define 8-bit, 16-bit, 32-bit, and 64-bit unsigned integers, respectively; on x86-64, these correspond to unsigned char , unsigned short , unsigned int , and unsigned long .

This general procedure is used to represent a multi-byte integer in memory.

• Write the large integer in hexadecimal format, including all leading zeros required by the type size. For example, the unsigned value 65534 would be written 0x0000FFFE . There will be twice as many hexadecimal digits as sizeof(TYPE) .
• Divide the integer into its component bytes, which are its digits in base 256. In our example, they are, from most to least significant, 0x00, 0x00, 0xFF, and 0xFE.

In little-endian representation, the bytes are stored in memory from least to most significant. If our example was stored at address 0x30, we would have:

In big-endian representation, the bytes are stored in the reverse order.

Computers are often fastest at dealing with fixed-length numbers, rather than variable-length numbers, and processor internals are organized around a fixed word size . A word is the natural unit of data used by a processor design . In most modern processors, this natural unit is 8 bytes or 64 bits , because this is the power-of-two number of bytes big enough to hold those processors’ memory addresses. Many older processors could access less memory and had correspondingly smaller word sizes, such as 4 bytes (32 bits).

The best representation for signed integers—and the choice made by x86-64, and by the C++20 abstract machine—is two’s complement . Two’s complement representation is based on this principle: Addition and subtraction of signed integers shall use the same instructions as addition and subtraction of unsigned integers.

To see what this means, let’s think about what -x should mean when x is an unsigned integer. Wait, negative unsigned?! This isn’t an oxymoron because C++ uses modular arithmetic for unsigned integers: the result of an arithmetic operation on unsigned values is always taken modulo 2 B , where B is the number of bits in the unsigned value type. Thus, on x86-64,

-x is simply the number that, when added to x , yields 0 (mod 2 B ). For example, when unsigned x = 0xFFFFFFFFU , then -x == 1U , since x + -x equals zero (mod 2 32 ).

To obtain -x , we flip all the bits in x (an operation written ~x ) and then add 1. To see why, consider the bit representations. What is x + (~x + 1) ? Well, (~x) i (the i th bit of ~x ) is 1 whenever x i is 0, and vice versa. That means that every bit of x + ~x is 1 (there are no carries), and x + ~x is the largest unsigned integer, with value 2 B -1. If we add 1 to this, we get 2 B . Which is 0 (mod 2 B )! The highest “carry” bit is dropped, leaving zero.

Two’s complement arithmetic uses half of the unsigned integer representations for negative numbers. A two’s-complement signed integer with B bits has the following values:

• If the most-significant bit is 1, the represented number is negative. Specifically, the represented number is – (~x + 1) , where the outer negative sign is mathematical negation (not computer arithmetic).
• If every bit is 0, the represented number is 0.
• If the most-significant but is 0 but some other bit is 1, the represented number is positive.

The most significant bit is also called the sign bit , because if it is 1, then the represented value depends on the signedness of the type (and that value is negative for signed types).

Another way to think about two’s-complement is that, for B -bit integers, the most-significant bit has place value 2 B –1 in unsigned arithmetic and negative 2 B –1 in signed arithmetic. All other bits have the same place values in both kinds of arithmetic.

The two’s-complement bit pattern for x + y is the same whether x and y are considered as signed or unsigned values. For example, in 4-bit arithmetic, 5 has representation 0b0101 , while the representation 0b1100 represents 12 if unsigned and –4 if signed ( ~0b1100 + 1 = 0b0011 + 1 == 4). Let’s add those bit patterns and see what we get:

Note that this is the right answer for both signed and unsigned arithmetic : 5 + 12 = 17 = 1 (mod 16), and 5 + -4 = 1.

Subtraction and multiplication also produce the same results for unsigned arithmetic and signed two’s-complement arithmetic. (For instance, 5 * 12 = 60 = 12 (mod 16), and 5 * -4 = -20 = -4 (mod 16).) This is not true of division. (Consider dividing the 4-bit representation 0b1110 by 2. In signed arithmetic, 0b1110 represents -2, so 0b1110/2 == 0b1111 (-1); but in unsigned arithmetic, 0b1110 is 14, so 0b1110/2 == 0b0111 (7).) And, of course, it is not true of comparison. In signed 4-bit arithmetic, 0b1110 < 0 , but in unsigned 4-bit arithmetic, 0b1110 > 0 . This means that a C compiler for a two’s-complement machine can use a single add instruction for either signed or unsigned numbers, but it must generate different instruction patterns for signed and unsigned division (or less-than, or greater-than).

There are a couple quirks with C signed arithmetic. First, in two’s complement, there are more negative numbers than positive numbers. A representation with sign bit is 1, but every other bit 0, has no positive counterpart at the same bit width: for this number, -x == x . (In 4-bit arithmetic, -0b1000 == ~0b1000 + 1 == 0b0111 + 1 == 0b1000 .) Second, and far worse, is that arithmetic overflow on signed integers is undefined behavior .

The C++ abstract machine requires that signed integers have the same sizes as their unsigned counterparts.

We distinguish pointers , which are concepts in the C abstract machine, from addresses , which are hardware concepts. A pointer combines an address and a type.

The memory representation of a pointer is the same as the representation of its address value. The size of that integer is the machine’s word size; for example, on x86-64, a pointer occupies 8 bytes, and a pointer to an object located at address 0x400abc would be stored as:

The C++ abstract machine defines an unsigned integer type uintptr_t that can hold any address. (You have to #include <inttypes.h> or <cinttypes> to get the definition.) On most machines, including x86-64, uintptr_t is the same as unsigned long . Cast a pointer to an integer address value with syntax like (uintptr_t) ptr ; cast back to a pointer with syntax like (T*) addr . Casts between pointer types and uintptr_t are information preserving, so this assertion will never fail:

Since it is a 64-bit architecture, the size of an x86-64 address is 64 bits (8 bytes). That’s also the size of x86-64 pointers.

To represent an array of integers, C++ and C allocate the integers next to each other in memory, in sequential addresses, with no gaps or overlaps. Here, we put the integers 0, 1, and 258 next to each other, starting at address 1008:

Say that you have an array of N integers, and you access each of those integers in order, accessing each integer exactly once. Does the order matter?

Computer memory is random-access memory (RAM), which means that a program can access any bytes of memory in any order—it’s not, for example, required to read memory in ascending order by address. But if we run experiments, we can see that even in RAM, different access orders have very different performance characteristics.

Our arraysum program sums up all the integers in an array of N integers, using an access order based on its arguments, and prints the resulting delay. Here’s the result of a couple experiments on accessing 10,000,000 items in three orders, “up” order (sequential: elements 0, 1, 2, 3, …), “down” order (reverse sequential: N , N −1, N −2, …), and “random” order (as it sounds).

Wow! Down order is just a bit slower than up, but random order seems about 40 times slower. Why?

Random order is defeating many of the internal architectural optimizations that make memory access fast on modern machines. Sequential order, since it’s more predictable, is much easier to optimize.

Foreshadowing. This part of the lecture is a teaser for the Storage unit, where we cover access patterns and caching, including the processor caches that explain this phenomenon, in much more depth.

The C++ programming language offers several collection mechanisms for grouping subobjects together into new kinds of object. The collections are arrays, structs, and unions. (Classes are a kind of struct. All library types, such as vectors, lists, and hash tables, use combinations of these collection types.) The abstract machine defines how subobjects are laid out inside a collection. This is important, because it lets C/C++ programs exchange messages with hardware and even with programs written in other languages: messages can be exchanged only when both parties agree on layout.

Array layout in C++ is particularly simple: The objects in an array are laid out sequentially in memory, with no gaps or overlaps. Assume a declaration like T x[N] , where x is an array of N objects of type T , and say that the address of x is a . Then the address of element x[i] equals a + i * sizeof(T) , and sizeof(a) == N * sizeof(T) .

Sidebar: Vector representation

The C++ library type std::vector defines an array that can grow and shrink. For instance, this function creates a vector containing the numbers 0 up to N in sequence:

Here, v is an object with automatic lifetime. This means its size (in the sizeof sense) is fixed at compile time. Remember that the sizes of static- and automatic-lifetime objects must be known at compile time; only dynamic-lifetime objects can have varying size based on runtime parameters. So where and how are v ’s contents stored?

The C++ abstract machine requires that v ’s elements are stored in an array in memory. (The v.data() method returns a pointer to the first element of the array.) But it does not define std::vector ’s layout otherwise, and C++ library designers can choose different layouts based on their needs. We found these to hold for the std::vector in our library:

sizeof(v) == 24 for any vector of any type, and the address of v is a stack address (i.e., v is located in the stack segment).

The first 8 bytes of the vector hold the address of the first element of the contents array—call it the begin address . This address is a heap address, which is as expected, since the contents must have dynamic lifetime. The value of the begin address is the same as that of v.data() .

Bytes 8–15 hold the address just past the contents array—call it the end address . Its value is the same as &v.data()[v.size()] . If the vector is empty, then the begin address and the end address are the same.

Bytes 16–23 hold an address greater than or equal to the end address. This is the capacity address . As a vector grows, it will sometimes outgrow its current location and move its contents to new memory addresses. To reduce the number of copies, vectors usually to request more memory from the operating system than they immediately need; this additional space, which is called “capacity,” supports cheap growth. Often the capacity doubles on each growth spurt, since this allows operations like v.push_back() to execute in O (1) time on average.

Compilers must also decide where different objects are stored when those objects are not part of a collection. For instance, consider this program:

The abstract machine says these objects cannot overlap, but does not otherwise constrain their positions in memory.

On Linux, GCC will put all these variables into the stack segment, which we can see using hexdump . But it can put them in the stack segment in any order , as we can see by reordering the declarations (try declaration order i1 , c1 , i2 , c2 , c3 ), by changing optimization levels, or by adding different scopes (braces). The abstract machine gives the programmer no guarantees about how object addresses relate. In fact, the compiler may move objects around during execution, as long as it ensures that the program behaves according to the abstract machine. Modern optimizing compilers often do this, particularly for automatic objects.

But what order does the compiler choose? With optimization disabled, the compiler appears to lay out objects in decreasing order by declaration, so the first declared variable in the function has the highest address. With optimization enabled, the compiler follows roughly the same guideline, but it also rearranges objects by type—for instance, it tends to group char s together—and it can reuse space if different variables in the same function have disjoint lifetimes. The optimizing compiler tends to use less space for the same set of variables. This is because it’s arranging objects by alignment.

The C++ compiler and library restricts the addresses at which some kinds of data appear. In particular, the address of every int value is always a multiple of 4, whether it’s located on the stack (automatic lifetime), the data segment (static lifetime), or the heap (dynamic lifetime).

A bunch of observations will show you these rules:

These are the alignment restrictions for an x86-64 Linux machine.

These restrictions hold for most x86-64 operating systems, except that on Windows, the long type has size and alignment 4. (The long long type has size and alignment 8 on all x86-64 operating systems.)

Just like every type has a size, every type has an alignment. The alignment of a type T is a number a ≥1 such that the address of every object of type T must be a multiple of a . Every object with type T has size sizeof(T) —it occupies sizeof(T) contiguous bytes of memory; and has alignment alignof(T) —the address of its first byte is a multiple of alignof(T) . You can also say sizeof(x) and alignof(x) where x is the name of an object or another expression.

Alignment restrictions can make hardware simpler, and therefore faster. For instance, consider cache blocks. CPUs access memory through a transparent hardware cache. Data moves from primary memory, or RAM (which is large—a couple gigabytes on most laptops—and uses cheaper, slower technology) to the cache in units of 64 or 128 bytes. Those units are always aligned: on a machine with 128-byte cache blocks, the bytes with memory addresses [127, 128, 129, 130] live in two different cache blocks (with addresses [0, 127] and [128, 255]). But the 4 bytes with addresses [4n, 4n+1, 4n+2, 4n+3] always live in the same cache block. (This is true for any small power of two: the 8 bytes with addresses [8n,…,8n+7] always live in the same cache block.) In general, it’s often possible to make a system faster by leveraging restrictions—and here, the CPU hardware can load data faster when it can assume that the data lives in exactly one cache line.

The compiler, library, and operating system all work together to enforce alignment restrictions.

On x86-64 Linux, alignof(T) == sizeof(T) for all fundamental types (the types built in to C: integer types, floating point types, and pointers). But this isn’t always true; on x86-32 Linux, double has size 8 but alignment 4.

It’s possible to construct user-defined types of arbitrary size, but the largest alignment required by a machine is fixed for that machine. C++ lets you find the maximum alignment for a machine with alignof(std::max_align_t) ; on x86-64, this is 16, the alignment of the type long double (and the alignment of some less-commonly-used SIMD “vector” types ).

We now turn to the abstract machine rules for laying out all collections. The sizes and alignments for user-defined types—arrays, structs, and unions—are derived from a couple simple rules or principles. Here they are. The first rule applies to all types.

1. First-member rule. The address of the first member of a collection equals the address of the collection.

Thus, the address of an array is the same as the address of its first element. The address of a struct is the same as the address of the first member of the struct.

The next three rules depend on the class of collection. Every C abstract machine enforces these rules.

2. Array rule. Arrays are laid out sequentially as described above.

3. Struct rule. The second and subsequent members of a struct are laid out in order, with no overlap, subject to alignment constraints.

4. Union rule. All members of a union share the address of the union.

In C, every struct follows the struct rule, but in C++, only simple structs follow the rule. Complicated structs, such as structs with some public and some private members, or structs with virtual functions, can be laid out however the compiler chooses. The typical situation is that C++ compilers for a machine architecture (e.g., “Linux x86-64”) will all agree on a layout procedure for complicated structs. This allows code compiled by different compilers to interoperate.

That next rule defines the operation of the malloc library function.

5. Malloc rule. Any non-null pointer returned by malloc has alignment appropriate for any type. In other words, assuming the allocated size is adequate, the pointer returned from malloc can safely be cast to T* for any T .

Oddly, this holds even for small allocations. The C++ standard (the abstract machine) requires that malloc(1) return a pointer whose alignment is appropriate for any type, including types that don’t fit.

And the final rule is not required by the abstract machine, but it’s how sizes and alignments on our machines work.

6. Minimum rule. The sizes and alignments of user-defined types, and the offsets of struct members, are minimized within the constraints of the other rules.

The minimum rule, and the sizes and alignments of basic types, are defined by the x86-64 Linux “ABI” —its Application Binary Interface. This specification standardizes how x86-64 Linux C compilers should behave, and lets users mix and match compilers without problems.

Consequences of the size and alignment rules

From these rules we can derive some interesting consequences.

First, the size of every type is a multiple of its alignment .

To see why, consider an array with two elements. By the array rule, these elements have addresses a and a+sizeof(T) , where a is the address of the array. Both of these addresses contain a T , so they are both a multiple of alignof(T) . That means sizeof(T) is also a multiple of alignof(T) .

We can also characterize the sizes and alignments of different collections .

• The size of an array of N elements of type T is N * sizeof(T) : the sum of the sizes of its elements. The alignment of the array is alignof(T) .
• The size of a union is the maximum of the sizes of its components (because the union can only hold one component at a time). Its alignment is also the maximum of the alignments of its components.
• The size of a struct is at least as big as the sum of the sizes of its components. Its alignment is the maximum of the alignments of its components.

Thus, the alignment of every collection equals the maximum of the alignments of its components.

It’s also true that the alignment equals the least common multiple of the alignments of its components. You might have thought lcm was a better answer, but the max is the same as the lcm for every architecture that matters, because all fundamental alignments are powers of two.

The size of a struct might be larger than the sum of the sizes of its components, because of alignment constraints. Since the compiler must lay out struct components in order, and it must obey the components’ alignment constraints, and it must ensure different components occupy disjoint addresses, it must sometimes introduce extra space in structs. Here’s an example: the struct will have 3 bytes of padding after char c , to ensure that int i2 has the correct alignment.

Thanks to padding, reordering struct components can sometimes reduce the total size of a struct. Padding can happen at the end of a struct as well as the middle. Padding can never happen at the start of a struct, however (because of Rule 1).

The rules also imply that the offset of any struct member —which is the difference between the address of the member and the address of the containing struct— is a multiple of the member’s alignment .

To see why, consider a struct s with member m at offset o . The malloc rule says that any pointer returned from malloc is correctly aligned for s . Every pointer returned from malloc is maximally aligned, equalling 16*x for some integer x . The struct rule says that the address of m , which is 16*x + o , is correctly aligned. That means that 16*x + o = alignof(m)*y for some integer y . Divide both sides by a = alignof(m) and you see that 16*x/a + o/a = y . But 16/a is an integer—the maximum alignment is a multiple of every alignment—so 16*x/a is an integer. We can conclude that o/a must also be an integer!

Finally, we can also derive the necessity for padding at the end of structs. (How?)

What happens when an object is uninitialized? The answer depends on its lifetime.

• static lifetime (e.g., int global; at file scope): The object is initialized to 0.
• automatic or dynamic lifetime (e.g., int local; in a function, or int* ptr = new int ): The object is uninitialized and reading the object’s value before it is assigned causes undefined behavior.

Compiler hijinks

In C++, most dynamic memory allocation uses special language operators, new and delete , rather than library functions.

Though this seems more complex than the library-function style, it has advantages. A C compiler cannot tell what malloc and free do (especially when they are redefined to debugging versions, as in the problem set), so a C compiler cannot necessarily optimize calls to malloc and free away. But the C++ compiler may assume that all uses of new and delete follow the rules laid down by the abstract machine. That means that if the compiler can prove that an allocation is unnecessary or unused, it is free to remove that allocation!

For example, we compiled this program in the problem set environment (based on test003.cc ):

The optimizing C++ compiler removes all calls to new and delete , leaving only the call to m61_printstatistics() ! (For instance, try objdump -d testXXX to look at the compiled x86-64 instructions.) This is valid because the compiler is explicitly allowed to eliminate unused allocations, and here, since the ptrs variable is local and doesn’t escape main , all allocations are unused. The C compiler cannot perform this useful transformation. (But the C compiler can do other cool things, such as unroll the loops .)

One of C’s more interesting choices is that it explicitly relates pointers and arrays. Although arrays are laid out in memory in a specific way, they generally behave like pointers when they are used. This property probably arose from C’s desire to explicitly model memory as an array of bytes, and it has beautiful and confounding effects.

We’ve already seen one of these effects. The hexdump function has this signature (arguments and return type):

But we can just pass an array as argument to hexdump :

When used in an expression like this—here, as an argument—the array magically changes into a pointer to its first element. The above call has the same meaning as this:

C programmers transition between arrays and pointers very naturally.

A confounding effect is that unlike all other types, in C arrays are passed to and returned from functions by reference rather than by value. C is a call-by-value language except for arrays. This means that all function arguments and return values are copied, so that parameter modifications inside a function do not affect the objects passed by the caller—except for arrays. For instance: void f ( int a[ 2 ]) { a[ 0 ] = 1 ; } int main () { int x[ 2 ] = { 100 , 101 }; f(x); printf( "%d \n " , x[ 0 ]); // prints 1! } If you don’t like this behavior, you can get around it by using a struct or a C++ std::array . #include <array> struct array1 { int a[ 2 ]; }; void f1 (array1 arg) { arg.a[ 0 ] = 1 ; } void f2 (std :: array < int , 2 > a) { a[ 0 ] = 1 ; } int main () { array1 x = {{ 100 , 101 }}; f1(x); printf( "%d \n " , x.a[ 0 ]); // prints 100 std :: array < int , 2 > x2 = { 100 , 101 }; f2(x2); printf( "%d \n " , x2[ 0 ]); // prints 100 }

C++ extends the logic of this array–pointer correspondence to support arithmetic on pointers as well.

Pointer arithmetic rule. In the C abstract machine, arithmetic on pointers produces the same result as arithmetic on the corresponding array indexes.

Specifically, consider an array T a[n] and pointers T* p1 = &a[i] and T* p2 = &a[j] . Then:

Equality : p1 == p2 if and only if (iff) p1 and p2 point to the same address, which happens iff i == j .

Inequality : Similarly, p1 != p2 iff i != j .

Less-than : p1 < p2 iff i < j .

Also, p1 <= p2 iff i <= j ; and p1 > p2 iff i > j ; and p1 >= p2 iff i >= j .

Pointer difference : What should p1 - p2 mean? Using array indexes as the basis, p1 - p2 == i - j . (But the type of the difference is always ptrdiff_t , which on x86-64 is long , the signed version of size_t .)

Addition : p1 + k (where k is an integer type) equals the pointer &a[i + k] . ( k + p1 returns the same thing.)

Subtraction : p1 - k equals &a[i - k] .

Increment and decrement : ++p1 means p1 = p1 + 1 , which means p1 = &a[i + 1] . Similarly, --p1 means p1 = &a[i - 1] . (There are also postfix versions, p1++ and p1-- , but C++ style prefers the prefix versions.)

No other arithmetic operations on pointers are allowed. You can’t multiply pointers, for example. (You can multiply addresses by casting the pointers to the address type, uintptr_t —so (uintptr_t) p1 * (uintptr_t) p2 —but why would you?)

From pointers to iterators

Let’s write a function that can sum all the integers in an array.

This function can compute the sum of the elements of any int array. But because of the pointer–array relationship, its a argument is really a pointer . That allows us to call it with subarrays as well as with whole arrays. For instance:

This way of thinking about arrays naturally leads to a style that avoids sizes entirely, using instead a sentinel or boundary argument that defines the end of the interesting part of the array.

These expressions compute the same sums as the above:

Note that the data from first to last forms a half-open range . iIn mathematical notation, we care about elements in the range [first, last) : the element pointed to by first is included (if it exists), but the element pointed to by last is not. Half-open ranges give us a simple and clear way to describe empty ranges, such as zero-element arrays: if first == last , then the range is empty.

Note that given a ten-element array a , the pointer a + 10 can be formed and compared, but must not be dereferenced—the element a[10] does not exist. The C/C++ abstract machines allow users to form pointers to the “one-past-the-end” boundary elements of arrays, but users must not dereference such pointers.

So in C, two pointers naturally express a range of an array. The C++ standard template library, or STL, brilliantly abstracts this pointer notion to allow two iterators , which are pointer-like objects, to express a range of any standard data structure—an array, a vector, a hash table, a balanced tree, whatever. This version of sum works for any container of int s; notice how little it changed:

Some example uses:

Addresses vs. pointers

What’s the difference between these expressions? (Again, a is an array of type T , and p1 == &a[i] and p2 == &a[j] .)

The first expression is defined analogously to index arithmetic, so d1 == i - j . But the second expression performs the arithmetic on the addresses corresponding to those pointers. We will expect d2 to equal sizeof(T) * d1 . Always be aware of which kind of arithmetic you’re using. Generally arithmetic on pointers should not involve sizeof , since the sizeof is included automatically according to the abstract machine; but arithmetic on addresses almost always should involve sizeof .

Although C++ is a low-level language, the abstract machine is surprisingly strict about which pointers may be formed and how they can be used. Violate the rules and you’re in hell because you have invoked the dreaded undefined behavior .

Given an array a[N] of N elements of type T :

Forming a pointer &a[i] (or a + i ) with 0 ≤ i ≤ N is safe.

Forming a pointer &a[i] with i < 0 or i > N causes undefined behavior.

Dereferencing a pointer &a[i] with 0 ≤ i < N is safe.

Dereferencing a pointer &a[i] with i < 0 or i ≥ N causes undefined behavior.

(For the purposes of these rules, objects that are not arrays count as single-element arrays. So given T x , we can safely form &x and &x + 1 and dereference &x .)

What “undefined behavior” means is horrible. A program that executes undefined behavior is erroneous. But the compiler need not catch the error. In fact, the abstract machine says anything goes : undefined behavior is “behavior … for which this International Standard imposes no requirements.” “Possible undefined behavior ranges from ignoring the situation completely with unpredictable results, to behaving during translation or program execution in a documented manner characteristic of the environment (with or without the issuance of a diagnostic message), to terminating a translation or execution (with the issuance of a diagnostic message).” Other possible behaviors include allowing hackers from the moon to steal all of a program’s data, take it over, and force it to delete the hard drive on which it is running. Once undefined behavior executes, a program may do anything, including making demons fly out of the programmer’s nose.

Pointer arithmetic, and even pointer comparisons, are also affected by undefined behavior. It’s undefined to go beyond and array’s bounds using pointer arithmetic. And pointers may be compared for equality or inequality even if they point to different arrays or objects, but if you try to compare different arrays via less-than, like this:

that causes undefined behavior.

If you really want to compare pointers that might be to different arrays—for instance, you’re writing a hash function for arbitrary pointers—cast them to uintptr_t first.

Undefined behavior and optimization

A program that causes undefined behavior is not a C++ program . The abstract machine says that a C++ program, by definition, is a program whose behavior is always defined. The C++ compiler is allowed to assume that its input is a C++ program. (Obviously!) So the compiler can assume that its input program will never cause undefined behavior. Thus, since undefined behavior is “impossible,” if the compiler can prove that a condition would cause undefined behavior later, it can assume that condition will never occur.

Consider this program:

If we supply a value equal to (char*) -1 , we’re likely to see output like this:

with no assertion failure! But that’s an apparently impossible result. The printout can only happen if x + 1 > x (otherwise, the assertion will fail and stop the printout). But x + 1 , which equals 0 , is less than x , which is the largest 8-byte value!

The impossible happens because of undefined behavior reasoning. When the compiler sees an expression like x + 1 > x (with x a pointer), it can reason this way:

“Ah, x + 1 . This must be a pointer into the same array as x (or it might be a boundary pointer just past that array, or just past the non-array object x ). This must be so because forming any other pointer would cause undefined behavior.

“The pointer comparison is the same as an index comparison. x + 1 > x means the same thing as &x[1] > &x[0] . But that holds iff 1 > 0 .

“In my infinite wisdom, I know that 1 > 0 . Thus x + 1 > x always holds, and the assertion will never fail.

“My job is to make this code run fast. The fastest code is code that’s not there. This assertion will never fail—might as well remove it!”

Integer undefined behavior

Arithmetic on signed integers also has important undefined behaviors. Signed integer arithmetic must never overflow. That is, the compiler may assume that the mathematical result of any signed arithmetic operation, such as x + y (with x and y both int ), can be represented inside the relevant type. It causes undefined behavior, therefore, to add 1 to the maximum positive integer. (The ubexplore.cc program demonstrates how this can produce impossible results, as with pointers.)

Arithmetic on unsigned integers is much safer with respect to undefined behavior. Unsigned integers are defined to perform arithmetic modulo their size. This means that if you add 1 to the maximum positive unsigned integer, the result will always be zero.

Dividing an integer by zero causes undefined behavior whether or not the integer is signed.

Sanitizers, which in our makefiles are turned on by supplying SAN=1 , can catch many undefined behaviors as soon as they happen. Sanitizers are built in to the compiler itself; a sanitizer involves cooperation between the compiler and the language runtime. This has the major performance advantage that the compiler introduces exactly the required checks, and the optimizer can then use its normal analyses to remove redundant checks.

That said, undefined behavior checking can still be slow. Undefined behavior allows compilers to make assumptions about input values, and those assumptions can directly translate to faster code. Turning on undefined behavior checking can make some benchmark programs run 30% slower [link] .

Signed integer undefined behavior

File cs61-lectures/datarep5/ubexplore2.cc contains the following program.

What will be printed if we run the program with ./ubexplore2 0x7ffffffe 0x7fffffff ?

0x7fffffff is the largest positive value can be represented by type int . Adding one to this value yields 0x80000000 . In two's complement representation this is the smallest negative number represented by type int .

Assuming that the program behaves this way, then the loop exit condition i > n2 can never be met, and the program should run (and print out numbers) forever.

However, if we run the optimized version of the program, it prints only two numbers and exits:

The unoptimized program does print forever and never exits.

What’s going on here? We need to look at the compiled assembly of the program with and without optimization (via objdump -S ).

The unoptimized version basically looks like this:

This is a pretty direct translation of the loop.

The optimized version, though, does it differently. As always, the optimizer has its own ideas. (Your compiler may produce different results!)

The compiler changed the source’s less than or equal to comparison, i <= n2 , into a not equal to comparison in the executable, i != n2 + 1 (in both cases using signed computer arithmetic, i.e., modulo 2 32 )! The comparison i <= n2 will always return true when n2 == 0x7FFFFFFF , the maximum signed integer, so the loop goes on forever. But the i != n2 + 1 comparison does not always return true when n2 == 0x7FFFFFFF : when i wraps around to 0x80000000 (the smallest negative integer), then i equals n2 + 1 (which also wrapped), and the loop stops.

Why did the compiler make this transformation? In the original loop, the step-6 jump is immediately followed by another comparison and jump in steps 1 and 2. The processor jumps all over the place, which can confuse its prediction circuitry and slow down performance. In the transformed loop, the step-7 jump is never followed by a comparison and jump; instead, step 7 goes back to step 4, which always prints the current number. This more streamlined control flow is easier for the processor to make fast.

But the streamlined control flow is only a valid substitution under the assumption that the addition n2 + 1 never overflows . Luckily (sort of), signed arithmetic overflow causes undefined behavior, so the compiler is totally justified in making that assumption!

Programs based on ubexplore2 have demonstrated undefined behavior differences for years, even as the precise reasons why have changed. In some earlier compilers, we found that the optimizer just upgraded the int s to long s—arithmetic on long s is just as fast on x86-64 as arithmetic on int s, since x86-64 is a 64-bit architecture, and sometimes using long s for everything lets the compiler avoid conversions back and forth. The ubexplore2l program demonstrates this form of transformation: since the loop variable is added to a long counter, the compiler opportunistically upgrades i to long as well. This transformation is also only valid under the assumption that i + 1 will not overflow—which it can’t, because of undefined behavior.

Using unsigned type prevents all this undefined behavior, because arithmetic overflow on unsigned integers is well defined in C/C++. The ubexplore2u.cc file uses an unsigned loop index and comparison, and ./ubexplore2u and ./ubexplore2u.noopt behave exactly the same (though you have to give arguments like ./ubexplore2u 0xfffffffe 0xffffffff to see the overflow).

Computer arithmetic and bitwise operations

Basic bitwise operators.

Computers offer not only the usual arithmetic operators like + and - , but also a set of bitwise operators. The basic ones are & (and), | (or), ^ (xor/exclusive or), and the unary operator ~ (complement). In truth table form:

In C or C++, these operators work on integers. But they work bitwise: the result of an operation is determined by applying the operation independently at each bit position. Here’s how to compute 12 & 4 in 4-bit unsigned arithmetic:

These basic bitwise operators simplify certain important arithmetics. For example, (x & (x - 1)) == 0 tests whether x is zero or a power of 2.

Negation of signed integers can also be expressed using a bitwise operator: -x == ~x + 1 . This is in fact how we define two's complement representation. We can verify that x and (-x) does add up to zero under this representation:

Bitwise "and" ( & ) can help with modular arithmetic. For example, x % 32 == (x & 31) . We essentially "mask off", or clear, higher order bits to do modulo-powers-of-2 arithmetics. This works in any base. For example, in decimal, the fastest way to compute x % 100 is to take just the two least significant digits of x .

Bitwise shift of unsigned integer

x << i appends i zero bits starting at the least significant bit of x . High order bits that don't fit in the integer are thrown out. For example, assuming 4-bit unsigned integers

Similarly, x >> i appends i zero bits at the most significant end of x . Lower bits are thrown out.

Bitwise shift helps with division and multiplication. For example:

A modern compiler can optimize y = x * 66 into y = (x << 6) + (x << 1) .

Bitwise operations also allows us to treat bits within an integer separately. This can be useful for "options".

For example, when we call a function to open a file, we have a lot of options:

• Open for reading?
• Open for writing?
• Read from the end?
• Optimize for writing?

We have a lot of true/false options.

One bad way to implement this is to have this function take a bunch of arguments -- one argument for each option. This makes the function call look like this:

The long list of arguments slows down the function call, and one can also easily lose track of the meaning of the individual true/false values passed in.

A cheaper way to achieve this is to use a single integer to represent all the options. Have each option defined as a power of 2, and simply | (or) them together and pass them as a single integer.

Flags are usually defined as powers of 2 so we set one bit at a time for each flag. It is less common but still possible to define a combination flag that is not a power of 2, so that it sets multiple bits in one go.

File cs61-lectures/datarep5/mb-driver.cc contains a memory allocation benchmark. The core of the benchmark looks like this:

The benchmark tests the performance of memnode_arena::allocate() and memnode_arena::deallocate() functions. In the handout code, these functions do the same thing as new memnode and delete memnode —they are wrappers for malloc and free . The benchmark allocates 4096 memnode objects, then free-and-then-allocates them for noperations times, and then frees all of them.

We only allocate memnode s, and all memnode s are of the same size, so we don't need metadata that keeps track of the size of each allocation. Furthermore, since all dynamically allocated data are freed at the end of the function, for each individual memnode_free() call we don't really need to return memory to the system allocator. We can simply reuse these memory during the function and returns all memory to the system at once when the function exits.

If we run the benchmark with 100000000 allocation, and use the system malloc() , free() functions to implement the memnode allocator, the benchmark finishes in 0.908 seconds.

Our alternative implementation of the allocator can finish in 0.355 seconds, beating the heavily optimized system allocator by a factor of 3. We will reveal how we achieved this in the next lecture.

We continue our exploration with the memnode allocation benchmark introduced from the last lecture.

File cs61-lectures/datarep6/mb-malloc.cc contains a version of the benchmark using the system new and delete operators.

In this function we allocate an array of 4096 pointers to memnode s, which occupy 2 3 *2 12 =2 15 bytes on the stack. We then allocate 4096 memnode s. Our memnode is defined like this:

Each memnode contains a std::string object and an unsigned integer. Each std::string object internally contains a pointer points to an character array in the heap. Therefore, every time we create a new memnode , we need 2 allocations: one to allocate the memnode itself, and another one performed internally by the std::string object when we initialize/assign a string value to it.

Every time we deallocate a memnode by calling delete , we also delete the std::string object, and the string object knows that it should deallocate the heap character array it internally maintains. So there are also 2 deallocations occuring each time we free a memnode.

We make the benchmark to return a seemingly meaningless result to prevent an aggressive compiler from optimizing everything away. We also use this result to make sure our subsequent optimizations to the allocator are correct by generating the same result.

This version of the benchmark, using the system allocator, finishes in 0.335 seconds. Not bad at all.

Spoiler alert: We can do 15x better than this.

1st optimization: std::string

We only deal with one file name, namely "datarep/mb-filename.cc", which is constant throughout the program for all memnode s. It's also a string literal, which means it as a constant string has a static life time. Why can't we just simply use a const char* in place of the std::string and let the pointer point to the static constant string? This saves us the internal allocation/deallocation performed by std::string every time we initialize/delete a string.

The fix is easy, we simply change the memnode definition:

This version of the benchmark now finishes in 0.143 seconds, a 2x improvement over the original benchmark. This 2x improvement is consistent with a 2x reduction in numbers of allocation/deallocation mentioned earlier.

You may ask why people still use std::string if it involves an additional allocation and is slower than const char* , as shown in this benchmark. std::string is much more flexible in that it also deals data that doesn't have static life time, such as input from a user or data the program receives over the network. In short, when the program deals with strings that are not constant, heap data is likely to be very useful, and std::string provides facilities to conveniently handle on-heap data.

2nd optimization: the system allocator

We still use the system allocator to allocate/deallocate memnode s. The system allocator is a general-purpose allocator, which means it must handle allocation requests of all sizes. Such general-purpose designs usually comes with a compromise for performance. Since we are only memnode s, which are fairly small objects (and all have the same size), we can build a special- purpose allocator just for them.

In cs61-lectures/datarep5/mb2.cc , we actually implement a special-purpose allocator for memnode s:

This allocator maintains a free list (a C++ vector ) of freed memnode s. allocate() simply pops a memnode off the free list if there is any, and deallocate() simply puts the memnode on the free list. This free list serves as a buffer between the system allocator and the benchmark function, so that the system allocator is invoked less frequently. In fact, in the benchmark, the system allocator is only invoked for 4096 times when it initializes the pointer array. That's a huge reduction because all 10-million "recycle" operations in the middle now doesn't involve the system allocator.

With this special-purpose allocator we can finish the benchmark in 0.057 seconds, another 2.5x improvement.

However this allocator now leaks memory: it never actually calls delete ! Let's fix this by letting it also keep track of all allocated memnode s. The modified definition of memnode_arena now looks like this:

With the updated allocator we simply need to invoke arena.destroy_all() at the end of the function to fix the memory leak. And we don't even need to invoke this method manually! We can use the C++ destructor for the memnode_arena struct, defined as ~memnode_arena() in the code above, which is automatically called when our arena object goes out of scope. We simply make the destructor invoke the destroy_all() method, and we are all set.

Fixing the leak doesn't appear to affect performance at all. This is because the overhead added by tracking the allocated list and calling delete only affects our initial allocation the 4096 memnode* pointers in the array plus at the very end when we clean up. These 8192 additional operations is a relative small number compared to the 10 million recycle operations, so the added overhead is hardly noticeable.

Spoiler alert: We can improve this by another factor of 2.

3rd optimization: std::vector

In our special purpose allocator memnode_arena , we maintain an allocated list and a free list both using C++ std::vector s. std::vector s are dynamic arrays, and like std::string they involve an additional level of indirection and stores the actual array in the heap. We don't access the allocated list during the "recycling" part of the benchmark (which takes bulk of the benchmark time, as we showed earlier), so the allocated list is probably not our bottleneck. We however, add and remove elements from the free list for each recycle operation, and the indirection introduced by the std::vector here may actually be our bottleneck. Let's find out.

Instead of using a std::vector , we could use a linked list of all free memnode s for the actual free list. We will need to include some extra metadata in the memnode to store pointers for this linked list. However, unlike in the debugging allocator pset, in a free list we don't need to store this metadata in addition to actual memnode data: the memnode is free, and not in use, so we can use reuse its memory, using a union:

We then maintain the free list like this:

Compared to the std::vector free list, this free list we always directly points to an available memnode when it is not empty ( free_list !=nullptr ), without going through any indirection. In the std::vector free list one would first have to go into the heap to access the actual array containing pointers to free memnode s, and then access the memnode itself.

With this change we can now finish the benchmark under 0.3 seconds! Another 2x improvement over the previous one!

Compared to the benchmark with the system allocator (which finished in 0.335 seconds), we managed to achieve a speedup of nearly 15x with arena allocation.

Quantitative Methods Using R

4 data representation.

Data representation refers to the process of presenting data in a visual or graphical format that makes it easier to understand and interpret. It is crucial to effectively represent data to communicate research findings, identify trends, and explore relationships between variables.

Examples of data representation techniques in educational psychology might include creating line graphs to show changes in student performance over time, using pie charts to compare student achievement across different demographic groups, or using scatter plots to explore the relationship between two or more variables in a study.

ggplot2 is a powerful data visualization package in R, created by Hadley Wickham. It is based on the Grammar of Graphics, a framework that allows you to build complex and customizable plots by layering components. ggplot2 enables the creation of a wide variety of visually appealing and informative graphics with a relatively concise and consistent syntax.

4.1 Loading your data

You can import or load your data as we discussed in Chapter 3. In these examples, we will be loading gapminder data set which is available as a package in R.

4.2 Frequency Tables

A frequency table displays the number of occurrences (frequencies) for each category or value in a data set. It is particularly useful for summarizing categorical data or discrete numerical data.

4.3 Histograms:

Histograms are used to visualize the distribution of continuous or discrete numerical data. They display the data using intervals (bins) along the x-axis and the frequency of observations within each bin on the y-axis.

The aes() function sets the aesthetic mappings for the plot. In this case, the x-axis is mapped to the lifeExp variable from the dataset, which represents life expectancy. geom_histogram(binwidth = 5) adds a histogram layer to the plot. The binwidth parameter is set to 5, which means that the data will be divided into bins of width 5. The height of each bar in the histogram represents the frequency (count) of data points within each bin. xlab("Life Expectancy") adds a label to the x-axis, naming it “Life Expectancy”. ylab("Frequency") adds a label to the y-axis, naming it “Frequency”.

4.4 Bar Graphs:

Bar graphs are used for displaying categorical data. Each category is represented by a bar, and the height (or length) of the bar indicates the frequency or count of that category.

4.5 Pie Charts:

Pie charts represent categorical data as slices of a circle. The size of each slice is proportional to the frequency of each category. Pie charts are useful for visualizing relative proportions of categories. Drawing piechart in ggplot2 package requires transforming a bar plot to polar coordinates, however, its much easier with plotrix package. You can install this package using install.packages(plotrix) .

4.6 Box Plots:

Box plots are used for visualizing the distribution of continuous or discrete numerical data. They show the median, quartiles, and outliers of the data, providing a compact and informative representation of the data distribution.

4.7 Scatter Plots

Scatter plots are used to display the relationship between two continuous variables. They can be particularly helpful in identifying trends, correlations, and potential outliers in the data.

4.8 Line Graphs

Line graphs are used to display the relationship between a continuous variable and a discrete or ordinal variable, often representing change over time. They can be particularly useful for identifying trends and patterns in time-series data.

Laboratory for Intelligent Probabilistic Systems

Princeton University Department of Computer Science

What is representation learning?

Roger Grosse · February 25, 2013

In my last post , I argued that a major distinction in machine learning is between predictive learning and representation learning. Now I'll take a stab at summarizing what representation learning is about. Or, at least, what I think of as the first principal component of representation learning.

A good starting point is the notion of representation from David Marr's classic book, Vision: a Computational Investigation  [1, p. 20-24]. In Marr's view, a representation is a formal system which "makes explicit certain entities and types of information," and which can be operated on by an algorithm in order to achieve some information processing goal (such as obtaining a depth image from two stereo views). Representations differ in terms of what information they make explicit and in terms of what algorithms they support. He gives the example of Arabic and Roman numerals (which Shamim also mentioned): the fact that operations can be applied to particular columns of Arabic numerals in meaningful ways allows for simple and efficient algorithms for addition and multiplication.

I'm sure a lot of others have proposed similar views, but being from MIT, I'm supposed to attribute everything to Marr.

I'm not going to say much about the first half of the definition, about making information explicit. While Marr tends to focus on clean representations where elements of the representation directly correspond to meaningful things in the world, in machine learning we're happy to work with messier representations. Since we've built up a powerful toolbox of statistical algorithms, it suffices to work with representations which merely correlate enough with properties of the world to be useful.

What I want to focus on is the second part, the notion that representations are meant to be operated on by algorithms. An algorithm is a systematic way of repeatedly applying mathematical operations to a representation in order to achieve some computational goal. Asking what algorithms a representation supports, therefore, is a matter of asking what mathematical operations can be meaningfully applied to it.

Consider some of the most widely used unsupervised learning techniques and the sorts of representations they produce:

• Clustering maps data points to a discrete set (say, the integers) where the only meaningful operation is equality.
• Nonlinear dimensionality reduction algorithms (e.g. multidimensional scaling, Isomap) map data points to a low-dimensional space where Euclidean distance is meaningful.
• Linear dimensionality reduction algorithms like PCA and factor analysis map data points to a low-dimensional space where Euclidean distance, linear combination, and dot products are all meaningful.

This list doesn't tell the whole story: there are a lot of variants on each of these categories, and there are more nuances in terms of precisely which operations are meaningful. For instance, in some versions of MDS, only the rank order of the Euclidean distances is meaningful. Still, I think it gives a useful coarse description of each of these categories.

There are also a wide variety of representation learning algorithms built around mathematical operations other than the ones given above:

• Greater/less than. TrueSkill is a model of people's performance in multiplayer games which represents each player with a scalar, and whoever has a larger value is expected to win. [2]
• $L_p$ distance . There is a variant of MDS which uses general $L_p$ distance rather than Euclidean distance. [3]
• Hamming distance. Semantic hashing is a technique in image retrieval which tries to represent images in terms of binary representations where the Hamming distance reflects the semantic dissimilarity between the images. The idea is that we can use the binary representation as a "hash key" and find semantically similar images by flipping random bits and finding the corresponding hash table entries. [4]
• Projection onto an axis. We might want the individual coordinates of a representation to represent something meaningful. As Shepard discusses, MDS with L1 distance often results in such a representation. Another MDS-based technique he describes is INDSCAL, which tries to model multiple subjects' similarity judgments with MDS in a single coordinate system, but where the individual axes can be rescaled for each subject. [3]
• Matrix multiplication. Paccanaro and Hinton [5] proposed a model called linear relational embedding, where entities are represented by vectors and relations between them are represented by matrices. Matrix multiplication corresponds to function composition: if we compose "father of" and "father of," we get the relation "paternal grandfather of."

Roger Shepard's classic 1980 paper, " Multidimensional scaling, tree fitting, and clustering ," [3] highlights a variety of different representation learning methods, and I took a lot of my examples from it. The paper is written from a cognitive science perspective, where the algorithms are used to model human similarity judgments and reaction time data, with the goal of understanding what our internal mental representations might be like. But the same kinds of models are widely used today, albeit in a more modern form, in statistics and machine learning.

One of the most exciting threads of representation learning in recent years has been learning feature representations which could be fed into standard machine learning (usually supervised learning) algorithms. Depending on the intended learning algorithm, the representation has to support some set of operations. For instance,

• For nonparametric regression/classification methods like nearest neighbors or an SVM with an RBF kernel, we want Euclidean distance to be meaningful.
• For linear classifiers, we want dot products to be meaningful.
• For decision trees or L1 regularization, we want projection onto an axis to be meaningful.

For instance, PCA is often used as a preprocessing step in linear regression or kernel SVMs, because it produces a low-dimensional representation in which Euclidean distance and dot products are likely to be more semantically meaningful than the same operations in the original space. However, it would probably not be a good preprocessing step for decision trees or L1 regularization, since the axes themselves (other than possibly the first few) don't have any special meaning. The issues have become somewhat blurred with the advent of deep learning, since we can argue endlessly about which parts of the representation are semantically meaningful. Still, I think these issues at least implicitly guide the research in the area.

The question of which operations and algorithms are supported is, of course, only one aspect of the representation learning problem. We also want representations which where both the mapping and the inverse mapping can be computed efficiently, which can be learned in a data efficient way, and so on. But I still find this a useful way to think about the relationships between different representation learning algorithms.

[1] David Marr. Vision: a Computational Investigation . W. H. Freeman and company, 1982.

[2] Ralf Herbrich, Tom Minka, and Thore Graepel,  TrueSkill(TM): A Bayesian Skill Rating System , in  Advances in Neural Information Processing Systems 20 , MIT Press, January 2007

[3] Roger Shepard. "Multidimensional scaling, tree-fitting, and clustering." Science, 1980.

[4] Ruslan Salakhutdinov and Geoffrey Hinton. "Semantic Hashing." SIGIR Workshop on Information Retrieval and Applications of Graphical Models, 2007.

[5] Alberto Paccanaro and Geoffrey Hinton. "Learning distributed representations of concepts using linear relational embedding." KDE 2000.

Share: Twitter , Facebook

Representation Learning Techniques: An Overview

• Conference paper
• First Online: 29 January 2020
• Cite this conference paper

• Hassan Khastavaneh   ORCID: orcid.org/0000-0002-1333-3838 7 &
• Hossein Ebrahimpour-Komleh   ORCID: orcid.org/0000-0002-9935-7821 7  

Part of the book series: Lecture Notes on Data Engineering and Communications Technologies ((LNDECT,volume 45))

Included in the following conference series:

• The 7th International Conference on Contemporary Issues in Data Science

1300 Accesses

2 Citations

Representation learning techniques, as a paradigm shift in feature generation, are considered as an important and inevitable part of state of the art pattern recognition systems. These techniques attempt to extract and abstract key information from raw input data. Representation learning based methods of feature generation are in contrast to handy feature generation methods which are mainly based on the prior knowledge of expert about the task at hand. Moreover, new techniques of representation learning revolutionized modern pattern recognition systems. Representation learning methods are considered in four main approaches: sub-space based, manifold based, shallow architectures, and deep architectures. This study demonstrates deep architectures are considered as one of the most important methods of representation learning as they cover more general priors of real-world intelligence as a necessity for modern intelligent systems. In other words, deep architectures overcome limitations of their shallow counterparts. In this study, the relationships between various representation learning techniques are highlighted and their advantages and disadvantages are discussed.

This is a preview of subscription content, log in via an institution to check access.

Access this chapter

• Available as PDF
• Read on any device
• Instant download
• Own it forever
• Available as EPUB and PDF
• Compact, lightweight edition
• Dispatched in 3 to 5 business days
• Free shipping worldwide - see info

Tax calculation will be finalised at checkout

Purchases are for personal use only

Institutional subscriptions

Bengio, Y., Lecun, Y.: Scaling learning algorithms towards AI. In: Large Scale Kernel Machines, pp. 321–360 (2007)

Google Scholar  

Bengio, Y., Courville, A., Vincent, P.: Representation learning: a review and new perspectives. IEEE Trans. Pattern Anal. Mach. Intell. 35 , 1798–1828 (2012)

Article   Google Scholar  

Cadima, J., Jolliffe, I.T.: Loading and correlations in the interpretation of principle components. J. Appl. Stat. 22 , 203–2014 (1995)

Article   MathSciNet   Google Scholar  

Zou, H., Hastie, T., Tibshirani, R.: Sparse principal component analysis. J. Comput. Graph. Stat. 15 , 262–286 (2006)

Schölkopf, B., Smola, A., Müller, K.-R.: Nonlinear component analysis as a kernel eigenvalue problem. Neural Comput. 10 , 1299–1319 (1998)

Zhao, J., Philip, L.H., Kwok, J.T.: Bilinear probabilistic principal component analysis. IEEE Trans. Neural Netw. Learn. Syst. 23 , 492–503 (2012)

Abdi, H.: Multidimensional scaling: eigen-analysis of a distance matrix. In: Encyclopedia of Measurement and Statistics, pp. 598–605 (2007)

Comon, P.: Independent component analysis, a new concept? Sig. Process. 36 , 287–314 (1994)

Hyvärinen, A., Hoyer, P.O., Inki, M.: Topographic independent component analysis. Neural Comput. 13 , 1527–1558 (2001)

Bach, F.R., Jordan, M.I.: Kernel independent component analysis. J. Mach. Learn. Res. 1 , 1–48 (2002)

MathSciNet   MATH   Google Scholar  

Fisher, R.A.: The use of multiple measurements in taxonomic problems. Ann. Eugen. 7 , 179–188 (1936)

Aliyari Ghassabeh, Y., Rudzicz, F., Moghaddam, H.A.: Fast incremental LDA feature extraction. Pattern Recognit. 48 , 1999–2012 (2015)

Belkin, M., Niyogi, P.: Laplacian eigenmaps for dimensionality reduction and data representation. Neural Comput. 15 , 1373–1396 (2003)

Roweis, S., Saul, L.: Nonlinear dimensionality reduction by locally linear embedding. Science 290 , 2323–2326 (2000)

Donoho, D.L., Grimes, C.: Hessian eigenmaps: locally linear embedding techniques for high-dimensional data. In: Proceedings of the National Academy of Sciences, pp. 5591–5596 (2003)

Tenenbaum, J., Silva, V., Langford, J.: A global geometric framework for nonlinear dimensionality reduction. Science 290 , 2319–2323 (2000)

De Silva, V., Tenenbaum, J.B.: Global versus local methods in nonlinear dimensionality reduction. In: Proceedings of the 15th International Conference on Neural Information Processing Systems, pp. 721–728. MIT Press, Cambridge (2002

Brand, M.: Charting a manifold. In: Advances in Neural Information Processing Systems, pp. 961–968 (2002)

Coifman, R.R., Lafon, S.: Diffusion maps. Appl. Comput. Harmon. Anal. 21 , 5–30 (2006)

Bengio, Y.: Learning deep architectures for AI. Found. Trends® Mach. Learn. 2 , 1–127 (2009)

Freund, Y., Haussler, D.: Unsupervised learning of distributions on binary vectors using two layer networks. In: Advances in Neural Information Processing Systems, pp. 912–919 (1992)

Zhang, C.-Y., Chen, C.L.P., Chen, D., Ng, K.T.: MapReduce based distributed learning algorithm for restricted Boltzmann machine. Neurocomputing 198 , 4–11 (2016)

Hinton, G.E.: Training products of experts by minimizing contrastive divergence. Neural Comput. 1800 , 1771–1800 (2002)

Hinton, G.E.: A practical guide to training restricted Boltzmann machines. Neural Net.: Tricks Trade 7700 , 599–619 (2012)

Van Tulder, G., De Bruijne, M.: Combining generative and discriminative representation learning for lung CT analysis with convolutional restricted Boltzmann machines. IEEE Trans. Med. Imaging 35 , 1262–1272 (2016)

Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning representations by back-propagating errors. Nature 323 , 533–536 (1986)

Japkowicz, N., Hanson, S.J., Gluck, M.A.: Nonlinear autoassociation is not equivalent to PCA. Neural Comput. 12 , 531–545 (2000)

Ranzato, M.A., Poultney, C., Chopra, S., Cun, Y.L.: Efficient learning of sparse representations with an energy-based model. In: Advances in Neural Information Processing Systems, pp. 1137–1144. MIT Press (2007)

Vincent, P., Larochelle, H., Bengio, Y., Manzagol, P.-A.: Extracting and composing robust features with denoising autoencoders. In: Proceedings of the 25th International Conference on Machine Learning – ICML, pp. 1096–1103. ACM Press, New York (2008)

Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: 2nd International Conference on Learning Representations (ICLR), pp. 1–14 (2014)

LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. Nature 521 , 436–444 (2015)

Bottou, L.: Large-scale machine learning with stochastic gradient descent. In: Proceedings of COMPSTAT, pp. 177–186 (2010)

Chapter   Google Scholar  

Bengio, Y., Lamblin, P., Popovici, D., Larochelle, H.: Greedy layer-wise training of deep networks. In: Advances in Neural Information Processing Systems, pp. 153–160. MIT Press (2007)

Erhan, D., Courville, A., Vincent, P.: Why does unsupervised pre-training help deep learning? J. Mach. Learn. Res. 11 , 625–660 (2010)

Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition (2014)

He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Computer Vision and Pattern Recognition (CVPR), pp. 770–778. IEEE (2016)

Schroff, F., Kalenichenko, D., Philbin, J.: FaceNet: a unified embedding for face recognition and clustering. In: Computer Vision and Pattern Recognition (CVPR), pp. 815–823. IEEE (2015)

Shichijo, S., Nomura, S., Aoyama, K., Nishikawa, Y., Miura, M., Shinagawa, T., Takiyama, H., Tanimoto, T., Ishihara, S., Matsuo, K., Tada, T.: Application of convolutional neural networks in the diagnosis of helicobacter pylori infection based on endoscopic images. EBioMedicine 25 , 106–111 (2017)

Chen, H., Dou, Q., Yu, L., Qin, J., Heng, P.-A.: VoxResNet: deep voxelwise residual networks for brain segmentation from 3D MR images. Neuroimage 170 , 446–455 (2018)

Motlagh, M.H., Jannesari, M., Aboulkheyr, H., Khosravi, P.: Breast cancer histopathological image classification: a deep learning approach, pp. 1–8 (2018)

Yuan, Y., Chao, M., Lo, Y.-C.: Automatic skin lesion segmentation using deep fully convolutional networks with jaccard distance. IEEE Trans. Med. Imaging 36 , 1876–1886 (2017)

Redmon, J., Farhadi, A.: YOLO9000: better, faster, stronger. In: Conference on Computer Vision and Pattern Recognition (CVPR), pp. 6517–6525. IEEE (2017)

Rokach, L.: Ensemble-based classifiers. Artif. Intell. Rev. 33 , 1–39 (2010)

Vincent, P., Larochelle, H.: Stacked denoising autoencoders: learning useful representations in a deep network with a local denoising criterion pierre-antoine manzagol. J. Mach. Learn. Res. 11 , 3371–3408 (2010)

Salakhutdinov, R., Hinton, G.: Semantic hashing. Int. J. Approx. Reason. 50 , 969–978 (2009)

Krizhevsky, A., Hinton, G.: Using very deep autoencoders for content-based image retrieval. In: Proceedings of the European Symposium on Artificial Neural Networks, pp. 1–7 (2011)

Download references

Author information

Authors and affiliations.

University of Kashan, Kashan, Esfahan, Iran

Hassan Khastavaneh & Hossein Ebrahimpour-Komleh

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Hassan Khastavaneh .

Editor information

Editors and affiliations.

Institute for Advanced Studies in Basic Science, Zanjan, Iran

Mahdi Bohlouli

Bahram Sadeghi Bigham

Zahra Narimani

Mahdi Vasighi

Ebrahim Ansari

Rights and permissions

Reprints and permissions

Copyright information

© 2020 Springer Nature Switzerland AG

About this paper

Cite this paper.

Khastavaneh, H., Ebrahimpour-Komleh, H. (2020). Representation Learning Techniques: An Overview. In: Bohlouli, M., Sadeghi Bigham, B., Narimani, Z., Vasighi, M., Ansari, E. (eds) Data Science: From Research to Application. CiDaS 2019. Lecture Notes on Data Engineering and Communications Technologies, vol 45. Springer, Cham. https://doi.org/10.1007/978-3-030-37309-2_8

Download citation

DOI : https://doi.org/10.1007/978-3-030-37309-2_8

Published : 29 January 2020

Publisher Name : Springer, Cham

Print ISBN : 978-3-030-37308-5

Online ISBN : 978-3-030-37309-2

eBook Packages : Intelligent Technologies and Robotics Intelligent Technologies and Robotics (R0)

Share this paper

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Publish with us

Policies and ethics

• Find a journal
• Track your research

• school Campus Bookshelves
• menu_book Bookshelves
• perm_media Learning Objects
• login Login
• how_to_reg Request Instructor Account
• hub Instructor Commons
• Download Page (PDF)
• Download Full Book (PDF)
• Periodic Table
• Physics Constants
• Scientific Calculator
• Reference & Cite
• Tools expand_more
• Readability

selected template will load here

This action is not available.

1.3: Presentation of Data

• Last updated
• Save as PDF
• Page ID 577

Learning Objectives

• To learn two ways that data will be presented in the text.

In this book we will use two formats for presenting data sets. The first is a data list, which is an explicit listing of all the individual measurements, either as a display with space between the individual measurements, or in set notation with individual measurements separated by commas.

Example \(\PageIndex{1}\)

The data obtained by measuring the age of \(21\) randomly selected students enrolled in freshman courses at a university could be presented as the data list:

\[\begin{array}{cccccccccc}18 & 18 & 19 & 19 & 19 & 18 & 22 & 20 & 18 & 18 & 17 \\ 19 & 18 & 24 & 18 & 20 & 18 & 21 & 20 & 17 & 19 &\end{array} \nonumber \]

or in set notation as:

\[ \{18,18,19,19,19,18,22,20,18,18,17,19,18,24,18,20,18,21,20,17,19\} \nonumber \]

A data set can also be presented by means of a data frequency table, a table in which each distinct value \(x\) is listed in the first row and its frequency \(f\), which is the number of times the value \(x\) appears in the data set, is listed below it in the second row.

Example \(\PageIndex{2}\)

The data set of the previous example is represented by the data frequency table

\[\begin{array}{c|cccccc}x & 17 & 18 & 19 & 20 & 21 & 22 & 24 \\ \hline f & 2 & 8 & 5 & 3 & 1 & 1 & 1\end{array} \nonumber \]

The data frequency table is especially convenient when data sets are large and the number of distinct values is not too large.

Key Takeaway

• Data sets can be presented either by listing all the elements or by giving a table of values and frequencies.

Data representation

Computers use binary - the digits 0 and 1 - to store data. A binary digit, or bit, is the smallest unit of data in computing. It is represented by a 0 or a 1. Binary numbers are made up of binary digits (bits), eg the binary number 1001. The circuits in a computer's processor are made up of billions of transistors. A transistor is a tiny switch that is activated by the electronic signals it receives. The digits 1 and 0 used in binary reflect the on and off states of a transistor. Computer programs are sets of instructions. Each instruction is translated into machine code - simple binary codes that activate the CPU. Programmers write computer code and this is converted by a translator into binary instructions that the processor can execute. All software, music, documents, and any other information that is processed by a computer, is also stored using binary. [1]

To include strings, integers, characters and colours. This should include considering the space taken by data, for instance the relation between the hexadecimal representation of colours and the number of colours available.

This video is superb place to understand this topic

• 1 How a file is stored on a computer
• 2 How an image is stored in a computer
• 3 The way in which data is represented in the computer.
• 6 Standards
• 7 References

How a file is stored on a computer [ edit ]

How an image is stored in a computer [ edit ]

The way in which data is represented in the computer. [ edit ].

To include strings, integers, characters and colours. This should include considering the space taken by data, for instance the relation between the hexadecimal representation of colours and the number of colours available [3] .

This helpful material is used with gratitude from a computer science wiki under a Creative Commons Attribution 3.0 License [4]

Sound [ edit ]

• Let's look at an oscilloscope
• The BBC has an excellent article on how computers represent sound

See Also [ edit ]

Standards [ edit ].

• Outline the way in which data is represented in the computer.

References [ edit ]

• ↑ http://www.bbc.co.uk/education/guides/zwsbwmn/revision/1
• ↑ https://ujjwalkarn.me/2016/08/11/intuitive-explanation-convnets/
• ↑ IBO Computer Science Guide, First exams 2014
• ↑ https://compsci2014.wikispaces.com/2.1.10+Outline+the+way+in+which+data+is+represented+in+the+computer

A unit of abstract mathematical system subject to the laws of arithmetic.

A natural number, a negative of a natural number, or zero.

Give a brief account.

• Computer organization
• Very important ideas in computer science

Home Blog Design Understanding Data Presentations (Guide + Examples)

Understanding Data Presentations (Guide + Examples)

In this age of overwhelming information, the skill to effectively convey data has become extremely valuable. Initiating a discussion on data presentation types involves thoughtful consideration of the nature of your data and the message you aim to convey. Different types of visualizations serve distinct purposes. Whether you’re dealing with how to develop a report or simply trying to communicate complex information, how you present data influences how well your audience understands and engages with it. This extensive guide leads you through the different ways of data presentation.

Table of Contents

What is a Data Presentation?

What should a data presentation include, line graphs, treemap chart, scatter plot, how to choose a data presentation type, recommended data presentation templates, common mistakes done in data presentation.

A data presentation is a slide deck that aims to disclose quantitative information to an audience through the use of visual formats and narrative techniques derived from data analysis, making complex data understandable and actionable. This process requires a series of tools, such as charts, graphs, tables, infographics, dashboards, and so on, supported by concise textual explanations to improve understanding and boost retention rate.

Data presentations require us to cull data in a format that allows the presenter to highlight trends, patterns, and insights so that the audience can act upon the shared information. In a few words, the goal of data presentations is to enable viewers to grasp complicated concepts or trends quickly, facilitating informed decision-making or deeper analysis.

Data presentations go beyond the mere usage of graphical elements. Seasoned presenters encompass visuals with the art of data storytelling , so the speech skillfully connects the points through a narrative that resonates with the audience. Depending on the purpose – inspire, persuade, inform, support decision-making processes, etc. – is the data presentation format that is better suited to help us in this journey.

To nail your upcoming data presentation, ensure to count with the following elements:

• Clear Objectives: Understand the intent of your presentation before selecting the graphical layout and metaphors to make content easier to grasp.
• Engaging introduction: Use a powerful hook from the get-go. For instance, you can ask a big question or present a problem that your data will answer. Take a look at our guide on how to start a presentation for tips & insights.
• Structured Narrative: Your data presentation must tell a coherent story. This means a beginning where you present the context, a middle section in which you present the data, and an ending that uses a call-to-action. Check our guide on presentation structure for further information.
• Visual Elements: These are the charts, graphs, and other elements of visual communication we ought to use to present data. This article will cover one by one the different types of data representation methods we can use, and provide further guidance on choosing between them.
• Insights and Analysis: This is not just showcasing a graph and letting people get an idea about it. A proper data presentation includes the interpretation of that data, the reason why it’s included, and why it matters to your research.
• Conclusion & CTA: Ending your presentation with a call to action is necessary. Whether you intend to wow your audience into acquiring your services, inspire them to change the world, or whatever the purpose of your presentation, there must be a stage in which you convey all that you shared and show the path to staying in touch. Plan ahead whether you want to use a thank-you slide, a video presentation, or which method is apt and tailored to the kind of presentation you deliver.
• Q&A Session: After your speech is concluded, allocate 3-5 minutes for the audience to raise any questions about the information you disclosed. This is an extra chance to establish your authority on the topic. Check our guide on questions and answer sessions in presentations here.

Bar charts are a graphical representation of data using rectangular bars to show quantities or frequencies in an established category. They make it easy for readers to spot patterns or trends. Bar charts can be horizontal or vertical, although the vertical format is commonly known as a column chart. They display categorical, discrete, or continuous variables grouped in class intervals [1] . They include an axis and a set of labeled bars horizontally or vertically. These bars represent the frequencies of variable values or the values themselves. Numbers on the y-axis of a vertical bar chart or the x-axis of a horizontal bar chart are called the scale.

Real-Life Application of Bar Charts

Let’s say a sales manager is presenting sales to their audience. Using a bar chart, he follows these steps.

Step 1: Selecting Data

The first step is to identify the specific data you will present to your audience.

The sales manager has highlighted these products for the presentation.

• Product A: Men’s Shoes
• Product B: Women’s Apparel
• Product C: Electronics
• Product D: Home Decor

Step 2: Choosing Orientation

Opt for a vertical layout for simplicity. Vertical bar charts help compare different categories in case there are not too many categories [1] . They can also help show different trends. A vertical bar chart is used where each bar represents one of the four chosen products. After plotting the data, it is seen that the height of each bar directly represents the sales performance of the respective product.

It is visible that the tallest bar (Electronics – Product C) is showing the highest sales. However, the shorter bars (Women’s Apparel – Product B and Home Decor – Product D) need attention. It indicates areas that require further analysis or strategies for improvement.

Step 3: Colorful Insights

Different colors are used to differentiate each product. It is essential to show a color-coded chart where the audience can distinguish between products.

• Men’s Shoes (Product A): Yellow
• Women’s Apparel (Product B): Orange
• Electronics (Product C): Violet
• Home Decor (Product D): Blue

Bar charts are straightforward and easily understandable for presenting data. They are versatile when comparing products or any categorical data [2] . Bar charts adapt seamlessly to retail scenarios. Despite that, bar charts have a few shortcomings. They cannot illustrate data trends over time. Besides, overloading the chart with numerous products can lead to visual clutter, diminishing its effectiveness.

For more information, check our collection of bar chart templates for PowerPoint .

Line graphs help illustrate data trends, progressions, or fluctuations by connecting a series of data points called ‘markers’ with straight line segments. This provides a straightforward representation of how values change [5] . Their versatility makes them invaluable for scenarios requiring a visual understanding of continuous data. In addition, line graphs are also useful for comparing multiple datasets over the same timeline. Using multiple line graphs allows us to compare more than one data set. They simplify complex information so the audience can quickly grasp the ups and downs of values. From tracking stock prices to analyzing experimental results, you can use line graphs to show how data changes over a continuous timeline. They show trends with simplicity and clarity.

Real-life Application of Line Graphs

To understand line graphs thoroughly, we will use a real case. Imagine you’re a financial analyst presenting a tech company’s monthly sales for a licensed product over the past year. Investors want insights into sales behavior by month, how market trends may have influenced sales performance and reception to the new pricing strategy. To present data via a line graph, you will complete these steps.

First, you need to gather the data. In this case, your data will be the sales numbers. For example:

• January: $45,000
• February: $55,000
• March: $45,000
• April: $60,000
• May: $ 70,000
• June: $65,000
• July: $62,000
• August: $68,000
• September: $81,000
• October: $76,000
• November: $87,000
• December: $91,000

After choosing the data, the next step is to select the orientation. Like bar charts, you can use vertical or horizontal line graphs. However, we want to keep this simple, so we will keep the timeline (x-axis) horizontal while the sales numbers (y-axis) vertical.

Step 3: Connecting Trends

After adding the data to your preferred software, you will plot a line graph. In the graph, each month’s sales are represented by data points connected by a line.

Step 4: Adding Clarity with Color

If there are multiple lines, you can also add colors to highlight each one, making it easier to follow.

Line graphs excel at visually presenting trends over time. These presentation aids identify patterns, like upward or downward trends. However, too many data points can clutter the graph, making it harder to interpret. Line graphs work best with continuous data but are not suitable for categories.

For more information, check our collection of line chart templates for PowerPoint and our article about how to make a presentation graph .

A data dashboard is a visual tool for analyzing information. Different graphs, charts, and tables are consolidated in a layout to showcase the information required to achieve one or more objectives. Dashboards help quickly see Key Performance Indicators (KPIs). You don’t make new visuals in the dashboard; instead, you use it to display visuals you’ve already made in worksheets [3] .

Keeping the number of visuals on a dashboard to three or four is recommended. Adding too many can make it hard to see the main points [4]. Dashboards can be used for business analytics to analyze sales, revenue, and marketing metrics at a time. They are also used in the manufacturing industry, as they allow users to grasp the entire production scenario at the moment while tracking the core KPIs for each line.

Real-Life Application of a Dashboard

Consider a project manager presenting a software development project’s progress to a tech company’s leadership team. He follows the following steps.

Step 1: Defining Key Metrics

To effectively communicate the project’s status, identify key metrics such as completion status, budget, and bug resolution rates. Then, choose measurable metrics aligned with project objectives.

Step 2: Choosing Visualization Widgets

After finalizing the data, presentation aids that align with each metric are selected. For this project, the project manager chooses a progress bar for the completion status and uses bar charts for budget allocation. Likewise, he implements line charts for bug resolution rates.

Step 3: Dashboard Layout

Key metrics are prominently placed in the dashboard for easy visibility, and the manager ensures that it appears clean and organized.

Dashboards provide a comprehensive view of key project metrics. Users can interact with data, customize views, and drill down for detailed analysis. However, creating an effective dashboard requires careful planning to avoid clutter. Besides, dashboards rely on the availability and accuracy of underlying data sources.

For more information, check our article on how to design a dashboard presentation , and discover our collection of dashboard PowerPoint templates .

Treemap charts represent hierarchical data structured in a series of nested rectangles [6] . As each branch of the ‘tree’ is given a rectangle, smaller tiles can be seen representing sub-branches, meaning elements on a lower hierarchical level than the parent rectangle. Each one of those rectangular nodes is built by representing an area proportional to the specified data dimension.

Treemaps are useful for visualizing large datasets in compact space. It is easy to identify patterns, such as which categories are dominant. Common applications of the treemap chart are seen in the IT industry, such as resource allocation, disk space management, website analytics, etc. Also, they can be used in multiple industries like healthcare data analysis, market share across different product categories, or even in finance to visualize portfolios.

Real-Life Application of a Treemap Chart

Let’s consider a financial scenario where a financial team wants to represent the budget allocation of a company. There is a hierarchy in the process, so it is helpful to use a treemap chart. In the chart, the top-level rectangle could represent the total budget, and it would be subdivided into smaller rectangles, each denoting a specific department. Further subdivisions within these smaller rectangles might represent individual projects or cost categories.

Step 1: Define Your Data Hierarchy

While presenting data on the budget allocation, start by outlining the hierarchical structure. The sequence will be like the overall budget at the top, followed by departments, projects within each department, and finally, individual cost categories for each project.

• Top-level rectangle: Total Budget
• Second-level rectangles: Departments (Engineering, Marketing, Sales)
• Third-level rectangles: Projects within each department
• Fourth-level rectangles: Cost categories for each project (Personnel, Marketing Expenses, Equipment)

Step 2: Choose a Suitable Tool

It’s time to select a data visualization tool supporting Treemaps. Popular choices include Tableau, Microsoft Power BI, PowerPoint, or even coding with libraries like D3.js. It is vital to ensure that the chosen tool provides customization options for colors, labels, and hierarchical structures.

Here, the team uses PowerPoint for this guide because of its user-friendly interface and robust Treemap capabilities.

Step 3: Make a Treemap Chart with PowerPoint

After opening the PowerPoint presentation, they chose “SmartArt” to form the chart. The SmartArt Graphic window has a “Hierarchy” category on the left.  Here, you will see multiple options. You can choose any layout that resembles a Treemap. The “Table Hierarchy” or “Organization Chart” options can be adapted. The team selects the Table Hierarchy as it looks close to a Treemap.

Step 5: Input Your Data

After that, a new window will open with a basic structure. They add the data one by one by clicking on the text boxes. They start with the top-level rectangle, representing the total budget.  

Step 6: Customize the Treemap

By clicking on each shape, they customize its color, size, and label. At the same time, they can adjust the font size, style, and color of labels by using the options in the “Format” tab in PowerPoint. Using different colors for each level enhances the visual difference.

Treemaps excel at illustrating hierarchical structures. These charts make it easy to understand relationships and dependencies. They efficiently use space, compactly displaying a large amount of data, reducing the need for excessive scrolling or navigation. Additionally, using colors enhances the understanding of data by representing different variables or categories.

In some cases, treemaps might become complex, especially with deep hierarchies.  It becomes challenging for some users to interpret the chart. At the same time, displaying detailed information within each rectangle might be constrained by space. It potentially limits the amount of data that can be shown clearly. Without proper labeling and color coding, there’s a risk of misinterpretation.

A heatmap is a data visualization tool that uses color coding to represent values across a two-dimensional surface. In these, colors replace numbers to indicate the magnitude of each cell. This color-shaded matrix display is valuable for summarizing and understanding data sets with a glance [7] . The intensity of the color corresponds to the value it represents, making it easy to identify patterns, trends, and variations in the data.

As a tool, heatmaps help businesses analyze website interactions, revealing user behavior patterns and preferences to enhance overall user experience. In addition, companies use heatmaps to assess content engagement, identifying popular sections and areas of improvement for more effective communication. They excel at highlighting patterns and trends in large datasets, making it easy to identify areas of interest.

We can implement heatmaps to express multiple data types, such as numerical values, percentages, or even categorical data. Heatmaps help us easily spot areas with lots of activity, making them helpful in figuring out clusters [8] . When making these maps, it is important to pick colors carefully. The colors need to show the differences between groups or levels of something. And it is good to use colors that people with colorblindness can easily see.

Check our detailed guide on how to create a heatmap here. Also discover our collection of heatmap PowerPoint templates .

Pie charts are circular statistical graphics divided into slices to illustrate numerical proportions. Each slice represents a proportionate part of the whole, making it easy to visualize the contribution of each component to the total.

The size of the pie charts is influenced by the value of data points within each pie. The total of all data points in a pie determines its size. The pie with the highest data points appears as the largest, whereas the others are proportionally smaller. However, you can present all pies of the same size if proportional representation is not required [9] . Sometimes, pie charts are difficult to read, or additional information is required. A variation of this tool can be used instead, known as the donut chart , which has the same structure but a blank center, creating a ring shape. Presenters can add extra information, and the ring shape helps to declutter the graph.

Pie charts are used in business to show percentage distribution, compare relative sizes of categories, or present straightforward data sets where visualizing ratios is essential.

Real-Life Application of Pie Charts

Consider a scenario where you want to represent the distribution of the data. Each slice of the pie chart would represent a different category, and the size of each slice would indicate the percentage of the total portion allocated to that category.

Step 1: Define Your Data Structure

Imagine you are presenting the distribution of a project budget among different expense categories.

• Column A: Expense Categories (Personnel, Equipment, Marketing, Miscellaneous)
• Column B: Budget Amounts ($40,000, $30,000, $20,000, $10,000) Column B represents the values of your categories in Column A.

Step 2: Insert a Pie Chart

Using any of the accessible tools, you can create a pie chart. The most convenient tools for forming a pie chart in a presentation are presentation tools such as PowerPoint or Google Slides.  You will notice that the pie chart assigns each expense category a percentage of the total budget by dividing it by the total budget.

For instance:

• Personnel: $40,000 / ($40,000 + $30,000 + $20,000 + $10,000) = 40%
• Equipment: $30,000 / ($40,000 + $30,000 + $20,000 + $10,000) = 30%
• Marketing: $20,000 / ($40,000 + $30,000 + $20,000 + $10,000) = 20%
• Miscellaneous: $10,000 / ($40,000 + $30,000 + $20,000 + $10,000) = 10%

You can make a chart out of this or just pull out the pie chart from the data.

3D pie charts and 3D donut charts are quite popular among the audience. They stand out as visual elements in any presentation slide, so let’s take a look at how our pie chart example would look in 3D pie chart format.

Step 03: Results Interpretation

The pie chart visually illustrates the distribution of the project budget among different expense categories. Personnel constitutes the largest portion at 40%, followed by equipment at 30%, marketing at 20%, and miscellaneous at 10%. This breakdown provides a clear overview of where the project funds are allocated, which helps in informed decision-making and resource management. It is evident that personnel are a significant investment, emphasizing their importance in the overall project budget.

Pie charts provide a straightforward way to represent proportions and percentages. They are easy to understand, even for individuals with limited data analysis experience. These charts work well for small datasets with a limited number of categories.

However, a pie chart can become cluttered and less effective in situations with many categories. Accurate interpretation may be challenging, especially when dealing with slight differences in slice sizes. In addition, these charts are static and do not effectively convey trends over time.

For more information, check our collection of pie chart templates for PowerPoint .

Histograms present the distribution of numerical variables. Unlike a bar chart that records each unique response separately, histograms organize numeric responses into bins and show the frequency of reactions within each bin [10] . The x-axis of a histogram shows the range of values for a numeric variable. At the same time, the y-axis indicates the relative frequencies (percentage of the total counts) for that range of values.

Whenever you want to understand the distribution of your data, check which values are more common, or identify outliers, histograms are your go-to. Think of them as a spotlight on the story your data is telling. A histogram can provide a quick and insightful overview if you’re curious about exam scores, sales figures, or any numerical data distribution.

Real-Life Application of a Histogram

In the histogram data analysis presentation example, imagine an instructor analyzing a class’s grades to identify the most common score range. A histogram could effectively display the distribution. It will show whether most students scored in the average range or if there are significant outliers.

Step 1: Gather Data

He begins by gathering the data. The scores of each student in class are gathered to analyze exam scores.

After arranging the scores in ascending order, bin ranges are set.

Step 2: Define Bins

Bins are like categories that group similar values. Think of them as buckets that organize your data. The presenter decides how wide each bin should be based on the range of the values. For instance, the instructor sets the bin ranges based on score intervals: 60-69, 70-79, 80-89, and 90-100.

Step 3: Count Frequency

Now, he counts how many data points fall into each bin. This step is crucial because it tells you how often specific ranges of values occur. The result is the frequency distribution, showing the occurrences of each group.

Here, the instructor counts the number of students in each category.

• 60-69: 1 student (Kate)
• 70-79: 4 students (David, Emma, Grace, Jack)
• 80-89: 7 students (Alice, Bob, Frank, Isabel, Liam, Mia, Noah)
• 90-100: 3 students (Clara, Henry, Olivia)

Step 4: Create the Histogram

It’s time to turn the data into a visual representation. Draw a bar for each bin on a graph. The width of the bar should correspond to the range of the bin, and the height should correspond to the frequency.  To make your histogram understandable, label the X and Y axes.

In this case, the X-axis should represent the bins (e.g., test score ranges), and the Y-axis represents the frequency.

The histogram of the class grades reveals insightful patterns in the distribution. Most students, with seven students, fall within the 80-89 score range. The histogram provides a clear visualization of the class’s performance. It showcases a concentration of grades in the upper-middle range with few outliers at both ends. This analysis helps in understanding the overall academic standing of the class. It also identifies the areas for potential improvement or recognition.

Thus, histograms provide a clear visual representation of data distribution. They are easy to interpret, even for those without a statistical background. They apply to various types of data, including continuous and discrete variables. One weak point is that histograms do not capture detailed patterns in students’ data, with seven compared to other visualization methods.

A scatter plot is a graphical representation of the relationship between two variables. It consists of individual data points on a two-dimensional plane. This plane plots one variable on the x-axis and the other on the y-axis. Each point represents a unique observation. It visualizes patterns, trends, or correlations between the two variables.

Scatter plots are also effective in revealing the strength and direction of relationships. They identify outliers and assess the overall distribution of data points. The points’ dispersion and clustering reflect the relationship’s nature, whether it is positive, negative, or lacks a discernible pattern. In business, scatter plots assess relationships between variables such as marketing cost and sales revenue. They help present data correlations and decision-making.

Real-Life Application of Scatter Plot

A group of scientists is conducting a study on the relationship between daily hours of screen time and sleep quality. After reviewing the data, they managed to create this table to help them build a scatter plot graph:

In the provided example, the x-axis represents Daily Hours of Screen Time, and the y-axis represents the Sleep Quality Rating.

The scientists observe a negative correlation between the amount of screen time and the quality of sleep. This is consistent with their hypothesis that blue light, especially before bedtime, has a significant impact on sleep quality and metabolic processes.

There are a few things to remember when using a scatter plot. Even when a scatter diagram indicates a relationship, it doesn’t mean one variable affects the other. A third factor can influence both variables. The more the plot resembles a straight line, the stronger the relationship is perceived [11] . If it suggests no ties, the observed pattern might be due to random fluctuations in data. When the scatter diagram depicts no correlation, whether the data might be stratified is worth considering.

Choosing the appropriate data presentation type is crucial when making a presentation . Understanding the nature of your data and the message you intend to convey will guide this selection process. For instance, when showcasing quantitative relationships, scatter plots become instrumental in revealing correlations between variables. If the focus is on emphasizing parts of a whole, pie charts offer a concise display of proportions. Histograms, on the other hand, prove valuable for illustrating distributions and frequency patterns. 

Bar charts provide a clear visual comparison of different categories. Likewise, line charts excel in showcasing trends over time, while tables are ideal for detailed data examination. Starting a presentation on data presentation types involves evaluating the specific information you want to communicate and selecting the format that aligns with your message. This ensures clarity and resonance with your audience from the beginning of your presentation.

1. Fact Sheet Dashboard for Data Presentation

Convey all the data you need to present in this one-pager format, an ideal solution tailored for users looking for presentation aids. Global maps, donut chats, column graphs, and text neatly arranged in a clean layout presented in light and dark themes.

Use This Template

2. 3D Column Chart Infographic PPT Template

Represent column charts in a highly visual 3D format with this PPT template. A creative way to present data, this template is entirely editable, and we can craft either a one-page infographic or a series of slides explaining what we intend to disclose point by point.

3. Data Circles Infographic PowerPoint Template

An alternative to the pie chart and donut chart diagrams, this template features a series of curved shapes with bubble callouts as ways of presenting data. Expand the information for each arch in the text placeholder areas.

4. Colorful Metrics Dashboard for Data Presentation

This versatile dashboard template helps us in the presentation of the data by offering several graphs and methods to convert numbers into graphics. Implement it for e-commerce projects, financial projections, project development, and more.

5. Animated Data Presentation Tools for PowerPoint & Google Slides

A slide deck filled with most of the tools mentioned in this article, from bar charts, column charts, treemap graphs, pie charts, histogram, etc. Animated effects make each slide look dynamic when sharing data with stakeholders.

6. Statistics Waffle Charts PPT Template for Data Presentations

This PPT template helps us how to present data beyond the typical pie chart representation. It is widely used for demographics, so it’s a great fit for marketing teams, data science professionals, HR personnel, and more.

7. Data Presentation Dashboard Template for Google Slides

A compendium of tools in dashboard format featuring line graphs, bar charts, column charts, and neatly arranged placeholder text areas. 

8. Weather Dashboard for Data Presentation

Share weather data for agricultural presentation topics, environmental studies, or any kind of presentation that requires a highly visual layout for weather forecasting on a single day. Two color themes are available.

9. Social Media Marketing Dashboard Data Presentation Template

Intended for marketing professionals, this dashboard template for data presentation is a tool for presenting data analytics from social media channels. Two slide layouts featuring line graphs and column charts.

10. Project Management Summary Dashboard Template

A tool crafted for project managers to deliver highly visual reports on a project’s completion, the profits it delivered for the company, and expenses/time required to execute it. 4 different color layouts are available.

11. Profit & Loss Dashboard for PowerPoint and Google Slides

A must-have for finance professionals. This typical profit & loss dashboard includes progress bars, donut charts, column charts, line graphs, and everything that’s required to deliver a comprehensive report about a company’s financial situation.

Overwhelming visuals

One of the mistakes related to using data-presenting methods is including too much data or using overly complex visualizations. They can confuse the audience and dilute the key message.

Inappropriate chart types

Choosing the wrong type of chart for the data at hand can lead to misinterpretation. For example, using a pie chart for data that doesn’t represent parts of a whole is not right.

Lack of context

Failing to provide context or sufficient labeling can make it challenging for the audience to understand the significance of the presented data.

Inconsistency in design

Using inconsistent design elements and color schemes across different visualizations can create confusion and visual disarray.

Failure to provide details

Simply presenting raw data without offering clear insights or takeaways can leave the audience without a meaningful conclusion.

Lack of focus

Not having a clear focus on the key message or main takeaway can result in a presentation that lacks a central theme.

Visual accessibility issues

Overlooking the visual accessibility of charts and graphs can exclude certain audience members who may have difficulty interpreting visual information.

In order to avoid these mistakes in data presentation, presenters can benefit from using presentation templates . These templates provide a structured framework. They ensure consistency, clarity, and an aesthetically pleasing design, enhancing data communication’s overall impact.

Understanding and choosing data presentation types are pivotal in effective communication. Each method serves a unique purpose, so selecting the appropriate one depends on the nature of the data and the message to be conveyed. The diverse array of presentation types offers versatility in visually representing information, from bar charts showing values to pie charts illustrating proportions. 

Using the proper method enhances clarity, engages the audience, and ensures that data sets are not just presented but comprehensively understood. By appreciating the strengths and limitations of different presentation types, communicators can tailor their approach to convey information accurately, developing a deeper connection between data and audience understanding.

[1] Government of Canada, S.C. (2021) 5 Data Visualization 5.2 Bar Chart , 5.2 Bar chart .  https://www150.statcan.gc.ca/n1/edu/power-pouvoir/ch9/bargraph-diagrammeabarres/5214818-eng.htm

[2] Kosslyn, S.M., 1989. Understanding charts and graphs. Applied cognitive psychology, 3(3), pp.185-225. https://apps.dtic.mil/sti/pdfs/ADA183409.pdf

[3] Creating a Dashboard . https://it.tufts.edu/book/export/html/1870

[4] https://www.goldenwestcollege.edu/research/data-and-more/data-dashboards/index.html

[5] https://www.mit.edu/course/21/21.guide/grf-line.htm

[6] Jadeja, M. and Shah, K., 2015, January. Tree-Map: A Visualization Tool for Large Data. In GSB@ SIGIR (pp. 9-13). https://ceur-ws.org/Vol-1393/gsb15proceedings.pdf#page=15

[7] Heat Maps and Quilt Plots. https://www.publichealth.columbia.edu/research/population-health-methods/heat-maps-and-quilt-plots

[8] EIU QGIS WORKSHOP. https://www.eiu.edu/qgisworkshop/heatmaps.php

[9] About Pie Charts.  https://www.mit.edu/~mbarker/formula1/f1help/11-ch-c8.htm

[10] Histograms. https://sites.utexas.edu/sos/guided/descriptive/numericaldd/descriptiven2/histogram/ [11] https://asq.org/quality-resources/scatter-diagram

Like this article? Please share

Data Analysis, Data Science, Data Visualization Filed under Design

Related Articles

Filed under Design • March 27th, 2024

How to Make a Presentation Graph

Detailed step-by-step instructions to master the art of how to make a presentation graph in PowerPoint and Google Slides. Check it out!

Filed under Presentation Ideas • January 6th, 2024

All About Using Harvey Balls

Among the many tools in the arsenal of the modern presenter, Harvey Balls have a special place. In this article we will tell you all about using Harvey Balls.

Filed under Business • December 8th, 2023

How to Design a Dashboard Presentation: A Step-by-Step Guide

Take a step further in your professional presentation skills by learning what a dashboard presentation is and how to properly design one in PowerPoint. A detailed step-by-step guide is here!

Leave a Reply

If you're seeing this message, it means we're having trouble loading external resources on our website.

If you're behind a web filter, please make sure that the domains *.kastatic.org and *.kasandbox.org are unblocked.

To log in and use all the features of Khan Academy, please enable JavaScript in your browser.

Unit 15: Represent and interpret data

About this unit.

This unit explores data representation through stem and leaf plots, line plots with fractions, and interpreting line plots. You'll also delve into the center and spread of data, focusing on median and range.

Stem and leaf plots

• Stem-and-leaf plots (Opens a modal)
• Reading stem and leaf plots (Opens a modal)
• Reading stem and leaf plots Get 3 of 4 questions to level up!

Line plots with fractions

• Measuring lengths to nearest 1/4 unit (Opens a modal)
• Read line plots (Opens a modal)
• Graphing data on line plots (Opens a modal)
• Making line plots with fractional data (Opens a modal)
• Interpreting line plots with fractions (Opens a modal)
• Line plots review (Opens a modal)
• Measure lengths to nearest 1/4 unit Get 3 of 4 questions to level up!
• Graph data on line plots Get 3 of 4 questions to level up!
• Graph data on line plots (through 1/8 of a unit) Get 3 of 4 questions to level up!
• Read line plots (data with fractions) Get 3 of 4 questions to level up!

Interpret line plots

• Reading a line plot with fractions (Opens a modal)
• Line plot distribution: trail mix (Opens a modal)
• Interpret line plots Get 3 of 4 questions to level up!
• Interpret line plots with fraction addition and subtraction Get 3 of 4 questions to level up!
• Interpret dot plots with fraction operations Get 3 of 4 questions to level up!

Center and spread of data

• Frequency tables & dot plots (Opens a modal)
• Statistics intro: Mean, median, & mode (Opens a modal)
• Mean, median, & mode example (Opens a modal)
• Median & range puzzlers (Opens a modal)
• Estimate center using dot plots Get 3 of 4 questions to level up!
• Calculating the median Get 3 of 4 questions to level up!
• Median & range puzzlers Get 3 of 4 questions to level up!

10 Methods of Data Presentation with 5 Great Tips to Practice, Best in 2024

Leah Nguyen • 05 Apr 2024 • 11 min read

There are different ways of presenting data, so which one is suited you the most? You can end deathly boring and ineffective data presentation right now with our 10 methods of data presentation . Check out the examples from each technique!

Have you ever presented a data report to your boss/coworkers/teachers thinking it was super dope like you’re some cyber hacker living in the Matrix, but all they saw was a pile of static numbers that seemed pointless and didn’t make sense to them?

Understanding digits is rigid . Making people from non-analytical backgrounds understand those digits is even more challenging.

How can you clear up those confusing numbers in the types of presentation that have the flawless clarity of a diamond? So, let’s check out best way to present data. 💎

Table of Contents

• What are Methods of Data Presentations?
• #1 – Tabular

#2 – Text

#3 – pie chart, #4 – bar chart, #5 – histogram, #6 – line graph, #7 – pictogram graph, #8 – radar chart, #9 – heat map, #10 – scatter plot.

• 5 Mistakes to Avoid
• Best Method of Data Presentation

Frequently Asked Questions

More tips with ahaslides.

• Marketing Presentation
• Survey Result Presentation
• Types of Presentation

Start in seconds.

Get any of the above examples as templates. Sign up for free and take what you want from the template library!

What are Methods of Data Presentation?

The term ’data presentation’ relates to the way you present data in a way that makes even the most clueless person in the room understand. 

Some say it’s witchcraft (you’re manipulating the numbers in some ways), but we’ll just say it’s the power of turning dry, hard numbers or digits into a visual showcase that is easy for people to digest.

Presenting data correctly can help your audience understand complicated processes, identify trends, and instantly pinpoint whatever is going on without exhausting their brains.

Good data presentation helps…

• Make informed decisions and arrive at positive outcomes . If you see the sales of your product steadily increase throughout the years, it’s best to keep milking it or start turning it into a bunch of spin-offs (shoutout to Star Wars👀).
• Reduce the time spent processing data . Humans can digest information graphically 60,000 times faster than in the form of text. Grant them the power of skimming through a decade of data in minutes with some extra spicy graphs and charts.
• Communicate the results clearly . Data does not lie. They’re based on factual evidence and therefore if anyone keeps whining that you might be wrong, slap them with some hard data to keep their mouths shut.
• Add to or expand the current research . You can see what areas need improvement, as well as what details often go unnoticed while surfing through those little lines, dots or icons that appear on the data board.

Methods of Data Presentation and Examples

Imagine you have a delicious pepperoni, extra-cheese pizza. You can decide to cut it into the classic 8 triangle slices, the party style 12 square slices, or get creative and abstract on those slices. 

There are various ways for cutting a pizza and you get the same variety with how you present your data. In this section, we will bring you the 10 ways to slice a pizza – we mean to present your data – that will make your company’s most important asset as clear as day. Let’s dive into 10 ways to present data efficiently.

#1 – Tabular 

Among various types of data presentation, tabular is the most fundamental method, with data presented in rows and columns. Excel or Google Sheets would qualify for the job. Nothing fancy.

This is an example of a tabular presentation of data on Google Sheets. Each row and column has an attribute (year, region, revenue, etc.), and you can do a custom format to see the change in revenue throughout the year.

When presenting data as text, all you do is write your findings down in paragraphs and bullet points, and that’s it. A piece of cake to you, a tough nut to crack for whoever has to go through all of the reading to get to the point.

• 65% of email users worldwide access their email via a mobile device.
• Emails that are optimised for mobile generate 15% higher click-through rates.
• 56% of brands using emojis in their email subject lines had a higher open rate.

(Source: CustomerThermometer )

All the above quotes present statistical information in textual form. Since not many people like going through a wall of texts, you’ll have to figure out another route when deciding to use this method, such as breaking the data down into short, clear statements, or even as catchy puns if you’ve got the time to think of them.

A pie chart (or a ‘donut chart’ if you stick a hole in the middle of it) is a circle divided into slices that show the relative sizes of data within a whole. If you’re using it to show percentages, make sure all the slices add up to 100%.

The pie chart is a familiar face at every party and is usually recognised by most people. However, one setback of using this method is our eyes sometimes can’t identify the differences in slices of a circle, and it’s nearly impossible to compare similar slices from two different pie charts, making them the villains in the eyes of data analysts.

Bonus example: A literal ‘pie’ chart! 🥧

The bar chart is a chart that presents a bunch of items from the same category, usually in the form of rectangular bars that are placed at an equal distance from each other. Their heights or lengths depict the values they represent.

They can be as simple as this:

Or more complex and detailed like this example of presentation of data. Contributing to an effective statistic presentation, this one is a grouped bar chart that not only allows you to compare categories but also the groups within them as well.

Similar in appearance to the bar chart but the rectangular bars in histograms don’t often have the gap like their counterparts.

Instead of measuring categories like weather preferences or favourite films as a bar chart does, a histogram only measures things that can be put into numbers.

Teachers can use presentation graphs like a histogram to see which score group most of the students fall into, like in this example above.

Recordings to ways of displaying data, we shouldn’t overlook the effectiveness of line graphs. Line graphs are represented by a group of data points joined together by a straight line. There can be one or more lines to compare how several related things change over time. 

On a line chart’s horizontal axis, you usually have text labels, dates or years, while the vertical axis usually represents the quantity (e.g.: budget, temperature or percentage).

A pictogram graph uses pictures or icons relating to the main topic to visualise a small dataset. The fun combination of colours and illustrations makes it a frequent use at schools.

Pictograms are a breath of fresh air if you want to stay away from the monotonous line chart or bar chart for a while. However, they can present a very limited amount of data and sometimes they are only there for displays and do not represent real statistics.

If presenting five or more variables in the form of a bar chart is too stuffy then you should try using a radar chart, which is one of the most creative ways to present data.

Radar charts show data in terms of how they compare to each other starting from the same point. Some also call them ‘spider charts’ because each aspect combined looks like a spider web.

Radar charts can be a great use for parents who’d like to compare their child’s grades with their peers to lower their self-esteem. You can see that each angular represents a subject with a score value ranging from 0 to 100. Each student’s score across 5 subjects is highlighted in a different colour.

If you think that this method of data presentation somehow feels familiar, then you’ve probably encountered one while playing Pokémon .

A heat map represents data density in colours. The bigger the number, the more colour intense that data will be represented.

Most U.S citizens would be familiar with this data presentation method in geography. For elections, many news outlets assign a specific colour code to a state, with blue representing one candidate and red representing the other. The shade of either blue or red in each state shows the strength of the overall vote in that state.

Another great thing you can use a heat map for is to map what visitors to your site click on. The more a particular section is clicked the ‘hotter’ the colour will turn, from blue to bright yellow to red.

If you present your data in dots instead of chunky bars, you’ll have a scatter plot. 

A scatter plot is a grid with several inputs showing the relationship between two variables. It’s good at collecting seemingly random data and revealing some telling trends.

For example, in this graph, each dot shows the average daily temperature versus the number of beach visitors across several days. You can see that the dots get higher as the temperature increases, so it’s likely that hotter weather leads to more visitors.

5 Data Presentation Mistakes to Avoid

#1 – assume your audience understands what the numbers represent.

You may know all the behind-the-scenes of your data since you’ve worked with them for weeks, but your audience doesn’t.

Showing without telling only invites more and more questions from your audience, as they have to constantly make sense of your data, wasting the time of both sides as a result.

While showing your data presentations, you should tell them what the data are about before hitting them with waves of numbers first. You can use interactive activities such as polls , word clouds , online quiz and Q&A sections , combined with icebreaker games , to assess their understanding of the data and address any confusion beforehand.

#2 – Use the wrong type of chart

Charts such as pie charts must have a total of 100% so if your numbers accumulate to 193% like this example below, you’re definitely doing it wrong.

Before making a chart, ask yourself: what do I want to accomplish with my data? Do you want to see the relationship between the data sets, show the up and down trends of your data, or see how segments of one thing make up a whole?

Remember, clarity always comes first. Some data visualisations may look cool, but if they don’t fit your data, steer clear of them. 

#3 – Make it 3D

3D is a fascinating graphical presentation example. The third dimension is cool, but full of risks.

Can you see what’s behind those red bars? Because we can’t either. You may think that 3D charts add more depth to the design, but they can create false perceptions as our eyes see 3D objects closer and bigger than they appear, not to mention they cannot be seen from multiple angles.

#4 – Use different types of charts to compare contents in the same category

This is like comparing a fish to a monkey. Your audience won’t be able to identify the differences and make an appropriate correlation between the two data sets. 

Next time, stick to one type of data presentation only. Avoid the temptation of trying various data visualisation methods in one go and make your data as accessible as possible.

#5 – Bombard the audience with too much information

The goal of data presentation is to make complex topics much easier to understand, and if you’re bringing too much information to the table, you’re missing the point.

The more information you give, the more time it will take for your audience to process it all. If you want to make your data understandable and give your audience a chance to remember it, keep the information within it to an absolute minimum. You should set your session with open-ended questions , to avoid dead-communication!

What are the Best Methods of Data Presentation?

Finally, which is the best way to present data?

The answer is…

There is none 😄 Each type of presentation has its own strengths and weaknesses and the one you choose greatly depends on what you’re trying to do. 

For example:

• Go for a scatter plot if you’re exploring the relationship between different data values, like seeing whether the sales of ice cream go up because of the temperature or because people are just getting more hungry and greedy each day?
• Go for a line graph if you want to mark a trend over time. 
• Go for a heat map if you like some fancy visualisation of the changes in a geographical location, or to see your visitors’ behaviour on your website.
• Go for a pie chart (especially in 3D) if you want to be shunned by others because it was never a good idea👇

What is chart presentation?

A chart presentation is a way of presenting data or information using visual aids such as charts, graphs, and diagrams. The purpose of a chart presentation is to make complex information more accessible and understandable for the audience.

When can I use charts for presentation?

Charts can be used to compare data, show trends over time, highlight patterns, and simplify complex information.

Why should use charts for presentation?

You should use charts to ensure your contents and visual look clean, as they are the visual representative, provide clarity, simplicity, comparison, contrast and super time-saving!

What are the 4 graphical methods of presenting data?

Histogram, Smoothed frequency graph, Pie diagram or Pie chart, Cumulative or ogive frequency graph, and Frequency Polygon.

Leah Nguyen

Words that convert, stories that stick. I turn complex ideas into engaging narratives - helping audiences learn, remember, and take action.

More from AhaSlides