Panel Studies

views updated May 23 2018

Panel Studies

The turnover table



The potentials of panel analysis were first developed at Columbia University under the aegis of Paul F. Lazarsfeld. The first major panel study carried out under the techniques pioneered by Lazarsfeld was a study of voter decision making during the 1940 presidential election campaign (Lazarsfeld et al. 1944). Since then there have been numerous panel studies in a variety of sociological, political, and economic areas. Panel studies have been made of other elections in the United States and other countries, of the socialization of medical students to professional norms, and of social climates in high schools.

The idea behind panel analysis is deceptively simple. Instead of comparing aggregates over time, panel analysis compares repeated observations ofindividuals. Insofar as the study of social change is limited to net changes in a social aggregate, the absence of such net changes is often assumed to be indicative of social stability. But constancy in the aggregate may obscure considerable compensatory change among individuals. For example, the distribution of income in a society may show no net change over a ten-year period, but various processes of economic mobility may be at work. In short, panel analysis gives rise to the study of an aspect of social change that tends to be neglected in studies of aggregate trends.

Origins. The technique of using repeated interviews with a constant sample of people was first attempted by Stuart Rice (1928) during the 1924 presidential election campaign. Rice’s panel consisted of students of sociology in the three upper classes at Dartmouth College. His purpose was to find out how preferences for candidates shifted during the campaign. Though his research design included few variables, Rice’s analysis was quite modern in that he did differentiate between net change and gross change. Theodore Newcomb (1943), in his classic study of Bennington College students, conducted repeated interviews for the four-year period 1935 to 1939. His main interest was in learning what effect a liberal college environment had upon the attitudes of girls coming from well-to-do, conservative families.

In both of these early efforts at panel analysis, the researchers were interested in identifying students who had changed attitudes and in discerning the reasons for the changes. Neither Rice nor Newcomb, however, developed an appropriate technique for handling repeated data. Newcomb described in detail the girls who showed a marked shift from a conservative to a liberal outlook during their stay at Bennington and, conversely, those who were impervious to the liberalism of the faculty. But although he collected a wealth of statistical information in his interviews with students, Newcomb tended to rely on qualitative analysis of the extreme cases of change.

The first systematic statement on the technique of panel analysis was made by Lazarsfeld and Fiske (1938). These researchers reported their experiences and problems with the new panel technique.

Applications. Formally, panel analysis is a research technique for collecting and analyzing data through repeated interviews with a sample of individuals in a natural, rather than a laboratory, setting. Before examining the characteristic modes of panel analysis, it is useful to compare panel analysis with other research techniques in social science that involve over-time data.

Econometrics, for example, makes use of quantitative time series data. Monthly, quarterly, or annual series may cover hundreds of time points, whereas panel studies seldom exceed six interviews. In econometrics, the data typically involve a single geographic or political entity. Though data may be collected from cross-sectional samples of individuals, families, or firms, the interest is usually in the aggregate dynamics of a single economic system. Panel studies typically range from 300 to 3,000 cases, and the interest is in the molecular changes of these hundreds or thousands of individuals.

In experimental psychology, learning studies involve longitudinal data covering hundreds of observations, but the focus is usually on the change in one criterion. The research tends to be highly experimental and theoretical, and involves the use of mathematical models. Panel analysis, on the other hand, is nonexperimental and descriptive. It emphasizes the interrelationships of many changing variables and is statistical, though not highly mathematical. In such panel studies, a set of interlocking variables may be investigated because they are thought to be theoretically fruitful, and there is considerable theoretical improvisation during the analysis of the data.

In educational research, follow-up studies involve over-time data. The interest, however, is not in periodically reobserving the same variables but in correlating a set of predictors with a criterion, as when a battery of aptitude tests is correlated with subsequent vocational success.

In social psychology, research design usually takes the form of controlled experiment involving before-after data obtained from subjects in a laboratory setting. The researchers decide which subjects will be exposed to what stimulus. In panel studies, by contrast, the researchers have no control over the individuals in the study, nor do they manipulate any stimuli. Experiments, unlike panels, are not intended to uncover new concepts or descriptively explore new terrain but, rather, to verify or nullify hypotheses dictated by psychological theory and formulated before the start of the experiments. However, a type of panel study known as the “impact” panel does parallel controlled experiments; the similarities between these two techniques will be considered at greater length.

The turnover table

The starting point, or central concept, of panel analysis is the turnover table, showing a categorical variable at time 1 cross-tabulated with itself at time 2. For a variable with n possible response categories, a turnover table is an n x n table summarizing the responses of each individual in the panel on a particular item at two successive interviews. In essence, such a table shows not only net change but also gross change.

Since the turnover table is basic, let us first examine the simplest case: a dichotomy (n = 2) observed at two time points. An example comes from a study by the Bureau of Labor Statistics, which in February 1963 carried out a nation-wide survey of men between the ages of 16 and 21 who were no longer enrolled in school. Two years later the bureau resurveyed them to analyze their early work experiences and problems. Table 1, which includes only those who were in the labor force at both times, shows the employment status of the youths at the two times.

Table 1 – Net change equals gross change
FEBRUARY 1963EmployedUnemployed1963 totals
Source: Perrella & Waldman 1966, table A.
1965 totals90.6%9.4%100.0%

From the totals, we observe that the employment rate for these youths rose from approximately 81 per cent to almost 91 per cent. Although there was a 10 per cent net change, there were actually 19 per cent whose employment status changed from working to being jobless, or the reverse. The

Table 2 –Net change equals gross change
FEBRUARY 1963EmployedUnemployed1963 totals
Source: Perrella & Waldman 1966, table A.
1965 totals1,926,000200,0002,126,000

information that the rate rose from 81 to 91 per cent is not enough to predict how many individuals experienced a change in employment status. Table 2 shows one possible situation—where net change and gross change are identical. Table 3 shows a contrasting possibility—where the gross change is almost three times as much as the net change. Thus, tables 1-3, with the same aggregate employment trend, represent very different economic situations.

Table 3— Gross change equals almost three times net change
FEBRUARY 1963EmployedUnemployed1963 totals
Source: Perrella & Waldman 1966, table A.
1965 totals1,926,000200,0002,126,000

I will return to this example in more depth when I come to discuss qualifier analysis. Meanwhile, it is useful to examine more complex multicategoried turnover tables.

The data in Table 4 come from a panel study of the 1948 presidential election campaign between President Truman and Governor Dewey (Berelson et al. 1954). The study was carried out in Elmira, New York, which in 1948 was a predominantly Republican community. Before examining the interior cells, we first examine the trend, as shown in Table 5.

Table 4 –Vote intention, August and October 1948
AUGUST VOTE INTENTIONRepublicanUndecidedDemocratAugust totals
Source: Adapted from Berelson et al. 1954, p. 23.
October totals397183180760
Table 5 –Trend in vote intention, August to October 1948 (per cent)
Source: Adapted from Berelson et al. 1954, p. 23.

Thus, as the campaign progressed, instead of more would-be voters reaching a decision, both parties appear to have lost adherents, with a corresponding increase in the undecided category. This information could of course have been obtained from two polls of separate samples. When we examine the interior cells of the turnover table, we see that there was more than a 6 per cent change. Dividing all of the entries in Table 4 by the grand total, we can obtain the distribution of changer types (see Table 6).

Table 6 –Distribution of vote intention (per cent)
AUGUST VOTE INTENTIONRepublicanUndecidedDemocratAugust totals
Source: Adapted from Berelson et al. 1954, p. 23.
October totals52.224.123.7100 = 760

The percentages on the diagonal running from upper left to lower right represent those whose vote intention remained constant, the total being 81.8 per cent. The gross change is thus about three times the net change. The individual changes, largely compensatory, can be classified into three types: (1) crystallizers (4.5 per cent), or those who were undecided but reached a decision in October; (2) waverers (10.7 per cent), or those who were decided but became doubtful in October; (3) converters (3 per cent), or those who switched

Table 7 –Changes in vote intention (per cent)
AUGUST VOTE INTENTIONRepublicanUndecidedDemocratTotals*Number of cases
* Details may not add to totals because of rounding.
Source: Adapted from Berelson et al. 1954, p. 23.

from Republican to Democrat or from Democrat to Republican.

Turnover tables are also frequently examined by dividing the entries in each row by the row total; these percentages, which provide estimates of “transition probabilities,” enable us to examine change, controlling for initial position (see Table 7).

To predict what decisions will be made by those who are undecided in August, we might advance one of the following four theories. (1) In a predominantly Republican community, those who are undecided might tend to lean toward the Democrats, since those leaning toward the Republicans would not hesitate to say so. (2) There are presumably equal forces pulling the undecided toward the Republican camp and toward the Democratic camp. Hence, those who do decide for whom to vote will split their votes 50-50 for Republicans and Democrats. (3) The undecided will vote in the same proportion as the rest of the community—2.4 Republicans to 1 Democrat. (4) Each undecided will be exposed, on the average, to 2.4 Republican stimuli to 1 Democratic stimulus and will vote for Republicans in most cases—or, at any rate, having no better criterion, most of the undecided will vote with the majority. Conjectures of this kind can best be tested by panel studies.

In Table 7, the transition probabilities on the main diagonal show that the undecided group is the most volatile: 16.2 per cent switched to Republican and 8.8 per cent switched to Democrat, distributing their votes to Republican and Democrat in a ratio of 1.8 to 1. This is less than the August ratio of Republicans to Democrats (440/184 = 2.4); it is also less than the October ratio (397/180 = 2.2).

Examination of Table 4 revealed that of 138 individual changes, all but 23 involved either switching from undecided to a party preference or from a party preference to undecided. Analysis of turnover using the Republican-Undecided-Democrat trichotomy, as in Table 7, tends to focus on the voters who are least involved politically. As was shown by an earlier panel, study of the 1940 presidential campaign (Lazarsfeld et al. 1944), the socalled independent voter who listened to both candidates and judiciously weighed the soundness of their respective programs before deciding on his vote was largely a mythical being. For the most part, those who oscillated from either Republican or Democrat to indecision, or who did not decide whom they would vote for until late in the campaign, were the politically uninterested who often

Table 8 – Change of vote intention from August to October (per cent)
  VOTE INTENTION IN OCTOBER 1948    August totals
VOTE INTENTION IN AUGUST 1948R+R?DD+Per cent*Number of cases
* Details may not add to totals because of rounding.
Source; Adapted from Berelson et al. 1954, p. 23.
Strong Republican75131021100241
Moderate Republican32471642100199
Moderate Democrat04165921100122
Strong Democrat1013157110062
October totals (number of cases)25414318310278 760

were persuaded to join one or another political camp for the flimsiest of reasons. Most of the changers were politically uninvolved and uninformed. Their shifts were more easily explained by social determinants than by political considerations. As a result of these findings, some political scientists charged the authors with supplanting traditional concepts of the political process with a type of sociological determinism. Accordingly, when the 1940 election study was replicated in 1948, the researchers sought a deeper analysis of the decision-making processes of politically involved voters. By combining the variable of party preference with that of strength of party preference, a five-point vote intention scale was constructed. The turnover table with this as the criterion shows much more of the dynamics of the campaign than the trichotomous turnover tables above (see Table 8).

Although a moderate Republican’s vote counts as much at the ballot box as a strong Republican’s vote, this analytic device of transforming the vote intention into a five-point scale allowed study of intraparty change and, hence, more study of the political issues. One further feature of the table is worth noting. The percentages on the main diagonal indicate for each political position the proportion who remained constant. Note that the moderate Republicans and the moderate Democrats are the least constant. They tend either to become more partisan or to become undecided. Thus, as the campaign progresses, there is an increase in the number of undecided, together with an increase in partisanship. In August, the ratio of strong to moderate Republicans is 241/199 = 1.2; by October the ratio rises to 1.8. Among the Democrats, the ratio is 0.51 in August and rises to 0.76 in October. It might be supposed that a minority which is outnumbered 2.4 to 1 would show intense partisanship, but this is clearly not the case here.


Turnover tables, by showing all of the gross change with respect to some criterion, serve the function of raising questions about the different types of changers. What are their characteristics, the processes that impel them to change, and the psychological and social effects of different patterns of change? In order to answer such questions, researchers have stratified turnover tables by other variables, which are sometimes called qualifiers. In panel studies, analysis by means of qualifiers constitutes the great bulk of the work performed.

Qualifiers can be classified according to whether they are changing or constant. Constant qualifiers can be further classified by precedence: those which occur prior to the first interview are called antecedent qualifiers, and those which occur between interviews 1 and 2 are called intervening qualifiers.

The constant qualifiers that precede interview 1 are the conventional demographic characteristics, such as age, education, nationality, sex, and social class. Of course, in some instances such characteristics might change; whether or not they are categorized as constant depends on the interval between interviews in the particular study. An individual can change his marital status or primary group attachments or socioeconomic status when the period between interviews is sufficiently long. The function of such qualifiers is to elaborate the original turnover table by showing the conditions under which there is more or less change, just as in survey analysis the researcher starts with the association between two variables and introduces other variables to find conditions under which the original relationship is heightened or diminished.

Tables 1-3 above showed changes in employment status over a two-year period. In order to learn more about the effects of different characteristics on employability, the same researchers examined the employment turnover by the educational level

Table 9 – Employment status of high school graduates (1963) in 1963 and 1965 (per cent)
 FEBRUARY 1963  1963 totals
FEBRUARY 1965EmployedUnemployedPercentNumber of Cases
Source: Adapted from Perrella & Waldman 1966, table A.
1965 totals1,065,00030,000 1,095,000
Table 10 –Employment status of high school dropouts (1963) in 1963 and 1965 (per cent)
 FEBRUARY 1965  1963 totals
FEBRUARY 1963EmployedUnemployedPercentof cases
Source: Adapted from Perrella … Waldman 1966, table A.
1965 totals861,000170,000 1,031,000

of the youths. The results are shown in tables 9 and 10.

The trend is the same for both groups. Dropouts increased from 74 per cent employed to 84 per cent, and high school graduates increased from 88 per cent employed to 97 per cent. The transition probabilities reveal more of the situation. Among graduates who were employed at the time of the first survey, only 1.6 per cent were jobless at the time of the resurvey, but of those who were unemployed at the time of the first survey, 11.4 per cent were still unemployed at the time of the resurvey. Among the dropouts, of those employed at the time of the first survey, 11.2 per cent were jobless at the time of the resurvey, whereas of those dropouts who were unemployed at the time of the first survey, fully 31.7 per cent were jobless at the time of the resurvey. These findings strongly suggest that there exists a hard-core group of jobless youth. It should also be noted that 26 per cent of the dropouts underwent some change in employment, compared with 12 per cent of the graduates.

To discover why some dropouts were employed at both times while some graduates were unemployed at both times, we might qualify the employment turnover table by race. There are income statistics which show that there is a considerable gap between the earnings of whites who did not graduate from high school and nonwhite high school graduates. In the course of a lifetime the average nonwhite high school graduate can expect to earn approximately $50,000 less than the white dropout. It would not be surprising to find that race accounts for some of the anomalous cases.

We might study other qualifiers. For example, among the graduates, did those who pursued a strictly academic program do better or worse than those who pursued a business education or vocational program? Were married youths more likely to hold their jobs? They might be more diligent or more motivated to work to pay for furniture or to accumulate savings for a prospective family. And there might be a tendency among employers to lay off unmarried workers first when work slackens.

Undoubtedly, employment and marital status are related. In fact, more dropouts than graduates were unmarried at the time of the first survey. If we were interested in the social effects of employment, we could use marital status as the criterion and examine turnover of marital status as qualified by employment. Were those who were employed and unmarried in 1963 more likely to become married by 1965 than those who were unemployed? And were those unemployed and married in 1963 more likely than the employed to become separated or divorced? We could go further and analyze the joint change of employment and marital status, which, with dichotomous variables, would involve analysis of a sixteenfold table.

Impact panels

Frequently the panel technique is used for evaluating the effects of information campaigns. A public health agency, for example, might be interested in learning the effects of an information campaign on the recognition and reporting of cancer symptoms; the federal government might want to determine the effectiveness of a consumer education program among the poor; a state commission might want to determine the effectiveness of a campaign aimed at reducing discrimination against minority groups. These questions and others like them can best be answered by a type of panel study called an impact panel, in which the qualifiers intervene between the first and second series of interviews and refer to exposure to some stimulus or succession of stimuli. The impact study thus represents one equivalent of the controlled experiment when observations must be made in a natural social setting and the individuals being studied cannot be randomly assigned to experimental or control groups. [SeeExperimental Design, article onquasi-experimental design;see alsoEvaluationresearch.]

One general observation should be made concerning the place of theory in experiments as compared with panel studies. In experiments virtually all of the theoretical thinking must be done in advance of the field work. In panel studies, however, theorizing takes place at both ends of the research process. A small set of interlocking variables must be decided upon at the outset; these variables determine the boundaries of the analysis. Even the most empirically oriented researcher must have some implicit theory of the relative fruitfulness of different variables. After the data have been coded and put on punch cards or tape, theory serves an organizing function, since without some notion about which tabulations to run, it would be possible to keep a computer busy for months before exhausting the astronomical number of possible cross tabulations.

Some problems connected with impact studies may be worth pointing out. The experiment seeks to measure the effects of a stimulus when the individual is exposed to it. The panel, on the other hand, seeks to explain what happens under non-laboratory conditions; it aims to assess effects of exposure under conditions where audience self-selection is operative. Under experimental conditions a captive audience is exposed to some prepared stimulus, such as a lecture, and often shows significant before-after differences in information or attitude. Usually, the shift is among the least educated and least interested individuals. When the same program is transferred from the laboratory to a mass medium, a study of effects frequently reveals little or no change: those who are least educated, least interested, and least in agreement with the communication rarely expose themselves to the stimulus in the first place. The panel, unlike the experiment, deals with a noncaptive audience; hence, not only the effects of exposure but also the nature of the audience self-selection are studied [seeCommunication, mass, article oneffects; Persuasion].

Determination of exposure to mass media or to personal influence is difficult, for it generally involves the use of retrospective questions. This, of course, violates the central conception of panel analysis, because we cannot be sure to what extent selective recall is operative. For example, in the study of political behavior during a national election campaign, it would be important to assess the role of local political parties. What, for instance, is the effect of receiving campaign literature? What influence does a personal visit by a party worker have on the turnout and vote intentions of potential voters? But if we ask respondents whether they received literature or were contacted by a party worker, how can we be sure that we are not dealing with memory biases—that is, the voters were contacted randomly by workers but those interested in politics tended to remember the contact? If responses to the queries on being visited or on receiving political literature are analyzed, we do find that the more interested respondents report greater personal and impersonal contact.

To some extent this result may reflect selective contact. Party workers may have approached wouldbe voters who seemed more receptive to political discussion, or they may have selected names of contacts from previous voting lists or made their contacts in various other nonrandom ways.

One safeguard that can be built into panel studies to check against the bias of selective recall is stratification. When effects of party contact are analyzed, for instance, it is necessary to control by level of political interest. Another safeguard is to try to measure exposure independently by finding out from party workers whom they contacted.

Inherent in panel study design is the possibility of examining the differential processes through which the stimulus is related to its effects. Experiments ordinarily do not permit such detailed study, for the simple reason that the number of cases is usually not sufficient, although there are occasional exceptions. In panel studies, more refined analyses of the relations between the stimulus and its effects are carried out by determining the characteristics of those respondents who are most influenced by the stimulus—whether they are men or women, old or young, educated or uneducated, and so on. The stimulus may have had a marked effect in the desired direction on certain groups of respondents and no effect or a “boomerang” effect on other groups.

Particularly important in this respect is the consideration of effects in terms of the respondents’initial position on the criterion. In the impact panel study, unlike the experimental study, the respondents who have been exposed to the stimulus are quite likely to differ from unexposed respondents on the criterion prior to the exposure. If they do, it is necessary, in order to impute any effect to the stimulus, to control the criterion prior to the campaign, for there is every reason to expect that the effects of a stimulus will not be the same for respondents who are initially favorable, indifferent, or unfavorable on the criterion. By controlling for initial position on the criterion, we do not altogether approximate experimental matching, although we do reduce its urgency. By holding initial position constant, the subgroups to be compared are made more homogeneous on everything except exposure to the stimulus.

For other material bearing on impact panels and controlled experiments, see the discussions by Hovland (1959) and Cohen (1964, chapter 9).

Mutual effects

The most interesting part of panel analysis is the analysis of mutually interacting variables. Some examples of two interacting variables are the following. (1) In the investigation of psychosomatic disorders we find that worry and anxiety produce hypertension or ulcers, and these disabilities in turn raise the level of anxiety. (2) Studies have been carried out relating chronic illness with poverty. Chronic illness drains a family’s resources, while low socioeconomic status reduces access to appropriate medical care, proper diet, and a congenial working environment. (3) In studies of family life, we often find that problem children and lack of parental love tend to be associated. But which causes which? Does lack of parental love induce aberrant behavior, or does aberrant behavior corrode parental affection? It seems likely that each reinforces the other. (4) The social psychologist may find that friendship and similarity in values feed into each other. People tend to select friends among those with similar values, and values tend to become convergent in the course of friendship. (5) The sociologist may observe that individuals who reveal greater conformity to company goals tend to have higher rates of promotion. Correspondingly, promotion tends to reinforce conformity with company goals. In each example, one cannot say which variable is the dependent variable and which the independent; each influences the other. The problem in research is to analyze their relative influence. Let us look at an example of the analysis of relative influence.

Suppose there is to be a referendum in some community on the issue of fluoridating the water, and we are studying how people influence each other on public health matters. We twice ask a panel of married couples: “As things stand now, how do you intend to vote on this issue?” The resultant sixteenfold table (Table 11) shows the hypothetical joint distribution of vote intention at two interviews of 671 hypothetical couples. Assuming these figures were empirical, what could we infer from them about the relative influence of the spouses?

Consider, on the one hand, those couples who at time 1 were in agreement but at time 2 were in disagreement. Altogether there were b + c + n + o =1+ 4 + 8 + 2=1 5 such changes. In 12 of them, it was the husband who “defected,” so to speak, while in only 3 did the wife “defect.” Thus, as far as preserving agreement between the spouses, the husband was about four times more influential than the wife.

Consider, on the other hand, the couples who were initially in disagreement but as a consequence of one spouse changing opinion were subsequently in agreement. Altogether there were e + i + h + I = 25 + 5 + 10 + 25 = 65 such cases. To what extent did each spouse influence the other to change in accord with his own opinion? Of the 65 cases, 50 agreements were generated as a result of the wife changing her opinion to accord with her husband’s; only 15 agreements resulted from the husband changing to harmonize with his wife’s opinion. Thus, as far as generating agreement, the husband was more than three times as influential as the wife. From the four agreement-to-disagreement cells of the sixteenfold table and the four disagreement-to-agreement cells, an index of relative influence might be devised. During the last twenty years, Lazarsfeld has proposed several indices of relative influence, all based on the single-change cells of the sixteenfold table. If the eight singlechange cells are labeled as in Table 11, one of Lazarsfeld’s indices is the following:

The terms in the first pair of parentheses represent H’s influence on W; the terms in the second pair of parentheses represent W’s influence on H.

Table 11 - Opinion of spouses at successive interviews on fluoridation issue*
  WIFEFavorOpposeFavorOpposeTime 1 total
* Hypothetical data.
 Favor ef9h 
TIME 1  ijkl 
 Oppose mnOV 
Time 2 totals 3335152235671

In the hypothetical husband-wife sixteenfold table (Table 11),

(in favor of husbands).

A second method of measuring relative influence is based on the relative magnitude of the cross-lagged correlations: rH1W2 versus rW1H1 (Campbell 1963; Pelz & Andrews 1964). If Yule’s Q is used as a measure of correlation, the cross-lagged correlations in the husband-wife example are QHlW2 =.93 >QW1H2 = .83. Since QH1W2> QW1H2, the inference would be that the husband is more influential. Pelz and Andrews (1964) also consider as a measure the relative magnitude of the partial crosslagged correlations:rH1W2.W1. versusrWlH2.Hl. Generally, the Lazarsfeld index, the simple cross-lagged correlations, and the partial cross-lagged correlations will agree as far as indicating which of two variables is more influential. [For a discussion of Yule’s Q, seeStatistics, Descriptive, article onassociation.]

In conclusion, it should be stated that much of the potential value of panel analysis for sociological inquiry remains untapped. The analysis of mutual effects has been directed almost entirely at two-variable interactions. The burgeoning literature on causal analysis in recent years attests to the readiness of the field to move beyond this. Moreover, panel methodology has developed around change of individuals, principally opinion change. Problems in which the unit of analysis is not the isolated individual, but individuals interlocked sociometrically or in a formal organization, have rarely been the subject of panel research.

Bernard Levenson

[See alsoCounted Data; Diffusion, article oninterpersonalinfluence; errors, article onnonsampling errors; Reason analysis; Statistics, descriptive, article onASSOCIATION; survey ANALYSIS; and the biography ofRice.]


Anderson, T. W. 1954 Probability Models for Analyzing Time Changes in Attitudes. Pages 17-66 in Paul F. Lazarsfeld (editor), Mathematical Thinking in the Social Sciences. Glencoe, 111.: Free Press.

Berelson, Bernard; Lazarsfeld, Paul F.; and Mcphee, William N. 1954 Voting: A Study of Opinion Formation in a Presidential Campaign. Univ. of Chicago Press. → Tables 2, 3, 4, and 5 are based on data from this book, by permission of the authors and the University of Chicago Press. Copyright 1954 by the University of Chicago.

Blumen, Isadore; Kogan, Marvin; and Mccarthy, Philip 1955 The Industrial Mobility of Labor as a Probability Process. Cornell Studies in Industrial and Labor Relations, Vol. 6. Ithaca, N.Y.: Cornell Univ. Press.

Campbell, Donald T. 1963 From Description to Experimentation: Interpreting Trends as Quasi-experiments. Pages 212-242 in Chester W. Harris (editor),Problems in Measuring Change. Madison: Univ. of Wisconsin Press.

Cohen, Arthur R. 1964 Attitude Change and Social Influence. New York and London: Basic Books.

Coleman, James S. 1964a Introduction to Mathematical Sociology. New York: Free Press. → See especially pages 132-188, “Relations Between Attributes: Over-time Data.”

Coleman, James S. 1964b Models of Change and Response Uncertainty. Englewood Cliffs, N.J.: Prentice-Hall.

Glock, Charles Y. (1951) 1955 Some Applications of the Panel Method to the Study of Change. Pages 242-250 in Paul F. Lazarsfeld and Morris Rosenberg (editors), The Language of Social Research: A Reader in the Methodology of Social Research. Glencoe, III.: Free Press.

Glock, Charles Y. 1952 Participation Bias and Reinterview Effect in Panel Studies. Ph.D. dissertation, Columbia Univ.

Goodman, Leo A. 1962 Statistical Methods for Analyzing Processes of Change. American Journal of Sociology 68:57-78.

Goodman, Leo A. 1965 On the Statistical Analysis of Mobility Tables. American Journal of Sociology 70: 564-585.

Hovland, Carl I. 1959 Reconciling Conflicting Results Derived From Experimental and Survey Studies of Attitude Change. American Psychologist 14:8-17.

Katona, George 1958 Attitude Change: Instability Response and Acquisition of Experience. Psychological Monographs 72, no. 10.

Kendall, Patricia L. 1954 Conflict and Mood: Factors Affecting Stability of Response. Glencoe, 111.: Free Press.

Lazarsfeld, Paul F.; Berelson, Bernard; and Gaudet, Hazel (1944) 1960 The People’s Choice: How the Voter Makes Up His Mind in a Presidential Campaign.2d ed. New York: Columbia Univ. Press.

Lazarsfeld, Paul F.; and Fiske, Marjorie 1938 The “Panel” as a New Tool for Measuring Opinion. Public Opinion Quarterly 2:596-612.

Lipset, Seymour M. et al. (1954)1959 The Psychology of Voting: An Analysis of Political Behavior. Volume 2, pages 1124-1175 in Gardner Lindzey (editor),Handbook of Social Psychology. Cambridge, Mass.: Addison-Wesley.

Mcdill, Edward L.; and Coleman, James S. 1963 High School Social Status, College Plans, and Academic Achievement: A Panel Analysis. American Sociological Review 28:905-918.

Newcomb, Theodore M. (1943) 1957 Personality and Social Change: Attitude Formation in a Student Community. New York: Dryden.

Pelz, Donald C.; and Andrews, F. M. 1964 Detecting Causal Priorities in Panel Study Data. American Sociological Review 29:836-848.

Perrella, Vera C.; and Waldman, Elizabeth 1966 Out-of-school Youth: Two Years Later. U.S. Bureau of Labor Statistics, Division 7, Labor Force Studies, Special Labor Force Report No. 71. Washington: Government Printing Office.

Rice, Stuart A. 1928 Quantitative Methods in Politics. New York: Knopf.

Underhill, Ralph 1966 Values and Post-college Career Change. American Journal of Sociology 72:163-172.

Wiggins, Lee M. 1955 Mathematical Models for the Interpretation of Attitude and Behavior Change: The Analysis of Multi-wave Panels. Ph.D. dissertation, Columbia Univ.

Zeisel, Hans (1947) 1957 Say It With Figures. 4th ed.,rev. New York: Harper. → See especially pages 215-254, ’The Panel.”

Panel Studies

views updated May 14 2018


A panel study is defined as a study that collects information on the same individuals at different points in time. The various data collections are often called waves. A panel study is therefore a longitudinal study; it differs from other studies that collect information over time, such as time series and cohort studies, in that it studies the same persons longitudinally.


Panel studies are the optimal design for addressing some of the core questions in aging research: first by measuring in each wave the same characteristics in the same persons, panel studies are able to provide descriptions of changes experienced by individual persons over time and differences between people in their individual change patterns. For example, measures of cognitive functioning and diagnosed diseases are collected at each wave in the Health and Retirement Study, a panel study of physical and mental health and their labor force and economic consequences. Change can then be defined either as transition from one state to another on a categorical variable (e.g., the transition from "no diabetes" to "diabetes") or as difference in level on a continuous variable (e.g., the extent of decline or improvement in cognitive functioning). The absence of change, or stability, may also be of interest.

A single panel study of persons of a limited age range can describe change or stability associated with aging for only one cohort. It is generally recognized that aging is conditioned by the historical, political, economic, and societal contexts in which it takes place. For example, the onset of diabetes depends on lifestyle factors such as obesity or lack of exercise, and lifestyle factors have been changing. Therefore panel studies of different cohorts who age through different historical times are required for generalizeable descriptions of change. But even a multiple-cohort design cannot provide a definitive distinction between aging and historical phenomena; direct investigations of specific causal factors such as lifestyle on diabetes onset represent a more fruitful approach.

Second, panel studies with several waves are the best quasi-experimental design for investigating the causes and consequences of change with high internal validity. While a true experimental design generally is considered the strongest design for investigating causal patterns, many potential causes of interest to aging researchers are not amenable to experimental manipulation. In the above example lifestyle as a potential cause of the onset of diabetes is not easily manipulated, nor is it feasible to manipulate the onset of diabetes in order to study its hypothesized effects on quality of life outcomes. Quasi-experimental designs that depend on the naturally occurring variation of potential causes are the next best alternative. Because naturally occurring variation depends on other causal factors (e.g., the fact that those with healthy lifestyles differ in many other ways from those with less healthy lifestyles), the data collection must be planned to ensure that those other factors are measured and careful statistical control of those factors must be applied during the analysis of the data from quasi-experimental designs. Panel studies have an important advantage over those that use the simpler quasi-experimental design of a single, cross-sectional data collection in that they allow for better specification of the time-ordering between presumed cause and effect.


Panel studies pose fairly formidable methodological challenges. The lag between waves needs to be consistent with the change patterns of interest. For example, whereas the impact of a medication's side effect on cognitive functioning may last only weeks, the effect of obesity and lack of exercise on onset of diabetes may not take place for ten or more years. Also, measurement must be consistent over waves in order not to confound change in the underlying characteristic with measurement change, and it needs to be inclusive in order to capture all possible causes and outcomes and controls.

Most of the big panel studies utilize population probability samples that permit generalization to the target population and provide for external validity. Such generalizations can be biased by missing data, including nonresponse to initial recruitment, missing information to specific measures, andparticularly critical for panel studiesdropping out of later waves, because those with missing data may be systematically different from those without. Potential biases can be reduced by minimizing nonresponse and attrition during conduct of the panel study, by imputing missing responses after data are collected, and by modeling missing information during data analysis.

All of these design considerations imply that a fair amount of prior knowledge about the phenomenon of interest and its explanations is required to optimally design a panel study. Moreover, relevant knowledge continues to accumulate over the often lengthy life of a panel study and can make an ongoing study obsolete. Finally, panel studies tend to be very expensive to conduct and therefore tend to be designed with several objectives in mind; the multipurpose nature can interfere with a design tightly focused on a specific hypothesis.

Data analysis

Many options for analysis of panel data are available. For outcomes in the form of simple transitions between two waves, logistic regression analysis may be used. For example, the transition from "no diabetes" at wave t to "diabetes" at wave t+1 may be so analyzed. The "competing risks" of dying and dropping out of the study may be modeled by using multinomial regression analysis that allows for multiple outcomes. In the case of transitions for which exact timing is available such as death or retirement, survival or hazard analysis techniques allow a refined assessment of the change patterns and the causal dynamics. For outcomes in the form of continuous variables such as a performance measure of cognitive functioning, ordinary least squares regression analysis can model the amount of change between two waves. Autoregressive structural equation modeling and growth curve analysis permit the specification of multiwave change and stability; in the former change is conceived as relative change, in the latter as absolute change. For both types of models statistical procedures exist that allow for the evaluation and control of measurement error. In all of these analytical techniques potential causes conceptualized either as status or changemay be evaluated as statistical predictors. In multi-wave panel studies different lags between cause and consequence can also be evaluated.

Examples of panel studies for the study of aging

Major panel studies for studying age-related changes, their causes and consequences include the Baltimore Longitudinal Study of Aging, the Panel Study of Income Dynamics, the Health and Retirement Study, the Americans' Changing Lives Survey, The Berlin Longitudinal Study, the National Long Term Care Survey, The Established Populations for Epidemiologic Studies of the Elderly, and the Longitudinal Study of Aging.

A. Regula Herzog

See also Developmental Psychology; Surveys.


Alwin, D. F., and Campbell, R. T. "Quantitative Approaches: Longitudinal Methods in the Study of Human Development and Aging." In Handbook of Aging and the Social Sciences, 5th ed. Edited by Robert H. Binstock and Linda C. George. New York; Academic Press. Forthcoming.

Kasprzyk, D.; Duncan, G. J.; Kalton, G.; and Singh, M. P., eds. Panel Surveys. New York: John Wiley & Sons, 1989.

Menard, S. Longitudinal Research. Newbury Park, Calif.: Sage, 1991.

Panel Studies

views updated May 14 2018

Panel Studies


A panel study is a type of nonexperimental longitudinal study design that samples (using random or stratified random methods) a population to identify a set of respondents, then recontacts the same respondents repeatedly over time. A traditional panel study differs from a trend study in that a trend study samples the same population repeatedly but with different respondents at each time point. A panel study is similar to a traditional epidemio-logic cohort study in that a sample or cohort is followed over time. However, the purpose of following participants in a traditional cohort study is to allow enough time to be able to observe a particular outcome and determine how exposure to a certain variable is associated with the outcome. In contrast, the purpose of following participants in a panel study is typically to observe changes on a particular variable over time to potentially relate those changes to an outcome of interest or explain what factors may lead to the changes in the variable over time. Thus, a panel study often includes multiple measurements of variables, usually at regular intervals, over a circumscribed period of time.

Compared to other nonexperimental study designs such as case-control or cross-sectional studies, a carefully conducted panel study can strengthen the case for inferring causation because temporality (i.e., the cause precedes the effect in time) is more accurately preserved. In panel studies, a researcher is also able to determine the strength of an association between an explanatory variable and an outcome. However, specificity or the degree to which a cause leads to a single effect is not as high as it is in a randomized controlled trial or experimental study. Thus, in panel studies statistical controls for group non-equivalence are often employed. A further advantage of panel studies is that they allow for complex statistical modeling of dynamic phenomena (e.g., latent growth modeling and other types of trajectory analyses) that may more accurately reflect reality.

One primary disadvantage of panel studies is the problem of selective attrition over time. Participants who remain in the panel over time may be qualitatively different from the participants who are lost to follow-up due to relocation, lack of interest, or death. Another disadvantage is that panel studies can be quite costly as they require measurements at multiple time points, and tracking highly mobile respondents can be a challenge. Panel studies are strengthened when the measurement protocol or instrument remains the same over time. However, this can be difficult to maintain because of emerging technological advances in instruments or measurements or the difficulty in measuring the same phenomena during developmentally different phases of life. A change in the measurement protocol or instrumentation makes it difficult to determine if an observed change in the data is the result of measurement protocol changes or of actual changes. Finally, panel surveys can also be fatiguing to participants, leading to selective attrition, or participants may become primed to the instrument or measurement protocol, leading to atypical responses. Large scale, well-conducted panel studies include the Panel Study of Income Dynamics, National Longitudinal Surveys, the Health and Retirement Survey, the National Longitudinal Study of Adolescent Health, Coronary Artery Risk Development in Young Adults, and the Framingham Heart Study.

SEE ALSO Data, Longitudinal; Data, Pseudopanel; National Education Longitudinal Study; National Longitudinal Study of Adolescent Health; National Longitudinal Survey of Youth; Panel Study of Income Dynamics; Sample Attrition; Survey; Surveys, Sample


Baltagi, Badi H. 2005. Econometric Analysis of Panel Data. 3rd ed. West Sussex, U.K.: Wiley.

Institute of Social Research. University of Michigan. Panel Study of Income Dynamics.

National Heart, Lung, and Blood Institute of the National Institutes of Health. Coronary Artery Risk Development in Young Adults (CARDIA).

National Heart, Lung, and Blood Institute of the National Institutes of Health. Framingham Heart Study.

National Institute on Aging. The Health and Retirement Study. A Longitudinal Study of Health, Retirement, and Aging.

University of North Carolina. The National Longitudinal Study of Adolescent Health: AddHealth.

U.S. Department of Labor. Bureau of Labor Statistics. National Longitudinal Surveys.

Bernard F. Fuemmeler