Sociological Methods & Research

In this article, a description of the test-retest study as part of the larger 1984 German General Social Survey (ALLBUS) is presented. First, the comparability of the test-retest mirrors the larger ALLBUS study with few exceptions, and an examination of the stability of demographic variables is presented. Although the stability of some of the items was poor (e.g., report of one's own first occupation and father's occupation), most of the other items of central importance to survey recorders (e.g., education or religious identification) had at best minimally acceptable stabilities. The data also show a consistent tendency for the t ₂-t ₃ stabilities to be higher than the t ₁-t ₂ stabilities, a finding that is discussed and for which a possible rationale is presented.

This article focuses on three main points: (1) Because a short-wave panel study such as the ALLBUS-Retest Study 1984 is suited optimally for reliability estimations, the quality of a larger number of items concerning attitudes about the welfare state and inequality and toward a minority group in West Germany (“guestworkers”) is investigated; (2) we show that most of the items are more reliable in the subgroup with high political interest than in the low-interest group; and (3) because we cannot rely on multiple indicators in most of the cases, methods and statistical models for the analysis of single indicators in multiple-wave panels are discussed. Because we can assume that the latent attitudes are continuous and the observed variables have ordered categories (at least), we confine our analyses to metric variable models. We show that (a) the assumptions underlying the classical models proposed by Heise (1969) and by Wiley and Wiley (1970, 1974) will often be violated if the times between the panel waves are very short; (b) frequently, models of parallel or tau-equivalent measurement, or some modifications of these models lead to parameter estimates that are more plausible from a theoretical point of view; (c) the ML- and ULS-estimators for these models can be easily calculated; (d) the use of polychoric correlation coefficients for these models is more appropriate if the observed variables have a badly skewed distribution and a small number of categories.

It has often been observed in experimental studies that the reliability of items increases if the same questions are asked of the same respondents more than once. This phenomenon, called the “Socratic effect,” also occurs in nonexperimental, short-wave panel studies. In the first section of this article a number of hypotheses presumed to underlie the “Socratic effect” are presented. It is argued that a distinction must be made between consistency processes at the structural level (latent attitudes) and the observational level (respondent behavior). Given this distinction, the hypotheses are tested within a LISREL framework that takes this differentiation into account. The hypotheses are then evaluated using four items to measure respondents' attitudes toward guestworkers in West Germany. By and large the central hypotheses are confirmed. It is also shown that two different models can be fit to the observed data equally well, and therefore a nonstatistical criterion has to be invoked to decide which model to use as the basic model for describing the stability and reliability of the attitude.

This article deals with the assessment of the reliability and stability of the respondent's attitudes toward guestworkers across selected standard demographic variables (i.e., age, education, gender) as well as across contacts with guestworkers. As a baseline model we use the “best” model developed by Jagodzinski et al. (this volume) and test its invariance across the subgroups using the multiple-group option of LISREL.

In this article, it is shown how one can estimate latent means in panel analysis. It is shown how simultaneous estimation of latent location parameters in different groups can be obtained and how latent means can be recalculated using the location parameters along with the stability coefficients. Latent means for attitudes toward guestworkers in Germany are computed for gender, age, and education and are compared with the means from composites based on the observed item means.

When using panel data important problems are often conveniently overlooked. These include model misspecification, asymptotic stability, unequally spaced panel waves, and the use of ordinal rather than metric data. While panel data may be useful to eliminate specification error, if the process generating data is in equilibrium, the problems of misspecification persist if a model with a lagged endogenous variable is formulated. Furthermore, the notion of asymptotic stability and its relation to the notion of inertia, which is often called stability by social scientists, is discussed. An important practical question is how to deal with data from panels with unequally spaced intervals between waves. Finally, the issue of the treatment of ordinal dependent variables is addressed. Using ordinal probit thresholds, the question of the variance of an ordinal variable across time is clarified and a quick estimator for this variance is given.