Causal Models in Experimental Designs

H. M. Blalock

• Publisher: Routledge
• Year: 2017
• Print ISBN: 9780202309729
• eBook ISBN: 9781351529808
• Topic: Social Sciences
• Subtopic: Sociology
• Index: Social Sciences

This is a companion volume to Causal Models in the Social Sciences, the majority of articles concern panel designs involving repeated measurements while a smaller cluster involve discussions of how experimental designs may be improved by more explicit attention to causal models. All of the papers are concerned with complications that may occur in actual research designs- as compared with idealized ones that often become the basis of textbook discussions of design issues.

II

The Use of Causal Models in Panel Designs

Change data afford special opportunities to disentangle causes and effects and to distinguish the relative magnitude of the effect of X upon Y from that of Y upon X. They also permit one to assess the stabilities of different variables over time and thereby to gain greater insights as to the exogenous factors producing social change. But all of this depends upon several crucial assumptions, especially those relating to measurement errors and the correct specification of lag periods in relation to intervals of observation. Frequently, for example, the observed “changes” in a given period are relatively small as compared with presumed measurement-error variances, so that it becomes difficult to separate out real change from apparent change due to measurement error. Furthermore, unless one has a good theory regarding lag intervals, the superimposition of measurement errors on improperly specified lag models may create an extremely ambiguous situation, or even one of underidentification.

As the chapters in Part II indicate, there are a host of potential complications, the implications of which are just beginning to be understood even in the case of models involving only one or two theoretical variables. For example, as Heise notes in Chapter 6, if one allows for random measurement errors and only a single measure of a given variable, then observations at three time points will be necessary to achieve identification unless some very strong assumptions can be made. Chapter 7 by Wiley and Wiley and Chapter 8 by Werts, Jöreskog, and Linn both elaborate on this relatively simple type of model but point to further complications in its implementation. As soon as one moves to multiple indicators that may be utilized at two or more points in time, one may begin to allow for sources of nonrandom measurement errors as well, provided that certain kinds of simplifications are allowed. Chapter 10 by Blalock suggests some of the tradeoffs that exist in this connection in instances where one has either three measures at two points in time or two measures at three points in time.

Unless one is willing to assume either no measurement errors or strictly random ones, analyses of panel data must confront the question of how one allows for nonrandom sources of measurement errors in instances of repeated measurements. Clearly, this requires a theory of measurement errors that is sufficiently realistic to allow for very different sources of error. Strictly constant errors, as for example the addition of 10 units to every score, can be cancelled out rather simply by taking change or difference scores. Indeed, if such constant errors were the only major sources of error, this would suggest the clear-cut superiority of panel designs over those that rely on comparative data for which one were unwilling to assume such constant errors across all observations.

As measurements are repeated, however, several things begin to happen with human subjects. Respondents may become increasingly accurate as they gain experience, in which case measurement-error variances may become smaller over time. But perhaps the measurement-error variances remain constant, whereas the variances in the true scores are diminishing (or increasing) over time. Since it is the ratio of the two kinds of variances that matters, if one utilizes standardized measures—such as correlations or path coefficients — one may be misled if these ratios were to shift. Since both the true scores and the measurement-error components will be unobserved, however, these shifts in the unknown variances may remain undetected and thereby lead one astray. This point is specifically noted by Wiley and Wiley in Chapter 7.

If there are artifacts or biases produced by specific measurement techniques, it may be convenient to assume that such biases remain constant over time or that their distorting effects are constant across a certain set of variables at each time point, though they may shift at a later point in time, perhaps because of the increased sophistication of the respondents. If these measurement-instrument effects shift in unknown ways across variables and across time, however, the measurement-error model is likely to become underidentified, which means that empirical estimation will be impossible. Thus, there will inevitably be tradeoffs between, on the one hand, the need for simplifications and, on the other, the objective of building in necessary complications to allow for different types of sources of bias. The larger the number of measures of each variable and the more time points, the more such complications one can handle while still achieving an estimable model. As can be imagined, however, there are many reasons why this very simple generalization must be qualified in particular instances, as will become clear from the discussions in the remainder of the volume.

There must also be a concern about the correct specification of lag periods, particularly if one wishes to use panel designs to sort out the direction of causation between two or more variables. If one allows for both simultaneous and lagged effects and also for reciprocal causation between X and Y, one’s model will become underidentified unless other kinds of restrictions are placed on the model. In Chapter 12, Greenberg and Kessler discuss the feasibility of making certain assumptions about the constancy of parameter values across adjacent time periods, if there are at ...

Table of contents

• Cover Page
• Title Page
• Copyright Page
• Contents
• Preface
• I The Use of Causal Models in Experimental and Nonexperiment al Designs
• II The Use of Causal Models in Panel Designs