Method Variance in Structural Equation Modeling: Living with "Glop"
G. R. Patterson
Oregon Social Learning Center
In the past 10 years there has been a marked increase in the use of structural equation modeling (SEM) in the study of family processes and their outcomes. As articulated by Martin ( 1987), SEM is not a panacea for the ambitious investigator, and of course, does not alleviate interpretive problems resulting from poor quality data, sampling biases, or internal validity problems resulting from faulty study design. On the other hand, SEM has a number of clear and far-reaching advantages. It does not require that error terms be uncorrelated among the independent and dependent variables. In fact, with SEM the question can be addressed statistically. In addition, both the measurement and causal models can be estimated simultaneously, thereby increasing our understanding of both aspects of model building and how they might interact. Thus, a more thorough understanding of the data as they relate to the confirmation or disconfirmation of theoretical hypotheses is possible. Furthermore, relatively complex theoretical structures can be posited and tested, while in a very real sense maintaining experiment-wise alpha levels at known and acceptable rates.
The analytic versatility of SEM also provides an opportunity to examine issues that have long haunted psychological researchers, such as the issue of how measurement method determines the outcomes to a research enterprise. For example, if one defines both the independent and dependent variable with a common measure (e.g., observer impressions or self-report) in a structural model, the estimated effect coefficient is much higher than when the variables are defined by nonoverlapping indicators. For example, reviews by Rutter ( 1979) and Emery ( 1982) showed consistent correlations between reports of marital discord and child adjustment problems. However as pointed out in a later study by Emery and O'Leary ( 1984), in most of these studies the same person (usually