Discussant Comments on &Quot;Issues in Structural Models&Quot;

Tom Page, Michigan State University
[ to cite ]:
Tom Page (1993) ,"Discussant Comments on &Quot;Issues in Structural Models&Quot;", in NA - Advances in Consumer Research Volume 20, eds. Leigh McAlister and Michael L. Rothschild, Provo, UT : Association for Consumer Research, Pages: 654.

Advances in Consumer Research Volume 20, 1993      Page 654


Tom Page, Michigan State University

The comments presented here serve to highlight the contribution of the papers presented in this session, as well as to point out some concerns that readers should be aware of. The only common thread among the three papers is the use of LISREL. The paper by Sauer and Dick is a very good tutorial on the use of moderator variables in structural equations. The paper by Mittal is a cautionary note on what LISREL can and cannot tell us, and the paper by McCarty and Shrum is an application of structural equations to the issue of recycling. The papers will be discussed in the order in which they were presented.


The objective of this paper is to provide the reader with a tutorial on moderator variables in structural equations. The authors first provide a useful discussion of the distinction between mediators and moderators, and even discuss mediated moderators and moderated mediators. This is a distinction that is often confused, and the authors' explanation of exactly what constitutes mediation and moderation is both clear and succinct.

The main contribution of the paper is the discussion of moderator variables in structural equations. Two types of moderators are discussed, discrete and continuous. For discrete moderators, such as nominal variables like gender, the authors advocate using multiple group analysis with the categories serving to identify the groups. Then, in one analysis, the relevant paths are constrained to be equal, and in the other analysis they are unconstrained. Then, the chi-square difference and the path coefficients are examined to determine the nature of the mediating relationship.

This is one area of the paper that should be expanded. In other words, exactly how does one tell if s/he has a moderator relationship or not? If the chi-square difference is significant, it means that the path coefficients are not equal across the groups. However, how big should the difference be before a moderating relationship is supported?

In the continuous case, the authors examine the case of both observed and latent moderator variables. For the observed case, they advocate forming an interaction term and refer to several sources for a description of the procedure. The test for moderation, however, is not clear in this case. If two analyses are performed as described, one of which contains the interaction term and one which does not, the chi-square difference test, suggested by the authors, is not appropriate since the models are not nested. That is, the two analyses are not based on the same set of variables. This point needs to be clarified. In the continuous latent moderator case, the authors do a good job of explaining what to look for in examining path coefficients.

The discrete case is one that every structural equation user should be especially aware of since it is also applicable to experimental design. Too often researchers attempt to analyze experimental data by modeling the manipulation as an exogenous variable and combining all of the groups into one analysis. Not only does this violate some of the assumptions of LISREL, but it also precludes a clear interpretation of the path coefficients. The preferred method of analyzing such data should be the technique described by Sauer and Dick for discrete moderators.


This paper explores the links among personal values, attitudes, and behavior concerning recycling. The objective is to determine whether personal values about recycling directly affect behavior or their effect is mediated by attitudes about the behavior. The authors make a good case for examining the effects of values in predicting behavior. Their results basically show that values are related to attitudes, but not directly related to behavior.

There are several suggestions that perhaps should be considered in future research in the area. First, it might be useful to include the construct of intention to perform recycling behavior as a mediator between attitudes and behavior. Most research has shown that the link between attitudes and intentions, and intentions and behavior is stronger than the direct link between attitudes and behavior. This may account for the lack of a significant effect of one of the attitude constructs on behavior.

Second, instead of using the frequency of recycling behavior as the final endogenous construct, it might be useful to measure percentages of items recycled. For individuals that do not have convenient methods of recycling (i.e., curbside pickup), frequency may not be a very relevant measure of recycling. For example, I recycle 100% of my bottles and newspapers, but I only go to a recycling center about three times a year.

A final minor point concerns the scale used to rate the values on the LOV instrument. This scale ranges from very unimportant to very important. While this response format is often employed, it is not clear how something can be "very unimportant." In other words, once something becomes unimportant, how can it be any less important. Something can have degrees of importance but not unimportance. Furthermore, it is questionable to expect respondents to be able to reliably interpret ten degrees of importance regardless of the endpoints.


The purpose of this paper is to illuminate some of the issues concerning LISREL's ability to test for causality in correlational data. The author proceeds to analyze various combinations of paths in a model examining the relationships among attitude toward the brand, attitude toward the ad, image beliefs, and utilitarian beliefs.

The paper makes two very important points concerning the use of LISREL. First, LISREL does not provide "proof" of causality. Second, a priori theory is the only true test of causality in correlational data. These are both useful caveats to remember when interpreting structural equation models. While the technique can be used in an exploratory sense, this should be done with caution, and should always be based on a priori theory. Just because a modification produces a better fit statistic or significant path coefficient does not, in and of itself, constitute the establishment of a causal link.

The only minor concern with the paper has to do with Figure 3. Both model B1 and B2 have the same degrees of freedom (17), but model B2 has one more parameter being estimated (one of the simultaneous paths between Aad and IB) so it must have one less degree of freedom. This needs to be clarified.


All three of the papers are well worth reading, and the reader can certainly benefit from them. The authors are to be commended for their work.