Article by Chin, Wynne, W., and Todd, Peter, A. (1995) in MIS Quarterly, June 1995.
Few weeks ago, one of my colleagues – Mr. H2 from the Faculty of Businesss Management – told me about his perception on SEM while we were having our breakfast together. He told me that he feels like cheating when he uses SEM, especially when he modifies his model using the ‘modification index’ (this index basically tells us what to do to improve the model’s fitness, such as linking ‘error’ of one ‘item’ with ‘error’ of another ‘item’. To a certain extent, some researchers delete few items in their initial model to get the better fit. Technically in SEM, we call this step as ‘specification search’[1] or in some articles they refer it as ‘model modification’[2].)
I think this paper provides the best answer to what my colleague has raised few weeks ago (I’m looking forward to give him the hardcopy of this article – until now I don’t have that opportunity, he is now very busy with his lectures and doesn’t have time to breakfast with me anymore:) ).
This paper is actually a critique to the work done by Segars and Grover (1993) (hereafter S&G). S&G’s work is also a critique to another work done by earlier researchers, Adams, et. al. (1992). So this paper is basically a critique on a critique.
This paper outlines 5 mistakes done by S&G and Adams. We can take these 5 issues as a note of caution and guideline for our future research using SEM.
-
Adams did not validate their measurement model prior to analyzing the ‘structural’ model. S&G argued that “unless the measurement model, which postulates the relationship between observed measures (or indicators) and their underlying constructs, is both reliable and valid, its application in testing structural relationships may lead to equivocal results… This can occur due to a confounding of substantive and measurement issues.” To overcome this, S&G suggested applying Confirmatory Factor Analysis prior to model testing [note: those who’re familiar with model development using SEM, they can always imagine the big problem that one will face if CFA is not performed at the initial stage – definitely a catastrophic ending!!!].
-
S&G used cross validation technique [note: note on what is validation techniques is provided at the end of this entry] to confirm their finding, but it was not based on a sample of independent respondents. Therefore, it exposed his model to a great probability of bias.
-
During the step of specification searches, the model’s fitness was tested after several modifications were done on the model. Therefore, no conclusion could be made as which amendment was actually contributing to the best model’s fit.
-
The sample size in the calibration set was too small to provide a stable solution for the cross validation. This paper refers to MacCallum (1986) that stated “specification searches typically show inconsistent and unstable outcomes for sample sizes of 100 to 400 observations”.[3]
-
During the step of specification searches, constructs and model modifications were guide largely by statistical considerations, and not supported by any sound and substantive theoretical rationale. So it was lack of substantive knowledge and theoretical justification.
Another valuable lesson I learn from this paper is the so called “distribution free resampling”. This approach will rest us from the assumptions of multivariate normality and the limitation of sample size. This paper provides a sample on how to conduct the “distribution free resampling” approach.
What is Cross Validation Technique
Cross validation addresses the question of how well a solution obtained by fitting a model to a given sample will fit an independent sample from the same population. It typically begins by randomly splitting a sample into two sub-samples. This provides two independent sub-samples sharing similar statistical properties. One sub-samples then used as a calibration set for model parameter estimation. These parameter estimates are then validated by holding them constant and applying them on the second sub-sample, which is referred to as the validation set. This is done to test the predictive accuracy of a fitted model, which may have provided a good fit to one data set by capitalizing on the peculiar characteristics of that data set. If the models valid, the exact parameter estimates from the first data set should predict relationships in the new sample as well.
Look at the illustration below for better understanding.
[1] This term used by MacCallum (1986) in his article “Specification Searches in Covariance Structure Modeling” Psychological Bulletin, 100:1, pp 107-120. Insya-ALLAH, I’ll put my notes about this article later in the coming post.
[3] I would personally prefer to use what Bartlett, et. al. (2001) has suggested for sample size determination. With regards to SEM, I usually combine Bartlett’s suggestions with suggestions from Benter and Chou (1987). I really hope I’ll have time to put my notes on these two papers in the coming post.