In search of multivariate associations: comparison of CHAID, log-linear analysis, and multiple correspondence analysis
Svetlana Zhuchkova has presented her graduation paper on the RSG seminar on 5th June. The paper was based on the comparison of three methods of searching for multivariate associations of categorical variables. The comparison was conducted in the paradigm of Data Mining.
Categorical variables, which are nominal or ordinal scales, are widely spread in sociological studies. However, the analysis of such data is usually limited to two-dimensional relations while there are many theoretical and empirical assumptions for multivariate analysis of such variables. It was the crucial issue of Svetlana’s graduation paper, and she tried to compare CHAID, log-linear analysis, and multiple correspondence analysis. The research was intended to find out the most effective method regarding interpretation and prediction quality. There were some hypotheses based on Svetlana’s investigative practice and depth scientific literature review. She presumed that the log-linear analysis would provide the most informative interpretation, and CHAID would be more suitable for prediction tasks.
After theoretical comparison of methods and operationalization of key variables for empirical comparison, Svetlana constructed multinomial regression models with interaction effects that were detected by CHAID, log-linear analysis and multiple correspondence analysis with the same categorical dependent variable. The results afford ground to reject the hypothesis of the study. Contrary to expectations, multiple correspondence analysis was the best in criteria of in-depth interpretation because this method detected more multivariate associations and their dimension was bigger than those found by another two methods. Log-linear analysis was more suitable for prediction tasks because its Nagelkerke R2 was statistically significantly larger than CHAID’s one. Svetlana suggested why these results were derived and gave some ideas for the future investigations. Thus, deputy head of the RSG presented the example of the research that is worth to follow and that other members of the group were interested in.