444 – Contributed Oral Poster Presentations: Survey Research Methods Section
Two-factor Interaction Effect Detection for the Generalized Linear Models
Jing Shyr
IBM
Jane Chu
IBM
Sier Han
IBM SPSS Predictive Analytics
This paper proposes a method of two-factor interaction effect detection for the generalized linear models. To test whether an interaction effect is significant, a likelihood ratio test is used because the traditional ANOVA type test, which is valid for linear models with the normality assumption, is not applicable. A likelihood ratio test is to compare the log-likelihood values between the full model (two main effects and an interaction effect) and the main effects only model. The log-likelihood value for the full model can be computed without estimating parameters, but parameter estimation is needed to obtain the log-likelihood value for the main effects only model. We propose to estimate parameters for the main effects only model under a linear model framework using only basic statistics. Since those basic statistics can be computed in a single data pass, the new method overcomes the drawback of many data passes needed in the traditional parameter estimation process for the generalized linear models. Implementation of such tests in a single data pass is important for the large and distributed data sources which become increasingly common in practice now.