st: Interaction in logit

[<Teemu.Kautonen@tse.fi>]

Dear All

My colleagues and I are working on an analysis where we have a dummy moderator and three continuous IVs which the dummy is hypothesised to moderate. The dependent variable is a dummy and we estimated a logit model. 

We have run the analysis both using the inteff command by Ed Norton and his colleagues and by analysing the odds ratios, as advised by Jaccard (2001). 

When we enter the interactions to the inteff command line one by one, the resulting interactions are highly significant and feasible in terms of theory. However, when all three interactions are added into a single model, the significances disappear. Could this be due to multicollinearity?  

The coefficients for the interaction terms in the logit results are not significant. According to Ai and Norton (2003), this should not matter since the interaction should be computed for each observation individually (if we have understood the procedure correctly); according to Jaccard (2001) it does. 

Thus, using the guidelines in Jaccard (2001), the interactions are not significant, while according to inteff they are - assuming they are analysed one by one. 

We also ran the same logit model for both groups individually to get a feel for the group differences. This suggests that there are differences which would support the inteff results, although the differences are not really that large in terms of the odds ratios (or marginal effects). 

Does anyone have more experience with this type of analysis and with these two approaches to analysing interactions in binary regression? My colleagues and I would be grateful for any advice. 

Many thanks

Teemu


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/