enter image description hereenter image description here

Hello!

I’m attempting to build a Logit Model to predict a binary response. I’ve compared the fitted full to the null, and then the fitted full to the reduced model (Determined using backwards reduction).

Using the Chi-Sq test- the reduced model had the lowest AIC score. Furthermore, I wanted to compare this to the reduced model (with interactions). I found that including interactions between the predictor variable provides an even lower AIC score.

I was surprised to find that the Accuracy of the model with interactions was lower than that of the reduced model with no interactions. I’m relatively new, so can someone explain why this is the case? My understanding was that the AIC score was a good indicator of a models predictive capabilities.

Thank you for your time!

Logit Modeling, Chi- Squared Test to compare AIC scores.

I was expecting the model with the lowest AIC score to have the best accuracy