I am running a mixed model with a binary dependent variable and a logit link using SPSS 23´s Generalized linear mixed model procedure. The data is multilevel (1 Subject, 5 questions/data points per subject) and I have included subject ID as a random factor with an intercept only. For the group under analysis, I have 65 subjects with a total of 325 observations and the full model includes 10 variables. 4 of these variables are dummy variables, 4 are likert scale confounding variables as well as 1 explanatory variable (interval scale) and 1 interaction term.
My problem is this: Looking at the BIC and AIC, SPSS considers the base model (intercept only) fits best as it comes with the lowest value (~1480). Every variable (whether established explanatory variables or new potential predictors I hypothesized) I add into the model improves prediction accuracy (~70% correctly classified with intercept only, >80% with all explanatory variables), but worsens information criteria despite strong statistical significance of predictors (up to ~1610).
Is this a problem with SPSS, did I misunderstand the SPSS output (should it be the highest value I use?) or could there be another explanation for this?