I am new to this forum and have trouble with my mixed logit regression model. I have conducted an experiment with 169 participants who each performed 5 brand choice decisions (Buy brand A vs. Buy brand B). In each decision, there was a different product type (e.g. laundry detergent, chocolate bars ...) used, so it is not just a repetitive experiment but 5 individual decisions per participant, each represented by 1 line in the data set. Each decision round comes with a cost to the participant and I am trying to understand how spending in previous rounds affects brand choice in the current round.

With 5 decisions per participant I used a mixed approach with a binary response variable and modeled everything using the GENLINMIXED procedure in SPSS.

DV: Binary Brand choice

IV: PrePaid (previous rounds spending aggregate in EUR)

CV: 10 Variables (5 Product Type dummies, 3 variables changing for each round (e.g. attractiveness of product), 2 subject-specific controls that do not change between rounds (e.g. income))

My problem is now that no matter which of these variables I introduce to the model beyond the basic intercept, AIC and BIC increase. Including highly significant constructs in the model (p<.001), improving model prediction rate etc. does not make any change - I add a variable and AIC + BIC increase. Is there any structural error I am making that I am unaware of? Thinking about what I´ve done so far I see these potential problems:

- Is it maybe not okay to include a random intercept per subject for logit regression?

- When using the random intercept per subject, base prediction accuracy is >75%. Could this high baseline be the problem?

- Do I maybe need to use a normal, i.e. non-mixed model for some reason? Or is there a way I can decide whether using a non-mixed model could be appropriate?

I would be absolutely grateful for any pointers and help as I am starting to become desperate here. If any additional information is necessary, please tell me, I am willing to provide anything that might be required for solving this issue.

Thanks and best,

Malte ]]>

I got slides from A.Pelsser presentation about replicating portfolios and there was a claim that R-squared na Standard Error of Weights focuses too much on centre of distribution. Do anyone of you know what that means? I don't have any idea about that, but I'm interested.

Best regards,

Rutra1992 ]]>