Controlling for variables in multilevel logistic regression modeling

Question

I am new to mixed models and want to calculate a binary mixed model. However, I can't make much sense of the results and am hoping someone can help me out.

So I want to calculate the probability of a hit (correctly recognizing an image) as a function of PE rating (how unexpected an image is) and CS score (chronic stress score). I expect that the more unexpected an image is rated to be (as compared to the preceding image), the higher the chance of remembering it. Additionally, I want to check whether chronic stress attenuates this effect.

The data consists of 160 trials per participant, so it is a multilevel repeated measures design, with participants and CS score (continuous variable) being level 2, and PE rating (also continuous) being level 1.

FM_SCSS <- glmer(Hit ~ PE_rating_cmc + SSCS_score_gmc + 
   (1 | participantID) + PE_rating_cmc:SSCS_score_gmc, 
   data = final_data, family = "binomial", na.action = na.exclude)

When I use this model, I get a slightly significant result of rating (p = 0.025). The OR is pretty low, OR = 1.08, 95% CI [1.01, 1.16]. Does it even make sense to interpret an OR of 1.08 as statistically significant?

However, I was also advised to control for the trial number, as with proceeding trial number, the PE rating goes down because participants get used to the task, and also their memory for the images might decline with the span of time. I hope that makes sense.

I am not sure how to control for trial Number however, I have heard that you should use it as a fixed and random effect, so I am doing this:

FM_SCSS <- glmer(Hit ~ PE_rating_cmc + trialNumber + 
    SSCS_score_gmc + (1 | participantID) + (1 | trialNumber) + 
   PE_rating_cmc:SSCS_score_gmc, 
   data = final_data, family = "binomial", na.action = na.exclude)

However, this turns PE rating insignificant and trialNumber becomes highly significant instead. Is this because of multicollinearity? Because I cannot imagine that image unexpectedness does not have an effect on memory, as this has already been demonstrated in quite a few studies before. But when use plotmodel() do illustrate the interaction, it does show a steady increase in p(Hit) of PE rating, which goes up with the PE rating, but the overall p(hit) is lower, the higher the trialNumber.

I am not really sure how to make sense of these results. Can anyone help me out please?

I tried multiple variations of the model, however, I still don't know how to make sense of it.

Answer 1

I think you are right that it is due to collinearity.

Since you don't seem to be interested in the learning effect (you know it's there, estimating it is not your goal) what I would do here is change to a count model. So, for each person, you have number of hits, CS and PE.

You could model this with a Poisson model or, more likely, negative binomial.