I am trying to fit a model with some pretty sparse data; I cannot collect more data so please refrain from that suggestion. I do not have a large sample size, so I have issues with singularities for complex models. What I have is 25 subjects (ID) each with two Trials (Trial). Trials are of varying duration (Duration), which means that the Behaviors (Grim) vary due to Trial and Duration. I need to control for Duration and Trial, but what I am really interested in is the ID variable and whether the effect of ID remains consistent across trials. That is, are subject scores repeatable while controlling for a duration effect and the mean trial effect. What I think is that I should be testing a slope variable: does the Behavior remain consistent across Trial 1 and 2 within each individual (slope of each individual’s behavior across trials), while acknowledging that Duration of trial has an influence on likelihood of a behavior occurring and that Trial 2 likely has a more muted response in general across all individuals.

I am having trouble parameterizing this in R, though. Essentially, I have n=1 per ID/Trial class with 25 IDs and 2 Trials per ID, since I was only able to run two Trials per ID. I cannot, therefore, run 1 + Trial|ID (Error: number of observations (=50) <= number of random effects (=50) for term (1 + Trial | ID)).

Where I am at, right now, is two run 3 models and then test if they significantly differ with an ANOVA. I am unsure if, though, this gets to the issue at hand.

Here are the three models, the ANOVA, and my rudimentary interpretation:

m1: Grim ~ Duration + Trial
m2: Grim ~ Duration + Trial + (1 | ID)
m3: Grim ~ Duration + Trial + ID
   npar    AIC    BIC  logLik deviance   Chisq Df Pr(>Chisq)   
m1    4 351.55 359.20 -171.78   343.55                        
m2    5 352.90 362.46 -171.45   342.90  0.6527  1   0.419136   
m3   28 355.98 409.51 -149.99   299.98 42.9183 23   0.007082 **
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

So, m1 is just the effect of Trial and Duration on the frequency of the Behavior. m2 adds ID as a random effect, and does not significantly differ from model 1. As a random effect, the slope of trials within each individual does not vary when taking Duration into account, right? So this is my model of interest? I also tried m3, which would emphasize between individual effects, yes? That between individual variation do significantly predict Behaviors, indicative of different intercept. Please correct me as needed so that I am modelling my variable of interest. I do know about rptR, but I fear that my data are not robust enough to use it without issues in singularity.




Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.