1
$\begingroup$

I've got a dataset where I try to predict a between-subject dependent variable (continuous), the average number of times a participant made a mistake on a cognitive task, with a within-subject predictor (categorical), hard vs easy trial, and a between-subject predictor (continuous), mean activation in a brain region for this trial. There are approximately 3000 observations, with approximately 70 observations per participant. I am not sure whether a mixed-effects model is the right approach or whether there is a better way to approach this? Does someone have any suggestions?

$\endgroup$
3
  • $\begingroup$ Could you elaborate a bit and explain exactly what is your dependent variable, and what are your predictor variables? It would be good to know how many observations there are per person, and what are your research questions $\endgroup$
    – user139190
    Aug 27, 2019 at 14:38
  • $\begingroup$ edited the information $\endgroup$ Aug 28, 2019 at 13:13
  • $\begingroup$ Hi Sebastian, given your description of the data, a mixed-effects model seems to be a great way to work with these data. Please see this vignette from package "afex" which might help you understand why it could help cran.r-project.org/web/packages/afex/vignettes/…. Many neuroscientific studies with many trials per participant have used lmms recently, and as some other papers show, it is a great way to deal with the non-independence of these types of data. $\endgroup$ Aug 28, 2019 at 21:40

1 Answer 1

1
$\begingroup$

As mentioned by Craig in the comment, mixed-effects model is appropriate. In short, multiple observations per participants = hierarchical data.

On another note, your outcome variable (DV) is unlikely to be normally distributed, so you should not really run a linear mixed-effects model assuming a normal distribution. You stated that you are taking an average of counts (i.e., number of times participants made a mistake on a cognitive task) -- aggregating data should be avoided as it artificially lowers the variance. Treating the outcome as counts and running a mixed Poisson model is likely to be more appropriate given your data.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.