For this particular study, I had 3 monkeys complete 3 tasks, each task has a binary outcome (Win/Lose). A task is considered complete when they reach a criterion of 85% accuracy in their most recent 120 trials. So for each monkey I have lists, varying in length, of W's and L's. Originally I planed on using a Fisher's exact test, for each individual monkey, to compare the number of Wins and Losses for Test 1 and 2, and then for Test 1 and 3, and for Test 2 and 3. However, I've read that it is not appropriate to use a Fisher's exact test for repeated measures data and was told it may be possible to use a binary logistic regression instead. I've been looking into the binary logistic regression and have found mixed reviews on whether this test is appropriate for repeated measures data. I suppose my question is if binary logistic regression would work for this dataset? and if not, what other model would be more appropriate?


If you want to use a binary logistic regression technique, you'd have to use a longitudinal regression technique which will handle repeated outcomes having values 0,1. Examples are GLM or GEE with a family or "link function" set to "logit," i.e., logistic. GEE has the advantage of not requiring uniform time periods, meaning the time intervals can vary across measurements and animals. There are also "mixed" regression modules which will treat each animal as its own cluster (random effects) in order to accommodate the within-subject correlations.

  • $\begingroup$ I've looked into using GEE with the logit function, however, sample sizes have to be even for that to work. Since I am analyzing each monkeys data individually, the trial counts for each task are the sample size. Due to the criterion they had to meet to finish the task, each task was completed in a different number of trials. The mixed regression models would work if I wasn't analyzing the monkeys data separately, however, that is not the case here. $\endgroup$ – Brooke Jackson Apr 12 at 17:49
  • $\begingroup$ You can have a different number of repeated measurements for each animal in a longitudinal regression model. It may be a limitation of the package you looked at. With highly imbalanced observations over time, then maybe look at separate Somer's D analyses for each group, using a cluster ID variable (for each animal), outcome (0,1) and time of the outcome. This analysis is like a correlation between time and outcome and will provide a handle on whether W or L go up or down over time within the treatment group analyzed. Somer's D coefficient ranges from -1 to 1, like correlation. $\endgroup$ – wrstks Apr 12 at 18:05
  • $\begingroup$ I’m not comparing all animals to each other. I’m comparing a single animals performance across 3 tasks. It’s not that each animal has different numbers of trials, it’s that all 3 tasks for each monkey has different trial numbers. The issue with RM tests is I can’t compare 145, to 215, to 160 trials of w/l for one animal. Im trying to compare monkey1s 3 tests to each other, which vary in size. All of the repeated measures tests, including longitudinal regressions models, require the 3 samples that are being compared to be even. I have not heard of the somer’s D, I will look into that. $\endgroup$ – Brooke Jackson Apr 12 at 21:30

Your Answer

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

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