So my data is looking at the incidence of emotions within therapy sessions. I am looking at three groups that I suspect have different pattern across several emotions

Group 1 n=36 Group 2 n=12 Group 3 n=6

The variables are measured in terms of time spent in that emotion (e.g., 10 minutes of anger, 5 minutes of sadness). I have also constructed a proportion for each emotion based on how much time spent in a given emotion divided by the total time spent experiencing any emotion (e.g., 10 minutes of anger divided by 20 minutes in emotion = 50% anger in that session).

I am aware that the group sizes are uneven, and that one group is quite small - these are exploratory analyses so I would just like to find a way to compare them statistically as a starting point.

Given that the groups are different across emotions - there is a low incidence of certain emotions in some groups, thus violating normality and homogeneity of variance.

What I want to know is whether these groups are different across each emotion type. Any ideas of what would be appropriate or other considerations before making a decision?

  • 1
    $\begingroup$ There is always a non-parametric alternative, which is randomization test (closely related to bootstrap). $\endgroup$
    – stans
    Commented Feb 18, 2023 at 2:04

1 Answer 1


If at each evaluation time there is either no emotion or exactly 1 emotion, then multinomial regression should be a good choice. That models the log-odds of each of a defined set of categorical outcomes against a reference category, in a situation where there is exactly one outcome per observation. This is effectively a set of binomial regressions modeled in a way that ensures the probabilities over all outcomes sum up to 1. Normality isn't an issue. Unequal group sizes will limit the ability to detect true differences but shouldn't matter otherwise.

You should include "emotion-free" as one of the outcomes, at least to start your model, as you want there to be exactly 1 outcome at each observation. "Emotion-free" might be a reasonable choice for the outcome reference category. Group membership would be the predictor/independent variable. The above link illustrates the procedure for single observations, but the help page for the multinom function used there shows how to set up the outcome as a set of columns of counts for each of the outcomes. Those values should be integers. For example, with evaluation of emotion done once per minute, they would be the number of minutes in each state.

After you've built that model, you can interrogate the results in terms of probabilities to perform the comparisons the you want. That could include comparisons among groups of proportion of any emotion time in each emotion. Post-modeling tools like those in the emmeans package can make that task easier.

  • $\begingroup$ I appreciate your answer! This sounds very useful - but would there be a benefit to using this rather than a Kruskal-Wallis test instead? $\endgroup$
    – user380249
    Commented Feb 18, 2023 at 19:38
  • $\begingroup$ @user380249 with Kruskal-Wallis (K-W) tests you would have to do a separate test for each emotion, leading to possible multiple-comparison problems. Also, unless the distributions within groups have the same shapes and scales (which doesn't seem likely here) it's hard to interpret K-W with respect to anything other than stochastic dominance. Multinomial models directly evaluate probabilities and probability differences. $\endgroup$
    – EdM
    Commented Feb 18, 2023 at 20:19
  • $\begingroup$ Thanks @EdM!! I have done logistic regressions before but didn't know the option of multinomial. This will work great for my data :) $\endgroup$
    – user380249
    Commented Feb 23, 2023 at 16:22
  • $\begingroup$ @EdM wouldn't the dependent variable be the group membership? Then the IV's would be predictors of group membership? $\endgroup$
    – Statsyy
    Commented Feb 24, 2023 at 17:21
  • $\begingroup$ @Statsyy you can use "emotion" as the dependent variable in a multinomial regression if, at any observation time, there is either no emotion or exactly 1 emotion among a list of possible emotions. Then Group could be used as a predictor variable. You could consider analyzing the other way around, as your comment suggests, although results might differ. For example, with a 2-class situation you could evaluate class membership as a function of predictor values, or distributions of predictor values among classes. This is just the multi-class extension of that principle. $\endgroup$
    – EdM
    Commented Feb 24, 2023 at 18:46

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.