0

I ran a brms model in R with a complex three-way interaction that I am trying to decompose. See below for code. I have a multinomial dependent variable with three levels (happy/angry/fear) and multiple categorical factors: group (three levels; social anxiety, other anxiety, non-anxious) / emotion (three levels; happy, angry, fear) / percentage (four levels: 10%, 30%, 50%, 70%). I am trying to figure out how I can answer the following question: Do groups differ in their responses to emotional faces (e.g., is the socially anxious group more likely than the other groups to pick angry as their response to happy or fearful faces at any of the percentage levels? The three-way interaction was significant. My question is, how should I use emmeans to get the output I would like to answer this question? Thanks so much in advance for your response.

model code: cormod_multi <- brm(Emochoice ~ f_group + f_emotion_r + f_percentage + f_groupf_emotion_rf_percentage + (1 + f_emotion_r*f_percentage | f_ppn), data = data_mult, family = 'categorical', warmup = 1000, iter = 10000, chains = 4, cores = 10, control = list(adapt_delta = 0.95, max_treedepth = 15))

user17821937
  • 1
  • 1
  • Look at the documentation for `contrast.emmGrid` and `vignette("interactions", "emmeans")`, especially the parts about interaction contrasts. – Russ Lenth Jan 04 '22 at 01:01
  • Thanks a lot. I have read the documentation and I understand how to dissect the fixed effects and their interactions. What I don't understand is how to get these effects separately for each level of the multinomial dependent variable (I have updated my question to make this clearer) which has three levels (happy/angry/fear). I am also running into an issue that I cannot pass a brms object to all emmeans commands (with some I can), I saw this has been posted before(https://discourse.mc-stan.org/t/using-emmeans-to-get-marginal-means-from-a-multinomial-logistic-regression-in-brms/24749). – user17821937 Jan 04 '22 at 09:04
  • 1
    The emmeans support for brms objects is in the brms package; and I believe that it is presently true that brms multinomial models are not completely supported. What is needed here, I think, is to get direct estimates of the multinomial probabilities at the three response levels. I am not sure what to suggest except to do a frequentist analysis using net::multinom, which is supported. – Russ Lenth Jan 04 '22 at 13:59
  • Thank you, that is helpful. I will follow it up with frequentist analyses. – user17821937 Jan 05 '22 at 08:53

0 Answers0