3
$\begingroup$

I've been looking all around the webs but cant find a conclusive answer. I have count data for a longitudinal study where subjects were grouped into three treatment groups (A,B,C) and blocked by litter and starting weight category (high, med, low). I'd like to use a linear mixed model (lme4 R package) to ask questions like: which factors are most indicative in differentiating each group of subjects?

I've only found examples using mixed models that use either a continuous response variable or a dichotic (0/1) response. In my case, my response is categorical with three groups. Is it possible to use mixed models (and more specifically glmer()) with a categorical response of more than two outcomes? Do I simply specify a binomial family (probit or logit)?

$\endgroup$
3
  • 1
    $\begingroup$ You seem confused about your response. You say that you have "count data", which usually implies that your response variable consists of integer counts, 0, 1, 2, 3, ... but then you talk about having three groups A, B, C and you say "my response is categorical with three groups". So is your response counts, 0,1,2,3,4,... or is it groups A, B, C? $\endgroup$ Nov 12, 2015 at 18:20
  • $\begingroup$ I agree with @Gregor. I don't think you will get any sensible answers without first clarifying what your independent and dependent variables are and communicating them to the forum. $\endgroup$ Nov 12, 2015 at 18:49
  • $\begingroup$ I see how that can be confusing. My count data is counts of numerous different types of bacteria. I'm trying to use this data in order to predict which category (A,B,C) the animal belonged to. Each category of animal was given a different treatment which should impact the counts of different bacteria. $\endgroup$ Nov 12, 2015 at 19:22

1 Answer 1

1
$\begingroup$

There is no reason that you cannot use a 3 level categorical response in a mixed model.

However, if your dependent variable is a count, then you may not want a linear mixed model, you may need a nonlinear one that is designed for count data.

$\endgroup$
2
  • $\begingroup$ any suggestions on which type of model to use then? $\endgroup$ Nov 12, 2015 at 19:22
  • $\begingroup$ Poisson, negative binomial or maybe one of the zero inflated versions of each. $\endgroup$
    – Peter Flom
    Nov 13, 2015 at 12:33

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.