2
$\begingroup$

My experimental design is roughly this: I have observations of 5 categories of behavior in a specific situation, with two individuals interacting at a time. In each dyad, I only collect data from one individual. Each time that situation is observed, I categorize the presence/absence of the 5 behaviors. The behaviors are not mutually exclusive. I also register: 1) if the individuals are facing each other or not (as this might influence the focal behavior), 2) what is the reaction of the non-focal individual, and 3) the intensity of the interaction at that point. I have 19 individuals and 248 instances of the situation I'm studying. What we want to know is if these 5 behaviors of the focal are influencing the non-focal or not, and what are the variables that have a stronger influence on this.

So I organized my variables like this:

  • n: 248 instances of the situation
  • predictor variables: facing each other (yes or no), non-focal reaction (a or b), interaction intensity (1 out of 8 categories)
  • outcome variables: 5 behaviors
  • random variable: individuals

My questions:

  1. Can I use a GLMM with this type of data/experimental design?

  2. I have a lot of missing values, since observations where not always done in good visibility conditions. I can either organize the outcome variables as present/absent (which would be binomial data, but would mean that I have 5 outcome variables) or I can combine them into a single categorical outcome variable, but would lose information about missing values, because missing values would either count as absences (and I think this is not correct) or I would have to dismiss a lot of my data. So is it possible to apply GLMM with any of these situations? If I go for the binomial data, I was under the impression that I would have to run models for each category of behavior separately?

  3. If I can use GLMM with the behaviors in categories, which would be the function and family to use?

  4. If the binomial data is more correct, I know the functions to use, but how do I connect the 5 categories of behavior for interpretation purposes?

$\endgroup$
  • $\begingroup$ How are you thinking about your response variables? Do you think of them as 5 different measures of essentially the same underlying construct, or do you think of them as independent? $\endgroup$ – gung - Reinstate Monica Nov 27 '12 at 14:26
  • $\begingroup$ @gung: the 5 behaviors are independent of each other, but ecologically speaking, it would make sense to analyze them together, since they are different states of the same thing. $\endgroup$ – purple_cc Nov 27 '12 at 15:22
  • 1
    $\begingroup$ If they're different versions of the same thing, we might not want to think of them as "independent". At any rate, being different versions of the same thing means that we analyze them together, as you note. $\endgroup$ – gung - Reinstate Monica Nov 27 '12 at 16:13
  • 1
    $\begingroup$ @gung: yes, but in the case of transforming it into a single categorical variable, could I still use GLMM? $\endgroup$ – purple_cc Nov 27 '12 at 16:30
  • $\begingroup$ @purple_cc I believe you could run a GLMM with a multinomial distribution in your case. $\endgroup$ – DirtStats Aug 29 '17 at 16:15

Your Answer

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

Browse other questions tagged or ask your own question.