I would need some help analyzing data that I'm not sure how to analyze. I tested ten times fifteen subjects in three different experimental conditions and my goal is to compare their behaviors between these three different conditions. Since each experimental trial does not have the same duration, and having recorded the duration of the behaviors of interest during trials, I can not directly compare the durations of behaviors by running LMEs. On the other hand, I can calculate proportions of time spent doing these behaviors but these are non-integer proportions. For example, a behavior lasting 7.88 seconds during a 30.51s test gave us a proportion of 0.26 for this behavior. And when I try to make a GLMM directly on the proportions, I get the following warning because my proportions are not integer and do not follow binomial law: "" In eval (family $ initalize, rho): non-integer # successes in a binomial glm! ". So my question is: how can I compare non-integer proportions with a GLMM if it is possible? Or via another statistical method to take into account the repetitions on individuals? I saw in other messages that beta regression would be a solution but it does not include random factor. Another possibility that I read is a GLMM TMB but I'm not sure that it can apply to my data? Thanks a lot for your help!
2
-
$\begingroup$ Sounds like this might be a case for beta regression. $\endgroup$– NatWHCommented May 22, 2018 at 15:02
-
$\begingroup$ I thought about beta regression but it seems that beta regression does not include random factor and I need to have the individuals as random factors since I have repetitions on individuals... $\endgroup$– CC_05_2018Commented May 23, 2018 at 12:34
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
1
The quasi-binomial model option on R's GLM will run on the proportions without giving the error message, but you'd need to check whether it is appropriate in other ways for your data.
model1 <- glm(proportion.data ~ predictor.data), data = Data, family = quasibinomial(link = 'logit'))
-
$\begingroup$ I need to run a GLMM instead of a GLM since I have random factors and quasi families cannot be used in glmer :/ $\endgroup$ Commented May 23, 2018 at 13:33