7
$\begingroup$

In order to analyse which factors have greater weight in the proportion of incidence (number of infected inidivuals against total individuals) difference within different habitats a Generalized linear mixed model was performed. Normality and homoscedasticity were tested by Shapiro-Wilk and Levene test. Incidence data did not follow a normal or homoscedastic distribution. The data showed that followed Poisson distribution, so the GLMM was adjusted for a Poisson distribution.

However, Poisson distribution is for rate or count data. Are there complications when using the Poisson with proportions?. Rates can be standardised by time. But, proportions are typically modelled using the binomial distribution? What alternative can I use to adjust the GLMM?

Thank in advance

$\endgroup$
1
  • $\begingroup$ Do you have repeated measures within individuals/subjects ? What are your random effects ? $\endgroup$ Commented Sep 27, 2020 at 11:26

1 Answer 1

7
$\begingroup$

Normality and homoscedasticity were tested by Shapiro-Wilk and Levene test.

Note that there is no requirement for the data to be normally distributed, nor homoscedastic. These tests are a waste of your time.

If you want to model the raw counts as the outcome then a count model such as Poisson or negative binomial would be appropriate.

If you want to model proportions as the outcome, then a count model will be inappropriate. A regular linear mixed model (LMM) might be sufficient. You can plot the residuals of the LMM and if they are approximately normally distributed, then all is good. If not then, you may consider a binomial GLMM or a beta GLMM.

$\endgroup$

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.