Dear Philip,

I started with binomial actually. My data are proportions (percentages actually - percent representation of a certain functional groups in the community) derived from annual counts, so yes, I only have a single data point for each year. When I fitted binomial, the data was overdispersed, so I directly moved to quasi-binomial. And here is the result. What do you suggest in this case? What should I do?

İbrahim Kaan Özgencil,

You might want to dig further into your dispersion issue. I like using the DHARMa package (https://cran.r-project.org/web/packages/DHARMa/index.html). The vignette does a good job of explaining why the package was created, how you can use it for glm/glmm, and helps with interpretation.

It might be helpful in understanding what could be causing your dispersion issues.

Dear Philip,

Thank you so much for your answer again. I checked your article and looked into "gamlss" package and and modelling beta distribution with it and I found it quite messy. Without someone to teach me I can't do it on my own. Instead, I used "glmmADMB" package to model with beta distribution. There aren't many ways to check model fit in a proper way though. I attached two plots above. The one at the top is a predicted VS observed plot and the one at the bottom is residuals plotted over the range of the response variable (residuals VS observed). The sample size is really low: only 16. What do you think about this? Is this model ok?

Dear Philip,

Thanks again for taking time to write a reply. My response variable is a proportional one. I used a "successes and failures" kind of approach so that I can use binomial (and quasi-binomial) error distributions. It's also possible to model proportional data that way. But I think we are both missing something here. My response variable is derived from COUNT DATA which is discrete (my response is percentage of a certain of group birds in a community). In such a case, I think, beta distribution should not be used. Am I wrong? I remember (from some source) that the underlying variable should be continuous one for beta to work. And if your proportional response is derived from a count, binomial is the answer. Isn't this right?

Dear Phillip,

I'm a newbie in the field of numerical ecology, so I really appreciate every piece of information I'm getting here. And thank you for the nice conversation, too. I read through the link that you sent and it actually is the best explanation regarding beta distribution I've ever come across with on the Internet. Thank you for that, too. So, yes, then beta distribution is still ok for (underlying) count data.

I don't know about the niche assumptions of this binomial assumptions model y the way. But your point here is an interesting one and I'm not equipped to answer that.

I think beta-binomial (also available from "gamlss" package) is a better fit.