Cross Validated
2026-11-03 16:25 UTC
By Kaila Frazer
AI-113-20261103-social-media-cb67676a
Beta distribution for ecological percent cover data
What I'm trying to do: I'm trying to build a generalized linear mixed model of some binned percentage data. A note on the data: The response can take on values of c(0,10,20,30,40,50,60,70,80,90,100) and is highly skewed towards smaller values (the frequency plot looks like a Poisson distribution with a lambda of 1, just binned). The data was binned before I received it and I can't access raw percentages. I'm looking for help on the following two issues: Dispersion issue: I'm struggling to build a model that accurately represents the spread of the response. I've tried a variety of random and fixed effects, as well as covariates for the dispformula, and I can't get a model to predict responses greater than 50 percent. Additionally, even though I've included a ziformula, the model doesn't appear to predict zeros at all. I've tried excluding the zero and one values and just modeling the 20-90% range and the model still underpredicts high values. Hmm! Residual diagnostics issue: DHARMa response vs. residuals plots consistently have a wavy and trended shape; this is true for 10+ covariates I've tried. I suspect this issue is related to the dispersion problem, but I'm at a bit of a sticking point. Here's my test data and code: # test data dat % # beta_family() can't accept values of one, so transform those to 0.99 mutate(response=case_when(response==1~0.99,response!=1~response)) # model mod Any advice?
What I'm trying to do: I'm trying to build a generalized linear mixed model of some binned percentage data. A note on the data: The response can take on values of c(0,10,20,30,40,50,60,70,80,90,100) and is highly skewed towards smaller values (the frequency plot looks like a Poisson distribution with a lambda of 1, just binned). The data was binned before I received it and I can't access raw percentages. I'm looking for help on the following two issues: Dispersion issue: I'm struggling to build a model that accurately represents the spread of the response. I've tried a variety of random and fixed effects, as well as covariates for the dispformula, and I can't get a model to predict responses greater than 50 percent. Additionally, even though I've included a ziformula, the model doesn't appear to predict zeros at all. I've tried excluding the zero and one values and just modeling the 20-90% range and the model still underpredicts high values. Hmm! Residual diagnostics issue: DHARMa response vs. residuals plots consistently have a wavy and trended shape; this is true for 10+ covariates I've tried. I suspect this issue is related to the dispersion problem, but I'm at a bit of a sticking point. Here's my test data and code: # test data dat % # beta_family() can't accept values of one, so transform those to 0.99 mutate(response=case_when(response==1~0.99,response!=1~response)) # model mod Any advice?
Full article content could not be extracted automatically. Read the original below.
Source:
Cross Validated
· stats.stackexchange.com