blog/2021/11/08/beta-regression-guide/index

[giscus[bot]]

blog/2021/11/08/beta-regression-guide/index

Everything you ever wanted to know about beta regression! Use R and brms to correctly model proportion data, and learn all about the beta distribution along the way.

https://www.andrewheiss.com/blog/2021/11/08/beta-regression-guide/index.html

[igorkf]

Hey, this is the best beta distribution post ever!

How about this approach for tricking with the (0, 1) bound?
Have you ever used this?

Smithson and Verkuilen (2006)
smithson_verkuilen <- function(y) {
  n <- length(y)
  return((y * (n - 1) + 0.5) / n)
}

For my data it "plays" pretty well, but I'm not sure whether should I use this approach:

> summary(data$resp)
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
0.00000 0.03774 0.17241 0.21562 0.34615 0.94118 
> summary(data$resp)
     Min.   1st Qu.    Median      Mean   3rd Qu.      Max. 
0.0003636 0.0380720 0.1726520 0.2158268 0.3462657 0.9408556

[KeterAlfred]

Nice presentation. How would you interpret the Group level effects in the random effects beta regression/zero-inflated beta regression model for both the mean and the precision parameters? How would you compute the intra-class correlation coefficient (ICC)?

[eggsbrown]

Love that you covered fractional logistic regression. I use this all the time. Usually I account for "0-1 inflation" using a fractional logistic regression + observation level random effects. This approach is described here -- https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4517959/

[sidmehta1975]

Thanks for a very clear tutorial professor! Also, very grateful for all the hard work you put into making your code available to the public. Can’t thank you enough.

[RaymondLTremblay]

Hello Andrew

Thank you VERY much for going through this step by step. It is really appreciated. Your Bayesian approach is super cool, mostly that you include a zero-inflated approach. I work with evolutionary processes in orchids and most plant have 0% fruit set, so it will be very useful. I'll clearly will use your website in the future.

On another note would you have a method of doing model selection and evaluating all the combination of variables, something like the dredge function in MuMIn package? That is my present challenge. Much appreciated Raymond

[tylerz-145]

I am applying this code to my own data. I am having an issue with the marginaleffects() in the code. It keeps giving me the error message:

Error in typical(Holding = c(13, 26), model = list(coefficients = list( :
could not find function "typical"

Do you happen to know why this may be? Any help is much appreciated!

[andrewheiss]

This is because {marginaleffects} has matured a lot since November 2021. It no longer uses typical(), but datagrid() instead. I've update the post to reflect that.

[tylerz-145]

May be a dumb question, but for the Beta regression or Bayesian Beta regression, is there a way to simply output the 95% CI around the mu estimates for each group? I know we have the the mu and phi values, but not sure how to code just simply getting the 95% CI for each mu.

[tylerz-145]

• credible intervals for the Bayesian model

[andrewheiss]

You can use something like median_hdi() or mean_qi() or other related functions from {tidybayes} (see http://mjskay.github.io/tidybayes/articles/tidybayes.html#point-summaries-and-intervals-with-the-point_interval-functions-medianmeanmode_qihdi):

model_beta_bayes %>% 
  gather_draws(`b_.*`, regex = TRUE) %>% 
  median_hdi()

[andrewheiss]

And that works with other dplyr functions, so you can group by specific columns:

model_beta_bayes %>% 
  gather_draws(`b_.*`, regex = TRUE) %>% 
  group_by(whatever) %>%
  median_hdi()

[Miki-E]

This is literally one of the best descriptions of beta and inflated beta models I have ever seen!

Well done.

Is it at all possible to add something of the same quality describing how we test if the zero inflated beta is a good fit, and possibly some bayesian comparison (BIC or DIC) for chosing which zero inflated beta model covariates to keep in the model?

[Crispy13]

Thank you for posting!

It helped me a lot in my work.

[Evattimo]

Hi. This explanation is amazing! However, I understood that the explanation regarding the 0-inflated beta regression only uses the categorical variable (quota) as an example. My variable is continuous, which is addressed by the regular beta regression in this tutorial. I did not see any example for a zero-inflated beta regression with a continuous covariate. It would be very helpful if we could discuss that. Thank you.