I am running glm on beetle counts data. My predictors are environmental variables and my response variable is the number of beetles.

I ran three glms:

  1. The response variable $Y_1$ is the total number of beetles.

  2. The response is a subset of $Y_1$ ($Y_2$).

  3. The response is also a subset of $Y_1$ ($Y_3=Y_1-Y_2$). In this $Y_3$, there are many zeros in my distribution, so the residual distribution is very far from normal.

How can I transform $Y_3$ to meet the assumption of normality? Does anyone have an equivalent robust non-parametric test?

  • 3
    $\begingroup$ If your outcome is a count is there any reason why you are avoiding Poisson regression? $\endgroup$
    – mdewey
    Commented May 29, 2018 at 10:56
  • 2
    $\begingroup$ Or negative binomial regression, etc... $\endgroup$
    – Glen_b
    Commented May 29, 2018 at 12:11
  • $\begingroup$ thanks for your answers. I log-transformed to normalize my response variable and diagnosed the model by checking for residuals normality... what is wrong. I will do a simple one glm(Y~X, family=poisson) and check for the goodness-of-fit the model. $\endgroup$ Commented May 30, 2018 at 6:16

1 Answer 1


You should use a Poisson regression (or in case of overdispersion, maybe negative binomial or quasi-Poisson regression). There is no assumption of normality in those models, so no need for transformations. See Normality requirements for GLM and GEE models


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.