GAMLj3 - robust standard errors for poisson regression

[diogohcc]

Dear colleagues.

Is the function for estimate robust standard errors not available for poisson regression? In the module GAMLj3, the option of robust standard errors is available only for General Linear models, but not for generalized linear model.

Best regards.

[mcfanda@gmail.com]

At present, they are not implemented. Nevertheless, what would be the theoretical rationale for employing robust standard errors in generalized linear models? In standard GLMs, robust estimators are typically used to correct for heteroscedasticity or related violations of model assumptions; however, heteroscedasticity is not an assumption of generalized linear models. If one wishes to move beyond standard errors based on model assumptions, a bootstrap confidence interval may be preferable, as it can provide insights into the stability of the estimates without relying on specific distributional assumptions.

[diogohcc]

Thank you very much for your reply.
I agree with the comments.
However, Poisson regression with robust standard error is a way to estimate the prevalence ratio for highly prevalent dichotomous outcomes (in this case do not assume Poisson distribution). It provides a better estimate compared to logistic regression, as well as in cases where negative log-binomial regression present problems of convergence. Robust adjustment for Poisson regression is available in STATA, but I have chosen to use jamovi because I appreciate the project. However, in this case, the absence of robust standard error prevent to use jamovi.

Best regards.

[mcfanda@gmail.com]

Well, everything can be done if can be useful. The question is: what robust standard error do you actually use? To be precise, do you mean the standard error STATA produces with vce(robust). If so, they are the HC estimators, aka heteroskedasticity-consistent covariance matrix estimators. If that's correct, we can implement them for the Poisson model in next version of GAMLj3.

[mcfanda@gmail.com]

consider testing the results you obtain with HC SE against the results of gamlj3 Poisson (overdispersion), which employs a quasi-poisson distribution that corrects for overdispersion, pretty much what the HC SE do.

[mcfanda@gmail.com]

Done, robust standard error for poisson and other generalized models are now implemented. Check version 3.6.4 on monday

[diogohcc]

Well, everything can be done if can be useful. The question is: what robust standard error do you actually use? To be precise, do you mean the standard error STATA produces with vce(robust). If so, they are the HC estimators, aka heteroskedasticity-consistent covariance matrix estimators. If that's correct, we can implement them for the Poisson model in next version of GAMLj3.

Dear colleague, thank you very much for your reply and help.
In Stata, I only enter the VCE (robust) code, with no other specifications.

[diogohcc]

consider testing the results you obtain with HC SE against the results of gamlj3 Poisson (overdispersion), which employs a quasi-poisson distribution that corrects for overdispersion, pretty much what the HC SE do.

I compared the results obtained using the overdispersion correction available in gamlj3 with those obtained using the robust standard error in stata. The differences in results were minimal, to the second decimal place, for both the irr and the confidence intervals.

[diogohcc]

Done, robust standard error for poisson and other generalized models are now implemented. Check version 3.6.4 on monday

The jamovi download site only lists versions 2.6.44 and 2.7.11. I downloaded and installed both, but the option for robust standard error in Poisson regression in GAMLJ3 or GAMLJ still doesn't appear. I also uninstalled and reinstalled the modules, but I still can't find the option.

[diogohcc]

Would version 3.6.4 be the module version, not the software version? However, I searched the program library today, and the only available GAMLJ3 module is version 3.6.3.