Beta regression in Statsmodels?

[Eridk Poliruyt]

Hi all,

It seems that Beta regression hasn't been included in statsmodels? The package now only includes those one-parameter exponential family likelihoods for generalised linear model, such as Poisson, logistic. I wonder if it is possible to implement Beta regression using existing functionalities of the package? I guess it is not that straightforward, otherwise it should have been implemented? Thanks!

Best regards,

Erick

[josef...@gmail.com]

On Fri, Sep 15, 2017 at 4:28 PM, Eridk Poliruyt <ep1...@gmail.com> wrote:

Hi all,

It seems that Beta regression hasn't been included in statsmodels? The package now only includes those one-parameter exponential family likelihoods for generalised linear model, such as Poisson, logistic. I wonder if it is possible to implement Beta regression using existing functionalities of the package? I guess it is not that straightforward, otherwise it should have been implemented? Thanks!

https://github.com/statsmodels/statsmodels/pull/2030

or maybe https://github.com/josef-pkt/statsmodels/commits/betareg_2030_3

I don't remember why it stalled (except that I get easily distracted and have been more interested in LEF, QMLE and misspecified models for the last few years, i.e. use Logit/Binomial QMLE for fractional regression)

One problem is that our standard design does not provide much support for two exog. (I recently opened https://github.com/statsmodels/statsmodels/issues/3903 )

start_params and convergence issues are also showing up in many models.

I'm currently at zero-inflated models, which are a bit similar in terms of extra parameters or extra exog. So, BetaRegression might be easier now.

Josef

 

Best regards,

Erick

[Eridk Poliruyt]

Thanks Josef! So I suppose the implementation of the simpler Beta regression where the precision parameter is assumed to be constant should be much easier with the existing tools?  

在 2017年9月15日星期五 UTC+1下午11:00:38，josefpktd写道：