CODEWITHSUNDEEP
pythonscipydistributionstatsmodels
Xtiaan
Xtiaan

Reputation: 296

Are these the same distributions in the statsmodels and scipy libraries?

I want to make a GLM model using statsmodels in python. statsmodels.glm supports the following distrubutions:

I want to find the same distributions in the scipy library. I think these are the same:

statsmodels (docs) scipy (docs)
Family expon
Binomial binom
Gamma gamma
Gaussian norm
InverseGaussian invgauss
NegativeBinomial nbinom
Poisson poisson
Tweedie tweedie

I am not entirely sure if "expon" is exactly the same as "Family". How could I check if this is the case? Also, I think it is wise to check them all, because I am determining the best fitting distribution using the scipy package, but then apply the statsmodels glm.

Update: I thought "Family" was a distribution (in my defense, the table is quite unclear).

Upvotes: 0

Views: 218

Answers (1)

Josef
Josef

Reputation: 22897

Family is the parent class and not a specific distribution family.

The families for GLM in statsmodels correspond to distributions in scipy.stats (except maybe for tweedie). However the parameterization is different between GLM families and the common parameterization of the distributions, such as those in scipy.

In the latest release, the GLM and most discrete models have a get_distribution method that returns and instance of the corresponding scipy or scipy-compatible distribution.

For example
https://www.statsmodels.org/devel/_modules/statsmodels/genmod/families/family.html#Binomial.get_distribution which has the same parameterization, and https://www.statsmodels.org/devel/_modules/statsmodels/genmod/families/family.html#NegativeBinomial.get_distribution
where GLM parameters need to be transformed to correspond to the scipy distribution.

Upvotes: 1

Related Questions