statsmodels.stats.outliers_influence.GLMInfluence.resid_score_factor

GLMInfluence.resid_score_factor()

Score residual divided by sqrt of hessian factor

experimental, agrees with GLMInfluence for Binomial and Gaussian. This corresponds to considering the linear predictors as parameters of the model.

Note: This might have nan values if second derivative, hessian_factor, is positive, i.e. loglikelihood is not globally concave w.r.t. linear predictor. (This occurred in an example for GeneralizedPoisson)