Using the Deviance Residual to select the target distribution.

downthesun01downthesun01
edited May 2024 in GLM.ModelSelection

Hi I'm confused by this. How would we calculate the predicted means from our GLM to be used in calculating the Deviance Residual if we didn't already select a target variable distribution?

Are we initially picking a target variable distribution (i.e., gamma for severity), building out the model, then calculating the deviance residuals, and then using that information to update the target variable distribution?

Thanks.

Comments

  • adminadmin

    Yes, building a GLM is an iterative process. You start by assuming a target variable distribution such as a gamma distribution, build the GLM and then look at the deviance residuals to assess if your choice was appropriate. If the deviance residuals were skewed significantly then you would iterate by updating to say an inverse Gaussian distribution for the target variable (greater skewness than the gamma) and re-fit the GLM. Then see if the new deviance residuals look closer to a normal distribution.

Sign In or Register to comment.