Compute ROC after Poisson Regression?

[Zhianni Yang]

Hello, I am using poisson regression to have relative risk estimation instead of logistic where it only computed odds ratio. My command is as follows:

poisson Gagal ib0.Rujukan ib0.Kategori_BBL ib0.Tidak_ASI ib0.Tumbuh ib1.Perubahan_Leukosit_NEW ib1.Perubahan_Neutrofil_NEW, vce(robust) irr base

I tried to compute the roc using lroc but to no avail. I read it's only avail for the logit/logistic/regress command.

So i tried to modify by change the poisson command to logit and ran the lroc command which successfully applied.


Is my method correct? Or is there any way to compute roc after poisson regression?

Thanks so much

[Rich Goldstein]

if you only want the C statistic (AUROC), then you can use the user-written -somersd- command (use -search- to find and download); the help file tells you how to get C; you can probably also use the official -roctab- command; again, read the help file; with -somersd-, you cannot automatically get a graph if you want one but you can build your own; with -roctab-, you can get a graph

[Leonardo Guizzetti]

A ROC curve would not be very informative from your model. Predicted probabilities from the logistic and Poisson models will not be the same, and there's no guarantee they will yield the same c-statistic. However, you can take your Poisson regression and compute predicted probabilities (assuming your dependent variable is binary 0/1 coded). Then you can use the predicted probability and observed outcome with the advice given by Rich.

[Zhianni Yang]

Thanks Rich Goldstein !! May I confirm:
1. roctab yields a roc for each variables (in an adjusted model where n variable enters the model, a total of n roc will be provided); and
2. lroc yields a roc for a model itself (in one adjusted model, only one roc will be provided)

is my udnerstanding correct?

[Rich Goldstein]

if the second variable you use in your -roctab- command is the predicted probabilities from the model, then it will give you the same result that lroc does; if you, instead, use one predictor, then it will give some form of "partial" AUROC; so, your first statement is incorrect if you use the predicted probabilities from your full model