[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: Percent correctly predicted after Asclogit or Clogit

From   Steven Samuels <>
Subject   Re: st: Percent correctly predicted after Asclogit or Clogit
Date   Thu, 3 Jul 2008 12:56:02 -0400

On Jun 30, 2008, at 11:24 AM, Luigi Curini wrote:
Hi Statalist!
any suggestions about calculating the percent correctly predicted by a given
model after Asclogit or Clogit?

This question comes up many times.

The predicted proportions from -clogit- will not resemble sample proportions, and therefore are not useful for estimating sensitivity, specificity, ROC Curves, or other measures of correct prediction.

The probability modeled by -clogit- is not the unconditional probability P(Y = 1 | X's), but a probability that is conditional on the number of "events" in the matched set. The underlying model has a different intercept for each set, and this intercept is not estimated. The default prediction (pc1) is for the probability of exactly one response in the subject's matched group, given the subject's covariates. The optional prediction (pu0) is the probability of response assuming that the group intercept is zero.

The predicted values and fit statistics produced after -clogit- can be useful for detecting model failures.


Here is code to illustrate the different kinds of predictions.

/*************CODE BEGINS*******************************/
use, clear
gen x3 = x2-1 // Create a zero level for x2

logit y x1 x3 // unconditional logit
predict phat

clogit y x1 x3, group(id)
predict p1, pc1
predict p0, pu0

sum y phat p1 p0
/* Only the unconditional logit prediction resembles the sample proportion */

sum p0 if x1==0 & x3==0 // p0 is 1/2 when the covariates equal 0

/***********CODE ENDS**************************/

*   For searches and help try:

© Copyright 1996–2017 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index