[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
st: Epidemiological Confounder Selection
Does anyone know of a Stata routine to perform confounder selection
using a selected change-in-estimate criterion and backwards deletion?
What?! I hear you say. In the rest of this message, the initiator of
this request, a friend and colleague, describes precisely what that
entails. Until now, she had done all of that "manually" in Stata and she
wonders if there is an easier way. Notice that the selection criterion
used is change in relative risk estimates and not change in P value (as
Dr. Salaheddin Mahmud
Division of Cancer Epidemiology
Department of Oncology, McGill University
Gerald Bronfman Centre
546 Pine Avenue West
Canada H2W 1S6
Ph: (514) 398-8191.
Fax: (514) 398-5002.
From: Anita Koushik [mailto:firstname.lastname@example.org]
Sent: Wednesday, July 30, 2003 9:29 AM
Subject: Epidemiological Confounder Selection
Procedure for covariate selection using a selected change-in-estimate
criterion using backwards deletion.
1. Start with a full model including the exposure variable of interest
potential covariates, determine the parameter estimate (or RR) for the
exposure variable of interest in that full model.
2. Determine the RR for the exposure variable when each covariate is
3. Calculate the absolute value of the % change in the RR between the
model and with the given covariate deleted.
4. After the % change has been calculated for each covariate dropped,
these % changes. Delete the variable with the lowest % change from the
model (the variable with the least impact on confounding).
5. Now start with a new full model with one less covariate (the variable
6. Repeat above steps. Continue until the change in estimate is at the
criterion level selected (e.g. 10% - will not drop any variable that
changes the OR by 10% or more).
7. Issue: Can a variable be forced into the procedure so that it never
drops out while other covariates are being assessed. For instance, might
want to leave age in the model regardless of whether it changes the OR
estimate substantially or not.
* For searches and help try: