|From||VISINTAINER PAUL <PAUL_VISINTAINER@nymc.edu>|
|Subject||st: RE: conditional logistic regression|
|Date||Mon, 24 Nov 2003 13:50:17 -0500|
I think either approach is okay. Since the number of strata is small (5-clinics), an unconditional approach, with clinics included as indicators, would model test whether the outcome varied by clinic relative to a reference clinic. In this case, the clinics represent different intercepts. If you stratified in your design, i.e., simply sampled patients within each clinic, then I think unconditional LR is fine.
In the case of conditional LR, you have explicitly established many "strata", each including only a few patients, through matching. You're not interested in explicitly modeling the intercepts because there would be many intercepts--possibly hundreds-- one for pair or cluster of patients. The intercepts are not informative (because there are so many), but they need to be modeled in order to get unbiased risk estimates.
So, in your case, either approach would appear to be okay, although conditional LR may represent an "overkill" and unnecessary complexity.
How different are the point estimates of risk? How much do the p-values differ?
From: firstname.lastname@example.org [mailto:email@example.com] On Behalf Of Ricardo Ovaldia
Sent: Monday, November 24, 2003 8:46 AM
Subject: st: conditional logistic regression
Dear Stata users,
I posted a message last week and did not receive any
replies. So let me ask it more simply:
How can I best analyze (estimate and adjust for
covariates) a dichotomous outcome measured in a 100%
sample of patients at 5 different clinics when there
is significant clinic to clinic variation in the
--- Ricardo Ovaldia <firstname.lastname@example.org> wrote:
> Dear Stata users,
> I have data from women seen at 5 outpatient clinics.
> One of the outcome variables of interest is whether
> not women are routinely performing self-breast
> We are interested in determining factors associated
> with this outcome. In the entire dataset African
> American women perform self-breast exams less often
> than Caucasian women (OR=1.96). However in one of
> clinics the rate among African American is higher
> in Caucasian women (an interaction between clinic
> race?). I had decided to estimate all odds ratios
> using -clogit- to account for the clinic effect.
> I do this for race I get an even more significant
> 2.5678 (same direction). Is it correct to use
> as I did or do I need to something different such as
> stratified analysis?
> Thank you,
> Ricardo Ovaldia, MS
> Oklahoma City, OK
> Do you Yahoo!?
> Free Pop-Up Blocker - Get it now
> * For searches and help try:
> * http://www.stata.com/support/statalist/faq
> * http://www.ats.ucla.edu/stat/stata/
Do you Yahoo!?
Free Pop-Up Blocker - Get it now
* For searches and help try: