Stata The Stata listserver
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: clogit or logistic for matched pairs


From   Jay Kaufman <Jay_Kaufman@unc.edu>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: clogit or logistic for matched pairs
Date   Fri, 07 Nov 2003 10:26:25 -0500

Joseph Coveney wrote:
> 
> In lieu of a Mantel-Haenszel command for risk ratio in Stata, I suppose that
> you could try something like -gllamm/xtgee , i(pair) family(binomial)
> link(log) eform- in a pinch, but is there a reason to prefer risk ratios
> over odds ratios?  Is it for ease of interpretation?  I don't have Rothman
> and Greenland's text, but it seems from Jay's post that they abide by the
> convention of odds ratios for case-control studies, risk ratios for cohort
> studies.  Does that represent universal thinking among experts currently for
> analysis and reporting of cohort studies with a binomial outcome?  I just
> was under the impression that odds ratios were a more "natural" metric for
> this kind of data regardless of whether from a case-control or cohort
> design.


The argument has indeed been made that odds ratios are a more 
"natural" metric.   For example:

	Walter SD.
	Choice of effect measure for epidemiological data.
        J Clin Epidemiol. 2000 Sep;53(9):931-9. 

However, these (statistical) arguments overlook an important 
deficiency of the odds ratio which is that it is not collapsible. 
This deficiency makes it useless as a CAUSAL measure (as opposed
to a statistical measure) unless it approximates the risk ratio
by virtue of either rare outcome or study design (e.g. density
sampling).  

For explanation and demonstration of non-collapsibility of the OR: 

	Greenland S, Morgenstern H.
        Confounding in health research.
        Annu Rev Public Health. 2001;22:189-212. 
	(especially pages 203-206) 

For explanation of why the OR is therefore deficient as a measure
of causal effect, regardless of its statistical properties
(unless it approximates the RR): 

	Greenland S.
        Interpretation and choice of effect measures in epidemiologic 
	analyses. Am J Epidemiol. 1987 May;125(5):761-8. Review. 

This deficiency of the OR arises because collapsibility is generally 
our empirical criterion for confounding, and so a measure that is
not collapsible even when there is no confounding is difficult to 
impossible causally.  The ramifications of this problem surface 
in countless ways, for example in the problem of how to decide if 
one should adjust or not adjust for covariates in a randomized
controlled trial (RCT): 

	ROBINSON LD, JEWELL NP
	SOME SURPRISING RESULTS ABOUT COVARIATE ADJUSTMENT IN 
	LOGISTIC-REGRESSION MODELS
	INT STAT REV 59 (2): 227-240 AUG 1991 

- JK

-- 
Jay S. Kaufman, Ph.D         
-----------------------------
email: Jay_Kaufman@unc.edu
-----------------------------
Department of Epidemiology   
UNC School of Public Health  
2104C McGavran-Greenberg Hall
Pittsboro Road, CB#7435   
Chapel Hill, NC 27599-7435  
phone:  919-966-7435         
fax:    919-966-2089         
-----------------------------
*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/



© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index