-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
Conditional logistic regression and small sample sizes: what to do?
Dear R-Gurus,
[I run R 2.8.0 on WinXP. I have no formal training in statistics.]
Please feel free to skip the 'Blah blah part' for questions below.
- ---Blah blah part--------
Why I use clogit (package 'survival')?: I am looking for risk factors
for cases during a recent small "flare-up" of disease in a hospital.
To
this end I matched 11 cases (no more of them, sorry!) with 33 controls
based on sex. The next step was to perform conditional logistic
regression to evaluate risk 'exposures' for cases.
Now clogit() (Package 'survival') produces some conflicting results:
Briefly, 10 of 11 cases have a certain exposure, compared to 9 of 33
controls. The odds ratio is significant (p-value = 0.000311) while the
output of clogit() is inconclusive (.95 confidence intervals for risk
factor range from 0 to infinity, conflicting 'likelihood ratio test' and
'Wald test') plus produces following warning message, "Warning
message:
In fitter(X, Y, strats, offset, init, control, weights = weights,: Ran
out of iterations and did not converge." Tampering with the data (i.e.
reducing the number of cases with the exposure to 7 of 11) renders the
odds ratio unsignificant (p-value = 0.06681), yet makes clogit() assign
a significant p-value to the exposure in question (0.043). Furthermore,
the warning message is gone after the tampering.
Interestingly reducing (instead of increasing) cases with the risk
factor yields a significant result with clogit().
I believe now, that this behaviour might be due to small sample size.
Now I am looking for an R package to maybe circumvent this limitation of
clogit(). (Alternatively, I could just chuck the project and order a
Margerita. Maybe the risk factor is plainly not significant after all.
But I digress...)
Six years ago, the following message was posted to this list:
-----old post 2002------------
Dear R folks,
We completed a matched case-control study that was analyzed using
conditional logistic regression. Because of the small sample size we
need to calculate exact confidence intervals. The quick solution is to
purchase LogExact by Cytel. However, we'd like to do this in R. Anyone
have experience with this?
Many thanks,
Tomas
____________________________________
Tomas Aragon, MD, DrPH
-----old post 2002 end------------
At that time, the answers were not too encouraging. Anyway, I might have
the very same query 2008.
Ok, I tried to do some homework and started to look for packages maybe
implementing different methods. I did run across 'elrm' (Zamar, D.,
McNeney, B. & Graham, J., 2007), but I am unsure whether this package is
appropriate for my original 1:3 matched design.
- ---Blah blah part end--------
- ---Questions, finally!---
So my questions to you are:
1) Is the problem I describe above likely related to small sample sizes?
2) Do you have experience with 'elrm' or packages finding a solution for
the above problem? Is 1:n matching as in clogit() possible?
- ---Questions, finally! end---
Thank you for your time and effort,
Johannes
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.9 (MingW32)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org
iEYEARECAAYFAkkUZLsACgkQeixPRMMDSjthxwCgpe0ACd6qrMVCz60yzVfzML2K
EpMAnRnrfbWDrZ29TLFKMCe+bH/YrNZI
=hb1P
-----END PGP SIGNATURE-----
