[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
Re: st: Meta-analysis
Marcello Pagano <email@example.com>
Re: st: Meta-analysis
Mon, 17 Mar 2008 13:54:24 -0400
Welcome to one of the few areas in statistics where people promote throwing data away with impunity. One could say that whether you discard experiments where there are no events on either arm, or not, depends which side of the argument you are on. Since no events on either arm is evidence towards equality of the two arms, then throw the data away if you are trying to show a difference :-) Otherwise, why throw the data away?
The usual reason proffered for discarding the data is that it is embarrassing when looking at odds ratios to have to divide zero by zero. But the question you have to ask yourself is why are you looking at odds ratios? If you have to, for example if you have a case-control study, then you have a problem, otherwise stay clear of odds ratios and you won't have a problem if you use appropriate methods for analysis.
On 3/17/2008 1:24 PM, Tom Trikalinos wrote:
this is a big topic.
see for example:
Stat Med. 2007 Jan 15;26(1):53-77. Much ado about nothing: a
comparison of the performance of meta-analytical methods with rare
events. Bradburn MJ, Deeks JJ, Berlin JA, Russell Localio A.
Stat Med. 2004 May 15;23(9):1351-75. What to add to nothing? Use and
avoidance of continuity corrections in meta-analysis of sparse data.
Sweeting MJ, Sutton AJ, Lambert PC.
and quite a few other papers that are out there. The ones discussing
the recent rosiglitazone meta-analysis are also relevant. Do a PubMed
search if you have not already.
My take: Assuming you do not go Bayesian and that you use the typical
garden variety meta-analysis methods:
1. Random effects per Der Simonian and Laird are probably a no for
main analyses (biased tau^2 in simulation studies).
2. Peto OR seems to do well in terms of bias and coverage
probabilities for the CI (!).
3. Mantel-Haenszel (MH) OR seems to do well, I presume the same for RR
though i think this is not as clear, or so i remember.
4. M-H RD is reported to give somehow biased estimates and conservative CI
5. If you use multiplicative effect sizes - e.g. an OR I would
calculate main analyses without 0% vs 0% studies, then add them in in
a sensitivity analysis
All the above allowing for the caveat that one an operational
knowledge of the relevant methods literature and knowledge of which
methods need fudge factors to correct for 0 cells...
hope this helps
On Mon, Mar 17, 2008 at 12:42 PM, Sripal Kumar <firstname.lastname@example.org> wrote:
I was wondering what are your thoughts on meta analysis of trials with
limited number of events. Should studies with no events be censored
from the analysis?
Any input is highly appreciated.
* For searches and help try: