Skip to content

[R-meta] Fwd: meta analysis package "metafor" --- one group survival outcome?

3 messages · Hong Zhao, Viechtbauer Wolfgang (STAT), Michael Dewey

Original
Hong Zhao
# 
Dear all,

I have a question for R package "metafor". On page 35 of the package file,
I saw the meta-analysis for a two sample survival analysis, does anyone
know whether I can use this package to do one-sample survival analysis?

Here is what I have: three 3 studies with all the survival information
(time to event and censor information). I want to combine the survival
rates (KM estimates) at 6 months for these 3 studies using meta-analysis.
Since the survival rates are not normally distributed and I cannot just
simply treat them as proportions to use meta-analysis method for normal
data or proportion data, I am not sure whether I can just transform them
using log(survival rate) (and get their variance accordingly) and use
meta-analysis for normal data. Does anyone know a better way to do it? Or
can the package perform the meta-analysis for a single group survival
outcome?

BTW, does anyone know any method for sample size calculation for
meta-analysis (outcome is one-sample binomial or survival outcome as I
mentioned above)? I want to perform non-inferiority test, but did not find
anything available from online for meta-analysis sample size
calculation for either binomial or survival outcome. I guess I may need to
just do simulation if there is no available sources.

Thanks so much for your kind help!

Best!
Viechtbauer Wolfgang (STAT)
# 
A method for pooling KM estimates (at a particular time point) across multiple studies is described in:

Coplen, S. E., Antman, E. M., Berlin, J. A., Hewitt, P., & Chalmers, T. C. (1990). Efficacy and safety of quinidine therapy for maintenance of sinus rhythm after cardioversion: A meta-analysis of randomized control trials. Circulation, 82(4), 1106-1116.

In essence, we can just pool the KM estimates in the usual manner (FE/RE model). The variance of the KM estimates can be computed using Greenwood's formula (https://en.wikipedia.org/wiki/Kaplan%E2%80%93Meier_estimator#Statistical_considerations).

You can do the pooling with metafor (or other packages). In metafor, just use rma(yi, vi) where yi are the KM estimates and vi the corresponding variances.

This approach uses the KM estimates directly. Indeed, the sampling distribution of KM estimates may not be normal, but that may not be a major issue unless you suspect that the underlying true survival rates are close to 0 or 1.

As for your question about a sample size calculation -- I am not quite sure what you are after. You (typically) do not have any influence on the number of studies included in a meta-analysis or the sample sizes thereof. So what are you trying to accomplish here?

Best,
Wolfgang
Michael Dewey
# 
Dear Hong

Comment in-line
On 28/06/2017 03:25, Hong Zhao wrote:
I think you are looking for incidence rates here. There are some options 
outlined in the documentation for escalc under the heading methods for 
event counts.
Why not just do your analysis and report confidence intervals? It is not 
like a trial where you can always recruit a few more cases so knowing 
you need 100 primary studies is hardly going to help if only seven exist.
In this case it did not matter but this is a plain text list and sending 
HTML mangles your message.