[jennifer Thornton]

I am following up on this post regarding matching two different groups of patients to evaluate survival . I matched the two groups on age at diagnosis (within 3 years), presumed diagnosis (within 1 year) and sex (exact). I then used the sts test command to perform the log rank test. Is the log rank test appropriate now that my data are "matched" or is it too conservative?

[Clyde Schechter]

Because you have posted this on a thread that began with matching on the observation level, I'm assuming that you are referring to that kind of matching (as opposed to just some sort of selection process that assures balance on age, presumed diagnosis within 1 year and sex between the two groups without actually pairing up individuals across groups).

That being the case, no, the log rank test is not appropriate. Nor is it necessarily too conservative: it could be wrong in either direction. For a semi-parametric analysis, I would use -stcox- with the -frailty()- option to handle this. Or, for a fully parametric analysis, -mestreg-. Fully non-parametric survival analysis on matched-pair data is also possible, but I am not familiar with it. The attached PDF has brief reviews and references for various approaches.

[jennifer Thornton]

Thank you very much for your reply; I sincerely appreciate it. And yes, I began with matching on the observation level. I will look into the options you suggested and review the PDF for alternatives to the log rank test. Thank you again.

[jennifer Thornton]

Thank you again for referring me to the technical report listing published options regarding the analyses of paired and clustered time to event data. I am in the process of trying to obtain copies of several of these articles to evaluate their methods.

I have already used the rangejoin command (followed by keep if) to match my cases and controls (1:3 ratio) on age at diagnosis (within 3 years), presumed diagnosis (within 1 year) and sex (exact). I had previously compared their "times to event" using sts test, log rank which you confirmed was not appropriate because the data are paired/clustered. Am I correct in assuming that the other STATA options besides log rank (eg, Wilcoxon, Peto) are also not appropriate?

Interestingly, many published papers that have similar designs and evaluate time to event in paired data use the log rank test...

Thank you very much.

[Clyde Schechter]

Am I correct in assuming that the other STATA options besides log rank (eg, Wilcoxon, Peto) are also not appropriate?

Yes, you are correct.

There is much in the published literature that is not valid.

[jennifer Thornton]

Thank you very much. I will continue to look into other alternatives.