Meta analysis: Harbord's test using Stata's official command 'meta'

[Roman Mostazir]

Hi all,

I am using Stata's official new command suit 'meta' for a meta-analsyis. The dataset, I have, came with pre-computed effect size in log odds ratios and their standard errors. All went well using 'meta set' command. However, when I want to carryout Harbord's test for small publication bias, Stata replies that the test can only be specified with binary data. In the scenario when I cannot access the event-cell information, how can I override this warning and let Stata understand that my effect sizes are sourced from binary data? Any suggestion is appreciated.

meta bias, harbord
options peters and harbord may be specified only with binary data

[Tiago Pereira]

Only Egger's test can be carried out with log(OR) and SE(logOR). To compute Harbord's test, you must have 2x2 tables (e.g., a,b,c,d)

[Roman Mostazir]

Yes and that's a shame that Stata did not keep an option through 'meta update' or 'meta set' to declare that the variable is indeed a ln(or). At the end of the day, that is what Stata estimates as effect size from the '2X2' matrix. Egger's test is not suitable for ln(or).

[Leonardo Guizzetti]

It appears that Harbord's test requires the individual cell counts of the 2x2 table, not because it needs to compute log(OR) and its variance, but because the counts are needed to compute the components of the score statistic which are then used for the test of slope. You may be interested to see the details in their paper in the Stata Journal., or their accompanying package called -metabias-. It's not clear (to me) that you can work backwards only from the log(OR) and its variance to derive the same statistics for the regression test. In short, I don't think it's Stata's fault, but just a matter of how the test is performed.

[Roman Mostazir]

Thanks Leonardo for pointing towards the score statistic, I didn't think about it. Surely that might be the case Stata does not have an option to declare a variable as binary with 'meta set' or 'meta update' command. I can't use 'metabias' for the same reason I can't use 'meta esize' in Stata (don't have the 2x2 information available). I am not sure if I understood you correctly on the latter part of your response. Did you mean from the ln(or) and its variance I can work backwards and derive the Harbord test? How would you derive the score statistics? Can you show an example? Thanks again.

[Leonardo Guizzetti]

Sorry, I was a bit unclear. I do not think you can derive the necessary statistics for the score statistics from the summary odds ratios. For this reason neither the official meta nor metabias commands will work. I simply pointed you in that direction to see how it was being calculated.

On a side note, Stata does have some awareness of what type of data you have when using meta esize, because you need to specify the type of effect size. However, this doesn't solve your immediate concern.