Help with sample size calculation for incidence rate ratios

[Lauren Cohee]

I am working on design of a 2-arm trial to compare incidence rates of a clinical outcome. I anticipate that the annual incidence rate is 0.7 cases person. I would like 80% power at alpha=0.05 to detect a 20% reduction in the annual incidence rate (irr in intervention arm 0.56) with follow-up over only 1 year. Arms will be compared by poisson regression using a log-link and offset equal to the log of the number of days a person was followed.

1) what is the best sample size calculation to use in this scenario as I don't see a clear answer in the basic power and samples size commands? (unless I treat the incidence rate as just a proportion which seems to simple and gives numbers lower than I anticipated)

2) if I need to enroll participants from multiple sites to achieve the sample size, how should I account for that in the sample size calculation?

Appreciate thoughts.
Lauren

[jameskb101]

Hi Lauren Cohee, I'm curious as to whether you ever figured out an answer to this question? I'm facing a similar problem and have got as far as what you say in your point (1). This older thread might be useful Re: st: power calculation and sample size - I wonder if you need to estimate the incidence proportion from the incidence rates and use that as the basis for estimates.

[Joseph Coveney]

If you cannot find an analytical solution, then perhaps you could use simulation for the power analysis. FAQ on the general method here and blog entries starting here, among other sources.

I take it that these are longitudinal studies. Are you marking only the first event each participant within the observation interval? If not, then to program the simulation you'll need to decide on a distribution of the random effect.

[Clyde Schechter]

This problem is beyond the scope of the sample size calculations that come built-in to Stata. But there are standalone sample size calculation programs that allow you to input the relevant parameters, click on a button, and get a result for many kinds of "advanced" sample size problems without the effort and difficulties of writing a simulation or finding an analytic solution on your own. These programs are not, as far as I know, available as freeware, and they are actually pretty expensive, definitely not a worthwhile investment for people who don't use them frequently.

But if simulating isn't something you are up to, my advice would be to find a statistician in your environment who has this kind of "high octane" sample size software and ask him or her to do this for you. Most academic or corporate departments that do research, particularly research that is externally funded (for which sample size analyses are always required), have at least one statistician who can do this. There are also some Statalist members who have this kind of software and might do it for you--perhaps one of them is following along and will jump in.

[jameskb101]

Hi Clyde, Joseph

Just a quick note to say thanks for the very pragmatic advice. Simulation is a good idea, that I will explore. Clyde, your advice is sensible and reassuring in many ways that I've not missed anything obvious. We are well supported by statisticians at my university, so will follow-up with them.

Thanks,

James