How to calculate sample size for median and interquartile range ?

[Daniel Jost]

Dear all
My Stata is SE.12.0
I search the command to calculate sample size in case of comparison between two median values (and IQR) . Thanks a lot .

[Nick Cox]

That's not an answerable problem without extra information on the distribution expected.

[Svend Juul]

I don't think we should consider sample size estimation an exact science. Here is a quick-and-dirty advice:

I assume that you will use the ranksum command for the comparison. Try to express your assumptions approximately in terms of means and standard deviations and use sampsi (you use Stata 12 and don't have access to the new power command). Now, since the requirements for parametric testing are hardly fulfilled (if they were, you would use a t-test), add 50%.

[Clyde Schechter]

Another, more complicated approach is to run a series of power simulations with varying sample sizes. Whether this is worth the effort depends on why you want the sample size estimate in the first place. If you're just in need of a ballpark estimate to see if some idea is feasible/credible, Svend Juul's suggestion is quite sufficient. If you are talking about committing hefty amounts of resources to launching a study, then you may need the greater precision that power simulation will bring.

[Rebecka Hess]

Does anyone have a paper I can quote for adding 50% to sample sizes calculated with parametric assumptions in order to estimate the sample size required under non-parametric conditions? Thanks.

[Nick Cox]

That sounds on all fours with how much slower will someone go in the mountains compared with the lowlands? Depends on the someone and on the mountains and the lowlands (and on the mode of transport!).

More directly put, I don't know what you mean in general by "parametric assumptions" on "non-parametric conditions". Even if we sharpen that up to (for example) "procedures for which normal distributions are ideal" and "non-normal distributions", there can't be anything more than a rule of thumb plucked out of the air. This sounds like a parody of bad textbooks from the 1950s and the 1960s that said something like this: if you have a normal distribution, you can do such and such; otherwise, you must use non-parametric tests. Hardly true at the time and quite amazingly wrong now in most respects.

Even if you find a reference stating this, it shouldn't be believed as meaning anything worth quoting. Sorry if I seem to pile it on, but any fault almost certainly lies in some bad teaching or textbook or advice you have been given.

There could be much longer answers than this, but they would fill entire books! One great book is Rupert Miller, Beyond ANOVA, Chapman and Hall 1986.

Also, you might well get a much better answer with a more specific question.