Minimum stata sample size in stratification?

[Statzor]

Hi,

I'm designing a methodology to track the effects of a price test at work, and I'm curious to know whether there are limits on how many strata I can use in a sample or on how many subjects need to be in each strata.

For context, we're taking about 80-120 products (the exact number will depend on what we wind up having good data for) and applying a new pricing algorithm to one half of them, selected at random. Then, we'll apply a differences-in-differences model to determine whether and to what degree our test group was more profitable than our control.

We have a wealth of information on these products' past performance and I'd like to use it to stratify our sample and shrink our confidence intervals a bit. In my imagination, I could stratify these products on a number of dimensions until I have strata containing no more than 2 to 6 products.

I never see such low-level stratification used in studies that I've reviewed, however, so I ask, is there any statistical drawback to creating such small strata, or are other conditions just preventing other researchers from being so thorough?

Thanks so much!

 

[terzi]

Hi Statzor,

There is no minimum number for strata. Actually, as a curious fact, a simple random sample can be regarded as a stratified sample with only one strata.

The only issue could be the number of observations per strata, but again, that will depend on the analysis you are planning. Particularly, some type of models may become problematic with small samples, but since you are talking about 2 to 6 strata I don't think that would be problematic. If you wish, you can share more details about your methodology once you get them for us to aid.