County Level Matched Pairs Treatment

#1
I usually try to write answers up here, but this time I'm asking the question. Hopefully, you all can go easy on me.

I am working on a project where our team is trying to implement a county-level treatment and measure its effects. They have chosen to do a matched pair analysis where they control for different things like race, income, gender, etc. The treatment is an educational program to reduce a particular potentially deadly habit. Our response variable is proportion being measured objectively through a survey.

I have concerns about the sample size which is about 20 counties (10 pairs) as well as whether or not the counties can truly be paired when they will clearly have differences in their demographic makeups. How would I determine whether these counties are suitable to be matched and if there is enough power to draw sound conclusions? Lastly, what is the final test we would run to determine the answer? A paired t-test? Any other advice on potential designs are welcome.
 

Karabiner

TS Contributor
#2
Do you measure the outcome on the individual level (something like: "do you engage in that particular activity yes/no"?),
If yes, then you could consider a multilevel approach, with respondents are clustered within counties. You can include
covariates on the individual level, as well as covariates on the county level. Admittedly, I do not know how matching is
carried out in such an approach, or whether it would be carried out at all. Maybe just a random assignment of counties
to treatments is sufficient (since, as mentioned before, you can include covariates such as race/income/gender/etc. ,
measured on the county level, to adjust for pre-treatment differences).

With kind regards

Karabiner
 
#3
Yes, so they will be surveyed again after a certain time has passed since the treatment. We planned to look at the difference in proportion though from there. This approach you're suggesting makes sense. Would you then input this through a logistic regression? I seem to be a little fuzzy on the last step of the analysis.
 
#5
Thank you for your replies! You have been helpful. One more question. Have you worked with studies where (in this case) we would give the treatment to say 2 counties and have 10 matched control counties for each of them? Say Counties A and B got the treatment, but counties A1-A10 and B1-B10 were controls? This almost seems like cheating, but I can't figure out why.