t-test or z-test for two proportions?

We did an intervention to reduce the frequency of use of a certain type of treatment and now I want to compare the data pre vs post intervention.

For each month I have the number of treatments per 1000 patient visits (because the monthly number of visits fluctuates). I have 6 months of pre-intervention data and 6 months of post-intervention data. Sample sizes are about 800 visits per month.

My first question: Is it better to
A. do a t-test comparing the means of the monthly rate per 1000 patient visits, or
B. take the raw data and calculate the rate per 1000 patient visits for the entire pre-intervention period, then do the same for the post-intervention period, and use the z-test of two proportions?

My second question: There are actually 3 subtypes of treatment. If I want to say "subtype A accounted for 50% of the total treatments given in the pre-intervention period but only accounted for 20% in the post-intervention period" what test would be best?



Less is more. Stay pure. Stay poor.
It a little more complicated then the approaches you listed, but interrupted time series is the proper approach. You have rolling admissions and patients are not matched, correct. Also, do you have patients in the pre that are also in the post group?


Less is more. Stay pure. Stay poor.
This would potentially mean those observations are not independent. Is the likelihood that they come back supposed to be affected by the intervention.

Not controlling for patients seen multiple times would mean they would be over-weighted. So if I was trying to figure out the average weight of a person at a gym and I just weighed everyone that comes through the door, I could get a biased estimate of the clients. That is a person may come to the gym twice a day. In addition their attendance bias the estimate since they are counted twice. In addition, this person may also be leaner than others, so not only do I count them twice, I am counting an extreme (low) value twice. This could happen to you if they have greater comorbidities and are sicker. I ran a study awhile back and I had a patient with 50+ ED encounters in a 3-year period. This is a totally different patient than most of them and they are contributing a lot of data.