Variable dropped because it predicts success perfectly

[Bill Austin]

Dear Stata users,

I found the topic mentioned in my subject lined to be covered on other threads, but to my best knowledge none of the previous discussions led to an answer that I think is suitable to my scenario.

I ran an RCT where individuals were allocated to a treatment or to a control group. I know want to measure whether being part of the treatment group increased the probability of an individual to take a certain action. As such, as I want to run the following command on Stata: "probit Action Treatment"

where both variables "Action" and "Treatment" are dummy binary variables taking the value of 1 if the individual took action and 1 if she/he was in the treatment group. This is the message that Stata gives me:

I understand this may be due to a number of issues, but I really need to measure whether my treatment was effective at achieving the intended outcome or not! I tried forcing it by adding "asis" but then the variables becoming insignificant (which can't be because I see from a graph that treatment did better than control by a very large margin). Adding other control variables also doesn't fix the problem (of course).

Any suggestions on how I can still measure the impact of treatment in my setting?

Many thanks,

Bill

[Clyde Schechter]

Well, put in other words, "Treatment != 0 predicts success perfectly" means that 100% of those in the Treatment = 1 group had Action = 1. If you run -tab Action Treatment- you will see that. Since it appears you are not looking to adjust for any other variables, you don't need to do a probit model anyway. Just show the results of -tab Action Treatment-. If you feel you need a p-value, then add the -chi2- or -exact- option to that command.

[Bill Austin]

Thanks! Super helpful.