Error when `treated` variable dtype is integer

[drbenvincent]

Known problem for regression discontinuity, possibly for other experiments...

When the treatment column data is integer (0/1) then we get an error, it currently only works when the dtype is boolean

[jpreszler]

I've looked into this and think there are two ways to proceed, but first, what's happening.

When treatment is a numeric column, patsy treats it different than something that's boolean or categorical so the design_info being used isn't what's expected at the end of the experiment.

This would be an issue with PrePostNEGD except for a small formula difference, which is why I think there's two good solutions:

1. The user just needs to tell patsy the variable should be treated categorically "y ~ 1 + x + C(treated) + x:treated" results in no error for the regression discontinuity with binary treated values. We should probably point this out in the documentation.
2. In _input_validation, we convert any 0/1 "treated" or "group" columns into booleans and then the experiments will run as expected regardless of the formula differences ("y ~ 1 + x + C(treated) + x:treated" and "y ~ 1 + x + treated + x:treated" would be the 'same'). The code change for this is simple, but more opinionated and may lead to confusion since we're technically not using the given formula in some cases.

Since we're already using the formula to handle binary data in PrePostNEGD, I'm a bit more in favor of 1 but would happily implement either if anyone has strong opinions in favor of 2.