I am running a regression discontinuity design for a project in the early stages. I'm unable to share the data or printouts at this point, for which I apologize, but hopefully the general issue I'm discussing will help. I am pretty comfortable with the method, but am running into a situation that I think indicates a problem, although I can't find a clear discussion of this anywhere.
I'm using Nichols' "rd" command in Stata 13, which implements a local linear regression for the RDD.
I'm running the test using three different similar outcome measures. On one measure, the treatment has no effect. On another outcome the treatment has an effect, which is sizable but not dramatic- a coefficient of 3 for a measure that runs from 0 to 10. This effect is also consistent across different bandwidths.
On the third outcome, the effect is huge- positive 150+ coefficient for a variable that runs from 0 to 20. And it isn't consistent-this is the effect at one bandwidth setting. For other bandwidths, it has negative effects, some of which are also unrealistically large (but not to the same extent as the positive one).
Normally, when I run into this huge of an effect, I'm suspicious. If it was a logit, I'd expect some sort of separation issue. I'm not as familiar with the mechanics of the RDD to know if this is the case, but given the huge effect size with one specification and opposite effect with others I imagine there's some issue with my data.
Has anyone run into this, or would you know where I can look for more information?