Timeline for Causal inference method for analyzing randomized control trial with covariates / pre intervention observations
Current License: CC BY-SA 4.0
12 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Aug 15, 2020 at 12:00 | history | tweeted | twitter.com/StackStats/status/1294604766526357504 | ||
| Aug 9, 2020 at 20:55 | answer | added | mc51 | timeline score: 0 | |
| Aug 7, 2020 at 9:01 | comment | added | dimitriy | The centered version can produce something like that as well. I would say centering is not strictly necessary, and if I was forced to do something like this, I would try to just use some fixed value. I guess centering does allow for easier comparisons with the other models. | |
| Aug 7, 2020 at 8:53 | comment | added | dimitriy | @mc51 Thanks, that does makes sense. I think I have an odd preference to keep them separate. It strikes me as strange to add heterogeneous treatment effects to the model, but then effectively evaluate the derivative at the mean score. I would rather plot a marginal treatment effects curve as a function or evaluate the derivative at some meaningful points. | |
| Aug 7, 2020 at 7:56 | comment | added | mc51 |
@DimitriyV.Masterov: Because of the interaction term, the coefficient for test depends on the value of pre: test = 13.5859 + (-0.9985 * pre). Pre has a range of values in the data. For each of those, the coefficient for test would be different. By centering pre (mean=0) we get the average treatment effect as a coefficient for test.
|
|
| Aug 7, 2020 at 6:03 | comment | added | dimitriy | The best approach depends on how correlated the outcomes are across time. There a nice paper by David McKenzie, with lots of practical advice. | |
| Aug 7, 2020 at 5:52 | comment | added | dimitriy | Could you explain why you need to center here to get marginal effects? | |
| Aug 6, 2020 at 22:22 | history | edited | mc51 | CC BY-SA 4.0 |
Added paper reference
|
| Aug 6, 2020 at 21:20 | history | edited | mc51 | CC BY-SA 4.0 |
print df.sample
|
| Aug 6, 2020 at 20:40 | comment | added | Noah | Just an FYI, when you include the interaction you need to center the covariate at the sample mean to be able to interpret the treatment effect as a marginal treatment effect. This equivalent to estimating the average marginal effect (which the other estimators do automatically). You should also be using robust standard errors for all of these if that is an option to protect against heteroscedasticity. Eager to see what others have to say about the problem itself, though. | |
| Aug 6, 2020 at 19:42 | history | edited | mc51 | CC BY-SA 4.0 |
fixed image
|
| Aug 6, 2020 at 19:36 | history | asked | mc51 | CC BY-SA 4.0 |