Timeline for Reciprocal Causation in Panel Data
Current License: CC BY-SA 3.0
9 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Sep 16, 2015 at 17:24 | comment | added | generic_user | Huh? I don't do a lot of time series, but I think that when you control for a lag DV in AR1, the autocorrelation clears up. Irrespective of that however, you'll want to use heteroskedasticity and autocorrelation-consistent standard errors (i.e.: robust or cluster-robust) standard errors, clustered at the level of your spatial unit. If you are afriad of spatial autocorrelation, you might need to account for such spillovers as well. Cluster-robust SE's are routine, spatial stuff is less so. | |
| Sep 16, 2015 at 16:31 | comment | added | Matteo | Thanks. Tedious follow up question: my data is also autocorrelated (according to Woolridge test [xtserial in Stata]) which I believe means that my independent variables might be biased downwards if I include a lagged dependent variable. Is this a concern? If so what are the available options? | |
| Sep 15, 2015 at 13:56 | vote | accept | Matteo | ||
| Sep 15, 2015 at 12:50 | comment | added | generic_user | Also, I'm not sure that this is a case where a OLS approximation of a non-negative integer process yields bad bias in marginal effects, though I could be missing something. | |
| Sep 15, 2015 at 12:48 | comment | added | generic_user | The mechanism determining the intensity of the stops. It is likely not entirely random, and we're supposing that it is correlated with prior crime. What else is it correlated with? If it is correlated with anything else that is also correlated with present crime, then you've got an omitted variable problem that isn't solved by lagging the DV. | |
| Sep 15, 2015 at 11:35 | comment | added | Matteo | - also sorry but could you clarify what you mean by "selection into stops"? Obviously the main driver of stops is police strategy/policy changes but it is those effects that I am interested in? | |
| Sep 15, 2015 at 11:29 | comment | added | Matteo | Thanks, very helpful! I had read that with large T Nickell bias shouldn't be a problem but wasn't sure if that also applied to nonlinear (negative binomial) models? If so then that seems like the simplest solution | |
| Sep 15, 2015 at 11:28 | history | edited | generic_user | CC BY-SA 3.0 |
added 39 characters in body
|
| Sep 15, 2015 at 11:22 | history | answered | generic_user | CC BY-SA 3.0 |