Timeline for Fixed effects in differences-in-differences
Current License: CC BY-SA 4.0
25 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Oct 18, 2022 at 2:56 | history | edited | Thomas Bilach | CC BY-SA 4.0 |
Edited the equation. Other minor textual edits.
|
| Apr 7, 2022 at 6:31 | history | edited | Thomas Bilach | CC BY-SA 4.0 |
edited body
|
| Aug 4, 2021 at 13:20 | vote | accept | Daniel Pinto | ||
| Aug 4, 2021 at 3:29 | comment | added | Thomas Bilach | If the data has a nested structure (i.e., $i$ within $s$), then I supposed this makes sense: $y_{it} = \beta D_{it} + \alpha_{i} + \nu_{st} + \epsilon_{it}$, where the policy varies at the $i$ level (e.g., firm, industry, county, etc.). Here, the state-year effect (i.e., $\nu_{st}$) doesn't absorb $D_{it}$. As indicated earlier, a state times year effect would eliminate your state-year policy variable(s) (i.e., $D_{st}$). | |
| Jun 2, 2020 at 16:10 | history | edited | Thomas Bilach | CC BY-SA 4.0 |
Edited the text/equation.
|
| Apr 10, 2020 at 9:57 | comment | added | Daniel Pinto | Thanks for the reply Tom. Indeed, I agree that state-by-year FE would be collinear with the treatment dummy. Would there by any reason to prefer state FE and year FE as opposed to state FE and industry-by-year FE (assuming degrees of freedom is not an issue, though)? | |
| Apr 8, 2020 at 18:24 | history | edited | Thomas Bilach | CC BY-SA 4.0 |
Further explication to understand why this is important.
|
| Apr 8, 2020 at 15:52 | history | edited | Thomas Bilach | CC BY-SA 4.0 |
Further explication to understand why this is important.
|
| Apr 8, 2020 at 2:25 | history | edited | Thomas Bilach | CC BY-SA 4.0 |
Edited the text.
|
| Apr 7, 2020 at 15:16 | history | edited | Thomas Bilach | CC BY-SA 4.0 |
Further explication to understand why this is important.
|
| Apr 7, 2020 at 15:07 | comment | added | Thomas Bilach | I will adjust my answer to reflect these concerns. However, I am not sure how you would obtain a DD estimate with a single 'state-year' fixed effect and a treatment variable that only varies at the 'state-year' level. I don't think you are wrong. I think you want to make comparisons across individuals/firms $i$. Is this fair to say? With a single 'state-year' fixed effect, I am not sure what two-group/two period comparisons we would be making. | |
| Apr 7, 2020 at 14:58 | history | edited | Thomas Bilach | CC BY-SA 4.0 |
Further explication to understand why this is important.
|
| Apr 7, 2020 at 14:23 | history | edited | Thomas Bilach | CC BY-SA 4.0 |
Further explication to understand why this is important.
|
| Apr 7, 2020 at 10:41 | comment | added | Daniel Pinto | When you include a state fixed effect, you control for all time-invariant state-specific unobservables. When you include state-year fixed effects, you also control for all time invariant state-specific unobservables as well as for state-specific unobservables that vary year-on-year. Why would this be wrong? Same idea if we think in terms of year versus industry-year FE. Why only control for common shocks to all firms within an year, if we can also control for industry-level shocks within the year? | |
| Apr 7, 2020 at 10:37 | comment | added | Daniel Pinto | My point is why would one use state and year FE separately, when one can tighten the identification further by using state-year FE? | |
| Apr 7, 2020 at 10:37 | comment | added | Daniel Pinto | Hi Tom. Yes, I do agree with the fake dataset. By state times year FE I mean indeed " 'state-year' effects (separate dummies for NY-2019, NY-2020, CA-2018, etc.)". I also agree that "And it will not return the same estimate of $\beta$ if estimated with separate state and time effects.". That is not my point. My point is that you are still getting a diff-in-diff estimate either way. The estimates change for the same reason that including firm-level control variables would change estimates: we get rid of some omitted variable bias. | |
| Apr 7, 2020 at 2:52 | history | edited | Thomas Bilach | CC BY-SA 4.0 |
added 450 characters in body
|
| Apr 6, 2020 at 16:04 | comment | added | Thomas Bilach | See my updated answer. Do we agree on the fake dataset I created? See the five variables appended to the end of the data frame. Would these be the 'state-year' effects (separate dummies for NY-2019, NY-2020, CA-2018, etc.) you are referring to in your question? Depending upon which 'state-year' we leave out, only five would enter the model. | |
| Apr 6, 2020 at 16:03 | history | edited | Thomas Bilach | CC BY-SA 4.0 |
Further explication to understand why this is important.
|
| Apr 6, 2020 at 15:39 | history | edited | Thomas Bilach | CC BY-SA 4.0 |
Further explication to understand why this is important.
|
| Apr 6, 2020 at 14:03 | comment | added | Thomas Bilach | Thank you for the follow-up. Hopefully we can find some common ground. Just so I’m not visualizing this improperly, you stated “state times year” in your previous comment. You are not referring to an interaction, correct? You mean separate dummies for each state-year (e.g., NY-2018) with multiple $i$ embedded within state $s$. And you’re saying the state-year dummies will absorb the individual state and year fixed effects in a ‘generalized’ DD equation? | |
| Apr 6, 2020 at 7:07 | comment | added | Daniel Pinto | So my question: why would this be a problem? I do not understand your statement: "I don't see how a single 'state-year' fixed effect would return the same estimate of 𝛽 β (i.e., DD coefficient) without the accompanying state and time effects.". As far as I understand it, the state and the year fixed effects separately would be collinear with state times year fixed effects, i.e., they would control for exactly the same thing (and more). | |
| Apr 6, 2020 at 7:04 | comment | added | Daniel Pinto | Now, a state fixed effect simply absorbs all within-state time-invariant common variation across states. This means we eliminate differences in levels across states. By using state-by-year fixed effects we do exactly the same thing but allow for time variation across different years: but we still eliminate the cross-state levels differences: which is the idea of the fixed effects. By the same token, I do not see the problem of using industry times year fixed effects instead of year fixed effects: the former will absorb all that the latter absorbs and a bit more. | |
| Apr 6, 2020 at 7:01 | comment | added | Daniel Pinto | Thanks Tom. That's. under the assumption. that you have 100 observations, one observation per state. Quite often, however, you have a treatment shock at group level (e.g. state) and many observations $i$. For instance, using your example, if you have 10 states, 10 years, and 100 firms per state-year, you have 10000 observations. In that case, there is no problem with degrees of freedom. Rest of reply below | |
| Apr 4, 2020 at 23:58 | history | answered | Thomas Bilach | CC BY-SA 4.0 |