1
$\begingroup$

Borusyak, 2021 has a sentence

While our baseline setting is for panel data with two-way fixed effects, we show how our formal results extend naturally in a number of ways. We allow for more general panel-type specifications that can include, for instance, unit-specific trends or time-varying covariates

It is quite contradictory to my understanding. Because to me, when you consider "two-way fixed effects", you also control for unit and period fixed effects. Why the author used the word "more general panel-type specifications that can include, for instance, unit-specific trends or time-varying covariates" here. I mean, the "more general panel-type specifications" as he described is "two-way fixed effects", not "more general"

Improve this question
$\endgroup$

1 Answer 1

Reset to default
3
$\begingroup$

A two-way FE model is:

$y_{it} = \beta_0 +\beta_1 x_{it} +\gamma_i + \delta_t +u_{it}$

The $\delta_t$ absorb average time effects, but there may still be unit-specific trends, which could be modelled with:

$y_{it} = \beta_0 +\beta_1 x_{it} +\alpha_i t +\gamma_i + \delta_t +u_{it}$

There is not a violation of multicollinearity and it is more general than before. If instead we had a general time trend, rather than unit specific, it would violate multicollinearity.

$y_{it} = \beta_0 +\beta_1 x_{it} +\alpha t +\gamma_i + \delta_t +u_{it}$

Also, there has been some literature that has cautioned against the use of individual-specific time trends because such can absorb part of a heterogeneous treatment effect. Meer and West 2016.

Improve this answer
$\endgroup$
6
  • $\begingroup$ Thank @Micheal , can I ask an example of controlling for "unit-specific trends" is unit's lag variable? $\endgroup$
    – Phil Nguyen
    Commented Sep 16, 2021 at 9:44
  • 1
    $\begingroup$ "general time trend" would be perfect colinear with time fixed effect, is it what you mean then? Thanks $\endgroup$
    – Phil Nguyen
    Commented Sep 16, 2021 at 9:46
  • $\begingroup$ Lag variable is different than time trends. A lag variable might be as follows: $y_{it} = \beta_0 +\beta_1 x_{it} + \beta_2 Z_{it-1} +\gamma_i +\delta_t +u_{it} $Z$ is used generically. $Z$ could be a lagged $X$ variable, or instead a lagged outcome variable. There are issues with lagged outcome variables, and I'm guessing that's not relevant to your case. $\endgroup$
    – Michael Gmeiner
    Commented Sep 16, 2021 at 9:59
  • 1
    $\begingroup$ ""general time trend" would be perfect colinear with time fixed effect, is it what you mean then?" Yes. $\endgroup$
    – Michael Gmeiner
    Commented Sep 16, 2021 at 9:59
  • 1
    $\begingroup$ The second equation in my answer does this. The coefficient, $\alpha_i$, differs for each individual, measuring an unit-specific time trend. $\endgroup$
    – Michael Gmeiner
    Commented Sep 16, 2021 at 10:15

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.