1
$\begingroup$

According to my understanding about the difference-in-differences (DID) model with fixed effects, there are two specifications

  1. y=a0 + a1*TREAT + a2*POST + a3*TREAT_POST + e
  2. y=a0 + a1*TREAT_POST + timeFE + individualFE + e

Here, the dependent variable is a count variable and TREAT is an indicator variable that represents multiple groups of individuals affected by an event/treatment. The treatment affects groups of individuals at different calendar years, so yrsfromtreatment is a dummy variable for year pre- and post-treatment.

TREAT represent 1 for treated individuals (regardless of group) and 0 for matched control individuals. POST represent 1 for all years after the treatment is introduced. TREAT_POST=1 for the treated individuals in the post-treatment period.

So, Model 2 is better if there are possible omitted time-invariant and time-specific variables. And TREAT and POST indicator variables should usually be dropped in Model 2.

However, I find that I can actually run a fixed-effect negative binomial regression with calendar year (i.year) and treatment year (i.yrsfromtreatment) dummies in Stata:

xtset panelid yrsfromtreatment
xtnbreg y i.TREAT##i.POST i.yrsfromtreatment i.year, fe robust

Stata reports positive signs and significance in TREAT and POST.

How do I explain the coefficients on TREAT and POST when fixed effects are included? Is this model incorrect?

$\endgroup$
1
$\begingroup$

I'm a bit confused. Shouldn't your regression be collinear?

Your command is: xtnbreg y i.TREAT##i.POST i.yrsfromtreatment i.year, fe robust

Let:

  • $I_{i,t,\tau}$ be a variable indicating whether individual $i$ is $\tau$ years from treatment at time $t$.
  • $\mathrm{Treat}_i$ is an indicator as to whether $i$ at any point receives treatment.
  • $\mathrm{Post}_{it}$ is an indicator as to whether $i$ has received treatment at time $t$ or in prior periods.

It looks that command is running the regression: $$ y_{it} = b \mathrm{Treat}_{i}\mathrm{Post}_{it}+ \sum_{\tau=0}^\bar{\tau} c_\tau I_{i,t,\tau }+ u_i + v_t + \epsilon_{it} $$

What confuses me is that shouldn't this be collinear? If $\tau$ takes discrete values $0$, $1$, $2$, etc... then shouldn't

$$ \mathrm{Treat}_{i}\mathrm{Post}_{it} = \sum_{\tau = 0}^\bar{\tau} I_{i,t,\tau}$$ That is, if individual $i$ had treatment prior to or in year $t$ (i.e. $\mathrm{Treat}_{i}\mathrm{Post}_{it} = 1$), then they were either treated this year (i.e. $I_{i,t,0}$ = 1), or they were treated last year (i.e. $I_{i,t,1} = 1$), or treated the year before that ($I_{i,t,2} = 1$), etc....

$\endgroup$
  • $\begingroup$ Yes, I think the regression with panelid and year fixed effects needs to be: xtnbreg y i.TREAT##i.POST i.year, fe robust $\endgroup$ – EconoQ Mar 25 '17 at 17:08

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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