2
$\begingroup$

I am doing a difference-in-difference analysis of an event that affected several states in the US. I am interested to understand the effects of this event on state-level unemployment rates. I have state level data on demographics for several years before and after this event. My question is what covariates should I include if I am estimating equation of the following form:

$$ \mbox{unempRate}_{jt} = \gamma_j + \alpha_t + \beta D_{jt} + \delta X_{jt} + \epsilon_{jt} $$

where $\gamma_j$ are state-fixed effects; $\alpha_t$ are time fixed effects; $D_{jt}$ are dummies - 1 if state $j$ is affected at time $t$, and 0 otherwise; and $X_{jt}$ are time-varying state level covariates. Coefficient of interest is $\beta$.

So far, I have included population and average household income at the state level.

What more covariates can I include? How does one decide which covariates to include in a setting like this? What is the guiding philosophy?

$\endgroup$
11
  • $\begingroup$ Covariates should be pre-determined $\endgroup$
    – Papayapap
    Feb 2, 2023 at 23:00
  • $\begingroup$ @Papayapap, could you clarify with some details? How does one 'pre-determine'? $\endgroup$
    – funcard
    Feb 3, 2023 at 2:00
  • 1
    $\begingroup$ They should not be affected by the treatment, either measured before or invariant. $\endgroup$
    – Papayapap
    Feb 3, 2023 at 9:05
  • $\begingroup$ Average household income could easily be argued to be affected by the treatment under the assumption that the treatment has a non-zero effect on the unemployment rate. However, variables describing the quality of the labour force could be a suggestion. Perhaps, something like average educational level, shares of education groups, degree of urbanization/labour market access variables (area of land per citizens with a road density higher than x ... or something like that). Immigrants can have a hard time achieving full labour market integration so share of certain type of immigrants. $\endgroup$ Feb 3, 2023 at 23:48
  • $\begingroup$ @JesperHybel, thanks. What's the general philosophy that guides the choice of a particular covariate? Trying to find a reference on this question, but no luck so far. $\endgroup$
    – funcard
    Feb 3, 2023 at 23:55

1 Answer 1

0
$\begingroup$

Understanding back-door criterion via Directed Acyclic Graphs (http://causality.cs.ucla.edu/blog/index.php/category/back-door-criterion/) helped me understand how to add covariates on my econometric model. While studying the material, I also found a wonderful website that tells which variables to condition on and how to test whether the DAG you draw is correct: http://www.dagitty.net/dags.html

$\endgroup$

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.