Dear Statalisters,
I'm estimating a Poisson DiD model. I have about 100k individuals over 5 states and over a 20 year period. I have a battery of outcome variables that I'm individually testing, all of which are count variables with a high frequency of zeros. My treatment corresponds with laws that changed in those states at different times and I'm interested in the causal effect of those changes on the behavior of individuals (as measured by my various outcome variables). We should expect that individual behavior is systematically different between states because of differing regulatory environments, but more homogeneous within a given state. My question is, should I use state fixed effects or individual fixed effects to control for unobservables? I can't use both simultaneously because they are nested (Individuals do not change states during the sample period).
Individual FEs intuitively seem the most robust in terms of omitted variable bias, but including FEs for 100k individuals diminishes my degrees of freedom substantially more than including FEs for 5 states. Also, using individual FEs cuts my usable sample size in half because many individuals have no variation over time (i.e., "0" in every year) and are thus dropped. For each of my outcome variables I get less significant results when using individual FEs rather than state FEs, but the treatment effect is similar in terms of direction and magnitude. My sense is that this is due to the reduced statistical power of the individual FE approach and not due to unobservable variables that are captured by the individual FEs but not captured by the state FEs. Is there an appropriate method for testing this intuition?
Thanks in advance
-Brent
I'm estimating a Poisson DiD model. I have about 100k individuals over 5 states and over a 20 year period. I have a battery of outcome variables that I'm individually testing, all of which are count variables with a high frequency of zeros. My treatment corresponds with laws that changed in those states at different times and I'm interested in the causal effect of those changes on the behavior of individuals (as measured by my various outcome variables). We should expect that individual behavior is systematically different between states because of differing regulatory environments, but more homogeneous within a given state. My question is, should I use state fixed effects or individual fixed effects to control for unobservables? I can't use both simultaneously because they are nested (Individuals do not change states during the sample period).
Individual FEs intuitively seem the most robust in terms of omitted variable bias, but including FEs for 100k individuals diminishes my degrees of freedom substantially more than including FEs for 5 states. Also, using individual FEs cuts my usable sample size in half because many individuals have no variation over time (i.e., "0" in every year) and are thus dropped. For each of my outcome variables I get less significant results when using individual FEs rather than state FEs, but the treatment effect is similar in terms of direction and magnitude. My sense is that this is due to the reduced statistical power of the individual FE approach and not due to unobservable variables that are captured by the individual FEs but not captured by the state FEs. Is there an appropriate method for testing this intuition?
Thanks in advance
-Brent
Comment