I have repeated cross-sectional data at the individual level and am attempting to estimate the effect of a policy (implemented at the state level and targeted at eligible==1 observations in my dataset) that, in many states, turns on and off multiple times during my study window. Currently, I am estimating a simple TWFE regression with the below structure:
...to estimate the impact of the policy (1.treatment##1.eligible).
Given that many states in my data have multiple times in which their treatment status changes (as opposed to just one), I know that I don't have a true staggered 2x2 DiD design, and my understanding is that this precludes use of something like csdid (Callaway-Sant'Anna) because I cannot define treatment cohorts. That said, I was hoping to implement something in the spirit of the many new DiD estimators that have been developed in recent years: that is, a more sophisticated estimation method that addresses the potential biases of TWFE models like the one above. I have used the Callaway-Sant'Anna approach (and other advanced DiD designs, like the Wing et al. stacked sub-experiment method) previously for other projects using data in which jurisdictions only change treatment status once, but I am naive to situations in which this is not the case (like my current predicament) -- and know that there are some resident experts on these kinds of causal methods on this forum! I have not yet been successful in identifying a promising approach.
So, I would greatly appreciate any thoughts/guidance on what my options might be! (Unfortunately, the nature of my data does not feasibly allow for restricting the study window to a time period in which each state might only have one change in treatment status, or anything of that sort.) Thanks in advance.
Code:
reg outcome i.treatment##i.eligible i.month i.state i.control1 i.control2 c.control3, vce(cluster state)
Given that many states in my data have multiple times in which their treatment status changes (as opposed to just one), I know that I don't have a true staggered 2x2 DiD design, and my understanding is that this precludes use of something like csdid (Callaway-Sant'Anna) because I cannot define treatment cohorts. That said, I was hoping to implement something in the spirit of the many new DiD estimators that have been developed in recent years: that is, a more sophisticated estimation method that addresses the potential biases of TWFE models like the one above. I have used the Callaway-Sant'Anna approach (and other advanced DiD designs, like the Wing et al. stacked sub-experiment method) previously for other projects using data in which jurisdictions only change treatment status once, but I am naive to situations in which this is not the case (like my current predicament) -- and know that there are some resident experts on these kinds of causal methods on this forum! I have not yet been successful in identifying a promising approach.
So, I would greatly appreciate any thoughts/guidance on what my options might be! (Unfortunately, the nature of my data does not feasibly allow for restricting the study window to a time period in which each state might only have one change in treatment status, or anything of that sort.) Thanks in advance.
