Estimates from Day-1 doesn't show in plot

[d0nghyunkang]

Hey, I've found another peculiar issue in the resulting plot. The command doesn't show the estimate from Day-1. For example, if I run the following code from your example_csv, the resulting plot omits the estimate from Day-1.

esplot paygrade, by(male) event(to_male_mgr, save) window(-20 30) estimate_reference

[dballaelliott]

Hi,

Thanks for highlighting this, this is actually by design!

The dynamic treatment effects (or event-study coefficients are only identified up to a constant). By convention, these coefficients are often normalized to 0 in the time t = -1. For example, that's what is done in the figures in Cullen & Perez-Truglia, 2020. Of course, the normalization is arbitrary, and e.g. Chetty, Friedman & Rockoff, 2014 normalize with respect to other years (-3 and +2).

I know that moving the reference period is a feature that has been requested in the past, and I'll flag that as a new feature to add (i.e. changing the omitted period to t=0 or t = -2).

There is currently an undocumented option called recenter that may be of interest to you. It is a workaround for the under-identification of the full set of dynamic effects (i.e. the coefficients need to be relative to something). What happens is the package will estimate the normalized event study with t= -1 set to 0 as usual. Then, it will estimate the mean of the outcome variable in t = -1 and shift the whole series by that amount. That is, instead of setting t=-1 to be zero, it sets it to the mean of the outcome in t=-1. To highlight that we've mechanically set t=-1 to it's mean, confidence intervals around that point are still ommitted.

[d0nghyunkang]

Thanks for the detailed explanation. I learned something!

[dballaelliott]

Great! Glad this was helpful!

I think this is a common source of confusion, so I'll leave this issue open until I can add this information to the package docs