csdid/did: significant IPW ATTs but event-study pretrend test p=0 with small cohorts

[Marco Di Luca]

Hi FernandoRios

Context
I’m studying whether receiving a Michelin star affects the opening of new business ventures associated with a restaurant.
Panel: restaurant-year, unbalanced (each restaurant from its opening year to present).

Vars: outcome = yearly count of new ventures (delta), event time = first star year (year_event, 0 if never), calendar time = year, unit id = cod_event. Covariates: province-year GDP and restaurant age (time-varying). I also have a 1:1 “never-treated” comparison set (restaurants never in the guide but otherwise similar).

What I see

• IPW estimates are statistically significant (both in Stata and R).
• But the pretrend test is extremely rejected: in Stata, estat pretrend χ² is huge with p = 0.000; in R, the “p-value for pre-test of parallel trends” in summary(att_gt) is 0.
• The data are small/unbalanced with very small g×t cells for some cohorts (e.g., some year_event have 1–2 observations in early years). Stata reports “Units always treated found… Panel is not balanced… Will use observations with Pair balanced” and the R output flags small groups and very wide simultaneous bands.
• If I bin cohorts (e.g., five adoption windows instead of 2000…2025), the problem eases and results are stable; but I’d prefer not to bin except as a robustness check.
• In R, when I switch to est_method="dr", I get The regression design matrix for post-treatment is singular, which looks like collinearity/lack of variation in covariates within some post samples.

If I plot with dynamic aggregation (event study) I can see no pre-trend: not significant and all close to 0.

Any hint or advise would me much appreciated!

(if you need more info I can provide all the context and results)

Thanks