[Zhixiao Yao]

Dear all,

I am trying to understand the “Years” row in Table 1 of Nagengast, Rios-Avila, and Yotov (2025).
Their sample period is 1980–2016, but in their main table Table 1 they report 34 years, not 37. So 3 years are missing.

When I replicate their estimation method (same dataset, different treatment variable), the last three years (2014–2016) drop out entirely from my analysis.

Initially, I thought this was because, in the ETWFE setup, the last few years after treatment for the latest cohorts might have no untreated observations left to serve as controls. But I checked my data and also there data, and there are still untreated dyads in those years. So that explanation does not seem to apply.

In my Stata output, all treatment-cohort × year interactions involving those last 3 years are shown as (empty) and the years themselves are not counted in the Years row of the summary table.

My questions:

1. Why would these years drop out if untreated observations are still present?
2. Is there something in the ETWFE/jwdid estimation routine (or the way the event-time dummies are constructed) that would cause the last few years to be omitted even when controls exist?

For reference:

• Screenshot from Table 1 of Nagengast et al. showing “Years = 34” (1980–2016 sample).
• Screenshot from my own ETWFE regression output showing (empty) for all interactions involving 2014–2016.

Thank you for any clarification!
Zhixiao Yao

[FernandoRios]

This is very specific to your data. and you do not provide enough information to figure out what could be happening. For instance, could you do something like
tab year year_treated

that typically shows if there is data or not for the analysis

[Zhixiao Yao]

This is very specific to your data. and you do not provide enough information to figure out what could be happening. For instance, could you do something like
tab year year_treated

that typically shows if there is data or not for the analysis

Hi FernandoRios ,

Thanks for your suggestion earlier. I’m trying to understand whether the “Years = 34” in Table 1 of Nagengast et al. could indeed be due to no untreated observations left to serve as controls in the last few years. I’ve been looking into this for quite a while but I’m still not fully sure.

I have now created the tables tab year year_treated (attached) for my dataset.

From my table, I see that my cohort == 0 group (never treated) still has more than 2,000 observations in 2014–2016, so I’m struggling to see why these years would drop entirely in ETWFE. ChatGPT suggested some possible reasons related to event-time matching and fixed effects structure, but I’m not convinced.

Could you help clarify how to correctly interpret this table and how one can tell from it that there are no untreated observations left to serve as controls? Any clarification would be greatly appreciated.

Thanks a lot!

Zhixiao

[FernandoRios]

next less obvious points. while less evident in JWDID setting, you need enough variation to use coefficients. for instance, because some cells =1, you cant add any controls to your model.
the other point, what about other controls?
you could run a simple regression with your model specification and fixed effects, keep the sample, then re do this table to see how it looks.

[Meng Zhang]

Dear Professor Rios-Avila,

I am currently working with jwdid and would greatly appreciate your advice on two questions:
1. How can I save the estimates of the fixed effects?
2.I noticed that when I use only never-treated groups as the control group, I can obtain pre-treatment estimates. However, when I use both never-treated and not-yet-treated groups as controls, I only get results for post-treatment periods. Could you please clarify why this happens?
Thank you very much for your time and guidance.

Best regards,
Meng Zhang