Different results from did package vs. "manual" diff-and-diff

[skuchukh]

Hello did community,

My analysis looks at the impact of a policy on HIV incidence in a cohort of women. I have person-time for each woman's exposure to treatment. I have only one exposure time point (year 2023). I am conducting a propensity score weighted poission regression. Oddly, the results I get from did are qualitatively different from "manual" diff and diff.

Here is my manual did:

didreg_adj <- glm( HIV ~ treated + pre_post + pre_post* treated + log (pt), data = data, weights = weight, family = poisson(link = "log"))
Here pre_post is a dummy for before or after 2023 and treated is a dummy for ever-treated vs. never treated. pt is person--time.

Here is my did using the package

resa <- att_gt(
  yname = "anc_retest_pos", tname = "year", idname = "area_id", gname = "first_exposure",
  xformla = ~ pt, weights = "weight", allow_unbalanced_panel = F,
  bstrap = TRUE,
  data = dt2 , est_method = "reg", control_group = "nevertreated"
)

first_exposure is 2023 for everyone. year is the time variable. pt is log of person-time. unfortunately changing the est_method to dr does not make much of a difference.

Thank you in advance for your help!

best
Salome

[pedrohcgs]

Hi Salome, You “manual DiD” is not really a standard DiD as it is not constructed using the same parallel trends assumption. Please check the paper by Jeff Wooldridge at the Econometrics Journal on non-linear DiD estimators. Thanks

…

---------------------------------------------------- *Pedro H. C. Sant'Anna* *https://psantanna.com <https://psantanna.com>*

---------------------------------------------------- Warning: This email may contain confidential or privileged information intended only for the use of the individual or entity to whom it is addressed. If you are not the intended recipient, please understand that any disclosure, copying, distribution, or use of the contents of this email is strictly prohibited. ----------------------------------------------------

On Fri, May 30, 2025 at 08:52 skuchukh ***@***.***> wrote: Hello did community, My analysis looks at the impact of a policy on HIV incidence in a cohort of women. I have person-time for each woman's exposure to treatment. I have only one exposure time point (year 2023). I am conducting a propensity score weighted poission regression. Oddly, the results I get from did are qualitatively different from "manual" diff and diff. Here is my manual did: didreg_adj <- glm( HIV ~ treated + pre_post + pre_post* treated + log (pt), data = data, weights = weight, family = poisson(link = "log")) Here pre_post is a dummy for before or after 2023 and treated is a dummy for ever-treated vs. never treated. pt is person--time. Here is my did using the package resa <- att_gt( yname = "anc_retest_pos", tname = "year", idname = "area_id", gname = "first_exposure", xformla = ~ pt, weights = "weight", allow_unbalanced_panel = F, bstrap = TRUE, data = dt2 , est_method = "reg", control_group = "nevertreated" ) first_exposure is 2023 for everyone. year is the time variable. pt is log of person-time. unfortunately changing the est_method to dr does not make much of a difference. Thank you in advance for your help! best Salome — Reply to this email directly, view it on GitHub <#227>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABE73443WRXF2XNIUEHKIST3BBIAHAVCNFSM6AAAAAB6H77IDGVHI2DSMVQWIX3LMV43ERDJONRXK43TNFXW4OZYGM4TGMRXGM> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[skuchukh]

Dr. Sant'Anna

thank you for your reply! I looked at the Wooldridge paper which suggests that when we have non-linear models (e.g. mine - poisson) and small number of time periods (N= 2 in my case), fixed effects strategies (my "manual-did" above) do not generally produce reliable estimators of parameters or partial effects.

Does the did package account for Woolridge's flexible strategy for non-linear outcomes? I thought did is also based on OLS suggesting that both of the above approaches would generate misleading inference.
Are there other packages I should consider instead? For example, using fixed-effects Poisson pseudo-maximum likelihood estimator (PPML) via fixest (fepois).

Thank you so much once again!
Salome

[skuchukh]

https://grantmcdermott.com/etwfe/ would be correct application, in case this question comes up again.