WooldridgeDiD never-treated OLS drops arbitrary cohort-time cells instead of g-1 references

[shawcharles]

WooldridgeDiD(method="ols", control_group="never_treated") currently builds every treated-cohort × time indicator, absorbs unit and time fixed effects, and relies on generic rank-deficiency handling to remove the resulting collinear columns.

That does not enforce the ETWFE normalization. The dropped column for each treated cohort should be its last pre-treatment period, t = g - 1, but the QR path can remove different cells. This can suppress genuine post-treatment effects and changes the interpretation and value of the remaining coefficients.

I found this while validating the genuine Callaway-Sant'Anna mpdta data for #723.

Reproducer

import warnings

from diff_diff import WooldridgeDiD
from diff_diff.datasets import load_mpdta

mpdta = load_mpdta()

with warnings.catch_warnings(record=True) as caught:
    warnings.simplefilter("always")
    result = WooldridgeDiD(
        method="ols",
        control_group="never_treated",
    ).fit(
        mpdta,
        outcome="lemp",
        unit="countyreal",
        time="year",
        cohort="first_treat",
    )

for warning in caught:
    if "Rank-deficient design matrix" in str(warning.message):
        print(warning.message)

for key in [(2004, 2004), (2004, 2005), (2007, 2007)]:
    print(key, result.group_time_effects.get(key))

Observed on current main plus the source-verified loader branch:

Rank-deficient design matrix: dropping 3 of 15 columns
('g2004_t2004', 'g2006_t2005', 'g2007_t2005').

(2004, 2004) None
(2004, 2005) -0.059919911882184825
(2007, 2007) -0.057141530108885445

For cohorts 2004, 2006, and 2007, the intended reference periods are 2003, 2005, and 2006. Instead, the current solve drops 2004, 2005, and 2005 respectively; the first is a post-treatment cell and the third is not cohort 2007's g-1 reference.

An explicit full-rank dummy OLS on the same observations, omitting t = g - 1 for each treated cohort before solving, gives:

g2004_t2004  -0.010503246221103879
g2004_t2005  -0.07042315810285515
g2007_t2007  -0.026054410719146937

The reference implementation makes this normalization explicit for the never-treated control group by setting .Dtreat = (time != cohort - 1) before forming the interactions:
https://github.com/grantmcdermott/etwfe/blob/2b47cfdf61b90140c8372574eece4ad6d43d37f9/R/etwfe.R#L399-L403

Would you accept a focused PR with this boundary?

Proposed scope

• For OLS with control_group="never_treated", omit each treated cohort's g-1 interaction deterministically before rank handling.
• Keep all other identified pre- and post-treatment cells and preserve the documented group_time_effects mapping.
• Apply the same interaction basis across within-transformed and full-dummy OLS paths, including bootstrap refits and covariate interactions.
• Add regression tests that verify:
  • each cohort's intended reference cell is absent and no post-treatment cell is dropped;
  • coefficients match an independently constructed full-rank dummy OLS on mpdta;
  • coefficient results are invariant to row and interaction-column ordering;
  • HC1 and full-dummy coefficient paths use the same normalization;
  • anticipation and unbalanced-panel reference semantics are explicit and tested.
• Add a brief changelog entry and update any docstring that currently implies all (g, t) cells are simultaneously identified.

Out of scope

• No change to control_group="not_yet_treated".
• No change to logit or Poisson interaction bases without separate methodology evidence.
• No change to generic solve_ols() rank-deficiency behaviour.
• No unrelated aggregation, inference, or result-object redesign.