Track A — incumbents. When a state changed its corporate law, did it move the firms already incorporated there? Track B — movers. Did reincorporating to Texas or Nevada move the stock? The grid below is the whole study; green = executed on real data, amber = pending.
An event study asks: on the day news hit, did the stock move more than it should have? The part of the move that the market and known risk factors cannot explain is the abnormal return — the market's verdict on the news.1
Every stock normally moves with the market. We first learn each firm's "normal" relationship to the market over a long quiet period before the event. Then on event day we compare what the stock actually did to what that normal relationship predicted. The gap is the reaction to the news.
A student usually scores about 5 points above the class average. This week the class averaged 80, so we predict ~85. She scores 92. The +7 is her "abnormal" score — the part not explained by how the class did. An event study does exactly this for a stock instead of a student.
Market model fit on the estimation window:
\[ R_{i,t}=\alpha_i+\beta_i' F_t+\varepsilon_{i,t} \]
Abnormal return on event day \(t\):
\[ AR_{i,t}=R_{i,t}-\big(\hat\alpha_i+\hat\beta_i' F_t\big) \]
Cumulated over a window and averaged across \(N\) firms:
\[ CAAR=\tfrac1N\sum_{i}\sum_{t\in W} AR_{i,t} \]
Our benchmark \(F_t\) is the Fama–French five-factor model plus momentum (FF5+UMD)2 — it strips out market, size, value, profitability, investment, and momentum, so what's left is closer to the news effect and not a style tilt.
Small firms, cheap firms, and profitable firms tend to move together for reasons that have nothing to do with corporate law. The factor model subtracts those common patterns so we don't mistake "this is a small-cap day" for "this is the law's effect."
Before saying a drug raised someone's heart rate, a doctor adjusts for age, caffeine, and exercise. FF5+UMD is that adjustment for a stock: subtract the known drivers, then see if anything is left.
\[ R_{i,t}-R_{f,t}=\alpha_i+\beta_{mkt}\,\mathrm{MKT}_t+\beta_{smb}\mathrm{SMB}_t \] \[ +\beta_{hml}\mathrm{HML}_t+\beta_{rmw}\mathrm{RMW}_t+\beta_{cma}\mathrm{CMA}_t+\beta_{umd}\mathrm{UMD}_t+\varepsilon_{i,t} \]
We report the test statistic robustly. Because event-day returns are fat-tailed and firms share a single event date, standard t-tests overstate significance; we report standardized and cross-sectionally adjusted statistics alongside the parametric tests.3
| Statistic | What it guards against |
|---|---|
| Patell Z (1976) | standardizes each firm by its own estimation-window noise |
| Boehmer–Musumeci–Poulsen (1991) | event-induced volatility (variance jumps on news days) |
| Kolari–Pynnönen (2010) | cross-correlation — firms sharing one event date are not independent |
| Corrado rank / sign (1989) | non-normal, outlier-heavy return distributions |
The interactive panel below shows the estimation window (used to learn each firm's normal behaviour) and the event window (the days the news effect is read), with no event-day data leaking into the baseline.
tristate_results_canonical.json (build v3, as of 2026-06-19) · BUILD_MANUAL §152.149–§152.164. Hover for values; drag/scroll where enabled; ⤢ for fullscreen inside the panel.Returns are CRSP CIZ daily (delisting-adjusted) through 2025-12-31, extended to 2026-06-18 on the reviewer-built canonical panel; fundamentals from WRDS wave-2 (130 parquet files, 1.68 GB); factors from Fama–French.4 Every excluded firm carries a named structural reason — there is no black-box dropout.5
A controller-protective statute should bite hardest on controlled firms — those with a founder or insider holding a large voting block. The Delaware reform's bright line is ≥ 33⅓% of the voting power.6
Our first run produced a candidate estimate: founder-controlled Delaware firms fell more than dispersed firms around the Senate-passage window (p≈), in the magnitude zone of Khoo & Tallarita's −1.4%.7 We then ran the standard robustness battery on it. The estimate did not survive seven independent diagnostics — each is reported below, because an estimate whose magnitude or significance is not invariant to the robustness battery is itself reported as a finding, consistent with the project's pre-registration.
The same collapse is shown dynamically below: as each inference layer is added, the candidate coefficient and its confidence band migrate toward zero.
tristate_results_canonical.json (build v3, as of 2026-06-19) · BUILD_MANUAL §152.149–§152.164. Interactive: hover for values, drag/scroll where enabled, ⤢ for fullscreen inside the panel.tristate_results_canonical.json (build v3, as of 2026-06-19) · BUILD_MANUAL §152.149–§152.164. Interactive: hover for values, drag/scroll where enabled, ⤢ for fullscreen inside the panel.tristate_results_canonical.json (build v3, as of 2026-06-19) · BUILD_MANUAL §152.149–§152.164. Interactive: hover for values, drag/scroll where enabled, ⤢ for fullscreen inside the panel.tristate_results_canonical.json (build v3, as of 2026-06-19) · BUILD_MANUAL §152.149–§152.164. Interactive: hover for values, drag/scroll where enabled, ⤢ for fullscreen inside the panel.tristate_results_canonical.json (build v3, as of 2026-06-19) · BUILD_MANUAL §152.149–§152.164. Interactive: hover for values, drag/scroll where enabled, ⤢ for fullscreen inside the panel.tristate_results_canonical.json (build v3, as of 2026-06-19) · BUILD_MANUAL §152.149–§152.164. Interactive: hover for values, drag/scroll where enabled, ⤢ for fullscreen inside the panel.tristate_results_canonical.json (build v3, as of 2026-06-19) · BUILD_MANUAL §152.149–§152.164. Interactive: hover for values, drag/scroll where enabled, ⤢ for fullscreen inside the panel.We labelled a firm "controlled" using the wrong number — total insider ownership, not the largest single holder's votes. Fix the column and the candidate estimate becomes , p≈.50: nothing.
Calling a company "family-controlled" because all employees together own 35% — when no single family member owns more than 10%. That is a different thing entirely.
Wrong: \( \mathbb{1}[\textstyle\sum_j \text{insider}_j \ge 33.3\%] \). Right: \( \mathbb{1}[\max_j \text{voting}_j \ge 33.3\%] \). The two constructions differ for many firms; the wrong column produced false-positive Delaware "controlled" flags and no false negatives.
We tried many windows, events, and cuts. With ~42 tries, one will look significant by chance. Correct for that and the p-value rises to 0.36.
Flip 42 fair coins ten times each; one will show 8 heads and "look" biased. That coin is not special — you just looked 42 times.
Romano–Wolf step-down max-\(t\) over the family \(\{\text{events}\times\text{windows}\times\text{cuts}\}\): \( p_{RW}=0.36 \). Bonferroni \(\approx\) .74–1.0; Benjamini–Hochberg \(q\approx\).07.8
We re-ran the test on random dates where nothing happened — and the same −1.6% to −1.8% gap appeared. The controlled firms were also already falling ~6% before the window opened. So it reads as a pre-existing trend, not a reaction to the law.
If patients on a placebo "improve" as much as patients on the drug, the drug is not doing the work. A no-news date is the placebo for an event study.
Pre-event run-up \(CAAR_{[-20,-3]}=-6.08\%\). Difference-in-differences nets it out: event-window increment \(=+3.68\%\) (t=1.31). Placebo dates: −1.75% (p=.07), −1.60% (p=.10).
Remove the single most extreme firm and significance is gone (p=.066). The result is carried by five small stocks that fell 20–28% — for reasons unrelated to a Senate committee vote.
A drug trial where the "benefit" disappears if you drop one patient is not a finding — it is that patient's story.
Jackknife: \( \hat\delta_{(-i)} \) range gives \(p\in[.019,.066]\); 10 of 84 single-firm deletions flip significance.
All firms react on the same day, so their "abnormal returns" are correlated — there are not 84 independent observations. Account for it and p rises to ~0.07.
Surveying 84 people who all watched the same advertisement together is not 84 independent opinions — it is closer to one room's reaction.
Effective N with mean correlation \(\bar r\): \( N_{\text{eff}}=N/[1+(N-1)\bar r] \); Kolari–Pynnönen adjusted \( p=.072 \).
As risk factors are added, the effect shrinks step by step (−2.94% → −2.11%). A genuine governance effect would not depend on which risk model you use; a style tilt does.
If a "diet effect" shrinks every time you also control for exercise, sleep, and age, much of it was never the diet.
\( \hat\delta \): CAPM −2.94% → FF3 −2.94% → FF5 −2.28% → FF5+UMD −2.11%. Monotone decay = benchmark-dependent = not invariant.
The journal-standard fix for a few extreme returns is to cap the top and bottom 1%. Do that and the corrected result stays null at every level; the candidate result decays further.
If one lottery winner makes the "average income" of a town look rich, you cap the extremes before averaging. Same idea.
Winsorize \( CAR \) at \([1,99]\): corrected −0.89% (p=.46); candidate wrong-column −1.85% (p=.048) → −1.32% (p=.114) at \([2.5,97.5]\).9
We ran the full endogeneity battery — none of it resurrects an effect.10
| Method | What it corrects | Result |
|---|---|---|
| Propensity-score matching (1:3 NN) | observable confounding | −2.10%, p=.072 (significance lost) |
| Entropy balancing | covariate imbalance (no exclusion needed) | marginal (p≈.05–.07) |
| Heckman two-step | self-selection into "controlled" | collapses: coef t=−0.17, p=.87 |
| Instrumental variables (2SLS) | endogenous incorporation choice | inadmissible (weak F=2.64; exclusion violated) |
If the candidate estimate (or K-T's −1.4%) were a real revaluation, controlled firms should have kept underperforming. With the panel extended through 2026-06-18 we held the portfolio over the full ~14–16 months after the event and tested two ways, with daily returns winsorized 1/99.11
tristate_results_canonical.json (build v3, as of 2026-06-19) · BUILD_MANUAL §152.149–§152.164. Hover for values; drag/scroll where enabled; ⤢ for fullscreen inside the panel.Over the full year-plus, did the controlled firms stay down, recover, or bounce? They reverted — the gap closed. To the data's edge (June 18, 2026, ~16 months) the cumulative gap is , not significant.
A patient dips after a treatment, then recovers to baseline over the year. The dip was transient, not a permanent decline.
Calendar-time hedge \(\alpha\): \[ R^{hedge}_t=\alpha+\beta'F_t+\varepsilon_t \] Through the full window to 2026-06-18 (334 td): unconditional \( \bar R^{hedge}=+0.005\%/\text{day} \), \(t_{NW}=0.14\), \(p=.89\), cum +0.9%; CAPM \(\hat\alpha=+0.003\%/\text{day}\), \(p=.93\). FF5+UMD \(\hat\alpha\) (\(p=.82\), cum +2.8%) is factor-bounded at 2026-04-30. BHAR = +3.1% (median −17.7%, 95% CI [−21.8, +34.8]). Returns winsorized 1/99.
On the dates S.B. 29 actually moved through the legislature, nothing happened. The only negative window (April 3) sits on the "Liberation Day" tariff crash and is concentrated in nine oil & gas names — remove them and it vanishes.12
tristate_results_canonical.json (build v3, as of 2026-06-19) · BUILD_MANUAL §152.149–§152.164. Hover for values; drag/scroll where enabled; ⤢ for fullscreen inside the panel.A "no detectable reaction" finding only means something if the test could have found a reaction. With Texas firms, the smallest effect we could reliably detect is ~2%. So we can say "no large reaction," not "the law is proven irrelevant."
A negative result is meaningful only if the test is sensitive enough at the stage you are testing. A weak test giving a negative tells you little. The Texas leg is a weak test.
Minimum detectable effect: \[ MDE=(z_{1-\alpha/2}+z_{power})\cdot \sigma_{CAR}/\sqrt N \approx 2\% \] Equivalence (TOST) against ±50 bp does not clear — the band we can support is ~±2%, not ±0.5%.13
We held the Texas-incumbent cohort from the S.B. 29 signing forward and asked whether any abnormal return accumulated over 1, 3, 6, and 12 months — a calendar-time portfolio (factor-model alpha) and buy-and-hold abnormal returns vs the market.13
For the firms whose reincorporation we have locked to an as-of-event date, we measure buy-and-hold abnormal returns vs the market at 1, 3, 6, and 12 months from each firm's own move date — the move is the event, not the legislative date. The mover track is reported separately from incumbents and is small, so read it as directional.15
tristate_results_canonical.json (build v3, as of 2026-06-19) · BUILD_MANUAL §152.149–§152.164. Hover for values; drag/scroll where enabled; ⤢ for fullscreen inside the panel.Destination by firm count: the DExit wave is overwhelmingly to Nevada, then Texas, with a small set going to other states and a counterflow back to Delaware. The flow diagram below maps each destination cohort to its current result status.
tristate_results_canonical.json (build v3, as of 2026-06-19) · BUILD_MANUAL §152.149–§152.164. Hover for values; drag/scroll where enabled; ⤢ for fullscreen inside the panel.Thirty-five tests across seven families. Status is reported as executed: nothing is marked run before it was.
Setting aside the retracted number, one structural fact stands: on the corrected single-holder ≥33⅓% measure, controller density runs Nevada (≈13.8%) > Delaware (≈2.6%) > Texas (0%). Texas's cohort is dominated by dispersed banks and royalty trusts, with no controlled firm at the bright line; Delaware sits in the middle; Nevada — the most controller-protective regime — is the densest.14
tristate_results_canonical.json (build v3, as of 2026-06-19) · BUILD_MANUAL §152.149–§152.164. Hover for values; drag/scroll where enabled; ⤢ for fullscreen inside the panel.The forest plot collects each leg's headline estimate with its status badge; the governance matrix records, per cohort, the legislative event and the per-leg result. Every value binds to the canonical dataset.
ff5_umd_daily_2020_2026.csv.v3plus_wrds_data_2026-06-19, 130 parquet, 1.68 GB). BUILD_MANUAL §152.144, §152.154–155.07_REDTEAM/redteam_grid.py.06_OUTPUTS/winsorization_sensitivity_2026-06-19.md; BUILD_MANUAL §152.166. Convention: 1/99 cross-sectional, disclosed ex ante, robustness reported across levels.07_REDTEAM/extrev_b_selection.py; BUILD_MANUAL §152.164.daily_returns_panel_2026-06-20_v3.parquet (CRSP-canonical →2025-12-31, Compustat-canonical 2026-01..06-18), keyed on gvkey, returns winsorized 1/99. The FF5+UMD factor-model alpha is bounded at 2026-04-30 by Ken-French daily factor availability; the unconditional long-short hedge and a panel-EW-market CAPM alpha extend to the data edge 2026-06-18 (cum +0.9%, p=.89/.93), as does the BHAR. BUILD_MANUAL §152.171, §152.173 (supersedes the ~10.4-month CRSP-only result, §152.167).tristate_results_canonical.json (single source). Provenance: BUILD_MANUAL §152.136–§152.167. SMU Cox · Corporate Governance Initiative · Cox & Dedman Law. Empirical research only — not legal, investment, or tax advice. All interpretations are the author's own and do not represent SMU. Working draft v3 — not for external distribution until the voting-power rebuild + as-of-event lock land.