TX vs NV deep-dive memo — correction memo

SMU Corporate Governance Initiative · The Reincorporation Tracker · v3.84-rev5p · 2026-04-30

What replicated cleanly

The auditor confirmed the equal-weighted arithmetic and the basic non-parametric tests:

Final TX/NV sample: 44 firms (19 TX / 25 NV) after excluding GLTO, NL, FORA (non-TX/NV destinations from the broader 47-firm deep-dive file). Confirmed.
TX equal-weighted mean Day-0 AR: −0.9905%. Confirmed.
NV equal-weighted mean Day-0 AR: +0.4378%. Confirmed.
Raw TX−NV difference: −1.4283 pp. Confirmed.
Welch t-test p-value: 0.0942. Confirmed.
Permutation-test p-value: 0.1086 (memo reported 0.1112 — close, within Monte-Carlo error). Confirmed.
Controlled-subsample comparison p-value: 0.2565. Confirmed.
TX total market cap: $2,220,693M. NV total market cap: $159,445M. Confirmed.

What did not replicate, with corrections

1. Value-weighted "Welch" — mislabeled inference, vastly different p-value

The original memo reported a value-weighted Welch test with p = 0.0006 supporting strong inference. The auditor showed that under a defensible WLS-of-AR-on-TX-dummy with market-cap weights:

Nonrobust p-value: 0.0046 (still significant, but ~7× the memo's 0.0006).
HC3-robust p-value: 0.2336 (no longer significant).

Correction. The value-weighted result is specification-fragile. We will not present it as a "Welch" test going forward — Welch's t-test does not natively accommodate market-cap weighting, and the memo conflated two estimators. Under HC3-robust WLS, the value-weighted difference is not statistically significant.

2. Propensity-score matching — does not replicate from the supplied input file

The memo's PSM result was −2.80 pp matched-pair difference. The auditor, running a standard logit on log market cap and sector with the supplied tx_vs_nv_regression_input.csv, obtained −3.68 pp. The paired test was still significant under the auditor's spec, but the magnitude does not match.

Correction. PSM is implementation-sensitive. We will not describe the PSM result as the "strongest evidence" until we publish: the exact logit specification, encoding of categorical sector dummies, matching algorithm (greedy / optimal / 1:1 / 1:k), caliper width, replacement rule, ties-breaking rule, bootstrap seed, and the matched-pair export. v3.85 will include the full replication artifact.

3. TSLA coding inconsistencies

(a) Sign error. The memo at one point referenced "TSLA with a large negative AR of +0.28%". This is internally inconsistent. Per the v3.84-rev5p database, TSLA's Day-0 AR is +0.28% — a small positive return, not a large negative one.

(b) Controlled-status definition violation. The memo's stated rule was that "controlled" requires >30% voting power. TSLA is coded controlled = 1 with 20.3% (Elon Musk per SEC 13G). 20.3% does not satisfy the stated >30% rule. Either (i) the rule should be restated to include known-effective-control under the 13G/13D characterization, or (ii) TSLA should be recoded controlled = 0 (which alters the controlled-subsample test).

Correction. v3.85 will publish a written controller-status protocol — what evidence we accept (13G filings, 13D filings, dual-class economic/voting splits, voting agreements, family blocks) — and apply it consistently across all 44 firms.

4. TX market-cap concentration claim

The memo stated TSLA + XOM constitute 99.2% of TX market cap. The auditor recomputed this share at 95.6%. The directional claim survives (TX market cap is dominated by two firms), but the precise figure was overstated.

5. TOST equivalence test — interpretation reversed

The memo interpreted TOST p = 0.2484 as evidence of practical equivalence between TX and NV. This is wrong. A TOST p-value of 0.2484 means we fail to reject the null of inequivalence within the ±2.0 pp band — it does not establish equivalence.

Correct sentence. "The raw TX–NV difference is not statistically significant, but the data also do not prove practical equivalence within a ±2.0 pp band. With N = 44, the TOST is underpowered to support an equivalence claim."

6. Controller-status sample coverage

Only 16 of 44 firms (36%) have verified controller status; 28 of 44 (64%) are unknown. The memo's headline "destination matters less than governance type" is not supportable with this coverage. The headline is preserved only as a hypothesis to be tested in v3.85 once the EDGAR controller-status audit is complete.

Required sensitivity table (v3.85)

The auditor's required sensitivity-table specification:

Specification	Description
Baseline	All 44 firms, equal-weighted TX vs NV mean comparison
Excl. TSLA	Drop TSLA (largest TX firm; +0.28% AR)
Excl. XOM	Drop XOM (second-largest TX firm; ≈ −0.71% headline AR)
Excl. TSLA + XOM	Drop both anchors
Excl. DDOG	Drop largest NV firm
Winsorized at 1%	Winsorize Day-0 AR at 1st / 99th percentiles
Winsorized at 5%	Winsorize at 5th / 95th
Verified-controller subsample only	n = 16
Non-controlled subsample only	n unknown until controller audit completes

For each row, report: equal-weighted mean (TX / NV / diff), value-weighted (with HC3-robust SE), permutation p, OLS p (HC3-robust), N.

Revised headline

In the equal-weighted TX/NV announcement-day sample (n = 44), Texas firms show a lower average Day-0 abnormal return than Nevada firms (TX −0.99% vs. NV +0.44%, difference −1.43 pp). The difference is not statistically significant at conventional levels (Welch p ≈ 0.094; permutation p ≈ 0.11). Some size- and matching-adjusted specifications produce a larger negative TX coefficient, but those results are sensitive to weighting, matching assumptions, sector encoding, and incomplete controller-status coding (28 of 44 firms unknown). The data do not yet support a causal claim that destination state, in isolation, drives announcement-day pricing. State choice appears endogenous to controller-ownership structure: in the verified-controller subsample, NV firms are nearly all controller-led and TX firms are nearly all not. Disentangling state choice from controller dynamics requires the v3.85 EDGAR controller-status audit and a pre-registered propensity-score matching protocol.

What this changes on the public site

The cohort event-study page (cohort_event_study.html) now links to this correction memo from the JF-referee-findings card.
The original TX_VS_NV_DEEP_DIVE_MEMO.md is preserved in the repository for transparency, prefaced with this disposition.
The headline copy on the cohort page is reverted to the underpowered-null framing; the value-weighted "p = 0.0006" claim is removed.
v3.85 will ship the full sensitivity table, revised PSM with documented design, and the EDGAR controller-status audit for the 28 unknown firms.

Acknowledgments

This correction was made possible by the external auditor whose detailed replication caught the errors, by the Grok+Harper+Benjamin+Lucas internal team whose v3.84-rev5p announcement-date audit ran in parallel, and by the JF/JFE-style internal hostile referee read.