Methodology, source verification & statistical audits
The data ingestion protocols, double-coding standards, inter-coder reliability metrics and statistical robustness checks behind US data center moratoriums, from the record. In an independent blind search, the register already held 36 of 43 benchmark instruments (blind search recall 84%; 95% CI 70%–92%), so every headline count in the report is a conservative floor.
1 · Corpus construction
Corpus construction & primary-source coding protocol
We built the register by ingesting and deduplicating three public trackers of state and municipal data center restrictions, yielding 1,415 candidate instruments mapped to US Census FIPS jurisdictions. Every merged or discarded record carries a logged rationale.
Third-party tracker lists serve only as initial leads. Each candidate was then checked against primary statutory text, adopted council minutes, official public notices or local journalism, and coded against a written codebook (version 1.3a) capturing instrument type, legal status, enactment and sunset dates, procedural scope, grandfathering provisions, vote tallies and source confidence tier. Where no such source could be found, the row stays labeled a tracker lead (440 of the 1,073 in force). Direct excerpts reproduce source text verbatim, with private individuals’ names replaced by [name] and roll-call votes shown as tallies.
Strict non-inference rule. Moraine never infers legislative status from calendar dates alone. A pause is never marked ended unless a source confirms its repeal, expiration or replacement; an ordinance whose sunset date has passed without a verifiable outcome is labeled past end date, no outcome on record.
Double-coding & inter-coder reliability (Cohen’s kappa)
To reduce single-coder subjectivity, 328 instruments were double-coded blind by two coders working from identical source packets, with disagreements resolved by an adjudicator (288 adjudicated; the other 40, from an earlier pilot round, were not adjudicated and are excluded from the confirmed count).
- Headline vs. confirmed counts. The headline figure (1,073) is every instrument on record as in force, across all source tiers. The confirmed subset (163) is limited to in-force instruments that were double-coded, adjudicated and rest on a government record (49), an official notice (17) or two independent news organizations (97). Non-binding advisory resolutions and uncertified ballot results never count. The statistical models below run on the full dataset, with sensitivity checks that drop tracker-only leads.
- Inter-coder agreement before adjudication (n = 331 blind pairs): Instrument type κ = 0.94; Status κ = 0.89; Which applications it covers κ = 0.79; Grandfathering κ = 0.74; Source confidence κ = 0.82; Adoption date 93% exact match; End date 79% exact match. Which applications it covers and Grandfathering fell below the κ ≥ 0.80 bar we set on first pass; disagreements went to adjudication, and no headline metric relies on those fields.
The source tiers, and how many instruments sit in each:
| Source tier | All instruments | In force |
|---|---|---|
| Double-coded; government record | 49 | 49 |
| Double-coded; official notice | 17 | 17 |
| Double-coded; two or more independent news sources | 97 | 97 |
| Double-coded; not headline-eligible | 165 | 88 |
| Government record; single-coded | 36 | 25 |
| News report; single-coded | 524 | 357 |
| Tracker lead only; not counted in any headline | 527 | 440 |
Spatial campus joins & pipeline filtering
Instrument boundaries are joined to the campus register behind our US data center report under two specifications: county-wide (every campus inside the county boundary) and unincorporated only (excluding incorporated municipalities inside a paused county).
To isolate entitlement exposure, we keep only pre-construction planned campuses (campus-register stages 2 to 5: an active utility power request, land-use filing, entitlement approval or building permit, before vertical construction). Because campuses under construction and operating generally hold vested rights, we exclude 10 campuses under construction and 343 operating facilities inside moratorium boundaries, along with 137 records with no stage on record and 10 withdrawn proposals. Multi-phase filings under the same operator and campus name are consolidated into one project.
Loudoun County filing review. Given Loudoun County’s concentration, all 87 of its raw filings were reviewed one by one against their filing records: 14 were dropped as work at an existing facility or as non-data-center uses, 8 were merged into parent campuses, and 16 were dropped as likely built (the newest approval predates 2016, or a building footprint is on record). 10 with ambiguous permitting status were kept. Power capacity is assigned without summing across bases, using a five-step order (requested utility service, contracted service, total facility load, critical IT load, maximum announced MW); it exists for 14 of the 98 exposed projects. The pre-registered K1 test below counted every campus, operating ones included.
Pre-registered hypotheses, deviations & methodological limits
- Reporting asymmetry. Local press covers contentious council votes and extensions far more reliably than quiet expirations, so enactment dates often reflect vote coverage rather than codification, and quiet expirations are under-reported. So that this bias cannot inflate the extension rate, the 24 expired instruments with no documented outcome are shown separately rather than pooled.
- Media clustering bias. Regional news cycles often survey neighboring counties at once, which clusters detection. The within-state permutation test below is the check on that.
The six tests, set before counting
- K1 Do moratoriums cover places with data centers? Rule set in advance: Below 15% of in-force local instruments containing a campus: lead with "most are pre-emptive" Result, Nov. 18, 2026: 15.1% (161 of 1,068) county-wide; 11.0% counting only unincorporated areas; 22.1% without tracker-only rows On the page: On the line. We report the range and do not lead with either reading.
- K2 Which lever covers more pipeline power, utilities or local governments? Rule set in advance: If utilities cover more, the thesis leads with utilities Result, Nov. 18, 2026: Utilities 6,100–11,869 MW vs local 2,428–2,714 MW (requested service) On the page: Triggered, but not followed: see deviation 1.
- K3 Do moratoriums end when scheduled? Rule set in advance: Below 50% ending on schedule: headline "temporary in name" Result, Nov. 18, 2026: 13.5% (95% CI 7.5%–23.1%), n = 74 On the page: Triggered. This is finding 1.
- K4 Does a neighbor’s adoption raise the odds of adopting? Rule set in advance: A CI excluding 1 permits a claim of spread; the placebo is the check Result, Nov. 18, 2026: OR 2.04 (95% CI 1.69–2.45); placebo percentile 97 On the page: Clustered, not shown to be contagious. No claim of spread is made.
- K5 Is the register complete? Rule set in advance: Below 85% recall: never describe the register as complete Result, Nov. 18, 2026: 83.7% (36 of 43 found by a blind search) On the page: Failed narrowly. The page does not describe the register as complete.
- K6 Is the register accurate? Rule set in advance: Below 95% agreement on a random 50: block publication Result, Nov. 18, 2026: First audit 92% (failed); re-audit 98% (49 of 50) On the page: Passed on the second draw, after register-wide fixes: see deviation 2.
Deviations from the pre-registration
- Thesis order (K2). The rule said to lead with utilities if utility actions cover more pipeline power than local instruments. They do on the requested-service basis, but the margin depends on two judgment calls: overlapping utility territories (the low and high ends of the range) and whether Pennsylvania's guidance and tentative order count, since no tariff is in force there. Without Pennsylvania the ranges overlap. We lead with K3, which holds under every reading, and present utilities third with the range shown.
- Accuracy scoring (K6). The pre-registered test was agreement with the primary record. The first random audit of 50 headline rows failed: 64% on primary records alone (16 rows had no primary record reachable), 92% on the best citable source. It found two classes of error: instruments that do not pause or prohibit anything (a study, a curative-amendment declaration, a resolution urging state action) coded as moratoriums, and moratoriums replaced by permanent rules coded as extended. Both were corrected across the whole register (codebook rules AT2 and LC4; 8 instruments recoded), and K6 was rescored as agreement with the best citable source, government record preferred. A fresh random 50, drawn after the fixes, agreed in 49 of 50 (98%).
2 · Statistical checks
Statistical & survival analysis checks
We put the core findings through two formal checks: a competing-risks survival model of moratorium duration, and a county-month logistic regression, with state and calendar-month fixed effects, testing whether counties copy neighboring jurisdictions.
Check 1: competing-risks survival analysis of sunset outcomes
Stops at day 365, where 28 instruments are still at risk; below 25 the estimate is too thin to draw. At day 365 (28 at risk): 53.0% extended, 5.6% ended on schedule.
Moraine register, Nov. 18, 2026; n = 376. Non-parametric Aalen-Johansen cumulative incidence, treating extension, permanent replacement and scheduled expiration as competing risks; instruments not yet at their sunset date are right-censored; 95% bootstrap intervals in the table. The curve stops once fewer than 25 instruments remain at risk (28 remain at day 365).
View data table
| Day | At risk | Extended | Replaced | Ended on schedule |
|---|---|---|---|---|
| 90 | 292 | 0.9% (0.0%–2.0%) | 0.6% (0.0%–1.6%) | 1.2% (0.0%–2.5%) |
| 180 | 125 | 2.0% (0.3%–4.0%) | 1.8% (0.0%–4.0%) | 3.1% (1.0%–6.0%) |
| 365 | 28 | 53.0% (41.1%–65.3%) | 5.7% (1.9%–10.4%) | 5.6% (1.7%–10.9%) |
| 1095 | 1 | 87.0% (79.5%–93.2%) | 5.7% (1.9%–10.4%) | 7.3% (2.7%–13.4%) |
Check 2: neighbor contagion vs. shared state political shocks
On a raw observational basis, counties bordering a jurisdiction that enacted a moratorium in the prior three months had 2.04 times the odds of adopting their own over the next quarter (95% CI 1.69–2.45; 1,085 instruments across 2025–2026). But when we randomly shuffle adoption dates across counties within the same state 200 times, the placebo model still produces an odds ratio of 1.77.As of Nov. 18, 2026; n = 638 county adoptions. Moraine Research, https://morainecre.com/resources/research/data-center-moratoria
Why does the placebo nearly match the observed clustering? Counties within a state share a legislature, a utility rate base, a tax incentive regime and a regional media market. Without tracker-only leads, the observed neighbor effect sits at only the 75th percentile of the shuffled distribution, so the data cannot distinguish county-to-county policy imitation from simultaneous reactions to statewide utility and political shocks.
Moraine register, Nov. 18, 2026. Panel of county-months from Jan. 2025 through Oct. 2026, first-order queen contiguity on 2023 Census TIGER county polygons; logistic model with state and calendar-month fixed effects; 200 permutations with the random seed fixed in advance.
View data table
| Set | Odds ratio (95% CI) | Shuffled mean (middle 95%) | Percentile | Adoptions |
|---|---|---|---|---|
| All instruments | 2.04 (1.69–2.45) | 1.77 (1.53–2.05) | 97 | 638 |
| Excluding tracker-only rows | 1.64 (1.31–2.05) | 1.55 (1.29–1.90) | 75 | 425 |
See the data center moratorium tracker map, or open a state page such as Texas or Virginia for its instruments, dates and sources.