Alabama historical WAR methodology
Race-residual WAR for Alabama legislative elections from 1994 through 2022.
1. Estimand
Race WAR = raw gap − fitted structural expected gap
Democratic WAR = race WAR; Republican WAR = −race WAR
Every score is a race differential in two-party margin points. No pooled candidate coefficient, career average, fundraising, or ideology adjustment enters WAR.
2. Modern training model
The backcast uses the selected decaying_lag ridge specification with alpha 100. It is trained on 3,658 strict Southern races after 2016. Predictors are incumbency balance, ticket margin and its square, time and cycle indicators, state, chamber, ticket-office family, prior presidential margin, ticket change, and the ticket-change-by-years interaction.
3. Historical backcast
For the 407 races from 1994 through 2014, the modern fitted relationship is applied backward to the historical Alabama ticket, incumbency, chamber, and prior-presidential context. The model is not refit on those historical outcomes.
4. Published modern scores
The 97 races in 2018 and 2022 retain the exact published Alabama WAR v1 same-cycle residuals. The pooled modern-model prediction is retained only as a diagnostic.
5. Missingness and names
Missing prior-presidential context remains labeled. For compatibility with the selected modern design, unavailable numeric lag inputs receive its zero-valued model encoding; that is not treated as an observed zero.
Candidate display names come from election identity records. Evidence-backed manual adjudications replace malformed identifier-shaped source labels, while archived election pages provide a spelling cross-check. Finance provider names, committee names, and committee IDs cannot supply display identity.
6. Interpretation and limitations
- WAR is retrospective, not a forecast probability.
- A race residual cannot distinguish candidate strength from opponent weakness or omitted local conditions.
- Historical baselines, district allocation, and map vintages carry source uncertainty. Pre-2010 precinct-to-district allocation uses provisional legislative-activity splits and county fallbacks, not certified boundaries; the drawn district outlines are display geometry.
- Same-cycle ticket baselines for a few districts rest on thin or conflicting source coverage and moved materially when the warehouse identity repairs were applied on 2026-09-11: 2014 House 52 and 56 (federal contested coverage about 55% that cycle) and 2002 House 26 (conflicting precinct observation sets under review). Their residuals should be read as source-limited; see the release card in the downloads for run identifiers and the full list.
- Backcast scores keep the within-cycle ranking of a same-era fit (correlation above 0.9 in every cycle) but not its level: a descriptive fit on 1994–2014 Alabama alone sits about 20 points lower on average and the incumbency effect is roughly twice the modern estimate. Compare backcast WAR within a cycle, not against 2018 and 2022 levels; see the release card in the downloads.
- The 2018 and 2022 scores are the exact Alabama rows of the independently reviewed Southern WAR run recorded in the published manifest; the 1994–2014 rows are backcasts of that run and are flagged as such in every download.
- The public forecast carries part of each matched nominee's prior Alabama WAR forward and retains the fitted WAR structure and incumbency effect; historical WAR itself is retrospective.