{
  "file": "data/turnout_model/gg26/election_levels.csv",
  "description": "Statewide election x party level alpha (logit scale) from the structure model; alpha_se from vcov(fit).",
  "generated_by": "scripts/turnout_model/02_fit.R",
  "run_date": "2026-10-17T17:07:09-0400",
  "sources": {
    "training": "data/turnout_model/inputs/training.csv",
    "covariates": "data/turnout_model/inputs/covariates.csv"
  },
  "source_ids": [],
  "checks_passed": [
    "training.csv has 432 rows",
    "covariates.csv has one row per county",
    "360 rows after dropping closed-primary rows",
    "0 <= ballots_cast <= active in every row",
    "every training row has covariates",
    "GG18 and GG22 fitted ballots sum by party to within 0.1% of observed",
    "invlogit(alpha_hat + s_hat + u_obs) reproduces fitted rates in every midterm cell",
    "midterm fitted rates within 1 point of observed in every cell",
    "draw means of s_cp within 0.02 of the point estimates",
    "structure_draws has 72,000 rows",
    "15 election x party levels (no OTH in primaries)",
    "bootMer statistic matches the point estimate on the fit",
    "every bootMer replicate is finite",
    "backtest refit without GG22 is non-singular and converges without messages",
    "backtest GG22 naive structure effects are finite",
    "backtest GG22 calibrated totals match actual statewide ballots",
    "backtest refit without GG18 is non-singular and converges without messages",
    "backtest GG18 naive structure effects are finite",
    "backtest GG18 calibrated totals match actual statewide ballots"
  ],
  "notes": "Model: cbind(ballots_cast, active - ballots_cast) ~ 0 + ep + z_share_ba_plus + z_median_age + z_share_moved_past_year + z_share_owner_occupied + (1 | county:party_group) + (1 | county:election_type) + (1 | obs) (binomial glmer, bobyqa). s_cp = covariate part + u_county:party_group + u_county:election_type at midterm_general; (1 | county) is not in the selected formula (its variance was estimated at 0). See fit_summary.json.",
  "caveat": "Not for use uncalibrated: the alpha draws and the random-effect draws behind s_cp are independent, so alpha + s_cp carries the uncertainty of the population mean (statewide GG22 ballots sd about 42,000), and the parametric bootMer (use.u = FALSE) agrees with that width for the same reason rather than validating it. The forecast calibrates the statewide level per draw and does not use the alpha spread."
}
