AK 2024 Primary — xgboost

Likely-voter model for AK 2024 PRIM, training target election held Aug 20, 2024.

Test ROC AUC

0.9232

Brier

0.0750

Log loss

0.2483

Positive rate (training)

19%

Eligible voters: 577107

Training sample: 461685

Test sample: 115422

Methodology

Likely-voter model for 2024 PRIM, fitted as an XGBoost gradient-boosted classifier on voter-history and demographic features from the AK Division of Elections statewide voter file. Trained on n=461,685 eligible voters (target positive rate 18.7%); test ROC AUC 0.9232; Brier 0.0750. Most-important features (gain): voted_2022_primary (0.559), ever_voted_primary (0.140), voted_2022_general (0.094), n_prior_elections (0.044), voted_2016_general (0.030), voted_2020_primary (0.027), years_since_last_vote (0.024), n_prior_primaries (0.019).

Feature importance (XGBoost gain)

Rank	Feature	gain
1	voted_2022_primary	0.5585
2	ever_voted_primary	0.1396
3	voted_2022_general	0.0938
4	n_prior_elections	0.0445
5	voted_2016_general	0.0297
6	voted_2020_primary	0.0270
7	years_since_last_vote	0.0238
8	n_prior_primaries	0.0187
9	share_mail	0.0161
10	n_prior_special	0.0102
11	n_prior_municipal	0.0045
12	years_registered	0.0033
13	party_U	0.0032
14	party_D	0.0030
15	share_absentee	0.0027
16	n_prior_generals	0.0023
17	age_at_election	0.0023
18	voted_2018_general	0.0019
19	party_R	0.0016
20	share_early	0.0015
21	voted_2018_primary	0.0015
22	years_at_district	0.0013
23	share_polls	0.0012
24	ever_voted_municipal	0.0011
25	ever_voted_general	0.0009
26	pra_flag_int	0.0009
27	gender_F	0.0008
28	party_OTHER	0.0008
29	share_questioned	0.0006
30	voted_2020_general	0.0006
31	party_L	0.0005
32	gender_M	0.0005
33	party_N	0.0005
34	party_G	0.0004
35	gender_U	0.0002

Calibration — predicted vs. actual turnout by decile

Decile (10 = highest LV)	n	Pred. p	Actual turnout
10	57711	0.887	0.904
9	57711	0.466	0.457
8	57711	0.237	0.235
7	57710	0.128	0.127
6	57711	0.072	0.071
5	57711	0.042	0.039
4	57710	0.021	0.020
3	57711	0.012	0.011
2	57711	0.007	0.006
1	57710	0.004	0.003

Eligible scored voters bucketed into 10 equal-size deciles by predicted probability. A well-calibrated model has predicted p ≈ actual turnout in every decile.

Hyperparameters

{'seed': 42, 'max_depth': 5, 'subsample': 0.9, 'test_size': 0.2, 'n_estimators': 400, 'learning_rate': 0.05, 'colsample_bytree': 0.8, 'early_stopping_rounds': 30}

Trained on AK Division of Elections statewide voter file snapshot AK DoE 2025-05 at 2026-05-23 05:00 UTC.