Illustrative Veteran Retention Logit

Summary

This model is an illustrative weighted logistic regression that predicts a synthetic 12-month retention flag for post-9/11 veterans employed at the time of the ACS interview. It exists to demonstrate an end-to-end analytics workflow — feature engineering, stratified split, weighted fit, calibration — using entirely public, individually de-identified data.

It is not a production retention predictor. The target is generated, not observed, and the sample design is not corrected for. This card is deliberately written to make those limits legible.

Target

retention_12mo_synthetic is a 0 / 1 flag produced by the Phase 3 pipeline using a fixed-seed logistic rule over observable cohort features. It does not come from a longitudinal employer record. Retention rates in this project are expressible as modeled-illustrative, not measured. The overall base rate is 67.9% (PWGTP-weighted) among the employed working-age cohort.

Downstream documents that cite any retention-model output must carry the Modeled — Illustrative label, consistent with the project convention established in Phase 3.

Cohort and universe

• Post-9/11 veterans identified via ACS PUMS variables MIL / VPS / MLPA.
• Ages 22–64 at interview, civilian status only (active-duty ESR 4 / 5 excluded).
• Employed at interview only (ESR ∈ {1, 2}). Retention is undefined for non-employed rows.
• Survey years 2019, 2021, 2022, 2023 pooled. 2020 intentionally omitted — ACS published a smaller experimental file for 2020 and pooling with the standard 1-year files introduces weighting inconsistencies.
• N = 108,902 unweighted rows (11.9 million PWGTP-weighted person-years) after feature engineering. 33 rows (0.03%) dropped for missing values on required features.

Features (47 including intercept)

No interaction terms. No regularization. The modeling choice on this pass is transparency over accuracy.

Split

• 70 / 15 / 15 train / validation / test.
• Seed: 20260417 (fixed at Phase 4a modeling-prep kickoff).
• Stratified on DRAT × SOC major, the two strongest non-demographic moderators in Phase 4a EDA.
• Split counts: train 76,145 · val 16,257 · test 16,500.

Fit

• Estimator: statsmodels GLM, family Binomial, link logit, IRLS, max_iter=200.
• Weights: PWGTP passed as freq_weights. Every coefficient and log-likelihood reflects the weighted design.
• Convergence: yes, 45 degrees of freedom on the model.

Metrics (held-out sets, PWGTP-weighted)

AUC is modest — by design. The synthetic target is dominated by disability, education, and state labor-market conditions, which the model captures, but the residual variance is large relative to signal. A reader who wants headline discrimination should not read this model; a reader who wants to understand how observable cohort features relate to a structured retention label should.

Calibration (test set, PWGTP-weighted deciles)

Leading signals

Ranked by absolute coefficient, dropping the two small-n occupation dummies:

• Farming / fishing / forestry occupation (+0.33). Small group, large positive — reference occupation is Management.
• Disability rating 50–60% (−0.38) and 70–100% (−0.35). Severe-rated veterans retain meaningfully less often than non-rated peers at the same demographic and occupational position.
• Personal care & service (−0.14) and food preparation & serving (−0.11) occupations both carry retention penalties.
• MOS ↔ SOC skill match (+0.12). Veterans whose military occupation maps onto a civilian job in their current SOC major retain at higher rates. This is the signal the Build v0.12 match feature was engineered to capture.
• Education: no HS diploma (−0.12).
• State unemployment rate (−0.09 per SD). Tight labor markets hold cohort retention up.
• Disability rating 0% (+0.15). Veterans rated at 0% retain at meaningfully higher rates than the non-rated reference — consistent with the EDA finding that rated-but-unimpaired veterans fare better than the never-filed subgroup.

All seven above are significant at p < 0.001.

Out of scope

• Causal claims. The target is synthetic; no causal inference language is permissible.
• True survey inference. freq_weights reproduces the point estimate but does not produce design-correct standard errors. Confidence intervals in coefficients.csv are therefore directional, not survey-exact.
• Individual-level predictions. The model is fit to population-weighted rates; applying it to a specific veteran and treating the output as their retention probability would be inappropriate.
• Decision support. Nothing in this model should be used to make staffing, hiring, or benefits-administration decisions. It is a demonstration artifact.

Ethical notes

• No personally-identifying information is present. ACS PUMS has already been run through the Census Bureau's disclosure-avoidance routine; no row maps back to a named individual.
• The MOS ↔ SOC match feature is derived from the public O*NET Military Crosswalk. It is not based on any single veteran's DD-214.
• The synthetic retention target was chosen explicitly to avoid exposing any real retention data from any employer. No company, government agency, or VA system contributed to the outcome.
• Feature selection reflects the modeler's own judgment on what Phase 4a EDA surfaced as important. A different modeler would reasonably make different choices. This card documents the choices made; it does not claim they are the only defensible set.

Artifacts

What would change the story

Only two upstream deliveries materially alter this model:

• A real observed retention outcome. Would replace the synthetic target and allow actual causal claims about which features predict retention in this cohort.
• O*NET work-context composites (physical demand, schedule variability, autonomy). Currently blocked on the sandbox proxy allowlist. Once joined, the occupation-family coefficients would be decomposed into a smaller set of continuous work-environment variables, making the story about why some occupations retain more cleanly interpretable than 22 dummy variables.

Both are tracked in the project backlog.

Related artifacts