Intersectional Patterns In SNAP Take-up Among Likely-Eligible Women: Preliminary Findings from a Maihda

Background and Purpose: The SNAP is the largest U.S. food-assistance program, yet an estimated one in six likely-eligible households does not participate. Prior research has documented take-up disparities by race/ethnicity, rurality, and family structure individually, yet seldom examined how they combine. Multilevel analysis of individual heterogeneity and discriminatory accuracy (MAIHDA) partitions outcome variance between intersectional strata and individuals via variance partition coefficients (VPCs) and proportional changes in variance (PCVs), distinguishing additive main effects from multiplicative excess. We ask three questions: how much between-stratum variation in SNAP take-up exists among likely-eligible adult women; how much is explained by additive main effects of race/ethnicity, rurality, and household structure; and how much residual intersectional excess remains.

Methods: Data were drawn from the 2015–2019 ACS PUMS. Likely-eligible women aged 18 and older were identified via state-specific Broad-Based Categorical Eligibility thresholds; one adult woman per household was retained by random selection. We estimated MAIHDA models using both frequentist and Bayesian Markov Chain Monte Carlo methods. Given the computational demands of fitting these models to the full dataset (~ 1.2 million records), results presented are based on preliminary analyses using a 10% random subsample (N = 120,759) while full-sample analyses are ongoing. Individuals were cross-classified into 72 intersectional strata (6 race/ethnicity categories, one of which aggregates Native Hawaiian or Pacific Islander, Some Other Race, and Multiracial respondents into a single cell, crossed with 3 rurality and 4 household-structure categories) and 51 state-level units. Three models were fit via R2MLwiN: a null model (M0); an additive main-effects model with year fixed effects (M1); and a nested refit omitting year (M1a) to decompose the M0→M1 variance reduction. Each used four parallel chains of 200,000 monitoring iterations after 50,000 burn-in.

Results: All three models satisfied the convergence gate (maximum potential scale reduction factor ≤ 1.004; minimum effective sample size = 490). The null-model between-stratum VPC was 13.50% (95% credible interval [CrI] 10.03%–18.54%), substantially above the published MAIHDA-corpus median of 5.5%. The null-model median odds ratio of 1.98 reflects a roughly two-fold gap between randomly chosen strata. Additive main effects absorbed 94.94% of between-stratum variance (M0→M1 PCV); the residual intersectional VPC was 0.78% (95% CrI 0.46%–1.29%), small yet clearly above zero and indicating a modest nontrivial multiplicative component. Year fixed effects absorbed no between-stratum signal (M1 and M1a σ²_u = 0.0260 vs. 0.0259). State-level variance absorbed approximately 2%–3% of total variance, stable across models.

Conclusions and Implications: Intersectional position is a moderately strong predictor of likely-eligible women’s SNAP take-up, yet the patterning is predominantly additive: race/ethnicity, rurality, and household structure combine in a largely separable fashion, leaving a small residual multiplicative component. For social welfare policy, outreach targeting the most disadvantaged positions on each axis would capture this additive patterning, while the modest residual excess identifies intersectional positions where single-axis approaches underperform and administrative-burden reduction becomes the more appropriate lever. State-level variance (~2%–3%) is the natural target for future state-policy-moderation analyses.