Does the Safety Net Bite Back? Medicaid Postpartum Coverage Extensions Through the Public Health Emergency and Beyond

Abstract

Between 2021 and 2024, 47 U.S.\ jurisdictions adopted extensions of Medicaid postpartum coverage from 60 days to 12 months, the most rapid expansion of maternal health coverage in decades. Using individual-level data on 237,365 postpartum women from the American Community Survey (2017–2019, 2021–2024; 2020 excluded due to non-standard data collection) and a staggered difference-in-differences design with the Callaway and Sant'Anna (2021) estimator, this paper evaluates whether these extensions increased insurance coverage among women who recently gave birth. The full-sample CS-DiD ATT for Medicaid coverage is $-0.50$ percentage points (SE $= 0.63$ pp, $p > 0.10$). The post-PHE specification (2017–2019 + 2023–2024), which excludes the period when continuous enrollment rendered the coverage cliff non-binding, yields a statistically significant negative Medicaid ATT of $-2.18$ pp (SE $= 0.74$ pp, $p < 0.01$). This negative estimate does not reflect policy harm; rather, it captures the secular Medicaid unwinding that disproportionately reduced enrollment in treated states after the PHE ended. A triple-difference (DDD) design—comparing postpartum to non-postpartum low-income women within treated and control states—absorbs these common unwinding shocks. The DDD CS-DiD estimate is $+0.99$ pp (SE $= 1.55$ pp), statistically insignificant but signed in the direction predicted by the policy's institutional logic. The overall story is that the standard DiD picks up the Medicaid unwinding confound; the DDD resolves it, yielding a small but imprecise postpartum-specific effect estimate. Event-study estimates through two post-treatment periods show flat dynamics, and permutation inference (200 randomizations of the full CS-DiD estimator), wild cluster bootstrap (including state-cluster bootstrap for CS-DiD), and Rambachan-Roth HonestDiD sensitivity analysis all confirm that the confidence interval includes zero under moderate violations of parallel trends. These findings constitute a well-identified result with substantially stronger methodology than the earlier analysis: even with post-PHE data, a DDD design, and multiple inference procedures, the postpartum extensions do not produce detectable coverage gains in survey data, suggesting that administrative enrollment mechanics, measurement limitations in the ACS, or the thin control group (4 states) may attenuate the policy's apparent impact.