class: center, middle, inverse, title-slide # Medicaid Expansion and Health Changes for Safety Net Patients in the South ### John Graves, Vanderbilt School of Medicine <br> Laura Hatfield, Harvard Medical School <br> Bill Blot, Vanderbilt School of Medicine <br> Nancy Keating, Harvard Medical School and Brigham and Women’s Hospital <br> J. Michael McWilliams, Harvard Medical School and Brigham and Women’s Hospital ### 2019 (updated: 2019-05-27) --- # One of these regions is not like the others ... .center[ <img src="01_medicaid-expansion-map.png" width="75%" /> ] --- # Does the extant safety net provide implicit insurance? - ## One (of several) common arguments against reform in nonexpansion states. - ## What is the impact of Medicaid above and beyond implicit insurance provided via uncompensated & charity care, safety net clinics, etc.? --- # We still don't have a handle on the overall health effects of the ACA's Medicaid expansion - ## Mazurenko (2018): 60% of assessments did not find evidence of changes in self-reported health outcomes. - ## Previous studies focus on general low-income population and may not be well positioned to detect health status changes. - ## Dichotomized (excellent / very good vs. not, fair/poor vs. not) outcomes focus on ends of health distribution. --- # We analyze longitudinal data on the largest ever recruited cohort of low-income Americans - ## Southern Community Cohort Study - ## Recruited primarily at community health clinics in twelve states in the South. - ## SCCS participants have articularly low incomes, are older, and have considerably higher mortality, morbidity and uninsured rates than the low-income population in general. --- .center[ <img src="01_physical-health-meps.png" width="60%" /> ] --- .pull-left[ # SCCS Study Population - ### From 2001 and 2009, recruited 84,513 participants aged 40-79 from twelve states. - ### Of these, 4 (AR, KY, WV, LA) expanded Medicaid during our study frame. - ### 86% recruited at community health centers, the remainder via random mailing. ] .pull-right[ .center[ <img src="01_sccs-map.png" width="50%" /> ] ] --- # Study Population - ## Baseline survey and **four** follow-up surveys to date. - ## Baseline through FU2 were fielded prior to 2014. - ## FU3 fielded 2015-2017, FU4 in field now. --- # Study Population - ## Only FU1 and FU3 included Medical Outcomes Study 12-Item Short-Form General Health Survey (SF-12). - ## Effective response rate of ~67% once we account for both survey response and death. - ## We focus on a sample of 15,356 SCCS non-elderly participants <400% FPL in expansion and non-expansion states. --- # Study Outcomes - ## Self-reported health insurance - ## *Changes* in composite self-reported overall, physical and mental health. - ### Maintenance of health status (i.e., no change) [39%] - ### Health improvement [23%] - ### Health decline (includes death) [38%] - ## Survival --- # We utilize a Difference-in-Difference study design, with some wrinkles.... - ## **Challenge**: we only have two points in time; parallel trends assessments difficult. - ## **Solution**: use (continuous) survival data to test for differences in mortality in pre-period. - ## **Solution**: Permutation inference as a "check" on parallel trensds. --- # We utilize a Difference-in-Difference study design, with some wrinkles.... .pull-left[ - ## **Challenge**: SF-12 answers are ordinal, and measured on different scales. - ## **Solution**: model *changes* ] .pull-right[ .center[ .middle[ <img src="01_sf12.png" width="150%" /> ] ] ] --- # We utilize a Difference-in-Difference study design, with some wrinkles.... - ## **Challenge**: Death truncates self-reported health response, possibly biasing naive estimates. - ## **Solution**: model death jointly with other self-reported health outcomes. - ## **Soluation**: Principal Stratification for DD as a robustness check. --- # Our statistical inferences draw on randomization methods - ## **Challenge**: "Small-cluster bias" from four expansion states and eight nonexpansion states. - ## **Solution** Randomization inference based on permuting state expansion status. - ## Our inferences are based on relaying the "extremeness" of the observed estimate relative to the distribution of estimates generated by all (495) possible permutations. --- # Our statistical inferences draw on randomization methods .pull-left[ - ## Quantile ranking as a summary measure of extremeness. 0.2 and 100.0 are the most extreme values possible. - ## Also use sparkline visualizations to show observed estimate and distribution of estimate. ] .pull-right[ .center[ .middle[ <img src="01_sparkline.png" width="50%" /> ] ] ] --- background-image: url(01_baseline-table.png) background-size: contain class: center,middle, --- background-image: url(01_insurance-results.png) background-size: contain class: center,middle --- background-image: url(01_health-results.png) background-size: contain class: center,middle --- # Summary and Context - ## Medicaid expansion associated with maintenance of health and reduction in health declines. - ## 257: Implied number needed to insure to stem off a health decline. - ## Reduction in health declines analogous to moving from 8th the 4th ranked state, or from 4th to 1st. - ## Among low-income Southerners, safety-net access an inadequate substitute for health insurance coverage. --- # Limitations - ## Baseline differences in race across expansion vs. nonexpansion states. - ### Findings were nearly identical in sensitivity analyses that excluded all patient controls (including race) and that adjusted for differences in health trends associated with differences in race - ## Only observe one pre- and post-period observation (though FU4 in field now). - ### Analyzed diffrences in pre-expansion mortality and permutation inference test also provides a check on parallel trends assumption.