Impact of ACA/Medicaid Expansion on Excellent / Very Good Health

Research Challenges

• Rarely do we have RCT evidence. Where we do, it is noisy.
• Difficult to focus on populations where health gains likely to occur (OHIE: 7% with diabetes).
• Evidence on ACA impacts from household surveys is (repeated) cross-sectional, not longitudinal.
• Death is both an outcome of interest, and a competing risk for other outcomes.
• In Medicaid expansion studies, small cluster bias leads to (gross) over-rejection of null hypotheses.

Study Significance: Setting and Sample

• Largest prospective cohort of low-income minority adults in US (baseline sample =84,513).
• Drawn from poor and rural areas in the south.
• Cohort has high prevalence of chronic disease (40% report diabetes, 50% obese, 8% with cancer Dx).
• High mortality: 4% / 16mo vs. 0.8% / 16 mo in OHIE.
• Five states (FL, GA, LA, MS, SC) rank among the 10 states with the highest (pre-ACA) uninsured rates in the US.
  
• 80% of uninsured adults potentially eligible for the ACA’s Medicaid expansions reside in the South.

Study Signficance: Research Contribution

• Longitudinal evidence on changes in health insurance, access and self-reported physical and emotional health.
• Explicit accounting of death via study design and principal stratification.
• Exact inference via cluster-based permutation.

The Southern Community Cohort

• National Cancer Institute-funded cohort (n = 84,513) recruited starting in 2001.
• Recruited to identify root causes of health disparities in the incidence and risk factors of cancer and other chronic diseases.
• 86% recruited at community health centers in 12 southern states.
• 14% recruited via general population mail survey.
• States: AL, AR, FL, GA, KY, LA, MS, NC, TN, SC, VA, WV.

Inclusion Criteria at Baseline

• 40 to 79 years of age.
• English speaker.
• No Hx of cancer treatment in last year.

Baseline survey

• Demographics, medical history, family history, environmental and occupational exposures, diet, physical activity, psychosocial variables, health insurance, and access.
  
• Administered via standardized computer-assisted personal interviews for community health center participants, and via self-administered mailed questionnaires for the general population participants.

Follow-Up

• Mailed follow-up interviews roughly every 2-3 years, with additional phone-based follow-up for mail nonresponders.
• Advance consent with administrative linkages to cancer & death registries, CMS claims (Medicare and Medicaid) and hospital discharge data (TN only).
• High mortality (2-3%/year) but good response rates among living (62%), particularly given population (vast majority have income <$20k year).

Study Sample

• We utilize the 2009-2011 FU1 survey (n = ) as “baseline,” as this was the first year SF12 health questions were assessed.
• We fielded an additional followup (“FU3”) survey in 2015-2017, which targeted 53,104 living FU1 responders.
• Analytic sample drawn from 29,502 responders to FU3 survey as of early 2017.

CONSORT

Generalizability and Limitations

• Aged >40 at baseline, so nonelderly sample has older age profile.
• Analytic sample is approximately one-quarter general population SCCS sample.
• Given enrollment source, self-reported access to care skews towards clinics as usual source (27% vs. 19% in NHIS)

Key question: What was the impact of expansion on health?

• Standard approaches dichotomize a health status variable (e.g.,1 if excellent or very good health, 0 otherwise) and then fit a regression (e.g., logit, linear probability) to that outcome.
• Dichotomization results in loss of information and may obscure effects. What if expansions improved health from fair to good?
• How should we interpret an increase in self-reported poor health?

Key question: What was the impact of expansion on health?

• We really want to understand how health/access/insurance status changed, not just whether it changed.
• Our analytic approach is crafted with this goal in mind.
• Our study design was chosen to give us credible causal inferences.

Outcomes: Insurance

• Insured (yes/no)
• Insurance types: Private, Medicaid, Medicare, Military, Other, Uninsured

Outcomes: Access

• Usual source of care (e.g., private MD, clinic/CHC, ED, hospital, VA, no USOC).
• Did not see doctor due to cost.

Outcomes: Health

• General health status (excellent, v. good, good, fair poor, [mortality])
• Does your health now limit you in moderate activities? (a litte, a lot, not at all)
• During the past 4 weeks, how much of the time were you limited in the kind of work or other activities? (all, most, some, little, none of the time)
• During the past 4 weeks, how much of the time have you accomplished less than you would like as a result of any emotional problems? (all, most, some, little, none of the time)

Overview of Approach

• At the most basic level we are using a differences-in-differences (DD) design.
• DD comparison of expansion vs. nonexpansion states, with individual state and year controls.
• Rich sociodemographic and health controls from pre-ACA (baseline and first follow-up) surveys, which don’t end up mattering…
• But it gets complicated …

Overview of Approach

1. Quantities of interest.
2. Inference with few clusters.
3. Dealing with death.

Health Status Changes: Expansion States

Does your health now limit you in moderate activities?

Health Status Changes: Expansion States

Health Status Changes: Expansion States

Pre-Post Transitions as a Discrete Time Markov Chain

For treatment variable \(Z \in (0,1)\): \[ \\ p'_{\textrm{post}}(Z) = p'_{\textrm{pre}} T(Z)^t \]

Effects on Marginal Distribution of Outcome

\[ p'_{\textrm{post}}(1) - p'_{\textrm{post}}(0) \]

Effects on Outcome Transitions

\[ T(1) - T(0) \]

Health Status Changes: Expansion States

Health Status Changes: Expansion States

Health Status Changes: Expansion States

For outcome category \(d\) (e.g., excellent health)

\[ h(E(Y_{istd})) = \beta_0 + \beta_1 \textrm{Expansion-State}_s + \beta_2 \textrm{Post-Expansion}_t + \\ \beta_3 \textrm{Expansion-State}_s \times \textrm{Post-Expansion}_t + \\ \sum_{c=2}^C \beta_{4c}1(Y_{i0}=c) + \sum_{c = 2}^C \beta_{5c} \textrm{Expansion-State}_x \times 1(Y_{i0}=c) + \\ \sum_{c=2}^C \beta_{6c} \textrm{Post-Expansion}_t \times 1(Y_{i0} = c) + \\ \sum_{c = 2}^C \beta_{7c} \textrm{Expansion-State}_s \times \textrm{Post-Expansion}_t \times 1(Y_{i0}=c) + \\ \beta_8 \textrm{Sociodemographics}_{i0} + \beta_9\textrm{Clinical}_{i0} + \beta_{10}\textrm{Age}_{it} \]

Estimation

• Model fit using multiple multivariate regression, i.e., multiple outcomes with shared set of RHS variables.
• One-stop shopping: equivalent to fitting a series of DD regressions for each health category, stratified by baseline health category.
• Predicted values used to fill in the transition probability matrix under observed vs. counterfactual (no expansion).
• Close analogue: Multinomial logit model in Polsky et al (HSR 2009).

Health Status Changes: Expansion States

Inference

• Well known issues with conduncting (asymptotic) inference when there is a small number of (treated and untreated) clusters (Cameron, Gelbach, and Miller).

• Simulation study of 10,000 replicates of n=10,000 study of 4 treated and 8 control states, alpha = 0.05.

  Rej	ection Rate
  Standard	92.90
  Cluster robust (state)	10.50
  Wild t percentile bootstrap	5.63

Inference

• There are 495 possible permutations state treatment status with 12 total states and 4 expansion states.
• We estimate each parameter under each of the possible permutations, and compare the observed estimate relative to the distribution of (permutation-based) estiamtes.
• This exact inference procedure yields a rejection rate of 4.98% (alpha =0.05) in our simulation study.

##                        method  rate
## 1                    Standard 92.90
## 2      Cluster robust (state) 10.50
## 3 Wild t percentile bootstrap  5.63
## 4         Cluster permutation  4.98

Permutation Inference Sparklines

Baseline Characteristics

TK

Parallel Health Trends

Insurance Coverage

Insurance Coverage

Insurance Coverage

Access to Care: Usual Source of Care

Access to Care: Usual Source of Care

Self-Reported Health Status

Self-Reported Health Status

Self-Reported Health Status: Include Mortality

Does your health now limit you in moderate activities?

Does your health now limit you in moderate activities?

Physical Limitation

Physical Limitation

Emotional Limitation

Emotional Limitation

TEST: Does your health now limit you in moderate activities?

TEST: Does your health now limit you in moderate activities?