Supplementary Appendix

Transcription

1 Supplementary Appendix This appendix has been provided by the authors to give readers additional information about their work. Supplement to: Sommers BD, Musco T, Finegold K, Gunja MZ, Burke A, McDowell AM. Health reform and changes in health insurance coverage in N Engl J Med 2014;371: DOI: /NEJMsr

2 Health Reform and Changes in Health Insurance Coverage in 2014 Online Appendix: Contents Appendix Methods Page 2 Appendix Table S1: Sensitivity Analyses of Changes in the Uninsured Rate in 2014 Page 6 Appendix Table S2: Changes in the Uninsured Rate Among Adults 18 to 64, Based on Income and State Medicaid Expansion Status (Excluding Income Non-Responders) Page 7 Appendix Table S3: Association between Uninsured Rate in Survey Data and ACA State-Level per Capita Enrollment Statistics from Government Reports Page 8 Appendix References Page 9 1

3 Appendix: Methods Regression Models Primary Analysis: Identifying Changes in Coverage Trend, with Multivariate Adjustment Uninsured ist = β 0 +β 1 TimeTrend t +β 2 Q t +β 3 Q t +β 4 Q t + β 5 Employed i + β 6 Income i + β x X i + Ω State s + ε ist Equation (1) where i indexes person, s state, and t date. X i is a vector of demographics (age, self-reported race/ethnicity, and sex). Ω is a vector of state fixed effects. TimeTrend is a linear variable measuring the number of months since the beginning of the study sample (January 2012 is 1, February 2012 is 2, etc.). The coefficients of interest are β 2 through β 4, which capture the quarterly changes in the uninsured rate compared to the pre-existing trend, associated with the beginning of open enrollment (October-December 2013; β 2 ), the second half of open enrollment combined with the initiation of new Marketplace and Medicaid eligibility (January-March 2014; β 3 ), and the first full quarter after the end of the open enrollment period (April-June 2014; β 4 ). Subgroup analyses in the paper (Tables 2 and 3) use the same model, with the sample limited to each subgroup one at a time. This is also the regression equation used to model the two access outcomes. Sensitivity Analyses: The sensitivity analyses in Appendix Table S1 use Equation 1, with the inclusion of a quadratic time trend (TimeTrend squared) in addition to the linear trend, or without any time trend variable. Alternative time frames for the sample were analyzed using Equation 1. We also considered models in which we did not adjust for employment and/or income, since the policy changes related to the ACA may have led to changes in employment options or decisions, which may be mediating factors for changes in health insurance. Differences-in-Differences Analysis of Coverage for Low-Income Adults: Uninsured ist = β 0 +β 1 Expansion State s * Year2014 t +β 2 Year2014 t +β 3 TimeTrend t + β 4 Employed i + β 5 Income i + β x X i + Ω State s + ε ist Equation (2) Expansion State is equal to 1 for states that expanded Medicaid between January and April 2014, and 0 for states that did not. Year2014 is a binary variable for the year The state fixed effects capture the direct impact of living in an expansion state. β 1 is the differences-in- 2

4 differences estimate for how much the uninsured rate changed in Medicaid expansion states in 2014, compared to non-expansion states. As is standard in such analyses, we use robust standard errors clustered at the state level. Q was excluded from this analysis as a wash-out period. Comparing Survey-Reported Coverage Changes by State with HHS ACA Enrollment Data Uninsured ist = β 0 +β 1 TimeTrend t +β 2 Q t +β 3 Q t +β 4 Q t + β 5 Q t * Per Capita HHS Enrollment s + β 6 Employed i + β 7 Income i + β x X i + Ω State s + ε ist Equation (3) Per Capita HHS Enrollment is the ratio of HHS s official Marketplace-Medicaid enrollment total (by state) to the state s non-elderly population see below for more details on this data source. This variable captures the approximate percentage of the state s population reported to have signed up for ACA-related coverage through the Marketplaces by the end of March β 5 estimates how much those state-level percentages (from administrative enrollment data) predict overall changes in the state uninsured rate in 2014 (from the WBI survey data). Given the statelevel variable of interest, we use robust standard errors clustered at the state level for these estimates. One limitation is that the WBI data are for adults 18-64, whereas HHS reports include children 0-17; differences in state-level enrollment between children and adults could bias these estimates. We also conducted a simple non-regression based comparison of the state-level survey estimates and the HHS enrollment statistics. For this approach, we estimated each states change in the uninsured rate from the pre-open enrollment period to Q2 2014, and then measured the population-weighted correlation coefficient between those state estimates and the HHS statistics. The result of this analysis was rho = (p<0.001). HHS Enrollment Data We linked survey data from the Gallup-Healthways Well-Being Index with official enrollment statistics for Medicaid/CHIP and Marketplace coverage in each state, published by the U.S. Department of Health and Human Services (HHS). These statistics are based on administrative data collected by the Federally-facilitated Marketplace and data submitted to the Centers for Medicare & Medicaid Services (CMS) by State-Based Marketplaces. The totals used in our analysis are the sum of the individuals determined or assessed eligible for Medicaid / CHIP by the Marketplace and the number of individuals who have selected a Marketplace plan (Columns 5 and 6 from Appendix C in the HHS report [ASPE 2014]). The Marketplace enrollment data are based on the total number of individuals who have selected a qualified health plan through the Marketplace during the reporting period, including those who have already paid a premium and those who have not yet paid a premium. These totals include all enrollment through March 31,

5 Our analysis only includes enrollment in Medicaid/CHIP that occurred via Marketplace eligibility determinations, and does not include enrollment that occurred through state Medicaid agencies, due to significant data limitations in the latter. CMS reports of Medicaid/CHIP enrollment have not been able to distinguish between new applications versus eligibility redeterminations for individuals already enrolled in Medicaid/CHIP the prior year for certain states, and are also subject to double counting if a person has had multiple eligibility determinations during the study period. One state s figures also included eligibility determinations for non-medicaid programs, such as the AIDS drug assistance program and statefunded insurance. For these reasons, we limit our analysis to Medicaid/CHIP enrollment figures from the Marketplace statistics. To convert the administrative figures into comparable units, we divided the total enrollment for each state by the non-elderly population in that state, based on the 2012 American Community Survey (ACS) conducted by the U.S. Census Bureau. Gallup-Healthways Well-Being Index (WBI) Survey Weights Gallup weights the WBI survey data daily to compensate for disproportionalities in selection probabilities and nonresponse. In addition, Gallup weights the data to match targets from the U.S. Census Bureau by age, sex, region, education, ethnicity, and race, as well as population density of self-reported location. Gallup also weights the data to phone status targets from the National Health Interview Survey (NHIS). 1 Gallup-Healthways Well-Being Index Survey Items for Study Outcomes We analyzed three outcomes in the WBI: the uninsured rate, having a personal doctor, and difficulties affording medical care. The relevant survey items were as follows: 1) Do you have health insurance coverage? 2) Do you have a personal doctor? 3) Have there been times in the past 12 months when you did not have enough money to pay for healthcare and/ or medicines that you or your family needed? Imputation of Income as a Percentage of the Federal Poverty Level 22% of respondents in the Gallup-Healthways WBI do not report their monthly income. This figure is quite similar to the rate of non-response for income in government surveys such as the National Health Interview Survey (24-33%) and the ACS (22.4%). Following the approach used by Skopec et al. 2014, we imputed income in the Gallup-Healthways WBI for missing values using a multivariate regression model as a function of age, race/ethnicity, sex, state of residence, household size, marital status, health insurance, self-reported health status, and employment. 1 See for additional background information. 4

6 After imputation, previous research (ibid.) shows that the Gallup sample has a similar proportion of low-income adults as the government surveys, but more middle-income and fewer highincome individuals. Household income was then converted into a percentage of the federal poverty level based on HHS poverty guidelines and household size. 17.4% of the weighted sample did not report their household size, so we used a similar approach as for income to impute these missing values, except with a Poisson distribution for count data. 2.2% of the weighted sample required imputation for both income and household size. Income categories in the WBI are reported in monthly terms. We multiplied these results by 12 to obtain annual income. The categories of annual income are as follows: Less than $720, $720- $6000, $6001-$12,000; $12,001-$24,000; $24,001-$36,000; $36,001-$48,000; $48,001- $60,000; $60,001-$90,000; $90,001-$120,000; and greater than $120,000. Income analyses using the Gallup data show higher baseline uninsured rates for low-income individuals than the ACS (56-60% in the WBI compared with ~40% in the ACS) and lower uninsured rates for high-income individuals (~2% in the WBI compared with ~5% in the ACS). The need to impute some of the Gallup income data, the over-representation of middle income individuals, the imprecision of the ranges in which income is reported, and the possibility that in complicated households respondents may provide information that does not match the family definition used in federal surveys may all contribute to this pattern. 5

7 Appendix Table S1: Sensitivity Analyses of Changes in the Uninsured Rate in 2014 MODEL / ANALYSIS FIRST QUARTER 2014 SECOND QUARTER % CI P value Change from 95% CI P value Prior Trend Change from Prior Trend Time Trend and Study Period Quadratic Time Trend , -2.6 < , -5.4 <0.001 No Time Trend , -1.2 < , -3.7 <0.001 Study period , -2.1 < , -4.6 <0.001 Study period , -1.4 < , -3.9 <0.001 Alternative Models No adjustment for employment , -1.9 < , -4.5 <0.001 No adjustment for employment or income , -1.1 < , -3.9 <0.001 Difference-in-Difference Analysis < 138% FPL in Expansion vs. Non-Expansion State N/A , Notes: All estimates are in percentage points. All models use nationally-representative survey weights and adjust for age, sex, race, ethnicity, employment, household income, and state of residence, unless otherwise noted. Analyses also included a binary variable for the fourth quarter of % CI = 95% confidence interval This model pools Q1 and Q2 of 2014 and omits Q as a wash-out period. 6

8 Appendix Table S2: Changes in the Uninsured Rate Among Adults 18 to 64, Based on Income and State Medicaid Expansion Status (Excluding Income Non-Responders) POPULATION Income Under 138% of FPL Non-Medicaid Expansion States Medicaid Expansion States Income % of FPL Non-Medicaid Expansion States Medicaid Expansion States Income Over 400% of FPL Non-Medicaid Expansion States Medicaid Expansion States Baseline Uninsured Rate* FIRST QUARTER 2014 SECOND QUARTER % CI P Change from value Prior Trend Change from Prior Trend 95% CI P value , , , , , -2.6 < , -3.9 < , -3.2 < , -7.5 < , , , , Notes: All estimates are in percentage points. All models use nationally-representative survey weights and adjust for age, sex, race, ethnicity, employment, household income, state of residence, and a linear time trend. Analyses also included a binary variable for Q4 2013; for brevity, those results are not reported here. Sample excludes individuals not reporting any income information. 95% CI = 95% confidence interval *Mean uninsured rate for the population, from Q1 of 2012 through Q

9 Appendix Table S3: Association between Uninsured Rate in Survey Data and ACA State-Level per Capita Enrollment Statistics from Government Reports Probability of Being Uninsured, Adults VARIABLE Coefficient 95% CI P value Q (-1.0, 0.3) 0.35 Q (-3.4, -1.8) <0.001 Q (-3.9, -1.0) Q * Per Capita HHS Enrollment (-0.77, -0.29) <0.001 Notes: All estimates are in percentage points. All models use nationally-representative survey weights and adjust for age, sex, race, ethnicity, employment, household income, state of residence, and a linear time trend. Q * Per Capita HHS Enrollment captures the relationship between the Q reduction in the uninsured rate and the per capita HHS enrollment statistics for each state. The coefficient of means that each percentage point increase in HHS enrollment for a given state was associated with a 0.53 percentage point decline in the surveymeasured uninsured rate. 95% CI = 95% confidence interval 8

10 Appendix References Health Insurance Marketplace: Summary Enrollment Report for the Initial Annual Open Enrollment Period. U.S. Department of Health and Human Services, Office of the Assistant Secretary for Planning and Evaluation (ASPE), (Accessed at t.pdf) Skopec L, Musco T, Sommers BD. A Potential New Data Source for Assessing the Impacts of Health Reform: Evaluating the Gallup-Healthways Well-Being Index. Healthcare: Journal of Delivery Science and Innovation 2014;2: