Sample and Data
We analyzed data from the Medical Expenditure Panel Survey (2006-2015), a nationally representative survey of the civilian non-institutionalized population. We had an eligible sample of 34,721 MEPS respondents who received Medicare and were age 65 years or older. The vast majority of our study covariates, including outcomes, main demographics, and chronic conditions, were virtually fully available for the entire sample (0% to <0.5% missing). Data for only 4 key variables (education, self-reported general and mental health status, and having a usual source of care) had missing for <1.5% of the eligible sample (Suppl. Table 1). With such low missing data rates, our final analytical sample included all respondents with complete data for all study covariates (n = 32,919). Excluded respondents (only 5% of the eligible sample) had greater ED and inpatient utilization, were more likely to have had a myocardial infarction or a stroke (and activity limitations), but had fewer chronic conditions overall. On average, excluded respondents were older, poorer, less likely to be white, and less likely to be married (Suppl. Table 2). Given the small size and worse characteristics of excluded participants, we did not expect their exclusion to materially bias our findings; if anything, our estimates might be slightly conservative.
We linked respondents’ data in MEPS annual files to their respective records from the Medical Conditions files, and then pooled linked datasets for years 2006-2015. Our data cover three distinct periods with respect to the ACA: pre-ACA (2006-2010), implementation period of ACA provisions relevant to older adults with MCCs (2011-2013), and post-ACA (2014-2015).
Outcomes: As primary outcomes, we first documented the prevalence of having any (at least one) emergency department (ED) visit, hospital inpatient visit, and overnight inpatient stay. As secondary outcomes, we analyzed counts of ED visits, inpatient visits, and LOS (total and average).
Chronic Conditions: We identified chronic conditions by using the definitions developed by Hwang and colleagues, and adopted by the Agency for Healthcare Research and Quality [26, 27], applied to International Classification of Diseases 9th Revision (ICD-9) 3-digit codes in the MEPS Medical Conditions files. We then calculated the total number of unique chronic conditions for each respondent, and categorized them as having 0, 1, 2, 3, 4, or 5+ chronic conditions. Those with ≥2 conditions were classified as having MCCs.
Covariates: Our analysis used data about respondents’ characteristics known to be related to ED visits, inpatient services, and having MCCs. Respondent sociodemographic characteristics included age, gender, race/ethnicity, language, marital status, Census region, income relative to the federal poverty line (FPL), and education. To measure respondents’ health status, we included self-rated general and mental health, activity limitations (physical and cognitive), and their chronic condition(s) (e.g., high blood pressure, diabetes, heart disease, stroke, and asthma). We also considered respondents’ access to care including types of payer (i.e., Medicaid, private insurance), having a usual source of care, receiving needed medical care, and getting needed prescription drugs. These factors are key determinants of ED use and hospitalization. Detailed levels of these covariates are reported in Table 2.
The goal of our analysis was to provide an update of where levels of ED visits and inpatient stays stand among older adults with MCCs following relevant ACA reforms, relative to the pre-ACA period. In our statistical models, this was accomplished by interacting a period indicator (pre-ACA = 0, post-ACA = 1) with chronic condition categories (having 5+, 4, 3, 2, 1, vs. 0), while including the main effects of these variables as well as the aforementioned confounding covariates. Since we are interested in the specific associations of having MCCs with ED/inpatient utilization, we adjusted for potential confounding by the following sets of covariates: 1) sociodemographic factors, which predispose (e.g., age) or enable (e.g., income) utilization; 2) particular conditions respondents had (e.g. stroke, myocardial infarction, asthma), which drive both the burden (count) of chronic conditions and the need for ED/inpatient utilization; and 3) additional insurance (Medicaid or private) and access-related factors (e.g., having a usual source of care), which also enable or create the need for ED/inpatient utilization. Our preferred model specification fully adjusts for these three sets of potential confounders. Additionally, we assessed the changes in model fit as we sequentially adjusted for these covariate sets.
We analyzed binary outcomes (prevalence of having ≥1 utilization event [i.e., visit or night]) in logit models. For count outcomes, we used a two-part, logit-negative binomial model, In the two-part model (known as a hurdle model for count data), a logit model is fitted for the probability of having ≥1 utilization event, and concurrently a negative binomial regression model is fitted for the actual count of events, conditional on a positive utilization event. By doing so, this two-part model handles the severely right-skewed nature of count distributions, with a concentrated mass of zeros on the left-hand side of the distribution and a very long right tail . Two-part models also allow recovering population-average estimates of change in outcome levels from the entire sample, as opposed to conditional estimates obtained from models fit only to the subsample with ≥1 event . After estimating each of our logit and two-part models, we recovered the adjusted, average marginal probability (of having ≥1 event) and count of events, by ACA period and MCC category. Finally, we estimated the pre-post-ACA changes in probabilities and counts for each MCC category.
For our logit models of binary outcomes, we assessed the goodness of fit using a modified version of the Hosmer-Lemeshow test for complex survey data . P-values for our preferred fully adjusted models were all between 0.3 and 0.4, indicating adequate fit. For the hurdle models of count data, we used Akaike and Bayesian Information Criteria (AIC & BIC) to compare model specifications. Our fully-adjusted models had the smallest AIC and BIC, indicating best fit among all tested specifications.
All models were estimated using maximum likelihood estimation. All estimates were also generated using Stata’s “svy” prefix, which uses survey weights to make estimates nationally representative. This prefix also calculates linearized standard errors, which account for MEPS’s complex, multi-stage sampling. All analyses were performed in Stata 14.2 (StataCorp, College Station, TX).