Data Source and Study Population
We obtained individual-level data from the Medicare Healthcare Effectiveness Data and Information Set (HEDIS) maintained by the Centers for Medicare and Medicaid Services (CMS) for the years 2007 through 2010. HEDIS contain individual-level data on Medicare Advantage (MA) enrollees’ use of hospital care. Individuals were matched to the Medicare beneficiary summary file to determine their demographic characteristics. Monthly information on health plan benefits for all Medicare plans was used to identify each plan’s cost-sharing requirement for inpatient hospitalizations. Information on health plan characteristics is publicly available on the CMS website.
We identified 33 plans that changed their inpatient benefit from a deductible at admission to a per diem (daily copayment), hereafter referred to as intervention plans. The intervention plans were identified across any two-year timeframe between 2007 and 2010 (e.g., 2007-2008, 2008-2009, or 2009-2010), with the intervention plans changing from a deductible in the first year of the two-year period to a per diem copayment in the second year of the two-year period. We found 223 plans that had no change in any inpatient or post-acute cost sharing across any one of the two-year timeframes between 2007 and 2010, hereafter referred to as control plans. Because changes in outpatient cost sharing can have an effect on hospital use and skilled nursing facility (SNF) or ambulatory care may substitute for hospital use, we limited the intervention and control plans to those that did not change, or made minimal change, to physician office or SNF cost sharing. Additionally, to mitigate any issues with co-insurance, we limited the intervention and control plans to those that did not impose co-insurance. In other words, the intervention plans imposed inpatient deductibles in year one and per diem copayments in year 2 whereas control plans imposed inpatient deductibles only on both years 1 and 2.
From the 33 intervention plans and 223 control plans, we utilized 1:n matching to match on the basis of contract year, tax status (i.e., for-profit or not-for-profit), geography, and deductible amount. We required plans to match based on contract year and tax status. Then, matching was prioritized by state, contract, neighboring state, division, region, and baseline inpatient deductible. Of the 33 case plans, 28 were matched to control plans. We excluded 5 pairs with incomplete data across the two analysis years or pairs with low volume (less than 150 admissions) in either of the analysis years. Our final sample consisted of 23 intervention plans matched to 36 control plans.
From our initial sample of 565,075 unique individuals, we limited our sample to those beneficiaries 65 years of age and older, excluding 99,303 individuals (17.8%), and who were not dually enrolled in Medicaid, excluding another 42,138 individuals (7.5%), resulting in our main analytic sample of 423,634 unique individuals enrolled in the intervention and control plans during our observation period.
The main outcome variables were inpatient utilization as measured by inpatient admissions per 100 enrollees, inpatient days per 100 enrollees, proportion hospitalized and the mean length of stay. Length of stay was calculated as the total number of inpatient days divided by the total number of inpatient admissions.
The primary independent variables were an indicator variable for whether the health plan changed from an inpatient deductible to a per diem (1 for intervention and 0 for controls), an indicator variable for time (0 for the year before the intervention plans changed the inpatient benefit and 1 for the year after), and an interaction term between those variables.
We determined whether each individual received a Part D subsidy, which can serve as a proxy covariate for low income. Since we do not have individual level income, the Part D subsidy can serve as a valid substitute since Part D low-income subsidy recipients have limited assets and a maximum income of 150% of the federal poverty level. (21) Those receiving part D subsidies were subject to inpatient and outpatient copayments, since we excluded dual eligible enrollees.
Covariates included age category (65 to 74 years or older than 74 years), sex, race or ethnic group (black, white, other), and low-income Part D subsidy. To account for differences in plan benefits, we added the copayment amount for primary care and specialist office visits and the monthly premium amount. To account for any temporal trends in inpatient utilization, we also included a fixed effect for the calendar year.
We used a difference-in-difference approach to assess the effect of plans changing from an inpatient deductible to a per diem benefit. This method accounts for time-invariant trends in outcomes by subtracting the change in inpatient utilization in control plans from the concurrent change in intervention plans that changed the inpatient cost-sharing benefit (hereafter referred to as difference-in-differences estimates). (22) (23)
We fitted one-part generalized linear models that included independent variables and covariates described above. We specified a negative binomial distribution and identity link for inpatient admissions and days per 100 enrollees and inpatient length of stay, and a binomial distribution for the proportion hospitalized. We ran each model using PROC GENMOD and clustered standard errors at the plan level to account for correlation among enrollees.
We conducted a sensitivity analysis that restricted the population to those who were continuously enrolled in the same plans for a full 24 months, the 12 months before and after the benefit change. These enrollees exhibited a much greater increase in utilization, perhaps indicating a sicker population with a higher likelihood of hospitalization in the second year. To account for exit and entrance of enrollees from health plans, we conducted an additional sensitivity analysis that considered all enrollees irrespective of the number of months of enrollment. Higher baseline utilization among these enrollees could be due to the inclusion of decedents who will often have high concentrations of hospital use at the end of life. Because there may be selection issues in the disenrollment from a plan, enrollment into a plan or the decision to stay in a plan based on the plan benefits, we also assessed the characteristics of enrollees that exited their plan, those that entered a plan after the intervention plans changed their benefit structures and those that remained with their plan as well as disenrollment rates from intervention and control plans.
To evaluate whether pre-policy trends in inpatient utilization were similar in intervention and control plans, we estimated difference-in-difference effects comparing annual changes in all outcomes during the two-year time period prior to the change in inpatient benefits. In other words, for an intervention plan that changed from a deductible in 2008 to a per diem in 2009, we analyzed the plan’s differences in inpatient utilization between 2007 and 2008. None of the estimates reached conventional levels of statistical significance at the 95% level. (Appendix Table 1) We also conducted a falsification test utilizing dual eligible enrollees that were excluded from our primary analysis since they are not subject to cost sharing. None of the estimates reached conventional levels of statistical significance at the 95% level. (Appendix Table 2)
All analyses were performed with the use of SAS software, version 9.4. Results are reported with two-tailed P-values or 95% confidence intervals. Our University’s Human Research Protections Office and the CMS Privacy Board approved the study protocol.