Healthcare Spending and Preventive Care in High-Deductible and Consumer-Directed Health Plans | Page 2
Published Online: March 22, 2011
Melinda Beeuwkes Buntin, PhD; Amelia M. Haviland, PhD; Roland McDevitt, PhD; and Neeraj Sood, PhD
measures, receipt of mammography, cervical and colorectal cancer screening, and glycosylated hemoglobin (A1C) testing for patients with diabetes mellitus. Dichotomous measures were created at the annual family level indicating whether some or none of the eligible family members had obtained the recommended care. We adapted HEDIS 2008 specifications to conform to the single-year windows in our analysis framework (discussed further in the eAppendices, available at www.ajmc.com). We created 2 child immunization measures indicating whether a child was on track to obtain the full set of recommended immunizations. Counting only care received in the current year caused these measures to be lower than typical HEDIS measures but consistently so in each year and across the treatment and control groups.
Covariates for analyses were derived from enrollment, claims, and geocoded location. Enrollment files provided family type, age of head of household, family size, geographic region, metropolitan statistical area status, and employer’s industry type. Claims data supplied prospective relative risk scores based on diagnostic cost group16,17 summed to the family level and indicators for whether a family received care in each of 23 major diagnostic categories. Actuarial values (percentage of allowed charges paid by the plan) were calculated for each plan using the plan provisions to simulate payment of claims for a standard population. The zip code–level geocoded characteristics are the median household income, percentage of adults with high school and college degrees, percentage unemployed, and percentage of Hispanic, black, and non- Hispanic white race/ethnicity.
For the cost-outcome models of the effects of enrollment in HDHPs or CDHPs, we used propensity score weighting to balance the distributions of numerous characteristics observed in 2004 between treatment and control families.18,19 Logistic regression analysis was used to model the odds of being a treatment family as a function of characteristics that predict both health plan selection and healthcare use (discussed further in the eAppendices). Predicted probabilities (propensity scores) were used to derive individual family weights for control families proportional to the conditional probability of being a treatment family. To check the adequacy of the propensity score model, we evaluated the balance of the weighted means of the measured characteristics. When balance is obtained, weighted analyses adjust for potential confounding owing to measured characteristics.
When estimating the effects for different types of high-deductible plans, treatment families were divided into different subgroups. Each subgroup was propensity score weighted to match the distribution of characteristics of the entire treatment sample (discussed further in the eAppendices).
Most of the covariates included in the propensity score model were also included in the weighted outcome models to provide estimates that are more efficient and “doubly robust” to misspecification of either model.20 We tested a range of generalized linear model specifications (identity and log links and constant, proportional to the mean, and proportional to the mean squared variance functions) to address the skewness and truncation at zero in healthcare costs.21,22 None of the other models tested outperformed the identity link and constant variance; hence, this is the model specification we use. Robust standard errors that account for clustering of family over time were used in all models. To estimate the effects of the employer decision to offer HDHPs or CDHPs on healthcare cost growth, we used parallel models but without propensity score weighting of those families not offered HDHPs or CDHPs.
For the HEDIS immunization outcomes, we computed unadjusted difference-in-difference estimates and then performed logistic regression analysis using the same framework and set of regressors as in the cost models aforedescribed.23,24 Unlike in the cost regression analysis, the same sets of families are not eligible for each measure in the pre-post years. For the remainder of the HEDIS outcomes, we stratified the sample into those who did or did not receive the recommended care in 2004 and within strata compared the rates of receiving the recommended care in 2005 by treatment status, both unadjusted and controlling for the same set of regressors as theprior models. The stratification was to address concerns that families who are about to transition into a high-deductible plan will try to obtain care that they anticipate needing before the transitions; that is, they will try to “stock up” on care.
Analyses were performed using commercially available statistical software (SAS version 9.1; SAS Institute, Cary, North Carolina; and STATA version 10; StataCorp Inc, College Station, Texas). We report statistical significance levels from 2-sided tests without adjustment for multiple testing. Full results are provided in the eAppendices.
Enrollees in high-deductible plans were more likely to be single men, were younger, had lower risk scores (better baseline health), and lived in areas with higher percentages of college graduates and non-Hispanic whites than families enrolled in control plans. After weighting using propensity scores, the samples have similar measured characteristics (Table 1 and eAppendix 1, available at www.ajmc.com).
At baseline, both groups were enrolled in plans with actuarial values that averaged 82%. The baseline monthly costs of the treatment and weighted control groups are given in Table 2 and eAppendix 2 (available at www.ajmc.com): both groups had similar monthly family healthcare costs of just over $500 and a distribution of costs by service type typical of those with employer-provided insurance. Before weighting, cost growth for the control group was 13%, similar to estimates for other data sources covering this period. Differences in the growth for the control group after weighting reflect the alignment of the controls to match those who enroll in treatment plans.
Effects of HDHP or CDHP Enrollment on Cost Growth
Costs grew for both the treatment families and the control families, but they grew more slowly in the higher-deductible group (Table 2). The monthly costs of the households enrolled in higher-deductible plans grew by $85 less than the comparable controls; in percentage terms, the expenditures of the control group grew by 20%, while the expenditures of the treatment group grew by 4%. Consequently, in the first year after enrolling in an HDHP or a CDHP, spending was 14% (95% confidence interval, 11.3%-16.9%) lower than that for comparable families in control plans (difference in the post-year mean monthly costs for treatment and control families divided by the mean post-year costs for control families). This was due to lower growth in inpatient, outpatient, and prescription drug costs. Growth in expenditures for emergency department care did not differ significantly between the 2 groups.
As shown in the Figure, cost growth for families in plans with moderately high deductibles ($500-$999) did not differ significantly from costs of those in control plans. However, cost reductions were greater (and significant) for families in plans with deductibles of $1000 or more. These cost reductions were maintained at a similar level when an account option was added with a low employer contribution (<$500 [mean, $399.32]). However, these cost reductions were attenuated when employers made generous contributions (>$500 [mean, $768.38]) to the accounts. This pattern of results across plan characteristics held for each of the individual care settings as well (discussed further in the eAppendices).
Some evidence was observed of increases in healthcare costs in the final quarter of 2004 among families who were about to enter an HDHP or a CDHP in 2005 (not significant for the treatment group as a whole but significant for those with deductibles >$1000), suggesting possible stocking up by those about to change insurance. Because the opportunity for stocking up is limited by the timing of information on health plan offerings for the upcoming year, insurance restrictions on the frequency of many procedures, and uncertainty about future health needs, we assume that any stocking up of services would occur near the time of the insurance change and would include services that would otherwise be obtained early at the start of the new plan-year. As a robustness check (discussed further in the eAppendices), we compared cost growth using the same framework but using only costs from the second and third quarters of 2004 and 2005. We obtained similar results.
Effects of HDHP or CDHP Offer on Cost Growth
PDF is available on the last page.