This week, four analysts from CBO's Health Analysis Division are presenting their work at the 12th Annual Conference of the American Society of Health Economists ("ASHEcon") in St. Louis, Missouri. These presentations are part of CBO’s engagement with the broader research community, which improves the quality of our analysis and makes our methods and findings more transparent and available. The presentations show the range and quality of the Health Analysis Division's work:
Estimating Group Effects
Christopher Adams (CBO)
This paper considers the problem of estimating relationships at the group level when the researcher has a large number of groups but the number of observations in each group may be small. This is a generalization of standard fixed effects estimators. The paper proposes a non-parametric empirical Bayesian estimator as described in Robbins (1956). While fixed effects estimators run into problems as the number of groups gets large or the size of the groups is small, empirical Bayesian estimators become better as the number of groups gets large and are generally agnostic to group size. The paper presents theoretical attributes of this classical estimator, the estimation algorithm, R code, Monte Carlo results and uses the estimator to analyze the demand for drugs. The paper presents demand elasticities for 2,600 drug-brand-payor-type pairs using data from the Medical Expenditure Panel Survey (MEPS, 2010 to 2019). We show that demand for drugs is very inelastic. For brands being bought with insurance coverage, the average elasticity is -0.09 to changes in out-of-pocket prices. For brands bought without insurance, the average elasticity is -0.08 to changes in the total price. In both cases, a 10 percent increase in how much the patient pays is associated with a less than a 1 percent decrease in their demand for the product.
The fixed effects model is one of the workhorse models of economics. It is a relatively simple way to account for possible confounding in the data. Unfortunately, there may be identification problems because there is not enough variation at the group level. There can also be a huge number of parameters that are difficult to estimate, computationally burdensome, and often of little direct interest. These problems are exacerbated when the number of groups is large and the number observations within a group is small. This paper argues that an idea originally proposed in Robbins (1956) is well suited for estimating parameters at the group level. Moreover, the procedure works better when the number of groups is large, and it will provide reasonable parameter estimates even for groups with little or no variation.
In demand estimation we have a standard problem that prices are determined endogenously and so the observed variation in prices and output reflects both changes in demand and changes in supply. A number of solutions have been presented that generally involve a mix of parametric restrictions, structural assumptions, and instrumental variables. One issue with these approaches is that they are computationally burdensome and require assumptions on the data generating process that may not be reasonable. Here we estimate the demand parameter at the brand/payor-type level and assume that conditioning upon this group accounts for confounding in the variation of prices. That is, we assume that variation in prices over time for a particular drug brand is "as if" random. Positive elasticity estimates for some drug brands suggest that even this within-brand variation in prices may be confounded.
The Effect of Expanded Access to Telehealth on Medicare Spending and Utilization
Michael Cohen (CBO), Caroline Hanson (CBO), and Ru Ding (CBO)
To promote social distancing and maintain beneficiaries’ access to care during the COVID-19 pandemic, the Centers for Medicare and Medicaid Services authorized several waivers in 2020 that expanded Medicare’s coverage of telehealth. Those waivers broadened the services that were eligible for reimbursement and the providers who could deliver them, equalized payments for in-person and remote visits, and eased Medicare’s restrictions on where telehealth services could be provided. Congress has since adopted legislation to extend many of those changes through 2024. An important policy question is whether such expansions in telehealth coverage increase health care spending incurred by Medicare beneficiaries, and in turn, the federal deficit. The answer depends on whether telehealth replaces existing in-person services with remote visits or whether it spurs additional utilization. To date, most evidence on the effect of telehealth on utilization and spending has come from studies that limit their scope to a narrow set of clinical conditions (such as upper respiratory infections) and that focus on a commercially insured population—a group that has historically had much broader access to telehealth compared with Medicare beneficiaries.
In this paper, we conduct a difference-in-differences analysis to examine how per-beneficiary spending and utilization compares between Medicare fee-for-service (FFS) beneficiaries with high- and low-access to telehealth, before and after Medicare broadened its coverage of telehealth. We use 100 percent Medicare FFS claims files, in addition to beneficiary characteristics from the Master Beneficiary Summary File and other zip-code characteristics from the American Community Survey.
We define high- and low-access to telehealth in a two-step process. We first estimate providers’ idiosyncratic tendency to use telehealth based on their encounters with Medicare beneficiaries between March 6 and June 30, 2020. Our estimate of each providers’ tendency to use telehealth accounts for differences in the local trajectories of COVID-19 across areas, a providers’ specialty, and differences in providers’ patient populations. We then assign each beneficiary to high- or low-access according to the average tendency to use telehealth for the providers used by that beneficiary, weighted by his or her visits to those providers in 2019. We evaluate covariate balance and test for the presence of parallel trends prior to March 2020. To examine whether access to telehealth increases or decreases spending and utilization, we measure the effect of high-access to telehealth on per-beneficiary spending and utilization on all Medicare services, telehealth services, and in-person services. We also examine whether those effects differ by the type of service, such as office visits, pharmacy claims, and inpatient visits.
The Effect of the Pandemic Maintenance of Eligibility Policy on Medicaid Spending and Enrollment
Nianyi Hong (CBO), Noelia Duchovny (CBO), Ru Ding (CBO)
Enacted in 2020, the Families First Coronavirus Response Act provided states with an increase in federal funding in exchange for providing continuous eligibility to Medicaid enrollees through the end of the Public Health Emergency (PHE). All states took up the option, with enrollment increasing significantly; between February 2020 and July 2022, enrollment increased by 19 million individuals or roughly 26 percent. Other factors accounting for the increase in enrollment include worsening economic conditions and Medicaid expansions in five states. Understanding changes in enrollment and spending during the pandemic may help inform the cost of the pandemic policy and future policy proposals related to continuous eligibility.
Using data from the Centers for Medicare & Medicaid Services’ Transformed Medicaid Statistical Information System (CMS T-MSIS), we investigate enrollment and spending patterns for Medicaid enrollees before and during the PHE. The T-MSIS includes detailed enrollment and spending data for all states, along with key demographic information. We compare outcomes for a pre-pandemic cohort of enrollees starting in February 2018 with a pandemic cohort of enrollees starting in February 2020, which we track for 18 months. Our analysis is done by eligibility group (children and adults) because we expect larger effects among adults as most states provided 12-month continuous eligibility for children before the pandemic. The analytical sample includes 34 states and the District of Columbia; we use CMS’s Data Quality Atlas, a tool that evaluates the validity of T-MSIS data, to exclude other states with data issues regarding enrollment or spending.
Our preliminary results comparing the two cohorts suggest that enrollment patterns changed significantly. Due to the maintenance of eligibility requirement, enrollees in all groups who obtained Medicaid coverage in 2020 stayed on Medicaid coverage for significantly longer than those before the pandemic, with larger differences among adults. The share of people still enrolled 18 months after the beginning of the cohort increased from 72 percent to 93 percent among kids and from 63 percent to 89 percent among adults. Even with continuous eligibility in effect for the pandemic cohort, some people can lose Medicaid eligibility if they move out of state, voluntarily withdraw, or die.
Spending patterns changed less substantially. In nominal dollars among consistently enrolled beneficiaries, spending per child increased 13 percent over the 18-month period in the pre-pandemic cohort and only 9 percent in the pandemic cohort. Spending per adult declined 4 percent over the 18-month period in the pre-pandemic cohort and 7 percent in the pandemic cohort. In other words, spending among children increased over the 18-month period in the pandemic cohort, but more slowly than in the pre-pandemic period, while spending for adults decreased over this period for both cohorts, with greater reductions in the pandemic cohort.
Although our analyses cannot separate the effects of the pandemic from those of continuous eligibility, our findings suggest that Medicaid enrollment persisted longer after the pandemic while Medicaid spending changed less significantly. Continuous coverage could have reduced the negative consequences associated with churn or loss of insurance coverage, including interruption of care.
How the Prices Nongroup and Small Group Health Plans Pay to Providers Compare to the Prices Medicare Pays
Karen Stockley (CBO), Caroline Hanson (CBO), Ian McCarthy (Emory University), and Eamon Molloy (CBO)
Past research finds that employer-sponsored health plans pay providers much higher prices than Medicare, on average, particularly for hospital services. Several factors—including the greater prevalence of narrow networks plans and the greater participation of low-cost insurers that also participate in Medicaid—suggest that the prices paid by nongroup plans may be lower than in the employer-sponsored market. However, little is known about how the prices paid by nongroup health plans compare with those paid by employer-sponsored plans or Medicare. In this paper, we estimate how prices for professional, outpatient hospital, and inpatient hospital services in the on- and off-marketplace nongroup and small group markets compare with Medicare prices. We use Medicare as a reference point because it is a well understood benchmark and facilitates comparisons to previous literature.
We estimate nongroup and small group prices using the 2019 EDGE Limited Data Set, which includes claims covering almost all people enrolled in nongroup and small group plans subject to risk adjustment under the Affordable Care Act. We estimate Medicare prices for fee-for-service Medicare enrollees using the 2019 carrier, outpatient, and inpatient claims.
We present price comparisons of the private market segments to Medicare using the ratio of mean prices at the national level, after reweighting the Medicare sample to reflect the geographic composition of the private samples. We present price ratios for selected services, such as specific Healthcare Common Procedure Coding System (HCPCS) procedure codes or inpatient hospital Diagnosis-Related Group (DRG) codes, and using a market basket index of high-volume services. Our methods address two challenges for comparing inpatient and outpatient prices: a lack of DRG codes in the EDGE data for inpatient claims and differences in the allocation of payments to different claim lines for outpatient claims across the two datasets. For inpatient claims, we use machine learning techniques to impute DRG codes using a random forest algorithm trained on the Medicare data. For outpatient claims, we explore several approaches for comparing prices, and present estimates from multiple methods to assess the sensitivity of the estimates.
We find that, for the top 100 services by spending, professional prices are 13 percent higher than Medicare in on-marketplace plans, 18 percent higher than Medicare in off-marketplace plans, and 20 percent higher than Medicare in small group plans. Additionally, we find that prices for a set of outpatient claims matched by HCPCS code are 81 percent higher than Medicare in on-marketplace plans, 98 percent higher in off-marketplace plans, and 109 percent higher in small group plans. Other work finds that plans in the broader employer-sponsored market pay professional and outpatient prices that are about 30 percent and 140 percent higher than Medicare, respectively. Those estimates suggest that factors specifically affecting the nongroup market lead to somewhat lower prices relative to employer-sponsored plans. Our estimates also indicate that a public option with prices similar to Medicare’s would have one type of cost advantage over private nongroup plans.
Chapin White is CBO's Director of Health Analysis.