Medicaid Subscription-Based Payment Models and Implications for Access to Hepatitis C Medications

Key Points Question Did the use of direct-acting antiviral hepatitis C virus (HCV) medications change after implementation of subscription-based payment models for these drugs in Washington and Louisiana? Findings In this cross-sectional study, Louisiana experienced a 534.5% increase in HCV prescription fills after implementation of a subscription-based payment model, but no significant change in prescription fills was observed in Washington. Meaning In this study, subscription-based payment models in Louisiana and Washington were differentially associated with use of Medicaid-covered HCV medications, which may reflect state-level differences in implementation, historical restrictions on access to these medications, and responses to the COVID-19 pandemic.

We calculated our outcome variables quarterly as follows: HCV prescription fills per 100,000 enrollees = × 100,000

Unadjusted Comparisons
Unadjusted changes in HCV prescription fills were identified by comparing the four calendar quarters preceding implementation of subscription-based payment-models (July 1 st , 2018 through June 30 th , 2019) with the following four quarters (July 1 st , 2019 through June 30 th , 2020). Significance levels and confidence intervals were calculated using two-sided t-tests.

Synthetic Control Comparisons
The basic idea of synthetic control methods is to use a weighted average of each state's donor pool, with the weights chosen so that pre-trends in outcomes are as similar as possible between the treated state(s) and synthetic control. 1 We followed the approach outlined in Robbins, Saunders, and Kilmer (2016)  Taylor series linearization (TSL) was then used to calculate estimates and 95% confidence intervals for the effects of SBPM implementation during the following year. This approach accounts for the complex weighting of control states used to create the synthetic control. TSL uses a linear function of the observed data to approximate the estimator, and the variance estimation formulae for a linear estimator can then be applied to the linear approximation. In general, this leads to a statistical consistent estimator of the variance. 3 The TSL method tells us whether any differences in outcome between the treated states and their synthetic controls are statistically significant.
We also conducted a series of permutation tests to determine placebo effect sizes. This procedure occurs in four steps. First, we subset the donor pool of control states for a given treated state and outcome. Second, we iteratively reassign the "treatment" to each control state. Third, weights are calculated to match the placebo treatment to a new synthetic control, and a placebo effect, sampling distribution and associated p-value are generated. Lastly, we use a 2-sided t-test to determine whether the observed effect in the treated state is likely to have occurred by chance, given the distribution of placebo effect estimates. 4

Robustness Tests
Several tests were performed to assess the robustness of results: (1) In our main specification, our primary outcome was the rate of HCV prescription fills per 100,000 Medicaid enrollees. We alternatively measured out outcome at the per-pill level to account for potential differences in the length of treatment between direct-acting antiviral medications.
(2) We re-estimated our synthetic control models excluding the second quarter of 2020. This enabled us to assess the sensitivity of our results to the emergence of COVID-19 and subsequent disruption in the delivery of health services.
(3) In our main specification, we curated the donor pools for Louisiana and Washington to include only those states which had similar liver damage and sobriety restrictions in place at the time of SBPM implementation. Alternatively, we estimated 'unrestricted' versions of our synthetic control models allowing all control states to be included in the donor pool.
(4) To test whether our results are driven by trends in individual control states, we estimated 'leave one out' versions of our synthetic control models. This was achieved by iteratively removing control states from the donor pool and re-estimating our synthetic control models.
(5) Louisiana removed restrictions on HCV medication access (i.e., liver damage and sobriety restrictions) concomitantly with SBPM implementation. The SBPM granted the state budget predictability and may have enabled the loosening of access restrictions; nonetheless we estimated a 'stacked' synthetic control model for Louisiana to assess to what extent changes in HCV medication utilization were associated wither SBPM implementation versus loosening of access restrictions.
The stacking procedure proceeded as follows. We first limited the donor pool to those states that also removed both liver damage and sobriety restrictions during our study period. Next, we manipulated the time variable in our data set, shifting states forward or backward so their loosening of access restrictions coincides with SBPM implementation (graphically depicted in eFigure 3). Additionally, states were removed if they lacked either two years of pre-period data or