Removal of Medicaid Prior Authorization Requirements and Buprenorphine Treatment for Opioid Use Disorder

This cross-sectional study assesses the association of removing prior authorization for buprenorphine to treat opioid use disorder with prescribing practices for Medicaid enrollees.


Main statistical analyses
Two-way fixed effects models: For our main two-way fixed effects (TWFE) difference-in-difference analyses, we estimated a log-linear regression model at the state-quarter level.For state i, in quarter t, we fitted: (1)     =  1   +  2   +  3   +  4   +   where Log Bup Rxit is the log number of buprenorphine prescriptions per 1000 Medicaid enrollees; PAit is 1 if a state removed their Medicaid prior authorization requirement for at least one formulation of buprenorphine, 0 otherwise; Xit is a vector of time-varying covariates: opioid overdose rate per 100,000 adults, log number of X-waivered clinicians qualified to prescribe buprenorphine per 100,000 population, number of individuals receiving methadone through an opioid treatment program per 100,000 adults (including quadratic term); percent of individuals aged 12 and older living below the federal poverty limit, Medicaid expansion status, percent of Medicaid enrollees covered under comprehensive managed care plans, and prescription drug monitoring program mandate (1 if yes, 0 if no); Statei and Yeart represent state and year fixed effects; and   is an error term.Because models with quarter fixed effects showed very similar results to those with year fixed effects, we utilized year fixed effects in the final version of our model to improve statistical power.Standard errors were clustered at the state level.The treatment effect for removing a buprenorphine prior authorization is given by the parameter  1 .To ease interpretation, effect estimates were exponentiated and can therefore be interpreted as the percent change in buprenorphine prescriptions per 1000 Medicaid enrollees after removing prior authorization, controlling for all other covariates in the model.
Difference-in-difference-in-differences ("Triple Difference") models: For models evaluating whether the effect of prior authorization removal differed according to baseline state characteristics, we estimated a log-linear regression model at the state-quarter level.For state i, in quarter t, we fitted: (2)     =  1   +  2   +  3   +  4   +  5   +  6   *   +   where all variables are the same as those listed above in equation 1, aside from Base which corresponds to baseline (i.e.quarter 1 of 2015) state-level variables hypothesized a priori to potentially modify the effect of prior authorization removal of buprenorphine prescriptions, including above vs below-median baseline buprenorphine prescriptions per 1000 Medicaid enrollees, Medicaid expansion status, and above vs below-median percent of Medicaid enrollees covered under comprehensive managed care plans.We again utilized year fixed effects in the final version of our models to improve statistical power.Standard errors were clustered at the state level.The parameter of interest was  6 which represents the difference in log buprenorphine prescriptions per 1000 Medicaid enrollees after removing a buprenorphine prior authorization in states with various levels of the baseline characteristics.Estimates were again exponentiated for ease of interpretation and can therefore be interpreted as the percent change in buprenorphine prescriptions per 1000 Medicaid enrollees, controlling for all other covariates in the model.

Bacon decomposition
The canonical two-way fixed effects (TWFE) model relies on several assumptions to identify the average treatment effect on the treated (ATT), including that the average outcome in both the treated and comparison groups would have followed parallel trends in the post-period even in the absence of treatment.Recent work has demonstrated that issues can arise with TWFE models when there are multiple time periods with variation in treatment timing. 11,12In these cases, the coefficients from a TWFE regression model may not be a weighted average of unit-level treatment effects as anticipated.Rather, in the setting of heterogeneous treatment effects, TWFE models can include desired comparisons between treated and not-yet-treated units, but also undesired comparisons between treated and previously treated units that can lead to biased estimates.To diagnose whether such issues are present in our study, we use the decomposition method developed by Goodman-Bacon. 12In short, this decomposition method allows us to visualize the various treatment group comparisons that give rise to our effect estimates, as well as the effect size and weight assigned to each treatment group comparison (see eFigure 7).

Sensitivity analyses to account for differential treatment timing
As highlighted above, TWFE models may produce biased treatment effect estimates in the setting of multiple time periods with variation in treatment timing.While the Bacon decomposition described above can help diagnose how big of an issue this is in our study, alternative estimators have recently been proposed that seek to minimize/eliminate the biases caused by variation in treatment timing. 11We employ two of these estimators to assess whether our findings are robust to these different approaches.
First, we use the estimator proposed by Callaway and Sant'Anna which calculates group-time average treatment effects using not-yet-treated states as the control group. 13This estimator excludes undesired state comparisons, which in our case would occur if a state that repealed a buprenorphine PA in a later year was compared to a state that repealed a buprenorphine PA in an earlier year.We employ this estimator in an event study framework controlling for opioid overdose deaths per 100,000 residents and Medicaid expansion status (additional covariate inclusion resulted in more extreme inverse probability of treatment weights) and calculate an overall ATT which averages treatment effects across all lengths of exposure to treatment (see eFigure 8 and eTable 2).
Second, we use the estimator proposed by Sun and Abraham which similarly excludes undesired state comparisons but uses last-to-be-treated states as the control group rather than not-yet-treated states as in the Callaway and Sant-Anna estimator. 14We again employ this estimator in an event study framework controlling for all covariates included in our main models, and calculate an overall ATT (see eFigure 8 and eTable 2).

Generalized synthetic control sensitivity analyses
As an alternative to the traditional difference-in-differences framework, we also employed a generalized synthetic control (GSC) method to estimate the effect of removing buprenorphine PA requirements.This method does not impose the parallel trends assumption required in difference-in-difference models and may better account for unobserved heterogeneity between treatment and control states. 15Synthetic control approaches reweight and combine information from control units into a "synthetic control" that matches the treated units in the preintervention period.The resulting synthetic control unit is then used to predict counterfactual outcomes in the posttreatment period which are then compared to the treated units to generate an effect estimate. 16The GSC method is an extension of the traditional synthetic control method that allows for multiple treated units and variation in treatment timing, as is the case with buprenorphine PA repeals.We employ the GSC model proposed by Xu which uses an interactive fixed effects model to predict post-PA repeal counterfactual buprenorphine prescriptions. 17We use a parametric bootstrap procedure clustered at the state level with 1000 iterations to generate confidence intervals.Models with and without time-varying covariates produced very similar results, so we present simplified models without time-varying covariates (see eFigure 9).Rhode Island and Illinois were excluded from analyses due to limited pre-period data.

Relevant R packages used
Decomposition of 2x2 difference in differences (Bacon decomposition): bacon v0.    a Models include the same covariates as the main difference-in-differences models:

eFigure 7 .
Decomposition of 2 x 2 Difference-in-Differences Estimates and WeightsPanel A: Difference-in-differences estimates for prior authorization removal in model with no covariates Panel B: Difference-in-differences estimates for prior authorization removal in full model with covariates © 2023 Christine PJ et al.JAMA Health Forum.

eFigure 8 .
Abbreviations: TWFE = two-way fixed effects a Models use the log of buprenorphine prescriptions as the outcome of interest.Event time refers to the number of quarters relative to prior authorization removal.TWFE and Sun and Abraham models control for all covariates in the main model of the paper.Callaway and Sant'Anna model controls only for opioid overdose deaths per 100,000 residents and Medicaid expansion status as additional covariate inclusion resulted in more extreme inverse probability of treatment weights.Model results are fairly similar across the estimators, with no significant pre-period non-parallel trends.

eFigure 9 .
Generalized Synthetic Control Sensitivity Analyses for the Effect of Prior Authorization Removal on the Mean Difference in Log Buprenorphine Prescriptions per 1000 Medicaid Enrollees aPanel A: Mean difference in log buprenorphine prescriptions (dark line) and 95% confidence intervals (shaded area) Panel B: Mean log buprenorphine prescriptions comparing average of treated states (black solid line) with counterfactual synthetic control average (blue dashed line)

Excluded states where fee for service and managed care are allowed to have different prior authorization policies n=16
states (GA, HI, IN, KY, MA, MN, ND, NH, NV, OR, RI, SC, SD, UT, WA, WI); n=288 state-quarter-years © 2023 Christine PJ et al.JAMA Health Forum.eTable 1.

Buprenorphine Medicaid Prior Authorization Status by State as of April 2021 1
© 2023 Christine PJ et al.JAMA Health Forum.© 2023 Christine PJ et al.JAMA Health Forum.© 2023 Christine PJ et al.JAMA Health Forum.156 eFigure 3.

Unadjusted Log Buprenorphine Prescriptions Over Time by State and by Prior Authorization Policy Status, 2015-2019 eFigure 5. Event Study Estimates of Percent Change in Buprenorphine Prescriptions per 1000 Medicaid Enrollees Comparing States That Repealed PAs vs Those That Maintained PAs in Sensitivity Analysis Restricted to States Where Medicaid FFS PA Policies for Buprenorphine Are Known to Apply to All Medicaid Enrollees and Including Illinois
a Same sample as main analyses in paper but also includes Illinois which is the only state contributing to the effect estimates after the 9th quarter of the follow-up period (n=24 states, 432 state-quarters).eFigure 6.

Event Study Estimates of Percent Change in Buprenorphine Prescriptions per 1000 Medicaid Enrollees Comparing States That Repealed PAs vs Those That Maintained PAs in sensitivity Analysis With Expanded Sample Not Restricted to States Where Medicaid FFS PA Policies for Buprenorphine Are Known to Apply to All Medicaid Enrollees
a Sample includes states where Medicaid FFS and MCO prior authorization policies are known to apply to all Medicaid enrollees, as well as states where prior authorization policies are allowed to differ (n=40 states, 720 state-quarters).

eTable 2. Summary of Sensitivity Analyses and Point Estimates for the Effect of Prior Authorization Removal on Buprenorphine Prescriptions per 1000 Medicaid Enrollees Using Different Methods
a Callaway and Sant'Anna model only controls for opioid overdose deaths per 100,000 residents and Medicaid expansion status as noted above.eTable 3.

Sensitivity Analyses for the Effect of Prior Authorization Removal on Buprenorphine Prescriptions Using Different Exposure Lag Periods a
Difference-in-difference estimates from two-way fixed effects models.All models control state fixed effects, year fixed effects, and the following state-level covariates: opioid overdose deaths per 100,000, log number of X-waivered buprenorphine providers, number of individuals on methadone through an opioid treatment program per 100,000, percent poverty, Medicaid expansion status, proportion of the state's Medicaid enrollees covered under a comprehensive managed care plan, and whether or not a state mandates use of a prescription drug monitoring program.Standard errors are clustered at the state level. a

Sensitivity Analyses for the Effect of Prior Authorization Removal on Buprenorphine Prescriptions per 1000 Medicaid Enrollees Using Different Study Samples and Prior Authorization Categorizations a
Difference-in-difference estimates from two-way fixed effects models.All models control state fixed effects, year fixed effects, and the following state-level covariates: opioid overdose deaths per 100,000, log number of X-waivered buprenorphine providers, number of individuals on methadone through an opioid treatment program per 100,000, percent poverty, Medicaid expansion status, proportion of the state's Medicaid enrollees covered under a comprehensive managed care plan, and whether or not a state mandates use of a prescription drug monitoring program.Standard errors are clustered at the state level.b Main models treat Alaska as a control state (never removed prior authorization).As explained in eTable 1, Alaska removed their prior authorization for the first 28-day fill of buprenorphine-naloxone in April 2017.This sensitivity analysis treats Alaska as removing their prior authorization in April 2017.c Main models treat Florida as control state (never removed prior authorization).As explained in eTable 1, Florida removed their prior authorization for the first 7-day fill of buprenorphine-naloxone in February 2018.This sensitivity analysis treats Florida as removing their prior authorization in February 2018.d Main models exclude Vermont which had very high use of buprenorphine per 1000 Medicaid enrollees, more than double the next closest state.This sensitivity analysis keeps Vermont in the sample.