We used inpatient discharge data from nonfederal community hospitals in Arkansas, California, Florida, Georgia, Iowa, Maryland, Massachusetts, Nebraska, Nevada, New York, Tennessee, Vermont, and Wisconsin from the Agency for Healthcare Research and Quality (AHRQ) Healthcare Cost and Utilization Project (HCUP) . We included states with encrypted patient linkage numbers to link records from the 2013, 2014, and 2015 (quarters 1 through 3) HCUP State Inpatient Databases (SID)  during the study period.Footnote 1 We also included only those states with data indicating whether diagnoses were present on admission (POA) to exclude index stays (i.e., initial hospitalizations) involving an opioid diagnosis that may have occurred solely because of hospital-related factors, such as iatrogenic complications of opioid use.
We obtained state-level data about the status and specific implementation dates of naloxone standing orders and Good Samaritan laws from The Policy Surveillance Program: A Law Atlas Project [6, 7]. For state Medicaid MAT policies, we could determine the status of these policies for the 2013–2014 period, but specific implementation dates were unavailable. Our key sources of MAT Medicaid policy information included ASAM state reports [4, 15,16,17,18,19,20,21,22,23,24,25,26,27], a 2016 article by Grogan and colleagues about state Medicaid MAT benefits , and personal communication with the Grogan article authors. When data from these sources were incomplete, we used several supplementary sources including two state Medicaid Preferred Drug Lists [28, 29], contacts at five state Medicaid agencies, and a Kaiser Family Foundation (KFF) report on rehabilitative services .
We gathered information about presence of hospital detoxification and psychiatry units from the American Hospital Association . For each data year, we obtained the state population capacity of facilities to treat SUDs (facilities offering care for SUDs including outpatient, residential, and inpatient hospital treatment for all payer categories), number of OTPs, and number of providers newly certified to administer buprenorphine/naloxone from the Substance Abuse and Mental Health Services Administration (SAMHSA) National Survey of Substance Abuse Treatment Services [32,33,34] and the SAMHSA Number of Drug Addiction Treatment Act (DATA)-Waived Practitioners Newly Certified Per Year tracker . We also used Grogan and colleagues (2016)  to obtain data about state Medicaid coverage of ASAM-recommended SUD treatment levels from 2013 through 2014. Finally, we obtained state rates of opioid overdose deaths for each data year from the Kaiser Family Foundation State Health Facts database .
The study population comprised a retrospective longitudinal sample of patients aged 18 years and older with an opioid-related index hospitalization between April 2013 and June 2015 and no preceding opioid-related hospitalization within 90 days.Footnote 2 Opioid-related stays were identified by any-listed International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) diagnosis codes that were present on admission for opioid abuse or dependence alone or in combination with other drugs (304.00–304.02, 304.70–304.72, 305.50–305.52) or for poisoning by opium, methadone, heroin, opiates and related narcotics, or opiate antagonists (965.00–965.02, 965.09, 970.1). We also included external cause of injury codes (E codes) for accidental poisoning by opium, methadone, heroin, and opiates and related narcotics (E850.0–E850.2) and adverse effects of heroin, methadone, opiates and other narcotics, and opiate antagonists (E935.0–E935.2, E940.1). We included any-listed opioid diagnoses for our index stays to capture the potential population of individuals who could be affected by naloxone, Good Samaritan, and MAT coverage policies. We excluded index hospitalizations in which the patient died or was transferred into or out of the hospital.
Outcome variable: Readmissions
The outcome variable was a readmission within a 90-day period with an opioid-related principal diagnosis or an opioid-related accidental poisoning or adverse effect diagnosis (E code). Consistent with other studies, we selected 90 days as the follow-up period because it would be sufficient time for patients discharged from the hospital to access potential outpatient rehabilitation services [37, 38]. This limited readmissions to hospitalizations that potentially would be most affected by our state policies of interest and excluded hospitalizations in which opioid-related diagnoses were only a secondary concern.
Key independent variables: State policies for opioid treatment
The key independent variables focused on the three state policies: (1) naloxone standing orders, (2) Good Samaritan laws, and (3) Medicaid MAT coverage and generosity.
The first key independent variable indicated whether a state had a naloxone standing order that allowed pharmacies to dispense naloxone without an individual provider prescription. The second indicated whether a state had a Good Samaritan law granting immunity to users from arrest, charge, or prosecution for possession of drugs or drug paraphernalia. For these first two independent variables, we classified an index stay as having a naloxone standing order or Good Samaritan law if the date of implementation was before or on the date of the index stay [6, 7].
The final two key independent variables represented Medicaid MAT coverage and generosity. Because the states in our sample had little variation in naltrexone and buprenorphine/naloxone coverage, we focused on two components of MAT coverage: whether a state had any coverage of methadone for Medicaid enrollees and whether a state had more or less generous coverage of buprenorphine/naloxone or naltrexone for Medicaid enrollees. Generosity of coverage of buprenorphine/naloxone or naltrexone was a composite variable based on the following four measures: (1) prior authorization requirement for buprenorphine/naloxone, (2) prior authorization requirement for injectable naltrexone, (3) buprenorphine/naloxone dosing limits (either limitation on total days coverage or maximum dosage restrictions of less than 24 mg/day), and (4) requirement for SUD counseling prior to treatment with buprenorphine/naloxone or naltrexone. If a state lacked restrictions on at least two of these four measures, it was categorized as more (vs. less) generous. As noted above, the Medicaid methadone coverage and MAT coverage generosity measures were based on data collected from a combination of sources. Because these data sources did not provide exact dates of implementation, we counted a state as having coverage in place during the years in which our sources collected the data (2013–2014) [4, 9, 15,16,17,18,19,20,21,22,23,24,25,26,27].
The analysis included covariates for patient-level factors, hospital stay and hospital characteristics, and state-level factors—all measured at the time of the index stay—that could have influenced the outcome of 90-day readmission. Patient-level factors included sociodemographic characteristics: age (continuous variable), sex, race/ethnicity (White, Black, Hispanic, other, missing), expected primary payer (Medicare, Medicaid, private insurance, uninsured/self-pay, other), community-level income based on the state-defined quartile for median household income of the ZIP Code of the patient’s residence, and urban/rural residency. To examine severity of illness, we identified whether the primary reason for the admission (principal diagnosis) was an opioid-related diagnosis or a non-opioid-related diagnosis, whether the patient was admitted with an opioid use disorder diagnosis (304.00–304.02, 304.70–304.72, 305.50–305.52) or a poisoning/adverse effect diagnosis (965.00–965.02, 965.09, 970.1, E850.0–E850.2, E935.0–E935.2, E940.1), and whether the patient had a continuous opioid use disorder (304.01, 304.71, 305.51). We used the HCUP Clinical Classifications Software (CCS) diagnosis categories  to define any-listed co-occurring mental health conditions (CCS 650–652, 655–659, 662, 670) or alcohol-related conditions (CCS 660). We used the Elixhauser Comorbidity Software  to create dichotomous variables indicating whether the stay involved a specific co-occurring physical (medical) condition(s); the count of the co-occurring physical condition(s) also was included.
Hospital stay characteristics included whether patients received any treatment for drug rehabilitation or detoxification during the index stay (ICD-9-CM procedure codes 94.64–94.69) and the length of the index stay. Because SUD treatment varies across hospitals, we included covariates for characteristics related to the hospital in which the index admission occurred: percentage of hospital discharges among patients with opioid-related conditions and whether the hospital had a SUD detoxification or psychiatric acute care unit.
State-level factors included the following measures of capacity for opioid treatment that were based on the year of the index stay: Medicaid coverage of all four levels of ASAM-recommended treatment services (outpatient, intensive outpatient, inpatient, intensive inpatient) during the 2013–2014 period ; newly certified provider capacity for buprenorphine/naloxone therapy, defined as the number of newly eligible DATA-waived practitioners approved to provide buprenorphine/naloxone treatment in a state per 100,000 population ; the number of OTPs per 100,000 population; and the number of SUD treatment facility beds per 100,000 population [32,33,34]. We included covariates for year of index stay, source of admission (emergency department vs. direct admission), and state overdose death rates to denote state-wide severity of opioid use. Because we hypothesized that the state MAT policies would have a direct effect on the Medicaid population and spillover effects on other insurance populations, we included interaction terms for Medicaid MAT coverage and generosity with each payer group.
We first conducted bivariate analysis to examine characteristics of our sample as well as the association between our key independent (policy) variables and the outcome variable, opioid-related readmission within 90 days after discharge. Next, we conducted multivariate logistic regression analysis to estimate the association between our key independent variables and an opioid-related readmission, taking into account the patient-, hospital- and state-level factors described above.
All data were analyzed using SAS version 9.4. The HCUP databases are consistent with the definition of limited data sets under the Health Insurance Portability and Accountability Act Privacy Rule and contain no direct patient identifiers. The AHRQ Human Research Protections Program has determined that research using HCUP data has exempt status.