The Association of Provider and Practice Factors With HIV Antiretroviral Therapy Adherence

David J. Meyers, MPH; Megan B. Cole, PhD, MPH; Momotazur Rahman, PhD; Yoojin Lee, MS, MPH; William Rogers, PhD, Roee Gutman, PhD; Ira B. Wilson, MD, MSc, FACP


AIDS. 2019;33(13):2081-2089. 

In This Article


Data Sources

Our primary data source was Medicaid Analytic Extract (MAX) claims from 2008 to 2012. MAX data provide 100% of Medicaid claims from each state, and are the only currently available source of Medicaid data that is unified across states. While our dataset ends in 2012, this represents the most recent year of complete data from these states as of the time of this study. For our analysis, we had data from 14 states (California, Florida, Georgia, Illinois, Louisiana, Massachusetts, Maryland, North Carolina, New Jersey, New York, Ohio, Pennsylvania, Texas, Virginia). These states were chosen because they are the 14 states with the highest HIV prevalence, accounting for approximately 75% of all HIV cases in the United States. We used the MAX Prescription Drug file to determine use of ART therapy, the Other Therapies file to determine use of outpatient services, primary care, and specialist office visits, and Inpatient Therapies (along with Other Therapies files) to determine a comorbidity count. We used the National Provider Identifier (NPI) registry file[13] published by the Centers for Medicare and Medicaid Services to attribute patients to their primary HIV care providers (PCPs-HIV), and for provider characteristics.

Study Sample

Our sample included patients aged 18–64 with HIV in the 14 states from 2008 to 2012. We classified a patient as having HIV if they had at least two claims at least 30 days apart that listed HIV as a diagnosis, or if they used ART, as ARTs are seldom prescribed for patients without HIV. This classification method has been used in prior work.[14,15] We excluded patients who were dually enrolled with Medicare, as we did not have Medicare claims available for this study, and those that were enrolled in Medicaid managed care, as the MAX claims may not contain complete managed care claims during the time period of interest. We required at least 6 months of Medicaid enrollment in a year to be included in our sample. Note that because patients were included based on ART use, only patients judged by their clinicians to be eligible for ART were studied.

Provider and Practice Attribution

For each calendar year of the study, we assigned patients to a PCP-HIV. To accomplish this, we designed a multistage attribution process based on similar approaches used in the literature.[16] We assigned patients to the providers that provided the plurality of their ART prescriptions and primary care visits. We then attributed providers to practices on the basis of information from the NPI registry file. A detailed description of the attribution process can be found in Appendices 1 and 2, In this study, we define providers as either physicians, Physicians Assistants, or Nurse Practitioners.


Our primary outcome of interest was the percentage of the calendar year that PLWH were exposed to ART. To calculate this, we use days supplied and fill dates from the MAX prescription file to count the number of days in the year the patient had ART available. This number became the numerator. The denominator was 365 days if the patient was alive and enrolled for the whole year (e.g. if the patient was enrolled in Medicaid for half the year, we would only count 6 months toward the denominator). This is often referred to as Proportion of Days Covered (PDC). More details on the calculation of the measure are found elsewhere.[17]

Patient and Provider Characteristics

Patient characteristics included age, sex, race/ethnicity, and basis of Medicaid eligibility. Patients were classified as having a chronic condition if they had two outpatient codes, or one inpatient code, for a diagnosis. We also classified patients as having a problem with drug use, alcohol use or both. Importantly, substance use claims were not redacted from our data. From the prescription file, we included a flag for what ART regimen the patient was primarily receiving during the year (nucleoside reverse transcriptase inhibitors, non-nucleaoside reverse transcriptase inhibitor, Boosted protease inhibitor, integraise, multiple/others) as regimen may be related to adherence. Based on the attribution, we further included the consecutive number of years the patient was attributed to the same provider.

At the individual-provider level, we included variables from the NPI registry file for primary taxonomy/specialty (Generalist vs. Infectious Disease vs. Other Specialty) and credential (MD/DO vs. Nurse practitioner/physician assistant/other). A full list of specialty classifications is included in Appendix 6, We also included number of patients that were attributed to a given individual-provider as a measure of experience treating patients with HIV, recognizing that the count of Medicaid patients only partially captures a provider's HIV patient load. At the practice-level, we examined the number of patients and number of providers attributed to that practice.

Statistical Methods

Our primary unit of analysis was person-year. We examined the distribution of patients' adherence at the provider and practice-level. Based on the mean adherence of patients attributed to a provider or practice, we plotted the distribution of adherence across providers or practices. We then fit a linear model adjusting for age, sex, race/ethnicity, regimen, chronic conditions, substance use, and state, year, and provider or practice fixed effects. From this model, we calculate the adjusted adherence level for each provider or practice and also plotted this distribution.

In our primary analysis, we fit multiple hierarchical models on the patient-year level data with PDC as the outcome, and with provider and practice identifiers specified as random effects in nested multilevel models. The use of random effects in multilevel models allowed us to account for the natural clustering of patients within providers and practices, and enabled us to partition the variance in patient outcomes that can be attributed to the patient vs. the provider.[17–23] We fit two unconditional models, the first with provider random effects alone, and the second with practice random effects alone. Next, we included both provider random effects nested within practice random effects. We then added provider and practice characteristics detailed above, and in our final model also included patient characteristics and a patient random effect to account for multiple observations from the same patient over time. From each subsequent model, we calculated the variance partition coefficient, which is the percentage of the total variance in the outcome explained by the provider or practice, accounting for all other included patient characteristics. We assessed how much variance was due to the provider practice, or patient, and using the full model, measured the variables that were most associated with adherence outcomes. All models were linear models with state and year fixed effects to adjust for differences in state Medicaid programs, and used robust standard errors. As a sensitivity check, we log transformed the outcome and found similar results. To determine if our results were not purely driven by differences in geography, we also fit a model with zip code fixed effects to control for geographic access to different types of providers. An alpha of P less than 0.05 was considered statistically significant. All analyses were conducted in Stata Version 15 (StataCorp, College Station, Texas, USA).