We included adults 19–63 years living with HIV and enrolled in Virginia Medicaid who had a late ART prescription refill by >30–90 days (Box 1). We excluded those with coverage from other health service payers, including older adults who are insured through Medicare, and those whose claims indicated ART discontinuation due to toxicity or resistance.

The provider-level inclusion criterion was that their linked potential patient-participant(s) meets all selection criteria except having a late prescription; instead, the potential patient-participant has no history of a prior ART prescription. There were no provider-level exclusion criteria.

Cluster randomization occurred at the provider level and concurrently with initial potential participant screening (Figure 2). For each identified Medicaid enrollee with HIV, an HIV provider was identified and linked using the National Provider Identifier (NPI) from each HIV-related service or prescription claim, provider taxonomy, and provider frequency in the enrollee’s claims. To prevent contamination across study arms, HIV providers were cluster-randomized to the program or usual care arm. Contamination could occur in cases where patients assigned to usual care have providers who received the provider-level support (type II statistical error). Imbalance in observed and unobserved factors across study arms was minimized by stratified cluster-randomization based on provider characteristics (e.g., HIV patient volume) and patient characteristics (e.g., distribution of providers’ patients by age, sex, race, ethnicity, urbanicity).

For Screening Step 1 (initial eligibility, Figure 3), we used the Virginia Medicaid claims to identify a preliminary sample of potential patient-participants, defined as current Virginia Medicaid enrollees 19–63 years with an HIV care marker (i.e., at least one claim with an HIV diagnosis code, HIV-related service code or national drug code for ART). For Screening Step 2 (confirming eligibility), HIV status of potential patient-participants was confirmed using an extract of HIV surveillance data from VDH. An additional eligibility check determined if a more recent ART prescription may have come from state health department programs (e.g., Ryan White and AIDS Drug Assistance Program). Screening occurred monthly.

Once confirmed as eligible, potential patient-participants were assigned to a study arm (program versus usual care) based on cluster randomization of their HIV provider (Figure 2). Among eligible potential patient-participants assigned to the program arm, those with a late ART prescription by >30–90 days were assigned to the patient-level support component of the program arm; those with no late ART prescription were assigned to the provider-level support component of the program arm.

Recruitment of patient-participants in the patient-level support program arm occurred after screening and assignment and throughout trial implementation. Recruitment occurred over approximately two weeks each month. Recruitment of providers for the provider-level support component followed the same timing.

Deferred recruitment, taking place after the program arm’s 12-month follow-up, occurred for potential patient-participants in the usual care (control) arm and for patients of providers in the program arm.

A linkage coordinator securely faxed an introductory letter and study information sheet to eligible provider-participants. Within ten business days, the linkage coordinator called the provider’s office, confirmed fax receipt, and requested a short informational appointment with the provider or provider’s clinical staff to provide information about the study and answer questions.

Informed consent and HIPAA authorization procedures were conducted verbally by a linkage coordinator and followed identity verification (Table 1).

Providers interested in participating were verbally consented by the linkage coordinator. Post-consultation activities included retrospective consent for the eligible patient-participant (i.e., the patient of the provider-participant receiving the provider-level support).

Usual care procedures involved informed consent and follow-up. To minimize response bias, consenting procedures for use of the controls’ health information were deferred, identical to patients of supported provider-participants in the program arm.

Patient-level support in the program arm sought to improve timely filling of ART prescriptions among participants with late ART prescription refill(s) by >30–90 days. The program arm consisted of up to two phases (Phase I for persons >30–60 days late filling their prescription, Phase II for persons >60–90 days late), which were distinguished by the barriers primarily addressed in each phase (Table 1). Both phases involved phone consultations with a linkage coordinator to discuss barriers to ART adherence (e.g., medication side effects, lack of transportation); both provided referrals to resources that addressed participant barriers to timely filling of ART prescriptions. Discussions with Phase II participants focused on more complex barriers to filling ART prescriptions (e.g., substance use, unstable housing). These participants were additionally offered PositiveLinks, a warm health technology support using evidence-based mobile health platforms to promote better health for PWH.^16-20

Provider-level support in the program arm sought to improve ART prescribing by providers and providers’ knowledge about adherence support. Provider-level support involved a 30-minute peer-to-peer clinician consultation by phone. Consultations were delivered by HIV clinical experts (physicians, nurse practitioners and physician assistants) with a high volume of HIV patients and who provided HIV leadership in Virginia. Provider-level support involved: introductory letter and study information sheet, informational appointment and provider-participant consent, peer-to-peer consultation, post-consultation activities related to the provider-participant, and post-consultation activities.

Potential patient-participants assigned to the usual care arm (controls) received no study-specific intervention, which was consistent with state-level care for missing or late ART prescription refill(s) for Virginia Medicaid enrollees with HIV.

Outcomes were at the patient-participant level. The primary outcome was HIV viral suppression, defined as the most recent HIV RNA viral load <200 copies/mL at 12-month follow-up. In the usual care arm, 12-month follow-up was the first date an eligible control was confirmed as having no prior ART prescription or a late ART prescription refill(s) by >30–90 days; in the program arm, the 12-month follow-up date varied by support and phase (Table 2). The secondary outcome was adherence, defined as a proportion of days covered (PDC) of ≥90%. The PDC was a pharmacy claims-based metric reflecting the proportion of days for which a person has medication available during a measurement period. We additionally explored retention in care (defined as two or more care markers [e.g., CD4 count, HIV RNA viral load, medical visit] over 12 months and ≥90 days apart) and ART persistence (defined as the number of days between the first ART prescription claim following the date of enrollment and a late ART prescription claim >30 days).

A sample size of 1,000 patient-participants (500 per arm) provided adequate power to detect meaningful differences in HIV viral suppression across arms. We conservatively considered a power of 80% to detect a 10-percentage point difference between study arms, assuming a 50% rate of viral suppression in the control group.²⁸ Assuming 30% of the sample will be missing viral load data,^28-30 a sample size of 500 participants per arm was required. We anticipated peer-to-peer consultations for 15–40 providers randomized to the program arm and who had eligible patient-participant(s) with no prior ART prescription claim. No sample size estimate was calculated for provider-participants receiving peer-to-peer consultations; patient-participants of these provider-participants were included in the sample size calculation and were analyzed as part of the program arm.

Using descriptive statistics, we evaluated the balance of demographics, other patient characteristics, and outcomes between the study arms. We compared means and medians between the treatment samples using t-tests and Kruskal-Wallis tests.

We performed an intent-to-treat analysis to test hypotheses. We evaluated the effect of the program on the binary outcomes—including HIV viral suppression (primary outcome), adherence (secondary outcome), and retention (exploratory outcome)—by comparing the proportions between the program and usual care arms. We used Chi-squared tests to evaluate the statistical significance of the effects of treatments on proportions. We also used logistic regression models to test for program effects, adjusting for potential confounders. The differences in the means of persistence between the program and usual care groups were tested using a t-test, with a linear regression model used to evaluate the effect after adjusting for confounders.

Administrative claim databases are subject to data missingness.³¹ Multiple imputation is a common approach to address this limitation. We applied multiple imputation using predictive mean matching to address missing data, assuming missingness at random.³² Multiple imputation typically assumes that variables are normally distributed, but we used a predictive mean matching approach that does not make assumptions regarding the distribution of the data.³³

We revised the consent form and procedure to address discordance between self-reported ART adherence and adherence in the claims data. Early on, multiple potential patient-participants whose insurance claims data indicated a late ART prescription reported being up-to-date on their prescription and therefore did not need to enroll in the study. Given patient self-report of a late prescription or adherence challenges was never an eligibility criterion, we revised the consent form and procedure scripts to indicate that the study seeks to understand patient experiences related to ART prescription filling. This change did not impact our interpretation of the secondary outcome (ART adherence), as our assessment of adherence was based on the claims data and not self-report. Notably, we did not find an explanation for the discrepancy between the claims data and self-reports from potential patient-participants. We hypothesize that the discordance may be due to internalized stigma regarding prescription filling or adherence, unrecognized adherence challenges, and/or stockpiling of untaken medications—all potential barriers the AIMS program is designed to address.

We also addressed data limitations in the claims data. We updated eligibility criteria to exclude patient-participants with missing or incorrect contact information.

Finally, we revised the recruitment protocol to include SMS text messages for patient-participants. The text message contained a website link (https://adhere.vcu.edu/) with introductory text to describe the study and promote study credibility. Up to 2 texts per potential patient-participant could occur across recruitment months.

We removed the provider intervention after nine months during which no provider-participants were recruited (Figure 2). We found that providers and their clinic staff were frequently unable to talk or take messages unless in regard to a specific patient. Yet due to privacy and confidentiality protections, we were unable to reveal names of specific patients who had not received an ART prescription. In removing the provider intervention, any potential impact of provider engagement on patient outcomes was no longer part of the intervention arm. We updated our definition of the program to include the patient intervention only, such that we examined the effect of the newly defined program on the outcomes, versus usual care.

Given continued difficulties in recruiting patient-participants, we also relaxed some eligibility criteria (Box 1), adapted the study design to a new quasi-experimental design and reduced target enrollment to 100 patient-participants per arm (Figure 2). For the new quasi-experimental design, the usual care arm was changed to an artificial control constructed using propensity score matching. The individuals included in the new usual care arm were de-identified for analysis and therefore required neither consent nor HIPAA authorization. Follow-up in the revised study design was decreased to 3 months. To adjust for the effect decreased sample size would have on the study’s statistical power, we adjusted the analysis plan by lowering the target power to 70%, increasing the required effect size for detection to 25%, and adopting a one-sided test with a higher significance level (0.10).

AIMS, as an early example of D2C-Rx, was built on traditional D2C approaches by using Medicaid claims to identify PWH without ART prescriptions and offer tailored support interventions. Various programs have used pharmacy data or prescription claims to improve medication adherence,³⁴ including Medicare or Medicaid Medication Therapy Management Programs;^35-37 another D2C-Rx program, Link-Up Rx, has used pharmacy data in conjunction with HIV surveillance data to implement a D2C re-linkage program through a specialty pharmacy.³⁸ AIMS was the first program to our knowledge to target prescription adherence for Medicaid enrollees with HIV using claims data.

AIMS itself had novel features: Distinguishing PWH who have never been prescribed ART from those who have missed an ART prescription refill facilitated tailored support for prescription adherence barriers at the provider- or patient-level. Additionally, given ready availability of pharmacy claims, AIMS identified individuals with late ART prescriptions in nearly real-time and at precise intervals of lateness. Partnering with a state Medicaid program allowed for implementation statewide, versus through a single clinic, pharmacy, or managed care organization, and leveraged Medicaid’s status as the largest single provider of health insurance for this population nationally, covering nearly half (45%) of all diagnosed PWH.³⁹ Finally, use of Medicaid claims facilitated rapid assessment of prescription filling behavior after delivering the intervention—and if necessary, re-intervention in a phased approach.

A multi-stage data access and transfer procedure was created to satisfy institutional and agency requirements for data security, which in many cases were determined directly by state legislation.⁴⁰ Specifically, the Virginia Administrative Code protects information about Medicaid enrollees, allowing release of member information for certain activities such as plan administration (12 VAC 30-20-90), and the Code of Virginia governs allowable purposes for cross-agency data sharing (Va. Code § 2.2-3801). The project depended on matching identifiable data between Virginia Medicaid and VDH by an academic third party. In accordance with state legislation, as well as institutional and agency requirements, our process (Figure 3) allowed for release of individual-level information.

While the cluster-randomized controlled trial design can provide internal validity, we could not guarantee trial results would generalize to other states or health departments. This is important in the context of Medicaid, which is funded jointly by federal and state governments but is administered at the state level. State-level administration allows Medicaid agencies flexibility to implement new program improvement initiatives and tailor services to their state populations, but it also introduces variation between states that can affect generalizability of Medicaid programs. The AIMS program may be considered yet another example of the wide variability across states in Medicaid Medication Therapy Management programs.³⁷

First, while the program offered holistic patient support, it did not address or isolate the effects of individual interventions to decrease barriers to ART prescription and adherence, such as stigma or logistical and financial barriers. Thus, as originally designed, the study could not test differences between the provider-level and patient-level components of AIMS, nor of the different phases within the patient-level component. Study design changes limited our ability to incorporate and test the AIMS provider-level component as part of the full intervention. Further, decreasing the sample size would make detecting smaller differences, between study arms, challenging. However, we believe this holistic program of support best represents how such an initiative like AIMS would be implemented at state health departments or Medicaid agencies. The multi-level AIMS program is in line with recommendations for adherence interventions to address both provider- and patient-level factors⁴¹ and draws upon evidence-informed strategies to improve retention in care, such as motivational interviewing, removing structural barriers to care, and linkage to community resources at the patient level, as well as brief messages delivered by providers.^42,43 AIMS also combines elements of existing Medicaid practices and recommendations, including quality improvement initiatives and continuing medical education which address provider behavior⁴⁴ and Medication Therapy Management programs that focus on patient behavior.⁴⁵

Second, minor variation in the benefits of the different managed care organizations with which Virginia Medicaid contracts, as well as in community resource availability across jurisdictions, may alter the specific support offered.

We faced significant challenges with phone-based enrollment. These challenges necessitated a dramatic reduction in recruitment targets from a planned initial sample size of 500 to 100 per arm. Further developmental work, such as pilot studies and focus groups, may have provided insights to study design that prevented the need for such changes.