Patient preferences for long-acting HIV treatment: a preference heterogeneity assessment | BMC Infectious Diseases

Previous studies from our research team describe in detail the steps taken to design, pilot test and administer the DCE survey [17,18,19]. Here we present the most important features.

Population and setting

We recruited participants from five University of Washington (UW) HIV clinics in western Washington State and from the Grady Health System Ponce de Leon Center affiliated with Emory University in Atlanta, Georgia. Recruitment occurred between March 2021 and June 2022 via email and telephone communication using patient registries at each site or in-person outreach during clinic appointments. All participants were PWH who had established care at one of the research sites and were aged ≥ 18 years, fluent in English, and able to provide informed consent. We excluded persons deemed under the influence of drugs or alcohol when screened, persons currently taking long-acting cabotegravir/rilpivirine, and “elite controllers” (who do not require ART) [20]. Interested patients were screened for eligibility and subsequently consented using REDCap, a secure, electronic data-capture platform hosted by the UW [21, 22]. The final sample included 699 participants, 350 residing in Washington State and 349 in Atlanta. The sample size was informed by published guidelines [23] and by prior information obtained during pilot testing [18].

DCE design

We developed a three-alternative DCE including two unlabeled hypothetical LA-ART profiles and a current-therapy (opt-out) alternative. Each hypothetical LA-ART profile was defined by seven attributes: treatment type (LAO pills, subcutaneous injections, intramuscular injections, implants), location of administration (home, clinic, pharmacy), frequency of dosing (1 week, 1, 2, 3, 6, and 12 months), pain level (none, mild, moderate), pre-treatment time undetectable (i.e., the length of time that viral suppression would be required before LA-ART initiation: none, 3, and 6 months), pre-treatment negative reaction testing (needed or not), and late-dose leeway (i.e., flexibility in timing next dose before breakthrough viremia: short or long period). Participants were asked to assume equal effectiveness, safety, and cost across LA-ART options and their current therapy [19]. Attributes and levels were identified through a literature review and 12 key informant interviews [17]. The opt-out alternative represented the participant’s current ART regimen, which could be any number of oral pills taken at least once daily. Hypothetical LA-ART profiles were constructed so the attribute levels available for each mode of administration were realistic; Supplementary Material 1 shows the feasible combinations.

We used Ngene software (ChoiceMetrics, Sydney, Australia) to design the experiment using a modified Federov algorithm and D-optimal main effects [24]. The optimal D-error (d = 0.053978) for a three-alternative design based on the previously described attribute levels was attained by dividing the 699 participants into one of four randomly assigned blocks of 16 choice-tasks. We added a 17th choice-task for every participant, which presented two distinct types of long-acting oral regimens that were compared to the same constant opt-out. Thus, we gathered information from 65 unique profile comparisons, which provided adequate statistical power to estimate a preference weight for each attribute level.

DCE components and data collection

After providing informed consent, participants were transferred from REDCap directly to the DCE survey, which was self-administered using SurveyEngine [25], an online data collection platform specifically designed for preference research. In addition to the DCE, the survey collected sociodemographic, health-related, and clinical characteristics as detailed below. The full text of the survey can be found in Supplementary Material 2. Study personnel extracted clinical data from medical records.

Individual-level characteristics

Participant-characteristics included in the multinomial logit model were selected based on the findings from a previous study by our team that identified associations of individual characteristics with treatment preference [26]. Individual characteristics were grouped into three categories: sociodemographic characteristics (study site, sex at birth, age, race/ethnicity, income, education), health-related characteristics (clinic access, HIV stigma index, experience with medical injections, aversion to injections), and clinical characteristics (years on ART, viral suppression, AIDS diagnosis history, ART adherence [defined as self-report that participants always or almost always took medicines as instructed], mental health disorders, and substance use disorders). Clinic access was dichotomized as “very easy” or “easy” versus other responses (i.e., “neutral,” “difficult,” “very difficult”) to the question “In general, how easy is it for you to get to the clinic where you receive HIV care and treatment?” The HIV stigma index was based on six questions from a validated scale [27], each scored from 1 (no stigma) to 4 (high-level stigma) and then averaged across questions. Aversion to injections was defined as somewhat or strongly agreeing with the statement “I HATE getting injections and try to avoid getting them whenever possible.”

Data analysis

The DCE data were fitted to a latent class analysis (LCA) model to group the participant into preference groups (i.e., classes), whose preferences are similar within but systematically different between classes [28, 29]. This method allowed us to identify the presence of preference heterogeneity in the sample and classify participants accordingly [16]. We followed a model-selection process in which we tested models with 2 to 10 classes and chose the final model based on goodness of fit (GOF), the adjusted-Bayesian and Akaike information criteria, and qualitative meaningfulness of class differences to ensure that each additional class increased our understanding of the sample’s preference heterogeneity and to avoid overfitting. Consistent with our previous study, all models included an interaction between mode of administration and pain, to feature only feasible combinations, and the two injection modes (subcutaneous and intramuscular) were collapsed into a single mode [19]. We conducted the analysis using Stata statistical software (version 14.2, StataCorp LLC) [30].

After selecting the appropriate number of classes, we derived the posterior estimates of the probability of belonging to each class for each participant, which were used to determine each participant’s most likely class membership. We then used a multivariable, multinomial logit model to identify which individual-level characteristics (i.e., the explanatory variables in the model), including study site, were associated with class membership (i.e., the dependent variable, which takes as many values as the number of classes identified in the LCA). This multivariable model estimates relative adjusted risk-ratios (RRR), which are ratios of the proportion of participants with a given characteristic in any class relative to the proportion in the reference class [31]. Thus, we were able to describe each class based on their distinct LA-ART preferences and identify the characteristics of participants that differed by class, after adjustment for participant-characteristics including study site.