Sustained HIV viral suppression among men who have sex with men in the Miami-Dade County Ryan White Program: the effect of demographic, psychosocial, provider and neighborhood factors

Data from adult (18 years or older) gay, bisexual and other men who have sex with men enrolled in the Miami-Dade County Ryan White Program (RWP) Part A/Minority AIDS Initiative (MAI) [8] any time before January 2017 were used to conduct a cross-sectional study. Enrollment in the RWP was defined as having at least one medical case management encounter or peer education support network service during 2017. RWP clients are required to have a comprehensive health assessment every 6 months. Thus, individuals without a client health assessment in 2016 or 2017 were considered not enrolled and excluded. Additionally, those referred for ancillary services to the RWP by a non-RWP HIV provider, and those whose client file was closed because of movement to another county/state, financial ineligibility, or incarceration were excluded (11.9%). Male RWP clients were classified as gay, bisexual and other men who have sex with men (herein after referred to as “MSM”) if they reported any MSM behavior.

HIV viral load data were obtained from viral load laboratory test results. Sustained viral suppression was defined as having a HIV viral load < 200 copies/ml in all viral load tests in 2017. A viral load of < 200 copies/ml was chosen to stay consistent with the Centers for Disease Control and Prevention (CDC) definition for an undetectable viral load and to be able to compare to national studies [9]. If any test in 2017 was ≥200 copies/ml (not suppressed), they were considered not to have achieved sustained viral suppression in 2017. If an individual had only 1 viral load test in 2017 (or more than 1 test but the tests were less than 3 months apart) and the test showed viral suppression, we looked to the last viral load test in 2016 in an effort to assess consistent viral suppression on at least 2 tests [7]. Individuals with only 1 suppressed viral load test in 2017 and no viral load tests in 2016, or with no viral load tests in 2017 were excluded (n = 216; 6.0%).

Demographic data were obtained from the client intake assessment conducted at time of entry into the RWP. Demographic variables considered included age, race/ethnicity, birth region, and preferred language. Race was categorized into non-Latino Black, Latino, and non-Latino White. Due to small numbers of Haitians (n = 46), those reported as Haitian were categorized based on data about their race and Latino ethnicity (non-Latino Black = 43, Latino = 2, White = 1). Also because of small numbers, those with race other than Black or White (Asian = 21, Native American/Alaskan Native = 5, Native Hawaiian/Pacific Islander = 2), and those with unknown Latino ethnicity (n = 3) were excluded.

Psychosocial data were obtained from the first comprehensive health assessment conducted in 2017 and pertain to the clients’ responses at that time [8]. We chose the psychosocial variables based on the Behavioral Model for Vulnerable Populations [10] and what was available from the comprehensive health assessment, and grouped the variables into need and vulnerable/enabling characteristics. Need characteristics included AIDS diagnosis, and current HIV-related symptoms. Vulnerable/enabling characteristics included 30 variables related to mental health, substance and alcohol use, HIV status disclosure, employment/disability status, income, household structure, access to transportation, social services need, and social support (see Additional File 1 for details). Individuals with missing data on psychosocial indicators were categorized into the lower risk groups (e.g. no alcohol use). We developed indices of vulnerable/enabling factors to reduce collinearity by conducting reliability analysis and exploratory factors analysis, followed by confirmatory factor analysis. In reliability analysis, we removed 13 vulnerable/enabling variables which increased the Cronbach’s alpha from 0.33 to 0.62. Factor analysis using Varimax rotation yielded 5 indices: substance use, mental health, housing and transportation needs, unemployment, and food insecurity and low social support. Factor loadings are provided in Additional File 1; one variable was deleted due to a factor loading < 0.4 (experience of domestic violence). See Table 1 for variables included in each index.

To assess provider factors, we calculated the number of RWP clients by HIV physician and assigned this number to all clients who are primarily served by that physician. Additionally, we considered whether the client was enrolled in the Affordable Care Act (ACA) health insurance exchange. Neighborhood data were obtained from the 2013–2017 American Community Survey (ACS) by zip code tabulation area (ZCTA) and merged using client’s residential ZIP code [11]. The number of homicides for each ZIP code in Miami-Dade County was obtained from SimplyAnalytics [12]. A total of 25 neighborhood variables were considered (see Additional file 2). Variables were related to 5 categories: socioeconomic status, racial/ethnic composition, language and US nativity, residential instability, and violent crime. We developed indices as described above which yielded 2 indices: neighborhood deprivation, and residential instability and crime.

All analyses were conducted in SAS Version 9.4. Chi-squared (or Fisher’s exact test when applicable) for categorical variables and Wilcoxon signed-ranked test for continuous variables were used to compare sustained viral suppression by demographic, psychosocial, provider, and neighborhood characteristics in bivariate analyses. Three-level (individual, medical case management site, and neighborhood) cross-classified mixed-effects models were used to estimate adjusted odds ratios (aOR) and 95% confidence intervals (CI) for sustained viral suppression in multivariate logistic regression analyses using a residual pseudo-likelihood estimation technique in the PROC GLIMMIX procedure. Individuals were allowed to cluster by medical case management site and neighborhood, but these were mutually exclusive groups; the intraclass correlation coefficient (ICC) was calculated for each possible combination. We first fitted a cross-classified model that included demographic, provider and need factors, and indices of psychosocial and neighborhood factors (each index was tested as the combined standardized score of the variables in the index). Variable selection was guided by the Behavioral Model for Vulnerable Populations [10] and the literature. Variables in Table 2 were entered all at once into the model. Model fit was assessed by ensuring the ratio of the generalized chi-square statistic and the degrees of freedom was close to 1. For each psychosocial or neighborhood index that was significant in the first model, we fitted separate cross-classified models that included demographic, need and provider factors, and all other indices with each variable in the index of interest one by one. We examined the influence of outliers by inspecting plots and estimates of the Pearson residuals, the deviance residuals, and the DFBETAS. We assessed multicollinearity by examining the Pearson correlation coefficients between all variables in the model and Type II Tolerance values. We conducted sensitivity analyses to assess the influence of missing data by comparing results to models excluding individuals with missing data, and by setting the missing values at all possible levels of a given variable.