County-level estimates of prevalence rates (per 100,000), as well as the number of PLWH in 2011 were obtained from the National Center for HIV/AIDS, Viral Hepatitis, STD, and TB Prevention Atlas at the U.S. Centers for Disease Control and Prevention (CDC) (NCHHSTP Atlas, 2014). The analysis was based on 16 states in the US south: Alabama, Arkansas, Delaware, Florida, Georgia, Kentucky, Louisiana, Maryland, Mississippi, North Carolina, Oklahoma, South Carolina, Tennessee, Texas, Virginia, and West Virginia.

To ensure confidentiality of cases, county-level totals were suppressed for counties with populations of less than 100 or fewer than 5 PLWH and in accordance with approved state requirements under data agreements with CDC (NCHHSTP Atlas, 2014). To gain an overall picture of PLWH in the south and facilitate spatial analysis, we imputed the missing data for each data-suppressed county. To impute missing prevalence rates, we calculated state-specific average prevalence rates using the total number of reported cases and the total population size among reporting counties (U.S. Census Bureau American Community Survey: 2008–2012 5-Year Estimates, 2014).

We obtained data for the demographic variables from the 2008–2012 American Community Survey 5-Year Estimates (U.S. Census Bureau American Community Survey: 2008–2012 5-Year Estimates, 2014). County-level SDH variables obtained included poverty status (percent of county population below the federal poverty level), educational attainment (percent of county population with less than high school education), income (median household income), urban indicator (assigned counties a value of 1 if they had at least 50,000 inhabitants, and 0 otherwise), unemployment rate, percent black, percent white, and percent Hispanic (Census, 2013).

We used global and local clustering analyses to detect overall and specific clusters of elevated counts of PLWH, respectively. Two measures of global spatial dependence, Moran’s I index (Moran, 1950) and Geary’s contiguity ratio c (Geary, 1954) were used to provide an indication of the presence of spatial clustering. The Besag and Newell cluster (BNC) detection (Besag & Newell, 1991) was used to identify the locations and extent of clusters. The BNC method considers windows with a predetermined number of cases, k, representing the size of the cluster. The analysis was repeated for different values of k to detect recurring clusters. For each value of k, we selected the top nine clusters defined as having a prevalence ratio greater than two and the smallest p-values (all < .05). We then selected the counties that appeared at least twice over different values of k.

We considered three regression approaches to examine associations between PLWH prevalence and demographic variables. First, we treated log-transformed county-specific rates as a continuous outcome in linear regression models for each covariate separately and jointly. We then extended the linear regression model to include spatially dependent random intercepts to account for potentially spatially varying unmeasured confounders using a conditional autoregressive (CAR) model where spatial dependence is based on county neighborhoods (Waller & Gotway, 2004). Lastly we modeled the cases as counts using Poisson regression, under a Bayesian hierarchical framework with CAR random effects. We assume uninformative normal priors for the coefficients and uninformative inverse gamma priors for the variances. All three approaches were performed on all counties in the US south and on the clusters identified by the BNC. All statistical analysis was conducted using R3.1.1 software (R Core Team, 2014).

A total of 1422 counties were included with imputed data for 168 counties (11.8%). Figure 1 shows the county-level distribution of PLWH rates in the south. Counties with the highest prevalence rates were found in Texas (Walker), Florida (Union) and Maryland (Baltimore City). The prevalence rate per 100,000 for PLWH in the US south in 2011 was 307.9 with Florida having the highest rate of 446.9 (Table 1). Florida also had the second highest percent of Hispanics and the highest unemployment rate. West Virginia had one of the lowest PLWH rates (72.0 per 100,000) and the lowest percent of blacks and Hispanics.

Results of the global clustering indicate positive spatial dependence in the rates of PLWH: Moran’s index I = 0.258 (p < .001) and Geary’s contiguity ratio c = 0.632 (p = .07). The BNC analysis identified 78 counties with high prevalence of PLWH for multiple values of k = 50, 100, 300, 500, 1000, 15,000, 20,000, and 25,000 (Figure 2). The clusters capture select regions of Tennessee, Alabama, Georgia, Florida, Maryland, Virginia, and Delaware. Using nonparametric Wilcoxon rank sum tests, all of the race and SDH indicators, with the exception of the unemployment rates, in the clusters identified by BNC, were significantly different from the regional measures (p < .05). A Wilcoxon rank sum test also indicated that the prevalence rate of PLWH in the BNC clusters (846.8 per 100,000) was significantly different (p < .001) from the regional prevalence rate of 307.9 per 100,000.

Table 2 gives the estimated risk ratios (RR) and 95% confidence intervals for the univariate associations between SDH factors and PLWH rates obtained by standard linear regression and normal CAR models, as well as the associations between SDH factors and PLWH cases in Poisson CAR models. All models indicate that percent black was significant and positively associated with PLWH (p < .05). Poverty status, urbanicity, and unemployment were statistically significant and positively associated with the prevalence of PLWH in all the regression models. Income was not statistically significant in the standard linear regression model; however, when spatial effects were considered, it was negatively associated with PLWH.

Educational attainment was statistically significant and negatively associated with PLWH in the linear regression model, not statistically significant in the normal CAR model, and significant and positively associated with PLWH in the Poisson CAR model. Two factors may have contributed to the differences in association estimates between the normal and Poisson CAR model. First, the two models make different distributional assumptions on the observed log rates or counts. Second, while the coefficients obtained from both models can be interpreted as RR, the normal model estimates the difference in expected log rates and the Poisson model estimates the log of the difference in rates.

Many of the SDH factors were significantly correlated with each other. Educational attainment was significantly correlated with income; unemployment; poverty; and percent white, black and Hispanic (p < .05). To address multicollinearity issues, we retained only variables that were not significantly correlated with each other in the multivariate model: percent black and the urban indicator. Table 3 reports the results for the linear, normal CAR, and Poisson CAR models. Percent black and the urban indicator were both statistically significant and positively related to PLWH for the US south in all models. We performed multiple linear regression analysis on the spatial clusters identified by BNC. The results for percent black are similar to those for the entire region (Table 3).