As an overview, we employed a multi-step process and three different data sources to address the study objectives. In the first step, discrete choice modeling informed the parameters included in a supply access equation. Second, we constructed the supply access equation, which described spatial access to the supply of Atlanta HIV providers by zip code tabulation area (ZCTA) as a function of factors determined by results from the discrete choice model. Finally, the supply access equation was applied to quantify two study outputs: (1) population-based estimates of supply access by ZCTA and (2) estimates of underserved communities in the Atlanta area. For this study, Atlanta was defined by the six county area, which included Clayton, Cobb, Dekalb, Douglas, Fulton, and Gwinnett counties. A schematic of this three step process, along with the data sources used in each step, is highlighted in Fig. 1. Information on the analytic methodology employed in this analysis are outlined in subsequent sections, and more details on model assumptions are discussed in Additional file 1: Appendix S1.
Data sources
HIV care provider database
We created a database of key characteristics about each of the major clinics or practices located in the six county Atlanta area, identified through multiple sources of data (including the Southeast AIDS Training and Education Center Key Contacts book, the Ryan White medical provider directory available through the HRSA data warehouse, the AIDS.gov resource directory, and the Georgia Care and Prevention in the United States (CAPUS) resource hub). We called every practice or clinic to verify that HIV primary care was provided on-site by at least one physician, physician’s assistant, or nurse practitioner. After excluding locations that did not meet this definition, 41 HIV clinics and practices remained. Information on facility type, accepted payment options, availability of ancillary services, and the number of available providers and weekly appointment hours was collected from each practice or clinic.
We classified each facility as a private practice, clinical research facility, community health center or community-based service organization, or state or local health department. To assess patient eligibility based on payment options, we asked each clinic whether it accepted private health insurance, Medicare, or Medicaid as forms of payment, offered a discounted pay structure for those who qualified based on income, and provided Ryan White services to patients. First established in 1990, the Ryan White HIV/AIDS program is a federal grant system which works with local providers to cover the cost of medical care and support services for persons living with HIV who would otherwise be unable to afford such services. Data were also collected on whether or not clinics offered the following ancillary services: HIV case management, mental health services, dental care, substance abuse treatment, transportation assistance, and an on-site pharmacy that fills prescriptions. The number of part-time and full-time providers and weekly appointment hours were quantified and used to estimate the number of provider-hours available to clients per week. We computed descriptive statistics on the availability of weekly provider-hours, accepted payment types, and availability of ancillary services overall, and by provider type. We evaluated differences in these characteristics by provider type using a Mantel-Haenszel chi-square test for categorical variables and a Wilcoxon-Mann-Whitney test for continuous variables.
The Engage Study
The Engage Study investigated structural and psychosocial barriers to HIV care among self-identifying HIV-positive men who have sex with men (MSM) living in the Atlanta area. Details on recruitment have been previously described (Dasgupta et al. 2014). All participants completed an online questionnaire that collected data on residential address at the time of the interview, the location of the last attended HIV care provider, and primary modes of travel taken to attend appointments. We used this information to estimate travel distance and commute time to attend HIV care visits based on reported mode of travel (by car versus public transit) using the Google Maps Direction application program interface (API) (Dasgupta et al. 2014, 2015b). Latitude-longitude coordinates for residence were anonymized before being entered into Google maps to protect confidentiality of participants. Only participants who were living in the six county Atlanta area and reported receiving HIV care from a provider in the six county area in the previous year were included in the analysis. Descriptive statistics on the sample included in the analysis are described.
AIDSVu
We obtained HIV case counts by ZCTA in the six county Atlanta area from AIDSVu.org, an online mapping tool which illustrates rates of HIV for several cities across the United States. Data reflected the number of prevalent cases reported through the end of 2011. Only data for ZCTAs whose population-weighted centroids were contained in the six county area were included in this study.
Discrete choice modeling
The Engage Study data were used in a discrete choice model to estimate characteristics most important in selecting a provider, by mode of transportation taken to attend HIV care visits. Characteristics spanning four of the dimensions of spatial access (availability, accessibility, affordability, and accommodation) were considered in describing the supply environment, and thus, were contributors to the individual’s choice utility function. Facility-specific characteristics included: provider type (private practice versus other), whether or not any ancillary services were offered (e.g., transportation assistance, substance abuse treatment), payment options (whether or not Ryan White patients were accepted), the number of available provider-hours during the week, and whether or not walk-in hours were offered. Two covariates directly related to the participant were also considered as potential descriptors of supply access, including travel distance between study participant residence and each HIV provider and the mode of transportation taken to attend HIV care visits. Because travel distance and provider-hours were non-normally distributed, the natural-log of these two variables were included in the discrete choice modeling and evaluation of supply access.
We employed a generalized estimating equations (GEE) model with a logit link function to investigate factors associated with choosing an HIV care provider. We accounted for clustering by patient using an exchangeable correlation structure. Because assessing potential differences in choice of HIV care provider by travel mode was of primary interest, we included two-way interaction terms between each provider characteristic and mode of transportation taken to attend HIV visits. Backward selection was used to determine which variables should be retained in the final model, using a cutoff of p < 0.05. The multivariable GEE model was built using SAS 9.4 (Cary, NC).
Development of supply access equation
Spatial gravity modeling assumes an exponential relationship between travel distance and access to care, such that increasing travel distance results in decreased access (Crooks and Schuurman 2012). The supply access equation represented a modified gravity model that generalized spatial access to HIV care across the entire Atlanta study area based on regression coefficients from the discrete choice model. These coefficients were used as corresponding weights of importance for each characteristic included in the equation. A single score generated from the equation represented supply access from the population-weighted centroid of a given ZCTA to an individual HIV provider, given the sum of individual provider-related characteristics, road distance between the centroid and provider, and the mode of transit.
Application of supply access equation
Study output 1: computing population-based estimates of supply access by ZCTA
For every ZCTA, 41 supply access scores were generated for travel by public transit, and 41 supply access scores were generated for travel by car. For each mode of transit, these individual supply access scores for ZCTA-provider pairs were summed by ZCTA, presenting a single estimate of the average supply access to HIV care providers available for a given ZCTA. Summary scores were computed separately for travel by car and travel by public transportation. Based on previously published results, we also transformed supply access scores to account for the barrier to HIV care attendance associated with traveling by public transit, versus by car, among people living with HIV in Atlanta (Dasgupta et al. 2015b). The differences in supply access scores between two modes of travel were highly sensitive to the estimate we used to transform scores. Thus, we also present results from a sensitivity analysis which explores potential consequences of using different values to transform spatial access scores in Additional file 2: Figure S1.
Using ArcGIS 10.2, supply access scores were mapped for travel by car and by public transportation. The geographic distribution of scores were compared across travel mode using quintiles of supply access for travel by car. Differences in scores across urban versus suburban and rural areas for each travel mode were also assessed. Urban areas were defined as areas inside the main auxiliary highway, or loop route, of Atlanta, while suburban and rural areas represented neighborhoods outside this highway.
Study output 2: identifying underserved areas
After evaluating supply access to HIV care providers by ZCTA, potentially underserved areas in the six county area were identified by mode of transit. We defined underserved areas as ZCTAs with overlapping areas of low supply access (one of the two lowest quintiles) and high HIV case count (one of the two highest quintiles). Subsequently, we estimated the proportion of all HIV cases in Atlanta living in these underserved areas for each mode of travel. Population-based estimates of poverty by ZCTA (areas in which >20 % of the population are living in poverty) were obtained from the American Community Survey (five year estimates, 2009–2013) to quantify potential overlap with underserved areas.