Data
This study involved a retrospective, multiyear, cross-sectional analysis of OUD treatment episodes that accounted for EDT between the clients’ ZIP code and treatment facility. We analyzed discharge survey data from 22,587 OUD treatment episodes in Los Angeles County with client ZIP codes. We relied on client administrative data from the Los Angeles County Participant Reporting System. The data came from a parent study funded by NIDA (R33 DA03563401) that focused on SUD treatment programs that served communities with more than 80% Latino or African American residents in Los Angeles County. The multiyear cross-sectional data included 12,247 clients aged 12 or older served by 125 unique SUD treatment programs. This sample included 96 (76.8%) SUD programs that offered outpatient counseling services to clients with OUD and 32 (25.6%) outpatient programs that offered methadone (Note: some programs offer both). These two types of programs serve more than 95% of all clients entering publicly funded OUD treatment in Los Angeles County. The analysis was conducted at the episode level such that each client’s characteristics and EDT were included. We analyzed two mutually exclusive services: medication-assisted treatment (i.e., methadone) and nonmedication outpatient counseling.
Geographical variables
We geocoded each client at the population-weighted centroid of their reported ZIP code using ArcGIS Pro (ESRI, 2021). We calculated the population-based centroid of ZIP codes using census block-level population data [22]. We considered using ZIP Code Tabulation Area (ZCTA – a generalized spatial representation of ZIP code service areas). However, we found that using the higher resolution ZIP codes is more precise for geolocating clients and that only 0.59% of study sample episodes would be mapped to larger ZCTA using cross-walk files. Concerns over using ZIP code are more common when creating spatial aggregates. Because our study objectives are to determine the role of travel time on client-level experiences, we expect less problematic experiences with ZIP codes.
EDT via automobile was determined from the ZIP code centroid to the treatment facility using Google Maps Distance Matrix API [23]. We calculated the EDT during mornings (9:30 a.m.) on weekdays. We included traffic in EDT calculations, unlike studies reviewed by Kelly et al., none of which accounted for traffic congestion [20]. Traffic congestion ought to be accounted for when estimating driving time in densely populated metropolises like Los Angeles, where travel time by car varies mainly depending on the time of day. The EDT from the Google Maps Distance Matrix API represents the best-guess estimate which includes historical traffic patterns during the specified date and time. Therefore, the drive-time estimate obtained for each episode includes traffic congestion by default. Additionally, we are able to obtain an optimistic drive-time estimate – or an estimate of driving time with minimal traffic. We use the difference between optimistic driving time and best guess driving time to quantify the proportion of driving time attributable to traffic congestion. We compared the EDT with traffic obtained from Google Maps Distance Matrix API for 100 randomly selected episodes to ESRI road network and traffic [24] to assess the validity of the seimates. Additionally, the single-mode transportation approach of focusing on travel time by car used in this analysis has been found to predict a similar pattern of health care accessibility compared to multimodal approaches, with a high correlation observed between single-mode and multimodal accessibility rates [25].
Treatment outcome variables for aim 2
The key outcome variable for Aim 2 was treatment plan completion based on six official discharge codes. The first two codes evaluated whether the client completed the treatment or recovery plan or was referred or transferred, whereas the next four codes defined clients who left without making progress, died, got incarcerated, or other discharge status [26]. For the first outcome, we coded participants as 1 if the clinician reported the client completed the treatment or recovery plan for that episode and 0 if not. These measures have been used to evaluate treatment completion in regional [27,28,29] and national [30,31,32,33,34] studies. They do not include information on the number, type, or description of the treatment plan or its goals.
Explanatory variables
The independent variables of interest included clients’ self-reported sex, measured as a dichotomous variable (1 = female, 0 = male). The study also examined race and ethnicity, using categories of Latino or Hispanic, Black or African American, non-Latino White, and other. We coded the category “other” to represent clients identifying as American Indian, Asian, or another race and ethnicity because our data did not have sufficient clients to analyze these groups separately. Clients also reported demographic and socioeconomic variables including age, education (completing high school), Medi-Cal eligibility, veteran status, and referral source.
Statistical analyses
Statistical analyses were run using R statistical software to address each aim [35]. Specifically, group comparisons were used to study patterns and disparities in EDT. For instance, t-tests were used to determine if men and women have different average EDTs for each service type. A similar approach was used to compare age groups, Medi-Cal clients, and other study covariates. We constructed generalized linear models to assess the significance of the association between study covariates and EDT while adjusting for potential confounders. Transformations were applied to the dependent variable (EDT) to minimize violations of linear regression assumptions.
For the study’s second aim, multiple statistical techniques were employed to examine the association between EDT and treatment outcomes. To ensure that EDTs did not convey false precision, we estimated a secondary categorical variable for EDT: less than 10 min, between 10 and 20 min, between 20 and 30 min, and more than 30 min. Then, we constructed a logistic regression model to examine the association between EDT categories and odds of completing treatment. The multivariable logistic regression model allowed for isolating the effects of EDT and potential interactions with study covariates. A complete model was first constructed, then models were constructed in an explorative manner such that variables and interactions were iteratively added and assessed for significance.
The data studied in this analysis had an average missing rate of less than 10%. Observations were used whenever the variables of interest in the analysis step were available (pairwise deletion) to maximize the sample size in each analysis step. Missing data analysis was conducted to ensure the randomness of missing data through several packages in R statistical software [35].