Robustness of estimated access to opioid treatment providers in rural vs. urban areas of the United States

Introduction

Reproducible code for our forthcoming paper, Robustness of estimated access to opioid treatment providers in rural vs. urban areas of the United States, which compares the impact of hypothetical changes in physical access to opioid treatment across urban and rural counties. The full citation is:

Kiang MV, Barnett ML, Wakeman SE, Humphreys K, & Tsai AC, Robustness of estimated access to opioid treatment providers in rural vs. urban areas of the United States, Drug and Alcohol Dependence (September 2021), doi: 10.1016/j.drugalcdep.2021.109081

Abstract

Background Effective, evidence-based treatments for opioid use disorder are not equally accessible to Americans. Recent studies have found urban/rural disparities in the driving times to the nearest opioid treatment providers. These disparities may be even worse than currently reported in the literature because patients may not be able to obtain appointments with their nearest provider. We examine the robustness of the opioid treatment infrastructure by estimating how driving times to treatment change as provider availability decreases.

Methods We used public data from the federal government to estimate the driving time from each census tract centroid to the nearest 15 treatment providers. We summarized the median and interquartile range of driving times to increasingly distant providers (i.e., nearest, second nearest, etc.), stratified by urban/rural classification.

Results The median driving time to the nearest provider was greater in rural areas than urban areas for both opioid treatment programs (12 minutes vs 61 minutes) and buprenorphine-waivered prescribers (5 minutes vs 21 minutes). Importantly, driving times in rural areas increased more steeply as nearer providers became unavailable. For example, the increase in driving time between the nearest provider and the fifth nearest provider was much greater in rural areas than in urban areas for both buprenorphine-waivered prescribers (23 minutes vs 4 minutes) and for opioid treatment programs (54 minutes vs 22 minutes).

Conclusions Access to treatment for opioid use disorder is more robust in urban areas compared with rural areas. This disparity must be eliminated if the opioid overdose crisis is to be resolved.

Issues

Please report any issues via email or this repo.

Structure

• All data are publicly available. The raw data (i.e., directly from the source) is available in the ./data_raw folder.
• All necessary code is presented in the ./code folder and designed to be run in sequential order. Note that the 90 and above files are sensitivity analyses and it is not necessary to run them to reproduce the primary plots or figures.
• The scripts in the ./code folder convert the data in the ./data_raw folder into our analytic data which is stored in the ./data folder.
• The ./intermediate_objects folder (not shared on Github) stores the temporary files. Specifically, we we calculate the distance between any census tract and the top 15 nearest providers by state, save each state’s result independently, and merge results. (Note that the nearest provider can still be across state boundaries so the OSRM server should have the entire US road network — see below for details.)
• The ./output folder contains all output (i.e., plots, tables, etc) relevant to the paper.
• The ./rmds folder contains the supplemental information such as session information (such as package version numbers) required to reproduce the analysis.

Preparation

Note that reproducing all figures and tables can be done by running the 08 to 10 code files. However, in order to rerun the entire pipeline start to finish (e.g., on new data), there are several additional requirements.

• OSRM Server. In order to calculate the driving times, we use OpenStreetMap-Based Routing Service OSRM. This will require you to set up an OSRM server on your computer. Instructions differ but we describe the process we used in the ./rmds/install_osrm.html file. The server and its relevant files should be stored and run in the ./data_raw/osrm/ folder or changed appropriately. (Warning: This is a memory intensive process and 64 GB of ram is recommended.)
• Google Maps API credentials. See, for example, this page on how to obtain Google Map credentials for converting messy street addresses to latitude/longitude. (WARNING: Google Map API pings cost money after ~40,000 queries per month.)
• US Census Bureau API credentials. Go to the Census.gov website to get your own API keys to access the ACS population data. This is only necessary if you want to pull different population estimates than what was used in the paper.
• Credential file (secrets.R). Both the Google Maps and USCB credentials should be stored in a file called ./code/secrets.R. See the ./code/secrets_example.R file.
• Decompressing the shapefiles. The shapefiles for census tracts of every state are stored in the ./data_raw/shp_files. These zip files need to be decompressed.

Authors (alphabetical)

• Michael Barnett (: @ml_barnett)
• Keith Humphreys (: @KeithNHumphreys)
• Mathew Kiang (: mkiang | : @mathewkiang)
• Alexander Tsai (: @drdrtsai)
• Sarah Wakeman (: @DrSarahWakeman)

Reproducible information

See the ./rmds/session_info.html file for full reproducible information including package version numbers.