During the past decade, the US opioid epidemic has fueled an increase in illicit, unsterile injection drug use (IDU) and new hepatitis C virus (HCV) infections, especially within nonurban communities.1–6 For the first time, late in 2014, injection of prescription opioids was linked to an outbreak of HIV infections in a rural US community (Austin in Scott County, Indiana). This event raised concerns about the vulnerability of similar communities to the rapid spread of HIV infection if introduced into a network of persons who inject drugs (PWID). The socioeconomic context within which this outbreak occurred was not unique: unemployment and poverty in the county exceeded national levels,7 educational attainment was low,7 and life expectancy was poor.8 During November 2014–October 2015, 181 persons were diagnosed with HIV infection, of whom 92% were coinfected with HCV.9 This county of approximately 24,000 persons reported fewer than 5 new HIV infections in the preceding 10 years.9,10
Rapid recognition of the outbreak and implementation of intensive control efforts likely reduced new HIV and HCV infections and limited their geographic spread.9 Identifying jurisdictions particularly vulnerable to a similar outbreak can increase awareness of current vulnerability and guide public health efforts to detect and prevent this type of event. To this end, we conducted a multistep analysis (Fig. 1) to identify which set of indicators (independent variables) is highly associated with unsterile IDU, then used each US county's values for these indicators to calculate a vulnerability score for each county, and finally identified which states and specific counties are most vulnerable.
Nationally, there is no reliable county-level measure of unsterile IDU, the outcome of interest for this analysis. We therefore relied on a proxy measure for this behavior. We chose the county-level rate of persons reported with confirmed acute HCV infection as the best proxy measure for the county-level rate of unsterile IDU for 3 reasons. First, the acute phase of HCV infection occurs very close in time to the injection event and is transitory.5,6 Although not all acute HCV cases are acquired through unsterile IDU and not all PWID get HCV, the county-level rates of acute HCV and IDU are assumed to have a fixed relationship that does not vary by time, place, or infection. Second, HCV is highly transmissible, especially through unsterile IDU, which is the primary risk factor for HCV infection.5,6 Third, acute HCV infections are reported nationally through the National Notifiable Disease Surveillance System (NNDSS).11 Although the sensitivity of the NNDSS is limited and cases are known to be underreported,12 it has been demonstrated sufficient for other national-level analyses and the large majority of states and the District of Columbia voluntarily report cases.2,4,13
We included NNDSS data from 2012 and 2013. Counties in states that reported no acute HCV cases to CDC in 2012 or 2013 were excluded from the analysis (ie, missing values). We applied the following rules to identify counties with zero cases of acute HCV infection from those with missing values in a given year. For states that reported >1 case, we assumed counties with no reported acute HCV cases had zero cases. For states that reported 1 case of acute HCV infection for 1 year and no cases in the other year, we assumed that all other counties had zero acute HCV cases during the 1-case year and all counties had missing values for the other year. Thus, that state was included for the analysis in the 1-case year and excluded in the other year because there was limited evidence the state was able to consistently identify and report acute HCV cases.
Potential Indicator Variables
We investigated county-level independent variables that were known or plausibly associated with unsterile IDU and that could thus potentially serve as indicators of vulnerability to rapid dissemination of HIV/HCV infection among PWID. We required that data for potential indicators be available nationally at the county level or transformable to the county level, reported at least annually, recent (ideally reported within the past 3 years), and complete (<10% of US counties missing values). Data were transformed, if needed, by converting data aggregated by 3-digit or 5-digit ZIP codes to 5-digit county Federal Information Processing Standard (FIPS) codes using area proportion method. Potential indicators were considered and identified from 6 domains: (1) drug-overdose mortality, (2) access to prescription opioids (eg, production, sales, prescriptions), (3) access to care (eg, evidence of use of care or treatment services related to IDU), (4) drug-related criminal activity (eg, arrests for drug possession or sales), (5) prevalence of IDU (eg, survey-based data), and (6) sociodemographic characteristics associated with geographic areas with higher IDU prevalence.
Forty-eight indicators, a combination of outcomes and various sources, were identified. The majority of the potential indicators were excluded from the analysis because county-level data were not available. Table 1 describes the 15 potential indicators that met our inclusion criteria (see Figure S1, Supplemental Digital Content, http://links.lww.com/QAI/A844, which shows maps of the county-level indicators). Data from 2012 and 2013 were used when available because the indicators are intended to reflect current areas of vulnerability; when not available, we used data from the most recent year available. All indicators were treated as continuous numerical variables except presence of urgent care facilities and access to interstate within 5 miles of county border, which were dichotomous (yes/no). Descriptive statistics, including median, mean, first quartile and third quartile, were calculated for the 13 continuous indicators; frequency and percentage were calculated for the 2 dichotomous indicators.
We used modeling to identify a parsimonious set of IDU-associated indicators with the strongest association with our proxy measure of IDU: acute HCV infection. We modeled the rate of acute infections by county using a multilevel Poisson model with year (2012, 2013) nested in county and county nested in state with the county population set as the offset. We treated the states and counties as random effects to account for heterogeneity. Each of the 15 indicators was modeled using a univariable Poisson random-effects model. Per capita income and population per square mile were modeled on a log10 scale. We entered all 15 indicators in a multivariable model and removed the indicator with the highest P-value, then removed and added indicators in a backwards stepwise procedure until all remaining variables had a P-value <0.05. We used this approach because we were aiming to identify indicators with the strongest association with county-level acute HCV infection rates rather than indicators causally associated with acute HCV infection, in which case step-wise modeling procedures are not recommended (see Supplemental Digital Content, http://links.lww.com/QAI/A844, Regression Modeling Analyses that describe the modeling procedure). We assessed the linearity assumption and collinearity of our continuous indicators (see Supplemental Digital Content, http://links.lww.com/QAI/A844, Regression Modeling Analyses that describe the continuous indicators linearity assessment and collinearity assessment of indicators). Standardized regression coefficients were calculated from the final multivariable model to determine the relative contribution of each indicator (see Supplemental Digital Content, http://links.lww.com/QAI/A844, Regression Modeling Analyses that describe the standardized regression coefficients).
Application of Indicators to Generate Vulnerability Scores
To calculate a vulnerability score for each county, we created a scoring dataset that contained the value for each of the final indicators by county. For drug-overdose mortality rate, we imputed the median value for the indicator from the values among the bordering counties for 133 (4.2%) counties, assuming that the surrounding counties were most similar to the county with the missing value.14 We averaged indicator values if 2 years of data were available. For each county, we then multiplied the county's value for each indicator variable by that indicator's regression coefficient from the final model and summed to produce a vulnerability score. We ranked counties by their vulnerability scores from 1 to 3143 with a higher score indicating higher vulnerability.
To account for uncertainty in each county's vulnerability score, we used simulation to estimate the 90% confidence interval (CI) for each county's vulnerability score by drawing 10,000 samples from a normal distribution for each regression coefficient. We identified a county as vulnerable when the upper 90% CI for that county's vulnerability score exceeded or matched the vulnerability score for the 2985th ranked county (ie, lower threshold for the top 5% of counties with the highest vulnerability scores) [see Supplemental Digital Content, http://links.lww.com/QAI/A844, Composite Index (Vulnerability) Score and Rank, which describes the methodology for the vulnerability scores].
For each vulnerable county, we sought to assess the likelihood that HIV might be introduced into a network of PWID in that county based on the prevalence of HIV infection in that county and nearby. We defined “HIV proximity” as the weighted average rate of people living with diagnosed HIV infection (PLWH) in and around that vulnerable county at year-end 2012. We demarcated a 20-mile buffer zone around the border of each vulnerable county based on data from the US Department of Transportation National Household Travel Survey that the average daily distance traveled per person per day was 36 miles (approximately 20 miles when rounded for a radius).15 Using a method of spatial smoothing,16 we calculated the area proportion of each county adjacent to the vulnerable county that intersected the vulnerable county's 20-mile buffer zone and the corresponding weighted numbers of PLWH and total population represented by the buffer zone. We summed these buffer zone values with the values for the vulnerable county to calculate an estimated HIV proximity rate of PLWH per 10,000 population. We used Jenks natural breaks to group counties by estimated HIV proximity.17
The number of acute HCV infection cases reported to CDC in 2012 and 2013 were 1778 and 2138, respectively. Of the reported cases, 1710 (96%) in 2012 and 2074 (97%) in 2013 were reported from 2970 counties with valid county FIPS codes and were included in modeling. There were 173 counties in 8 states and the District of Columbia with no data during 2012–2013; these jurisdictions were excluded from modeling analyses (Alaska, Arizona, Delaware, District of Columbia, Hawaii, Mississippi, New Hampshire, Rhode Island, and Wyoming). Descriptive statistics are presented in Table 2.
In the univariable regression models, all indicators were significant at P < 0.1, except mental health providers and percentage of the population without high school diploma (Table 2). We found neither substantial departures from linearity nor collinearity between the continuous indicators. Based on the final multivariable model, the following indicators were most significantly associated with the county-level acute HCV infection rate: drug-overdose deaths per 100,000 population [regression coefficient (β) 0.016; P < 0.0001], prescription opioid sales per 10,000 population (β 0.024; P = 0.0015), median per capita income, log10 (β −2.181; P < 0.0001), percent of population of white, non-Hispanic race/ethnicity (β 0.027; P < 0.0001), percent of population aged ≥16 years unemployed (β 0.038; P = 0.0095), and buprenorphine prescribing potential by DATA 2000 waiver per 10,000 population (β 0.002; P = 0.0095) (see Supplemental Digital Content, http://links.lww.com/QAI/A844, Supplemental Results and Supplemental Figs. 2 and 3, which describe the model fit results, show the reported rate of acute HCV infection and model-estimated rate of acute HCV infection by county, and illustrate the model fit and 90% confidence intervals for the counties identified as vulnerable).
Vulnerability scores were calculated for all counties using the final model regression coefficients. We identified 220 counties in 26 states as vulnerable communities (Fig. 2) [see Supplemental Digital Content, http://links.lww.com/QAI/A844, Supplemental Results and Tables 1 and 2, which describe the composite index (vulnerability) score and rank and counties identified as vulnerable].
The estimated HIV proximity at year-end 2012 ranged from 0.9 to 37.6 per 10,000 population (Fig. 3). With one exception, the estimated HIV proximity for all vulnerable counties was lower than the national rate of 29.3 per 10,000 population.
We have developed a method to identify US counties potentially vulnerable to rapid spread of HIV, if introduced, and new or continuing high numbers of HCV infections among PWID. The method employed a proxy measure for county-level unsterile IDU (ie, acute HCV infection). Ecological-level measures have been used previously to create CDC's Social Vulnerability Index. That index was designed to identify socially vulnerable populations that are more likely to be adversely affected during disaster events (eg, elderly, people living in poverty).18 It assesses overall census tract vulnerability on the basis of 14 census variables and calculates an overall percentile rank of social vulnerability for each area.18 Our method builds on this approach by first identifying the specific indicators associated with HCV (proxy for unsterile IDU). We then calculated an index score to rank counties and account for the relative contribution of each indicator to identify the most vulnerable communities. The evolution of IDU in the United States is dynamic and this method can be applied periodically to account for those changes. Improved reporting of acute HCV infections to NNDSS and the identified indicators will increase the accuracy of the method for identifying vulnerable counties over time.
Deaths from drug overdose have grown exponentially in the past decade, surpassing motor vehicle traffic accidents as the leading cause of unintentional injury-related death in the United States.19 Overdose death rates because of prescription opioids increased nearly 4-fold from 2000 to 2013.20 Prescription opioids are the most commonly abused prescription drug, and an estimated 10%–20% of people who abuse prescription opioids escalate to injection of prescription opioids or heroin,21–23 creating opportunity for IDU-associated outbreaks of HIV or HCV infection.24,25 Rates of acute HCV infection have increased steadily nationwide from 2006 to 2012, most notably east of the Mississippi River and particularly among states in central Appalachia.4 The potential for an HIV outbreak within an opioid-injecting population was realized in Scott County, Indiana,26 which ranked 32nd among the top 220 counties in our vulnerability analysis.
The demography of PWID and persons diagnosed with acute HCV infection in the United States has changed substantially. From 1999 to 2002, HCV diagnoses in the United States were greater among men, non-Hispanic blacks, and persons aged 40–49 years.27 Persons diagnosed with acute HCV infection now are equally likely to be male as female are predominantly of white, non-Hispanic race/ethnicity, and younger age.2,4,13,28,29 Our finding that white, non-Hispanic race/ethnicity was significantly associated with county-level acute HCV infection reflects these changes.
The counties identified in our analysis were overwhelmingly rural. Since 2006, rates of acute HCV infection have increased faster in rural than in urban areas4 consistent with introduction and spread of this infection into populations made newly vulnerable by the expansion of IDU in rural America. The outbreak in Scott County, IN was notable for the absence or minimal availability of harm-reduction strategies to prevent IDU-associated HIV and HCV infections, such as addiction treatment and rehabilitation, medication-assisted therapy (MAT), and syringe service programs (SSPs).9 This outbreak illustrated the need for harm-reduction strategies suited to the rural context. Our analysis can help identify those rural areas at highest risk of infectious complications from IDU where interventions can be prioritized. Unemployment and income were also significantly associated with risk for acute HCV infection. It has been hypothesized that financial stressors increase vulnerability to drug use, so that young adults in economically deprived areas may accumulate more risk factors for drug use and be more likely to establish drug dependencies at a younger age than persons in more economically privileged areas.30 Since the Great US Recession of 2007–2009, rural areas have experienced persistent and greater unemployment and poverty than urban areas.31
Buprenorphine is approved for MAT of opioid-use disorder. The indicator “buprenorphine prescribing potential by waiver per 10,000 population” represents the potential availability of this form of MAT to the population of a county, reflecting the capacity to treat in response to demand for services and not the actual rate of people receiving treatment.32 In 2011, 43% of US counties had no buprenorphine-waivered physicians,32 indicating that capacity to prescribe MAT likely lags behind the need. It is important that these indicators are not interpreted as causally associated, rather are understood as indicators of potential vulnerability for this type of outbreak.
The risk for introduction and spread of HIV infection into a vulnerable community of persons will vary according to the estimated HIV proximity of the community and the community's IDU practices. We found that the estimated HIV proximity was lower than the national average for nearly all of the 220 counties identified as most vulnerable. Although we could not measure IDU practices systematically for each US county, it is important to consider the estimated HIV proximity in the context of local IDU practices. For example, Scott County, IN was in the second highest estimated HIV proximity group relative to the 220 vulnerable counties and lower than the national average, but the injection practices in Scott County, IN were notable for the high frequency of injections and combined with the relatively large and dense network of PWID9 in the absence of effective strategies to reduce IDU-associated infection leaving the community vulnerable to this type of outbreak.
All health officials can review these results along with the most recent sources of data on HIV and acute HCV diagnoses available to them. Additional local insight may be gained by examining data sources about other factors associated with IDU that were not available for inclusion in our analysis. There are a number of indicators we would like to consider that are not available nationwide but may provide substantial local insight on vulnerability to rapid spread of HIV infections associated with unsterile IDU. These data include emergency department and hospital admissions for drug overdose or intoxication; arrests for drug possession or sales; and opioid prescribing patterns from state prescription drug monitoring programs and data on local injection practices and availability of effective strategies to reduce IDU-associated infection (eg, drug type injected, injection frequency).33 CDC is preparing additional advice to assist health departments, so they can assess whether their jurisdiction is experiencing or at risk of new HIV and HCV infections potentially related to IDU (personal communication).
This analysis is subject to limitations. First, we excluded potential indicators, as noted above, because we established strict inclusion criteria to assess vulnerability for each county. Second, this analysis was intended to identify areas currently vulnerable; however, some of the data were ≥3 years old. Third, all ecologic analyses based on reported cases of notifiable disease and deaths are subject to potential underreporting (eg, HCV data) or misclassification (eg, cause of death) biases. Reporting of acute HCV infections throughout the United States to NNDSS is susceptible to underreporting12 not only because reporting is passive and voluntary but also because the surveillance case definition captures only persons with symptoms of an illness that is often asymptomatic. The implementation of complete and timely acute HCV surveillance by all states will increase the accuracy of this method. Fourth, the HIV proximity measure might not accurately measure risk of IDU-associated HIV transmission. Our estimate is limited by the distance of the buffer zone selected and may not reflect the population distribution within each county or underlying characteristics of PLWH in these areas (eg, percent with suppressed viral load). Lastly, the indicators used may result in bias against identifying urban areas as vulnerable because the percentages of certain characteristics (eg, percentage white, non-Hispanic) are lower in urban counties. However, the historic concentration of IDU in urban areas has likely resulted in these areas having already taken preventive actions (eg, SSPs, access to MAT) to reduce risk for infections from IDU.
Prompted by a recent outbreak of HIV infection among PWID in a rural community, we developed a method to identify other communities vulnerable to rapid dissemination of IDU-associated HIV, if introduced, and new or continuing high numbers of HCV infections. Identification of a county as vulnerable to this type of outbreak does not mean an outbreak is eminent; rather jurisdictions identified as at-risk might use potentially informative local data that were not available nationally and take action as recommended in the April 2015 CDC Health Alert Network advisory.34 Expanding epidemics of prescription opioid abuse and unsterile IDU and shifting demography and geography, heralded by national changes in patterns of acute HCV infection, are critical driving forces of vulnerability. To reduce vulnerability, targeted interventions in accordance with efforts to prevent and treat substance use disorder and to reduce risk of infectious complications of IDU are warranted.
The authors would like to thank Paul Sutton, PhD, at the National Center for Health Statistics, and Jinhee Lee, PharmD, at the Substance Abuse and Mental Health Services Administration, for providing data and expertise on indicator data. We thank Debra Houry and Grant Baldwin at the CDC National Center for Injury Prevention and Control for their leadership, guidance, and support of this analysis.
1. Kolodny A, Courtwright DT, Hwang CS, et al. The prescription opioid and heroin crisis: a public health approach to an epidemic of addiction. Annu Rev Public Health. 2015;36:559–574.
2. Zibbell JE, Iqbal K, Patel RC, et al. Increases in hepatitis C virus infection related to injection drug use among persons aged ≤30 years—Kentucky, Tennessee, Virginia, and West Virginia, 2006–2012. MMWR Morb Mortal Wkly Rep. 2015;64:453–458.
3. Lipari RN, Hughes A. The NSDUH: Trends in Heroin Use in the United States, 2002 to 2013. Rockville, MD: Center for Behavioral Health Statistics and Quality, Substance Abuse and Mental Health Services Administration; 2015.
4. Suryaprasad AG, White JZ, Xu F, et al. Emerging epidemic of hepatitis C virus infections among young nonurban persons who inject drugs
in the United States, 2006-2012. Clin Infect Dis. 2014;59:1411–1419.
5. Centers for Disease Control and Prevention. Viral Hepatitis Surveillance—United States. 2013. CDC, Atlanta, GA. Available at: http://www.cdc.gov/hepatitis/statistics/2013surveillance/index.htm
. Accessed 8, October 2015.
6. Centers for Disease Control and Prevention. Viral Hepatitis Surveillance—United States. 2012. CDC, Atlanta, GA. Available at: http://www.cdc.gov/hepatitis/statistics/2012surveillance/index.htm
. Accessed 8, October 2015.
7. U.S. Census Bureau. State and county Quick Facts. Scott County, Indiana; 2015. Available at: http://www.census.gov/quickfacts/table/PST045214/18143,00
. Accessed December 10, 2015.
8. Conrad C, Bradley HM, Broz D, et al. Community outbreak of HIV infection linked to injection drug Use of Oxymorphone–Indiana, 2015. MMWR Morb Mortal Wkly Rep. 2015;64:443–444.
9. Peters PJ, Pontones P, Hoover KW, et al. An Outbreak of HIV Infection Linked to Injection Drug Use of Oxymorephone—Indiana, 2014–2015. N Eng J Med. 2016;375:229–239.
10. United States Census Bureau. United States Census Bureau. Available at: http://www.census.gov/
. Accessed February 18, 2016.
11. Centers for Disease Control and Prevention. National notifiable disease surveillance System (NNDSS). Available at: http://www.cdc.gov/nndss/
. Accessed October 7, 2015.
12. Onofrey S, Aneja J, Haney GA, et al. Underascertainment of acute hepatitis C virus infections in the U.S. Surveillance System: a case series and chart review. Ann Intern Med. 2015; 163:254–261.
13. Centers for Disease Control and Prevention. Use of enhanced surveillance for hepatitis C virus infection to detect a cluster among young injection-drug users—New York, November 2004–April 2007. MMWR Morb Mortal Wkly Rep. 2008;57:517–521.
14. Gelman A, Hill J. Chapter 25-Missing-Data Imputation in Data Analysis Using Regression and Multilevel/Hierarchical Models. Cambridge University Press, Cambridge, United Kingdom; 2006:529–544.
15. US Department of Transportation. Summary of Travel Trends—2009 National Household Travel Survey. Available at: http://nhts.ornl.gov/2009/pub/stt.pdf
. Accessed February 18, 2016.
16. Tiwari C, Rushtin G. Using Spatially Adaptive Filters to Map Late State Colorectal Cancer Incidence in Iowa, in Developments in Spatial Data Handling. Heidelberg, Germany: Springer Berlin; 2005:665–676.
17. Jenks GF. The Data Model Concept in Statistical Mapping. vol 7. International Yearbook of Cartography; 1967:186–190.
18. Flanagan BE, Gregory EW, Hallisey EJ, et al. A social vulnerability index for disaster management. J Homeland Security Emerg Management. 2011;8:Article 3.
19. Warner M, Chen LH, Makuc DM, et al. Drug poisoning deaths in the United States, 1980–2008. NCHS Data Brief. 2011;81:1–8.
20. Hedegaard H, Chen LH, Warner M. Drug-poisoning deaths involving heroin: United States, 2000–2013. NCHS Data Brief. 2015;190:1–8.
21. Neaigus A, Miller M, Friedman SR, et al. Potential risk factors for the transition to injecting among non-injecting heroin users: a comparison of former injectors and never injectors. Addiction. 2001; 96:847–860.
22. Lankenau SE, Teti M, Silva K, et al. Patterns of prescription drug misuse among young injection drug users. J Urban Health. 2012;89:1004–1016.
23. Substance Abuse and Mental Health Services Administration, Center for Behavioral Health Statistics and Quality, Treatment Episode Data Set (TEDS): 2002–2012. National Admissions to Substance Abuse Treatment Services, in BHSIS Series. Rockville, MD: Substance Abuse and Mental Health Services Administration; 2014.
24. Jones CM. Heroin use and heroin use risk behaviors among nonmedical users of prescription opioid pain relievers—United States, 2002–2004 and 2008–2010. Drug Alcohol Depend. 2013;132:95–100.
25. Jones CM, Logan J, Gladden RM, et al. Vital Signs: Demographic and substance use trends among heroin users - United States, 2002-2013. MMWR Morb Mortal Wkly Rep. 2015; 64:719–725.
26. Indiana State Department of Health. HIV Outbreak in Southeastern Indiana. Available at: https://secure.in.gov/isdh/26649.htm
. Accessed September 16, 2015.
27. Armstrong GL, Wasley A, Simard EP, et al. The prevalence of hepatitis C virus infection in the United States, 1999 through 2002. Ann Intern Med. 2006; 144:705–714.
28. Centers for Disease Control and Prevention. Notes from the field: hepatitis C virus infections among young adults - rural Wisconsin, 2010. MMWR Morb Mortal Wkly Rep. 2012;61:358.
29. Centers for Disease Control and Prevention. Hepatitis C virus infection among adolescents and young adults: Massachusetts, 2002-2009. MMWR Morb Mortal Wkly Rep. 2011;60:537–541.
30. Keyes KM, Cerda M, Brady JE, et al. Understanding the rural-urban differences in nonmedical prescription opioid use and abuse in the United States. Am J Public Health. 2014;104:e52–e59.
31. United States Department of Agriculture, Rural America at a Glance. United States Department of Agriculture, Washington, DC; 2014. Available at: http://www.ers.usda.gov/media/1697681/eb26.pdf
. Accessed February 18, 2016.
32. Stein BD, Gordon AJ, Dick AW, et al. Supply of buprenorphine waivered physicians: the influence of state policies. J Subst Abuse Treat. 2015;48:104–111.
33. Brandeis University, Pdmp Center of Excellence. Prescription Drug Monitoring Program Center of Excellence. Available at: http://pdmpexcellence.org/content/coe-prescription-behavior-surveillance-system-0
. Accessed September 16, 2015.
34. Centers for Disease Control and Prevention. Outbreak of Recent HIV and HCV Infections Among Persons Who Inject Drugs
. Health Alert Network; 2015. CDC, Atlanta, GA. Available at http://emergency.cdc.gov/han/han00377.asp
. Accessed November 6, 2015.