The geography of opioid use disorder: a data triangulation approach

INTRODUCTION

The opioid use epidemic in the United States came into broad public view during a localized outbreak of HIV infection in southern Indiana in 2015.¹ By the time of that outbreak, the rate of drug-poisoning deaths had nearly tripled since 2000, with most of that increase attributed to opioid-related deaths², and 1.9 million American adults met criteria for a prescription opioid use disorder (OUD).³ The outbreak in Scott County brought attention to a national problem through the lens of a specific community; however, the geography of problematic opioid use is much broader than rural midwestern communities, and follows a very different geographic pattern of impact compared to many chronic diseases (e.g., diabetes, hypertension, obesity) or to many infectious diseases (HIV, sexually transmitted infections) both of which have disproportionate impact in the South and major urban areas throughout the country. The geographic distribution of OUD speaks to underlying characteristics of opioid users and settings of opioid use, and, most importantly, to the need for interventions in parts of the United States that are often underserved by preventive health services.

There are a variety of indicators that can be used to understand the geography of opioid use disorder. In the past, data from existing surveillance systems for infectious diseases that can be transmitted by needle sharing have been used to identify trends in injection drug use. For example, trends in HIV diagnoses related to injection drug use and Hepatitis C virus (HCV) diagnoses (especially in younger people) have been used to provide context to evolving opioid epidemics.⁴ Relying on existing infectious disease surveillance systems is cost-efficient, but HIV and HCV diagnoses are not specific to OUD. HIV is also relatively insensitive as an indicator of injection drug use because the efficiency of transmission of HIV is less than for HCV.⁴ Other proxies for injection drug use behaviors, including skin infections and infectious endocarditis, also lack specificity for opioid injection, may be transient and are not reportable conditions. Infectious endocarditis is an especially important indicator of OUD, because it is the most serious infectious consequence of injection drug use but is not monitored in any systematic way. The Drug Abuse Warning Network Emergency Department data system collects data on drug-related emergency department visits but focuses only on drugs related to the reason for the medical visit, which reduces sensitivity of this system for understanding the scope of OUD.⁵ Cause of death data from the National Vital Statistics System can provide high quality data about specific drug use, but may be variable in quality because the patterns of testing for specific drugs in overdose cases vary by coroner or medical examiner jurisdiction and budgetary constraints associated with drug testing. ⁶

Here, we review four indicators in terms of contributions toward our understanding of geographic patterns and temporal trends in OUD: percent of new HIV diagnoses attributable to injection drug use, hepatitis C virus (HCV) infections, the opioid prescription rate, and opioid-related overdose death rate. We discuss what is currently known about how each indicator varies geographically, at which level of geographic granularity the indicator can be meaningfully analyzed, and potential biases arising from use of each indicator as a signal for geographic patterns and temporal trends in the opioid epidemic. We conclude with a summary of geographic patterns from all four indicators and discussion of additional data needs.

DATA SOURCES

New HIV diagnoses attributable to IDU

For number of HIV diagnoses attributed to injection drug use and proportions of HIV diagnoses among PWID by state, we used data from the NCHHSTP Atlas and from AIDSVu.org. New diagnoses of HIV infection are reportable by law in all US jurisdictions to state/territorial health departments, who submit name-based diagnoses to CDC⁷, which deduplicates this information and makes summaries available on public repositories such as NCHHSTP Atlas+⁸ and AIDSVu.org⁹. HIV diagnostic data are typically presented as 1-year totals, with a less than 2-year lag in national reporting¹⁰ and often more timely data availability at the state-level.¹¹

HCV infections among persons under 50 years

For estimated numbers of people living with HCV we used a combination of NHANES data and census data to produce population-based estimates as previously described. ^12–14 Briefly, NHANES estimates for HCV viral detection were used to calculate direct weighted estimates of national Hepatitis C prevalence. The 2012–2016 American Community Survey microdata samples were used to generate estimated population totals within each state, stratified by sex, race/ethnicity, and birth cohort. These state-by-strata population totals were multiplied by the stratified Hepatitis C prevalence estimates to generate crude state-level estimates. NVSS mortality data and intercensal population totals were used to model the HCV-related and narcotic overdose death rates in the same strata by state.

Within strata, these two mortality rates were combined using weights computed from data-driven assumptions about the proportion of HCV prevalence in a given birth cohort that was likely to be attributable to injection drug use. The combined mortality rates were multiplied by state-by-strata prevalence estimates to generate mortality-adjusted HCV prevalence totals in each stratum in each state. These were summed within states across strata, and the numbers of additional prevalent HCV infections estimated from populations unsampled by NHANES were added to state-specific estimates. For the stratified estimates, state- and strata-specific hepatitis C prevalence among populations unsampled by NHANES were estimated in a way that reflected state- and strata specific hepatitis C estimates derived from NHANES.

Opioid prescription rates

Data on opioid prescription rates per population size are publically available through https://www.cdc.gov/drugoverdose/ and are sourced from the IQVIA Transactional Data Warehouse. IQVIA data are collected from a sample of approximately 50,000 retail pharmacies that together dispense 90% of retail prescriptions nationally. Prescriptions are recorded regardless of initial or refill status and payment type. Opioid prescriptions, include buprenorphine, codeine, fentanyl, hydrocodone, hydromorphone, methadone, morphine, oxycodone, oxymorphone, propoxyphene, tapentadol, and tramadol. U.S. Census data were used for population size denominators.

Opioid-related overdose death rate

Data for opioid-related deaths were obtained from CDC WONDER, using ICD-10 codes consistent with CDC reporting of drug overdose deaths (e.g., X40-X44, X60-X64, X85, and Y10-Y14)¹⁵ with T-codes indicative of an opioid-related cause (e.g., T40.1-T40.4, T40.6).^2,16

INDICATORS

New HIV diagnoses attributable to IDU

In most states, 7% or fewer of HIV diagnoses were attributable to IDU in 2017. States with higher proportions of HIV infections attributed to injection drug use were concentrated in central Appalachia and the Northeast. Many of the states with large proportions of HIV infections attributable to IDU are also states with overall low HIV prevalence. West Virginia, Kentucky, New Hampshire, Vermont, and Maine all had less than 200 persons/100,000 population living with HIV in 2016⁹, yet, at least 10% of new infections were attributed to IDU.

The high levels of risk-factor completeness for HIV cases render HIV infections attributed to IDUs a highly-specific signal of individuals’ IDU history. Yet given the long latency of HIV infection, in conjunction with infrequent testing for PWID^17,18, may render this less specific in terms of the timing of infection and may in turn result in geographic misclassification, as a persons’ residence at diagnosis may be discordant from that at the time of infection. These issues are even more extreme for the potential indicator of persons living with an HIV diagnosis, due to the lifelong nature of both HIV as a chronic infection and classification of IDU as risk factor.

Although data on new HIV diagnoses overall are publicly-available in most states at the county level, new diagnoses attributable to IDU is available at the state level (AIDSVu), although some states/cities individually release these data at finer levels. This lack of geographic resolution is in part due to HIV infection being a relatively rare consequence of IDU, primarily due to the low per-act acquisition risk associated with injection¹⁹, compared to other bloodborne viruses such as Hepatitis B or C.⁴

Thus, given the rareness of a new IDU HIV diagnoses, this indicator functionally may be suppressed or not very informative given the sparseness of the information. HIV diagnoses in a specific subgroup are suppressed for groups with 1–4 cases in a year in a specific geographic unit, resulting in missingness either because areas have small populations (even if infection rates are higher), or in areas that have strong prevention programs. For example, data in Indiana and New York City are both sparsely reported, despite very different epidemic situations. Indiana is more sparsely populated with IDU and has a higher rate of new infections in IDU, and New York City has a higher population of IDU with lower acquisition rates. On the other hand, this sparseness is amenable to detection of HIV transmission clusters and outbreaks (for example in West Virginia²⁰, Indiana¹, and less urban parts of Massachusetts²¹).

There is the potential for bias in eliciting IDU as a risk factor. The setting for the test (CBO, physician, lab), provider comfort/ability in eliciting risk factors, and a patient’s comfort in disclosing risk may lead towards undercounting of IDU risk factor.²² Some of these biases may be differential by urbanicity given variations in provider types, training, and the landscape of stigma towards PWID.^23,24,25

HCV prevalence among persons <50 years

HCV surveillance data quality differs across state and local jurisdictions, and standardized estimates of HCV infection by state are only available through modeling approaches.^12–14 Here, we present prevalent HCV infections among persons aged less than 50 years by state as previously published modeled estimates¹² and as described in “data sources.” Estimated rates of HCV infection among persons aged less than 50 years were concentrated in central Appalachia (Pennsylvania, Ohio, Kentucky, West Virginia, Tennessee), and West Virginia had the highest rate of infection among this age group in the country. This indicator is limited to persons aged less than 50, because these infections are more likely to be associated with injection drug use compared to infections among Baby Boomers (born between 1945–1965). HCV infections among Baby Boomers were generally acquired in the distant past via blood products or previous injection drug use. Among estimated 44,700 acute HCV infections in 2017, a majority were among those less than 40 years and associated with injection drug use.²⁶

Given the higher per-act risk for HCV, compared to HIV, HCV infections are in theory a more abundant signal for tracking IDU epidemics. HCV has a higher per act transmission rate per episode of needle sharing (250/10,000 acts)^27,28 than does HIV (63/10,000 acts)¹⁹. Relatedly, HCV infections are often a leading and more prevalent indicator for subsequent HIV infections in outbreak settings.^29,30 The use of HCV diagnoses as markers for underlying HCV incidence in the population are complicated by factors that include the ability for at least a quarter of infected persons to spontaneously clear infection, a multi-decade latency period of chronic infection, and possibly lower awareness and screening rates among PWID. ³¹

Another challenge for using HCV infections an indicator of OUD and associated IDU behavior is inconsistencies in how diagnoses are reported to states and to CDC. HCV diagnoses are submitted to local health departments and reported to CDC under two case definitions, one for its acute phase and the other for chronic phase, via the National Notifiable Diseases Surveillance System. As of 2017, 44 states reported acute HCV cases and 39 reported chronic HCV cases to CDC, with the remainder not including HCV as a legally reportable condition or not submitting data to CDC. Unlike for HIV, submissions to CDC are not name-based, precluding deduplication across jurisdictions and national-level mortality matching. Because of stringent criteria and low levels of diagnosis during acute phase, only about 3,000 reported by CDC annually and most new diagnoses reported are chronic (143,286 in 2017). These differences in reporting across jurisdiction make geographic comparisons of HCV infections using surveillance data difficult.

Because of the heterogeneity in state submissions to CDC, stringent case criteria, and limited stratifications available nationally, individual state-level reports often offer more detail into local HCV trends that might inform the OUD epidemic (ref: state epi reports and profiles). This comes with the drawback of less standardization of how case definitions are operationalized and reporting formats, between jurisdictions. Complicating this all substantially is that relative to HIV, fewer federal resources are made available to jurisdictions for HCV surveillance, resulting in lower levels of data quality, in terms of complete reporting, completeness of demographic and risk factors, case investigations. For example, among 2017 acute diagnoses reported CDC, 48% were missing risk-factor data.

Acute HCV infections overall, among those <50 years, and among those with reported IDU risk-factor are all useful indicators for understanding trends in new IDU-associated infections. The total acute infections is available from CDC at the state-level, whereas age and IDU-risk factor estimates are provided only for the national total. Given the low levels of acute infections reported to CDC relative to chronic ones, trends in reported chronic cases among persons < 50 years is another marker for IDU-associated transmission, but this is not summarized by CDC at any level of geography.

Both HIV and HCV diagnoses attributable to IDU have a strong correlation to OUD, as many PWID are a subset of those with OUD. Yet there are several key sources of potential sources of misalignment. First, harm reduction activities like syringe exchange lower HCV/HIV acquisition risk but do not directly address addiction. Thus in places with legal and more widespread access to syringe exchange³², one would expect greater discordance between HIV/HCV risk and OUD, controlling for drug treatment access. Conversely in places with fewer harm reduction resources (e.g. parts of Appalachia and South), one expects tighter relationship between HIV/HCV risk and OUD. Additionally, HIV/HCV risk attributable to IDU is partly independent of the specific substance being injected, other than the pathways by which specific substances may influence injection frequency and other risk behaviors.³³ The makeup of the drug supply varies geographically and by urbanicity. Stimulant use and injection, notably methamphetamine, is on the rise in a number of states and these non-opioid substances may accordingly disproportionately influence HCV/HIV risk in those places.³⁴ This is likely to induce meaningful geographic variation in the relationship between HCV/HIV diagnoses and underlying prevalence of persons with OUD.

Opioid prescription rate

In 2017, the rate of opioid prescriptions per 100 population were generally highest in the Southeast and central Appalachia. Nearly one-quarter of states had prescription rates higher than 74 opioid prescriptions dispensed per 100 persons during 2017. The opioid prescription rate has declined nationally since 2012, from a high of 81/100 persons in 2012 to 59/100 in 2017, but this decline has been differential by geography. ^35,36

Opioid prescription data are publically available at both state and county levels, currently from 2006 – 2017. Completeness of data is high and has improved over time, and the percentage of counties with available data increased from 88% in 2006 to 94% in 2017. Opioid prescription data are robust, relatively timely, and are collected from a sample of pharmacies representing most prescriptions dispensed in the U.S. However, the opioid prescription rate was likely a more informative geographic indicator of opioid use disorder prior to 2010 when most drug overdose deaths were attributable to more commonly prescribed opioids versus rarely prescribed synthetic opioids, such as fentanyl, or illicit opioids such as heroin. ^37–39 As the percentage of drug overdose deaths attributable to non-prescription opioids has increased, this indicator may be a less specific signal of opioid use disorder, increasingly attributed to illicit, non-prescription opioids.

OUD attributable to prescription opioids is also somewhat under-captured by this indicator due to prescription opioids that are obtained illegally, outside the pharmacy setting. It is difficult to impossible to measure how much illegal sales of prescription opioids reduces specificity of this indicator but highly likely that this issue with specificity varies geographically. Additionally, the extent to which prescription versus non-prescription opioids are responsible for OUD prevalence varies considerably by geography³⁷, which further complicates geographic comparisons. Sensitivity of this indicator for signaling OUD is compromised by legitimate use of prescription opioids for pain management, but it is difficult to know to what extent this issue may affect sensitivity for geographic comparisons.

Opioid-related overdose death rate

In 2018, the highest state-specific opioid overdose death rate was reported in West Virginia (38.2/100,000). By region, higher opioid overdose death rates were reported in Appalachia and the Northeastern United States; of the top 10 death rates by state, 7 were in the Northeast, and 4 were Appalachian states (Pennsylvania is included in both groups). Of the lowest 10 opioid overdose death rates by state, 5 were in the Midwest, 3 were in the west, and 2 were in the South.

The opioid-related overdose death rate is available at both state and county-levels by year through CDC WONDER, which includes multiple cause of death data by cause and decedent characteristics.⁴⁰ This indicator can be used to assess geographic patterns and trends down to the county level. Importantly, opioid-related overdose deaths can be further stratified by those attributed to prescription, synthetic, and illicit opioids.⁴¹

Opioid-related drug overdose is a highly specific indicator of OUD. Deaths in the U.S. are robustly recorded in vital records, and those indicating opioid-related overdose are based on toxicology screening on autopsy and are likely to be true representations of OUD. Specificity is diminished, however, when multiple drugs are detected on toxicology, making it unclear how to attribute an overdose death. ⁴² This indicator may have sub-optimal sensitivity for signaling geographic patterns and trends in OUD, however, due to both geographic variability in both completeness of recording drug-specific overdose deaths and resources invested in overdose prevention among persons with OUD.

Drug overdose deaths are ascertained using a multistep process with several opportunities for misclassification. Physicians, coroners, or medical examiners record multiple causes of deaths on death certificates, and these causes of death are classified using standardized algorithms by CDC using International Classification of Diseases, 10th Revision (ICD-10) cause of death and toxicology codes. However, when postmortem toxicology testing does not occur, T-codes will not be recorded, and drug overdose deaths will not be classified according to a specific drug type. This happens disproportionally in settings with limited death investigation resources or non-standardized toxicology testing procedures, complicating geographic comparisons.^43,44 Completeness of T-codes has increased over time as resources invested in monitoring drug overdose deaths has increased ⁴⁵, but a disparity in completeness remains in urban versus rural areas and in areas with medical examiners versus coroners, which are disproportionately rural.⁴³

Geographic variability in overdose prevention interventions also compromises the sensitivity of opioid-overdose deaths for signaling geographic patterns in OUD. Harm reduction strategies such as community-level distribution of naloxone, provision of test strips for detection of contaminated drugs, and polices such as Good Samaritan laws may all prevent overdose deaths among persons with OUD. ^46–48 These interventions and policies are supported and implemented differently by geography ^49,50, however, and the lack of standardized data on availability of overdose death prevention interventions by geography is a limitation of using this indicator to monitor geographic patterns in OUD.

DISCUSSION

The different types of opioid data available by geography show some commonalities, and some distinctions. For example, Appalachian states are disproportionately heavily impacted in all indicators, but Tennessee is in the lowest quartile of HIV diagnoses. Such discrepancies might result from other prevention modalities that would mitigate HIV transmission, but not HCV transmission or overdose deaths – for example, reductions in prevalence of unsuppressed HIV viral load levels in PWID. Similarly, some Northeastern states like Massachusetts, Vermont and New Hampshire have high levels of HIV diagnoses among PWID, but relatively low opioid prescription rates. In the West, Oregon, Idaho and Montana stand out as states that in the upper quartiles of multiple indicators; these consistent patterns of indicators suggest concerns across dimensions of opioid prescriptions, overdose and associated infectious diseases. In such areas, there are opportunities both to address the prevalence of OUD, and to broadly implement harm reduction programs (such as syringe exchange) to decrease opioid misuse and to reduce infectious consequences of needle sharing. We recognize that such interpretations of state across indicators are subject to ecological fallacy, in that we do not know that the people prescribed opioids, dying from opioid overdose, and acquiring HCV and HIV are the same people. ⁵¹

Despite these limitations, ecological comparisons across indicators by geography provide signals about where the OUD epidemic is concentrated and where it is heading. Ultimately, data at finer geographic resolutions are needed to better understand individual and community risk factors for OUD and to inform local resource allocation and intervention implementation. For example, supplemental surveillance systems such as the National HIV Behavioral System^17,52,53 and the Medical Monitoring Project⁵⁴ use targeted sampling and individual interviews to routinely assess risk factors among persons with HIV risk and health status and behaviors among persons living with HIV. These data are collected by states and cities and allow jurisdictions to understand local patterns while also allowing data to be pooled for national estimates. Similarly, the STD Surveillance Network collects geographically-specific data from sexually transmitted disease (STD) clinics and individuals diagnosed with gonorrhea to provide enhanced data on demographic characteristics, risk behaviors, and unmet needs for care among persons diagnosed with gonorrhea.^55,56

Supplemental surveillance systems such as these provide a depth of information at the person-level not available from core HIV and STD case reporting. OUD may be less likely to be diagnosed than either HIV or gonorrhea and is not a reportable condition, complicating sampling for similar supplemental surveillance systems, but creative solutions are needed to collect routine information from persons with high OUD risk. Data from currently on-going networks of research studies such as those funded by NIH’s Helping to End Addiction Long-term (HEAL) Initiative ⁵⁷, will also be helpful for understanding geographic patterns in OUD. Standardized data collection systems and a centralized repository for research data, particularly for multi-site research projects, could make information for geographic comparisons more readily accessible.

In addition to geography, characteristics of place are also important considerations for understanding patterns in the OUD epidemic. Social determinants of health such as per capita income, workforce characteristics, and urbanicity are associated with both OUD risk and prevention and treatment outcomes when measured at the county-level. ^29,58–60 An emerging literature^59–61 suggests these kinds of population characteristics cluster by geography and can help identify places with high OUD risk. Findings suggest, for example, that areas with large proportions of the population involved in the mining industry, with high levels of unemployment, or with high incarceration rates have higher risk for opioid-related overdose deaths. Particularly when these characteristics are observed in non-urban areas, they are also associated with less availability of OUD treatment services.⁶¹ This literature indicates structural factors interact with opioid availability to determine the trajectory of the OUD epidemic in a particular place.^29,62 Research elucidating factors underlying geographical differences in OUD prevalence provides context for understanding OUD risk, as well as critical information for resource allocation and tailoring prevention and treatment interventions.

CDC has recently implemented a number of novel surveillance systems to address drug overdose that, as they mature, will provide important new insights about opioid and other drug epidemics.⁶³ Enhanced State Opioid Overdose Surveillance (ESOOS) funding provides support for states to collect more timely and complete data on both fatal and non-fatal overdose. ESOOS was launched in 12 states in 2016 and expanded since that time to include 47 states. A major focus of ESOOS is increasing the use of syndromic data from emergency rooms and emergency medical services for a better understanding of the number and characteristics of non-fatal overdoses. ESOOS-funded states also conduct extensive death-scene investigations for overdose deaths and report information on toxicology, route of drug administration, and other pertinent information about decedents to the State Unintentional Drug Overdose Reporting System (SUDORS). These data sources are likely to provide more drug-specific information about both non-fatal and fatal overdoses than was previously available through NVSS and CDC WONDER.

Understanding the geography of OUD is important at the state and local level to identify failures of prevention, and to construct a public health program that addresses the specific needs indicated by the aggregate indicators for the jurisdiction. For example, many Appalachian states likely need a range of public health interventions, including interventions with providers to reduce opioid prescribing, increase in the number of providers allowed to prescribe buprenorphine, SSP to reduce needle sharing, improved Naloxone access and increasing access to trained responders, and improved HCV screening and treatment programs. Although the indicators presented here share underlying causes and public health needs, individual indicators also suggest specific public health programmatic opportunities. For example, high HIV diagnoses in PWID suggest opportunities for scaling up PrEP among PWIDs, prompt referral to HIV care and programs to support adherence in ART care, and high HCV diagnoses suggest opportunities for HCV treatment programs, coupled with harm reduction programs to prevent reinfection. For ID physicians, awareness of the specific indicators that document OUD in their communities and states should influence the frequency of screening for infectious consequences of OUD and indicate the need for robust referral networks for specific types of prevention services outside of the clinical setting.

OUD is a complex health condition with a myriad of negative health outcomes. Given the complexity of addiction and the multiple forms of opioid misuse that occur, understanding OUD required the triangulation of different data sources. Because different infectious disease surveillance systems have different strengths and weaknesses, ecological analyses can increase the sensitivity of a process of characterizing where services are needed compared with analysis of any single indicator. Moreover, more granular data are needed because available evidence indicates that there are subtle but important local differences in local populations, in the availability of local interventions, and in the types and routes of drugs used. Surveillance is the cornerstone of advocacy, awareness and evidence-based public health response. We call for increased investments in multiple aspects of surveillance, especially surveillance for Hepatitis C infection, and for sentinel surveillance activities to advance understanding geographic characteristics of the epidemic at the level of people, rather than at the level of jurisdiction.

Figure 1:

Percent of HIV diagnoses attributed to injection drug use, by US State, 2017

Figure 2:

Estimated rate of persons <50 years of age/100,000 population living with Hepatitis C, by state, United States, 2013–2016

Figure 3.

Opioid prescription rate per 100,000 population by state, 2017

Figure 4.

Death rate attributable to opioid overdose/100,000 population by state, United States, 2018

Synopsis.

Opioid use disorder is a complex condition that is not easily quantified among US populations as there are no reporting systems in place. We review indicators of OUD that are available at the state and county levels (HIV diagnoses among people who inject drugs, HCV diagnosis in people <50 years, opioid overdose death rates and opioid prescription rate), discuss their strengths and limitations as indicators, and present visualizations of four indicators. The interpretation of the ecological results, while subject to ecological fallacy, and the visualization of indicators at the local level will provide actionable insights for clinicians and public health officials seeking to mitigate the consequences of OUD at the patient and community levels.