Unclassified drug overdose deaths in the opioid crisis: emerging patterns of inequity

Abstract

Objective

Examine whether individual, geographic, and economic phenotypes predict missing data on specific drug involvement in overdose deaths, manifesting inequities in overdose mortality data, which is a key data source used in measuring the opioid epidemic.

Materials and Methods

We combined national data sources (mortality, demographic, economic, and geographic) from 2014–2016 in a multi-method analysis of missing drug classification in the overdose mortality records (as defined by the use of ICD-10 T50.9 on death certificates). We examined individual disparities in decedent-level multivariate logistic regression models, geographic disparities in spatial analysis (heat maps), and economic disparities in a combination of temporal trend analyses (descriptive statistics) and both decedent- and county-level multivariate logistic regression models.

Results

Our analyses consistently found higher rates of unclassified overdoses in decedents of female gender, White race, non-Hispanic ethnicity, with college education, aged 30–59 and those from poorer counties. Despite the fact that unclassified drug overdose death rates have reduced over time, gaps persist between the richest and poorest counties. There are also striking geographic differences both across and within states.

Discussion

Given the essential role of mortality data in measuring the scale of the opioid epidemic, it is important to understand the individual and community inequities underlying the missing data on specific drug involvements. Knowledge of these inequities could enhance our understanding of the opioid crisis and inform data-driven interventions and policies with more equitable resource allocations.

Conclusion

Multiple individual, geographic, and economic disparities underlie unclassified overdose deaths, with important implications for public health informatics and addressing the opioid crisis.

INTRODUCTION

The evolving and worsening opioid crisis

The US opioid epidemic continues to exact an enormous human toll and a profound financial burden.¹ Tracking the impact of the opioid crisis, a collective of individual experiences surrounded by shame and stigma, primarily relies upon end-stage indicators. Death as a result of overdose is 1 of the most prominent measurable attributes.² Despite the significance of the epidemic, opioid misuse and abuse are challenging to capture as large-scale data measurements,³ largely because overdose patterns have remained dynamic, with new sources of illicit opioids emerging as medical use has decreased.⁴ Illicit opioid sources, including potent synthetic opioids, fueled a six-fold increase in the opioid-associated death rate from 1999 to 2017.⁴ One of the greatest challenges of the opioid crisis is the lack of effective opioid risk modeling to predict misuse, abuse, and overdose.⁵ The medical informatics community is uniquely poised to navigate this challenge, and this work seeks to support that effort by highlighting the pattern of inequity in overdose mortality reporting, which is a critical outcome and metric in the opioid crisis.

Opioid mortality reporting

Despite being 1 of the starkest quantifications of this public health crisis, opioid-related deaths face well-delineated data collection problems.⁶ Factors impacting death certificate reporting are complex and have been poorly addressed to date.^7–9 Whereas the opioid crisis has impacted individuals and communities across socioeconomic lines, significant health disparities exist regarding access to care, treatment, and community support resources.¹⁰ Focusing on opioid-related mortality is critical to better understand patterns and to develop targeted prevention methods so as to support local communities.^11,12 Ideally, real-time data tracking and monitoring for opioid-related overdoses and deaths would allow for focused responses.¹³ In order to achieve this ambitious goal and to support the US Health and Human Services’ goal of improved data collection for the opioid crisis,¹⁴ it is necessary that we enhance our understanding of data collection itself, extraction challenges, and missing data patterns.¹⁵

Opioid drug categorization

An important facet of data collection ambiguity is how to best categorize the drugs involved in fatal overdoses.^7,16 Opioid deaths can be related to individual or combined prescription or nonprescription medications. Opioid deaths have been more recently categorized by Center for Disease Control and Prevention (CDC) designations of synthetic opioids (eg, fentanyl and codeine), commonly prescribed opioids (eg, natural and semisynthetic opioids), and heroin.¹⁷ However, these designations are neither widely adopted nor do they recognize whether those medications were prescribed or obtained illicitly. Compounding this issue, fatal and nonfatal overdose assessments have been challenged by data analysis and resource barriers.¹⁸

Unclassified drug overdose deaths

Overdose deaths with unclassified drug involvement constitute a substantial portion of all fatal overdoses, thus hampering our understanding of the true extent of the opioid-related crisis. Under representing the number of opioid deaths is a significant problem, with toxicology data demonstrating 86.8% of unclassified cases testing positive for an opioid.¹⁶ With 17.2% of fatal drug overdoses in 2015 having a missing drug classification (ICD-10 T50.9),⁸ the true opioid overdose death rate could be substantially higher than reported based on classified drug mortality only. The T50.9 code is used by medical examiners and coroners when the drug that caused the overdose is unclear. For example, this may occur in cases where the examiners may suspect drug involvement but are not able to specify the causal drug. This could be due to the drug not being included in the test assay, either because it was not requested or the panel has not been updated to reflect the most recent profiles of the drug.⁴ Additionally, the level of misreporting has been related to state-level death certificate systems and other data frameworks.¹⁵ Factors associated with relying solely on state-level death certificate systems and other data frameworks have previously been identified.^15,17

Patterns of inequity

Our work focuses on developing a better understanding of the economic, geographical, and individual-level phenotypes in missing drug classifications in US overdose death data, which is currently the primary data source used to measure the scale of the opioid crisis. The health equity implication of missing data is a subsequent disparity in resource support during the opioid crisis.¹ There is potential for using information science and technology to better understand disparities as part of developing solutions to the opioid epidemic.¹⁹ From a large-scale data analytic perspective, we focused on addressing inequities in the opioid crisis by examining individual, geographic, and economic disparities in missing information on specific drug involvement. By investigating 1) individual-level disparities, 2) geographic disparities, and 3) economic disparities in classified versus unclassified drug overdose deaths in a multi-method approach, we uncovered inequities in overdose mortality data, which has implications for our current understanding of the opioid epidemic and the development of solutions to the crisis.

MATERIALS AND METHODS

Methods

This project was reviewed by University of Rochester’s IRB and found to be exempt as not human subjects research (IRB # RSRB00073395).

Study design and data sources

To provide a comprehensive picture of disparities in missing data of specific drug involvement in fatal overdoses, we combined multiple sources of data and used several methods, both at the individual decedent and county levels. This multi-pronged approach allows us to identify inequities by individual characteristics as well as geographic disparities and economic disparities, which makes this study relevant for both policy and informatics solutions.

We performed a secondary analysis of the 2014–2016 Multiple Cause of Death data from the CDC. These data contain information on decedents’ age, sex, race/ethnicity, marital status, educational attainment, underlying cause of death and multiple causes of death (specified in ICD-10 codes), county of residence and county of death, year and month of death, day of the week of death, and place of death—among other characteristics. We linked mortality records to data on median household income from the Small Area Income and Poverty Estimates (SAIPE) program, population estimates from the Surveillance, Epidemiology, and End Results (SEER) program, and rural–urban county classification from the National Center for Health Statistics. In trend analyses, we augmented the period of study with 2008–2013 data.

Subjects

In individual-level analyses, our Population of interest is all decedents from drug overdoses. Following the CDC definition, we identified drug overdose deaths by the underlying cause of death with ICD-10 codes X40–44, X60–64, X85, and Y10–14. There were 163 091 fatal drug overdoses in 2014–2016. In county-level analyses, we focused on counties that had at least 1 drug overdose in a year (N = 7244 county-years).

Measurements

The outcome of interest in individual-level analyses is an indicator variable for missing information on specific drug involvement in a fatal drug overdose (further referred to as unclassified drug overdose). Following CDC, we defined an unclassified drug overdose according to the following criteria: 1) at least 1 of the multiple causes of death is overdose by other and unspecified drugs, medicaments and biological substances (ICD-10 T50.9); and 2) none of the other multiple causes of death indicate involvement of a specified drug, medicament, or biological substance (ICD-10 T36–T50.8). In county-level and spatial analyses, we calculated the annual rate of unclassified drug overdoses relative to all drug overdoses as well as the indicator for whether a county had any unclassified drug overdoses in a year.

Our main individual predictors included race (White, Black, Asian, American Indian), ethnicity (Hispanic), sex, age (0–19, 20–29, 30–39, 40–49, 50–59, and 60 and older), marital status, and educational attainment (no high school diploma, high school graduate, some college completed, college graduate). For county characteristics, we calculated tertiles of population count, median household income, number of total drug overdoses, nonzero rate of unclassified drug overdoses, and proportions of the population that are Black, Other race, and Hispanic. All 6 categories of rural–urban classification were used.

Analysis

To address the 3 objectives of the study, we developed the following analysis strategy: 1) individual disparities were examined in individual-level regression analyses, 2) geographic disparities were analyzed using spatial mapping techniques, and 3) economic disparities were investigated in individual-level regression models (both overall and stratified by county income tertiles), in graphical visualization of temporal trends by income tertiles, and in county-level regression models. These approaches are described in more detail later.

Corresponding to our first study objective, we estimated individual-level logistic regression models. We hypothesized that the main individual predictors described previously reveal disparities in whether a specific drug involvement in a drug overdose is identified. We adjusted for time-varying county characteristics such as county population, the number of overdoses, and median household income. We also accounted for observed and unobserved state characteristics and secular trends by including state, year, and month fixed effects. We also adjusted for the place of death, whether an autopsy was performed, whether the decedent’s county of residence is not the same as the county of death, whether the death occurred on the weekend, whether data on age, education, and place of death are missing, and the rural–urban county classification. We further performed stratified analyses by tertiles of nonzero rate of unclassified drug overdoses to investigate whether individual disparities differ by counties with relative differences in the rate of missing data on specific drug involvement.

With respect to our second objective, we examined geographic disparities by mapping unclassified drug overdose rates in order to visualize geographic zones where missing data on drug involvement is common. Finally, we analyzed economic disparities by examining coefficients on the county-level median household income tertiles in individual-level models, stratifying individual-level models by these income tertiles, examining temporal trends in unclassified drug overdose deaths by these income tertiles, and modeling missing drug information at the county-level. County-level models were estimated using logistic regression. The main predictor of interest is the county’s income tertile; we adjusted for tertiles of county population, proportions of minority populations (Black, other race, Hispanic) in a county, and the county rate of drug overdoses. We also adjusted for the rural–urban county classification, as well as state and year fixed effects. Statistical analyses were done in R and Stata; maps were created in ArcGIS.

RESULTS

Descriptive statistics

In Table 1, we display summary statistics of decedent characteristics by drug overdose groups: opioid overdoses, nonopioid overdoses, and unclassified drug overdoses. We observe that opioid overdoses and unclassified overdoses have similar age distributions, whereas decedents of nonopioid overdoses are generally older. Opioid overdoses and unclassified overdoses also have similar race and ethnicity profiles (ie, more White and less Black, Asian, or American Indian, and are less likely to be Hispanic) relative to nonopioid overdoses. Unclassified overdoses are more likely to be female than either opioid or nonopioid overdoses. Opioid and unclassified overdoses have similar rates of the death location being decedent’s residence. Autopsies were less likely to be performed for unclassified overdoses relative to either of the classified drug overdose groups.

Table 1.

Decedent Characteristics by Drug Overdose Group, 2014–2016

Characteristics	Opioid Related Overdoses (N = 103 987)	Non-Opioid Related Overdoses (N = 31 915)	Unclassified Drug Overdoses (N = 27 189)
Age: 0–19 (%)	1.67	1.86	1.35
Age: 20–29 (%)	20.56	9.62	15.73
Age: 30–39 (%)	25.65	17.07	23.24
Age: 40–49 (%)	21.65	23.45	23.23
Age: 50–59 (%)	21.60	29.20	25.36
Age: 60 and above (%)	8.86	18.78	11.08
Age: Unknown (%)	0.01	0.03	0.02
Female (%)	34.04	39.05	43.78
Race: White (%)	88.83	79.97	89.85
Race: Black (%)	9.36	15.86	8.47
Race: Asian (%)	0.76	2.44	0.82
Race: American Indian (%)	1.05	1.74	0.86
Hispanic (%)	7.75	9.50	6.49
Married (%)	21.60	23.91	25.45
Education: < High school (%)	18.45	18.50	18.27
Education: High school (%)	47.47	42.50	46.19
Education: Some college (%)	22.81	22.55	24.02
Education: College (%)	7.96	12.19	9.47
Education: Unknown (%)	3.31	4.26	2.05
Death Location: Hospital DOA (%)	1.99	1.44	2.17
Death Location: Hospital inpatient (%)	8.29	17.63	12.51
Death Location: Hospital outpatient (%)	14.28	15.89	11.39
Death Location: Nursing home (%)	0.18	0.44	0.36
Death Location: Decedent’s residence (%)	53.54	45.04	55.05
Death Location: Unknown (%)	21.53	18.90	17.95
Autopsy done (%)	79.84	72.14	66.92
County of death = County of residence (%)	84.75	83.53	85.11

Abbreviation: DOA, dead on arrival.

Individual disparities

Our individual-level regression estimates are presented in Table 2, Column 1. We found that racial and ethnic minorities are less likely to have an unclassified drug overdose: Black decedents are 17.6%, Asians 20.9%, American Indians 18.8% less likely than White decedents, and Hispanic decedents are 15.4% less likely than non-Hispanic decedents to have missing drug information. Relative to those aged 20–29, those in the age groups of 0–19 years old and 60 years and older are less likely to have an unclassified drug overdose, whereas those in age groups 30–39, 40–49, and 50–59 are more likely. Decedents who are married are 10.7% more likely to have missing drug information. Relative to those without a high school diploma, decedents with at least some college education are more likely to have an unclassified drug overdose: 8.0% and 11.7% for those with some college education and with a college degree, respectively. Although we observe variation in point estimates for our decedent-level characteristics across tertiles of county-level unclassified drug overdose rates, most associations are consistent with the overall model. One exception is that the youngest age group (age 0–19) is more likely to have an unclassified overdose than the reference group (age 20–29) in counties with the lowest rates of unclassified overdoses, reflecting a substantial change from the overall model and models by higher tertiles. Additionally, married decedents are more likely than unmarried decedents to have an unclassified overdose in our overall model and in counties in the 2 lower tertiles of unclassified overdose rates; this relationship reverses in counties in the highest tertile, although this estimate is only significant at 0.10 level of significance.

Table 2.

Individual predictors of unclassified drug overdoses, overall and by county unclassified drug overdose rate tertiles, 2014–2016

	Overall (N = 163 091)	County unclassified rate tertile 1 (N = 84 325)	County unclassified rate tertile 2 (N = 30 988)	County unclassified rate tertile 3 (N = 17 299)
	OR (95% CI)	OR (95% CI)	OR (95% CI)	OR (95% CI)
White	Ref.	Ref.	Ref.	Ref.
Asian	0.791***	1.100	0.721***	0.552***
	(0.676–0.925)	(0.794–1.524)	(0.568–0.917)	(0.373–0.816)
Black	0.824***	0.698***	0.810***	0.738***
	(0.782–0.869)	(0.615–0.793)	(0.736–0.891)	(0.655–0.832)
American Indian	0.812***	1.126	0.764**	0.642*
	(0.695–0.950)	(0.838–1.513)	(0.601–0.969)	(0.404–1.020)
Non-Hispanic	Ref.	Ref.	Ref.	Ref.
Hispanic	0.846***	0.969	0.824***	1.029
	(0.799–0.897)	(0.851–1.104)	(0.752–0.904)	(0.876–1.209)
Male	Ref.	Ref.	Ref.	Ref.
Female	1.396***	1.413***	1.590***	1.371***
	(1.354–1.438)	(1.323–1.509)	(1.513–1.672)	(1.270–1.479)
Age: <20	0.898*	1.388**	0.847	0.650***
	(0.790–1.020)	(1.080–1.784)	(0.682–1.052)	(0.494–0.856)
Age: 20–29	Ref.	Ref.	Ref.	Ref.
Age: 30–39	1.083***	1.063	1.193***	1.209***
	(1.033–1.135)	(0.951–1.187)	(1.100–1.293)	(1.083–1.351)
Age: 40–49	1.135***	1.217***	1.255***	1.238***
	(1.082–1.191)	(1.090–1.359)	(1.157–1.362)	(1.103–1.390)
Age: 50–59	1.116***	1.225***	1.246***	1.185***
	(1.063–1.171)	(1.097–1.368)	(1.149–1.352)	(1.056–1.330)
Age: 60 and above	0.896***	1.082	0.941	0.827***
	(0.844–0.952)	(0.948–1.235)	(0.853–1.039)	(0.718–0.951)
Age: Unknown	1.558	–	3.907	0.454
	(0.485–5.001)		(0.701–21.76)	(0.0477–4.319)
Not married	Ref.	Ref.	Ref.	Ref.
Married	1.107***	1.120***	1.223***	0.926*
	(1.068–1.146)	(1.039–1.208)	(1.155–1.295)	(0.849–1.011)
Education: < High School	Ref.	Ref.	Ref.	Ref.
Education: High School	1.002	1.082*	1.011	0.956
	(0.962–1.044)	(0.986–1.188)	(0.944–1.082)	(0.867–1.055)
Education: Some College	1.080***	1.207***	1.108***	0.977
	(1.031–1.131)	(1.089–1.338)	(1.026–1.196)	(0.872–1.095)
Education: College	1.117***	1.333***	1.185***	0.958
	(1.052–1.186)	(1.171–1.517)	(1.074–1.308)	(0.824–1.114)
Education: Unknown	0.948	1.035	0.898	0.804
	(0.855–1.051)	(0.828–1.294)	(0.762–1.059)	(0.616–1.048)
County income tertile 1	Ref.	Ref.	Ref.	Ref.
County income tertile 2	0.495***	0.888*	0.930	0.688***
	(0.472–0.519)	(0.784–1.006)	(0.848–1.020)	(0.616–0.768)
County income tertile 3	0.385***	0.701***	0.881**	0.574***
	(0.366–0.405)	(0.611–0.804)	(0.793–0.979)	(0.501–0.657)

Abbreviation: CI, confidence interval.

Notes: The sum of observations in unclassified rate tertiles does not equal the number of observations in the overall model, because the tertiles are based on non-zero unclassified rate. P values are indicated using stars: ***P < .01, **P < .05, *P < .1.

Geographic disparities

We evaluated geographical variation in drug overdose rates and the proportion of unclassified drug overdoses in 2016. Figure 1 shows county-level drug overdose rates per 100 000 people. There is variation within and across states. Areas that appear to be most impacted by the drug epidemic include the Appalachian states of Kentucky, Ohio, Pennsylvania, and West Virginia; rural Western United States; and New England.

Figure 1.

Drug overdose rates, by county, 2016.

In Figure 2, we highlight county-level variation in the percentage of drug overdoses without a classified drug. Although we observe significant variation within states, there are stark between-state differences. States with high levels of unclassified overdoses include Alabama, Arkansas, Indiana, Louisiana, Mississippi, Missouri, and Pennsylvania. It is unclear what exactly is driving this between-state variation, but it appears that most of the states with high rates of missing classification data rely, at least to some extent, on county coroners for death investigation rather than trained medical examiners.¹⁵

Figure 2.

Unclassified drug overdose proportions, by county, 2016.

These observations are generally supported by Figure 3, which shows the distribution of state-level proportions of unclassified drug overdoses in 2016. Consistent with Figure 2, Alabama, Indiana, Louisiana, and Pennsylvania stand out for their low drug classification rates. However, it is now more obvious that states like Delaware, Montana, and Nebraska are also marked by high rates of unclassified overdoses. From Figures 2 and 3, it appears that counties and states in the northeastern and mid-Atlantic regions of the US suffer less from low classification rates, with Pennsylvania being an obvious exception.

Figure 3.

State-level overdose rates and proportions of unclassified drug overdoses, 2016.

Economic disparities

We assessed economic disparities in individual-level and county-level regression models as well as visualized them in trend graphs. In individual-level regression analyses, we found that a county’s economic standing is a powerful predictor of specific drug identification. Decedents from higher-income counties are less likely to have an unclassified drug overdose, with decedents from counties in the highest-income tertile being 61.5% and from counties in the middle tertile being 50.5% less likely to have missing drug information, relative to decedents from counties in the lowest income tertile (Table 2, Column 1). This relationship holds in all 3 tertiles of unclassified overdose rates (Table 2, Columns 2–4). In individual-level analyses stratified by county income tertiles (Table 3), individual predictor estimates are consistent with our findings in the overall model (Table 2, Column 1).

Table 3.

Individual predictors of unclassified drug overdoses by county-level income tertiles, 2014–2016

	County Income tertile 1 (N = 21 367)	County Income tertile 2 (N = 48 274)	County Income tertile 3 (N = 93 389)
	OR (95% CI)	OR (95% CI)	OR (95% CI)
White	Ref.	Ref.	Ref.
Asian	0.758	0.675*	0.742***
	(0.369 – 1.557)	(0.451 – 1.011)	(0.614 – 0.898)
Black	0.630***	0.718***	0.863***
	(0.557 – 0.712)	(0.647 – 0.798)	(0.796 – 0.936)
American Indian	1.053	0.935	0.783**
	(0.736 – 1.506)	(0.718 – 1.217)	(0.618 – 0.993)
Non–Hispanic	Ref.	Ref.	Ref.
Hispanic	0.979	0.643***	0.914**
	(0.828 – 1.158)	(0.567 – 0.729)	(0.845 – 0.988)
Male	Ref.	Ref.	Ref.
Female	1.427***	1.387***	1.426***
	(1.320 – 1.543)	(1.314 – 1.463)	(1.366 – 1.488)
Age: <20	0.921	0.915	0.874
	(0.646 – 1.314)	(0.721 – 1.161)	(0.732 – 1.043)
Age: 20–29	Ref.	Ref.	Ref.
Age: 30–39	1.161**	1.082*	1.083**
	(1.027 – 1.312)	(0.993 – 1.180)	(1.014 – 1.157)
Age: 40–49	1.187***	1.107**	1.161***
	(1.048 – 1.344)	(1.013 – 1.209)	(1.086 – 1.242)
Age: 50–59	1.106	1.173***	1.085**
	(0.973 – 1.256)	(1.074 – 1.280)	(1.015 – 1.161)
Age: 60 and above	0.833**	0.878**	0.906**
	(0.707 – 0.980)	(0.788 – 0.979)	(0.833 – 0.985)
Age: Unknown	–	2.272	1.090
		(0.205 – 25.13)	(0.283 – 4.205)
Not married	Ref.	Ref.	Ref.
Married	1.056	1.141***	1.115***
	(0.967 – 1.153)	(1.074 – 1.213)	(1.060 – 1.173)
Education: < High school	Ref.	Ref.	Ref.
Education: High school	1.040	0.976	1.019
	(0.948 – 1.143)	(0.909 – 1.048)	(0.958 – 1.084)
Education: Some college	1.118*	1.071*	1.077**
	(1.000 – 1.249)	(0.988 – 1.161)	(1.006 – 1.153)
Education: College	1.177*	1.075	1.114**
	(0.992 – 1.398)	(0.964 – 1.200)	(1.024 – 1.211)
Education: Unknown	1.093	1.023	0.871*
	(0.821 – 1.456)	(0.852 – 1.230)	(0.749 – 1.012)

Abbreviations: CI, confidence interval; OR, odds ratio.

Notes: The sum of observations across county income tertiles (N = 163 030) is less than the number of observations in the overall model (N = 163 091) because of collinearity with several of the county fixed effects. P values are indicated using stars: ***P < .01, **P < .05, *P < .1.

In county-level regression analyses, there was no evidence of association between counties’ economic standing and having any unclassified drug overdoses (Table 4, Column 1). However, in analyses stratified by drug investigation system, we found strong evidence that highest-income counties with the decentralized medical examiner system are 60.7% less likely than lowest-income counties to have unclassified drug overdoses. We also find some evidence that high-income counties with decentralized coroners have a higher likelihood of having an unclassified drug overdose.

Table 4.

County-level predictors of unclassified drug overdoses, overall and by drug investigation system, 2014–2016

	Overall (N = 7235)	Centralized Medical Examiner (N = 1519)	Decentralized Medical Examiner (N = 940)	Hybrid System (N = 3401)	Decentralized Coroner (N = 1375)
	OR (95% CI)	OR (95% CI)	OR (95% CI)	OR (95% CI)	OR (95% CI)
Income tertile 1	Ref.	Ref.	Ref.	Ref.	Ref.
Income tertile 2	1.083	0.813	0.756	1.230*	1.348
	(0.926 – 1.266)	(0.547 – 1.209)	(0.509 – 1.123)	(0.984 – 1.538)	(0.911 – 1.994)
Income tertile 3	0.925	0.652	0.393***	1.068	1.547*
	(0.760 – 1.125)	(0.392 – 1.085)	(0.227 – 0.680)	(0.810 – 1.407)	(0.964 – 2.481)
Population tertile 1	Ref.	Ref.	Ref.	Ref.	Ref.
Population tertile 2	2.202***	2.055***	1.916***	2.250***	2.285***
	(1.878 – 2.581)	(1.278 – 3.304)	(1.252 – 2.933)	(1.806 – 2.803)	(1.571 – 3.325)
Population tertile 3	7.062***	5.564***	6.613***	7.271***	13.24***
	(5.670 – 8.794)	(3.165 – 9.782)	(3.526 – 12.40)	(5.315 – 9.946)	(7.130 – 24.60)
Black tertile 1 (%)	Ref.	Ref.	Ref.	Ref.	Ref.
Black tertile 2 (%)	0.815**	0.995	0.679*	0.819*	0.788
	(0.696 – 0.956)	(0.628 – 1.577)	(0.442 – 1.042)	(0.658 – 1.020)	(0.539 – 1.151)
Black tertile 3 (%)	0.817*	0.912	0.578*	0.851	1.154
	(0.661 – 1.010)	(0.503 – 1.654)	(0.330 – 1.014)	(0.639 – 1.134)	(0.615 – 2.166)
Other race tertile 1 (%)	Ref.	Ref.	Ref.	Ref.	Ref.
Other race tertile 2 (%)	1.021	0.884	1.442*	1.036	0.949
	(0.859 – 1.215)	(0.486 – 1.608)	(0.944 – 2.204)	(0.818 – 1.313)	(0.619 – 1.454)
Other race tertile 3 (%)	1.248**	1.226	1.117	1.382**	1.236
	(1.004 – 1.550)	(0.650 – 2.311)	(0.635 – 1.965)	(1.015 – 1.880)	(0.725 – 2.109)
Hispanic tertile 1 (%)	Ref.	Ref.	Ref.	Ref.	Ref.
Hispanic tertile 2 (%)	1.166*	1.178	1.092	1.117	1.526**
	(0.988 – 1.375)	(0.550 – 2.521)	(0.719 – 1.657)	(0.895 – 1.394)	(1.043 – 2.232)
Hispanic tertile 3 (%)	1.505***	1.900	1.205	1.151	2.422***
	(1.211 – 1.870)	(0.840 – 4.298)	(0.698 – 2.079)	(0.843 – 1.572)	(1.441 – 4.070)
Overdose rate tertile 1	Ref.	Ref.	Ref.	Ref.	Ref.
Overdose rate tertile 2	1.660***	1.795***	1.619**	1.703***	1.626***
	(1.444 – 1.908)	(1.174 – 2.746)	(1.097 – 2.389)	(1.403 – 2.065)	(1.185 – 2.231)
Overdose rate tertile 3	2.863***	3.735***	2.769***	2.725***	2.895***
	(2.450 – 3.346)	(2.449 – 5.697)	(1.839 – 4.168)	(2.181 – 3.405)	(2.013 – 4.165)

Abbreviations: CI, confidence interval; OR, odds ratio.

Notes: Tertiles were calculated at the county level. Tertile ranges for each predictor are: Income 1 = 22 045–42 540, Income 2 = 42 542–51 694, Income 3 = 51 697–134 609; Population 1 = 87–23 809, Population 2 = 23 817–65 585, Population 3 = 65 622–10 100 000; % Black 1 = 0.0–0.018, % Black 2 = 0.018–0.081, % Black 3 = 0.081–0.859; % Other race 1 = 0.003–0.014, % Other race 2 = 0.014–0.031, % Other race 3 = 0.031–0.950; % Hispanic 1 = 0.005–0.027, % Hispanic 2 = 0.027–0.069, % Hispanic 3 = 0.069–0.963; Overdose rate 1 = 9.5e⁻⁶–0.00011, Overdose rate 2 = 0.00011–0.00019, Overdose rate 3 = 0.00019–0.0115. Number of observations in the overall estimation model (N = 7235) is smaller than in our data (N = 7244), because of collinearity with 1 of the county fixed effects. P values are indicated using stars: ***P < .01, **P < .05, *P < .1.

In Figures 4 and 5, we explore annual county-level trends in drug overdoses and unclassified drug overdose rates by county’s median household income tertile (based on 2008 median household income). In Figure 4, we highlight annual trends in drug overdose rates at the county level from 2008 to 2016 by income tertile. We observe increasing trends in drug overdose rates for all 3 income tertiles. Low-income counties experience the highest drug overdose rates across the entire time period. However, the drug overdose rates of the 2 highest-income groups of counties appears to converge over time with the drug overdose rates of the lower income tertiles. In Figure 5, we show annual county-level trends in the proportion of unclassified drug overdoses by income tertile. We observe relatively consistent declines in the unclassified proportion in all 3 income tertiles, although counties in the highest-income tertile have lower rates of unclassified drug overdose deaths across the entire time period.

Figure 4.

Overdose death rate over time, by county income tertiles.

Notes: Income tertiles are based on 2008 county median household income: tertile 1: $19 182–39 392; tertile 2: $39 392–47 376; tertile 3: $47 376–$111 582.

Figure 5.

Proportion of unclassified overdose deaths over time, by county income tertiles.

Notes: Income tertiles are based on 2008 county median household income: tertile 1: $19 182–39 392; tertile 2: $39 392–47 376; tertile 3: $47 376–$111 582.

DISCUSSION

One important role medical informatics plays in the opioid crisis is supporting public health and epidemiological efforts. Moving forward, advanced informatics methods and modeling will play a role in the emerging challenges of mining and analyzing opioid-related data, with the National Institutes of Health Helping to End Addiction Long-term funding plan specifically identifying a need for data coordination and implementation.¹ Focusing on the analytic challenges of opioid-related data, this work sought to iteratively advance our understanding about unclassified overdose data to improve future studies by fully incorporating the context of this data point. A significant proportion of drug overdoses in the US mortality data is missing information about what specific drugs are involved, thus impacting our current understanding of the opioid epidemic. Existing literature demonstrates a significant increase in opioid overdose rates when unclassified drug overdose rates are incorporated into overall counts,⁹ doubling opioid overdose rates to as high as 86%.¹⁶ This knowledge gap undermines large-scale data automated monitoring, analysis, and effective implementation of population level interventions. To our knowledge, no published literature has sought to illuminate patterns of disparity in missing drug data for overdose mortality records. This study examined whether such inequities exist by investigating individual, geographic, and economic phenotypes of unclassified drug overdose deaths. Our overarching hypothesis was that these 3 factors combined to undermine reporting of specific drug involvement in fatal overdoses, reflecting systemic health inequities.

Our combined data approach identified significant decedent- and county-level predictors of unclassified drug overdose deaths. Specific individual factors associated with unclassified drug overdoses included White race, non-Hispanic ethnicity, female gender, ages 30–59, and a college education. The decedent being from a poorer county is associated with a greater likelihood of having an unclassified drug overdose, although individual-level predictors remain mostly unchanged across county-level income tertiles. There are clear geographic differences in unclassified drug mortality, with substantial between-state variation. Over time, counties in the highest income tertile consistently had lower rates of unclassified drug overdoses than the lower tertiles, although all 3 show substantial improvements in specific drug identification. Our findings make a significant contribution to the existing knowledge on the patterns of missing information in drug overdoses, as previous work has focused on factors associated with reporting, such as centralized coroner systems, rather than individual decedent factors or county-level economic characteristics.^14,16

The disparities in the missing drug information that our research uncovered have important implications for our overall understanding of the opioid crisis as well as for the development of interventions. In this vein, informatics solutions can play a vital role. First, reports on the opioid epidemic may mask the true picture if missing drug information is simply disregarded. Although the CDC reports a national rate of unclassified drug overdoses at 12% (2017), we find statistically significant differences among income tertiles at the county level, with 15% for the wealthiest tertile, 21% for the mid-tertile, and 23% for the lowest tertile in 2016. Together with our other findings on the influence of county income levels on missing information in drug overdoses, this implies that poorer counties may be marked by higher rates of underreported opioid overdoses, masking an even deeper crisis in poverty-stricken areas than we currently realize. In addition, assessments of the opioid crisis by demographic characteristics may be skewed if a nontrivial portion of unclassified drug overdoses does in fact involve opioids. For example, the impact of the opioid epidemic on women may be underestimated as they are more likely than men to be prescribed and use opioid analgesics²⁰ and we find that females are approximately 40% more likely to have missing drug information in the case of overdosing. The Fast-Track Action Committee on Health Science and Technology Response to the Opioid Crisis (Opioid FTAC) recently made a recommendation to assess opioid morbidity and mortality more quickly and accurately.¹⁹ To address this, our findings may be used by large-scale data analytics systems to make more accurate assessments and predictions in this intensifying public health crisis.

Second, our findings on health inequities in missing data on specific drug involvement can inform more targeted interventions and an associated allocation of resources across communities and geographies. With critical decisions being made regarding resource allocation for community programs to address the opioid crisis, gaps in reporting may be underappreciated. Understanding opioid addiction patterns requires indirect measurements, such as mortality data, that are vulnerable to analytic issues. Identifying the patterns and factors associated with individual life expectancy and environmental risks builds the foundation for community-level opioid risk models and an opportunity to support vulnerable health groups. Current national community approaches to the opioid crisis involve programs that are highly dependent on state and federal funding, including syringe service programs, prevention programs in K–12 schools, and medication-assisted treatment for populations.¹⁴ By focusing on individual and community phenotypes, new data-driven policy approaches could meaningfully direct resource allocation toward disparate populations that have been most affected by this data gap,²¹ especially as it has been highlighted in the literature that a quantitative and qualitative understanding of the data gaps may directly influence policy and funding support.²² County- and state-level differences in the missing drug involvement data that we report also point to the need for jurisdiction-based policy and informatics solutions.

Limitations

Counties that do not have any overdoses are not visible in any of our analyses or maps to study the prevalence of unclassified drug overdoses. This means that we cannot provide a full geographic picture of individual, geographic, or economic factors for these counties. If these counties experience overdoses in the future, they could be included in further studies. We hypothesize that opioid involvement in unclassified drug overdoses is likely and support it with evidence from prior literature, but our data does not allow us to confirm a particular drug type. Further, we are unable to include a rich set of time-varying county-level controls due to data unavailability. In our models, we accounted for state fixed effects to account for time-invariant geographic factors that were unobservable, as were year fixed effects to capture secular trends, but there may be additional residual confounding that we cannot account for that is time-varying at the county-level. Additionally, mortality data does not provide information on decedents’ income level, and individual-level economic disparities could have been described more accurately if such data was available for each decedent. Finally, we cannot include a longer period of time in trend analyses because we found data quality issues in years before 2008 associated with whether states had adapted to the national death certificate standard released in 2003.

CONCLUSION

Our work found disparities among individual, geographic, and economic factors associated with unclassified drug overdose deaths. Emerging literature demonstrates that the extent of the opioid crisis may be underestimated due to unclassified drug overdose deaths with opioid involvement. Public health informatics approaches rely on accurate data collection to develop surveillance and analytic approaches for the opioid crisis. Our work sought to determine whether patterns of health inequity exist in this missing information problem. Findings from our multi-method analytic approach consistently demonstrated higher rates of unclassified drug overdoses in decedents of female gender, White race, non-Hispanic ethnicity, with college education, of ages 30–59, and from poorer counties. Despite trend analyses demonstrating that unclassified drug overdose deaths rates have reduced over time, gaps persist between the richest and poorest counties. There is also clear geographic variation in unclassified drug mortality. Given the essential role of mortality data in measuring the extent of the opioid epidemic, it is important to understand the inequities underlying the missing data on specific drug involvement. Our findings can inform policy and informatics solutions to more accurately assess opioid-related mortality and to develop targeted interventions with more equitable resource allocations.

FUNDING

Research reported in this article was supported by NIH National Institute of Biomedical Imaging and Bioengineering of the National Institutes of Health under grant number K08 EB022631 and NIH Office of the Director under grant number 5DP 5OD021338.

The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

CONTRIBUTORS

Conception or design of the work: MA, AB, AD, EH.

Data analysis and interpretation: MA, AB, AD, EH.

Drafting the article: MA, AB, AD, EH.

Critical revision of the article: MA, AB, AD, EH.

Final approval of the version to be published: MA, AB, AD, EH.

CONFLICT OF INTEREST STATEMENT

None declared.