Association between undetermined intent death classification and county death investigation system for drug poisoning deaths in nine US jurisdictions

Abstract

Background

Both suicides and unintentional drug overdose deaths have risen dramatically over the last two decades in the United States. However, the classification of death manner intent can be a challenge for death investigators in drug poisoning deaths. This study assessed whether county death investigation system type was a predictive factor in the likelihood of “undetermined” intent death classifications.

Methods

This study examines the association between undetermined intent classifications and death investigation systems among drug poisoning deaths in the United States. With novel data from the Centers for Disease Control and Prevention and from the State Unintentional Drug Overdose Reporting System in nine states, we used logistic regression models to analyze whether county coroner systems were differentially associated with “undetermined” intent death classifications as compared to county medical examiner systems.

Results

County coroner systems were associated with increased odds of undetermined intent classifications in drug poisoning deaths as compared to county medical examiner systems [odds ratio: 1.79; 95% confidence interval: 1.43, 2.27]. Even after a full set of county and individual decedent controls, the association remained [adjusted odds ratio: 1.60; 95% confidence interval: 1.25, 2.05].

Conclusions

Future work should examine ways to identify suicides in drug poisoning deaths, as the consequences of misclassification echo forward into data used for prevention. The type of death investigation system may be an important factor in how drug poisoning deaths are classified.

Background

Death investigation is an essential tool for epidemiology. Mortality data influence research, policy, and prevention strategies.^1–4 Accurate death classification is therefore central to public health, but it can also be difficult to achieve. Studies show substantial variation in how death investigators classify death across many domains.^5,6

In the United States, two recent trends highlight the importance of accurate death classification. First, drug overdose deaths have risen dramatically over the last two decades.⁷ Annual overdose deaths have only recently dropped below 100,000.^8,9 Second, suicide deaths also rose over the same period, and annual deaths now total nearly 50,000.^10,11 Because of the strong correlation between substance use and suicidal ideation, death trends over the period may be difficult to accurately identify.^1,12,13 For example, drug overdoses form the majority of deaths for whom coroners/medical examiners classify intent as “undetermined.”² Researchers have long believed suicide is underreported, and many deaths with intent recorded as undetermined may, in fact, be suicides.^2,14,15

To address the difficulty of classification in drug poisoning deaths, Rockett et al. (2014) proposed a new definition of “drug self-intoxication” deaths for the Centers for Disease Control and Prevention (CDC) to add to the National Violent Death Reporting System (NVDRS).¹⁶ Several years later, the CDC expanded NVDRS with the State Unintentional Drug Overdose Reporting System (SUDORS) to include unintentional drug overdose deaths. Although the expansion of NVDRS differs from the vision outlined in Rockett et al. (2014), the expansion has substantially increased the ability to identify the determinants of death manner classifications in drug poisoning deaths through the addition of unintentional drug overdose deaths in SUDORS.

Every state reports data to the CDC, but death investigation across the United States is not standardized. Individual states are responsible for their own investigation systems.¹⁷ Twenty-two states currently employ either centralized or county/district-based medical examiner offices, and the remainder employ county/district-based coroner offices or a county-based mixture of coroner and medical examiner offices.¹⁸ Although occupational standards vary by jurisdiction, medical examiners are typically appointed physicians, and coroners are elected laypeople.^17,19 However, requirements vary by state. Several states require coroners to be physicians, and others require autopsies to be performed by a forensic pathologist.²⁰ Medical examiner and coroner offices may also differ in their average number of referred deaths, computer management systems, and access to forensic technology, among other factors.²¹ These differences may influence death classification at the investigation level.

This study examines the association between undetermined intent classifications and death investigation systems among drug poisoning deaths. The novel, multi-state SUDORS data allow for county level examination. The county focus allows for the control of both county-level potential confounders and individual decedent characteristics.

Methods

This study used logistic regression to analyze the association between undetermined death manner intent and county death investigation system type (coroner vs. medical examiner) among all drug poisoning deaths of either undetermined intent, suicide, or unintentional overdose classifications. Suicide and undetermined intent drug poisoning deaths were sourced from the Restricted Access to Data (RAD) file of the National Violent Death Reporting System (NVDRS), maintained by the Centers for Disease Control and Prevention (CDC). Unintentional drug overdose deaths were sourced from the State Unintentional Drug Overdose Reporting System (SUDORS) in nine states (Georgia, Illinois, Indiana, Louisiana, Maryland, Minnesota, North Carolina, Washington, and Wisconsin). Undetermined intent and suicide deaths are only available in NVDRS, and unintentional drug overdose deaths are only available in SUDORS. The division of deaths across data systems therefore required the combination of both systems.

Because individual-level SUDORS data are unavailable to researchers at the national level, the nine sample states were chosen both as a convenience sample and for their variation in death investigation systems. The SUDORS data were obtained through state data sharing agreements. The NVDRS data were also restricted to the nine sample states for comparability. The sample included decedents of any age with a death date in the years 2019–2021.

County death investigation system information was sourced from the CDC.²² An indicator variable for a county coroner system was set to one for any “Non-Medical Examiner Death Investigation System.” To address potential confounders, the regression analysis controlled for decedent age, race, sex, and indicators for whether an autopsy was performed, an opioid forensic toxicology test was performed, an alcohol toxicology test was performed, a suicide note was recorded in the death investigation, and whether the investigation included a free-form text narrative from the death investigator. Response categories of “no” and “unknown” were set to zero for the alcohol and opioid test variables, as either response indicates a potential lack of information available to data abstractors. Each of these control variables was sourced from NVDRS/SUDORS.

Regression models also controlled for county population size grouped into categories of 0–25K residents, 25K-250K residents, or greater than 250K residents, sourced from the Census Bureau; annual county poverty rates, sourced from the Census Bureau; annual county unemployment rates, sourced from the Bureau of Labor Statistics; and the 2023 rurality index, sourced from the Department of Agriculture. The county poverty rate is defined as the percentage (from 0–100) of households in a county whose total income falls below the national poverty threshold as determined by the US Census Bureau. The annual unemployment rate is defined as the percentage (from 0–100) of the total labor who was employed in a county (total unemployed persons divided by total labor force, multiplied by 100). The Bureau of Labor Statistics estimates annual unemployment rates from seasonally adjusted monthly rates in the Current Population Survey. The rurality index is a discrete scale from 1–9 in ascending order of rurality. All analyses were conducted in R version 4.4.1.

Summary statistics were reported for death manner, decedent-level covariates, and county-level covariates. Death manner included undetermined intent, suicide, and unintentional drug overdose. Decedent-level covariates included age, sex, race/ethnicity, whether opioid or alcohol forensic toxicology test was performed, whether a suicide note was present, whether an autopsy (full or partial) was conducted, and whether a coroner/medical examiner narrative was available in NVDRS/SUDORS. County-level covariates included death investigation system type (coroner or medical examiner), population, the unemployment rate, the poverty rate, and the rurality index.

Control variables were successively added in groups to identify their role as potential confounders in regression models. Models were run with (a) no controls, (b) county-level controls (population, poverty rate, unemployment rate, rurality index), (c) county and death investigation controls (autopsy, alcohol tested, opioid tested, suicide note, investigator text narrative availability), and (d), county, death investigation, and decedent-level controls (age, sex, and race). Preferred models also included state indicator variables to adjust for average variation by state and identify variation across county systems within states.

Patient and Public Involvement

This research was conducted without patient and public involvement because all study participants were deceased.

Results

Figure 1 contains a flowchart with the sample size by death manner, county death investigation system, and state. The sample contained 3,080 suicide deaths, 3,887 undetermined intent deaths, and 48,463 unintentional drug overdose deaths.

Figure 1.

Sample Size by Death Manner, Death Investigation System, and State for United States Drug Poisoning Deaths in Nine States from 2019–2021.

Note: NVDRS=“National Violent Death Reporting System”, SUDORS=“State Unintentional Drug Overdose System”, C=“coroner system”, ME=“medical examiner system”.

Figure 2 displays a map of county death investigation systems in the United States for the nine sample states. Counties in the sample states span a range of death investigation systems. For example, Indiana relies exclusively on county coroners, while North Carolina has a centralized medical examiner system.

Figure 2.

United States Death Investigation Systems in Nine Sample States.

Table 1 displays summary statistics for the sample at the decedent level. Most deaths were categorized as unintentional drug overdoses (87%), followed by undetermined intent (7%) and suicides (6%). Approximately 33% of the sample died in a county with coroner system. The mean decedent age was 43 years old. Just over 69% of decedents were male. The most common race was White at 70% of the sample, followed by Black at 26%. A slightly higher proportion of investigations recorded an opioid forensic toxicology test (79%) than an alcohol toxicology test (76%). A suicide note was recorded in slightly under 2% of the investigations.

Table 1.

Summary Statistics for United States Drug Poisoning Deaths by Death Manner in Nine States from 2019–2021.

Investigation System	Any (N=55,430)		Medical Examiner (N=37,198)		Coroner (N=18,232)
Variable	n	SD	n	SD	n	SD
undetermined intent	3,887	25.5	3,329	28.6	558	17.2
unintentional overdose	48,463	33.2	31,983	34.7	16,480	29.5
suicide	3,080	22.9	1,886	21.9	1,194	24.7
coroner	18,232	47.0	0	0.0	18,232	0.0
population <25K	1,290	15.1	505	11.6	785	20.3
population 25K-250K	16,682	45.9	8,795	42.5	7,887	49.5
population ≥250K	37,458	46.8	27,898	43.3	9,560	49.9
autopsy	37,346	46.8	24,969	47.0	12,467	46.5
opioid tested	44,011	40.4	29,499	40.5	14,512	40.3
alcohol tested	42,128	42.7	31,444	36.2	10,684	49.3
suicide note	1,148	14.2	747	14.0	401	14.7
no CME report	1,270	15.0	508	11.6	762	20.0
male	38,177	46.3	26,036	45.8	12,141	47.2
American Indian	1,166	14.4	944	15.7	222	11.0
Asian	620	10.5	504	11.6	116	8.0
Black	14,469	43.9	11,181	45.9	3,288	38.4
Pacific Islander	76	3.7	59	4.0	17	3.1
White	38,885	45.8	24,398	47.5	14,487	40.4
unspecified race	542	9.8	451	10.9	91	7.0
	Mean	SD	Mean	SD	Mean	SD
age	43.0	13.5	43.3	13.7	42.3	13.1
unemployment rate	5.7	2.1	5.8	2.1	5.6	2.2
poverty rate	13.3	5.0	12.8	4.8	14.3	5.2
rurality	1.9	1.6	1.7	1.5	2.3	1.8

Note: SD = standard deviation. For binary variables, n is the number of cases of the variable equal to 1, and the standard deviation is expressed in percent terms (with variables multiplied by a factor of 100). The ‘no CME report’ variable is set to 1 if no coroner/medical examiner narrative was available.

Table 2 presents the results of logistic regressions of the undetermined intent classification on various sets of control variables. The simplest model (1) only includes an indicator for whether a death occurred in a county with a coroner investigation system. The odds ratio on the indicator was 0.32 and statistically significant, which means deaths in coroner county systems were less likely to be designated with undetermined intent than deaths in medical examiner systems. However, this result is driven by the inclusion of Maryland, which has a centralized medical examiner system but unusually high rates of undetermined death classifications.^5,6 Model (2) includes state fixed effects to remove average variation by state and focus instead on variation across county systems within states. After controlling for state fixed effects, the odds ratio on the coroner system indicator rose to 1.79. For the interpretation, drug poisoning deaths had 79% higher odds of undetermined intent classifications in counties with a coroner system as compared to a medical examiner system.

Table 2.

Logistic Regression Results (Odds Ratios) for Undetermined Death Classification among United States Drug Poisoning Deaths in Nine States from 2019–2021.

	(1)	(2)	(3)	(4)	(5)
coroner	0.32***	1.79***	1.52***	1.61***	1.60***
	(0.29–0.35)	(1.43–2.27)	(1.20–1.95)	(1.26–2.06)	(1.25–2.05)
unemployment rate			0.93**	0.94**	0.94*
			(0.89–0.99)	(0.89–0.98)	(0.90–0.99)
rurality			0.96*	0.96*	0.96*
			(0.92–0.99)	(0.93–0.99)	(0.93–1.00)
poverty rate			1.01	1.01	1.01
			(1.00–1.02)	(1.00–1.02)	(1.00–1.02)
population 25K-250K			0.97	0.96	0.94
			(0.69–1.37)	(0.68–1.35)	(0.67–1.34)
population ≥250K			0.62*	0.62**	0.61**
			(0.43–0.90)	(0.43–0.89)	(0.42–0.89)
autopsy				1.29***	1.27***
				(1.14–1.46)	(1.12–1.44)
opioid tested				1.15*	1.14
				(1.01–1.31)	(1.00–1.29)
alcohol tested				0.99	1.00
				(0.84–1.15)	(0.86–1.17)
suicide note				0.46***	0.46***
				(0.29–0.74)	(0.28–0.73)
no CME report				0.52*	0.52*
				(0.29–0.91)	(0.30–0.92)
age					0.99***
					(0.99–1.00)
male					0.77***
					(0.71–0.83)
American Indian					0.73
					(0.49–1.09)
Asian					1.10
					(0.75–1.61)
Black					1.03
					(0.95–1.12)
Pacific Islander					1.06
					(0.40–2.86)
unspecified race					0.93
					(0.58–1.49)
constant	0.12***	0.01***	0.02***	0.02***	0.03***
	(0.11–0.13)	(0.01–0.02)	(0.02–0.04)	(0.01–0.03)	(0.02–0.05)
Year FEs	Y	Y	Y	Y	Y
State FEs	N	Y	Y	Y	Y
n	55,430	55,430	55,430	55,430	55,423
Log likelihood	−13,677	−10,648	−10,600	−10,575	−10,537

Note: ***, **, and * respectively denote significance at the 0.05, 0.01, and 0.001 levels. 95% confidence intervals are reported in parentheses below each odds ratio point estimate.

The next three models successively add controls for county and decedent characteristics. Model (3) includes the county unemployment rate, the rurality index, the percentage of the population below the poverty line, and population size indicators. With the addition of the controls, the odds ratio on the coroner system indicator fell from 1.79 to 1.52 but remained statistically significant with a 95% confidence interval of 1.20–1.95. The odds ratio on the unemployment rate was 0.93 and significant. For the interpretation, every one percentage point increase in the unemployment rate reduced the odds of an undetermined death classification by a factor of 0.93 after controlling for the other county regressors. The odds ratio on the rurality index was 0.96 and significant. Note that a higher index value indicates a more rural county. Lastly, deaths in counties with a population of ≥250K had reduced odds of an undetermined classification by a factor of 0.62 compared to deaths in counties with a population of 0–25K.

Models (4) and (5) add decedent controls. Model (4) adds controls meant to capture possible differences in informational resources available to death investigators: whether an autopsy was performed (either full or partial), whether opioid or alcohol forensic toxicology tests were reported as conducted in the investigation, whether a suicide note was present, and whether a free-form text narrative on the individual circumstances of the death was available from a coroner or medical examiner. The odds ratio on the coroner system indicator rose from 1.52 in model (3) to 1.61 in model (4) after the addition of the controls. An autopsy raised the odds of an undetermined death classification by a factor of 1.29. In contrast, the absence of an investigator narrative report was associated with reduced odds by a factor of 0.52, and the presence of a suicide note was associated with reduced odds by a factor of 0.46.

Lastly, model (5) also includes decedent age, sex, and race. The odds ratio on the coroner investigation system indicator, as well as the odds ratios on the other regressors, remained largely unchanged from those in model (4) after the controls. Decedent race was not associated with the odds of an undetermined classification, but age and male sex were. Each one-year increase in decedent age reduced the odds by a factor of 0.99. The odds of an undetermined death classification were 23% lower for male than female decedents.

Discussion

The type of county death investigation system is strongly associated with the odds of an undetermined intent death classification. In this discussion, we focus on the results for models (2)-(5), which remove average state variation through fixed effects. The inclusion of state fixed effects shifts the interpretation of results to within state variation in county investigation systems.

In the logistic regression, the odds ratio on the coroner investigation system indicator ranged from a high of 1.79 without control variables to a low of 1.52 with only county-level controls. Across models, the odds ratio was statistically significant with a lower bound of 1.20 on the 95% confidence intervals. The results mean deaths in counties with a coroner system have at least 52% increased odds of an undetermined death classification compared to deaths in counties with a medical examiner system.

One advantage of the case-level data is the ability to control for both county-level and decedent-level factors. Coroner and medical examiner offices may differ in the types of resources available to them, such as funding levels and access to forensic technology. We chose to divide population into categories of 0–25K, 25–250K, and ≥250K to match a Bureau of Justice Statistics survey of death investigation offices, which found differences in funding across offices in counties with these population thresholds.²¹ As another example, Tote et al. (2019) found that coroner systems were associated with higher rates of incomplete toxicology reports.²³ Potential resource differences may confound a simple association between the county death investigation system and the rate of undetermined death classification. However, controlling for county-level factors only reduced the odds ratio on the coroner system indicator from 1.79 to 1.52.

The individual-level data also contain information about individual deaths that may further address the question of resource differentials. Model (3) adds, for example, controls for whether an autopsy was conducted and whether opioid or alcohol forensic toxicology tests were recorded as conducted. The rate of autopsy in the US has fallen by approximately 50% since the 1970s, and the decline may be driven in part by financial constraints.²⁴ In the international context, Kapusta et al. (2011) found a strong positive association between autopsy rates and suicide rates, which provides suggestive evidence of improved death manner classifications through higher autopsy rates.²⁵

The odds ratio on the autopsy indicator in the full model (5) was 1.27, which means autopsy was positively associated with undetermined death classifications. The association likely captures a response of death investigators to order an autopsy when they are uncertain about the nature of a death. Conversely, the presence of a suicide note reduced the odds of an undetermined classification by 54%, which likely reflects an increase in investigator certainty.

We also included covariates for whether an opioid or alcohol forensic toxicology test were reported as conducted in death investigations. One prior study found a negative association between undetermined death classifications and the citation of one or more specific drugs on a death certificate.²⁶ We included the toxicology test indicators as proxies for information about drug involvement. For example, if toxicology tests inform death manner classifications, coroner systems could have higher rates of undetermined intent classifications because coroners have, on average, fewer resources to conduct toxicology tests. However, the odds ratios on the indicators for whether a test was available (regardless of the test result) were largely insignificant.

Investigation offices may also differ in the types of cases they review. Model (5) controlled for decedent race, age, and sex because prior studies have found differences in death classifications among different demographic groups.^27,28 For example, Huguet et al. (2012) found that Black decedents had higher rates of missing information in the NVDRS than White decedents, and Huguet et al. (2015) found that female decedents were more likely to be classified as undetermined deaths if they had lower educational attainment or a substance use problem.^29,30 Ali et al. (2022) found that Black adolescents had significantly higher odds of undetermined intent death classifications than White adolescents in potential suicide deaths.³¹ Rockett et al. (2006) also found suggestive evidence that suicide rates in Black individuals may be relatively underreported compared to White individuals.³²

In our results, race was not associated with undetermined death classification, but age and sex were. Every-one year increase in age was associated with a 1% decrease in the odds of undetermined death classifications, while males had 23% lower odds than females. One possible explanation for the difference in the results here as compared to prior work is the additional control variables, which may have been confounders in prior work, but the sample selection criteria also differed across studies.

Several studies have identified an association between the type of investigation and either suicide designations^33–35 or the rate at which individual drugs are specified in overdose deaths.^5,23 Our results focused instead on undetermined intent death classifications. We found that coroner systems were strongly associated with undetermined intent classification relative to medical examiner systems, and the association remained after accounting for potential confounders. Suicide may therefore be relatively underreported in counties with coroner systems if a non-zero fraction of undetermined deaths are actually suicides.^36,37

Researchers have proposed several potential paths to improved death classification.² Training and death investigator philosophies appear to be important factors in how investigators classify death, and more standardization across jurisdictions could potentially reduce rates of undetermined intent classification.¹⁷ Additional resources for autopsies and forensic toxicology tests could also aid investigators in more definitive conclusions.²⁶ As a final example, litigation reform could help reduce investigator reluctance to label a death as a suicide in the avoidance of legal liability risk.³⁸

Because SUDORS individual-level data are unavailable to researchers at the national level, the states in this study were chosen as a convenience sample. Results may not generalize to the entire country. Both “no” and “unknown” were coded to zero for the indicators of whether an opioid or alcohol forensic test was recorded in a death investigation, but these tests may have been conducted in some instances without the information made available to data abstractors. The logistic regression results controlled for several potential confounders in the relationship between death investigation system and undetermined death classifications, but other confounders may remain. For example, we were unable to directly measure differences in funding to death investigation offices. This study defined sex as it appears in the NVDRS/SUDORS data systems, but the variable may not accurately capture or distinguish sex from gender.

Lastly, the sample period of 2019–2021 overlapped with the onset of the COVID-19 pandemic. Excess deaths strained the capacity of death investigation offices in the US.³⁹ Although the regression models controlled for state fixed effects and other potential confounders, these controls may not fully address differential capacity constraints induced by the pandemic. Future work should revisit these results as more recent data become available.

Conclusions

This study assessed the relationship between undetermined intent death classifications and county death investigation systems in drug poisoning deaths. The results suggest that coroner and medical examiner offices may differ in how they classify drug poisoning deaths. Coroner systems had higher odds of classifying deaths as undetermined, and these differences remained after controlling for a range of potential confounders. Self-injury mortality has risen dramatically over recent years in the US.⁴⁰ Deaths of undetermined intent may conceal true suicides. Future work should continue to examine ways to identify suicides in drug poisoning deaths. The consequences of misclassification echo forward into data used for prevention. Work to improve these classifications will therefore be foundational in the prevention of future deaths.

Key Messages.

What is already known on this topic

The United States employs a mixture of coroner and medical examiner offices to perform individual death investigations, and these offices may differ in how frequently they classify drug poisoning deaths as “undetermined intent.”

What this study adds

In regression analysis based on a novel data linkage in nine jurisdictions, coroner offices had higher odds of undetermined intent classifications for drug poisoning deaths than did medical examiner offices.

How this study might affect research, practice or policy

Because death classification is central to prevention efforts, future work should identify ways to reduce the rate of undetermined intent classifications in drug poisoning deaths. Coroner and medical examiner offices may face different investigation challenges in drug poisoning deaths.

List of Abbreviations

CDC

Centers for Disease Control and Prevention

NVDRS

National Violent Death Reporting System

SUDORS

State Unintentional Drug Overdose Reporting System

Data Statement

SUDORS data are not available due to stipulations in state data sharing agreements. However, NVDRS data can be obtained by researchers with permission from the Centers for Disease Control and Prevention.