An Epidemic of Complexity: Polysubstance Use Fatal Overdose Typologies Across Cook County, Illinois, Neighborhoods, 2017‒2022

Abstract

Objectives. To examine geospatial polysubstance use overdose typologies and socio-structural determinants among overdose decedents in Cook County, Illinois, from 2017 to 2022.

Methods. Decedents’ addresses were geocoded and aggregated to census tract level (n = 1331). We conducted latent profile analysis to identify census tract profiles based on overdose fatality rates per 1000 census tract population across 9 drug types (alcohol, fentanyl, prescription opioids, heroin, cocaine, benzodiazepines, xylazine, methamphetamine/amphetamine, and other drugs). We used Bayesian spatial multinomial logistic regression to determine what census tract socio-structural determinants could predict polysubstance profile membership.

Results. Three clusters of polysubstance use—low, moderate, and high—emerged with fentanyl, cocaine, and heroin dominating all clusters. In the adjusted model, census tracts with lower median household income were associated with increased odds of belonging to the moderate polysubstance use profile (odds ratio [OR] = 0.90; 95% Bayesian credible interval [CrI] = 0.85, 0.96) compared with the referent group. Also, census tracts with high Black isolation (quartile 4: OR = 2.93; 95% CrI = 1.44, 6.02) and higher population density (OR = 1.24; 95% CrI = 1.13, 1.37) were more likely to belong in the high polysubstance use profiles.

Conclusions. Findings support the need for targeted resources tailored for particular neighborhoods. (Am J Public Health. 2026;116(8):1210–1219. https://doi.org/10.2105/AJPH.2026.308446)

US drug overdose deaths more than doubled from 2015 to 2023, increasing from 52 404 to 105 007.¹ The estimated cost for the opioid epidemic is $1.5 trillion annually² from expenses related to emergency medical services, hospitalizations, loss of productivity, and criminal justice expenditures.³

The overdose crisis has evolved over multiple waves. The introduction of illicitly manufactured fentanyl to the black-market drug supply in 2013 caused an exponential increase in fatal overdoses. Nationally, fentanyl overdose deaths rose from 3105 in 2013 to 73 654 in 2022, representing a 23-fold increase.^4,5 Fatal overdoses involving intentional and unintentional combined use of opioids (primarily fentanyl) with stimulants (primarily cocaine, methamphetamine, and MDMA) constitutes the fourth wave of the crisis.⁶ Overdose deaths involving both fentanyl and stimulants increased more than 50-fold since 2010, from 0.6% in 2010 to 32.3% in 2021.⁷ Other new and emerging adulterants or contaminants such as xylazine, medetomidine, and nitazenes contribute to increased morbidity and mortality for people who use drugs.^6,8,9 Adulterants are pharmacologically active or inactive ingredients that are often added to drug supplies to increase the bulk and help save on cost while producing synergistic or improved drug effects, reduce the amount of drug needed to acquire the desired effect, or aid in drug absorption.^9,10 Contaminants are unintentionally added to the drug supply because of inadequate storage practices and low-quality preparation processes that lead to polydrug product unknowingly.⁸ While the landscape of adulterants and contaminants in drug manufacturing is evolving faster than public health surveillance can keep up with, intentional polysubstance use by the consumer further contributes to the complex multidimensional aspects of the fatal overdose crisis.^11–13 Individuals may mix substances for the synergistic and preferred effects that are created when substances are combined over the effects created by a single substance.

In addition to individual preferences and market forces, several studies have found that overdose risk is fundamentally linked to neighborhood physical, social, and geographic characteristics^14–16; however, spatial patterns of polysubstance use‒related overdose typologies are understudied. Previous research has shown significant neighborhood- and population-level differences in fatal overdose risk determinants. A recent study found significant associations between location of fentanyl-involved fatal overdoses and neighborhood resource deprivation.¹⁷ Fentanyl-involved overdoses were also more likely to occur among men and Black and Hispanic/Latino subgroups as the overdose crisis has disproportionately affected Black racial subgroups in recent years.¹⁸ Ecologically, structural determinants associated with increased overdose risk includes low neighborhood annual income and educational attainment,¹⁹ residential instability,²⁰ and income inequality.¹⁵ However, it is unclear whether these factors are similarly associated with polysubstance use‒involved overdose, especially within the context of heightened adulterated and contaminated substances. Most studies do not consider the spatial distribution of latent subpatterns of use or how patterns are affected by structural determinants.²¹ Understanding the spatial distribution of polysubstance use‒related overdose will help optimize resource allocation and public health intervention planning.

In this exploratory study, we present a spatial analysis of polysubstance use overdose typologies and spatial relationships with socio-structural determinants of these heterogenous patterns among fatal overdoses. We hypothesized that polysubstance use‒involved fatal overdoses would follow identifiable geographic patterns and use profiles based on previous research of spatial patterns of fentanyl-involved fatal overdoses.

METHODS

Cook County Medical Examiner data are publicly available and updated daily and include full toxicology reports and Global Positioning System (GPS) coordinates for where an overdose occurred. Cook County, Illinois, is the second most populous county in the United States and includes Chicago and surrounding suburbs. We included records from January 1, 2017, to December 30, 2022. We used Python to conduct a comprehensive keyword search of the “primary cause of death” field, which includes toxicology reports, to categorize substances involved in overdose. ²² The search covered common substances, metabolites, and misspellings, including fentanyl or fentanyl analogs and metabolites (e.g., fentanyl, carfentanil, 4-ANPP, U-47700),^23,24 prescription opioids, heroin, alcohol, cocaine, benzodiazepines, methamphetamine/amphetamine, xylazine, and other drugs (e.g., ketamine, hallucinogens, inhalants or solvents, and other club drugs; n = 9846). Unflagged terms were manually reviewed to expand the keyword database. Overdose records were then assigned binary codes indicating the presence of that substance (e.g., 0/1 for heroin). The location of fatal overdose was used as a proxy measure for where overdose decedents used substances. Demographic information included age, sex, and 2 race and ethnicity categories with a separate designation for Latino. We recoded race/ethnicity into one category with non-Latino White, non-Latino Black, Latino, and other (combining Asian, American Indian, “other,” and “unknown”).

Reporting is consistent with the Strengthening the Reporting of Observational Studies in Epidemiology guidelines (Appendix A, available as a supplement to the online version of this article at https://ajph.org).

Measures

We pooled census tract‒level drug mortality rate per 1000 population over the periods between 2017 and 2022 to reduce annual fluctuations in small census tracts. The substance use indicators included drug mortality rates per 1000 census tract population across 9 drug types (alcohol, fentanyl, prescription opioids, heroin, cocaine, benzodiazepines, xylazine, methamphetamine/amphetamine, and other drugs).

We obtained median household income in $10 000s from 5-year American Community Survey (ACS) estimates. We calculated population density by dividing the total census tract population by tract area in square miles then multiplying by 1000. Black Isolation Index captures the extent to which Black people living in a census tract are exposed to only each other.²⁵ The Black Isolation Index is calculated using the formula

$$\left\{\left(a/A\right)\left(a/\left(a+b\right)\right)\right\}$$ (1)

where a is the Black population in a census tract, b is the non-Black population in the census tract, and A is the total Black population in Cook County. We calculated neighborhood deprivation with the formula

$$\left\{\left(\left(c/10+d/10\right)-\left(a/10+b/10\right)\right)/4\right\}$$ (2)

using census block tract‒level items from ACS: a is adults aged 25 years or older with a college degree, b is owner-occupied housing, c is households with incomes below the federal poverty threshold, and d is female-headed households with children (percentages are entered as whole numbers, not decimals; range = −5 is very low/little deprivation; +5 is very severe deprivation.²⁶ See Appendix B, available as a supplement to the online version of this article at https://ajph.org, for more information on formulas and data sources.

Data Analysis

We mapped GPS coordinates of overdose locations in R version 4.2.1 (R Foundation for Statistical Computing, Vienna, Austria) and aggregated them to the census tract (n = 1332). The drug mortality rates were square root transformed to address skewness and manage outliers. We conducted latent profile analysis (LPA) to identify salient subgroups of census tract‒level profiles based on overdose fatality rates (see LPA model in Appendix C, available as a supplement to the online version of this article at https://ajph.org). The estimated density function for LPA is presented in equation 3, where x_i are the 9 overdose fatality variables described for census tract i, λ_k are the mixture weights for each variable in class k, ${\theta }_k$ represents the class-specific parameters (e.g., means and covariances), and $f\left(x_i|{\theta }_k\right)$ is the class-specific density function and is typically assumed to follow a multivariate normal distribution²⁷:

$$f\left(x_i|{\theta }_k\right)={\sum }_{k=1}^k{\text{λ}}_{k}f\left(x_i|{\theta }_k\right)$$ (3)

We used multiple fit statistics to guide optimal model selection along with interpretability, including the log-likelihood, Akaike information criterion (AIC), consistent AIC, Bayesian information criterion (BIC), sample adjusted BIC, entropy, and bootstrapped likelihood ratio test (BLRT). The BIC is preferred for model selection because it balances fit and simplicity. The BLRT is generally used to determine the optimal number of profiles because it compares the fit of a model with k classes to a model with k‒1 classes with a significant P value indicating that the model with k classes is a better fit; however, BLRT is sensitive to large sample sizes and small improvements can yield significant results and lead to overfitting. We heavily considered BIC and sample size‒adjusted BIC for model selection, reviewed entropy to ensure clear separation and stability in profile assignments, and class size to prevent overfitting with less than 5% sample size. To verify the optimal number of profiles, we plotted the BIC values to visually inspect the elbow where the BIC plateaued (Appendix D, available as a supplement to the online version of this article at https://ajph.org).

Most studies only use 1 methodology to identify subgroups (e.g., LPA only); however, given the absence of a criterion standard for statistically validating data clustering findings,²⁸ the applicability and stability of subgroup findings are often unclear. To verify the consistency of the classification of the census-tract level LPA and robustness of results, we conducted k-means clustering using Euclidean distance (see Appendix E and Appendix F, available as supplement to the online version of this article at https://ajph.org, for visualizing the principal components by k-means clustering and the Sankey diagram comparing the LPA classes with the k-means clusters). We used the elbow plot method to identify the optimal number of profiles to discover the point in the graph where the within-cluster sum of squares starts to level off linearly.²⁹ We used multinomial logistic regression to determine which socio-structural determinants predicted different census tract memberships in the polysubstance use profiles. We conducted formal analysis by using R, with the standard .05 as the significance level.

After fitting the multinomial regression model under the ordinary least squares framework, we assessed residual spatial autocorrelation. Global Moran’s I on polysubstance use latent clusters showed moderate spatial clustering (Moran’s I = 0.18; P < .001), indicating that census tracts with similar latent profile cluster memberships are clustered together.^30,31 Although after adjusting for the neighborhood-level structural determinants we saw a substantial reduction in spatial dependence on the residuals, denoting the spatial patterns were partially explained by our covariates, a small but statistically significant degree of spatial autocorrelation remained (Moran’s I = 0.06; P < .001). Therefore, the initial OLS model was deemed unsuitable as a standalone model because it did not adequately account for spatial dependence, and the assumption of residual independence was not sufficiently met.³² Appendix I (available as a supplement to the online version of this article at https://ajph.org) shows the local spatial clustering of polysubstance use profiles and the residual obtained from the local indicator of spatial association statistics. To address this, we conducted a spatial multinomial logistic model using the integrated nested Laplace approximation approach with a Besag‒York‒ Mollié 2 (BYM2) spatial prior.³³ Results are presented as odds ratio (OR) with 95% Bayesian credible intervals (CrIs).^31,34

RESULTS

Table 1 describes sociodemographic characteristics and substances used at time of death. A plurality of decedents was non-Latino Black (n = 4846; 49.2%), followed by non-Latino White (n = 3525; 35.8%). The decedents were predominantly male (n = 7568; 76.9%); the mean age was 48 years (SD = 20), with 25.8% falling within the range of 45 to 54 years of age. Approximately 24.4% (n = 2367) of decedents had 1 drug on their toxicology report, while 38.6% (n = 3748) had 2 drugs, and 26.2% (n = 2538) had 3 drugs. More than 3 drugs were less common, with 9.0% (n = 876) of decedent records reporting 4 drugs, and 1.6% (n = 156) and 0.2% (n = 17) had 5 and 6 drugs on their toxicology report, respectively. This distribution highlights the ubiquity of polysubstance use, with the majority of decedents having multiple substances at the time of death. Fentanyl and its analogs were the most commonly identified substance (n = 7411; 75.3%), followed by cocaine (n = 4297; 43.6%) and heroin (n = 3838; 39.0%). Xylazine (n = 368; 3.7%) was the least commonly observed.

TABLE 1—

Individual Sociodemographic and Substance Use Characteristics: Cook County, IL, 2017–2022

Characteristics	No. (%) or Mean ±SD
Total	9846
Race/ethnicity
Non-Latino White	3525 (35.8)
Non-Latino Black	4846 (49.2)
Latino	1373 (13.9)
Other	102 (1.0)
Sex
Male	7568 (76.9)
Female	2278 (23.1)
Age, y	48 ±20.0
< 18	26 (0.3)
18–24	498 (5.1)
25–34	1702 (17.3)
35–44	1996 (20.3)
45–54	2542 (25.8)
55–64	2440 (24.8)
≥ 65	635 (6.4)
Drug count present on toxicology report
1	2367 (24.4)
2	3748 (38.6)
3	2538 (26.2)
4	876 (9.0)
5	156 (1.6)
6	17 (0.2)
Substance use indicators
Prescription opioids	1145 (13.7)
Heroin	3838 (39.0)
Fentanyl and analogs	7411 (75.3)
Alcohol	2513 (25.5)
Cocaine	4297 (43.6)
Methamphetamine and amphetamine	509 (5.2)
Xylazine	368 (3.7)
Benzodiazepines	1441 (14.6)
Other drugs	385 (3.9)

Note. The counts and percentages represent the fatal overdoses that tested positive for each substance on the toxicology report. Age does not add to 100% because of 7 decedents missing age information. Substance use indicators are not mutually exclusive and so percentages do not add up to 100 for those categories.

Census Tract‒Level Cluster Solution

Figure 1 presents the results of an LPA examining mean fatal overdose rates per 1000 population at the census tract level (see Appendix H, available as a supplement to the online version of this article at https://ajph.org, for overdose incidence across Cook County census tracts). The low polysubstance use profile (in purple) exhibits consistently low use rates across all substances, with fentanyl peaking at a mean of approximately 0.8 per 1000 population, while the other substances including cocaine, heroin, and prescription opioids remaining below 0.5 per 1000 population. The moderate polysubstance use profile (in green) shows moderate levels of polysubstance use across all substances, with highest mean rates observed for fentanyl and cocaine, at 1.3 and 1.2 per 1000 population respectively. The high polysubstance use profile (in yellow) shows the highest mean rates of polysubstance use, with peaks in fentanyl, cocaine, and heroin, indicating areas with the highest risk of multiple substances at the time of overdose. See Appendix H for spatial distribution of socio-structural determinant factors.

FIGURE 1—

Three Latent Profile Clusters of Drug Rates Across Cook County, IL, Census Tracts, 2017‒2022, by (a) Geospatial Distribution and (b) Rates per 1000 Population

Structural Determinant Correlates of Profiles

In the unadjusted analysis (Table 2), the census tracts in the moderate polysubstance use profile were associated with neighborhood deprivation (below-average deprivation: OR = 1.55 [95% confidence interval [CI] = 1.05, 2.29] and most deprivation: OR = 2.77 [95% CI = 1.95, 3.92]), increasing Black isolation (quartile 2: OR = 1.49 [95% CI = 1.04, 2.16]; quartile 3: OR = 2.11 [95% CI=1.46, 3.05]; quartile 4: OR = 2.48 [95% CI = 1.71, 3.60]) compared with the low polysubstance use profile as the referent group. However, for every $10 000 increase in median household income, the odds of a census tract belonging to the moderate polysubstance use profile decreased. Similarly, the odds of belonging to the high polysubstance use profile were significantly increased for census tracts with the most deprivation (OR = 4.28; 95% CI = 2.87, 6.36) and higher Black isolation (quartile 3: OR = 2.42 [95% CI = 1.57, 3.73]; quartile 4: OR = 3.68 [95% CI = 2.41, 5.61]).

TABLE 2—

Associations Between Census Tract Social Contextual Factors and Latent Polysubstance Use Profiles: Cook County, IL, 2017–2022

	Class 2: Moderate Polysubstance Use	Class 3: High Polysubstance Use
OLS model 1 (unadjusted), a OR (95% CI)
Neighborhood deprivation index
Least deprivation (Ref)	1	1
Below-average deprivation	1.55 (1.05, 2.29)	0.83 (0.48, 1.43)
Above-average deprivation	1.44 (0.99, 2.11)	1.33 (0.83, 2.12)
Most deprivation	2.77 (1.95, 3.92)	4.28 (2.87, 6.36)
Black isolation index
Quartile 1 (high integration; Ref)	1	1
Quartile 2	1.49 (1.04, 2.16)	1.15 (0.72, 1.83)
Quartile 3	2.11 (1.46, 3.05)	2.42 (1.57, 3.73)
Quartile 4 (high isolation)	2.48 (1.71, 3.60)	3.68 (2.41, 5.61)
Median household income (in $10 000s)	0.89 (0.86, 0.92)	0.85 (0.81, 0.88)
Population density (in 1 000s)	1.06 (0.99, 1.14)	1.06 (0.97, 1.15)
OLS model 2 (adjusted), OR (95% CI)
Neighborhood deprivation index
Least deprivation (Ref)	1	1
Below-average deprivation	0.85 (0.53, 1.34)	0.46 (0.25, 0.85)
Above-average deprivation	0.58 (0.35, 0.98)	0.55 (0.29, 1.04)
Most deprivation	0.87 (0.47, 1.62)	1.45 (0.70, 3.01)
Black isolation index
Quartile 1 (high integration; Ref)	1	1
Quartile 2	1.34 (0.92, 1.95)	1.06 (0.65, 1.71)
Quartile 3	1.64 (1.11, 2.43)	1.81 (1.13, 2.91)
Quartile 4 (high isolation)	1.54 (0.97, 2.44)	1.87 (1.10, 3.18)
Median household income (in $10 000s)	0.88 (0.83, 0.94)	0.90 (0.84, 0.97)
Population density (in 1 000s)	1.15 (1.06, 1.25)	1.23 (1.13, 1.35)
BYM2 model 3 (adjusted), OR (95% CrI)
Neighborhood deprivation index
Least deprivation (Ref)	1	1
Below-average deprivation	1.06 (0.67, 1.69)	0.67 (0.34, 1.33)
Above-average deprivation	0.73 (0.43, 1.24)	0.84 (0.41, 1.73)
Most deprivation	0.80 (0.43, 1.52)	1.56 (0.68, 3.60)
Black isolation index
Quartile 1 (high integration; Ref)	1	1
Quartile 2	1.31 (0.89, 1.93)	0.83 (0.48, 1.42)
Quartile 3	1.46 (0.96, 2.22)	1.31 (0.75, 2.28)
Quartile 4 (high isolation)	1.30 (0.77, 2.19)	2.93 (1.44, 6.02)
Median household income (in $10 000s)	0.90 (0.85, 0.96)	0.94 (0.88, 1.02)
Population density (in 1 000s)	1.09 (1.00, 1.17)	1.24 (1.13, 1.37)

Note. BYM2 = Besag‒York‒Mollié 2; CI = confidence interval; CrI = Bayesian credible interval; OR = odds ratio. The results are from ordinary least squares (OLS) and Bayesian spatial (BYM2) multinomial models.

^a Model 1 (unadjusted) shows the results from 4 separate models, and model 2 (adjusted) shows the results from 1 model. Model 3 (adjusted) shows results from 1 Bayesian multinomial logistic regression model with spatial random effects, which was specified using the BYM2 model. For all 3 models, cluster 1 served as the referent group.

In the adjusted analysis, the unadjusted relationships were attenuated but still significant. Census tracts with higher Black isolation were more likely to be members of moderate polysubstance use (quartile 3: OR = 1.64; 95% CI=1.11, 2.43) and high polysubstance use profiles (quartile 3: OR = 1.81 [95% CI = 1.13, 2.91]; quartile 4: OR = 1.87 [95% CI = 1.10, 3.18]). Census tracts with higher median household incomes were less likely to be in the moderate polysubstance use (OR = 0.88; 95% CI = 0.83, 0.94) and high polysubstance use profiles (OR = 0.90; 95% CI = 0.84, 0.97). However, census tracts with higher population density were more likely to be members of the moderate polysubstance use (OR = 1.15; 95% CI = 1.06, 1.25) and high polysubstance use profiles (OR = 1.23; 95% CI = 1.13, 1.35).

Bayesian Spatial Multinomial Logistic Model

Bayesian multinomial spatial results showed census tracts with lower median household income was associated with increased odds of belonging to the moderate polysubstance use profile (OR = 0.90; 95% CrI = 0.85, 0.96) compared with the referent group. Also, census tracts with high Black isolation (quartile 4: OR = 2.93; 95% CrI = 1.44, 6.02) and higher population density (OR = 1.24; 95% CrI = 1.13, 1.37) were more likely to belong in the high polysubstance use profiles.

DISCUSSION

This study describes the spatial distribution of polysubstance use‒involved fatal overdoses and examined neighborhood-level sociodemographic correlates of overdose fatality. We found 3 profiles of polysubstance use—low, moderate, and high polysubstance use—across census tracts in Cook County, with fentanyl, cocaine, and heroin playing dominant roles in polysubstance use across all profiles. We also found significant associations between neighborhood sociodemographics and rate of polysubstance use‒involved fatal overdose. Neighborhoods with higher neighborhood deprivation, racial segregation, and population density and lower median household income were more likely to be in the high polysubstance use profile. Although the 3 emergent latent profiles share some similarities across the census tract‒level mean substances use overdose rate, they represent meaningfully different polysubstance overdose typologies. The profiles differ in the overall intensity of substance involvement but also in the specific combinations and patterns of substances contributing to fatal overdoses rate. Our findings also observed lower methamphetamine involvement in the high polysubstance use profile compared with the moderate polysubstance use cluster, and benzodiazepine and xylazine rates were similar across moderate and high profiles as well. Therefore, the 3 latent profiles can be broadly conceptualized as representing low, moderate, and high levels of polysubstance use overdose in this research study; however, the specific shapes of these profiles were not determined to be linear a priori. Rather, the clusters saliently show the specific ways in which substance combinations co-occur across neighborhoods.

The high polysubstance use profile demonstrated the highest drug involvement patterns, identifying neighborhoods at highest risk for polysubstance use‒involved fatal overdoses. Although the low polysubstance use profile experienced lower fentanyl use rates compared with the high-use profile (0.8 per 1000 population vs 1.8 per 1000), fentanyl remained present in fatal overdoses in low polysubstance use areas. This finding supports the ubiquity of fentanyl in the illicit drug supply and widespread fentanyl co-use with heroin and cocaine, even in areas with lower overall polysubstance use.^17,35 Despite the focus on the rise of fentanyl in the illicit drug supply, it is important to note that heroin was still present in all 3 polysubstance use profiles at relatively high rates. Heroin often serves as a base substance for adulterated mixtures, and some people who use drugs prefer heroin or seek it out specifically.³⁶ Our findings support projections that heroin will not entirely disappear from the illicit drug supply despite the economic incentives from fentanyl’s increased retailer revenue.³⁷

Our findings also support previous national studies on the increased prevalence of cocaine‒fentanyl coinvolvement in fatal overdoses.⁷ Previous studies of people who use drugs found that cocaine‒opioid co-use (e.g., “speedball”) may be desirable as the stimulated sensation complements the opioid’s sedating effect.³⁸ At the same time, fentanyl adulteration could lead to an opioid overdose among people who use cocaine, especially if the presence of fentanyl is unknown.³⁹ Targeted messaging and harm-reduction outreach promoting drug self-testing to people who use cocaine, whether intentionally or unintentionally combined with other substances, may help reduce overdose fatality in high and moderate polysubstance use neighborhoods.

In contrast to national studies, we did not find high rates of methamphetamine‒ fentanyl coinvolvement as methamphetamine was comparatively rare in the high and moderate polysubstance use profiles (0.3 per 1000 population and 0.4 per 1000, respectively) and practically absent in the low polysubstance use profile (0.0 per 1000).^7,35,40 It is possible that the relatively low presence of methamphetamine co-use was related to the urban study setting as previous national research found that rurality was a significant positive predictor of methamphetamine co-use.⁴¹ Although use has increased in Cook County within recent years, methamphetamine has historically been less prevalent in Chicago than other drugs such as cocaine and opioids.⁴² Methamphetamine use often spreads through specific social networks (e.g., particular nightlife scenes, housing clusters, or peer groups); neighborhoods with strong ties to those networks may see earlier and steeper increases in use, which may explain the slightly increased methamphetamine use rate in moderate profile neighborhoods compared with high polysubstance use profile neighborhoods.^43–45 These discrepancies highlight the importance of localized inquiry into overdose risk factors to best tailor community overdose prevention interventions.

We found a nonlinear pattern between neighborhood segregation and polysubstance use profiles. Census tracts in the highest Black isolation quartile were 3 times more likely to be in the high polysubstance use profile compared with the least segregated tracts, after we adjusted for other neighborhood sociodemographics. We also found that median household income was protective in the spatial model and inversely associated with moderate polysubstance use profile. For each $10 000 increase in median household income, the odds of being in the moderate polysubstance use profile decreased 10%. Spatio-structural inequalities, including social, economic, and infrastructure disparities, have been increasingly theorized as fundamental drivers of the overdose crisis.^46,47 Previous research has shown that fatal opioid and polysubstance overdoses are often geographically concentrated and tend to cluster in predominantly Black neighborhoods.^48,49 Also, we found census tracts with increasing population density were associated with membership in the high polysubstance‒use clusters in our spatial model, which aligns with the extant literature noting that population density within the context of urbanicity influences substance use and related overdose risk.⁵⁰ More research is needed into the mechanisms by which neighborhood spatiotemporal dynamics affect polysubstance use and associated overdose risk.²¹

Strengths and Limitations

This cross-sectional study was limited to fatal overdose locations in 1 major US city. Neighborhood risk factors for polysubstance use‒involved fatal overdose may differ across cities and by urbanicity. We did not have access to nonfatal overdose data and cannot account for their geographic variability. There is potential for misclassification bias as certain substances may be missing from toxicology reports depending on the stringency of the medical examiner’s investigation and when routine testing for a specific substance began.¹⁷ For example, during the height of the severe acute respiratory syndrome coronavirus 2 pandemic in 2020, there was an increase in hospital-certified cases (the medical examiner had jurisdiction but did not bring the body to their facilities for autopsy to preserve morgue space and, instead, had medical examiner staff visit the hospital to sign the death certificate) and cases where the deceased were not found until decomposition started as social distancing caused drug use in isolation.^51,52 In this analysis, we assumed that the absence of a substance on a decedent’s toxicology report represents a value of zero. Future studies should incorporate individual-level data using multilevel latent profile analysis to gain more insights into these nuanced relationships.

Public Health Implications

Our study identified 3 typologies of polysubstance use profiles localized to specific neighborhoods in 1 major US city. Fentanyl, cocaine, and heroin emerged as key substances driving the higher-risk neighborhood profiles. Our study supports findings from previous national studies on stimulant and opioid co-use^6,7,35,40 but also identified important localized differences in prevalence of substances present at the time of fatal overdose. Findings support the need for localized harm reduction and overdose prevention interventions tailored to the unique neighborhood polysubstance typologies. Possible targeted interventions include harm reduction vending machines offering naloxone and drug self-testing kits, introducing overdose prevention centers in high polysubstance‒use neighborhoods, and expanding mobile overdose response programs that are able to dynamically respond to the needs of people who use drugs in a specific neighborhood.^53–55 Better understanding of how and why neighborhood sociodynamics affect fatal polysubstance overdoses may identify novel strategies for overdose prevention.