Identifying demographic predictors of increased non-fatal opioid overdose risk among New York State Medicaid enrollees after the COVID-19 Pandemic: An analysis of heterogeneous treatment effects

Abstract

Background:

Overdose rates in the U.S. rose dramatically during the COVID-19 pandemic. Well-documented racial and sociodemographic inequities in the impact of the pandemic suggest the potential for similar inequities for overdose. Our objective was to identify subgroups of New York State Medicaid enrollees who experienced the greatest increases in non-fatal opioid overdose risk following onset of the COVID-19 pandemic.

Methods:

Data are from a retrospective cohort of 1,021,889 people enrolled in New York State Medicaid from 2019-2020. To identify subgroups with the greatest increased risk of non-fatal overdose following onset of the COVID-19 pandemic, we used Heterogeneous Treatment Effect (HTE)-Scan, a novel machine learning method developed for accurate and computationally efficient discovery of heterogeneous treatment effects in complex data.

Results:

In the total sample, risk of non-fatal opioid overdose increased 22% after onset of the pandemic. We also identified two subgroups with elevated risk relative to the total sample: subgroup 1 (Black and Hispanic males aged 45-64 years old with no baseline documentation of opioid use disorder (OUD); N = 53,065) and subgroup 2 (people aged 45-64 years old with documented aged/blind/disabled status and no baseline documentation of OUD; N = 73,694). These subgroups experienced a 54% and 57% increase in non-fatal overdose risk, respectively.

Conclusions:

We estimated heterogeneous effects of onset of the COVID-19 pandemic on nonfatal overdose, with elevated risks estimated for older working-aged, structurally disadvantaged adults without documented OUD. These findings illustrate the importance of structural factors in driving heterogeneous risk of overdose following complex social events.

INTRODUCTION

Despite broad efforts to address opioid misuse and related harms,¹ drug overdose continues to be a leading cause of mortality in the United States. In 2022, there were more than 105,000 drug overdose deaths in the U.S.² Overdose deaths are largely driven by opioid overdoses, which in recent years have been fueled by a combination of synthetic opioid (e.g., fentanyl) use, contamination of the drug supply, and broader polysubstance use.^3,4 In addition to overall increases in overdose, alarming inequities have emerged across important individual-level sociodemographic axes, such as racial group membership (i.e., disproportionate increases among Black and Indigenous American populations),^5–8 as well as across community-level factors, such as county-level income inequality.⁹ In order to effectively intervene on the overdose crisis, it is vital that we understand the factors that influence changing patterns of risk so that we can appropriately target prevention efforts and resources.

A primary factor that has shaped overdose risk since 2020 has been the emergence of the Coronavirus disease 2019 (COVID-19) pandemic.^10–12 According to a report from the American Medical Association, as early as October 2020, 40 states had reported increases in overdose deaths following the onset of the pandemic; with 2021 seeing even higher overdose rates.^13,14 Non-fatal overdoses also increased during this period, as indicated by increases in non-fatal overdose emergency department visits.^15,16 Multiple mechanisms may have led to increased risk of non-fatal overdose, especially during the early stages of the pandemic. Some individuals may have been more likely to initiate drug use, or may have increased frequency or amount of drug use as a means of coping with high levels of psychological distress resulting from deaths of loved ones, social isolation, or increased economic hardship resulting from the pandemic and/or pandemic mitigation measures.^17,18 Additionally, disruptions to drug supply chains and trafficking during the pandemic may have led to volatility in the illicit drug market impacting the potency of available illicit opioids.^19–21 Lastly, disruptions to medical care could have increased risk of opioid overdose either through interruptions to treatment for opioid use disorder (OUD),^22–24 which could lead to increased relapse, or through interruptions to and/or difficulty initiating medical care for people with chronic pain,²⁵ which could result in increased initiation of illicit opioid use.

As has been the case for many consequences of the COVID-19 pandemic,^26,27 there are reasons to suspect that the impacts on overdose risk are not distributed equally. In particular, members of structurally marginalized groups such as Black and Hispanic individuals and others with low socioeconomic resources experience greater burdens during major societal crises due to their lack of material resources and social capital.^28,29 This manifests in fewer resources to prevent exposure (e.g., ability to work from home; online shopping), greater economic vulnerability (e.g., employment in service industry jobs that were more impacted by shelter-in-place orders), and fewer resources to mitigate disruptions to medical care (e.g., transportation services, geographically accessible healthcare settings). Other individual-level factors, such as age, disability status, and chronic pain may also be especially relevant, as each is associated with increased risk of COVID-19-related complications^30,31 and may also increase overdose risk. For example, individuals with chronic pain may be more severely impacted by pandemic-based interruptions to medical care than the general public due to physical disability, comorbidities, and need for ongoing pain management. Furthermore, community-level factors such as urbanicity, population density, and neighborhood markers of socioeconomic status (e.g., proportion unemployed, proportion living in poverty) may further influence how COVID mitigation measures and overdose prevention services are experienced in ways that can impact overdose risk.

Prior research has shown mixed evidence of heterogeneity in overdose risk following the onset of the COVID-19 pandemic, though methods and samples vary greatly across studies.^32–35 A study of drug overdose in Rhode Island found that there was an overall increased rate of overdose mortality from 2019 to 2020, with the greatest increases observed among men, non-Hispanic white people, single people, deaths involving synthetic opioids, and deaths occurring in personal residences.³⁴ Importantly, 58% of all overdose deaths during the observation period were among Medicaid beneficiaries, highlighting particularly pressing concerns regarding overdose within this population.³⁴ Conversely, a study in Philadelphia found that increases in fatal and non-fatal overdose rates following onset of the pandemic increased most for non-Hispanic Black people.³⁵

An important limitation of the aforementioned studies and others that have assessed differential risk in the effects of the pandemic on overdose is that they estimated heterogeneity across a single sociodemographic characteristic at a time, without accounting for the potentially complex combination of factors that may increase risk. This limitation of traditional epidemiologic methods may explain the lack of consistency in findings across those studies. However, a novel method that facilitates assessing heterogeneity across multiple risk factors simultaneously can account for the complex interplay of multiple individual- and community-level factors in producing heterogeneous effects. Applying this method may yield a more nuanced understanding of the mechanisms driving increased overdose risk and may suggest novel mechanisms for further exploration and opportunities for intervention.

As the impact of COVID-19 continues to evolve, especially in the context of the ongoing overdose crisis, it is vital that we identify those groups for whom overdose risks may be especially elevated. Doing so can provide better understanding of the underlying mechanisms and inform future prevention and intervention efforts in the event of future pandemics and other “Big Events” (major natural or unnatural environmental, economic, or political disruptions that lead to large scale social instability) that may interrupt access to care.^36,37 Additionally, focusing analysis on a high-risk population, such as those enrolled in Medicaid, has the potential to have a large public health impact due to the considerable burdens of both overdose and COVID-19 (and related mitigation factors) among this population.^38,39 The aim of this study was to identify subpopulations of the New York State Medicaid population who experienced the most significant increases in risk of a non-fatal opioid overdose following the onset of the COVID-19 pandemic, using a novel machine learning method designed to identify heterogeneous effects across combinations of relevant covariates.

METHODS

Data for the present study come from a retrospective cohort of New York State Medicaid beneficiaries derived from a parent study focused on chronic pain. In the parent study, people were eligible for inclusion if they were enrolled in New York State Medicaid as of February 2020, enrolled for ≥10 months between March 2019 and February 2020, and aged 18-64 years old as of February 29, 2020. Individuals were excluded if they were dual-eligible for Medicaid and Medicare, had missing information for county of residence, or if they had a recent history of cancer, palliative care, or long-term care between March 1, 2019 and February 29, 2020. Among the eligible population, individuals were selected into the parent study cohort on the basis of chronic pain status, as determined via previously published algorithms applied to Medicaid claims from March 1, 2019 through February 29, 2020.^40–45 All individuals with chronic pain, defined as ≥2 claims (between 90 and 366 days apart) in the inpatient or outpatient settings, within the same category of chronic pain based on ICD-10-CM diagnosis codes were selected (eAppendix 1 http://links.lww.com/EDE/C343). People without chronic pain were selected using Mahalanobis distance 5:1 matching criteria on demographic characteristics, including age, race and ethnicity, sex, health service area, and aged/blind/disabled status as of February 2020.⁴⁰ Although the matched study design was motivated by the aims of the parent study, matching on chronic pain status is compatible with the aims of the present study as well due to the role of chronic pain in patterns of overdose risk. The study protocol was approved by the New York University Langone Health Institutional Review Board.

For the present study, we followed the sample population through December 31, 2020, or until death/disenrollment. Our exposure of interest was onset of the COVID-19 pandemic in New York State, which we operationalized as beginning March 1, 2020. Additional exclusion criteria were applied to the parent cohort to create our analytic sample. Individuals not enrolled in Medicaid during either March 2019 (baseline of the unexposed period) or March 2020 (baseline of the exposed period) were excluded, as well as anyone who died in March of 2020 (baseline of the exposed period) to ensure the same denominator for the exposed and unexposed periods.

Outcome measure

The binary outcomes for this analysis were, for each individual, any non-fatal opioid poisoning diagnosis in the pre-COVID and COVID periods, respectively (ICD-10-CM diagnostic codes T40.0x-T40.4x and T40.6x). We excluded adverse effects in therapeutic use, underdosing, sequelae following the initial poisoning event, and concurrent outcomes of death in the same month as an opioid poisoning. We focused exclusively on non-fatal overdoses to ensure comparability between the pre-COVID and COVID periods, as the parent study excluded any participants who died before onset of the pandemic. Observed outcome data were split into two categories for each individual: “unexposed” pre-COVID onset observations which included all overdoses from March 1, 2019 – December 31, 2019 (e.g., prior to onset of the pandemic in New York state), and “exposed” COVID observations which included all overdoses from March 1, 2020 – December 31, 2020 (e.g., during the pandemic).

Covariates

Data on individual-level predictors were compiled from participants’ claims data at baseline of the study (March 2019). Included predictors were categorized age (18-26, 27-35, 36-44, 45-54, 55-64); race/ethnicity as defined by Medicaid (non-Hispanic Black, non-Hispanic white, Hispanic, Asian, other racial/ethnic group, unknown); gender (male, female), aged/blind/disabled status, health service area (Central NY, Finger Lakes, Mid-Hudson, Nassau-Suffolk counties, New York City, Northeastern NY, NY-Penn, Western NY), chronic pain, OUD, opioid treatment program visits, and pharmacy claims for medications for opioid use disorder (MOUD). Chronic pain diagnosis categories included arthritis, back pain, neck pain, unspecified back and/or neck pain, neurologic pain, chronic headache/migraine, or miscellaneous chronic pain. Participants were categorized as having chronic pain if they met criteria for any one or more chronic pain categories. We defined OUD as an ICD-10-CM diagnoses of F11.x and/or claims for any opioid treatment program visit during the baseline month of March 2019. Pharmacy claims for MOUD were operationalized as separate indicator variables for claims for buprenorphine, naltrexone, or vivitrol (National Drug Codes in the Appendix) during the baseline month.

Data on community-level predictors were compiled from the American Community Survey 5-year estimates (2015-2020) and operationalized at the census-tract level. Included variables were population density; proportion of population that is male; age distribution of the population (proportion under 18 years, proportion 18-64 years, proportion 65+); racial/ethnic density (proportion of population that is Black, proportion of population that is non-white); proportion that has never been married; proportion unemployed; median household income; proportion of owner occupied housing units; proportion of housing units occupied by same resident for 5+ years; proportion of adult population at or below 100% of the poverty line; and proportion of housing units with 4+ residents. Each community-level predictor was operationalized into five quintiles.

All covariate values were set statically to their value during the baseline month (March 2019) to ensure the composition of subgroups remained unchanged between the pre-COVID and COVID periods. This maximizes the comparability of the unexposed and exposed groups in the study by ensuring that each subgroup is comprised of the same individuals during the treatment period as it is during the control period.

Analytic Strategy

To identify subpopulations with increased risk of non-fatal opioid overdose following the onset of the COVID-19 pandemic, we used a machine learning algorithm called Heterogeneous Treatment Effect (HTE)-Scan, which is a recent extension of the Treatment Effect Subset Scan.⁴⁶ HTE-Scan was developed to provide accurate and computationally efficient discovery of heterogeneous treatment effects affecting subgroups in complex data. HTE-Scan iteratively searches the data for the subgroup (as defined by a non-empty subset of attribute values for each observed attribute, e.g., Black and Hispanic males aged 45-64) with the most heterogeneity in treatment effect (defined according to the highest value of a penalized log-likelihood ratio score). In doing so, HTE-Scan provides a computationally efficient means of searching across all 1.64 X 10²⁸ possible combinations of covariate values in the data calculated by taking the product of the 2^|v|-1 non-empty subsets of values (v) for each of the 23 total covariates (6 binary; 17 5-level categorical). The log-likelihood ratio score of a subgroup S, F(S) = log(Pr(Data | H₁(S)) / Pr(Data | H₀)), can be interpreted as a measure of the extent to which the data in S deviates from the null hypothesis H₀ or “uniform effect of treatment for all individuals”. Under this null hypothesis of uniformity, all individuals are assumed to have the same additive shift Δ₀ in log-odds of overdose during COVID (as computed by maximum likelihood estimation, Δ₀ = 0.20 for our data). The treatment effect is operationalized as an additive shift in the log-odds of the binary outcome in the treatment condition as compared to the control condition. Thus, under the alternative hypothesis H₁(S), individuals in S are assumed to have an additive shift in log-odds of overdose Δ₀ + Δ_S, where Δ_S ≠ 0.

Penalizing the score proportional to the number of specified attribute values in the subgroup (for this analysis, the penalty is twice the number of identified attribute values) prevents overfitting of the data and increases interpretability of the discovered subgroup. Given that each individual is present in both treatment and control populations, we do not need to account for selection into treatment. The steps of the HTE-Scan algorithm are as follows:

1. Compute the maximum likelihood estimate of Δ₀, the change in the log-odds of overdose under the null hypothesis H₀ of a uniform treatment effect.

2. Learn a logistic regression model Pr(overdose | exposure = 0, covariates) from the individuals’ pre-COVID outcomes.

3. Use the learned model to predict Pr(overdose | exposure = 1, covariates) for the individuals’ during-COVID outcomes, under the null hypothesis H₀: log-odds(overdose | exposure = 1, covariates) = log-odds(overdose | exposure = 0, covariates) + Δ₀.

4. Iteratively scan across sections of the data to identify subgroups (defined by their covariate profile) with the highest penalized log-likelihood ratio score, comparing the observed outcomes (non-fatal overdose) during the treatment period to the expected probability of overdose in that same period under the null hypothesis.

5. Perform permutation testing to assess the significance of discovered subgroups. To do so, we generate 100 simulated datasets where each individual’s treatment and control outcome are randomly shuffled (i.e., swapped with probability 0.5) and re-run the scan to compute the maximum log-likelihood ratio score for each simulated dataset. The threshold score for statistical significance (at α = .05), accounting for the multiple hypothesis testing resulting from the search, is the 95^th percentile of the maximum loglikelihood ratio scores across the simulated datasets. As is the case with all statistical significance testing, being above the threshold does not guarantee that the value was not produced by chance. Rather, it represents a probabilistic statement that, if the null hypothesis H0 were true, we would only see a score this high or higher less than 5% of the time. The choice of 5% (α = .05) is a standard, but arbitrary, choice of significance threshold.

For this analysis, we ran 100 iterations of HTE-Scan across the data, assessing heterogeneity in the effect of onset of the COVID-19 pandemic in NYS on risk of overdose, across subgroups defined by all 1.64 X 10²⁸ possible combinations of individual- and community-level covariates, measured at baseline. Analyses were conducted using Python and STATA 17.

RESULTS

Sample characteristics for the study are described in Table 1. The final analytic sample included 1,021,889 New York State Medicaid enrollees, of whom 29% were non-Hispanic white, 62% were female, and 23% were 45-54 years old. During the pandemic period, 2,277 individuals experienced at least one non-fatal opioid overdose, compared to 1,871 during the pre-pandemic period, illustrating a 22% overall increase in non-fatal overdose risk (95% CI: 1.15, 1.30) following onset of the COVID-19 pandemic.

Table 1.

Frequency of non-fatal overdose by demographic characteristics of New York State Medicaid Cohort before and during the pandemic, 2019-2020 (N = 1,021,889). All covariates are measured at baseline (March 2019).

	Pre-COVID			During COVID			RRa (95% CI)
	Overdose	No Overdose	Total	Overdose	No Overdose	Total
	N (%)b	N (%)b	N	N (%)b	N (%)b	N
Full Sample	1,871 (0.18)	1,020,018 (99.82)	1,021,889	2,277 (0.22)	1,019,612 (99.78)	1,021,889	1.22 (1.14, 1.29)
Patient Characteristics
Chronic Pain
Yes	805 (0.36)	225,408 (99.64)	226,213	903 (0.40)	225,310 (99.60)	226,213	1.12 (1.02, 1.23)
No	1,066 (0.13)	794,610 (99.87)	795,676	1,374 (0.17)	794,302 (99.83)	795,676	1.29 (1.19, 1.40)
Race/Ethnicity
Unknown	209 (0.08)	259,015 (99.92)	259,224	252 (0.10)	258,972 (99.90)	259,224	1.21 (1.00, 1.45)
Non-Hispanic White	909 (0.31)	296,124 (99.69)	297,033	1,062 (0.36)	295,971 (99.64)	297,033	1.17 (1.07, 1.28)
Non-Hispanic Black	313 (0.18)	173,504 (99.82)	173,817	431 (0.25)	173,386 (99.75)	173,817	1.38 (1.19, 1.59)
Asian	17 (0.01)	136,444 (99.99)	136,461	14 (0.01)	136,447 (99.99)	136,461	0.82 (0.41, 1.67)
Other Race	65 (0.13)	51,154 (99.87)	51,219	86 (0.17)	51,133 (99.83)	51,219	1.32 (0.96, 1.83)
Hispanic	358 (0.34)	103,777 (99.66)	104,135	432 (0.41)	103,703 (99.59)	104,135	1.21 (1.05, 1.39)
Sex
Female	743 (0.12)	633,513 (99.88)	634,256	864 (0.14)	633,392 (99.86)	634,256	1.16 (1.06, 1.28)
Male	1128 (0.29)	386,505 (99.71)	387,633	1,413 (0.36)	386,220 (99.64)	387,633	1.25 (1.16, 1.35)
Age
18-26	123 (0.09)	141,837 (99.91)	141,960	156 (0.11)	141,804 (99.89)	141,960	1.27 (1.00, 1.61)
27-35	472 (0.23)	201,324 (99.77)	201,796	512 (0.25)	201,284 (99.75)	201,796	1.08 (0.96, 1.23)
36-44	421 (0.21)	200,186 (99.79)	200,607	469 (0.23)	200,138 (99.77)	200,607	1.11 (0.98, 1.27)
45-54	421 (0.18)	237,880 (99.82)	238,301	592 (0.25)	237,709 (99.75)	238,301	1.41 (1.24, 1.59)
55-64	434 (0.18)	238,791 (99.82)	239,225	548 (0.23)	238,677 (99.77)	239,225	1.26 (1.11, 1.43)
Aged, Blind, or Disabled
Yes	480 (0.39)	122,977 (99.61)	123,457	645 (0.52)	122,812 (99.48)	123,457	1.34 (1.19, 1.51)
No	1391 (0.15)	897,041 (99.85)	898,432	1,632 (0.18)	896,800 (99.82)	898,432	1.17 (1.09, 1.26)
Health Service Area
Central NY	158 (0.28)	56,945 (99.72)	57,103	183 (0.32)	56,920 (99.68)	57, 103	1.16 (0.94, 1.43)
Finger Lakes	215 (0.4)	53,951 (99.60)	54,166	239 (0.44)	53,927 (99.56)	54,166	1.11 (0.92, 1.34)
Mid-Hudson	194 (0.24)	79,430 (99.76)	79,624	248 (0.31)	79,376 (99.69)	79,624	1.28 (1.06, 1.54)
Nassau-Suffolk	204 (0.25)	81,458 (99.75)	81,662	235 (0.29)	81,427 (99.71)	81,662	1.15 (0.96, 1.39)
New York City	754 (0.12)	624,647 (99.88)	625,401	935 (0.15)	624,466 (99.85)	625,401	1.24 (1.13, 1.36)
Northeastern NY	124 (0.26)	46,923 (99.74)	47,047	149 (0.32)	46,898 (99.68)	47,047	1.20 (0.95, 1.52)
NY-Penn	38 (0.41)	9,325 (99.59)	9,363	34 (0.36)	9,329 (99.64)	9,363	0.89 (0.56, 1.42)
Western NY	184 (0.27)	67,339 (99.73)	67,523	254 (0.38)	67,269 (99.62)	67,523	1.38 (1.14, 1.67)
Baseline Documentation of Opioid Use Disorder
Yes	754 (3.1)	23,566 (96.90)	24,320	820 (3.37)	23,500 (96.63)	24,320	1.09 (0.99, 1.20)
No	1,117 (0.11)	996,452 (99.89)	997,569	1,457 (0.15)	996,112 (99.85)	997,569	1.30 (1.21, 1.41)
Baseline Documentation of Opioid Use Treatment
Yes	291 (2.54)	11,188 (97.46)	11,479	336 (2.93)	11,143 (97.07)	11,479	1.15 (0.99, 1.35)
No	1,580 (0.16)	1,008,830 (99.84)	1,010,410	1,941 (0.19)	1,008,469 (99.81)	1,010,410	1.23 (1.15, 1.31)
Community Characteristics c
Population Density
1	277 (0.25)	110,047 (99.75)	110,324	319 (0.29)	110,005 (99.71)	110,324	1.15 (0.98, 1.35)
2	402 (0.32)	124,017 (99.68)	124,419	497 (0.40)	123,922 (99.60)	124,419	1.24 (1.08, 1.41)
3	480 (0.26)	184,168 (99.74)	184,648	597 (0.32)	184,051 (99.68)	184,648	1.24 (1.11, 1.40)
4	227 (0.1)	218,628 (99.90)	218,855	279 (0.13)	218,576 (99.87)	218,855	1.23 (1.03, 1.46)
5	482 (0.13)	382,747 (99.87)	383,229	581 (0.15)	382,648 (99.85)	383,229	1.21 (1.07, 1.36)
Proportion male
1	452 (0.19)	243,400 (99.81)	243,852	532 (0.22)	243,320 (99.78)	243,852	1.18 (1.04, 1.33)
2	365 (0.18)	199,313 (99.82)	199,678	422 (0.21)	199,256 (99.79)	199,678	1.16 (1.01, 1.33)
3	289 (0.15)	191,585 (99.85)	191,874	342 (0.18)	191,532 (99.82)	191,874	1.18 (1.01, 1.38)
4	361 (0.2)	183,242 (99.80)	183,603	444 (0.24)	183,159 (99.76)	183,603	1.23 (1.07, 1.41)
5	401 (0.2)	202,067 (99.80)	202,468	533 (0.26)	201,935 (99.74)	202,468	1.33 (1.17, 1.51)
Proportion <18 yrs
1	360 (0.22)	162,199 (99.78)	162,559	485 (0.30)	162,074 (99.70)	162,559	1.35 (1.18, 1.54)
2	309 (0.18)	170,951 (99.82)	171,260	355 (0.21)	170,905 (99.79)	171,260	1.15 (0.99, 1.34)
3	316 (0.17)	183,622 (99.83)	183,938	387 (0.21)	183,551 (99.79)	183,938	1.22 (1.06, 1.42)
4	358 (0.18)	198,114 (99.82)	198,472	411 (0.21)	198,061 (99.79)	198,472	1.15 (1.00, 1.32)
5	525 (0.17)	304,721 (99.83)	305,246	635 (0.21)	304,611 (99.79)	305,246	1.21 (1.08, 1.36)
Proportion 18-64 yrs
1	275 (0.16)	172,482 (99.84)	172,757	348 (0.20)	172,757 (99.80)	172,757	1.27 (1.08, 1.48)
2	361 (0.21)	175,690 (99.79)	176,051	436 (0.25)	176,051 (99.75)	176,051	1.21 (1.05, 1.39)
3	403 (0.18)	217,658 (99.82)	218,061	477 (0.22)	218,061 (99.78)	218,061	1.18 (1.04, 1.35)
4	397 (0.17)	237,822 (99.83)	238,219	487 (0.20)	238,219 (99.80)	238,219	1.23 (1.07, 1.40)
5	432 (0.2)	215,955 (99.80)	216,387	525 (0.24)	216,387 (99.76)	216,387	1.22 (1.07, 1.38)
Proportion 65+ yrs
1	587 (0.19)	300,839 (99.81)	301,426	720 (0.24)	300,706 (99.76)	301,426	1.23 (1.10, 1.37)
2	427 (0.17)	249,547 (99.83)	249,974	512 (0.20)	294,462 (99.80)	249,974	1.20 (1.05, 1.36)
3	295 (0.16)	187,581 (99.84)	187,876	353 (0.19)	187,523 (99.81)	187,876	1.20 (1.03, 1.40)
4	292 (0.2)	149,034 (99.80)	149,326	351 (0.24)	148,975 (99.76)	149,326	1.20 (1.03, 1.40)
5	267 (0.2)	132,606 (99.80)	132,873	337 (0.25)	132,536 (99.75)	132,873	1.26 (1.08, 1.48)
Proportion Black
1	225 (0.16)	140,014 (99.84)	140,239	263 (0.19)	139,976 (99.81)	140,239	1.17 (0.98, 1.40)
2	232 (0.15)	155,899 (99.85)	156,131	257 (0.16)	155,874 (99.84)	156,131	1.11 (0.93, 1.32)
3	314 (0.18)	177,168 (99.82)	177,482	413 (0.23)	177,069 (99.77)	177,482	1.32 (1.14, 1.52)
4	560 (0.22)	251,456 (99.78)	252,016	609 (0.24)	251,407 (99.76)	252,016	1.09 (0.97, 1.22)
5	537 (0.18)	295,070 (99.82)	295,607	731 (0.25)	294,876 (99.75)	295,607	1.36 (1.22, 1.52)
Proportion Other than Non-Hispanic White
1	246 (0.21)	115,105 (99.79)	115,351	292 (0.25)	115,059 (99.75)	115,351	1.19 (1.00, 1.41)
2	320 (0.26)	121,113 (99.74)	121,433	388 (0.32)	121,045 (99.68)	121,433	1.21 (1.05, 1.41)
3	402 (0.25)	159,316 (99.75)	159,718	465 (0.29)	159,253 (99.71)	159,718	1.16 (1.01, 1.32)
4	417 (0.15)	270,495 (99.85)	270,912	514 (0.19)	270,398 (99.81)	270,912	1.23 (1.08, 1.40)
5	483 (0.14)	353,578 (99.86)	354,061	614 (0.17)	353,447 (99.83)	354,061	1.27 (1.13, 1.43)
Proportion Never Married
1	194 (0.15)	128,508 (99.95)	128,702	205 (0.16)	128,497 (99.84)	128,702	1.06 (0.87, 1.29)
2	235 (0.15)	159,163 (99.95)	159,398	317 (0.20)	159,081 (99.80)	159,398	1.35 (1.14, 1.60)
3	322 (0.16)	198,371 (99.84)	198,693	398 (0.20)	198,295 (99.80)	198,693	1.24 (1.07, 1.43)
4	506 (0.19)	268,057 (99.81)	268,563	571 (0.21)	267,992 (99.79)	268,563	1.13 (1.00, 1.27)
5	611 (0.23)	265,508 (99.77)	266,119	782 (0.29)	265,337 (99.71)	266,119	1.28 (1.15, 1.42)
Proportion Unemployed
1	271 (0.17)	155,349 (99.83)	155,620	350 (0.22)	155,270 (99.78)	155,620	1.29 (1.10, 1.51)
2	278 (0.16)	169,940 (99.84)	170,218	336 (0.20)	169,882 (99.80)	170,218	1.21 (1.03, 1.42)
3	321 (0.17)	190,333 (99.83)	190,654	360 (0.19)	190,294 (99.81)	190,654	1.12 (0.96, 1.30)
4	395 (0.19)	208,766 (99.81)	209,161	469 (0.22)	208,692 (99.78)	209,161	1.19 (1.04, 1.36)
5	603 (0.2)	295,219 (99.80)	295,822	758 (0.26)	295,064 (99.74)	295,822	1.26 (1.13, 1.40)
Median household income
1	754 (0.21)	360,349 (99.79)	361,103	957 (0.27)	360,146 (99.73)	361,103	1.27 (1.15, 1.40)
2	353 (0.15)	234,076 (99.85)	234,429	395 (0.17)	234,034 (99.83)	234,429	1.12 (0.97, 1.29)
3	269 (0.15)	177,459 (99.85)	177,728	304 (0.17)	177,424 (99.83)	177,728	1.13 (0.96, 1.33)
4	264 (0.17)	150,703 (99.83)	150,967	329 (0.22)	150,638 (99.78)	150,967	1.25 (1.06, 1.47)
5	207 (0.22)	93,202 (99.78)	93,409	252 (0.27)	93,157 (99.73)	93,409	1.22 (1.01, 1.46)
Proportion Owner Occupied Housing
1	636 (0.17)	375,113 (99.83)	375,749	805 (0.21)	374,944 (99.79)	375,749	1.27 (1.14, 1.40)
2	425 (0.17)	252,059 (99.83)	252,484	523 (0.21)	251,961 (99.79)	252,484	1.23 (1.08, 1.40)
3	328 (0.19)	168,749 (99.81)	169,077	375 (0.22)	168,702 (99.88)	169,077	1.14 (0.99, 1.33)
4	285 (0.22)	131,260 (99.78)	131,545	317 (0.24)	131,228 (99.76)	131,545	1.11 (0.95, 1.31)
5	194 (0.21)	92,426 (99.79)	92,620	253 (0.27)	92,367 (99.73)	92,620	1.30 (1.08, 1.57)
Proportion 5+ Yr Housing Residents
1	563 (0.27)	208,519 (99.73)	209,082	726 (0.35)	208,356 (99.65)	209,082	1.29 (1.16, 1.44)
2	440 (0.18)	241,457 (99.82)	241,897	497 (0.21)	241,400 (99.79)	241,897	1.13 (0.99, 1.28)
3	327 (0.15)	218,909 (99.95)	219,236	391 (0.18)	218,845 (99.82)	219,236	1.20 (1.03, 1.38)
4	310 (0.16)	196,136 (99.84)	196,446	343 (0.17)	196,103 (99.83)	196,446	1.11 (0.95, 1.29)
5	228 (0.15)	154,586 (99.85)	154,814	316 (0.20)	154,498 (99.80)	154,814	1.30 (1.17, 1.64)
Proportion Adult Pop in Poverty
1	201 (0.22)	92,217 (99.78)	92,418	237 (0.26)	92,181 (99.74)	92,418	1.18 (0.98, 1.42)
2	228 (0.16)	141,038 (99.81)	141,266	287 (0.20)	140,979 (99.80)	141,266	1.26 (1.06, 1.50)
3	321 (0.17)	184,485 (99.83)	184,806	383 (0.21)	184,423 (99.79)	184,806	1.19 (1.03, 1.38)
4	382 (0.15)	251,189 (99.85)	251,571	447 (0.18)	251,124 (99.82)	251,571	1.17 (1.02, 1.34)
5	736 (0.21)	350,678 (99.79)	351,414	919 (0.26)	350,495 (99.74)	351,414	1.25 (1.13, 1.38)
Proportion Housing w 4+ Residents
1	471 (0.29)	163,443 (99.71)	163,914	584 (0.36)	163,330 (99.64)	163,914	1.24 (1.10, 1.40)
2	405 (0.22)	180,389 (99.78)	180,794	460 (0.25)	180,334 (99.75)	180,794	1.14 (0.99, 1.30)
3	341 (0.17)	200,048 (99.83)	200,389	469 (0.23)	199,920 (99.77)	200,389	1.38 (1.20, 1.58)
4	377 (0.17)	222,991 (99.83)	223,368	456 (0.20)	222,912 (99.80)	223,368	1.21 (1.06, 1.39)
5	274 (0.11)	252,736 (99.89)	253,010	304 (0.12)	252,706 (99.88)	253,010	1.11 (0.94, 1.31)

^aRisk ratio of overdose during COVID compared to pre-COVID periods.

^bPercentages are row percents.

^cCommunity characteristics are compiled from American Community Survey 5-year estimates (2015-2020), and operationalized as quintiles at the census tract level.

After permutation testing, HTE-Scan identified two subgroups for whom the estimated effect of the pandemic onset on their overdose risk was higher than expected under the null hypothesis of a uniform effect (Table 2). Subgroup 1 (N = 53,065) is characterized by males who identify as either Black or Hispanic, aged 45-64 years old who had no baseline documentation of OUD. Following onset of the pandemic in New York State, Subgroup 1 experienced an additional 168 overdoses per 100,000 people, or a 54% increased risk of a non-fatal opioid overdose. Subgroup 2 (N = 73,694) is characterized by people aged 45-64 years old with documented aged/blind/disabled status and no baseline documentation of OUD. Subgroup 2 experienced an additional 155 overdoses per 100,000 people, or a 57% increased risk of nonfatal overdose following onset of the pandemic.

Table 2.

Subgroups identified by HTE-Scan with heterogeneous effects of COVID-19 pandemic onset on risk of non-fatal opioid overdose per 100,000 people.

Subgroup	Defining Attributes	N	Penalized log likelihood score	Risk per 100,000 (95% CI)		RRa (95% CI)
				Pre-COVID	During COVID
1	45-64 years old; no baseline documentation of OUD; Black or Hispanic; male	53065	44.15	311 (264, 358)	479 (420, 538)	1.54 (1.27, 1.87)
2	45-64 years old; no baseline documentation of OUD; Aged/Blind/Disabled	73694	41.46	270 (233, 307)	425 (378, 472)	1.57 (1.32, 1.88)
Total Sample		1021889	n/a	183 (175, 191)	223 (214, 232)	1.22 (1.14, 1.29)

^aRisk ratio of overdose during COVID compared to pre-COVID periods.

Figure 1 illustrates the estimated effects of onset of the COVID-19 pandemic on risk of non-fatal opioid overdose for: each of the two subgroups identified above, the union of the subgroups (i.e., the combined grouping of anyone in subgroup 1 or subgroup 2) (N = 109,392; RR: 1.56; 95% CI: 1.35, 1.80), people belonging to neither subgroup (N = 912,497; RR: 1.15; 95% CI: 1.08, 1.23), and the total sample (N = 1,021,889; RR: 1.22; 95% CI: 1.15, 1.30). Whereas subgroups 1 and 2 experienced a 54% and 57% increased risk of non-fatal overdose respectively, the increased risk for people in the sample belonging to neither subgroup was only 15% following onset of the pandemic.

Figure 1.

Risk of non-fatal overdose pre- and post- onset of COVID-19 pandemic in New York State, among Subgroup 1 (n = 53,065), Subgroup 2 (n = 73694), the union of identified subgroups (n = 109,392), individuals in neither subgroup (n = 912,497), and total New York State Medicaid Cohort (n = 1,021,889), 2019-2020.

DISCUSSION

In our study, we identified two subpopulations of New York State Medicaid recipients who experienced heterogeneity in the estimated effects of onset of the COVID-19 pandemic on their risk of experiencing a non-fatal opioid overdose. The overall increases in overdose risk we observed in our sample align with those reported by many other groups following the onset of the pandemic;^10–13 however, we cannot fully rule out secular trends in increased overdose risks. Though we estimated a 22% increased risk of overdose in the total sample, we estimated a 54% increased risk for Subgroup 1 (Black or Hispanic males aged 45-64 years old who had no baseline documentation of OUD) and a 57% increased risk for Subgroup 2 (people aged 45-64 years old with documented aged/blind/disabled status and no baseline documentation of OUD). The pre-pandemic risk of opioid overdose was relatively low for both subgroups (311 and 270 overdoses per 100,000 for Subgroups 1 and 2, respectively). We estimated an increase of 169 and 155 overdoses per 100,000 individuals in Subgroups 1 and 2 respectively following onset of the pandemic. On average, there are almost 8 million people enrolled in New York State Medicaid in a given month, many of whom are Black or Hispanic. Given the large number of people in each subgroup, these increases may translate into a considerable public health burden and a large absolute number of overdoses. Since our observation period only extended through the first 10 months of the pandemic, these results should be interpreted as short-term effects of the pandemic on overdose risk.

Our approach allowed us to identify two large population subgroups that experienced dramatic increases in their risk of overdose during the pandemic from low pre-pandemic risk levels. While our analytic strategy is limited in that it does not necessarily identify the subgroups with the largest excess number of cases, groups with higher baseline overdose risk have been well characterized and are already the subjects of ongoing intervention and prevention strategies. Our approach highlighted groups who, prior to the pandemic, may have been overlooked in the development of prevention and intervention strategies but who nevertheless may benefit substantially from public health mitigation efforts.

A key attribute of both subgroups identified in our analysis was age (45-64 years), highlighting the possible impact of older age on overdose risk after the onset of the pandemic. There are a number of ways that older age could impact overdose risk. Older age was a key risk factor for more severe complications of COVID-19,^30,47,48 which means older adults, including older working-aged adults, may have been more likely to adhere to strict social isolation as a means of reducing COVID-19 risk; the steep age gradient in COVID-19 risk may have also meant greater loss of friends and loved ones for older adults.⁴⁹ Both of these factors could have influenced increased overdose risk, either through increased drug use in isolation or through increased drug use to cope with stressors or grief. Older adults may also be more susceptible to overdose due to the body’s decreased ability to metabolize opioids with age.^50,51 All of these factors may be especially relevant for adults 65 and older, but may also impact older working-aged adults, especially those in their late 50s and early 60s.

Another key attribute of both subgroups we identified is that both were comprised of individuals without documented OUD at baseline, e.g., individuals who might not normally be considered high risk for opioid overdose. One possible explanation for this finding is that these subgroups could include people who were unknowingly exposed to opioids through contamination of non-opioid drugs (e.g., fentanyl-contaminated cocaine). This may be especially relevant for Subgroup 1 due to the racial/ethnic makeup of the group (Black and Hispanic males). In 2021, 88% of cocaine-involved overdoses in New York City involved opioids, and Black people had the highest rate of overdose deaths attributed to cocaine and fentanyl, followed by Hispanic people.⁵²

A recent study of overdose rates among adults 55 and older found that non-Hispanic Black men had the steepest increase in overdose mortality rates from 1999 to 2019 – twice the increase of the next highest group.⁵³ This is just one of many studies over the last few years reporting evidence of disproportionate increases in overdose risk among structurally marginalized racial groups,^{5–8,54–56} including particularly dramatic increases in overdose deaths among older Black adults in New York city⁵² as well as disparities in access to treatment for opioid use disorder.^57,58 Our analysis adds to this body of literature and highlights not only that overdose risk is increasing for older Black people, but that pre-existing trends appear to have been amplified by the pandemic. Our findings thus reinforce the vital need to identify the mechanisms driving increased risk for racially and ethnically marginalized groups in order to develop effective prevention and treatment interventions.

Finally, our results for Subgroup 2 highlight the potential role of disability status in increased overdose risk following the onset of the pandemic. Prior research has highlighted people with disabilities as a population that may experience heightened overdose risk due to their disproportionate risk of opioid dependency as well as specific barriers to accessing medication for opioid use disorder.^59–61 Both of these issues may have been exacerbated by the pandemic as disruptions to care could simultaneously limit medical oversight of opioid use while also reinforcing barriers to treatment access. Our findings are consistent with this body of literature and further suggest the importance of targeted prevention and treatment efforts for people with disabilities.

The relevance of race/ethnicity (Subgroup 1) and disability status (Subgroup 2) to heightened overdose risk following onset of the pandemic suggests that societal-level structural mechanisms may impact population-level overdose risk more fundamentally than markers of individual-level behavior. These societal mechanisms (i.e., structural racism, social correlates of disability, health care access, contamination/toxicity of the illicit drug supply) are not unique to COVID-19 and could manifest as important drivers of overdose risk in future “Big Events,” which may impact both short- and long-term health.^37,62 Therefore, understanding the mechanisms underlying heterogeneity in overdose risk is paramount to public health preparedness.

The results of this study should be considered with key limitations in mind. First, our sample did not contain individuals 65 years or older, so we cannot determine to what extent these older populations experienced increased overdose risk; it is possible that increases in risk following the onset of the pandemic could be even greater for those over 65. Second, because our sample was selected based on survival until onset of the COVID-19 pandemic (any individual who experienced a fatal overdose prior to March 2020 was excluded from the cohort), we restricted our analysis to risk of non-fatal overdose. This further limits the generalizability of our findings as our sample was not fully representative of the New York State Medicaid population. Additionally, increases in non-fatal overdose rates do not necessarily reflect reductions in the case fatality rate, or a combination of reduced case fatality and increased overdose risk. Though a reduced case fatality rate could have contributed to the increased non-fatal overdose risks we observed in our study (e.g., due to increased awareness and distribution of naloxone), fatal overdose rates in New York state increased from 16.6 to 23.3 per 100,000 from 2019 to 2020,⁸ suggesting that the increases in our data do at least partially reflect increased overdose risk. It is also worth noting that many of the characteristics associated with increased overdose risk in our sample were also associated with increased risk of death due to COVID-19 (i.e., male gender, Black or Hispanic racial/ethnic identity).⁶³ To the extent that deaths attributable to COVID-19 may have been a competing risk, our estimates may underestimate the risk of overdose during COVID. Furthermore, overdose was assessed based on Medicaid claims, so we did not have the ability to capture any non-fatal overdoses that did not result in medical treatment (e.g., those reversed by naloxone in the community). Similarly, we could only capture OUD diagnoses documented via Medicaid claims, and therefore cannot rule out misclassification of those individuals without a formal diagnosis of OUD. As we analyzed data designed for a separate parent study, we could not assess OUD diagnoses or opioid treatment program and MOUD access prior to baseline. As all individuals were exposed to the onset of the COVID-19 pandemic, we cannot rule out the possibility that temporal changes in non-COVID related factors may have influenced overdose risk. Finally, though HTE-Scan controls for observed covariates in the discovery of subgroups, it is still prone to unmeasured confounding, especially by factors not included due to lack of predefined subgroups. Therefore, these results should be viewed as exploratory in nature and should be further explored through robust causal methods. Despite these limitations, a major strength of this study was use of a novel machine learning approach that allowed us to maximally leverage data from a large cohort of New York State Medicaid enrollees, and to efficiently explore complex patterns of heterogeneity across a large number of subgroups while correcting for multiple comparisons. Both of these objectives would have been impractical using standard regression techniques.

CONCLUSION

In this study, we attempted to estimate heterogeneous treatment effects of onset of the COVID-19 pandemic on risk of non-fatal overdose among a cohort of New York State Medicaid enrollees. Two subgroups, characterized by older age, structural disadvantage (i.e., Black or Hispanic racial/ethnic identity; disability status), and no baseline documentation of OUD, experienced a 54% and 57% increase in their risk of experiencing an overdose following onset of the pandemic, as compared to a 22% increase for the study population as a whole. These findings highlight the substantial public health burden of increased overdose among older Medicaid enrollees and points towards the importance of considering structural factors when developing prevention and intervention strategies to proactively and equitably respond to future societally disruptive Big Events.

Data availability

Data are not available outside of an executed data use agreement with the New York State Centers for Medicare and Medicaid Services. Code for running the analyses is available at https://github.com/JohnPamplinII/HTE-Scan-COVID-Overdose.