Opioid-related diagnoses and HIV, HCV, and mental disorders: Using Pennsylvania hospitalization data to assess community-level relationships over space and time

Abstract

Background.

We assessed the community-level spatiotemporal connections between hospitalizations for common opioid comorbidities (HIV, hepatitis C (HCV), and mental disorders) and opioid-related hospitalizations in the current and previous year.

Methods.

We used Bayesian hierarchical spatiotemporal Poisson regression with conditionally autoregressive (CAR) spatial effects to assess counts of HCV-, HIV-, and mental disorder-related hospitalizations at the ZIP code level from 2004 to 2014 in Pennsylvania. Models included rates of current- and previous-year opioid-related hospitalizations as well as covariates measuring demographic and environmental characteristics.

Results.

After adjusting for measures of demographic and environmental characteristics, current- and previous-year opioid-related hospitalizations were associated with higher risk of HCV, HIV, and mental disorders. The relative risks and 95% credible intervals for previous-year opioid-related hospitalizations were 1.092 [1.078, 1.106] for HCV, 1.098 [1.068, 1.126] for HIV, and 1.020 [1.013, 1.027] for mental disorders.

Conclusion.

Previous-year opioid-related hospitalizations are connected to common comorbid conditions such as HCV, HIV, and mental disorders, illustrating some of the broader health-related impacts of the opioid epidemic. Public health interventions focused on the opioid epidemic must consider individual community needs and comorbid diagnoses.

INTRODUCTION

Rates of opioid use disorder (OUD) and opioid poisonings have risen dramatically in the past decade. In 2016, unintentional injuries became the third leading cause of death in the United States¹, with drug poisonings as the primary contributor. Less direct outcomes related to OUD are equally troubling. Infectious diseases such as the hepatitis C virus (HCV) and the human immunodeficiency virus (HIV), as well as mental disorders, are major public health concerns connected to opioid misuse^2,3. Moreover, abundant evidence points to trends in each of these health conditions that vary across both space and time. Timely identification of vulnerable geographic regions can help us effectively direct preventive measures to reduce the burden of these key comorbidities.

HCV incidence has steeply increased since 2004, a pattern closely tied to the opioid epidemic⁴. Nationally, rates climbed from 0.6 to 1.0 per 100,000 people from 2012 to 2016. Yet, not all were equally impacted: in some states, rates declined (e.g., from 2.1 to 0.8 in Oklahoma); in others, they remained stable (e.g., 0.2 in California and Illinois); still in others, they soared (e.g., from 0.1 to 1.6 in Ohio and from 0.5 to 1.8 in Pennsylvania)⁵. Furthermore, there is evident spatial and demographic heterogeneity both between and within states. Most rapid rate increases were among young people in nonurban Appalachian regions⁶. Newly diagnosed individuals are commonly of non-Hispanic white or Hispanic race/ethnicity and/or people who inject drugs (PWID)^4,6.

Conversely, HIV incidence has declined and remained fairly stable since the 1990s due to significant advances in HIV treatment and prevention^7,8. However, declining trends are not universal, equal, or enduring across all communities; many continue to grapple with high rates of existing or emerging HIV cases⁸. Socioeconomic status, race/ethnicity, and location play major roles. Vulnerable communities tend to have some of the following characteristics: higher urbanicity^9,10, higher rates of poverty^9,11,12, limited access to healthcare¹². These communities also tend to have greater proportions of African American and/or Hispanic residents^11,13 and PWID⁹. Unfortunately, the opioid epidemic threatens to reverse HIV prevention and treatment efforts, as demonstrated by the 2015 outbreak in a Scott County, IN, community. In this rural community of about 4,200 people, fewer than five cases of HIV were reported prior to 2015. By the end of 2015, 181 new cases emerged^2,12, among whom 87.8% reported having injected opioid oxymorphone, and 92.3% had comorbid HCV¹².

The connection between mental disorders and OUDs is arguably more nuanced. Explanatory pathways are not clear-cut, and estimates of mental disorder prevalence trends have been mixed¹⁴. Yet, high comorbidity between substance abuse and mental disorders undeniably exists^15,16,17. For example, non-prescription opioid use is overrepresented among those with mental disorders, including depressive, bipolar, and anxiety disorders^3,18. In fact, though comprising 16% of Americans, those with mental disorders receive half of all opioid prescriptions in the United States¹⁹. Rates of certain mental disorders tend to be spatially autocorrelated²⁰. Characteristics of neighborhood disadvantage such as poverty, noisiness, and crime have been extensively linked to symptoms of depression^21,22 and, to a lesser extent, anxiety^23,24.

The relationship between OUD and poisonings and the spread of HIV, HCV, and mental disorders over space and time, as well as heterogeneities in high-risk areas, is not well understood. When available, spatially and temporally specific and consistently collected data are fundamental to understanding the spread and geographic distribution of health outcomes²⁵. Unfortunately, such data are commonly lacking for substance use-related outcomes. Hospital discharge data (HDD) are unique as they exist on a fairly well-resolved (ZIP code) level and contain consistently collected information for relevant diagnoses. A recent study used space-time HDD to identify associations of specific covariates by ZIP code, thus emphasizing idiosyncrasies of neighborhoods in connection to opioid-related hospitalizations²⁶. These results highlight the importance of considering individual needs of specific communities rather than offering blanket solutions.

In this study, we used HDD to examine spatiotemporal trends of HCV, HIV, and mental disorders. We were particularly interested in the effects of current- and previous-year OUD and poisoning diagnosis rates in connection to these outcomes. We also considered relative contributions of factors related to ZIP code-level medical needs, economic conditions, and other demographic and environmental covariates to the spatial patterns of these opioid comorbidities. We hypothesized that, beyond the contribution of current-year rates, past-year OUD and poisoning rates would be related to greater counts of current-year HIV, HCV, and mental disorders.

METHODS

We used Bayesian hierarchical misalignment models to assess three outcomes: diagnoses of HIV, HCV, and mental disorders (depressive, anxiety, and/or bipolar disorders), as indicated by International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes in hospital discharge records. We assessed the relationship between each outcome and opioid-related diagnoses, while adjusting for economic, demographic, and environmental conditions. To do this, we analyzed 11 years (from 2004 to 2014) of ZIP code-level HDD in Pennsylvania (n = 16,275 space-time units). As no human participants were involved in this study, institutional review board approval was not required.

Variables

We used Pennsylvania Healthcare Cost Containment Council (PHC4)²⁷ HDD to obtain counts of all relevant ICD-9-CM diagnoses. PHC4 data consist of approximately 1.9 million patient-level records per year across Pennsylvania, including ZIP code of residence, admitting diagnosis, principal diagnosis, up to 17 secondary diagnoses, and up to three E-codes (external cause of injuries and poisoning). Records or hospitalizations are overnight hospital stays recorded as hospital discharges. We excluded 951,478 (4.6% of observations available from 2004 through 2014) records that had missing or non-Pennsylvania residential ZIP codes. We also excluded secondary diagnoses past the ninth slot since only up to eight were permitted prior to 2011.

For HIV counts, we included diagnoses of HIV disease (042) or asymptomatic immunodeficiency virus infection (V08). For HCV, acute or chronic hepatitis C with or without hepatic coma (07041, 07051, 07054), unspecified viral hepatitis (07070, 07071), and/or hepatitis C carrier (V0262) codes were used. Mental disorder diagnoses included depressive disorder (311), major depressive disorder (296.2x, 296.3x), dysthymic disorder (300.4), anxiety states (300.0x), phobic disorders (300.2x), bipolar I disorder (296.0x, 296.4x, 296.5x, 296.6x), and other and unspecified bipolar or mood disorder (296.8x, 296.90, and 296.99). Individual records with multiple mental disorder diagnoses contributed only once to the total count of mental disorders.

Opioid-related diagnoses represented OUDs and opioid poisoning incidents. OUDs were measured by opioid abuse (305.5x), dependence (304.0x), or dependence with combinations of opioids with other drugs (304.7x) codes. Poisoning codes included poisoning by heroin (965.01), accidental heroin poisoning (E850.0), and non-heroin opioid-related poisoning codes (965.00, 965.02, 965.09, E850.1 and E850.2). Total counts of opioid-related diagnoses were converted to rates per total hospitalization count in each space-time unit. Overall hospitalization rates were included to control for inpatient care access.

Our primary interest was to relate each outcome to opioid-related diagnosis rates in the current and previous (temporally lagged) year. Due to the temporal instability of ZIP code boundaries, we could not directly link spatial units over time. Consequently, we matched all space-time units to previous-year units by greatest shared area. The temporally lagged value of a space-time unit was the previous year’s opioid-related diagnoses density of the unit with the greatest area overlap.

Block group-level U.S. Census variables were extrapolated to the ZIP code level²⁸ to reflect demographic conditions²⁹. Economic variables included median household income (per $10,000), percent of the population below 150% of the poverty line, and the unemployment rate. Other characteristics included age (percent of the population in age ranges 0–19, 20–24, 25–44, and 45–64), race/ethnicity (percent non-Hispanic white, black, and Hispanic), percent male, and population density per square mile.

Models included measures of manual labor and retail establishment densities; these counts originated from the North American Industry Classification System (NAICS) codes from Zip Code Business Patterns data³⁰. Manual labor industries included: Agriculture, Forestry, Fishing and Hunting (NAICS Sector 11); Mining, Quarrying, and Oil and Gas Extraction (Sector 21); Construction (Sector 23); Manufacturing (Sectors 31–33); Wholesale Trade (Sector 42); Transportation and Warehousing (Sectors 48–49); and Utilities (Sector 22). For retail density, we used Retail Trade (Sectors 44–45) and Accommodation and Food Services (Sector 72) counts. Counts were converted to densities per square mile and multiplied by 10 (credible intervals were otherwise too narrow to interpret).

Boundaries and counts of ZIP codes change with time, resulting in spatiotemporal misalignment. In addition to using a flexible spatial random effects framework (described in the data analysis section), we included a measure of geographic instability, calculated as the proportion of each space-time unit’s population that would have fallen outside of the boundaries of the previous-year best-matched unit²⁸.

Finally, 83 ZIP codes had a population count of zero. To allow for non-zero risk in all regions, the population count was amended to five people in such cases. These areas were then assigned mean values of all Census-derived variables in the same year.

Data analysis

We used Bayesian hierarchical space-time models with conditionally autoregressive (CAR) spatial random effects to estimate these models. ZIP code boundaries and counts (ranging from 1,469 in 2004 to 1,490 in 2014) changed over time, requiring use of methods that could handle spatial units not nested across time³¹. To account for small area issues, spatial dependence, and overdispersion, observation-level CAR spatial and spatially unstructured random effects were included³². Non-informative priors were specified for all fixed and random effects. Poisson regression was used:

$$Y_{i,t}|{\mu }_{i,t}~\mathit{\text{Poisson}}\left(E_{i,t}\text{exp}\left({\mu }_{i,t}\right)\right).$$

Y_i,t is the count of each outcome. E_i,t is the expected count based on the total population of each spatial unit i at time t. exp μ_i,t is the relative rate for each unit:

$${\mu }_{i,t}=\beta {\mathit{X}}_{i,t}+{\varphi }_{i,t}+{\theta }_{i,t}.$$

X_i,t is the matrix of observed fixed effects, and β is a vector of coefficient etimates for each fixed effect. ϕ_i,t and θ_i,t are vectors of observation-level spatial and non-spatial random effects, respectively.

Each outcome was considered separately with the same set of covariates. All models adjusted for rates of prior- and current-year opioid-related hospitalizations, economic conditions, demographic and environmental variables, and ZIP code misalignment.

In reporting results, we consider an effect “well-supported” if the proportion of its posterior distribution falling above an RR of 1.000 is outside the range of (0.025, 0.975)—that is, it is not within the 95% credible interval. As this cutoff is fairly arbitrary, assessing just how much of a posterior distribution is above or below “equal risk” is often more informative. This statistic can be used to elucidate findings; for example, although the percentage of Hispanic residents in a ZIP code is not a well-supported effect in Model 3, only 0.032 of this posterior distribution is above 1.000.

To fit these models, we used the R package R-INLA³³. This package has been widely used to fit Bayesian hierarchical spatiotemporal models using Integrated nested Laplace approximation (INLA). INLA is a deterministic approach that has been demonstrated to provide accurate approximations of posterior distributions of complex Bayesian integrals³⁴.

RESULTS

There were an average of 18.25 (SD=48.25) HCV hospitalizations, 6.66 (SD=29.78) HIV hospitalizations, and 191.50 (SD=284.27) mental disorder hospitalizations per ZIP code-year in Pennsylvania from 2004 to 2014 (Table 1). During this time, the mean percentage of patients with opioid-related primary or secondary diagnoses was 1.21, which increased by 0.93 from 2004 to 2014. Whereas mean counts of HCV and mental disorders increased during the study period (by 0.41 and 41.40, respectively), counts of HIV diagnoses dropped by an average of 1.08. An average of 21.33% (SD=10.03) of residents lived below 150% of the poverty line, and an average of 7.42% (SD=5.24) were unemployed. The median household income was $44,812.44. Pennsylvania has a fairly small minority population (mean 89.12% white). Between two large urban areas on its borders, it is a rural state (mean population density 1,278.49/mi²; SD=3,111.11). There were an average of 10.42 (SD=58.46) retail establishments and 6.52 (SD=17.09) manual labor establishments per square mile.

Table 1.

Descriptive statistics, ZIP codes in Pennsylvania, 2004–2014 (n = 16,275 ZIP codes)

Variable	Mean	Standard Deviation	Minimum	Maximum	Mean change, 2014 vs. 2004
Population size	8490.78	11258.67	5.00	73131.91	221.08
Number of HCV hospitalizations	18.25	48.25	0	712	0.41
Number of HIV hospitalizations	6.66	29.78	0	478	−1.08
Number of depression, anxiety, and bipolar disorder, hospitalizations	191.50	284.27	0	2421	41.40
Opioid-related diagnoses, per 100 hospitalizations	1.21	1.40	0.00	50.00	0.93
Demographic and environmental covariates
Economic conditions
Below 150% of the poverty line, %	21.33	10.03	0.00	86.65	3.50
Median household income ($10,000)	4.48	1.67	0.00	18.46	0.859
Unemployment, %	7.42	5.24	0.00	100.00	2.30
Race/ethnicity, %
Black	4.28	11.77	0.00	98.01	−0.07
Hispanic	2.24	5.50	0.00	71.45	0.37
White	89.12	15.27	0.40	100.00	−6.27
Age, %
0–19	24.02	3.71	0.00	50.00	−1.69
20–24	6.60	1.34	0.00	35.53	0.19
25–44	24.81	3.56	0.00	77.43	−1.83
45–64	27.41	3.15	0.43	68.75	0.80
Male, %	49.42	2.73	5.26	100.00	0.01
Population density, per mi2	1278.49	3111.11	0.10	36466.28	21.50
Number of retail establishments, per mi2 (x10)	104.20	584.60	0.00	18565.47	−6.49
Number of manual labor establishments, per mi2 (x10)	65.22	170.89	0.00	2697.41	−12.40
Overall hospitalization rate, per 100 people	18.77	53.75	0.00	178.00	−0.23

The percentage of previous-year opioid-related diagnoses was positively associated with each outcome (RRs [95% CIs] were 1.092 [1.078, 1.106], 1.098 [1.068, 1.126], and 1.020 [1.013, 1.027] for the HCV, HIV, and mental disorders models, respectively), even after adjusting for current-year opioid-related diagnoses (Table 2). As a sensitivity analysis, we ran similar models for other health-related outcomes, including counts of arthritis and cancer hospitalizations. In these models, the effect of opioid-related hospitalizations was not well-supported (data not shown).

Table 2.

Relative rates (RRs) [95% credible intervals] and proportion of the marginal distributions falling above 1, HCV, HIV, and mental disorders (depression, anxiety, and bipolar disorder) hospitalizations, Bayesian Spatial Misalignment (n = 16,275 ZIP codes)

	Model 1: HCV		Model 2: HIV		Model 3: Mental Disorders
Variable	RR [95% credible interval]	Proportion above 1.000	RR [95% credible interval]	Proportion above 1.000	RR [95% credible interval]	Proportion above 1.000
Year	1.000 [0.995, 1.005]	0.445	0.984 [0.974, 0.994]a	0.001	1.003 [1.001, 1.005]a	0.994
Opioid-related diagnoses
Current-year diagnoses, % of total hospitalizations	1.175 [1.162, 1.189]a	>0.999	1.102 [1.073, 1.129]a	>0.999	1.055 [1.049, 1.062]	>0.999
Previous-year diagnoses, % of total hospitalizations	1.092 [1.078, 1.106]a	>0.999	1.098 [1.068, 1.126]a	>0.999	1.020 [1.013, 1.027]	>0.999
Demographic and environmental conditions
Economic conditions
Median household income, per $10,000	0.890 [0.877, 0.902]a	<0.001	0.931 [0.908, 0.953]a	<0.001	0.940 [0.934, 0.946]a	<0.001
Below 150% of the poverty line, %	1.016 [1.014, 1.019]a	>0.999	1.018 [1.014, 1.023]a	>0.999	1.010 [1.009, 1.011]a	>0.999
Unemployment, %	1.001 [0.998, 1.004]	0.630	0.993 [0.988, 0.999]a	0.006	1.001 [0.999, 1.002]	0.868
Race/ethnicity, %
Black	1.012 [1.008, 1.015]a	>0.999	1.012 [1.006, 1.019]a	>0.999	1.001 [<1.000, 1.002]	0.891
Hispanic	1.002 [0.999, 1.006]	0.898	1.005 [0.998, 1.011]	0.916	0.998 [0.997, >1.000]	0.032
White	1.000 [0.997, 1.004]	0.564	0.982 [0.975, 0.988]a	<0.001	1.003 [1.001, 1.004]a	>0.999
Age, %
0–19	0.981 [0.977, 0.986]a	<0.001	0.972 [0.963, 0.981]a	<0.001	0.987 [0.985, 0.990]a	<0.001
20–24	0.932 [0.921, 0.944]a	<0.001	0.912 [0.890, 0.934]a	<0.001	0.957 [0.951, 0.966]a	<0.001
25–44	0.986 [0.981, 0.991]a	<0.001	0.989 [0.979, 0.998]a	0.009	0.976 [0.973, 0.978]a	<0.001
45–64	0.995 [0.988, 1.003]	0.111	0.988 [0.974, 1.003]	0.054	0.994 [0.991, 0.998]a	<0.001
Male, %	1.001 [0.996, 1.007]	0.680	0.998 [0.988, 1.009]	0.381	0.992 [0.989, 0.995]a	<0.001
Population density, per mi2 b
Quintile 2	1.353 [1.283, 1.428]a	>0.999	1.597 [1.400, 1.820]a	>0.999	1.204 [1.180, 1.229]a	>0.999
Quintile 3	1.503 [1.419, 1.592]a	>0.999	1.872 [1.641, 2.141]a	>0.999	1.340 [1.309, 1.373]a	>0.999
Quintile 4	1.761 [1.655, 1.874]a	>0.999	2.619 [2.290, 3.002]a	>0.999	1.518 [1.478, 1.559]a	>0.999
Quintile 5	1.874 [1.745, 2.013]a	>0.999	3.307 [2.861, 3.830]a	>0.999	1.546 [1.494, 1.600]a	>0.999
Overall hospitalization rate, per 100 people	1.005 [1.005, 1.005]a	>0.999	1.004 [1.003, 1.005]a	>0.999	1.005 [1.005, 1.006]a	>0.999
Manual labor, number of establishments per mi2 (x10)	1.003 [1.002, 1.004]a	>0.999	1.003 [1.001, 1.005]a	>0.999	1.005 [1.004, 1.005]a	>0.999
Retail clutter, number of establishments per mi2 (x10)	1.000 [<1.000, >1.000]	0.420	1.001 [>1.000, 1.001]a	0.992	1.000 [0.998, <1.000]a	0.003
Misalignment effects
ZIP code instability	1.000 [0.997, 1.004]	0.584	0.986 [0.976, 0.995]a	0.001	1.003 [1.002, 1.005]a	>0.999
Random effects
Spatial random effects (s.d. CARc process)	0.504 [0.499, 0.509]		0.421 [0.417, 0.425]		0.346 [0.343, 0.349]
ZIP code-level random effects (s.d.)	0.304 [0.285, 0.324]		0.691 [0.667, 0.716]		0.161 [0.151, 0.177]
Spatial to total random variability ratiod	0.733 [0.707, 0.758]		0.270 [0.256, 0.286]		0.821 [0.793, 0.841]

^aIndicates findings that are well-supported by the data as evidenced by credible intervals that exclude one for relative risks;

^bpopulation density was divided into approximately equal quintiles as follows: 1 (referent; < 52.5 people / sq. mile), 2 (52.6 – 124.1), 3 (124.2 – 333.3), 4 (333.4 – 1414.2), and 5 (1414.3+);

^cCAR = conditional autoregressive;

^dcalculated as the variance ratio of spatial to spatial and non-spatial random effects

Lower median household income and higher percentage of residents living below 150% of the poverty line were positively associated with greater RR for all outcomes. The “protective” effect of greater median household income was especially pronounced in the HCV model (0.890 [0.877, 0.902]). Unemployment was associated with lower HIV RRs; this effect was not well-supported in the other models. ZIP codes with higher proportions of black residents tended to have greater RRs of HCV and HIV (>99.9% above 1.000), whereas areas with greater proportions of white residents tended to experience higher RRs for mental disorders (>99.9% above 1.000) but lower RRs for HIV (>99.9% below 1.000). Relative to the percentage of residents over age 64, higher percentages of younger age groups tended to be negatively associated with each outcome.

Areas with greater population density experienced greater RRs for all outcomes, particularly in the HIV model (e.g., the fifth density quintile had an RR [95% CI] of 3.307 [2.861, 3.830] compared to the first quintile). Density of manual labor establishments was associated with higher RRs for each outcome, whereas the effect of retail clutter density varied. ZIP code instability related to proportion of population change in a given ZIP code by year differed by outcome as well: it was not well-supported in the HCV model, negatively associated with HIV, and positively associated with mental disorders. In contrast to the HIV model (0.270 [0.256, 0.286]), the ratio of spatial to total random variability in the HCV (0.733 [0.707, 0.758]) and mental disorders (0.821 [0.793, 0.841]) models was quite high (Table 2), suggesting more spatial dependence exists in these models.

DISCUSSION

Our analyses demonstrate a population-level connection between the opioid epidemic and commonly comorbid conditions. Risks for HCV, HIV, and mental disorders were universally connected to higher rates of current- and previous-year opioid-related hospitalizations. In fact, each percent increase in previous-year opioid-related hospitalizations was associated with a 9.2%, 9.8%, and 2.0% median RR increases in the HCV, HIV, and mental disorders models, respectively. This finding supports a temporal connection between opioid-related hospitalizations and future comorbidities.

Consistent with other studies^9,11,12, increased risk for each of these outcomes is connected to economic conditions, such as lower median household income and higher poverty rates. However, our analyses also indicated patterns inconsistent with other studies. For example, in contrast to other findings^4,11, HCV and HIV outcomes were not associated with greater proportions of Hispanic residents.

To distinguish whether opioid-related hospitalizations were related specifically to HIV, HCV, and mental disorders rather than simply being correlated with any other health-related outcome, we considered analogous models with other outcomes, including arthritis or cancer. Interestingly, neither current- nor previous-year opioid diagnoses were well-supported in these models. This further corroborates the link between opioid misuse and future risk for related disorders. Furthermore, the independent relationship of past-year OUD with current-year comorbid conditions, above and beyond cross-sectional associations, strengthens the temporal ordering: it appears that greater OUD impact subsequent rates of common comorbidities.

Many areas within Appalachia and surrounding regions (e.g., Ohio, Pennsylvania) have experienced increased HCV rates in recent years⁵. While rates are increasing most rapidly among nonurban youth⁶, our analyses suggest that urban areas are still at greatest risk (e.g., compared to the least densely populated areas, areas with greatest population density have an RR [95% CI] of 1.874 [1.745, 2.013]) (Table 2). Increased HIV risk in urban areas has been more consistently established^9,10. Among our three outcomes, HIV was most closely associated with areas of highest population density (RR [95% CI] of 3.307 [2.861, 3.830] compared to least dense areas).

Bayesian hierarchical CAR models employed in this study have been widely used to address spatial autocorrelation, an important artifact that, when ignored, can heavily bias findings by falsely presuming independence among units. Overall spatial clustering was evident for HIV and HCV counts, with Moran’s I values of 0.633 and 0.571, respectively. This implies that substantial autocorrelation exists at the ZIP code level for these outcomes. For mental disorders, the Moran’s I value was much lower (0.138) but still significant; furthermore, the ratio of spatial to total random variability in this model was high (0.733 [0.707, 0.758]) (Table 2).

Our analyses have several limitations. First, we cannot extend findings to finer (e.g., individual-level) or grosser (e.g., state-level) scales. Furthermore, this study was conducted in Pennsylvania, a predominantly Appalachian state with many high-risk regions. Our results point to vast spatial heterogeneity, suggesting that many regions outside of Pennsylvania will also have unique characteristics. Finally, we assessed overnight hospital stays, which likely underestimate all outcomes. This limits the generalizability of our findings. However, there is no better spatially-resolved (i.e., below the county level) way to measure these outcomes as no surveillance systems are in place with finer geographic resolution. Though HDD are not accurate representations of overall burden of these comorbidities, they provide consistent measures across ZIP codes and years, enabling our examination of the spread of these conditions across both space and time.

Due to misalignment of ZIP codes between years, we cannot directly assess differential growth using methods employed in these analyses. For example, while we know that youth in nonurban Appalachian regions have experienced particularly high increases of HCV rates (CDC, 2018b), our analyses indicate that urban regions are still at greatest risk, but we cannot compare rates of growth between ZIP codes. Future studies may consider other methods, such as continuous spatial frameworks implied by the stochastic partial differential equation approach³⁵, which would allow us to trace local trends through time.

Community needs in combating the opioid epidemic differ. Comorbidity prevalence, economic conditions, and demographics are some of the factors that must be considered in tailoring interventions to meet these specific needs. The connection between opioid-related hospitalizations and future risk of HCV, HIV, and mental disorders at the ZIP code level can aid in these efforts. Particularly, tracking rates of comorbid conditions can assist in proper allocation of resources.

What is already known on this subject?

It is known that opioid use disorders are connected to comorbid conditions such as HIV, hepatitis C, and mental disorders.

What does this study add?

This study finds a connection between opioid-related hospitalizations and future risk of these conditions at the ZIP code level. These findings highlight the need to track rates of comorbid conditions in communities hit hardest by the opioid epidemic, which can aid public health efforts in proper allocation of resources. Furthermore, as community needs differ, efforts to combat the opioid epidemic must consider demographic and economic differences between areas.