Tobacco-, Alcohol-, and Drug-Attributable Deaths and Their Contribution to Mortality Disparities in a Cohort of Homeless Adults in Boston

Abstract

Objectives. We quantified tobacco-, alcohol-, and drug-attributable deaths and their contribution to mortality disparities among homeless adults.

Methods. We ascertained causes of death among 28 033 adults seen at the Boston Health Care for the Homeless Program in 2003 to 2008. We calculated population-attributable fractions to estimate the proportion of deaths attributable to tobacco, alcohol, or drug use. We compared attributable mortality rates with those for Massachusetts adults using rate ratios and differences.

Results. Of 1302 deaths, 236 were tobacco-attributable, 215 were alcohol-attributable, and 286 were drug-attributable. Fifty-two percent of deaths were attributable to any of these substances. In comparison with Massachusetts adults, tobacco-attributable mortality rates were 3 to 5 times higher, alcohol-attributable mortality rates were 6 to 10 times higher, and drug-attributable mortality rates were 8 to 17 times higher. Disparities in substance-attributable deaths accounted for 57% of the all-cause mortality gap between the homeless cohort and Massachusetts adults.

Conclusions. In this clinic-based cohort of homeless adults, over half of all deaths were substance-attributable, but this did not fully explain the mortality disparity with the general population. Interventions should address both addiction and non-addiction sources of excess mortality.

Over 2 million people experience homelessness annually in the United States.¹ US-based studies of homeless adults have found varied but generally high rates of substance use, including a 68% to 80% prevalence of cigarette smoking,^2–6 a 29% to 63% lifetime prevalence of alcohol use disorders,^7–13 and a 20% to 59% lifetime prevalence of drug use disorders.^7–13 Mortality rates among homeless individuals exceed those in non-homeless people by a considerable margin,^14–26 but the extent to which substance use contributes to this disparity is uncertain.

In a prior study of homeless adults in Boston, Massachusetts, we found the rate of deaths caused by drug overdose to be approximately 20 times higher than the rate in the general population.¹⁴ Similar findings have been documented in New York City, New York, and San Francisco, California.^27,28 Overdose deaths draw attention to the overt complications of drug use among homeless people but represent only a single dimension of substance-attributable mortality. Tobacco, alcohol, and drug use may contribute to other common causes of death among homeless people, such as cancer, heart disease, liver cirrhosis, and HIV.^14–17 The population attributable fraction (PAF) represents the proportion of these deaths that would not have occurred in the absence of tobacco, alcohol, or drug use.^29,30 PAF methods form the basis of studies conducted by the US Centers for Disease Control and Prevention (CDC)^31,32 and the Global Burden of Disease investigators³³ to quantify the impact of substance use and other risk behaviors on health outcomes.

In this study, we describe a novel application of PAF-based methods to estimate tobacco-, alcohol-, and drug-attributable mortality rates among 28 033 adults who used Boston Health Care for the Homeless Program (BHCHP) services from 2003 to 2008. We compared our findings with the Massachusetts general population to examine the degree to which mortality disparities between these groups are driven by substance-attributable deaths. Quantifying the burden of tobacco-, alcohol-, and drug-attributable deaths among homeless individuals has the potential to inform upstream efforts at disease prevention and guide clinical priority setting around the delivery of addiction services for this group of people. Additionally, understanding the extent to which these deaths contribute to the mortality gradient between homeless adults and the general population may provide insight about the ways to reduce this disparity.

METHODS

We assembled a cohort of all adults aged 18 years and older who had an in-person encounter at BHCHP between January 1, 2003, and December 31, 2008. BHCHP serves more than 11 000 currently and formerly homeless individuals annually in over 90 000 outpatient medical, oral health, and behavioral health encounters through a network of more than 70 service sites based in emergency shelters, transitional housing facilities, hospitals, and other social service settings in greater Boston.^34,35 Individuals must be homeless to enroll in services at BHCHP; however, some patients continue receiving care at the program after they are no longer homeless. We did not have housing status data for the study cohort, but an internal analysis of individuals seen at BHCHP in 2011 found that 16% of patients were housed, most without supportive services and therefore at potentially high risk for recurrent homelessness.

We cross-linked the BHCHP cohort to the Massachusetts Registry of Vital Statistics 2003 to 2008 death occurrence files using probabilistic record linkage methods detailed elsewhere.¹⁴ Individuals were observed from the date of first contact within the study period until the date of death or December 31, 2008. For those who died, we based the cause of death on the International Statistical Classification of Diseases and Related Health Problems, 10th Revision^35a (ICD-10) code listed in the underlying cause of death field in the death occurrence file.

Population Attributable Fraction Estimation

We estimated PAFs for 57 causes of death etiologically related to the use of tobacco, alcohol, or drugs (Table A, available as a supplement to the online version of this article at http://www.ajph.org). We used 1 of 3 methods to estimate PAFs.

Formula method.

Our principal PAF estimation method relied on the following formula:

for which p_i is the prevalence of substance consumption at the i th level, RR_i is the relative risk for a given cause of death at the i th consumption level, and the counterfactual exposure distribution is no substance use.^{30–33,36–38} We used this method primarily for causes of death associated with a chronic pattern of substance use (e.g., lung cancer related to tobacco smoking).

Case series method.

Certain causes of death lack robust prospective data to quantify the risk association with substance use (e.g., HIV related to injection drug use) or correlate more closely with acute rather than chronic patterns of substance use (e.g., homicide related to alcohol use). In these instances, we followed the precedent of the CDC^32,39 and others^38,40–48 in using morbidity and mortality case series data to approximate the PAFs based on the proportion of cases where the use of a particular substance was causally implicated in the condition or injury.

Definitional attribution method.

Certain ICD-10 codes assign substance attribution by definition (e.g., accidental drug poisoning [X40-X44]). We considered these deaths 100% attributable to the named substance and assigned them a PAF of 1.^38,44,45

Tobacco-Attributable Mortality

We relied chiefly on the formula method to estimate PAFs for tobacco-attributable mortality.³¹ Because smoking status data were unavailable for the BHCHP cohort, we used age- and gender-specific current and former smoking prevalence data from the nationally representative 2003 Health Care for the Homeless (HCH) User Survey.^6,49 Because smoking prevalence did not vary significantly by geographic region, we used the full national sample for improved precision. Reweighting the HCH User Survey to match the race distribution of the BHCHP cohort did not meaningfully alter the prevalence or PAF estimates, so we used the original analysis weights developed by RTI International.⁴⁹ We used age- and gender-specific RR estimates for smoking-related causes of death, adjusted for race and education, from updated analyses of the Cancer Prevention Study II (1982–1988) and the pooled contemporary cohorts (2000–2010) published in the 2014 US surgeon general’s report.⁵⁰ We followed the CDC precedent in estimating smoking-attributable deaths only for adults 35 years and older.^31,50

Alcohol-Attributable Mortality

We used the formula method to estimate PAFs for causes of death related to chronic patterns of alcohol use. For these calculations, we used age- and gender-specific prevalence data for low (0.01–25g/day), medium (25.01–50g/day), and high (> 50g/day) average daily alcohol consumption from the 2003 HCH User Survey.⁴⁹ The HCH User Survey deployed items from the Behavioral Risk Factor Surveillance System (BRFSS) to assess alcohol drinking frequency, typical drinking quantity, and binge occasions over the past 30 days. We assumed a standard drink to contain 14 grams of ethanol.⁵¹ We applied an indexing method to incorporate binge episodes into the calculation of average daily alcohol consumption.⁵² We used pooled RR estimates from meta-analyses by the Global Burden of Disease studies^33,53 and Corrao et al.,⁵⁴ which were based on effect sizes that had been adjusted for measured confounders when possible.

The risk for alcohol-related injury deaths is influenced more by heavy drinking occasions than by average daily consumption.^44,46 Therefore, we followed the CDC³² and others^{38,40–42,44–47} in using the case series method to estimate PAFs for these deaths. For traffic injuries, we used customized data on the age- and gender-specific proportions of traffic fatalities involving a blood alcohol concentration (BAC) of 100mg/dL or greater in Massachusetts in 2003 to 2008, provided by the National Highway Traffic Safety Administration Fatality Analysis Reporting System (http://www.nhtsa.gov/FARS). For nontraffic injuries, we used data from a meta-analysis of 331 medical examiner studies to estimate PAFs based on the proportions of decedents with a BAC of 100mg/dL or greater.⁵⁵ The use of a BAC cutoff of 100mg/dL assumes that alcohol levels above this threshold are causally related to injury while conservatively treating levels below this value as noncontributory.⁴⁶

Various other causes of death were 100% attributable to alcohol by definition. Although there are ICD-10 codes for alcohol-related liver cirrhosis and pancreatitis, we followed the precedent of other studies^38,56 in applying the formula method to these deaths because prior evidence has suggested that relying solely on the alcohol-specific ICD-10 codes may underestimate alcohol attribution. We did not examine the protective effect of alcohol for certain causes of death (e.g., ischemic heart disease) because our primary goal was to enumerate the adverse consequences of alcohol consumption.

Drug-Attributable Mortality

The methods for estimating drug-attributable mortality are less well-defined because of the large number of potentially abusable drugs, the challenges in estimating drug use prevalence, and the paucity of prospective RR estimates for drug-related causes of death.⁴⁵ We selected a conservative list of causes of death that other studies have considered drug-related.

Consistent with the CDC’s approach,³⁹ we used the case series method to estimate PAFs for deaths caused by hepatitis B, hepatitis C, HIV, and homicide. As an approximation of the mortality PAFs for hepatitis B and C, we used CDC surveillance data on the proportion of acute hepatitis B and C cases diagnosed in 2007 where injection drug use (IDU) was a transmission risk factor.⁵⁷ As an approximation of the mortality PAF for HIV, we used CDC surveillance data on the proportion of acquired immunodeficiency syndrome (AIDS) cases assigned a transmission category of IDU cumulative through 2008,⁵⁸ reasoning that deaths from HIV would most likely proceed through the development of AIDS. For homicide, we based the PAF on 2003 to 2007 murder circumstance data from the US Department of Justice.⁵⁹ The remainder of drug-related deaths were 100% attributable by definition.

Overlapping Attribution

Certain causes of death can be attributed to more than 1 substance. When estimating the combined mortality burden of tobacco, alcohol, and drug use, a simple sum overestimates attributable deaths because of double-counting. To avoid this, we calculated the summary PAF for a given cause of death attributable to S substances using the following formula^30,60–62:

This approach assumes that the exposure distributions for tobacco, alcohol, and drug use are independent and that there is no risk ratio effect modification. These assumptions are conservative when the exposure distributions are positively correlated,⁶² as is the case for tobacco, alcohol, and drug use,⁶ and when there is supra-multiplicative risk ratio interaction,⁶² such as with tobacco and alcohol use and the risk for upper aerodigestive tract cancers.^63–66

Massachusetts Comparison

We used the methods detailed previously to estimate the individual and combined mortality burden attributable to tobacco, alcohol, and drug use in the 2003 to 2008 Massachusetts adult population. For these analyses, we obtained cause of death data from the CDC Wide-Ranging Online Data for Epidemiologic Research (WONDER) detailed mortality files,⁶⁷ substance use prevalence estimates from our own analysis of 2001 to 2005 Massachusetts BRFSS data,⁶⁸ and RR estimates and case series data from the sources in Table A (available as a supplement at www.ajph.org).

Statistical Analysis

We used Monte Carlo simulations to estimate age- and gender-specific PAFs for each substance-related cause of death while incorporating uncertainty in the data elements used to calculate the PAF.^30,33,53,69 For PAFs estimated using the formula method, we took 1000 random draws from the log-normal distribution of RR estimates and 1000 random draws from the binomial distribution of substance use prevalence estimates to generate 1000 age- and gender-specific PAFs for each cause of death. We used the same RR draw for all age and gender categories when the reported RR did not vary across these demographic groups, and we used the same substance use prevalence draw across all causes of death that were related to the same substance.³³ For PAFs estimated using the case series method, we took 1000 random draws from the binomial distribution of PAFs based on the sample sizes of the case series from which they were derived. PAFs based on the definitional attribution method were fixed at 1.

We multiplied the resulting 1000 PAFs for each age, gender, and cause of death stratum by the stratum-specific number of deaths to generate 1000 estimates of attributable deaths. These were summed at the simulation level across all age, gender, and cause of death strata to generate 1000 estimates of deaths attributable to a single substance, and then summed across substances to generate a composite estimate of all substance-attributable deaths. We ordered the simulation results and reported the 25th, 500th, and 975th values as the lower confidence bound, point estimate, and upper confidence bound, respectively. We divided the number of attributable deaths by the person-years at risk to generate substance-attributable mortality rates.

We compared substance-attributable mortality rates in the BHCHP cohort to those in the Massachusetts population using rate ratios (RtRs) and rate differences (RtDs). To examine the contribution of substance-attributable mortality disparities to all-cause mortality disparities between the BHCHP cohort and Massachusetts adults, we divided the substance-attributable mortality rate difference by the all-cause mortality rate difference within each age and gender stratum. This ratio, which we term the substance-attributable disparity fraction, reflects the estimated proportion of excess mortality in each age–gender stratum of the BHCHP cohort that is attributable to tobacco, alcohol, or drug use. To generate an overall age- and gender-adjusted estimate of the substance-attributable disparity fraction for 20- to 64-year-old BHCHP patients in comparison with similarly aged Massachusetts adults, we divided the standardized substance-attributable mortality rate difference by the standardized all-cause mortality rate difference, where the age and gender distribution of the BHCHP cohort served as the standard. We included individuals 65 years and older in estimating the overall number of tobacco-, alcohol-, and drug-attributable deaths in the BHCHP cohort, but we did not estimate age-specific substance-attributable mortality rates for this group as a result of their less reliable substance use prevalence estimates.

We used SAS software, version 9.3 (SAS Institute, Cary, NC), and Microsoft Excel 2003 (Microsoft Corporation, Redmond, WA) to conduct our analyses.

RESULTS

Overall, 28 033 adults were followed for a median of 3.3 years, totaling 90 450 person-years of observation. The mean age at cohort entry was 41.0 years, and 76.4% were under the age of 50. Sixty-six percent of participants were male. Forty-three percent were White, 28.8% were Black, and 18.9% were Hispanic.

Numbers and Causes of Substance-Attributable Deaths

Of 1302 deaths, 236 (95% CI = 215, 254) were tobacco-attributable, 215 (95% CI = 199, 232) were alcohol-attributable, and 286 (95% CI = 285, 287) were drug-attributable. The leading cause of tobacco-attributable death was trachea, bronchus, and lung cancer, followed by ischemic heart disease (Table 1). The leading cause of alcohol-attributable death was alcohol use disorder. The leading cause of drug-attributable death was accidental drug poisoning.

TABLE 1—

Leading Causes of Tobacco-, Alcohol-, and Drug-Attributable Deaths: Boston Health Care for the Homeless Program, MA, 2003–2008

Cause of Death (ICD-1035a Codes)	Deaths, No.	Attributable Deathsa (95% CIb)
Tobacco
Trachea, bronchus, and lung cancer (C33-C34)	74	69 (66, 70)
Ischemic heart disease (I20-I25)	114	68 (52, 80)
Nonischemic heart disease (I00-I09, I26-I47, I49-I51)	43	21 (13, 27)
Chronic obstructive pulmonary disease (J40-J44)	18	15 (14, 16)
Total		236 (215, 254)
Alcohol
Alcohol use disorder (F10)	71	71 (NA)
Drug poisoning, accidental (X40-X44)	183	53 (41, 66)
Chronic liver disease and cirrhosis (K70, K73-K74)	58	36 (30, 40)
Hypertension (I10-I15)	47	12 (6, 18)
Total		215 (199, 232)
Drugs
Drug poisoning, accidental (X40-X44)	183	183 (NA)
Drug poisoning, undetermined intent (Y10-Y14)	36	36 (NA)
Drug use disorder (F11-F16, F18-F19)	28	28 (NA)
HIV (B20-B24)	76	19 (19, 19)
Total		286 (285, 287)

Note. CI = confidence interval; NA = not applicable.

^aAttributable deaths are the estimated number of deaths as a result of each cause that would not have occurred in the absence of tobacco, alcohol, or drug use. They are calculated by multiplying the age- and gender-specific population attributable fraction (PAF) for each cause of death by the corresponding count of these deaths in each age–gender stratum and then summing across all strata.

^b95% CIs were derived using simulation methods. Certain causes of death (e.g., alcohol use disorder) have no variance in the attributable death estimate because these deaths were 100% attributable to the named substance by definition. There was negligible variance in the attributable number of deaths caused by HIV because PAF estimates were based on population surveillance data with large sample sizes.

After accounting for overlap in tobacco-, alcohol-, and drug-attributable causes, 676 (95% CI = 655, 698) deaths were attributable to any of these substances, representing 51.9% (95% CI  = 50.3%, 53.6%) of all deaths in the BHCHP cohort (Figure 1). Among the 587 adults who died before the age of 50 years, 57.2% (95% CI = 55.7%, 58.5%) of deaths were substance-attributable, with the individual and combined effects of alcohol and drug use accounting for the vast majority of these deaths (Figure 1). Among the 715 decedents 50 years and older, 47.7% (95% CI = 45.1%, 50.3%) of deaths were substance-attributable and tobacco use accounted for over half of these deaths (Figure 1).

FIGURE 1—

Proportion of deaths attributable to tobacco, alcohol, and drug use among decedents (a) of all ages (N = 1302), (b) younger than 50 years (n = 587), and (c) 50 years and older (n = 715): Boston Health Care for the Homeless Program, MA, 2003–2008.

Substance-Specific Mortality Rates

The majority of tobacco-attributable deaths occurred among 50- to 64-year-old men and women at rates over 3-times higher than rates in comparably aged adults in Massachusetts (Figure 2). Although relatively fewer tobacco-related deaths occurred among 35- to 49-year-old BHCHP patients, tobacco-attributable mortality rates in this age group were still about 4- to 5-times higher than in the general population.

FIGURE 2—

Age- and gender-stratified tobacco-, alcohol-, and drug-attributable mortality rates for (a) men aged 20–34 years, (b) women aged 20–34 years, (c) men aged 35–49 years, (d) women aged 35–49 years, (e) men aged 50 –64 years, and (f) women aged 50–64 years: Boston Health Care for the Homeless Program, MA, and the Massachusetts adult population, 2003–2008.

Note. RtD = rate difference; RtR = rate ratio. The lower 95% confidence bounds are greater than 1 for all rate ratios and greater than zero for all rate differences.

Alcohol-attributable mortality rates tended to increase with age. In comparison with Massachusetts adults, alcohol-attributable mortality rates in the BHCHP cohort were about 6 to 10 times higher among women and 6 to 8 times higher among men (Figure 2).

Drug-attributable mortality rates peaked in the 35- to 49-year age range but were relatively high across all age and gender categories, exceeding those in the Massachusetts population by a factor of 8 to 17 for women and 10 to 14 for men (Figure 2).

Combined Substance-Attributable Mortality Rates

Attributable mortality rates for all substances combined increased with age (Figure 3). In comparison with Massachusetts adults, stratified substance-attributable mortality rates were about 5 to 11 times higher in the BHCHP cohort, whereas non–substance-attributable mortality rates were about 3 to 6 times higher. After age- and gender-standardization, the substance-attributable mortality rate for 20- to 64-year-old BHCHP patients was 6.2-times higher (95% CI = 5.8, 6.6) than the rate for Massachusetts adults, and this disparity accounted for 57.4% (95% CI = 55.2%, 59.5%) of the all-cause mortality rate difference between these groups. The proportion of all-cause mortality disparities explained by substance-attributable deaths ranged from 41.2% (95% CI = 39.8%, 42.4%) among 20- to 34-year-old women to 63.8% (95% CI =  61.2%, 65.9%) among 35- to 49-year-old men (Figure 3).

FIGURE 3—

Substance- and non–substance-attributable mortality rates, stratified by age and gender, for adults aged 20–64 years among BHCHP participants and Massachusetts (MA) adults: 2003–2008.

Note. BHCHP = Boston Health Care for the Homeless Program; DF = disparity fraction; RtD = rate difference; RtR = rate ratio. Mortality rate is expressed as deaths per 100 000 person-years for BHCHP adults and as annual deaths per 100 000 persons for MA adults. The mortality rate for all MA adults aged 20–64 years is standardized to the age and gender distribution of the BHCHP cohort. The lower 95% confidence bounds are greater than 1 for all rate ratios and greater than zero for all rate differences. The DF is the substance-attributable rate difference divided by the all-cause mortality rate difference.

DISCUSSION

In this large cohort of homeless adults in Boston, more than half of all deaths were attributable to tobacco, alcohol, or drug use. These deaths occurred at rates that far exceeded those in the Massachusetts general population. Overall, nearly 60% of the all-cause mortality disparity between BHCHP patients and Massachusetts adults was explained by substance-attributable deaths. These findings are intermediate to the disparate results of prior studies in Sweden²⁴ and Canada¹⁵ that used differing non-PAF approaches to estimate tobacco-, alcohol-, and drug-related mortality among homeless and marginally housed people. Our use of PAF-based methods offers the advantage of incorporating substance use prevalence estimates, risk relations, and other epidemiological data to generate a more nuanced assessment of the burden of tobacco-, alcohol-, and drug-attributable mortality in this vulnerable population.

The relationship between homelessness and substance use is complex and bidirectional.^70–73 The disproportionate burden of substance-attributable mortality among homeless people may reflect the adverse selection of people with substance use disorders into the condition of homelessness. Conversely, being homeless may trigger or exacerbate substance use. Regardless of the direction of this association, our results suggest that substance use interventions could have a major impact on the health of homeless people. Although the optimal approach is uncertain, several strategies have shown promise. A systematic review⁷⁴ concluded that case management services⁷⁵ and post-detoxification stabilization programs⁷⁶ appear more effective than usual care in reducing alcohol and drug use among homeless people with substance use disorders. Certain housing interventions may also improve substance use outcomes while concomitantly addressing homeless individuals’ housing needs. A meta-analysis⁷⁷ of 4 randomized controlled trials^78–81 of abstinence-contingent housing for cocaine-dependent homeless people found consistently superior drug abstinence rates in comparison with usual care. A “housing first” intervention for chronically homeless adults with severe alcohol use disorders reported reduced alcohol consumption among participants despite no programmatic requirement for abstinence.^82,83

Although addiction-related interventions for homeless people have historically focused on alcohol and illicit drug use, our findings support the need to address with equal vigor the less-studied issue of tobacco use, which accounted for about the same number of deaths as alcohol. To date, interventions for homeless smokers have reported modest cessation rates,^84–87 reinforcing the need for additional research on the management of this pervasive addiction.⁸⁸

Because substance-attributable deaths contribute disproportionately to the mortality gap between homeless and nonhomeless adults, addiction treatment might be viewed as a crucial component of efforts to reduce health disparities for homeless people. At the same time, the residual disparity in non–substance-attributable deaths suggests that the problem of excess mortality among homeless adults is not merely an addiction-related phenomenon. This underscores the need for comprehensive services to address the full spectrum of medical and behavioral health issues seen frequently among homeless people.⁸⁹ Increasing the reimbursement for behavioral health care and reducing the barriers to its integration with primary care services would be important steps toward achieving the dual aims of improving health in both addiction and nonaddiction domains of illness.

Limitations

Study participants were patients of a large HCH program in Boston, so the findings may not be generalizable to homeless people in other cities or to those who do not use such services. We were unable to capture the dynamic nature of participants’ homeless experiences and how this may have impacted their risk of substance-attributable death. The accuracy of death certificates has been questioned,⁹⁰ but they appear to have high specificity in identifying deaths caused by unintentional poisoning⁹¹ and lung cancer,⁹² which were heavy contributors to substance-attributable mortality in this cohort.

Our PAF estimation methods require several assumptions. We assumed that substance use prevalence estimates from a 2003 national survey of HCH users were a valid approximation of prevalence estimates for the BHCHP cohort across all study years. As a result of a paucity of prospective epidemiological data on the health of homeless people, we assumed that RR estimates and case series data derived from the general population for substance-related deaths were valid in homeless individuals. The use of general population data likely imparts a conservative bias to PAF estimates derived using the case series method, given the generally higher prevalence of substance use disorders among homeless individuals.

Conclusions

Tobacco, alcohol, and drug use accounted for more than half of all deaths in this cohort of homeless adults but did not fully explain the mortality disparity with the general population. Efforts to improve the health of homeless individuals should focus on the expansion of treatment services for substance use disorders in conjunction with comprehensive health care and social policy interventions to address nonaddiction sources of excess mortality.