Human Health Benefits from Fish Consumption vs. Risks from Inhalation Exposures Associated with Contaminated Sediment Remediation: Dredging of the Hudson River

Background: Billions of dollars are spent on environmental dredging (ED) to remediate contaminated sediments, with one goal being reduced human health risks. However, ED may increase health risks in unanticipated ways, thus potentially reducing net benefits. Objectives: To assess the ways that ED may increase health risks in unanticipated ways, thus potentially reducing net benefits, we quantitatively assessed a subset of population health benefits and risks of ED, using the 2009–2015 remediation of the Hudson River Polychlorinated Biphenyls (PCBs) Superfund Site as a case study. Three remediation scenarios were evaluated: No Action (NA), Source Control (SC), and ED. Methods: We quantified health benefits for each scenario from reduced PCB levels in Hudson River fish, and health risks from ED operations due to increased inhalation exposures to PCBs and fine particulate matter (PM2.5), using disability-adjusted life years (DALYs) as a common metric. Occupational health risks were also considered in a separate sensitivity analysis. Estimates of population-level benefits and risks included Monte Carlo simulation-based uncertainty analysis. Results: Under NA, fish consumption would result in an estimated health burden of 112 DALYs, and ED would lead to a reduction of 15 DALYs in excess of SC. ED operations were estimated to induce a total burden of 33 DALYs, dominated by PM2.5 impacts from rail transport emissions (32 DALYs). Including uncertainty, the net health benefit of ED ranged from –138 to +1,326 avoided DALYs (90% confidence), with a median of –11 avoided DALYs. Conclusions: For the considered impacts, ED in the Hudson River might not have led to an overall net positive human health impact. The benefits and risks of ED, however, have different degrees of uncertainty and involve different populations. Reducing long-distance transport of dredged sediment is a priority. This comparative approach could be used prospectively to better determine trade-offs involved in different remediation scenarios and to improve remediation design to maximize benefits while minimizing risks. https://doi.org/10.1289/EHP5034


Introduction
Many surface waters throughout the United States have sediments contaminated with legacy pollutants from former industrial, mining, and/or agricultural activities (U.S. EPA 2004a). Some of these chemicals can bioaccumulate in aquatic organisms, including fish consumed by local anglers and their families. Depending on the chemical, its concentration and bioaccumulation potential, and the frequency and duration of exposure, these individuals can be at risk of developing adverse health outcomes. Thus, the potential negative effects of contaminated sediments on human health are compelling reasons to consider remediating them (NRC 2007).
Removal of sediments by dredging is one commonly used, yet costly, method of remediation. It is unclear, however, to what extent this remedy is directly responsible for reduced human health risk, with a primary reason being shortcomings in available monitoring data (NRC 2007). There is also a need to comprehensively investigate the trade-offs of dredging in terms of health impacts beyond those associated with the reduced sediment concentrations that typically drive these assessments. For instance, health benefits from reduced sediment concentrations may be partly offset by resuspension of sediment and release of contaminants into the water column during dredging (Bridges et al. 2010). Additionally, environmental dredging (ED) is a highly complex industrial operation that itself involves potential health risks. For example, dredging and transportation equipment can lead to emission and/or formation of considerable amounts of primary and secondary fine particulate matter (PM 2:5 ) as well as nitrogen oxides (NO x ) (Anderson 2008;U.S. EPA 2004b). Ambient PM 2:5 is an important contributor to cardiovascular and respiratory diseases, including lung cancer (Brook et al. 2010;IARC 2013;Lepeule et al. 2012;Pope and Dockery 2006). Dredging projects may also expose workers to chemical and physical hazards. In sum, appropriate risk assessments of any remediation strategy, especially dredging, should consider and compare respective risks and benefits.
The Hudson River Polychlorinated Biphenyls (PCBs) Superfund Site (the Site) provides a promising case study for evaluating trade-offs between health risks and benefits of ED. Approximately 200 river-miles long, the Site consists of the Upper Hudson (approximately 40 river-miles between Hudson Falls and the Federal Dam at Troy, New York) and the Lower Hudson (approximately 160 river-miles between the Federal Dam at Troy and Battery Park). From 1947 to 1977, two General Electric capacitor-manufacturing facilities discharged an estimated 1.3 million pounds of PCBs into the Upper Hudson (U.S. EPA 2017). PCBs are carcinogenic to humans (IARC 2016), and consistent positive associations have been observed in epidemiological studies for a variety of noncancer adverse health outcomes (Faroon and Olson 2000;Faroon and Ruiz 2016). Between 2009 and 2015, with an off-season in 2010, a dredging remedy costing $1:7 billion USD (Haggard 2017) removed approximately 2.65 million cubic yards of PCB-contaminated sediment from the Upper Hudson (U.S. EPA 2017). Extensive data sets of PCB measurements in fish, sediment, water, and air are available (U.S. EPA 2017), complemented by mechanistic modeling of PCB concentrations in various media (TAMS Consultants et al. 2000a, 2000b. Furthermore, rigorous documentation and monitoring has generated a wealth of relevant information. Focusing on the Site, this study sought to extensively evaluate and compare the potentially avoided and induced population health burden of this major ED project. We first evaluated overall human health impacts associated with fish consumption under three scenarios: No Action (NA), Source Control (SC), and the selected remedy combining SC with ED (SC&ED). The details of these scenarios are summarized in Table 1. In each case, we estimated the potential cancer and noncancer health burden on recreational anglers and their families due to consumption of PCB-contaminated fish. These risks constituted the basis for response action for the Site (U.S. EPA 2002a). For ED, we also estimated the health burden induced by this operation from increased air emissions of PCBs and from primary and secondary PM 2:5 . To enable comparison of benefits and risks involving different health effects, we expressed health impacts in terms of disability-adjusted life years (DALYs), representing the number of healthy years of life lost in a population accounting for premature death and disability (Forouzanfar et al. 2016). Finally, we conducted a separate sensitivity analysis estimating the potential health burden of ED on project workers due to inhalation exposures and fatal incidents (physical hazards).

General Framework
The general framework of our analysis assesses the average damage and associated long-term net health benefit under each scenario as the sum of several individual components. First, the induced health burden from consumption of Hudson River fish (IB fish,PCB ) was calculated under each scenario. In this paper, "fish" represents both fish and shellfish/crabs. Then, the "avoided burden" of SC (AB SC ) was represented by the difference between the induced burdens under SC and NA. Similarly, the avoided burden of ED was represented by the difference between the induced burdens under SC&ED and SC: AB ED = IB fish,PCB,SC − IB fish,PCB,SC&ED The benefits of ED from reduced PCB levels in fish were then offset by the burden induced by the ED operation (IB ED ). For this, we focused on risks to surrounding communities due to increased air emissions of PCBs (IB air,PCB,communities ) and on risks to the regional and U.S. populations (including surrounding communities) due to air emissions of primary and secondary PM 2:5 from diesel-powered equipment (IB PM 2:5 ,general population ). The total induced health burden was then calculated as follows: IB ED = IB air,PCB,communities + IB PM 2:5 ,general population The net health benefit of ED was represented by the difference between avoided and induced burdens: Equations 3 and 4 assume that the induced burden of ED on project workers is negligible; i.e., their counterfactual scenario in the absence of dredging would be to work in jobs with similar risk. As a separate sensitivity analysis, we conducted a fully attributional assessment that quantified potential health risks of ED to project workers due to increased inhalation of PCBs and PM 2:5 (IB air,PCB,workers and IB PM 2:5 ,workers ), and fatal occupational incidents (IB fatal incidents,workers ), yielding an upper-bound induced health burden: IB ED,sensitivity = IB air,PCB,communities + IB PM 2:5 ,general population + IB air,PCB,workers + IB PM 2:5 ,workers + IB fatal incidents,workers The effective marginal induced health burden of ED would likely fall within these two bounds (i.e., between IB ED and IB ED,sensitivity ).
The methods for calculating each of these components are further described below. Table 2 and Table S1 summarize the primary model input parameters for all considered health impact pathways. Assuming mutual independence among these parameters, we accounted for both variability and uncertainty using Monte Carlo simulation implemented in R (version 3.4; R Development Core Team 2017). For some parameters, we adapted an approach from MacLeod et al. (2002), assuming log-normal probability distributions. Table 3 summarizes these  distributions, and the Supplemental Material (Tables S2-S4) summarizes the remaining input data used in the uncertainty analyses.
where CðtÞ fish,PCB represents a time-dependent, species-and river-section weighted average, wet-weight, fish tissue Tri + PCB concentration (kg PCB =kg fish ), adjusted for cooking losses. "Tri + " refers to the sum of trichloro-through decachloro-PCB homologs. IR p is an annual average individual fish ingestion rate from the Site for subpopulation (p) (kg fish =person-year).
Specifically, we defined six subpopulations, three consuming fish from the Upper Hudson at annualized ingestion rates of twice per year, twice per month, and twice per week, and similarly for the Lower Hudson. DR ADD p,fish PCB ,e is a dose-response factor for oral intake of PCBs (cases per kg PCB intake ) for health effect (e). SF PCB,e is the corresponding severity factor converting cases of effect (e) into DALYs (DALY per case). N p is the annual number of consumers.

Exposure Assessment
For  Six-year averaging was based on the estimated exposure duration for an average adult angler, consistent with the prior human health risk assessment for the Site. We also applied the same central tendency PCB cooking loss factor of 20% used in that prior risk assessment (TAMS Consultants and Gradient Corporation 2000).
In our uncertainty analyses, we considered additional FISHRAND model forecasts. Because the central estimate forecasts under NA and SC are highly uncertain and may overestimate the rate of decline of PCBs, we derived geometric standard deviations (GSDs) for CðtÞ fish,PCB , based on the ratios of estimated upper bound and central estimate forecasts (Table   S4). Similarly, for the SC&ED scenario, we derived GSDs based on the forecasts under the "REM (6-y 2.5% resuspension)" scenario as an equivalent estimated upper bound.
To evaluate population-level fish consumption from the Site, we used data from several site-specific surveys. We first used data on fishing effort from the 2007 New York Statewide Angler Survey [provided by N. Connelly (personal communication); Connelly and Brown 2009], specific to counties surrounding the Site. In particular, we estimated the annual average number of anglers who fish from the Site, both for the Upper and Lower Hudson. We then coupled these estimates with original data on consumption habits of anglers and their families from two surveys administered by the New York State Department of Health (U.S. EPA 2017). These calculations accounted for the proportions of anglers who reported consuming fish from the Site, and those who also reported sharing fish with their families, at different annualized ingestion rates (twice per year, twice per month, and twice per week). We assumed an average household size of three persons (U.S. Census Bureau 2017). Table S5 summarizes these results. To address limitations of these surveys, we corrected estimates of N p , such that corresponding estimates of total populationlevel consumption (kg fish =y) matched those calculated using data from two more comprehensive, site-specific creel surveys Table 3. Uncertainty analysis input data for estimating the health burden induced by increased air emissions of PCBs, and primary and secondary PM 2:5 . Note: Presented values include geometric standard deviations (GSDs), sensitivity coefficients, and percent contributions to total output uncertainty. For these input parameters, we adapted an approach from MacLeod et al. (2002), assuming independent lognormal probability distributions (e.g., see Slob 1994). Following their approach, variance in each model output (cumulative health burden, DALYs) was calculated as a weighted sum of variances contributed by each input parameter with sensitivities as weights. Sensitivities were based on a 10% change in each input parameter relative to the total induced health burden in DALYs.
a Based on the average variability within dredging seasons in the site-specific ambient air PCB monitoring results used for this study (Anchor QEA and Environmental Standards, Inc. 2009;Environment 2004, 2017).
b Accounts for uncertainty in the extrapolation of rodent data to humans as calculated by Huijbregts et al. (2005). d Based on the greatest 95th uncertainty interval for the corresponding DALY and incidence data as calculated by the Institute for Health Metrics and Evaluation (IHME 2017a). This assumes that the relative fractions of incidence for the three cancer types in these exposed populations are similar to those for the greater U.S. population (age-and sex-adjusted). Assuming these fractions are unknown would result in a maximum GSD 2 of 1.3 for this parameter. This would have a negligible (1%) effect on the total uncertainty in cancer health risk, since this uncertainty is driven by uncertainty in the interspecies conversion factor. e Total uncertainty is displayed on Figure S2. Separate uncertainty distribution was applied in allometric scaling by body weight, accounting for chemical-specific interspecies differences. Interindividual variability was addressed by assuming a lognormal distribution for human variation, with an additional uncertainty distribution for the GSD of human variation. No subchronic uncertainty factor was applied, because the duration of the study by Tryphonas et al. (1991) was 55 months.
f Based on Huijbregts et al. (2005) with considerably greater uncertainty than for cancer arising from use of an average severity factor, in DALY per case, across 49 diverse, noncommunicable diseases.
g Accounts for uncertainty in the use of emission factors from Cao et al. (2016), based on variability across equipment types deemed to be most representative for this study. Furthermore, a separate uncertainty distribution was programmed in the Monte Carlo simulation to assign equal likelihood of Tiers 3 and 4 equipment. k Uncertainty distribution as calculated by Gronlund et al. (2015).
l Based on the variability of exposure levels reported by Lewné et al. (2007) for "construction machine operators" and "other outdoor workers exposed to diesel exhaust." Worker impacts were considered as part of a separate sensitivity analysis.
(Normandeau Associates 2003Associates , 2007. Table S6 summarizes these population-level estimates. Furthermore, we characterized the uncertainty of IR p assuming uniform distributions ranging from one to three meal-per-week equivalents. To convert meals into mass equivalents, we assumed a serving size of 0:5 lbð227 gÞ=meal, consistent with the prior human health risk assessment (TAMS Consultants and Gradient Corporation 2000). Tables S7 and S8 further summarize these input data.

Dose-Response and Severity
The term DR ADDp,fish PCB ,e × SF PCB,e in Equation 5 includes two general health effects: cancer and noncancer. In both cases, DR ADD p,fish PCB ,e is a function of average daily dose (ADD), ADD p,PCB = CðtÞ fish,PCB × IR p =ðBW × DtÞ, with body weight BW (kg) and Dt = 365 d=y. We derived the cancer DR ADD p,fish PCB ,e from the central estimate "High Risk and Persistence" Cancer Slope Factor for PCBs from the Integrated Risk Information System (IRIS) (U.S. EPA 1996). This estimate was consistent with the prior human health risk assessment (TAMS Consultants and Gradient Corporation 2000). We converted the slope factor into a human-equivalent lifetime dose, the inverse of which yielded a linear DR ADDp,fish PCB,e accounting for PCB carcinogenicity. We considered three cancer types based on conclusions of the International Agency for Research on Cancer (IARC 2016). These types included malignant melanoma, non-Hodgkin's lymphoma, and breast cancer. Then, using data from the 2015 Global Burden of Disease Study for these cancer types (IHME 2017a), we calculated an incidence-weighted average SF PCB,e for the U.S. population, adjusted for age and sex.
For noncancer, we adapted recent work by the World Health Organization's International Programme on Chemical Safety to probabilistically incorporate a nonlinear dose-response relationship (Chiu et al. 2018;Chiu and Slob 2015;WHO 2014). This approach enabled us to predict the human population dose-response relationship based on experimental animal data, incorporating a probabilistic characterization of uncertainty in extrapolating from animals to humans and in the degree of variability in the human population.
We based our prediction on the critical study by Tryphonas et al. (1991) that forms the basis of IRIS's current reference dose for Aroclor 1254 (U.S. EPA 1994). The U.S. Environmental Protection Agency (U.S. EPA) determined that Aroclor 1254 ( ∼ 54% chlorine by weight) was the commercial PCB mixture that most resembled the homolog distribution in Hudson River fish (TAMS Consultants and Gradient Corporation 2000). Using data from Tryphonas et al. (1991) (summarized in Table S9), we conducted benchmark dose modeling of decreases in immunoglobulin M. This is a sensitive end point, selected based on high statistical significance and relatively large effects reported in the study. We used the web portal benchmarkdose.org , setting the benchmark response at 50%, a value similar to the lowest observed adverse effect-level for multiple effects reported in the study (Tryphonas et al. 1991). To account for model uncertainty, we used Bayesian model averaging (three Markov Chains for eight different models) as described by Shao and Shapiro (2018). The remaining probabilistic extrapolations were conducted following approaches described by Chiu and Slob (2015) and Chiu et al. (2018). Figure 1 displays the nonlinear dose-response relationship for noncancer PCB effects, the slope of which is equal to DR ADD p,fish PCB ,e ( Figure S2). Also shown are 95% confidence intervals (CI) of average daily doses (mg/kg-d), which are functions of CðtÞ fish,PCB and IR p , under the No Action scenario (2004)(2005)(2006)(2007)(2008)(2009), for the three Upper Hudson subpopulations. These subpopulations were predicted to have had the highest exposures because they were consuming fish from the Upper Hudson, for which average predicted PCB concentrations were about a factor of two higher than those for the Lower Hudson (Table 2) median was therefore used as the measure of central tendency. Median estimates of noncancer DR ADDp,fish PCB ,e ranged several orders of magnitude between the twice-per-year (1 × 10 −4 cases= kg PCB intake ) and twice-per-week (690 cases=kg PCB intake ) consumer subpopulations. In comparison, the median cancer DR ADDp,fish PCB,e was 0:6 cases=kg PCB intake for all subpopulations, assuming linearity (Table 2). Thus, noncancer risk could greatly surpass that of cancer risk for highly exposed individuals. Because the specific noncancer end points anticipated are not as clear as they are for cancer (Faroon and Olson 2000;Faroon and Ruiz 2016), for the noncancer SF PCB,e we used an incidenceweighted average of 2:7 DALYs per case, based on the work of Huijbregts et al. (2005). This incidence-weighted average accounts for 49 diverse, noncommunicable diseases.
We estimated the induced burden of increased air emissions of PCBs on surrounding-community subpopulations (p) across all dredging seasons (s) using the following equation: IB air,PCB,p = X s C air,PCB,s × BR p × D p,s × DR ADD p,air PCB ,e × SF PCB,e where C air,PCB,s represents an above-baseline, dredging-season average, ambient air concentration of total PCBs (kg PCB =m 3 ); BR p is the average individual breathing rate (m 3 =d); D p,s is a spatially differentiated, cumulative exposure duration (person-d).
The dose-response (DR ADDp,air PCB,e ) and severity factor (SF PCB,e ) have the same interpretations as in Equation 5.

Exposure Assessment
C air,PCB,s is based on ambient air data from a comprehensive monitoring program (Anchor QEA and Environmental Standards, Inc. 2009;Environment 2004, 2017). Briefly, General Electric employed 24-h samplers that continuously monitored airborne PCBs at the processing facility and along the dredging corridor. We stratified these data by dredging season and location (dredging corridor and processing facility), after excluding background data and treating nondetects as one-half the method detection limit. We then applied arithmetic mean total PCB concentrations for each stratum as C air,PCB,s . To estimate the health burden on surrounding communities induced by the remediation (i.e., above the baseline risk), we used air PCB data from the Predredging, Background Monitoring Results (2005)(2006) reported by . For this calculation, we set BR communities to 0:66 m 3 =h or 16 m 3 =d, an average long-term daily inhalation rate for the general population (U.S. EPA 2011).
To derive D p,s for the surrounding communities, we used census-based population count data (CIESIN and Columbia University 2016) and a geographic information system of the Site [provided by M. Cheplowitz (personal communication)]. Specifically, we estimation the number of individuals residing within 500 m of the dredging for each dredging season. This was the distance from the Site at which the median ambient air PCB concentration reached a level near background ( Figure S3). Population count estimates differed for each dredging season (Table S10) because dredging progressed in a general north-tosouth fashion (U.S. EPA 2017), thus affecting different community subpopulations.

Dose-Response and Severity
We evaluated cancer-related health impacts associated with PCB inhalation using a DR ADDp,air PCB,e derived from the central estimate "Low Risk and Persistence" Cancer Slope Factor for PCBs (U.S. EPA 1996). For noncancer, we assumed the same nonlinear doseresponse relationship for inhalation exposures as the doseresponse relationship for ingestion, due to a lack of dose-response data for this exposure pathway. SF PCB,e is also the same as that for fish consumption.

Health Burden of Primary and Secondary PM 2:5 Emissions
Although diesel exhaust is a complex mixture of organic and inorganic chemicals in gas and particulate phases (Harbison et al. 2015), we focused our assessment on the impacts of primary and secondary PM 2:5 for three primary reasons: a) reducing ambient air concentrations of PM 2:5 has been a fundamental aim of national emission-control legislation for nonroad vehicles in the United States (U.S. EPA 2004b); b) the particulate fraction of diesel exhaust has been considered to be the "risk driver" of its health effects (Hesterberg et al. 2011); and c) available epidemiological data for ambient PM 2:5 are much more reliable for quantitative risk assessment than those for the preferred surrogate of exposure to diesel exhaust, elemental carbon (Möhner and Wendt 2017;Morfeld and Spallek 2015;Pronk et al. 2009). For the regional and U.S. populations, we used the following equation for all PM 2:5 precursors (i) and emission source categories (j): where M i,j represents the total emitted precursor mass (kg i emitted ); iF i is the spatially differentiated PM 2:5 intake fraction for each considered precursor (kg PM 2:5 intake =kg i emitted ); DR PM2:5 is a doseresponse factor for PM 2:5 (deaths per kg PM 2:5 intake ); and SF PM 2:5 is the corresponding severity factor (DALYs per death).

Exposure Assessment
We derived M i,j for primary PM 2:5 and NO x (i) using an emission-factor based approach for three source categories (j): nonroad diesel-powered heavy equipment, barge traffic, and rail transport (line-haul operation). Tables S11-S13 summarize the input parameterization. For heavy equipment, we obtained data on the number and types of equipment from an estimated project inventory of primary diesel-powered, nonroad heavy equipment [provided by M. Cheplowitz (personal communication)] and data on equipment specifications from manufacturers and distributors. For each equipment type, we then estimated the total number of hours of effective operation and idling from data in Weekly Productivity Summaries. For similar equipment types, we applied in-use load factors, a measure of how hard an engine is working, and emission factors from a recent study (Cao et al. 2016). Because emission control technology can substantially reduce emissions (Clark et al. 2002;Khalek et al. 2009Khalek et al. , 2013, and because the project began around the time when the U.S. EPA phased in the latest emission control standard, Tier 4 (U.S. EPA 2004b), we probabilistically assessed emissions and impacts for Tier 3 (U.S. EPA 1998) and Tier 4 heavy-equipment scenarios, assuming equal likelihoods. Unlike for heavy equipment, we did not have detailed data on equipment characteristics and operating times for barge traffic. Accordingly, we applied emission factors for this source category in units of mass of emitted pollutant per (short) ton-mile of transported sediment and water (g per ton-mile). We obtained riverbarge emission factors from the U.S. EPA SmartWay Carrier Performance database (U.S. EPA 2016). For the 92,590 reported total barge miles traveled (Louis Berger Group et al. 2017), we assumed equal distances to and from the processing facility. For trips to the facility, we calculated an average mass of transported load per barge based on the reported tonnage of material shipped off-site, the total volume of water treated and returned to the river, and the total number of unloaded barges for each dredging season. To account for unloaded barge returns, we scaled the emission factors proportional to an estimated weight reduction. For this, we obtained gross tonnage data from a local distributor (Sterling Equipment, Inc.) for equipment that reasonably matched project photographs and specifications in the Phase 1 Evaluation Report (Louis Berger Group 2010) and Remedial Action Work Plans.
For rail transport, we applied weighted-average emission factors (g per ton-mile) representing U.S. Class 1 railroad companies (2011)(2012)(2013)(2014)(2015). These data were also from the EPA SmartWay Carrier Performance database (U.S. EPA 2016). We calculated the mass of sediment delivered to each of seven reported hazardous waste landfills, using project manifest data for all remediation seasons and reported trips. We then estimated the distance traveled to and from each landfill as the shortest path, using the Network Analyst tool in Esri ArcMap (version 10.5) on the U.S. Railroad Lines data set (U.S. Department of Transportation 2017). This estimation resulted in seven distinct railroad routes. We accounted for each unloaded return as we did for barge traffic, using project data on tare weights of empty project gondolas, the To link emissions of PM 2:5 and NO x with corresponding intakes of primary and secondary PM 2:5 by the regional and U.S. populations, we derived spatially differentiated estimates of iF i for both precursors. iF i represents the incremental PM 2:5 mass intake, summed over all exposed individuals over time, per unit mass of emitted precursor (kg PM 2:5 intake =kg i emitted ) (Bennett et al. 2002). For the Site and for each of the seven routes, we derived estimates of iF i ( Table 2 and Table S13) by coupling 2015 population density data for North America (CIESIN and Columbia University 2016) with marginal increases in ambient PM 2:5 concentrations associated with ground-level emissions. These concentrations were simulated by a mechanistic air pollution model for the greater region of North America, known as Intervention Model for Air Pollution version 1.2.0 (Tessum et al. 2017). We assumed the same average breathing rate of 16 m 3 =d as for the PCB-inhalation exposure pathway (U.S. EPA 2011).

Dose-Response and Severity
To quantify the health burden associated with population intakes of primary and secondary PM 2:5 , we applied a combined DR PM2:5 and SF PM 2:5 of 78 DALY=kg PM 2:5 intake from Gronlund et al. (2015), accounting for cardiopulmonary and lung-cancer mortality.

Sensitivity Analysis: Health Risks to Project Workers
To assess the potential health burden of ED on project workers, we considered two worker subpopulations to account for their estimated differences in exposure. These subpopulations included processing facility workers and dredging corridor workers.

Health Burden of Inhalation of PCBs and PM 2:5
The general framework for estimating occupational health risks from PCB inhalation is the same as for surrounding communities (Equation 6). For this estimation, we set BR workers to 1:6 m 3 =h, an average hourly inhalation rate for outdoor workers (U.S. EPA 2011). For facility workers, we derived D p,s from projections in Remedial Action Work Plans and from operation dates reported in Weekly Productivity Summaries. For dredging corridor workers, we applied a similar method using the estimated project inventory supplied by M. Cheplowitz mentioned previously. Tables S10 and S14 summarize these estimates.
To quantify occupational health risks due to PM 2:5 inhalation from the use of diesel-powered heavy equipment, we used the following equation for all worker subpopulations (p) and dredging seasons (s): IB PM2:5,workers = X p,s C PM 2:5 ,p × BR p × D p,s × DR PM 2:5 × SF PM 2:5 (8) where C PM 2:5 ,p represents an above-background, personal exposure concentration of diesel PM 2:5 during a work shift (kg=m 3 ); individual breathing rate (BR p ) and cumulative exposure duration (D p,s ) are the same as for the PCB-inhalation exposure pathway; DR PM 2:5 and SF PM 2:5 have the same meanings as for the regional and U.S. populations.
We obtained estimates of C PM 2:5 ,p from Lewné et al. (2007) for a) operators of nonroad heavy equipment and b) support workers frequently in close proximity to heavy equipment. As previously described, we probabilistically assessed emissions and impacts for Tier 3 and Tier 4 heavy equipment scenarios, assuming equal likelihoods. To assess the increased risk attributable to the remediation, we subtracted from each C PM 2:5 ,p an estimated regional background concentration of 8:3 lg=m 3 for Albany/ Schenectady, New York (Shaddick et al. 2018).

Health Burden of Fatal Occupational Incidents
To estimate the health burden of fatal occupational incidents, we used the following equation for all general labor categories (c) and dredging seasons ( To derive PF c , we adapted an approach used by the U.S. Bureau of Labor Statistics (U.S. BLS) for their Census of Fatal Occupational Injuries. Each year, U.S. BLS publishes national fatal injury rates using the Standard Occupational Classification (SOC) system (U.S. BLS 1997). We matched project general labor categories to detailed SOC occupations based on work descriptions in Remedial Action Work Plans. We then used data on fatal occupational injury counts and the number of at-work employees for years 2011-2015, as provided by J. Kang (personal communication), calculating for each general labor category an average fatal injury rate (PF c ) for one full-time equivalent worker (2,000 h/y).
To estimate N c,s , we used projections from the Remedial Action Work Plans, converting these to a full-time equivalent basis. We then estimated LE c based on median ages of at-work employees in matched SOC categories as provided by J. Kang (personal communication) using the Global Burden of Disease 2015 Reference Life Table (IHME 2017b). Table S2 presents estimates of PF c , N c,s and LE c for each general labor category and the corresponding SOC categories.  1, 2,618). For all three scenarios, noncancer risk for Upper Hudson anglers and family members comprised most of the total burden (62%-75%), due to a relatively small subpopulation of consumers with frequent consumption (twice per week). The rest of the total burden was attributed to noncancer risk for Lower Hudson anglers and family (15%-21%), cancer risk for Lower Hudson anglers and family (7%-14%), and cancer risk for Upper Hudson anglers and family (2%-3%).

Health Burdens from PCBs in Fish
The first bar in Figure 3 shows the avoided burden due to ED (i.e., SC&ED over SC), including the 90% CI. As shown, ED resulted in 15 avoided DALYs (90% CI: 0, 1,376) of health benefit due to long-term reductions in fish tissue PCB concentrations. Figure 3 summarizes estimates of induced burdens from ED on the surrounding communities and regional and U.S. populations associated with increased air emissions of PCBs and primary and secondary PM 2:5 . Results for each impact are described later in this paper.

Inhalation of PCBs
Using Equation 6 with parameters summarized in Table 2 yielded an induced burden of 2 ×10 −3 DALYs (90% CI: 1 ×10 −4 , 2 ×10 −2 ) for the surrounding communities attributed to increased PCB emissions to air during the dredging (Figure 3). From Table 2,  Table S10, and Figure S3, it is evident that annual averages of total airborne PCB concentrations measured during the remediation were well below the U.S. EPA's level of concern of 110 ng=m 3 . This finding suggests the applied best management practices were successful overall in limiting increases in air PCB concentrations. This finding was also noted during the second Five Year Review for the Site . For the surrounding communities, estimates of inhalation exposures are likely conservative, given that the underlying data represent nearest-receptor locations. In any case, the burden induced by increased air emissions of PCBs during the ED operation appears to have been minor in comparison with other induced burdens and to the potential health benefits of long-term reductions in fish tissue PCB concentrations.

Inhalation of PM 2:5
As summarized in Table 2 (with more detailed results in Table  S11), total emissions of primary PM 2:5 and NO x from project heavy equipment under the Tier 3 scenario amounted to 3,300 kg and 71,000 kg, respectively. These yielded, in combination with values from Table 2, a total PM 2:5 intake of 0:007 kg across the regional and U.S. populations. Relative contributions of primary PM 2:5 and NO x to this total were 28% and 72%, respectively. When assuming better emission-control technology under the Tier 4 standard, emissions of PM 2:5 and NO x from heavy equipment were considerably reduced to 100 kg and 30,000 kg, respectively. This assumption led to a PM 2:5 intake of 0:002 kg across the regional and U.S. populations, with a much lower relative contribution of 3% from primary PM 2:5 vs. 97% from NO x . Assuming equal likelihood for both heavy equipment scenarios resulted in a median health burden of 0.3 DALYs (90% CI: 0.05, 2) for these populations. Barge traffic emissions (Table 2 and  Table S12) resulted in an additional cumulative PM 2:5 intake of 0:002 kg and a health burden of 0.1 DALYs (90% CI: 0.03, 1).
Site-specific emissions from project heavy equipment and barge traffic were limited in comparison with emissions from rail transport of sediment across the United States. Assuming project trains used the shortest path to and from each landfill, we estimated that they traveled a total of 1,021,255 railroad miles to ship the reported 3,165,820 tons of sediment (Table S13). As shown in Table 2, this estimation corresponded to 73,000 kg of primary PM 2:5 and 2,586,000 kg of NO x emitted by project locomotives in total. These emissions resulted in a substantial cumulative population PM 2:5 intake of 0:410 kg, over 60 times higher than the total intake from all other sources. The corresponding health burden was 32 DALYs (90% CI: 7, 145). The nearest landfill used for the project was CWM Chemical Services in Model City, New York. This landfill was an estimated 334 railroad miles from the Site, yet only two project railcars were shipped to this destination out of an estimated 30,000, while approximately 70% of project unit trains delivered waste to landfills over 1,500 railroad miles from the Site (Table S13). If all project trains delivered to the nearest landfill, the corresponding health burden would have been about a factor of 5 lower [7 DALYs (90% CI: 1, 32)]. Figure 4 summarizes all estimated adverse health impacts and health benefits of ED, including uncertainty evaluated via Monte Carlo simulation. As previously mentioned, ED achieved an estimated avoided burden of 15 DALYs (90% CI: 0, 1,376) ( Figure  4A). The large right skew in the distribution of total avoided burden reflects the large uncertainty associated with the noncancer dose-response for PCBs. On the other hand, potential health burdens induced by ED on surrounding communities and the regional and U.S. populations amounted to 33 DALYs (90% CI: 8, 146) ( Figure 4B). As shown in Figure 3, most of these adverse impacts were attributed to PM 2:5 from rail transport [32 DALYs (90% CI: 7, 145)]. Figure 4C shows the net health benefit of ED [i.e., the distribution of the difference between the avoided burden (benefit) and the induced burden (risk)]. The median of this distribution was −11 DALYs of benefit (90% CI: −138, + 1,326) or 11 DALYs of induced burden. Although the upper confidence bounds suggest the possibility of larger benefits, only 39% of these Monte Carlo samples corresponded to a positive net health benefit for the ED operation.

Sensitivity Analysis: Health Burdens for Workers
The net health benefit of the ED operation would be further diminished if it posed increased risks to project workers. Equations 6 and 8 with parameter values summarized in Table 2 yielded induced burdens of 1 × 10 −4 DALYs (90% CI: 9 × 10 −6 , 1 × 10 −3 ) and 0.4 DALYs (90% CI: 0.03, 5) on workers due to inhalation of PCBs and PM 2:5 , respectively. As shown in Figure 3, these burdens were of similar orders of magnitude as their general-population counterparts, though they may be reduced if effective personal protective equipment were used. Worker impacts from barges and rail transport were not assessed, due to a lack of available exposure data.
According to Equation 9, the health burden of fatal occupational incidents was 14 DALYs (90% CI: 0, 46), which is considerably higher than the other considered worker impacts. Because the estimated median life expectancy of this workforce is 47 y (Table S2; IHME 2017b), this central tendency health burden corresponds approximately to a 30% chance (14=47 × 100%) of a fatal occupational incident occurring for this project. Including workers in the overall assessment would result in a net health benefit of −20 DALYs (90% CI: −163, + 1,312) or 20 DALYs of induced burden for the ED operation ( Figure S4).

Strengths and Limitations
In this study, we conducted a comparative human health risk assessment to evaluate trade-offs in health benefits and risks accompanying one of the largest remediation projects in U.S. history. Use of DALYs as a common metric facilitated extensive accounting of population-level effects beyond local communities. Other strengths of our approach include: a) use of site-specific data in all parts the analysis; b) incorporation of noncancer health benefits from reduced exposure; and c) detailed application of uncertainty analysis. Considering both avoided and induced health burdens in our analysis suggests that the dredging performed at the Hudson River PCB Superfund Site might not have led to an overall net positive human health impact in excess of source control. Health risks primarily resulting from long-distance transport of sediment appear to negate the benefits of dredging. For the considered impacts, the results suggest there is <50% likelihood that the long-term population health benefits of the dredging remedy would exceed health risks to surrounding communities and the regional and U.S. populations (median in Figure 4C).
The difference between this study's and the U.S. EPA's results at the time of remedy selection (U.S. EPA 2002a) relates primarily to differences in the scope of the health risk assessments. Both approaches, however, have the potential to inform future decision-making in a complementary way.
• First, the U.S. EPA's assessment focused on individual health risk to community members, whereas this study assessed long-term, population health risks to communities and the regional and U.S. populations. Our assessment included risks due to PM 2:5 precursor emissions from longdistance transport of sediment. The individual perspective is useful for identifying and protecting individuals at high risk. Our analysis of survey data from the New York State Department of Health suggested that a small percentage of Upper Hudson anglers may be at high risk of noncancer health effects from frequent fish consumption. In the absence of active remediation, it would thus be imperative to maintain institutional controls that target these individuals. The population perspective, on the other hand, is useful for comparing potential benefits and risks associated with such a large remediation project. • Second, in contrast to the U.S. EPA's assessment, this study took a cumulative temporal perspective rather than focusing on the predicted years at which fish tissue PCB concentrations would reach levels deemed acceptable for individuals. Like the contrast between individual-level vs. populationlevel risk, both perspectives may be informative in future decision-making. This case study also has several important limitations, mostly related to large uncertainties in both the data and analysis. One of the largest uncertainties was in estimating noncancer health benefits of long-term reductions in fish tissue PCB concentrations, primarily due to uncertainties in the noncancer dose-response relationship spanning more than three orders of magnitude. This uncertainty is driven largely by the use of experimental animal data in deriving the predicted human population dose-response. IRIS is currently reassessing the noncancer human health hazards of PCBs and corresponding dose-response information. It is possible that these uncertainties can be reduced based on updated toxicity data and dose-response analyses.
Another significant uncertainty in estimating health benefits of ED was in quantifying PCB exposures from fish consumption. For instance, the second Five Year Review discusses several important differences between the FISHRAND model assumptions and the remediation as implemented ). In addition, our approach for quantifying fish consumption was limited to combining data from several data sources, deemed to be the best available, each with their own limitations (Tables Þ = total health burden of ED on the regional and U.S. populations (including surrounding communities) from increased air emissions of PCBs, and primary and secondary PM 2:5 ; C) Net Avoided Health Burden ðNet health benefit ED Þ = AB ED − IB ED . Dotted or dashed vertical lines correspond to the fifth, 10th, 25th, 50th, 75th, and 90th percentiles when read from left to right. The solid (red) vertical line through zero denotes a net of 0 avoided DALYs (i.e., benefits = risks). Values to the left of this line represent net risks while values to the right of this line represent net benefits. S7 and S8). Nonetheless, our estimates of limited consumption from the Upper Hudson are generally consistent with the only measurement-based assessment of PCB exposures from fish consumption conducted for the Site. Specifically, Fitzgerald et al. (2007) reported no significant differences in measured total serum PCB concentrations between older residents of Upper Hudson communities and an upstream control population. Both populations had low rates of reported fish consumption and PCB body burdens similar to populations with no unusual exposures (Fitzgerald et al. 2007).
Total uncertainty in the induced health burden of ED was driven by only two parameters: the intake fractions for PM 2:5 (mostly, for secondary PM 2:5 from NO x ) used to estimate impacts from train emissions (68%), and the corresponding dose-response factor (21%) ( Table 3). Nonetheless, there were several sources of uncertainty beyond our means to quantitatively evaluate. With respect to PM 2:5 , we used a linear dose-response factor because the nonlinear dose-response relationships are very uncertain in regions with relatively low background concentrations.
Our assessment also relied on national statistics to assess the health burden of fatal occupational incidents. The specific jobs for this remediation project may have been more or less hazardous than the more general categories to which we matched (Leigh 2011). These limitations, coupled with the possibility of similar risks occurring for workers in the absence of dredging, prompted us to first conduct the assessment without considering potential occupational health risks and then consider them separately in a sensitivity analysis. Our estimate of a 30% chance of a fatal occupational incident is consistent with the actual occurrence of one during the remediation. Specifically, a 39-y-old worker who had been involved in a support activity aimed at protecting cultural resources drowned after his boat went over a dam (Nelson 2009). This occurrence underscores the importance of rigorous occupational health and safety programs in future projects with careful attention to physical hazards.
Several other types of impacts were not included in our analysis. First, we did not address the trade-off between health risks associated with PCB exposure and health benefits of nutrients, such as the longer-chain omega-3 fatty acids, docosahexaenoic acid and eicosapentaenoic acid, which also bioaccumulate in fatty tissues. PCBs and omega-3 fatty acids may have counteractive effects on similar end points, such as those pertaining to early cognitive development and cardiovascular health (Turyk et al. 2012). Regarding other potential exposure pathways, previous work determined that baseline risks from incidental ingestion of sediment and water, dermal contact with sediment and water, and inhalation of volatilized PCBs were generally below levels of concern (TAMS Consultants and Gradient Corporation 2000). In future work, it may be worth evaluating potential health risks associated with consumption of resident waterfowl. This exposure pathway is complicated, however, by an inability to readily distinguish between resident and migratory waterfowl of the same species (HRNRT 2013). There is also an ongoing investigation of PCB contamination along the Hudson River floodplain (U.S. EPA 2017), which was not evaluated in this study. Additionally, we were unable to assess potential health risks of the source control action, and we did not evaluate the general benefit of full use of the Site without institutional controls. Finally, because this study focused on human health risk, potential ecological, social, and economic effects of environmental dredging were not addressed.