Enhancing a job exposure matrix with subject-specific information to assess combined exposure to benzene, toluene, and xylene in a case-control study

Abstract

Objectives

We modified an existing job-exposure matrix (CANJEM) to estimate the combined exposure to benzene, toluene, and xylene, which are highly correlated (hereafter, BTX), and integrated subject-specific occupational information to assess BTX exposure for participants in a bladder cancer case-control study.

Methods

We linked CANJEM to the lifetime occupational histories of subjects in a population-based case-control study of bladder cancer. We derived CANJEM-based estimates of the probability, intensity, and frequency of BTX exposure by assigning the highest rating observed among the benzene, toluene, and xylene CANJEM metrics. We used subject-specific exposure information in the occupational histories and exposure-oriented modules to refine the CANJEM-based BTX metrics when confirmatory exposure information was identified (hereafter, hybrid BTX metrics). We compared agreement between the CANJEM-based and hybrid BTX metrics at the job- and subject-level using kappa for the ordinal probability metrics and Spearman correlation for the continuous intensity and cumulative exposure metrics.

Results

The hybrid BTX approach increased 7% and decreased 5% of the CANJEM-based BTX probability ratings at the job-level and 16% and 7% at the subject-level, respectively. The CANJEM-based and hybrid BTX metrics identified 3.2% and 6.8% of the job records and 13% and 24% of the subjects as having a high probability of BTX exposure, respectively. The BTX-exposed subjects identified through the hybrid approach generally had lower cumulative exposures than those identified solely using CANJEM. CANJEM-based and hybrid BTX metrics had moderate agreement at the subject-level (probability: kappa = 0.62; cumulative metrics, Spearman correlation = 0.61).

Conclusions

Supplementing a JEM with subject-specific exposure information identified within-job exposure heterogeneity that was not captured by using only a JEM.

What's Important About This Paper?

Job-exposure matrices (JEMs) are an efficient approach for assigning exposure in epidemiologic studies but are unable to capture subject-specific exposure differences commonly observed within jobs. In this study, an existing JEM was modified to group highly correlated agents and account for subject-specific occupational information that was available in the occupational history and exposure-oriented modules collected by personal interviews. Accounting for subject-specific information changed 12% of the exposure probability ratings at the job-level and 23% at the subject-level. The cumulative exposure metrics were only moderately correlated, suggesting that incorporating subject-specific information captured exposure heterogeneity that was not captured by using only a JEM.

Introduction

The mononuclear aromatic hydrocarbons benzene, toluene, and xylene are generated from petroleum feedstock during the petroleum refining process. They are used in the production of numerous industrial chemicals and have been components of many products, including degreasing and cleaning agents, fuels, glues and adhesives, paints, lacquers, stains, varnishes, paint thinners, and printing inks (Table S1). In addition, other industrial-grade solvents produced during the petroleum refining process, especially through the early 1980s, contain mixtures of these compounds, both by design and as product impurities (Fishbein 1985a, 1985b; Tharr and Kudla 1997). For example, historically toluene was commonly contaminated with benzene, ranging from <0.01% benzene in highly purified toluene to as much as 25% in some industrial-grade toluene (Fishbein 1985a, 1985b). These aromatic compounds are also natural constituents of gasoline and have been added to gasoline as anti-knock agents since the phase-out of lead additives starting in the 1970s (IARC 1989; ATSDR 1995). Exposure assessment of these chemicals in epidemiologic studies has been challenging because of their co-occurrence and frequent use as substitution agents (Lynge et al. 1997; Gérin et al. 1998; Blanc-Lapierre et al. 2018; Warden et al. 2018). Due to this considerable overlap, these agents have high correlations across exposure metrics used in population-based epidemiologic studies (Hadkhale et al. 2017; Blanc-Lapierre et al. 2018; Warden et al. 2018), where exposure metrics are based on occupational information obtained from questionnaires (eg job classification; work tasks) rather than measurements. This high correlation is problematic when the correlated exposures are included within the same statistical model, making it difficult to estimate the effect of each agent after adjusting for exposure to the other agents.

To address this problem and to complement the analysis of single-agent metrics for benzene, toluene, and xylene, we developed 2 sets of exposure metrics that evaluated their exposure as a group for use in epidemiologic analyses within a population-based case-control study of bladder cancer. Hereafter, we refer to this group as BTX when describing the combined metrics and use the agents' names when describing metrics based on individual agents. For both sets of BTX metrics, we started with metrics available in CANJEM, a job-exposure matrix developed in Canada (Sauvé et al. 2018). The first set was derived solely from the CANJEM metrics (hereafter, CANJEM-based BTX metrics), where exposure metrics were assigned to subjects solely based on their 2010 US Standardized Occupational Classification codes. The second set was derived by revising the CANJEM-based BTX metrics to integrate subject-specific occupational information collected in the lifetime occupational histories and exposure-oriented modules (hereafter, BTX hybrid metrics).

In this paper, we describe the development of the CANJEM-based BTX and hybrid BTX metrics for the participants in the New England Bladder Cancer Study (NEBCS) and compare the resulting metrics.

Methods

Study population and available occupational information

The NEBCS study population has been previously described (Colt et al. 2011; Xie et al. 2024). We focus here on the occupational components of the study. Overall, 2,621 subjects between the ages of 30 and 79 reported 14,983 jobs held for at least 6 mo in a lifetime occupational history questionnaire completed during personal interviews conducted in 2001 through 2005. This questionnaire contained open-ended questions on job title, name/location of employer, type of service/product provided, year started/stopped, work frequency, principal work tasks and duties, tools and equipment used, and chemicals and materials handled. For 64% of the jobs, we also obtained the subjects' responses to job- or industry-specific modules, which collected detailed information on exposure-related tasks. All jobs were coded by experts using the 1980 US Standard Occupational Classification (SOC) system. We later coded these jobs to the 2010 system using crosswalks; jobs with multiple code matches were reviewed and assigned a single code by experts.

CANJEM metrics

We first assessed solvent exposure for this population using CANJEM, a job-exposure matrix developed from expert judgment applied to jobs reported in multiple Canadian case-control studies (Sauvé et al. 2018). In this paper, we focus on 4 agents available within CANJEM: the individual agents benzene, toluene, and xylene, as well as the broad group of mononuclear aromatic hydrocarbons (MAH), which includes these 3 agents as a subset, as well as styrene, phenol, ethyl benzene, and other single benzene ring compounds. We used the default CANJEM parameters for all parameters except calendar year to define exposure as broadly as possible: reliability interpretation R1 ≥ exposed; minimum frequency 0.5 h; minimum intensity = 1; minimum no. of jobs = 10; minimum number of subjects = 3, minimum frequency weighted intensity = 0.05; frequency-weighted threshold = 5. However, we stratified the time window into 4 calendar periods based on broad periods of changes in solvent use (1920 to 1970, 1971 to 1980, 1981 to 1990, and 1991 to 2005). We split NEBCS job records that straddled time periods into multiple records; the 14,983 jobs became 21,220 time period-specific job records. We linked each job record to the 4 CANJEM agents (benzene, toluene, xylene, MAH) based on its SOC-2010 code. We used a hierarchical approach because CANJEM did not provide an estimate if that combination of SOC code and time period had less than 10 records and/or less than 3 subjects. Whenever possible, we used the time period-specific estimate for each 6-digit code; if unavailable, we used its broader 1920 to 2005 estimate. We repeated the linkage at the 5-digit level, followed by the 3-digit level, until all records had a set of metrics for the 4 agents.

For each agent and SOC code, CANJEM provides the proportion of subjects exposed, frequency of exposure, intensity of exposure, and the frequency-weighted intensity (FWI) on a continuous scale. The proportion exposed metric was provided on a 0 to 100 scale representing the proportion of jobs with that code that would be assigned exposed to that agent. We used the proportion exposed as a measure of probability of exposure (P) and categorized it into 4 categories: unexposed (proportion exposed = 0); unlikely exposed (proportion exposed >0 and <25); low (proportion exposed ≥25 and <50); and high (proportion exposed ≥50). The frequency metric was provided as the median hours exposed per week. The intensity metric was provided as a weighted median based on the CANJEM-recommended 1-5-25 weighting scale to convert the low, medium, and high ordinal ratings intensity ratings to a numeric scale. The choice of the 1-5-25 scale was consistent with the high contrast observed in a recent retrospective benzene exposure assessment that incorporated publicly available measurements (Dopart et al. 2019), supporting the use of a high contrast scale over the lower contrast scales that are also provided by CANJEM (ie 1-3-9; 1-2-3). The FWI metric adjusted the intensity by the frequency of the exposure.

CANJEM-based BTX estimates

To account for the potential co-exposure of benzene, toluene, and xylene, but to also maintain these exposures as a distinct subset of MAH, we derived a set of exposure metrics that grouped these 3 agents. Probability of BTX exposure was assigned to each job/time period-specific record based on the highest probability level observed among benzene, toluene, and xylene. Similarly, the frequency, intensity, and FWI metrics for BTX were assigned the highest value of those respective metrics from the 3 agents but based only on those agents with a high probability. For example, if benzene and toluene both had high probability ratings and xylene had a low probability rating, we excluded the xylene metrics when we determined the highest values for frequency, intensity, and FWI. We refer to this set of metrics as the CANJEM-based BTX estimates to reflect that they were derived from CANJEM metrics but were not provided by CANJEM.

Hybrid BTX estimates

The CANJEM metrics for the 4 agents and the CANJEM-based BTX metrics were assigned based on the SOC code, which does not consider any of the other occupation information available. To account for subject-specific occupational information, we derived a set of BTX metrics that revised the CANJEM metrics based on expert-based decision rules applied to the additional information. We describe our approach below.

Step 1: identifying BTX exposure scenarios

Our first step was to identify BTX exposure scenarios. We reviewed the published and gray literature to identify scenarios in which exposure to benzene, toluene, or xylene has been reported. This included a previous literature review of benzene (van Wijngaarden and Stewart 2003) and a recent benzene exposure assessment based on exposure decision rules and similar occupational information in a case-control study (Dopart et al. 2019). Although Dopart et al. (2019) were focused on assessing benzene exposure, that assessment included a thorough identification of jobs and exposure scenarios with potential exposure to other aromatic hydrocarbons and solvents such as those containing toluene and xylene because of the potential for benzene co-occurrence. We summarize the main sources of BTX in Table S1.

We conducted a programmed search of the tasks, tools/equipment, and chemicals/materials questions from the lifetime occupational history responses using a previously developed list of keywords that may be associated with potential BTX exposure. This keyword search approach was first described by Friesen et al. (2014) and expanded for benzene and other aromatic hydrocarbons (including toluene and xylene) by Dopart et al. (2019). This search captured 30 tasks (eg painting), 16 tools (eg vehicle mechanic's tools), and 24 chemicals (eg gasoline) with potential exposure to aromatic hydrocarbons.

We also reviewed all NEBCS modules for questions that directly or indirectly contained information on BTX exposure. Examples of questions include: “Did you use aromatic hydrocarbons, such as benzene,” “How often did you use toluene to clean these tools or equipment,” “How often did you use xylene or xylol,” “What type of paint did you use,” “Did you use gasoline to clean your hands,” “Did you repair vehicles,” and “Did you pump gasoline”. We also identified variables related to the frequency of activities, such as “How often did you clean your hands with gasoline?”

Step 2: job/time-period assessment of probability

We assumed that the CANJEM-based BTX probability rating was likely correct for most job/time period-specific records and used a targeted approach for identifying those needing review. These job-level reviews were conducted by 2 experts with over 25 years of professional experience in occupational hygiene, retrospective exposure assessment, and epidemiological research (M.C.F.; S. Viet); they consulted a third expert (S.J.L.) for exposure scenarios when they desired additional expertise. They reviewed each job together and assigned a consensus estimate for the hybrid BTX metrics based on the criteria and exposure assignment decisions listed in Table 1, which were calibrated to align with the CANJEM-based metrics. For jobs with a module response, the experts reviewed any jobs that met 1 of 3 criteria: (i) a CANJEM-based BTX-exposure probability rating of high; (ii) responses to 1 or more module questions indicating potential BTX exposure; or (iii) a free-text response on tasks, tools, or chemicals that identified possible BTX exposure. For jobs without a module (36% of jobs), only jobs with a free-text response on tasks, tools, or chemicals that identified possible BTX exposure were reviewed. In all assessments, if neither confirmatory (which may increase a probability rating) or contradictory information (which may lower a probability rating) was identified, the CANJEM BTX rating was maintained. The experts were blinded to case-control status.

Table 1.

Exposure scenarios identified in modules with probability of BTX exposure of ≥25% pre-1990 and their hybrid BTX ratings.

Question	Criteria	Probability rating	Intensity rating	Frequency approach	Applicable modules
Questions with specific BTX chemicals, sources, or activities
Used aromatic hydrocarbons?	>0 h/wk	High	Medium	Subject report	Chemist
Used naphtha?	>0 h/wk	High	Medium	Subject report	Janitor, Cabinet Maker
Used toluene?	>0 h/wk	High	Medium	Subject report	Janitor, Cabinet Maker
Used xylene?	>0 h/wk	High	Medium	Subject report	Janitor, Cabinet Maker
Used naphtha to clean equipment?	Yes	High	Medium	Subject report, adjusteda	Painter
Used toluene to clean equipment?	Yes	High	Medium	Subject report, adjusteda	Painter
Used xylene to clean equipment?	Yes	High	Medium	Subject report, adjusteda	Painter
Used gasoline or cleaning solvents to clean equipment?	≥0.5 h/wk	High	Medium	Subject report	Farm Worker
Used gasoline or cleaning solvents to clean hands?	≥0.5 h/wk	High	High	Subject report	Farm Worker
Used gasoline to clean hands?	Yes	High	High	Subject report, adjusteda	Welder, Electiran, Plumber, Tile Setter, Carpenter
Pump gasoline?	≥0.5 h/wk	High	Low	Subject report	Gas Station Attendant
Gasoline used to clean paint equipment?	Yes	High	Medium	Subject report, adjusteda	Mechanic
Drive or operate gasoline-powered equipment in tunnel?	>0.5 h/wk	Low	Medium	Subject report	Heavy Construction
Repair vehicles?	>0.5 h/wk	High	Low	Subject report	Gas Station Attendant
Use oil-based paints	>0.5 h/wk	High	Medium	Subject report	Painter
Use oil or solvent-based wood stains	>0.5 h/wk	High	Medium	Subject report	Painter
Cemented or glued cores together? (Foundry)	Yes	Low	Medium	Subject report, adjusteda	Foundry Industry
Worked in furnace area? (Foundry)	≥0.5 h/wk	Low	Low	Subject report	Foundry Industry
Questions with free-text answers that could contain BTX chemicals
What chemicals were being made?	BTX-containing chemical	High	Medium	Expert judgmentb	Chemical Industry, Engineer
…	Solvent-containing chemicals, but too generic to determine if BTX-containing	Low	Medium	Expert judgmentb	Chemical Industry, Engineer
What chemicals did you smell?	BTX-containing chemical	High	Low	Expert judgmentb	Engineer, Foundry Industry
…	Solvent-containing chemicals, but too generic to determine if BTX-containing	Low	Low	Expert judgmentb	Engineer, Foundry Industry
What chemicals did you use in class?	Oil-based paints, varnish, paint thinner, benzene	High	Low	Expert judgmentb	Teacher
…	Solvent-containing chemicals, but too generic to determine if BTX-containing	Low	Low	Expert judgmentb	Teacher
What chemicals did you use?	BTX-containing chemical	High	Medium	Expert judgmentb	Janitor, Cabinet Maker, Sheet Metal Worker
…	Solvent-containing chemicals, but too generic to determine if BTX-containing	Low	Medium	Expert judgmentb	Janitor, Cabinet Maker, Sheet Metal Worker
What chemicals did you use to clean or degrease parts?	Gasoline, naphtha, toluene, gunk, paint thinner, lacquer thinner, aviation fuel	High	Medium	Subject report, adjusteda	Plumber, Sheet Metal Worker, Assembly Worker, Plumber, Mechanic
What chemicals used to clean tanks, vats, chemical sewers, or other equipment or machines?	Gasoline, naphtha	High	High	Subject report, adjusteda	Laborer, Machinist
What chemicals were mixed with dyes?	Toluene	High	Medium	…	Pulp and Paper Industry
What chemicals were used to clean pipes and other equipment?	Benzene, gasoline, naptha, paint thinner	High	High	Subject report, adjusteda	Industrial Mechanic
What chemicals were used to clean plates, blankets, presses, other?	Gasoline, lacquer thinner	High	High	Subject report, adjusteda	Printing Industry
What type of finish was used?	Lacquers	Low	Medium	Subject report, adjusteda	Cabinet Maker
What did you use to clean paint/maintenance equipment?	Gasoline, lacquer thinner, toluene, thinnex	High	Medium	Subject report, adjusteda	Painter
Strip paint with chemicals?	Lacquer	High	Medium	Subject report, adjusteda	Painter
What else did you do?	Maintenance work, mechanical work, painting	Low	Medium	Subject report, adjusteda	Bus Driver

^aSubject-reported time spent cleaning (or other stated activity), which we divided by the number of agents that the subject reported using.

^bExpect judgment considered the subject-reported frequency from other questions in module and the number of agents reported.

Step 3: assignment of frequency-weighted intensity

We assigned every job/time period-specific record that did not have a hybrid BTX probability rating of high a hybrid BTX intensity, frequency, and FWI estimate of 0. For records with a high probability rating, we assigned a nonzero intensity and FWI based on the information available in the module. We used subject-specific frequency (typically reported in hours per week on a continuous scale) when it was available and multiplied it by the corresponding intensity estimate from Table 1. When we had information on the frequency of the task, such as time spent cleaning, but not for the frequency of using a specific BTX-containing agent, we divided the time spent on the task by the number of agents reported. When no specific frequency information on the BTX-exposed activity was available in the module, we used expert judgment. For jobs without modules, if the CANJEM-based BTX probability rating was low or high, we assigned the CANJEM-based BTX FWI value; if the CANJEM-based BTX probability rating was unexposed or unlikely exposed, we used expert judgment.

Cumulative exposure

For the 4 CANJEM metrics and 2 BTX metrics, each subject's cumulative exposure was calculated by multiplying the FWI by the job/time period-specific job duration and summing across all jobs held by the subject. Jobs with an unlikely BTX exposure probability were assigned an FWI of 0 and thus did not contribute to the cumulative exposure calculation; however, subjects with uncertain BTX exposure were treated as a separate exposure group (and not considered unexposed) in etiologic analyses (Xie et al. 2024).

Comparisons between metrics

Descriptive statistics for the 4 CANJEM metrics and 2 BTX metrics was calculated at the time period-specific job/time period-specific level and at the subject-level. We compared the probability metrics using unweighted kappa because the planned epidemiologic analyses would use the probability metric as a categorical, not ordinal metric, and as binary (high vs. not high) when calculating the cumulative exposure metric. We compared the cumulative exposure metrics using Spearman correlation. Cross-tabulations of the 2 BTX metrics were used to determine the direction of rating change when we applied the hybrid BTX exposure assessment approach. Comparisons were also stratified by case-control status.

Results

A cross-tabulation of the CANJEM-based BTX and hybrid BTX probability ratings is shown in Table 2. At the job-level, the hybrid BTX approach resulted in no change of the CANJEM-based BTX probability rating for 18,817 (89%) of the 21,220 job/time period records, a higher rating for 1,409 (7%) records, and a lower rating for 994 (5%) records. Overall, 755 (3.6%), 513 (2.4%), and 141 (0.7%) records were raised 1, 2, or 3 rating categories, respectively. Similarly, 886 (4.2%), 98 (0.5%), and 10 (0.05%) were lowered 1, 2, or 3 rating categories, respectively. The changes occurred mostly among those with module responses, where subject-specific occupational information raised the BTX probability rating for 1,111 (9%) of the 11,874 records and lowered it for 846 (7.1%) records (data not shown). Most of the increased ratings were among those receiving the following modules: Mechanic (n = 215 raised, 47% of records with module), Farmer (n = 101, 26.2%), Gas Station (n = 87, 92%), Diesel Engine (n = 83, 12%), Painter (n = 71, 60%), Printing Industry (n = 64, 53%), and Cabinet Maker (n = 51, 11%). Most of the lowered ratings occurred among those receiving the Office Professional (n = 409, 30%), Teacher (n = 122, 19.2%), and Cabinet Maker (n = 113, 24%) modules. Of those without module responses, we raised the probability rating for 298 (3.2%) of the 9,346 records and lowered it for 148 (1.6%) records. The patterns in the direction of probability change were similar by case-control status. The rating was unchanged for 89% of the jobs held by both cases and controls; lowered for 4.1% and 5.2% of jobs held by cases and controls, respectively; and raised for 6.5% and 5.5% of jobs held by cases and controls, respectively.

Table 2.

Comparison between CANJEM-based BTX and hybrid BTX probability ratings at the job and subject-level.

	Hybrid BTX probability rating, number of jobs/Subjects
CANJEM-based BTX probability rating	Unexposed		Unlikely		Low		High		All
	N	%	N	%	N	%	N	%	N	%
Time period-specific job-level
Unexposed	8,291	39	219	1.0	48	0.2	141	0.7	8,699	41.0
Unlikely	622	2.9	9,494	45	219	1.0	465	2.2	10,800	50.9
Low	7	0.03	218	1.0	503	2.4	317	1.5	1,045	4.9
High	10	0.05	91	0.4	46	0.2	529	2.5	676	3.2
All	8,930	42	10,022	47	816	3.8	1,452	6.8	21,220	100
Subject-level
Unexposed	181	7.0	5	0.2	1	0.04	2	0.1	189	7.3
Unlikely	61	2.4	1,363	53	60	2.3	167	6.4	1,651	64
Low	0	0.0	53	2.0	157	6.1	172	6.6	382	15
High	1	0.0	37	1.4	20	0.8	310	12	368	14
All	243	9.4	1,458	56	238	9.2	651	25	2,590	100

At the subject-level, the 2 BTX metrics had the same probability rating for 2011 (78%) of the 2,590 subjects (Table 2). The hybrid BTX rating resulted in a higher rating than the CANJEM-based BTX rating for 407 (16%) subjects and a lower rating for 172 (7%) subjects. The majority (64%) of these changes were within 1 rating category. In total, 341 subjects with unexposed, unlikely, or low CANJEM-based BTX probability had confirmatory BTX exposure evidence in at least one of their jobs that increased their ever probability rating to high with the hybrid approach. These patterns were independent of case-control status, with the same proportions raised to a high BTX probability rating (controls, 13.3%; cases, 13.0%) or lowered from a high CANJEM BTX probability rating (controls 2.3%, cases, 2.2%).

MAH had the highest prevalence of subjects with a high probability rating (28.5% of subjects), then the hybrid and CANJEM BTX metrics (23.7% and 13.3%, respectively) (Table 3). The median cumulative exposure among exposed subjects was lower for the hybrid BTX metric than the CANJEM-based BTX metric, at 6.0 vs. 7.5 dimensionless unit-years, respectively. Exposure contrast ratios based on the cumulative exposure, calculated as the ratio of the 75th to 25th percentiles, were similar across all metrics (range 7.2 to 8.7), except xylene (5.2). We compare the 2 BTX cumulative metrics for the NEBCS participants in Fig. 1. Although there is a strong diagonal line identifying identical exposure values using both approaches, one can see that the hybrid BTX approach resulted in estimates that were both higher and lower than the CANJEM-based BTX metric across the full exposure range.

Table 3.

Subject-level comparisons: prevalence of each probability category by agent and agreement between exposure metrics for 2,590 NEBCS subjects.

Agent	Proportion (%) assigned to probability category, based on highest rated job(s)			Cumulative exposure if >0,		Kappa of probability metrics		Spearman correlation of cumulative exposure
	Unexposed	Unlikely/Low	High	Median (IQR)	Contrast Ratio (75th/25th percentile)	vs. Hybrid BTX	vs. CANJEM BTX	vs. Hybrid BTX	vs. CANJEM BTX
BTX: Hybrid	9.4	66.9	23.7	6.0 (2.3 to 17.8)	7.7	Reference	0.62	Reference	0.61
BTX: CANJEM-based	7.3	79.4	13.3	7.5 (2.6 to 22.6)	8.7	0.62	Reference	0.61	Reference
CANJEM Metrics
MAH	4.2	67.3	28.5	9.2 (3.2 to 24.0)	7.5	0.47	0.49	0.62	0.67
Benzene	11.7	77.3	11.0	6.0 (2.5 to 18.0)	7.2	0.49	0.75	0.56	0.90
Toluene	8.7	83.7	7.6	10.0 (3.8 to 29.3)	7.7	0.49	0.77	0.45	0.76
Xylene	13.4	79.6	7.0	5.6 (2.5 to 13.1)	5.2	0.36	0.51	0.44	0.73

MAH, mononuclear aromatic hydrocarbons.

Figure 1.

Scatterplot comparing the CANJEM-based BTX and hybrid BTX cumulative exposure metrics.

The hybrid BTX metric had fair-to-moderate subject-level agreement in probability ratings with the other metrics, with kappa ranging from 0.36 for xylene to 0.62 for the CANJEM-based BTX metric (Table 3). The CANJEM-based BTX probability metric had high agreement with the benzene and toluene metrics (kappas of 0.75 and 0.77, respectively). The cumulative hybrid BTX metric was slightly more correlated with the CANJEM-based BTX and MAH metrics (0.61 and 0.62, respectively), than with benzene, toluene, or xylene (0.44 to 0.56). In contrast, the CANJEM-based BTX cumulative metric was more highly correlated with benzene, toluene, and xylene (0.73 to 0.90) than the MAH metric (0.67).

The observed differences overall between CANJEM and hybrid BTX metrics did not change when stratified by case-control status. There was no difference in agreement between the probability ratings of the 2 BTX metrics by case status (controls, kappa = 0.61; cases, kappa = 0.63). Similarly, regardless of case status, the median cumulative exposure metrics was lower for the hybrid BTX metric (exposed controls, median = 5.3, IQR 1.9 to 14.7; exposed cases, median = 6.7, IQR 2.5 to 21) than for the CANJEM BTX metrics (exposed controls, median = 7.5, IQR 2.7 to 20; exposed cases, median = 8.4, IQR 2.7 to 23). Spearman correlations comparing the 2 BTX cumulative exposure metrics was 0.58 for controls and 0.63 for cases.

Discussion

In this study, we demonstrated how a JEM can be used to derive additional study-specific metrics for epidemiologic analyses. In particular, we derived group-based metrics representing exposure to benzene, toluene, and/or xylene from existing JEM metrics, to which we then integrated subject-specific occupational information. This allowed us to revise the CANJEM-based probability ratings when information confirming or refuting use of BTX-containing agents was observed in the lifetime occupational history or module responses from personal interviews with all subjects. It also allowed us to modify group-based frequency estimates with subject-specific frequency of exposure.

Our hybrid approach combined 3 common exposure assessment approaches—JEMs, decision rule-based assessment linking questionnaire responses to exposure decisions, and expert job-by-job review. Although JEMs are an efficient exposure assessment tool, they do not account for within-job heterogeneity in exposure that may lead to exposure misclassification. Here, we reduced the exposure assessment burden by using existing tools and accounted for subject-specific information when it was available (Dopart and Friesen 2017; Sauvé and Friesen 2019; Borghi et al. 2020). Our approach has similarities to the hybrid expert approach described by Sauvé et al. (2019), which used a JEM as a starting point and focused expert review to jobs in occupations that were expected to have heterogeneity in exposures, thus reducing the time of the expert assessors by half compared to traditional expert assessment.

Subject-specific information increased the JEM-based probability rating for many jobs and subjects because of confirmatory information observed in the responses to the structured questionnaires. Much of this confirmatory information was available in free-text responses, to which participants reported their most frequently used products. We did not assume no exposure simply because BTX-containing products were not mentioned. Demotions in exposure probability predominantly occurred when the subject responded “no” to a series of solvent-related questions, indicated no use of chemicals, conducted only office-based work tasks, and/or clearly described only unrelated chemicals (eg a chemist working in an inorganic chemistry lab). The free-text responses made it challenging to apply programmable decision rules. Existing keyword lists were used in a programmed search of the occupational history questions; however, because of frequent misspellings entered by the interviewers, we also conducted a review of all unique entries to the question “what chemicals and materials did you use in this job?” Scanning the unique responses of a single question was efficient and improved the sensitivity of our BTX exposure identification. The programmed keyword search was not applied to the free-text module responses; however, these were available for much fewer jobs and were reviewed in batches with a response to a module-specific question.

The hybrid BTX metrics had a higher prevalence of high probability ratings than the CANJEM-based BTX metrics. Subjects identified as having a high probability of BTX exposure via the hybrid BTX metric but not the CANJEM-based metric had on average lower cumulative exposure to BTX. This suggests that the hybrid approach was more sensitive in identifying lower intensity and/or frequency of BTX exposures. Both BTX metrics had a lower prevalence of high probability ratings than MAH, but a higher prevalence than benzene, toluene, or xylene. Given that benzene, toluene, and xylene have been frequently found in varying quantities within the same products, our findings suggest that the CANJEM metrics may miss capturing products with lower levels of an individual agent. MAH's higher prevalence demonstrates that it is capturing a broader set of agents than solely the BTX group. These distinctions between exposure metrics will be useful in interpreting results from epidemiologic analyses that use the existing CANJEM metrics and new BTX metrics.

The agreement observed between the hybrid and CANJEM-based BTX exposure metrics was at the upper end of the agreement seen in other studies that compared JEM and expert-based exposure estimates, which reported weighted kappas ranging from 0.10 to 0.70, with a median of ∼0.36 [ref 20]. Weighted kappas in this study were similar to the unweighted kappas reported here and did not change any interpretation (not shown). The higher agreement we observed is unsurprising as the BTX metrics were not independent. The agreement between BTX metrics was similar to the agreement previously observed between estimates from expert judgment vs. expert-derived decision rules, which reported weighted kappas ranging between 0.49 and 0.82 (Ge et al. 2018).

The strengths of our work include maximizing the use of existing exposure resources that are geographically relevant to the study population, including CANJEM, keyword search lists, and decision rules developed to identify exposure to benzene and other aromatic hydrocarbons. Our approach removed the need for expert review of all jobs, which made it feasible for the review to be efficiently conducted by 2 experts. To minimize recall bias, the occupational information was collected using structured questionnaires by trained interviewers. The interview and the coding of jobs and expert review of the questionnaires were conducted blind to case-control status, with our stratified analyses showing that the exposure estimates for cases and controls were as equally likely to be raised or lowered when we incorporated the subject-specific questionnaire responses using our hybrid approach.

Our approach also had several limitations. First, we assumed that the CANJEM single-agent assessments were correct for most jobs and calibrated our application of exposure decision rules and expert judgment to align with those estimates. However, we may have missed accounting for potential within-occupation exposure differences among the unreviewed jobs. Second, much of the BTX-related exposure information was provided in free-text responses that required job-level review; these were reviewed in batches to maintain internal consistency and were based on exposure decisions described in Table 1. Third, the exposure-oriented modules were developed specifically to capture exposures that were suspected bladder carcinogens (eg diesel fuels/exhausts, metal-working fluids), so information related to aromatic hydrocarbons was often indirect or based on broad groups of solvents; additionally, subject-specific frequency was not always available. Other NCI studies have used modules with more targeted solvent questions, such as the case-control study of non-Hodgkin lymphoma for which decision rule-based estimates for benzene were previously developed (Dopart et al. 2019). Fourth, the intensity estimates from CANJEM were based on expert judgment from Canadian studies which assessed intensity as low, medium, and high, to which a 1-5-25 weighting was later applied; thus, they were not calibrated to a measurement scale or modified to a US exposure setting, which may have some differences in workplace regulations. Lastly, with the lack of opportunities to validate these exposure metrics, we are limited to indirect validation by examining the sensitivity of exposure–response associations to the choice of exposure metric; however, by developing exposure metrics to examine these agents individually and combined, we may gain insight into the potential causative agents.

We recently reported the impact of these exposure assessment decisions on exposure–response associations with bladder cancer risk (Xie et al. 2024). In Fig. 2, we show the main findings for the metrics described by Xie et al. (2024), after applying a 20-yr lag to the cumulative exposure metrics. Exposure–response trends were observed only with the CANJEM-based BTX and hybrid BTX metrics, but not with the MAH, benzene, toluene, or xylene CANJEM metrics. The hybrid BTX metrics showed a steeper and stronger trend than the CANJEM-based BTX metric, which is likely because the hybrid BTX approach identified more subjects with BTX exposure; these subjects more often had exposure levels at the lower end of the cumulative exposure scale. These exposure trends provide only an indirect validation that should be interpreted cautiously; however, they suggest that considering these exposures as a group to account for their co-exposures in population-based studies, where the individual agents are difficult to disentangle, as well as enhancing the JEM to capture subject-specific exposure differences, likely reduced exposure misclassification in this study. It also suggests that the BTX metrics captured different exposure features than the original benzene, toluene, xylene, and MAH metrics from CANJEM. Moreover, the observed differences between the hybrid and CANJEM-based BTX metrics, and their impact on exposure–response associations, support the need to capture subject-specific exposure differences when possible. Recently, the International Agency for Research on Cancer classified automotive gasoline as “carcinogenic to humans” (Group 1) based on sufficient evidence for bladder cancer and acute myeloid leukemia (Turner et al. 2025). Since gasoline and its use as a cleaning solvent was a common BTX source in our population, further refinement of the exposure metrics to separately consider gasoline-specific sources may be warranted.

Figure 2.

Exposure–response associations with bladder cancer incidence and cumulative exposure metrics related to benzene, toluene, and xylene, and the combined CANJEM-based BTX and hybrid BTX metrics (20-yr lag) in the New England Bladder Cancer Study (details of analyses reported in Xie et al. 2024).

Author contributions

M.C.F., S.K., S.X., and D.S. conceptualized and developed the exposure assessment approach. S.J.L. contributed exposure-specific knowledge. D.B., M.S., A.J., N.R., and M.R.K. conceived, designed, and implemented the case-control study. M.C.F. conducted the statistical analyses and wrote the first draft of the manuscript. All authors provided feedback on the manuscript.

Supplementary material

Supplementary material is available at Annals of Work Exposures and Health online.

Funding

This research was supported by the Intramural Research Program of the National Institutes of Health (NIH), National Cancer Institute, Division of Cancer Epidemiology and Genetics (ZIA CP010125-27).

Data availability

Requestors interested in the data supporting the findings in this paper can access these data through a data transfer agreement with NCI. For more information on the data sharing policy of the NCI Division of Cancer Epidemiology and Genetics, follow this link: https://dceg.cancer.gov/tools/data-sharing.

Ethical approval

The study protocol of the New England Bladder Cancer Study was approved by the National Cancer Institute Special Studies Institutional Review Board and the human subjects review boards of each participating institution.