Validity of Expert Assigned Retrospective Estimates of Occupational Polychlorinated Biphenyl Exposure

Abstract

Assessment of retrospective exposures based on expert judgment in case-control studies is usually of unknown validity because of the difficulty in finding gold standards for comparison. We investigated the relationship between expert-assigned retrospective occupational polychlorinated biphenyl (PCB) exposure estimates and serum PCB concentrations. Analyses were conducted on a subset of cases (n = 94) and controls (n = 96) in the multi-center National Cancer Institute, Surveillance, Epidemiology, and End Results Case-Control Study of non-Hodgkin lymphoma. Based on the subjects’ lifetime work histories, an industrial hygienist assigned each job a probability of PCB exposure [<5% (unexposed), 5–<50% (possibly exposed), ≥50% (probably exposed)]. Ordinary least squares regression was used to investigate associations between the probability rating and log-transformed lipid-adjusted serum levels of 14 PCB congeners and total PCBs (ΓPCBs). Compared to unexposed participants (n = 163), those with a probably exposed job (n = 7) had serum levels that were 87% higher for ΓPCBs (95% confidence interval: 1.33–2.62) and 38% of serum level variability was explained by the probability rating. Statistically significant associations between probability ratings and serum levels for 12 of 14 individual congeners were also observed. In summary, the observed contrast in PCB serum levels by probability rating provides support for the occupational PCB exposure assessment.

INTRODUCTION

Polychlorinated biphenyls (PCBs) were widely used in electrical equipment, machinery, and construction materials until they were banned in the USA in the late 1970s due to ecological and health concerns (EPA 2011). The International Agency for Research on Cancer has recently classified PCBs as a human carcinogen (Group 1), although epidemiologic evidence is limited for some cancer sites, including non-Hodgkin lymphoma (NHL) (Lauby-Secretan et al., 2013). Retrospective occupational PCB exposure assessments in case-control studies could provide useful new evidence to clarify the carcinogenicity of PCBs for such malignancies. The assessment of occupational exposures, especially in case-control studies, often relies on expert judgment based on review of job history information. The validity of such retrospective occupational exposure assessments is, however, usually unknown due to the limited availability of gold standards with which to compare the expert estimates, yet the validity of the assessments is critical to the understanding of true exposure-disease associations. Serum and other biological measurements that may represent long-term exposure are often considered the gold standard exposure metric for chemicals with long half-lives, such as PCBs. To date, the four population-based studies that have compared expert assessment to biomarker measurements (Hertzman et al., 1988; Tielemans et al., 1999; Allen et al., 2006; Chen et al., 2014) have found experts were able to reasonably distinguish between low and high exposures based on comparisons to urine measurements. Our aim, therefore, was to assess the validity of expert-assessed retrospective occupational PCB exposure estimates compared to serum PCB concentrations.

MATERIALS AND METHODS

We conducted analyses comparing retrospective occupational exposure assessment to serum PCB levels among a subset of participants from the National Cancer Institute Surveillance, Epidemiology, and End Results (SEER)-NHL study, a population-based case-control study conducted in four study sites covered by SEER cancer registries: Detroit, Michigan area (Macomb, Oakland, and Wayne counties); state of Iowa; Los Angeles County, California; and Seattle, Washington area (King and Snohomish counties). The study population has been described in detail previously (Chatterjee et al., 2004; Colt et al., 2005). Briefly, cases were men and women aged 20–74 years newly diagnosed with a first primary NHL between July 1998 and June 2000. In all, 1321 cases (76% of the 1728 eligible cases) participated. Of 2046 eligible population controls selected using random digit dialing (<65 years of age) or from Center for Medicare and Medicaid Services files (≥65 years), 1057 (52%) participated.

Demographic and occupational information was obtained from computer-assisted personal interviews conducted in each participant’s home. Overall, 2173 (91%) participants worked qualifying jobs and completed a lifetime occupational history (Purdue et al., 2011). The occupational questionnaire asked open-ended questions about each job held at least 12 months from age 16, including job title, employer, dates of employment, main work tasks, and tools and chemicals used. Modules that asked more detailed exposure questions were triggered based on certain occupational history responses, but none directly related to PCBs. A subset of participants were mailed a dietary questionnaire that asked about daily food item intakes, including items with potentially high PCB levels: fish, shellfish, dairy (butter, cheese, cream, milk), and red meat. The study was approved by human subjects review boards at all participating institutions, and written informed consent was obtained from each participant.

We measured serum PCBs among a random sample of untreated cases (N = 100) and controls (N = 100), matched by birth date, date of blood draw, sex, and study site, who provided at least 7.1ml serum. Detailed information on the selection of serum samples has been described by De Roos et al. (2005). Final analyses were limited to participants with serum PCB measures who provided a lifetime occupational history (N = 94 cases; N = 96 controls).

Estimates of PCB exposure for each job held were determined by an industrial hygienist who assigned a probability rating based on a detailed review of each job record as follows: not exposed (<5% of workers likely exposed), possibly exposed (5–<50% of workers likely exposed), and probably exposed (≥50% of workers likely exposed). The industrial hygienist was blinded to case/control status and PCB serum concentrations when assigning probability ratings. During the assessments, the industrial hygienist found insufficient information available to assign differences in exposure intensity. Final exposure probability ratings were assigned to each participant based on their cumulative occupational history, categorized as: not exposed (held no job or never worked a possibly or probably exposed job), possibly exposed (ever worked a possibly exposed job and never worked a probably exposed job), and probably exposed (ever worked a probably exposed job). Lifetime duration working a probably exposed job was also calculated.

PCBs in serum were measured by high-resolution gas chromatography/isotope-dilution high-resolution mass spectrometry at the Dioxin and Persistent Organic Pollutants Laboratory of the Centers for Disease Control and Prevention in Atlanta, GA (Turner et al., 1994). Details of the laboratory methods and quality control have been described (De Roos et al., 2005). The usual analytic limit of detection for PCBs in samples was 6.4ng g⁻¹ lipid and we restricted our analyses to PCB congeners that were detected in at least 30% of all samples measured in the entire study population, which included 12 of 36 noncoplanar congeners (74, 99, 118, 138–158, 146, 153, 156, 170, 180, 183, 187, and 194) and 2 of 4 coplanar congeners (126 and 169).

We estimated levels below the limit of detection using an unbiased multiple imputation procedure (Lubin et al., 2004; De Roos et al., 2005). Using a maximum likelihood method, we estimated values below the limit of detection by sampling from the appropriate log-normal distribution observed for each PCB congener, adjusting for age, sex, study site, and year of blood draw. The imputation procedure was repeated five times to account for random variation in the sample selection.

We used ordinary least squares regression to investigate the associations between the probability rating and log-transformed lipid-adjusted serum levels of the 14 PCBs and the sum of all 14 PCB congeners (ΓPCB). Models were adjusted for case-control status, age, sex, and study site. Adjustment for PCBs measured in house dust, which have been associated with occupational exposure in this study population (DellaValle et al., 2013), was examined but it did not change effect estimates (not shown). Beta coefficients were exponentiated to calculate the ratio of mean log-transformed PCB levels between probability rating categories, which represent the ratio of geometric means. Separate models were fit using the five imputation data sets. The results were combined using the MIANALYZE procedure in SAS version 9.2 (SAS Institute, Inc., Cary, NC, USA) to create a single ratio and 95% confidence interval (CI) accounting for the variability between the imputed values.

Additional analyses were performed by stratifying the probably exposed rating category by job start year (</≥1965), job stop year (</≥1988), and duration worked in a probably exposed job (</≥15 years). The proportion of total variability explained by the probability rating, i.e. between-group variance, adjusting for case-control status, age, sex, and study site, was calculated as between-group variance divided by total variance, accounting for all five imputed data sets.

To examine the potential impact of dietary PCB intake, we compared dietary intakes of food items with potentially high PCB levels (fish, shellfish, dairy products, red meat) by exposure probability rating among the subset of participants that completed a food frequency questionnaire (N = 66 unexposed, 8 possibly occupationally exposed, 2 probably occupationally exposed) (Lim et al., 2005).

RESULTS

Our retrospective exposure assessment classified 163 unexposed, 20 possibly exposed, and 7 (2 controls) probably exposed participants. Serum ΓPCB concentrations were higher among probably exposed participants (median: 528.7ng g ⁻¹ lipid, 25th–75th percentile = 352.0–965.1ng g ⁻¹ lipid) compared to possibly exposed (median: 287.1ng g ⁻¹ lipid, 25th–75th percentile = 177.0–359.6ng g ⁻¹ lipid; P _difference < 0.01) and unexposed participants (median: 314.8ng g ⁻¹ lipid, 25th–75th percentile = 221.4–425.3ng g ⁻¹ lipid; P _difference < 0.01) (Fig. 1).

Figure 1.

Boxplot of serum ΓPCB concentrations by occupational exposure probability rating. Boxes represent 25th, 50th, and 75th percentiles. Boxplot whiskers are 1.5 +/- interquartile range. Dots (□) represent serum means.

In the regression model, the geometric mean for ΓPCB serum levels was 87% higher among probably exposed participants (95% CI: 1.33–2.62) than unexposed participants (Table 1). Similar patterns were observed for individual PCB congeners (Table 1). Elevated serum concentrations were consistently observed for all noncoplanar and coplanar PCBs, with all but two (PCB 99 and 146) statistically significant. Those who started probably exposed jobs later than 1965 (n = 2) had higher mean serum levels (ΓPCB: ratio = 3.70; 95% CI: 2.04–6.72) than those starting before 1965 (n = 5) (ΓPCB: ratio = 1.35; 95% CI: 0.90–2.04). We did not observe any consistent association between year stopped working a probably exposed job and PCB level. PCB levels among participants working probably exposed jobs for <15 years (n = 2) were elevated (ΓPCB: ratio = 2.02; 95% CI: 1.10–3.71), but not significantly different from those working in probably exposed jobs for ≥15 years (n = 5) (ΓPCB: ratio = 1.81; 95% CI: 1.22–2.69). For ΓPCBs, 38% of the variability was explained by the probability rating after adjusting for case-control status, age, sex, and study site. Analyses restricted to controls yielded similar findings.

Table 1.

Association between serum PCB concentration and retrospective occupational exposure probability rating.

	Measured above detection limit (%)	Ratioa (95% CI)
		Probability rating
		Unexposed	Possible	Probable
ΓPCBb		1.00	0.91 (0.73–1.13)	1.87 (1.33–2.62)
Median (ng g −1 lipid)		314.8	287.1	528.7
PCB 74	84	1.00	0.95 (0.72–1.25)	1.72 (1.10–2.68)
Median (ng g −1 lipid)		13.3	9.6	16.7
PCB 99	68	1.00	0.92 (0.63–1.35)	1.54 (0.85–2.80)
Median (ng g −1 lipid)		9.3	6.8	12.2
PCB 118	80	1.00	0.84 (0.55–1.28)	2.02 (1.12–3.65)
Median (ng g −1 lipid)		13.3	8.1	22.3
PCB 126	79	1.00	0.57 (0.33–0.97)	2.60 (1.23–5.48)
Median (ng g −1 lipid)		30.1	15.8	80.4
PCB 138/158	99	1.00	0.95 (0.72–1.26)	1.86 (1.18–2.94)
Median (ng g −1 lipid)		38.2	36.7	64.0
PCB 146	54	1.00	1.10 (0.82–1.47)	1.51 (0.99–2.28)
Median (ng g −1 lipid)		6.1	6.9	11.1
PCB 153	98	1.00	0.91 (0.69–1.19)	1.90 (1.23–2.95)
Median (ng g −1 lipid)		54.0	52.8	84.6
PCB 156	68%	1.00	1.06 (0.82–1.36)	1.49 (1.35–2.16)
Median (ng g −1 lipid)		7.7	7.8	13.2
PCB 169	89	1.00	0.93 (0.73–1.17)	1.49 (1.02–2.17)
Median (ng g −1 lipid)		27.9	26.4	49.4
PCB 170	97	1.00	0.92 (0.74–1.15)	1.72 (1.22–2.44)
Median (ng g −1 lipid)		17.9	17.6	34.2
PCB 180	99	1.00	0.95 (0.75–1.20)	1.73 (1.19–2.51)
Median (ng g −1 lipid)		42.7	45.0	77.1
PCB 183	28c	1.00	0.75 (0.45–1.23)	2.52 (1.13–5.60)
Median (ng g −1 lipid)		4.4	3.5	8.6
PCB 187	93	1.00	0.92 (0.71–1.18)	2.19 (1.47–3.25)
Median (ng g −1 lipid)		12.8	12.4	25.0
PCB 194	89	1.00	0.95 (0.76–1.19)	1.63 (1.14–2.34)
Median (ng g −1 lipid)		11.6	12.5	26.4
Γnoncoplanar PCBd		1.00	0.94 (0.75–1.19)	1.83 (1.28–2.62)
Median (ng g −1 lipid)		249.6	224.4	394.4
Γcoplanar PCBe		1.00	0.77 (0.59–1.00)	1.89 (1.24–2.88)
Median (ng g −1 lipid)		59.3	46.3	134.3

^aRatios are the exponentiated beta coefficients from regression models. Ratios represent the relative difference in geometric mean PCB levels from the reference category (unexposed), adjusted for case-control status, age, sex, and study site.

^b Sum of PCB congeners 74, 99, 118, 126, 138–158, 146, 153, 156, 169, 170, 180, 183, 187, and 194.

^cPCB congener 183 was detected in >30% of samples in the entire study population.

^d Sum of noncoplanar PCB congeners 74, 99, 118, 138–158, 146, 153, 156, 170, 180, 183, 187, and 194.

^e Sum of coplanar PCB congeners 126 and 169.

Probably exposed participants’ mean intakes of potentially contaminated food items were generally similar or lower than intakes among unexposed participants: fish (probably exposed = 4.9g day ⁻¹ versus unexposed = 5.2g day ⁻¹ ), shellfish (0.0g day ⁻¹ versus 1.9g day ⁻¹ ), dairy (198.2g day ⁻¹ versus 271.5g day ⁻¹ ), and red meat (54.9g day ⁻¹ versus 94.3g day ⁻¹ ). Dietary intakes for possibly exposed participants were also similar to unexposed participants (not shown).

DISCUSSION

In this study, the validity of an expert assessment of occupational PCB exposure in a case-control study was assessed by comparison to serum PCB concentrations, a gold standard exposure metric for cumulative burden of PCBs. We found significantly higher serum PCB levels among probably exposed participants compared to unexposed and did not find any evidence this difference was driven by PCB exposure via dietary sources. In addition, the median serum ΓPCB concentration among probably exposed participants in this case-control study (528.7ng g ⁻¹ lipid) was higher than median levels reported in the general US population from the National Health and Nutrition Examination Survey 2001–2002 (160–376ng g ⁻¹ lipid) among individuals ages 20–69 (Nichols et al., 2007). Although probability ratings were assigned by only one expert and we had a small number of occupationally exposed participants, our results provide support that an expert can identify participants who were occupationally exposed to PCBs.

No significant differences in serum PCB levels were observed between possibly exposed and unexposed participants. In instances when prevalence of exposure is low, specificity is more crucial than sensitivity to reducing potential bias to estimates of health risk in epidemiologic studies (Dosemeci and Stewart, 1996; Kromhout and Vermeulen, 2001). Thus, assessments of low prevalence exposures in this study reflect general guidelines to maximize specificity in order to minimize potential misclassification bias. By prioritizing specificity, jobs that are truly possibly exposed may be misclassified as unexposed, which would attenuate differences in serum PCB levels between possibly exposed and unexposed groups. In addition, when there is a high degree of heterogeneity with job exposures or low contrasts between job exposures pose particular challenges to an expert’s ability in distinguishing exposure categories (Chen et al., 2014). These difficulties may be exacerbated in case-control studies where information to capture within job exposure variability can be limited.

Higher PCB levels observed in participants who started probably exposed jobs later than 1965 may reflect temporal changes in occupational exposure patterns. These were mostly electricians in industrial settings who were likely to service and decommission equipment possibly containing PCBs. In part, because PCBs were generally contained in sealed compartments, replacement was required only when the performance of the liquid deteriorated. Exposures therefore occurred both in the normal repair of this equipment and when these sealed compartments degraded over time causing them to leak the PCBs. The higher PCB levels among those with shorter (<15 years), rather than longer (≥15 years) durations in probably exposed jobs, although not significant, may be because these participants were observed to have later start dates for probably exposed jobs and/or a result of natural variability among the small number of probably exposed participants. Given our limited sample size and the long half-life of PCBs in serum, which can range from a few years to over 40 years for heavily chlorinated congeners (Seegal et al., 2011), it is not surprising that we did not observe differences in serum levels by the year a participant stopped working a probably exposed job.

In summary, despite a small number of occupationally exposed participants and only 38% of total variance (ΓPCB) occurring between exposure groups, serum measurements of PCBs support the industrial hygienist’s retrospective assessment of occupational PCB exposure by job classification in this case-control study.

FUNDING

This study was funded by the Intramural Research Program of the Division of Cancer Epidemiology and Genetics, National Cancer Institute, National Institutes of Health (Z01 CP10122; Z01 CP010120).