Occupational Exposure to Benzene and Non-Hodgkin Lymphoma in a Population-Based Cohort: The Shanghai Women’s Health Study

Background The association between benzene exposure and non-Hodgkin lymphoma (NHL) has been the subject of debate as a result of inconsistent epidemiologic evidence. An International Agency for Research on Cancer (IARC) working group evaluated benzene in 2009 and noted evidence for a positive association between benzene exposure and NHL risk. Objective We evaluated the association between occupational benzene exposure and NHL among 73,087 women enrolled in the prospective population-based Shanghai Women’s Health Study. Methods Benzene exposure estimates were derived using a previously developed exposure assessment framework that combined ordinal job-exposure matrix intensity ratings with quantitative benzene exposure measurements from an inspection database of Shanghai factories collected between 1954 and 2000. Associations between benzene exposure metrics and NHL (n = 102 cases) were assessed using Cox proportional hazard models, with study follow-up occurring from December 1996 through December 2009. Results Women ever exposed to benzene had a significantly higher risk of NHL [hazard ratio (HR) = 1.87, 95% CI: 1.19, 2.96]. Compared with unexposed women, significant trends in NHL risk were observed for increasing years of benzene exposure (ptrend = 0.006) and increasing cumulative exposure levels (ptrend = 0.005), with the highest duration and cumulative exposure tertiles having a significantly higher association with NHL (HR = 2.07, 95% CI: 1.07, 4.01 and HR = 2.16, 95% CI: 1.17, 3.98, respectively). Conclusions Our findings, using a population-based prospective cohort of women with diverse occupational histories, provide additional evidence that occupational exposure to benzene is associated with NHL risk. Citation Bassig BA, Friesen MC, Vermeulen R, Shu XO, Purdue MP, Stewart PA, Xiang YB, Chow WH, Zheng T, Ji BT, Yang G, Linet MS, Hu W, Zhang H, Zheng W, Gao YT, Rothman N, Lan Q. 2015. Occupational exposure to benzene and non-Hodgkin lymphoma in a population-based cohort: the Shanghai Women’s Health Study. Environ Health Perspect 123:971–977; http://dx.doi.org/10.1289/ehp.1408307


Introduction
Benzene is a common occupational solvent that has been used in multiple industries, particularly as a chemical intermediate in the production of plastics and in rubber manu facturing, and has been used as a solvent in common consumer products worldwide (Wilbur et al. 2008). Moreover, exposure to benzene is ubiquitous in the general popu lation because it is present in gasoline, auto mobile emissions, and cigarette smoke (Wilbur et al. 2008). There has been continued concern regarding potential adverse health effects arising from airborne benzene exposures, particularly in light of evidence that benzene is hemato toxic at exposure levels as low as < 1 ppm (Lan et al. 2004;Qu et al. 2002) and is associated with an increased risk of myelo dysplastic syndrome at relatively low exposure levels (Schnatter et al. 2012). In China, benzene is regularly used in manufacturing industries, and historically high levels of occu pational benzene exposure have been reported, with a number of reports describing cases of benzene poisoning in exposed Chinese workers (Liu et al. 2009).
Exposure to benzene, which is classified by the International Agency for Research on Cancer (IARC) as a Group 1 carcinogen, has been concluded to cause acute myeloid leukemia (AML), based in part on results from several case-control and occupational cohort studies in a variety of human popula tions (IARC 2012). In addition, a variety of other adverse health effects associated with benzene have been observed in animal and human studies, including developmental, neurological, and hemato poietic toxicities (Wilbur et al. 2008). Benzene is also a suspect lymphomagen based on positive associations with nonHodgkin lymphoma (NHL) overall or with specific NHL subtypes observed in some occupational cohort and case-control studies, although the strength of the evidence is not as strong as the association with AML (Baan et al. 2009;IARC 2012). In particular, a significant exposure-response association with NHL was previously identified for benzene exposure in a large occupational cohort study of benzeneexposed workers in China (Hayes et al. 1997). Several more recent case-control studies conducted in the United States and in Europe have provided some additional evidence for an associa tion with NHL overall (Dryver et al. 2004;FabbroPeray et al. 2001) or with specific subtypes including follicular and diffuse large Bcell lymphoma and chronic lymphocytic leukemia (CLL) (Costantini et al. 2008;Miligi et al. 2006;Wang et al. 2009). Based in part on these epidemiologic studies, an IARC working group recently concluded that there is evidence for a positive association between benzene exposure and NHL risk in humans (IARC 2012).
To our knowledge, no previous studies have evaluated the association between benzene exposure and NHL in the context of a populationbased prospective cohort study Background: The association between benzene exposure and non-Hodgkin lymphoma (NHL) has been the subject of debate as a result of inconsistent epidemiologic evidence. An International Agency for Research on Cancer (IARC) working group evaluated benzene in 2009 and noted evidence for a positive association between benzene exposure and NHL risk. oBjective: We evaluated the association between occupational benzene exposure and NHL among 73,087 women enrolled in the prospective population-based Shanghai Women's Health Study. Methods: Benzene exposure estimates were derived using a previously developed exposure assessment framework that combined ordinal job-exposure matrix intensity ratings with quantitative benzene exposure measurements from an inspection database of Shanghai factories collected between 1954 and 2000. Associations between benzene exposure metrics and NHL (n = 102 cases) were assessed using Cox proportional hazard models, with study follow-up occurring from December 1996 through December 2009. results: Women ever exposed to benzene had a significantly higher risk of NHL [hazard ratio (HR) = 1.87, 95% CI: 1.19, 2.96]. Compared with unexposed women, significant trends in NHL risk were observed for increasing years of benzene exposure (p trend = 0.006) and increasing cumulative exposure levels (p trend = 0.005), with the highest duration and cumulative exposure tertiles having a significantly higher association with NHL (HR = 2.07, 95% CI: 1.07, 4.01 and HR = 2.16, 95% CI: 1.17, 3.98, respectively). conclusions: Our findings, using a population-based prospective cohort of women with diverse occupational histories, provide additional evidence that occupational exposure to benzene is associated with NHL risk.  (Friesen et al. 2012). This approach combined a jobexposure matrix (JEM) with shortterm area air measurements of benzene in factories in Shanghai over the period 1954-2000, and has been demonstrated to increase contrast in exposure levels between different indus tries and jobs, as well as over time, compared with using the JEM alone (Friesen et al. 2012). Here, we applied this exposure assess ment framework to evaluate the association between occupational exposure to benzene and risk of NHL in this populationbased cohort consisting of Chinese women living in Shanghai.

Study population and outcome ascertainment.
Details concerning the study population enrolled in the Shanghai Women's Health Study (SWHS) have been described previ ously (Zheng et al. 2005). Briefly, the study is a prospective cohort of 74,942 women living in urban Shanghai who were enrolled between the ages of 40 and 70 years, with subject enrollment occurring between December 1996 and May 2000. Eligible women for the study were identified from a roster of all women from the resident offices in the study communities and were approached by a trained interviewer and a local community health worker. The overall participation rate for the study was 92.7% (defined as women who agreed to participate and completed the baseline survey among all eligible women who lived in the study communities during the time period of the baseline survey). Enrolled women completed a lifetime occupational history questionnaire that assessed job title, name of the work place, type of process or business, description of work tasks, and employment dates for all jobs held for at least 1 year. Other components of the questionnaire included an assessment of demographic char acteristics, personal lifestyle habits, medical history, and a residential history (Zheng et al. 2005). The Standard Chinese Classification of Industries and Occupations for the third national census (World Bank 2014) was used by investigators from the Shanghai Cancer Institute to code and standardize each of the employer types and freetext job titles obtained from the occupational history questionnaire (Friesen et al. 2012).
All hospitals in Shanghai are legally required to notify the Shanghai Cancer Registry (Shanghai Cancer Institute 2015) of all newly diagnosed cancer cases. Thus, iden tification of disease outcomes for the study, including incident cancers and vital status, was conducted through annual record linkage to these populationbased Shanghai cancer and vital statistics registries. In addition, inperson followup study surveys that have taken place every 2 to 3 years have collected disease and hospital information. Medical charts were reviewed from each diagnostic hospital to verify the cancer diagnoses and to collect information on tumor characteristics. All cancer site information was based on the International Classification of Diseases, Ninth Revision (ICD9). All participants provided informed consent, and the study was approved by the institutional review boards of all participating institutions.
Exposure assessment. The exposure assess ment used a benzene JEM that was composed of two separate JEMs, one based on occupa tion and one based on industry. The JEM was calibrated with shortterm area benzene exposure measurements in Shanghai factories and has been described in detail elsewhere (Friesen et al. 2012). Briefly, the JEM using the same industry and job coding system as the questionnaire data was developed based on experts' rating of the probability and intensity of benzene exposure using an ordinal scale (coded 0-3) for 1980, the median year of employment in the study population. The probability ratings were assigned based on the estimated proportion of workers exposed to benzene, with scores of 0 to 3 corresponding to no exposure, < 5% exposed, 5-50% exposed, and > 50% exposed, respectively. Intensity ratings were assigned based on the maximum allowable concentration (MAC) for occupational benzene exposure in China in the year 1980 (i.e., 40 mg/m 3 ), with scores of 0 to 3 corresponding to very low or negli gible exposure, above background but < 10% of the MAC, 10-100% of the MAC, and exposures above the MAC, respectively.
The benzene exposure database consisted of shortterm area air measurements (n = 70,937) that were collected as part of health and safety inspections from factories in Shanghai that fell under the Shanghai Centers for Disease Control (CDC) inspection program. Measurements were collected over the period 1954-2000 by the Shanghai CDC and included the sampling date, factory name, and type of industry and job, all of which were abstracted from records maintained at municipal and district stations. Each measure ment was coded to its relevant industry and occupational classification, using the same clas sification system as used in the occupational histories and JEM. Information pertaining to the analytical method that was used to measure benzene exposure was not avail able in the database. As described previously (Friesen et al. 2012), in sensitivity analyses the measurements were calibrated to a single measurement method based on the prevalence of syringe measurement and gas chromatog raphy methods over time and the magnitude of the exposure difference between methods observed in another Shanghai study. The rank order of the benzene estimates derived from separate models adjusting or not adjusting for the analytical method was demonstrated to be very highly correlated. The benzene measure ment data were combined with the JEMs to derive cumulative exposure estimates (milli grams per cubic meter) for each woman as further described below.
The job/industry-specific benzene exposure estimates for each threedigit job/ industry code were derived using mixed effects modeling that combined both the JEM and the inspection measurements, as described elsewhere (Friesen et al. 2012). Briefly, the estimated geometric mean exposure levels were derived from the fixed effect parameters for year (incorporated as a bspline term) and JEM intensity ratings (incorporated as ordinal metrics, 0-3 scale) and from the best linear unbiased prediction estimates for the randomeffects of job group, and industry class nested within job group, which allowed for job and industryspecific deviations from the pooled estimates asso ciated with the JEM ratings (Friesen et al. 2012). Thus, this framework provided a mechanism to account for a 20fold differ ence in exposure concentrations between job/ industry groups, as well as a 13fold difference in exposure concentrations between 1954 and 2000, that would not have been possible to capture using the exposure matrices only (Friesen et al. 2012).
For each woman, we applied the time varying exposure levels predicted by the models above to each year of a job in her work history that was classified as exposed, with exposed jobs defined as those with a job code probability rating of 3 (i.e., > 50% exposed), or as jobs with an industry code probability rating of 3 (i.e., > 50% exposed) and a job code probability rating ≥ 1 (i.e., > 0% exposed). All other jobs were classified as unexposed and assigned a benzene exposure level of zero (Friesen et al. 2012). Occupational histories were available for each woman up until the year of entry into the study.
Statistical analysis. Analyses included all women currently enrolled in the SWHS with no prevalent cancer at baseline and a valid occupational history. A total of 73,087 women were included in the analyses, with followup for each woman starting at the time of completion of the baseline questionnaire. To evaluate the association between the job industry-specific benzene exposure metrics and risk of NHL (ICD9 codes 200, 202, 204.0-204.1), we estimated hazard ratios (HR) and calculated 95% confidence inter vals (CIs) using Cox proportional hazard models using entry and exit age as the under lying time factor. Of the total of 102 patients diagnosed with primary incident NHL during followup, 24 were ever exposed to benzene. Results are presented for NHL overall because of the lack of detailed histology data for NHL subtypes (i.e., 73% of the 102 cases are classified as ICD9 code 202.8, "other lymphoma"). Exit age for the analyses was based on the date of NHL diagnosis, the date of death, or the last followup date, whichever event came first. Given the previously estab lished relationship between benzene exposure and risk of myeloid leukemia (IARC 2012), we similarly evaluated the association between myeloid leukemia and the benzene exposure metrics, although the number of cases in our cohort was small (n = 33 overall, 4 exposed).
Metrics of occupational benzene exposure evaluated in the study included ever/never exposure, duration of exposure in years, and cumulative exposure in milligrams per cubic meter for total years (mg/m 3 years). Women were classified as ever exposed if they had any job/industry combination in their occupational history that was classified as exposed according to our criteria, regardless of duration. Total exposure duration was calcu lated by summing the number of exposed years for each woman (i.e., the number of years spent working in jobs classified as exposed). Cumulative exposure was calculated by summing all yearly exposure estimates for exposed jobs across the full occupational history for each subject. For the duration and cumulative exposure metrics, indicator variables based on the tertiles of exposure in the exposed population were included in the Cox regression models, with women having no exposure used as the reference group in all analyses. For these continuous exposure metrics, the highest tertile was further subdi vided based on the intracategory median to evaluate associations with NHL across a greater range of exposure levels. We tested for a linear trend in the HR for each of the benzene exposure metrics where applicable by including a continuous variable in the model representing the median values of each of the exposure tertiles. The assumption of propor tional hazards was evaluated and verified by a graphical check on the log cumulative hazard versus time and by modeling a product term between the timescale for the model (attained age) and the timefixed benzene exposure variables at baseline.
For each of the metrics described above, results are presented from a model adjusted only for age (i.e., using age as the time scale) and a model additionally adjusted for other covariates. The fully adjusted models included categorical variables for education, body mass index (BMI), ever smoking, and ever use of alcohol. All models were stratified by birth cohort in 5year intervals to account for possible calendar effects. For the myeloid leukemia analyses, results from a model adjusted only for age and a model further adjusted for ever smoking and education are presented for ever exposure, duration (based on the median and tertiles) and cumulative exposure based on the median. Further, we evaluated the association between cumula tive benzene exposure and NHL risk using a 5 and 10year lag by treating cumulative exposure as a timevarying explanatory variable in the model and using age as the time scale, as implemented using the stcox procedure in Stata version 13.1 (StataCorp). All other analyses were conducted using SAS version 9.3 (SAS Institute Inc.), with a twosided α = 0.05 considered statistically significant.

Results
The demographic characteristics at baseline of the 73,087 subjects included in the analyses both overall and stratified by benzene exposure status, are shown in Table 1. Overall, the mean age of women in the cohort at baseline was 52 years, and the mean attained age at the end of followup was 64 years. The mean BMI at baseline among all women was 23 kg/m 2 , and most women did not report ever smoking or alcohol use (≥ 97%). These characteristics were similar in workers exposed to benzene and in the unexposed workers (Table 1). About 86% of women had less than a college education.
Of the 73,087 women included in the analysis, 15% (n = 10,788) were ever exposed to benzene. Approximately 85% of everexposed women did not have an exposed job during the year of cohort entry, which is consistent with the decreasing distribution of exposed personyears from 1980 to 2000 for women in the cohort. The median cumulative exposure level and duration of exposure in exposed workers was 59 mg/m 3 years (75th percentile: 136 mg/m 3 years; 95th percentile: 491 mg/m 3 years) and 17 years, respectively. Among workers exposed to benzene, the predominant industries of employment in terms of personyears from the occupa tional histories included manufacturing of rubber products (13%), organic chemicals (10%), motor vehicles (8%), miscellaneous electronics (7%), and television sets and audio equipment (6%). The majority of exposed NHL cases (~ 71%) were exposed before 1990.
The mean and median length of time between first year of exposure to benzene and NHL diagnosis were both approximately 38 years (range, 13-59 years). Associations with NHL for each of the benzene exposure metrics are shown in Table 2 for the models adjusted only for age, as well as for the fully adjusted models. For all of the benzene metrics evaluated, further adjustment of the models for education, ever smoking, ever use of alcohol, and BMI had little or no impact on the results for NHL (Table 2, model 2); therefore, HRs and 95% CIs described here are based on the models with age as the time factor and no further adjustment for covariates. Women ever exposed to benzene had a significantly higher risk of NHL relative to unexposed women (HR = 1.87, 95% CI: 1.19, 2.96). Table 1. Demographic characteristics of women enrolled in the Shanghai Women's Health Study included in the benzene exposure analysis.

Characteristic
Total n = 73,087 Exposed n = 10,788 Not exposed n = 62,299 Age at baseline (mean ± SD) 52 ± 9 51 ± 9 52 ± 9 Attained age after follow-up (mean ± SD) 64 ± 9 63 ± 9 64 ± 9 BMI at baseline (mean ± SD) 23 ± 3 23 ± 3 23 ± Significantly higher associations of similar magnitude were apparent for women in the highest two duration tertiles compared with unexposed women (tertile 2, HR = 2.10, 95% CI: 1.01, 4.35; tertile 3, HR = 2.07, 95% CI: 1.07, 4.01; p trend = 0.006), but the estimated risk for women in the lowest tertile was not significantly different from unexposed women (Table 2). Further strati fication of the top duration tertile (n = 10 exposed cases) based on the intra category median revealed that the majority of the cases within this tertile were exposed to benzene for > 27 years. A significantly higher risk of NHL was observed in the women who were exposed for the longest period of time (i.e., > 27 years; HR = 2.65, 95% CI: 1.21, 5.77), but no significantly higher risk was observed in women in the lower half of the top tertile, which only included three exposed cases (Table 2).
Compared with unexposed women, the risk of NHL was also significantly higher for women in the two highest tertiles of cumula tive benzene exposure (tertile 2, HR = 2.22, 95% CI: 1.12, 4.44; tertile 3, HR = 2.16, 95% CI: 1.17, 3.98; p trend = 0.005), whereas there was no significant association for women in the lowest cumulative exposure tertile (Table 2). Similar to results for the duration metric, further stratification of the highest tertile showed that the largest magni tude of risk for NHL was in women with cumulative exposures greater than the median exposure (HR = 2.31, 95% CI: 1.06, 5.02) within the highest tertile, compared to unexposed women (Table 2). A similar though non significant higher risk of NHL was observed for women with exposure levels in the lower half of the top tertile (HR = 1.98, 95% CI: 0.80, 4.91). Further, results for cumulative exposure were gener ally consistent in analyses using a 5 and 10year lag, with the highest risks of similar magnitude observed in the second (5year lag, HR = 2.08, 95% CI: 1.00, 4.33; 10year lag, HR = 2.16, 95% CI: 1.04, 4.49) and third tertiles (5year lag, HR = 2.16, 95% CI: 1.17, 4.00; 10year lag, HR = 2.04, 95% CI: 1.08, 3.86) compared with unexposed women, whereas there was no significant asso ciation with NHL among women in the first tertile for either analysis. Significant trends were observed in both analyses (p trends = 0.007 and 0.02 for 5 and 10year lags, respectively) (see Supplemental Material, Table S1).
A higher risk of myeloid leukemia was not observed based on four women ever exposed to benzene (see Supplemental Material, Table S2). However, the four benzeneexposed myeloid leukemia cases in our cohort were all highly exposed (i.e., greater than median cumula tive exposure and exposed > 21 years) and a higher risk of myeloid leukemia was observed for women in the highest exposure duration tertile (HR = 2.62, 95% CI: 0.92, 7.50) and for workers with cumulative exposure levels higher than the median of 59 mg/m 3 years (HR = 1.65, 95% CI: 0.58, 4.71). These risks were slightly attenuated after further adjust ment for ever smoking and education (see Supplemental Material, Table S2).

Discussion
To our knowledge, this is the first study to investigate the risk of NHL in relation to benzene exposure in the context of a populationbased prospective cohort study of women with diverse occupational histories and job tasks. We observed that women who were ever exposed to benzene had a significantly higher risk of NHL compared with unexposed women, and significantly increased risks were also observed among women in the highest two tertiles of exposure duration and cumula tive exposure. Our study, which combined job and industryexposure matrices and shortterm area air measurements in Shanghai factories, provides evidence that occupational benzene exposure is associated with NHL among these women from Shanghai.
The lymphomagenic potential of benzene exposure has been the subject of much debate. A recent comprehensive evaluation by IARC concluded that there is evidence in humans for a positive association between benzene and NHL overall, as well as for acute lymphocytic leukemia (ALL), CLL, and multiple myeloma (IARC 2012). Notably, a relative risk of 3 was previously observed for NHL in relation to ever exposure to benzene in a cohort of Chinese workers (Hayes et al. 1997), whereas the majority of other occupational cohorts conducted to date have not observed signifi cantly higher associations for NHL (IARC 2012). The estimated benzene exposure levels in our study were lower, on average, than those reported in the previous Chinese occu pational cohort, which observed the highest risk for NHL associated with cumulative benzene exposures > 100 ppmyears, but were higher, on average, compared to several cohorts conducted in North America and Australia, which reported mean or median cumulative benzene exposures generally < 10 ppmyears (Collins et al. 2003;Glass et al. 2003;Schnatter et al. 1996). It is also a HRs (hazard ratios) and 95% CIs estimated from Cox proportional hazard models with entry and exit age as the time scale. b HRs and 95% CIs estimated from Cox proportional hazard models with entry and exit age as the time scale, and further adjusted for ever smoking, ever use of alcohol, BMI, and education. c Women were defined as ever exposed if they had any job/industry combination in their occupational history that was classified as exposed according to our specified criteria, regardless of duration. d A linear trend in the HR for each of the benzene exposure metrics was evaluated by including a continuous variable in the model representing the median values of each of the exposure tertiles.
notable that most previous cohorts have not included lymphoid leukemia cases in their NHL case definition, whereas we included cases of both CLL/small lymphocytic lymphoma and ALL, which are grouped with NHL in the most recent classifications (Turner et al. 2010). Differences between cohorts with respect to evaluating NHL incidence versus mortality also present challenges when comparing findings across studies. Recent case-control studies conducted in Caucasian populations have also provided some sugges tive evidence for an association between benzene exposure with either NHL overall or with specific NHL subtypes (Costantini et al. 2008;Dryver et al. 2004;FabbroPeray et al. 2001;Miligi et al. 2006;Wang et al. 2009). A previous hospitalbased case-control study of 649 NHL cases and 1,298 controls in Shanghai found a significantly higher risk of follicular lymphoma for ever exposure to benzene based on a limited number of cases but not for other subtypes or NHL overall (Wong et al. 2010).
A particular challenge in the evaluation of factors associated with NHL risk is the increasing evidence of etiologic heterogeneity for specific NHL subtypes (Morton et al. 2008). To that end, a recent metaanalysis of occupational cohort studies evaluating benzene exposure and cancer risk found evidence for an association with CLL and ALL, while an asso ciation with NHL overall was less pronounced (Vlaanderen et al. 2011). Further, evidence from some case-control studies has indicated that solvent use more generally may be associ ated with some NHL subtypes but not others (Cocco et al. 2010;Wang et al. 2009). Because we lacked detailed histology data on the NHL cases in our cohort, we were unable to evaluate NHL subtypespecific associations.
Exposure to benzene has been demon strated in animal studies to result in develop ment of lymphomas, including in one study of Trp53deficient mice that also observed an increase in benzeneinduced AML (IARC 2012;Kawasaki et al. 2009). The IARC working group evaluating benzene exposure concluded that there are two probable mecha nisms of benzeneinduced lymphoma genesis in humans, namely, immuno suppression leading to decreased immuno surveillance and chromosomal rearrangements (IARC 2012). In support of these conclusions, studies of benzene have shown a variety of genotoxic effects in humans resulting from exposure, including both structural and numerical chro mosomal abnormalities; these genotoxic effects appear to affect the lymphocytes of exposed populations (IARC 2012;Zhang et al. 2002). Previous molecular epidemiologic investiga tions conducted in the Chinese population in particular, including among healthy workers in Shanghai (Rothman et al. 1996), have demonstrated that benzene is associated with cyto genetic abnormalities in cultured peripheral lymphocytes, which are frequently associated with lymphomas, including the long arm deletion of chromosome 6 (Zhang et al. 2007). Further, we have shown among workers in China that benzene exposure is associated with increased mitochondrial DNA (mtDNA) copy number (Shen et al. 2008). Interestingly, increased mtDNA copy number was positively associated with NHL in a dosedependent manner in a prospective cohort of Finnish men, and thus the effect of benzene on this marker may represent an early biologic effect that is relevant to the subsequent development of NHL (Lan et al. 2008). Finally, a crosssectional study of occu pationally exposed workers in China found a decrease in lymphocyte cell counts including CD4 + T cells, which has been associated with an increased risk of NHL, at benzene expo sures as low as < 1 ppm, with greater reduc tions observed for higher levels of benzene (Lan et al. 2004). The inverse association of benzene exposure on levels of CD4 + T cells is particularly notable given the established rela tionship between immune deficiency, various immune disorders, and NHL risk (Grulich et al. 2007). The increased risk of NHL associ ated with immune deficiency may result from loss of viral control resulting from reduced immuno surveillance or due to dysregulation of Blymphocyte activity (Grulich et al. 2007). Although the relationship between subclinical immune deficiency and NHL is not as clear, the apparent lack of a threshold between levels of immuno deficiency and subsequent NHL risk (Grulich et al. 2007), as well as more recent prospective studies suggesting that subclinical alterations in various immune markers are associated with NHL (De Roos et al. 2012;Purdue et al. 2013), provides evidence for a role of more subtle immune alterations in lymphomagenesis.
The most recent IARC assessment of benzene also confirmed that there continues to be sufficient evidence in humans for an association with AML (IARC 2012). The number of exposed myeloid leukemia cases in our cohort was small (n = 4), and only 2 of these were AML; therefore, we had limited ability to evaluate this association. Because occupational benzene exposure had ceased for the vast majority of cohort members ever exposed to benzene by the time of enrollment into the cohort, it is possible that deaths due to benzenerelated AML had already taken place among people living within the cohort catchment area well before enrollment began. This possibility is based on some evidence suggesting that benzene exposure during the 5 years prior to diagnosis may be the most important for determining risk for benzene related AML (Hayes et al. 1997).
A strength of our populationbased study of Chinese women was the use of an exposure assessment framework with improved contrast in exposure levels across time and between jobs/industries compared with using the JEM alone (Friesen et al. 2012). Second, we used stringent definitions for defining exposure, and our a priori criteria were chosen to maximize specificity in the exposure assessment, as recommended previ ously (Kromhout and Vermeulen 2001). A further strength of our study was the acquisition of a full occupational history for women in the cohort and the availability of questionnaire data for women in the SWHS, although we lacked data on potential coexpo sures in the workplace and therefore cannot exclude the possibility of confounding from other chemical exposures. Nevertheless, the SWHS includes women with diverse job tasks across a variety of industries; thus, the common use of a confounding chemical(s) may be less likely, compared with occupa tional cohorts that include subjects employed in only one or a few industries with similar exposure patterns.
Given that the exposure measurements from the Shanghai database used in our framework reflected shortterm area samples that were collected for health and safety inspection purposes, as opposed to using fullshift personal monitors, some degree of exposure misclassification is likely, as has been discussed in detail elsewhere (Friesen et al. 2012). The use of area measurements could result in either an over or under estimation of an individual worker's exposure level. The exposure measurements used in our study were from factories included in the Shanghai CDC inspection program, which did not include all factories in Shanghai. However, there were a relatively large number of measurements from this database that were below the limit of detection (~ 50%). Moreover, our exposure estimates were, on average, lower than those from some previous studies conducted in China that measured benzene exposure in factories under inves tigation for benzenepoisoning events or factories with suspected highlevel exposures under consideration for use in molecular epidemiology studies (Friesen et al. 2012). These observations provide some evidence that the database reflected a wide range of exposure scenarios rather than worstcase scenarios only. Comparisons between the experts' intensity estimates and our model predictions for the year 1980 found that the predicted levels for job intensity ratings of 1, 2, and 3 were 4 mg/m 3 , 6 mg/m 3 , and 11 mg/m 3 , respectively, which were consistent with the expert cut points for intensity ratings of 1 [arithmetic mean (AM), < 4 mg/m 3 ] and 2 (AM, 4-40 mg/m 3 ) but were lower volume 123 | number 10 | October 2015 • Environmental Health Perspectives than the cut point for intensity rating 3 (AM > 40 mg/m 3 ) (Friesen et al. 2012). This may not be surprising because the measurements would reflect to some extent both the inten sity of the exposure and the probability of exposure, whereas the cut points were based on intensity of exposure only. However, our exposure assessment approach that combined the JEM with the benzene exposure measure ments provided a mechanism to calibrate the ratings to a concentration scale and across time, and thus the measurement driven values-rather than the theoretical cut points-were used.
Moreover, the measurements applied in our study were not directly associated with any individual subject but rather were used in the models to estimate geometric mean exposure levels for a given job/industry. Berkson error resulting from the assignment of grouplevel estimates to the individual level has been demonstrated to result in minimal bias to the risk estimates in the presence of small measurement error (Küchenhoff et al. 2007). Second, because occupational history data in the SWHS was not available after the cohort entry year, the estimated benzene exposure levels would therefore slightly under estimate cumulative exposure and duration of exposure for the minority (n = 1,822) of benzeneexposed women who still had an exposed job at the start of followup, if these women had continued exposure during the followup period. For these reasons, we placed greater weight on reporting asso ciations with NHL based on the rankorder exposure tertiles rather than emphasizing specific exposure levels in mg/m 3 years. This analytic approach would be less sensitive to slight under estimates in exposure relative to associations based on continuous analyses. Moreover, this approach combined with the decreasing personyears and concentration of benzene exposure over time observed in our study (Friesen et al. 2012), as well as the fact that elevated associations with NHL were still apparent in the lagged analyses, suggest that any continued exposure during the followup period is unlikely to have materially changed our conclusions. Third, the limited histology data in the cohort precluded an evaluation of the risk of specific NHL subtypes in relation to benzene exposure. Thus, given that subtype specific associations have been suggested in some previous case-control studies (Cocco et al. 2010;Wang et al. 2009), future studies that are able to obtain tumor samples on a high percentage of cases and that are able to conduct pathology review are needed to deter mine whether associations with specific NHL subtypes may be driving the observed asso ciation with overall NHL risk. Finally, given that NHL is a rare tumor in China, another limitation is the relatively small number of NHL cases in the cohort, which contrib uted to relatively broad confidence intervals associated with the observed effect estimates. Nonetheless, we note that the number of benzeneexposed cases is similar or larger to those of other cohorts that have evaluated this association (IARC 2012); despite the relatively small number of exposed cases, we observed significant elevations in NHL risk for three different benzene metrics, including ever exposure.

Conclusions
We estimated the relative risk of NHL in a populationbased cohort study of Chinese women living in Shanghai using an exposure assessment framework that maximized infor mation from two sources of exposure infor mation. We found evidence for an increased risk of NHL for women ever exposed to benzene and for those with higher cumula tive exposure levels and durations of occu pational benzene exposure. Our findings are consistent with several previous studies that have found an elevated risk of NHL overall or of specific NHL subtypes associated with benzene exposure, and they provide addi tional evidence that occupational exposure to benzene is associated with NHL risk.