The number of case subjects and mean cumulative benzene exposures for the five LH subtypes and matched control subjects are shown in . Benzene exposure was relatively low, with average cumulative exposure below 10 ppm-years derived from mean exposure intensities of 0.2–0.3 ppm and average durations of exposure close to 20 years. Median cumulative exposure was higher in MDS case subjects vs control subjects (median = 3.4 [interquartile range (IQR) = 0.4 to 8.9] ppm-years vs 1.4 [IQR = 0.2 to 3.8] ppm-years), but the same in CLL case subjects vs control subjects (median = 1.1 [IQR = 0.2 to 4.4] ppm-years vs 1.1 [IQR = 0.2 to 4.5] ppm-years). Average exposure certainty scores were the same in all case subjects combined and all control subjects combined (score = 2.2), but differed by study (Australian, score = 2.9; Canadian, score = 2.7; UK, score = 1.5). Approximately 74% of case subjects had a medium or high diagnostic certainty score, with CML and CLL outcomes having greater certainty than MPD and MDS (data not shown).
We analyzed LH subtype risks relative to different benzene exposure metrics by tertiles of exposure (). The association of cumulative exposure with LH subtype increased monotonically for AML (0.348–2.93 vs ≤0.348 ppm-years [referent], OR = 1.04 [95% CI = 0.50 to 2.19]; ≥2.93 vs ≤0.348 ppm-years [referent], OR = 1.39 [95% CI = 0.68 to 2.85]; P
global = .62), MDS (0.348–2.93 vs ≤0.348 ppm-years [referent], OR = 1.73 [95% CI = 0.55 to 5.47]; ≥2.93 vs ≤0.348 ppm-years [referent], OR = 4.33 [95% CI = 1.31 to 14.3]; P
global =.03), and MPD (0.348–2.93 vs ≤0.348 ppm-years [referent], OR = 1.28 [95% CI = 0.47 to 3.48]; ≥2.93 vs ≤0.348 ppm-years [referent], OR = 1.79 [95% CI = 0.68 to 4.74]; P
global =.49) (, ). The trend was statistically significant for MDS (P
trend = .01). Dose–response trends were weaker in tertile analyses of other exposure metrics (, ), although MDS was the only outcome that showed consistent monotonic trends for all metrics (, ). Peak exposures greater than 3 ppm showed an increased risk of MDS (ever peak exposure >3 ppm vs never peak exposure >3 ppm, OR = 2.48 [95% CI = 0.97 to 6.35]) (, ) but were unremarkable for other LH cancer subtypes (, ). When we restricted exposures to 2–15 years before diagnosis (recent exposures), the associations did not strengthen for AML, CML, and MDS (eg, third tertile cumulative exposure ≥2.93 ppm-years for all exposures vs recent exposures: for AML, OR = 1.39 [95% CI = 0.68 to 2.85] vs OR = 1.11 [95% CI = 0.37 to 3.34]; for CML, OR = 2.20 [95% CI = 0.63 to 7.68] vs OR = 1.70 [95% CI = 0.17 to 16.9]; and for MDS, OR = 4.33 [95% CI = 1.31 to 14.3] vs OR = 2.04 [95% CI = 0.40 to 10.3]). In contrast, recent exposures did strengthen the association between cumulative exposure and MPD (eg, third tertile cumulative exposure ≥2.93 ppm-years for all exposures vs recent exposures: OR = 1.79 [95% CI = 0.68 to 4.74] vs OR = 3.66 [95% CI = 0.81 to 16.6]), although dose–response relationships were generally not monotonic and results were less stable because of exclusion of a large fraction of the work history (data not shown).
Figure 2. Funnel plots of dose–response relationships between five lymphohematopoietic (LH) cancer subtypes and six discrete benzene exposure metrics. Metrics were calculated over the entire work history as well as a 2–15-year exposure window for (more ...)
The P-spline curves for the five LH subtypes showed monotonic dose–response relationships for MDS for all exposure metrics (, ) and reached statistical significance for maximum exposure intensity (P
spline = .03) (, ). No dose–response relationship was observed for CLL, except for a statistically significant association with duration of employment (P
spline = .03); the dose–response curve reached a plateau after 15 years (, ). The dose–response curves for AML, CML, and MPD did not show a compelling relationship with benzene exposure (, ), although the cumulative exposure metrics for AML and CML indicate a possible relationship (AML, P
spline = .14; CML, P
spline = .12) (, ).
Figure 3. Penalized regression smoothing spline (P-spline) functions showing log odds ratio of risk of lymphohematopoietic (LH) cancer subtypes and specific benzene exposure metrics. We used conditional logistic regression models with P-splines to examine dose–response (more ...)
Because smoking has been associated with LH cancers, we attempted to adjust for its potential effect. For MDS, an ever or never smoker categorization was known for 15 of 29 MDS case subjects and 78 of 129 MDS control subjects. When P-spline analyses were analyzed with smoking as an additional independent variable, the dose–response relationship between benzene exposure and MDS was stronger in workers with known smoking histories compared with all workers (P
spline = .02 vs .07 for cumulative exposure; P
spline = .003 vs .07 for average exposure intensity; and P
spline = .004 vs .03 for maximum exposure intensity), suggesting that smoking is unlikely to be a confounder responsible for the association between MDS and benzene exposure.
We assessed dose–response relationships for peak exposure, dermal exposure, and for the highest cumulative exposure tertile for each of the three studies (Table 1
). Patterns for average and maximum intensity of exposure and duration of exposure (data not shown) were similar to patterns for cumulative exposure. Dose–response relationships between studies were not statistically significantly heterogeneous for the six exposure metrics (eg, for MDS, P
= .18 for peak exposure, .16 for dermal exposure, .30 for cumulative exposure, .96 for maximum exposure intensity, .60 for average exposure intensity, and .92 for duration of exposure); thus data pooling was justified. However, some non- statistically significant differences regarding dose–response relationships between studies (Table 1
) were noted. Specifically, CLL was related to exposure more strongly in the Australian study (eg, third tertile cumulative exposure >2.93 vs ≤0.348 ppm-years [referent], OR = 5.2 [95% CI = 0.98 to 27.0]), and the relationship between peak exposure and MDS was less apparent in the UK study (ever peak exposure >3 ppm vs never peak exposure >3 ppm, OR = 0.80 [95% CI = 0.19 to 3.43]). However, MDS showed consistently increased associations for cumulative exposure in each study (third tertile cumulative exposure >2.93 vs ≤0.348 ppm-years [referent]: Australian study, OR = 3.6 [95% CI = 0.60 to 22]; Canadian study, OR = 3.0 [95% CI = 0.14 to 61]; UK study, OR = 3.4 [95% CI = 0.55 to 21]).
Associations between four LH cancer subtypes and specific benzene exposure metrics by study*
Facility and worker subgroup analyses indicated higher MDS risk at terminals (terminal facility type vs all other facility types, OR = 5.04 [95% CI 1.58 to 16.08]) and among tanker drivers employed for at least a year (ever tanker driver for 1 year vs never tanker driver for 1 year, OR = 2.16 [95% CI = 0.79 to 5.88]). There was also a higher risk of AML for tanker drivers (ever tanker driver for 1 year vs never tanker driver for 1 year, OR = 2.02 [95% CI = 1.08 to 3.78]), whereas refinery operators and craftsmen (primarily from the Australian study) showed a higher CLL risk (ever a refinery operator or craftsman for 1 year vs never a refinery operator or craftsman, OR = 2.26 [95% CI = 0.92 to 5.58]) (Table 2
Associations between the risk of five endpoints and worker subgroup*
For AML, CML, CLL, and MPD, results for more certain case subjects (ie, medium and high diagnostic certainty) and exposures (ie, career weighted average exposure certainty score ≥2) were generally similar vs results for all workers but with wider confidence intervals (data not shown). However, the relationship between MDS and benzene exposure strengthened (ie, a steeper slope and lower P-value were obtained in the dose–reponse curve, despite being based upon only 51% of subjects) for all subjects vs subjects with more certain diagnoses: for cumulative exposure, P
spline = .07 vs P
spline = .02; for average exposure, P
spline = .07 vs P
spline = .03; and for maximum exposure, P
spline = .03 vs P
spline = .02) (, ). Workers having more certain exposures, which accounted for 75% of subjects, showed a higher risk of MDS (ie, a steeper slope in the dose–response curve) for maximum exposure vs all workers, but this result was not statistically significant (P
spline = .063) (, ), whereas other metrics (, , , and ) showed a similar risk of MDS vs all workers.
Figure 4. Penalized regression smoothing spline (P-spline) functions showing log odds ratio of risk of myelodysplastic syndrome (MDS) and specific benzene exposure metrics for more certain cases and more certain exposure history. We defined more certain case subjects (more ...)
When we examined sensitivity results for medium or high certainty diagnoses and jobs with weighted exposure certainty scores of 2 or more, worker subgroups showed some clear patterns despite being based on fewer study subjects. The risk of MDS was statistically significant among tanker drivers (ever a tanker driver for more than 1 year vs never a tanker driver for more than 1 year, OR = 7.2 [95% CI = 1.37 to 37.4]). Also, jobs with peak exposure showed a statistically significant risk for more certain case subjects (peak exposure vs no peak exposure, OR = 6.32 [95% CI = 1.32 to 30.2]) and more certain exposure histories, OR = 5.74 [95% CI = 1.05 to 31.2]. Similar patterns of MDS risk were found among workers with more certain exposure histories (Table 3
Sensitivity analyses on associations between risk of five endpoints vs job and peak exposure*
Models of MDS risk that simultaneously included peak exposure and other exposure metrics suggested that peak exposure was the more robust metric, which means for highly certain case subjects, the P-values for the cumulative exposure term increased when including peak exposure in the model, yet the P-value for peak exposure remained statistically significant and unchanged when including cumulative exposure in the model. We also examined spline models to assess whether a threshold of exposure could be identified for cumulative, average, and maximum exposure. Initial models with unrestricted degrees of freedom suggested a potential threshold at 0.99 ppm maximum exposure, but this value was not confirmed when the degrees of freedom were restricted to more biologically justified values (eg, values that are not prone to overfitting the data). A simple plot of MDS case subject vs control subject exposures seemed to indicate an over-representation of case subjects (beyond the percentage predicted by the baseline matching ratio) starting at approximately 0.7 ppm maximum exposure ().
Plot of MDS case subjects and control subjects by maximum exposure intensity (ppm) and duration of employment (years).