Case-control studies on risk factors for myeloid leukemias and myelodysplastic syndromes

Sammanfattning: Case-control studies comprising telephone interviews for 333 cases of acute myeloid leukemia (AML), 226 cases of Philadelphia chromosome-positive chronic myeloid leukemia (Ph+ CML), 330 cases of myelodysplastic syndromes (MDS), and matched controls were conducted. Information had to be obtained from next-of-kin much more often for cases (85%) than for controls (18%). Tobacco smoking was associated with an elevated risk for MDS and, at high cumulative smoking doses, with AML. In contrast, no association with Ph+ CML was discerned. Among nine other exposures evaluated for AML and Ph+ CML (but not for MDS) based on the interview data, consistently increased risks were only observed for exposure to organic solvents. Analyses based on registry data combined with a Swedish re-coding of the Finnish job exposure matrix (FINJEM) did not confirm the findings for organic solvents. This discrepancy may be attributed to underestimated exposure proportions for low-prevalent exposures in FINJEM. For MDS, and possibly for AML, specific associations between smoking and aberrations involving chromosome 7 were suggested. A strong effect from organic solvents on AML with trisomy 8 as sole aberration was observed. The methodological aspects mainly concerned partially ecologic case-control studies, a study setting where group-level exposure data, obtained from an exposure database such as FINJEM are combined with individual-level data on disease status, group membership, and covariates. If the exposure is binary on the individual-level, then the corresponding group-level measure is the exposure proportion (exposure probability). Confidence limits calculated for the log transformed OR under a linear odds ratio (OR) model lead to satisfactory coverage in simulated scenarios without errors in the estimated exposure probabilities. An estimator of the attributable fraction (AF) with 95% confidence intervals was proposed and was found to perform well. Assessment errors that systematically distort the estimated exposure probabilities may produce severe bias of the effect estimates under the linear OR model in either direction. The corresponding AF estimates are generally unbiased unless exposed subjects are present in occupational groups assessed as unexposed, which may lead to pronounced bias towards the null.