| Gene-environment interaction in the etiology of disease is a topic of on-going interest. While there has been increasing use of the case-only study design to investigate gene-environment interaction in cancer, as well as other disease areas, concerns about the underlying assumption that the genetic and environmental exposures are independent in the underlying population (the independence assumption) have not been adequately addressed. The case-only study design requires only cases, no population controls or cohort, to estimate statistical interaction. This design has obvious cost advantages, as well as some methodological and ethical advantages. However, for results to be valid the independence assumption must be met. There has been little investigation into the frequency and magnitude of independence assumption violation for DNA repair genes and smoking, an interaction of particular interest in cancer. Nor have optimal methods for validating the independence assumption received much attention.Empirical data of two types were used to evaluate the independence assumption for selected genetic variants and smoking behavior. A systematic review of the literature identified 55 studies that presented the joint distribution of smoking and SNPs in 3 DNA repair genes (XRCC1 Arg399Gln, Arg194Trp, or Arg280His, XPD Lys751Gln, and Asp312Asn, and XRCC3 Thr241Met). Measures of smoking were ever/never smoking, current/not current smoker, duration of smoking (<=10 years, 11-20 years, >20 years), intensity (<1/2 pack/day, ½-1 pack/day, >1 pack/day), and pack-years (<=35 pack-years, >35 pack-years). The odds ratio for SNP-smoking association in controls (ORz) was used to estimate the gene-environment association in the underlying population. Results showed that ORz was not reliably null for any of the SNP-smoking combinations. Studies with XRCC1 399/ever-never smoking and XPD 751/pack-years were too heterogeneous for summary estimates [ranges, ORz (95% confidence interval (CI)): 0.7 (0.4, 1.2) -- 1.9 (1.2, 2.8) and 0.8 (0.5, 1.3) -- 2.3 (0.8, 6.1), respectively). In addition, estimates for studies considered homogeneous (Cochran's Q p-value <0.10) varied 2- to 5-fold within meta-analysis. No study characteristics were identified that could explain heterogeneity.Data from two population-based control groups, the Carolina Breast Cancer Study and the North Carolina Colon Cancer Study, were used to evaluate the independence assumption for smoking and a panel of eight metabolic and 26 DNA repair genes plausibly related to smoking behavior. ORz was not consistent across smoking measures precluding the use of one smoking measure (e.g. ever-never) as a substitute for evaluating other measures such as duration and dose. In particular, results for smoking status were most often near the null, while measures of smoking amount for the same SNPs were of sufficient magnitude to cause appreciable bias in the case-only estimates (ORz<=0.7 or >=1.4) approximately half of the time. There were no strong patterns of the magnitude or direction of ORz differing by race, age, gender or biological pathway (xenobiotic metabolism, DNA repair).Taken together, results suggest that ORz should be considered population-specific. Therefore, the independence assumption should be evaluated in the population underlying a case-only study, rather than in a proxy control group(s) or pooled controls. A systematic search for relevant literature and control data, in addition to a comprehensive evaluation of all smoking measures used in the case-only analysis are essential for evaluation of the independence assumption. |
