Smoking and Selected DNA Repair Gene Polymorphisms in Controls: Systematic Review and Meta-Analysis

The case-only study design as proposed by Prentice et al. (1) and promoted by Piegorsch et al. (2) has been increasingly used to estimate the magnitude of statistical interaction between 2 measured exposures with respect to a given outcome, most commonly a genetic and an environmental exposure. This method requires only cases, no controls or defined cohort. Provided the relevant exposures are independent in the underlying source population, the case-only study can estimate a specific form of statistical interaction, but not main effects of the 2 exposures.

There are potential advantages to the case-only method in several settings. Estimation of the interaction parameter from case-only analyses is more efficient than for a traditional case–control study (i.e., fewer cases are required for similar precision of estimate) and with no need for controls, there are fewer participants overall (3). Not using controls may mitigate selection biases due to, for example, differential recruiting success between cases and controls, or differential recall of environmental exposures by case–control status. Invasive procedures that are part of cases' diagnosis or treatment often cannot be done ethically in healthy volunteers, especially vulnerable groups such as pediatric populations (4). But these advantages come at a cost. A case-only study only estimates interaction on a multiplicative scale (deviation of the rate ratio for those having both the genetic and environmental exposures from the product of rate ratios for those with either the genetic or the environmental exposure, but not both). It cannot estimate the independent effect of either exposure, or interaction on the additive scale. This limits its use to situations in which the independent effects of the 2 exposures are not of interest, nor are synergism or antagonism of the exposures (5–6). Control-selection bias is the only validity threat the case-only design avoids, in comparison with the case–control design. Consequently, case-only studies have been proposed by several investigators as an initial screening method to identify candidate gene–environment (G–E) or gene–gene (G–G) interactions (7–9).

However, the increase in precision and avoidance of control-selection bias in the case-only method requires a major assumption: that the 2 exposures are independent in the source population (Z = 1; refs. 1–2). Although the constancy of rate ratios between different strata of exposure in the underlying source population is the true parameter of interest, control groups from case–control studies are frequently used to estimate Z using OR_z. Data simulations have demonstrated that even when violations of the independence assumption are of small magnitude they can strongly bias the case-only interaction parameter (7). Chance can also play a role. Since the expectation that OR_z = 1 when Z = 1 is a large sample asymptotic approximation, as sample size decreases, OR_z will deviate from the null with increasing frequency through random error alone (7). Furthermore, when control-group G–E associations are of similar magnitude but opposite in direction to the interaction effect, a case-only study may not detect interaction effects, a Type II error (7, 10).

However, published control group data on the associations of interest for G–E interaction research are limited. Therefore, we undertook a systematic review and meta-analysis of selected DNA repair gene polymorphisms in XRCC1, XPD and XRCC3, and smoking behavior in control groups, using OR_z to estimate Z. The purpose in estimating OR_z was to determine the degree of bias in the COR, relative to the interaction estimate from a case–control analysis, assuming no control-selection bias. Heterogeneity was explored using stratified analysis and meta-regression of study characteristics. The primary aim of this project was to evaluate the independence assumption for selected SNPs and smoking behavior. This will enable investigators considering a stand-alone case-only study of G–E interaction with these exposures to evaluate the independence assumption more rigorously, potentially identifying situations in which case-only estimates may be more or less valid.

PubMed, ISI Web of Science and the CDC Genomics and Disease Prevention databases were searched up to March 6, 2007 for peer-reviewed literature likely to contain noncase data on the joint distribution of any of the polymorphisms of interest. Polymorphisms of interest were nonsynonymous single nucleotide changes (SNPs) in XRCC1 [Arg399Gln (rs25487), Arg194Trp (rs1799782), and Arg280His (rs25489)], XPD [Asp312Asn (rs1799793) and Lys751Gln (rs13181)], and XRCC3 [Thr241Met (rs861539)] (11–15).

Noncase groups were defined as any group not selected on disease status (e.g. cohorts, convenience samples, and control groups from case–control studies). For simplification noncase groups will be referred to as controls throughout this article. There were no language restrictions on searches. A list of keywords for PubMed and the ISI Web of Science was developed in consultation with an information specialist from UNC Health Science Library to ensure that searches would be as inclusive as possible. Keywords for smoking were “smoking,” “tobacco,” “tobacco smoke,” “tobacco smoke pollution,” and “smoker.” The SNPs were searched by combining “polymorphism” and “polymorphism genetic” with the SNP-specific keywords “XRCC1,” “XPD,” “xeroderma pigmentosum group d protein,” “ERCC2” and “XRCC3.” ISI Web of Science keywords were “smoke*” and “tobacco,” and “XRCC1,” “XPD,” “ERCC2,” and “XRCC3”. GDPInfo was searched by limiting by factor menu terms to “smoking behavior,” “smoking (tobacco) passive,” “smoking (tobacco) bidi,” “smoking (tobacco),” “smoking (tobacco) maternal,” “tobacco,” “indoor air pollution,” “nicotine (nasal spray),” and “nicotine (transdermal),” and gene menu terms to “XRCC1,” “XPD,” “ERCC2” and “XRCC3.”

Inclusion criteria were deliberately broad. To be included, an article had to contain original control group data on the joint distribution of any genotype of interest and any aspect of tobacco smoking behavior. Abstracts were excluded.

Abstracts were screened for controls with relevant genotype and smoking data. SNP designations considered equivalent are shown in Table 1. If an abstract passed the initial screening, the full paper was reviewed for data appropriate for construction of a 2 × 2 table for genotype-smoking association in controls (OR_z). If a genotype-smoking OR_z could be calculated, the following data were abstracted: SNP, genotype categories (3 level additive, dominant, and/or recessive models), smoking status, and dose categories [ever/never, current/not current, smoker/nonsmoker, ever/former/current, pack-years (PY), duration, and/or intensity], and cell counts for all genotype and smoking categories.

The following study characteristics were also abstracted: year of publication, study design (case–control, cohort, cross-sectional, convenience, other), source of control group (for case–control: population, hospital, friends and non—blood-related family, convenience, community, neighborhood, other; for cohorts: population, occupational, convenience, other), type of clinic that hospital- or clinic-based control groups were from (disease clinics, checkup clinics), study outcome [cancer (type), noncancer disease, nondisease], full study/control group size (N), country, percent male participants, ethnicity, Hardy Weinberg equilibrium P-value, full study/control group size (N), and minor allele frequency (MAF). An estimate of central tendency for participants' age in years (“average age”) was derived for each study using, in order of preference: median, mean, weighted average across study categories, midpoint of range. One non-English language article could not be evaluated.

No study population contributed to any analysis more than once, maintaining independence of observations. Analyses focused on associations with genotype categorized using a dominant model (i.e., homozygotes of the most common allele were the referent group, compared with heterozygotes plus homozygotes of the minor allele) due to the small number of studies that provided sufficient information to assess recessive or additive models. The minor allele did not vary across studies for any of the included SNPs. Smoking status was categorized as (1) ever/never (referent), and (2) current/not current (referent). Smoking amount was analyzed as (1) pack-years [PY, highest vs. lowest nonzero category (referent)], (2) duration [years, longest vs. shortest nonzero category (referent)] and (3) smoking intensity [cigarettes/day, heaviest vs. lightest nonzero category (referent)].

Studies that did not provide sufficient data to include “passive only” smoking in the never smoking group were excluded. For analyses of current/not current smoking, never + former smokers and “nonsmokers” (unless identified as never smokers) were considered not current smokers. Pack-years of smoking (PY = pack-years = number of packs smoked per day multiplied by years smoked; 20 cigarettes = 1 pack) were collected as categorical variables with different cutpoints. We analyzed PY as relative categories [heaviest vs. lightest smokers regardless of PY cutpoints] and absolute categories [high PY (all categories with a cutpoint above a specified minimum) vs. low PY (all categories with a cutpoint below a specified maximum)]. The minimum and maximum cutpoints varied by SNP but were all chosen to maximize the number of included studies while keeping the range of cutpoints small enough that no study would have >1 cutpoint between the minimum and maximum cutpoints. Similar to PY, smoking intensity was categorized by relative [heaviest vs. lightest smokers regardless of cigarette/day cutpoints] and absolute [> = 20 vs. <20 cigarettes/day] measures. The smoking duration cutpoints were 20 and 40 years, inclusive, for all SNPs.

Crude ORs and 95% confidence limits were calculated from cell counts (Stata 9.2, using Metan STB-44: sbe24). Funnel plot asymmetry, an indicator of possible publication bias (16), was considered suggestive of study characteristics associated with variance and Z. When data were sufficient (N_studies> = 5), asymmetry was formally assessed using Begg and Mazumdar's test (17) and Egger's test (18) at α = 0.10. Cochran's Q 2-sided homogeneity P-values (α = 0.10 due to low power of the test) were used to assess overall heterogeneity in odds ratios (19).

Study characteristic analyses: Key study characteristics hypothesized to influence variation in the strength of control group SNP-smoking associations across studies were assessed using stratified meta-analysis and random-effects meta-regression, with the among-study variance estimated by restricted maximum likelihood (20). Stratified meta-analysis produces a summary OR_z estimate for each stratum of a study characteristic. Meta-regression provides a formal comparison of the stratified estimates in the form of an estimated ratio of odds ratios.

Study characteristics were selected a priori. They included (1) study design [case–control, cohort, or convenience; patient-based control groups, healthy control groups], (2) continent, (3) ethnicity, (4) Hardy–Weinberg equilibrium P-value, (5) average age, (6) gender (% male), (7) study outcome (lung cancer, other cancer, noncancer disease, nondisease), (8) MAF and (9) smoking prevalence. Study design was examined for all SNP-smoking combinations; additional study characteristics were examined for XRCC1 399, XPD 751, and XRCC3 241. Stratified random-effects meta-analyses were used when the overall SNP-smoking association had a Cochran's Q P -value α < 0.10, otherwise fixed effects meta-analysis was used, regardless of the homogeneity P-values of individual strata. To reduce the possibility of results being confounded by ethnicity (population stratification) in overall analyses and when examining the study characteristics likely to vary strongly by ethnicity (Hardy–Weinberg equilibrium P-values and minor allele frequency) studies were stratified by ethnicity and treated as separate studies if possible. Sample size was not formally examined as a study characteristic because that would have essentially reproduced funnel plot analyses of variance. Variance (or precision) is the more important measure here, and sample size is not the sole determinant of precision. However, because the independence assumption is a large sample approximation, a sensitivity analysis was conducted in which large (N> = 1,000) studies were examined separately. Stata 9.2 was used for all analyses. Results for study characteristics were assessed for consistency across smoking categories and across SNPs.

The literature searches identified 228 articles for evaluation. Of these, 55 articles were eligible for inclusion. The primary reason for exclusion was that an article did not present the genotype-smoking distribution in controls (N = 98, 57% of exclusions). Exclusion reasons for the remainder included: review article or abstract only (13%), did not assess any relevant SNPs (9%), and did not have any noncases (10%). Finally, of the 55 studies eligible for inclusion, 5 were not included in final summary estimates because no data were presented using the dominant genetic model (21, 22), there was no measure of adult-smoking behavior (23), former smokers were excluded (24), or never smokers were included in lowest PY category (25). Fifty articles representing 46 distinct study populations were included in the final meta-analyses (brief study descriptions in Supplementary Table 1). The number of study populations included for each polymorphism was: XRCC1 Arg399Gln (N = 32), XRCC1 Arg194Trp (N = 16), XRCC1 Arg280His (N = 8), XPD Lys751Gln (N = 16), XPD Asp312Asn (N = 9), and XRCC3 Arg241Gln (N = 13). Thirty-seven studies presented the control distribution of genotype and ever/never smoking, 16 for current/not current smoking and 14 for PY. Far fewer presented duration (N = 4) and/or intensity (N = 4). Case–control studies predominated with 12 population-based (26–37) and 23 hospital-based (38–60), 4 studies nested within cohorts (61–64) and 2 other case–control studies (65–66). Most control groups were from cancer case–control studies (N = 39), 1 was from a case–control study of rheumatoid arthritis. Nine cohort or convenience sample studies examined noncancer outcomes, predominantly measures of DNA damage (67–75).

Across SNPs there was more variation in ORs assessing control-only G–E associations (OR_zs) for measures of smoking amount (PY, duration, intensity) than for measures of smoking status (ever-never, current-not current; Table 2). Ten of 11 summary estimates of smoking status fell between 0.9–1.1. Summary estimates for smoking amounts were distributed more broadly, with only 5 of 10 summary estimates between 0.9 and 1.1; the most extreme measures were found for duration and intensity. Although only 2 of 18 genotype-smoking groups were too heterogeneous for a fixed effects summary estimate, based on Cochran's Q at α = 0.10, nearly all groups had study estimates above and below the null. Individual study OR_z (95% CI) are presented in Figure 1 and Supplementary Tables 3–8.

For XRCC1 399 any Gln and ever-never smoking, OR_zs ranged from 0.7 (95% CI: 0.3, 1.7; ref. 49) to 1.9 (95% CI: 1.0, 3.7; ref. 34; see Fig. 1); 3 other measures of smoking behavior were homogeneous enough for a summary estimate of OR_z: current smoker/not current (N = 11), PY (N = 9) and intensity (N = 4; Table 2). Higher PY and heavier-smoking intensity, but not current versus not current smoking, were associated with XRCC1 Arg399Gln (any Gln) [OR (95%CI): 1.2 (1.0, 1.5) and 1.5 (1.2, 1.9), respectively]. For XRCC1 194 and 280, having the variant allele was associated with smoking duration [XRCC1 194: 0.7 (0.5, 0.9), XRCC1 280: 1.2 (0.6, 2.3)] and current smoking [XRCC1 280: 1.2 (0.6, 2.3)] though the number of studies was small. For the 2 XPD SNPs (751, 312) there was considerable variation in the association between XPD 751 variant allele and higher PY. Study estimates ranged from 1.4 (0.8, 2.6; ref. 41) to 0.5 (0.3, 1.0); ref. 53; Table 2). Higher PY were associated with the variant allele for XRCC3 241 although the number of studies was small (N = 4).

Among the studies that were assessed for current-not current smoking, a subset could also be assessed for never, former or current smoking (Supplementary Table 2). No consistent pattern emerged for comparisons of never smoking with former or current smoking. Absolute measures of PY, intensity, and duration were calculated and compared with relative measures for consistency. Genotype-PY estimates for absolute cutpoints were comparable with estimates using relative categories although strata were sparse (Supplementary Table 2). Additionally, when studies with only smokers were dropped and never smoking was used as the reference category, results were essentially the same for relative and absolute measures of PY.

Genotype-smoking association between XRCC1 Arg399Gln and smoking intensity (cigarettes/day) could be estimated in 4 studies. There was an association between XRCC1 399 any Gln and greater-smoking intensity, which was consistent across methods of smoking intensity categorization. Results did not change appreciably when studies without smoking amount were excluded from ever-never analyses, indicating that articles that presented dose were not driving estimates of smoking status. Results did not change appreciably when studies without smoking amount were excluded from ever-never analyses, indicating that articles that presented dose were not driving estimates of smoking status (data not shown).

There were 6 large (N> = 1,000) study populations, 4 each with data for XRCC1 399 and 194 ever-never smoking. OR_zs for XRCC1 399 ever-never smoking showed evidence of heterogeneity [range of OR_z (95% CI): 1.0 (0.8, 1.2)–1.6 (1.2, 2.0); Supplementary Table 3; refs. 32, 35, 42, 60]. However, OR_zs for XRCC1 194 were consistently null across studies (Supplementary Table 4; refs. 32, 37, 42, 61). In the 3 large study populations with the relevant measures of smoking (32, 42, 60) the magnitude of OR_z was consistently different for status and amount for XRCC1 399 ever-never smoking; the 1 study population with smoking status and amount for XRCC1 194 had OR_zs at the null (42). Data were too sparse for further evaluation across large studies.

There was no evidence of funnel plot asymmetry for overall genotype-smoking associations (data not shown). In formal testing, a majority of P-values (75%) were > = 0.3; the lowest P-value was P = 0.14.

Stratified associations and univariate meta-regression were evaluated across SNPs and smoking categories on the basis of consistency and direction. Study design was examined for all 6 SNPs for ever/never, current/not current smoking, and PY. For smoking status, genotype-smoking associations for XRCC1 399 and 194 and XPD 751 and 312 were generally stronger for population-based case–control studies than for hospital-based or patient-based control groups, although the magnitude of the differences was small; the range of RORs was 0.7 to 0.9 for hospital/patient-based compared to population-based controls (referent; Table 3). However, for smoking amount as measured by PY (2 evaluable SNPS, XRCC1 399, and XPD 751) the hospital-based/patient-based control groups showed stronger genotype-smoking associations than population-based control groups (range of RORs: 1.2–1.5). When examining PY, for all SNPs, the genotype-smoking association for population-based control groups was below the null. The remaining study characteristics were examined only for XRCC1 399, XPD 751, and XRCC3 241 (Tables 4–6, respectively) due to sparse data for the other SNPs.

For PY, lung cancer studies were above the null for all 3 SNPs. When compared to studies of other cancers the genotype-smoking association was stronger for lung cancer studies (referent) compared to other cancer studies [ROR = 0.8 (0.5, 1.2) and 0.5 (0.3, 0.9) for XRCC1 399 and XPD 751, respectively]. All studies with PY were cancer studies. Older average age of study participants weakly but consistently showed stronger associations between ever smoking and variant allele for XRCC1 399, XPD 751, and XRCC3 241 than did younger age. For XRCC1 399 only, this was evident across all 3 smoking categories. Also, for XRCC1 399 current-not current smokers and PY only, studies with lower minor allele frequencies (N = 3) showed stronger associations (∼2.0) than those with higher MAF. These 3 studies had only African American or Asian participants. No strong and/or consistent patterns emerged for other study characteristics examined.

This systematic review and meta-analysis of DNA repair genotypes and smoking behavior in control data was conducted with the goal of examining the independence assumption of case-only studies of G–E interaction. There was considerable variation in estimates of Z for XRCC1 399 ever-never smoking and XPD 751 PY of smoking. Point estimates of OR_z varied as much as 5-fold, even when studies were homogeneous enough for a summary estimate. Summary estimates for individual SNPs varied across smoking categorizations, with larger magnitudes of association generally found for measures of smoking dose (PY, intensity, duration) than for smoking status (ever-never, current-not current). There was a weak association between XRCC1 399 and higher-smoking dose (PY, intensity). No study characteristics examined strongly predicted the magnitude of association although study outcome (lung cancer vs. other cancer for PY), study design (population-based vs. hospital/patient-based), and age warrant further investigation.

Although the validity of case-only estimates rests on the independence assumption (2, 76), literature on independence assumption verification is limited. Data simulations have demonstrated that small violations of the independence assumption can strongly bias the case-only interaction parameter (7). Even an OR_z of 1.2 biased the COR by nearly 30%. Furthermore, when Z ≠ 1 in population subgroups, the COR for those subgroups will be biased as well.

However, little empirical work has been conducted to quantitatively assess the magnitude of control-only associations (OR_z) between DNA repair gene variations and smoking. A population-based study (N = 339) of Japanese males assessed association between “habitual smoking” (ever/never) and a panel of 153 SNPs in 40 candidate genes, including the DNA repair genes OGG1 and NUDT1 (MTH1; ref. 77). Association was found between smoking and 3 of 4 of the SNPs in OGG1 (0.4–0.6, borderline statistical significance).

Smoking amount (PY and/or intensity) may be causally associated with variation in XRCC1 399, or with a polymorphism in linkage disequilibrium with XRCC1 399. There is evidence that the XRCC1 399 and XPD 751 variants are functional (78–80). Different aspects of smoking behavior (smoking initiation, smoking cessation, intensity, etc.) operate through multiple overlapping pathways (81) therefore would not be expected to be identically affected by DNA repair variation. This is supported by the differing results for smoking status and amount for several SNPs (XRCC1 399, XRCC1 280, XPD 751, XRCC3 241). There is some evidence that variation in DNA repair activity may affect neurological and/or respiratory outcomes, which could in turn affect smoking behavior (82–86). If the variants are functional, or linked to functional variants, heterogeneity could be due to G–E interaction in specific populations.

There are also several possible noncausal explanations for these finding. Although publication bias is a concern with meta-analyses, visual inspection of funnel plots and formal tests of asymmetry argue against this. Spurious results for XRCC1 399 and smoking amount could be caused by selection bias in a subsample of studies. Just over half of the studies with smoking amount information for XRCC1 399 were lung cancer studies (8 of 14) and lung cancer studies had on average higher OR_zs than other cancer studies for all PY analyses. The connection between smoking and lung cancer is well known, possibly leading to more variation in response rates or recall by smoking history and/or family history of cancer, but the direction of possible bias is unpredictable. The OR_z for the one XRCC1 399 study that explicitly excluded participants with smoking-related diseases was essentially the same as the summary estimate (42).

Population stratification could have contributed to the heterogeneity in XRCC1 399 ever-never and XPD 751 PY estimates since the variant alleles are found at different frequencies in different ethnic groups within the same study, and smoking behavior may also differ by ethnicity. Although this cannot be rigorously assessed without individual level data, there were no clear patterns in OR_z for any SNP for study-level ethnicity, either by stated ethnicity, when stratified by single-ethnicity versus multiethnicity studies, or when MAF was used as a crude proxy to assign ethnicity for studies with unknown ethnic makeup. Finally, chance could play a role, particularly given the large number of associations examined and sparse data for many analyses. However, in the 4 studies with large sample sizes (N> = 1000) for XRCC1 399 ever-never smoking, OR_zs ranged from 1.0 (0.8, 1.2) to 1.6 (1.2, 2.0), while OR_z was essentially null across the 4 large studies that examined XPD 194 ever-never smoking [OR_zs = 0.9, 1.0, 1.1 and 1.1]. Furthermore, the magnitudes of OR_z differed across smoking status and amount for XRCC1 399 (3 populations) but not for XRCC1 194 (1 population). This large-sample sensitivity analysis is consistent with the overall interpretation of OR_z's population-specificity for each SNP, rather than chance alone driving the heterogeneity among studies.

Z is a measure of the magnitude of bias in the COR. If Z = 1, the case-only estimate of interaction is not biased by genotype–environment association in the underlying population (76). Commonly, this assumption is assessed in control data from a small number of outside studies, using significance testing. Significance testing alone is not sufficient for assessment of potential bias (87). Rarely is Z estimated and/or adjusted for, analogous to other forms of bias such as confounding.

Results from this project illustrate some of the pitfalls of this approach. For instance, for XRCC1 399 ever-never smoking, 18 of the 21 included studies have estimates that are not statistically significantly different than the null value of 1.0. Considering any of these in a statistical significance testing framework would lead to the conclusion that the independence assumption was valid; therefore a case-only study estimate of interaction would not be biased, at least from independence assumption violation. However, the range of OR_zs for these 18 studies is 0.7–1.6, many with wide CIs, indicating the substantial range of potential bias of the effect estimate. Given that different conclusions can be drawn from subsets of smoking behavior and that less than half of the studies that collect control genotype and smoking information present it in publications, this ever-never approach seems inappropriate.

In the estimation framework, results from this project demonstrate the difficulty of using ancillary data to assess the independence assumption. Even when the Cochran's Q P-value is high, such as for XRCC1 399 current-not current smoking (P = 0.4), point estimates of OR_z can vary as much as 5-fold [2.1(1.1, 3.9) for African Americans (27) to 0.4 (0.1, 1.2; ref. 75)]. Without further information that certain study characteristics might be influential, there is no good way to decide which of the available ancillary control groups might best represent the underlying (unmeasured) population for a proposed case-only study. Furthermore, it is necessary to do a broad literature search to even to be aware of the possible values of OR_z and range of bias in the COR. Additionally, since both summary estimates and individual study estimates vary across smoking categories, it is important that the independence assumption be evaluated for all smoking categories that will be used in the case-only analyses. For investigations of smoking amount, it will be difficult for many SNPs to locate enough published control group data to even assess the possible range of the magnitude of bias.

This study has several strengths. Using a comprehensive search strategy in collaboration with information specialists increased power to detect and investigate heterogeneity between studies. Sample size was large for smoking status analyses and relatively large for XRCC1 399 and XPD 751 PY analyses. There were sufficient data for many studies to compare OR_z for smoking status and amount within studies, and by smoking category across multiple SNPs. However, of the searched studies that collected the appropriate information only about 1/3 presented it such that it could be abstracted for meta-analysis, limiting sample size, especially for measures of smoking amount.

Only unadjusted odds ratios could be calculated so study estimates may have been confounded. Although some study characteristics could be determined accurately from articles, others were more likely to be misclassified. In particular, average age of study participants was difficult to determine. However, the fact that age was not a central study feature for any of the studies makes it likely that misclassification is nondifferential with respect to smoking and genotype. Several potentially informative study characteristics could not be examined because too few articles presented the relevant information using the same metric. In particular, response rates, which may vary by smoking behavior and family/personal history of cancer (88–91), and control group exclusion criteria, were presented very differently. Only 2 of the 12 articles with multiethnic study populations presented data stratified by ethnicity, complicating interpretation of HWE P-value, ethnicity and MAF as study characteristics. Few studies presented enough control group information to examine multiple measures of smoking in the same study population.

This systematic review of control-group associations between smoking and widely studied polymorphisms in DNA repair genes was conducted to accomplish several objectives. The overarching goal was to enable investigators to make more effective use of ancillary data to evaluate the independence assumption prior to launching a stand-alone case-only study. Results from this study suggest that the independence assumption is frequently violated and caution is warranted before proceeding with any case-only interaction analysis. At a minimum, the independence assumption should be more rigorously evaluated than is often done. For a case-only analysis of a case–control study, separate OR_zs should be calculated for each anticipated COR in the relevant subgroup before proceeding. Evaluation of the independence assumption for a proposed stand-alone case-only study should include, whenever possible, results from studies similar to the current study, relevant literature reviews, and a thorough search for individual studies with control or cohort data to ascertain at least the range of OR_zs, both overall and in relevant subgroups. Finally, it serves as a reminder that in the traditional case–control study, interaction is a contrast between control-only association and case-only association, and interaction can be driven by unanticipated associations in controls.

Evaluation of the independence assumption for case-only interaction studies would be greatly improved with more transparency and finer detail in published articles. This could perhaps be accomplished by expanding supplementary online tables to include selected joint genotype-smoking distributions in noncase groups. If it could reliably be shown that Z = 1 across individual studies, better use could be made of data pooling from control groups and cohorts for selected SNPs and exposures, especially where individual level data on potential confounders can be provided. However, despite the current emphasis on pooling controls, our results indicate that investigators should not proceed with case-only studies without rigorously evaluating the independence assumption in individual studies.

No potential conflicts of interest were disclosed.