Is the Association between Cigarette Smoking and Breast Cancer Modified by Genotype? A Review of Epidemiologic Studies and Meta-analysis

Animal experiments and in vitro studies have shown that compounds found in tobacco smoke, such as polycyclic hydrocarbons, aromatic amines, and N-nitrosamines, may induce mammary tumors (1-3). The finding of smoking-specific DNA adducts (4-7) and p53 gene mutations (8) in the breast tissue of smokers also support the biological plausibility of a positive association between cigarette smoking and breast cancer. Nonetheless, >100 epidemiologic studies on smoking and breast cancer have shown inconsistent results, and the association remains the focus of debate (1-3). Epidemiologic studies have examined the association between smoking and breast cancer risk according to genotype with increasing frequency, commensurate with the growing awareness of the roles genes play in detoxifying or activating chemicals found in cigarette smoke and in preventing or repairing the damage caused by those compounds. Whether studies that considered genotype have helped to clarify the association is unclear. Our aim here is to summarize the results of epidemiologic studies of smoking and breast cancer risk that considered genotype and discuss their implications for future investigations.

We obtained relevant articles through searches of Medline and by cross-matching the references of relevant articles. The review of literature is organized by gene category and by individual gene. The presentation of epidemiologic findings in each section is preceded by a brief description of gene function. Results of individual studies are presented for the highest category of smoking duration compared with never smokers. If estimates for smoking duration were not available in a study, then results are presented for pack-years, cigarettes per day, current smokers, or ever smokers, in that order of priority. Nearly all of these studies have been of case-control design, including those nested within a cohort. Nine studies (9-17) did not report the association between smoking and breast cancer within strata of certain genotypes but provided the necessary counts of cases and controls so that crude odds ratios (OR) and corresponding 95% confidence intervals (95% CI) could be calculated by the authors of this review. Nearly all of these studies evaluated interaction between smoking and genotype on a multiplicative scale (e.g., using likelihood ratio tests); some studies also evaluated interactions on an additive scale (refs. 16, 18-25; e.g., using interaction contrast ratios). The results of two case-only studies (26, 27) are discussed in the text but are not shown in the tables.

Whenever possible, we examined each study with respect to its power to detect a relative risk (RR) estimate (main effect) of at least 2.0 (or 0.5 in one cohort study, where the frequency of disease among of nonexposed participants was 58%) and present these calculations in the tables. Power calculations were not done when necessary data were not available. All power calculations were conducted using statistical software Epi Info (Centers for Disease Control and Prevention, Atlanta, GA) for cohort studies and unmatched case-control studies and statistical software Quanto version 0.5 (University of Southern California, Los Angeles, CA) for matched case-control studies.

A summary meta-RR estimate for the association between smoking and breast cancer risk was calculated for genotypes for which estimates were available from three or more studies with nonoverlapping samples (NAT2, NAT1, CYP1A1, GST, SOD2, and XRCC1). We calculated meta-RRs as weighted averages of the individual study results using inverse variances as the weights (28). These summary analyses were based on random effects models, which assume that study-specific effect sizes arise from a random distribution of effect sizes with a certain mean and variance. Where the variability among studies is negligible (high levels of homogeneity), the random effects model reduces to a fixed effects model, which attributes all observed variations among results to sampling error alone (29). Because these studies examined different smoking measures and included a mixture of ages and ethnicities, the meta-RRs are intended to serve only as one possible summary of the available data. In instances where the meta-RR estimates were significantly different from null, we evaluated the potential role of publication bias by calculating the weighed fail-safe n as described by Rosenberg (30). The fail-safe n is equivalent to the number of studies of null effect and mean weight necessary to change the observed type I error from <0.05 to 0.05. These calculations were done using the Fail Safe Number Calculator software (Arizona State University, Tempe, AZ).

Cancer risk is, at least in part, an integrated function of carcinogen exposure and polymorphisms in genes involved in carcinogen metabolism, including cytochrome P450s (CYP), glutathione S-transferases (GST), N-acetyltransferases (NAT), and sulfotransferases (SULT; ref. 31). Studies that examined the association between smoking and breast cancer according to variation in carcinogen metabolism genes are discussed in the next four subsections (Tables 1-4).

N-acetyltransferases. Aromatic and possibly heterocyclic amines, constituents of tobacco smoke, may be detoxified or activated by NATs, including NAT1 and NAT2 (32). Regarding NAT2 genotype, individuals carrying homozygous wild-type, homozygous variant, and heterozygous genotypes are considered to be rapid, slow, or intermediate acetylators, respectively. These genotypes can determine the rate of metabolism or activation of carcinogenic aryl or heterocyclic amine substrates (32). For example, most studies indicate that slow acetylators, particularly individuals homozygous for NAT2 slow acetylator alleles, have an increased risk of arylamine-induced bladder cancer (33). However, whether this is true for breast cancer is unclear.

Effect modification by NAT2 genotype has been evaluated in 14 epidemiologic studies of smoking and breast cancer risk (Table 1). These studies tend to show increased risks due to smoking among postmenopausal slow acetylators. This suggests that rapid acetylators more efficiently detoxify the carcinogenic compounds in tobacco smoke and that greater exposure of breast tissue to activated compounds is likely in slow acetylators. Among all subjects combined, summary analyses produced meta-RR estimates of 1.2 (95% CI, 1.0-1.5) for rapid acetylators and 1.5 (95% CI, 1.2-1.8; fail-safe n = 54) for slow acetylators. After stratification by menopausal status, the meta-RR for postmenopausal slow acetylators was 2.4 (95% CI, 1.7-3.3; fail-safe n = 57), whereas similar analyses for the other categories showed no association (Table 1). However, these studies generally did not find the NAT2 genotype itself to be independently associated with breast cancer risk, and a recent case-only study of breast cancer (n = 502; ref. 26) found little evidence of interaction between smoking and NAT2 genotype. In addition, seven of these studies considered passive smoking in analyses according to NAT2 status (21, 26, 34-38), showing no clear pattern of effect modification of the association by genotype.

Five studies have examined the association between cigarette smoking and breast cancer risk according to NAT1 genotypes (Table 2). The NAT1*10 allele was considered in most of these studies, because it has been associated with elevated NAT1 activity (39). However, there has been no clear effect modification by this or any of the NAT1 alleles examined in epidemiologic studies of smoking and breast cancer risk. Summary analyses showed no association between smoking and breast cancer irrespective of NAT1 genotype (Table 2).

Cytochrome P450. Unlike aromatic and heterocyclic amines, polycyclic aromatic hydrocarbons (PAH) are metabolized primarily by enzymes in the CYP family and by GSTs, a process that can produce highly reactive DNA-damaging metabolites (40). The CYP1A1 gene encodes microsomal CYP1A1, a phase I enzyme that contributes to aryl hydrocarbon hydroxylase activity, catalyzing the metabolism of PAHs and other carcinogens (41). Four CYP1A1 polymorphisms have been studied in relation to breast cancer: 3801T→C (M1, a MspI RFLP in the 3′-noncoding region), Ile⁴⁶²Val (M2), 3205T→C (M3, a MspI RFLP in the 3′-noncoding region), and Thr⁴⁶¹Asp (M4). These polymorphisms generally have not been shown to alter breast cancer risk and their functional significance remains unclear (41). Regarding the M1 polymorphism, some studies have shown higher levels of DNA adduct levels in breast tissue, or urinary biomarkers of PAH exposure, among individuals with the variant allele (41). However, the results of these studies, and similar studies of the Ile⁴⁶²Val polymorphism, have been largely inconsistent (41).

Four studies have examined the association between smoking and breast cancer risk according to CYP1A1 genotypes to date (Table 3). There seems to be no clear effect modification by CYP1A1 generally, although the results of two studies (42, 43) suggest that long-term smoking may increase risk among women with CYP1A1-M1 variant genotype. Summary analyses produced meta-RR estimates for smoking of 1.3 (95% CI, 1.0-1.6) for women with M2 wild-type and 1.2 (95% CI, 0.6-2.1) for carriers of the variant allele.

The association between smoking and breast cancer risk has also been examined according to polymorphisms of other genes in the CYP family. In an early case-control study (272 cases and 334 controls, a subset of a larger study with information on genotype), Shields et al. (44) showed smoking increased breast cancer risk among premenopausal but not postmenopausal women with 6CYP2E1 variant (a DraI restriction enzyme site in intron 6). Because CYP2E1 enzyme is involved in the metabolic activation of N-nitrosamines, which are contained in tobacco smoke, it is possible that the polymorphism will modify susceptibility to breast cancer due to smoking. However, these findings were based on small numbers of subjects in some of the exposure categories and remain to be replicated. More recently, a case-only analysis (282 cases; ref. 27) suggested that CYP1B1 polymorphism (Val⁴³²Leu) may increase breast cancer susceptibility among smokers. CYP1B1 activates a variety of procarcinogens, including those found in tobacco smoke (45). This study also examined the association between smoking and breast cancer risk according to a polymorphism in SULT1A1 (see below).

Glutathione S-transferases. The GSTs are phase II enzymes that play key roles in detoxification of many potentially carcinogenic compounds, including PAHs, which are contained in tobacco smoke. There are eight distinct gene families encoding GSTs in humans: α, μ, 𝛉, π, ζ, σ, κ, and χ (also called ω; ref. 46). Polymorphisms have been described in several GST genes, although mostly in the μ, 𝛉, and π families (46). A recent pooled analysis showed no association between several of these polymorphisms and breast cancer risk (47).

Seven studies have examined the association between smoking and breast cancer risk according to GST genotypes (Table 4). These studies tend to show positive associations with smoking among women with GSTM1-null genotype. Individuals with this genotype express no protein (46), a protein that has been shown to modulate cytogenetic damage in smokers (48). Summary analyses produced meta-RR estimates for smoking of 1.1 (95% CI, 0.8-1.4) when the GSTM1 gene was present and 1.5 (95% CI, 1.1-2.1; fail-safe n = 11) when the gene was deleted (Table 4). Similar analyses for GSTT1 produced meta-RRs of 1.3 (95% CI, 1.1-1.6; fail-safe n = 7) and 1.2 (95% CI, 0.9-1.7) for GSTT1-present and GSTT1-null, respectively. An increased risk due to smoking among individuals with the GSTT1-present genotype is not necessarily inconsistent with similar findings for GSTM-null because GSTT1 may not only detoxify but may also metabolically activate carcinogens (49). Nonetheless, our findings for smoking according to GST genotypes differ from those of a recent analysis of data pooled from seven previous studies (2,048 cases and 1,969 controls)¹

Sulfotransferase 1A1. SULTs are enzymes involved in the metabolism of several classes of compounds through sulfonate conjugation, including the activation and inactivation of PAHs and heterocyclic amines found in cigarette smoke. A common polymorphism, a G→A transition (Arg²¹³His) in SULT1A1, results in reductions in enzyme activity, particularly among individuals homozygous for the His allele (51). A case-only analysis (282 cases; ref. 27) showed that carrying any His allele may increase breast cancer susceptibility among smokers, suggesting increased exposure to unsulfated tobacco carcinogens among His allele carriers. However, two case-control studies that examined the association between smoking and breast cancer risk according to this polymorphism found no clear evidence of effect modification (34, 52). A population-based case-control study in Finland (483 cases and 482 controls) reported no interaction between smoking status and Arg²¹³His polymorphism (ref. 52; relevant ORs were not reported). A population-based case-control study in Germany (419 cases and 884 controls) showed women who smoked ≥16 years had ∼50% increased breast cancer risk compared with nonsmokers, again with no clear effect modification by the SULT1A1 polymorphism (34). The latter study further examined the association between smoking and breast cancer by SULT1A1 after additional stratification by NAT2 genotype, also showing no clear effect modification when NAT2 was considered. Although that study had >90% power to detect ORs of at least 2.0 among strata of the SULT1A1 polymorphism, power fell to <35% for all analyses that were further stratified according to NAT2 genotype.

Superoxide dismutase 2. SOD2 encodes manganese superoxide dismutase (MnSOD), a mitochondrial enzyme that protects the cell against damage from superoxide free radicals. A common polymorphism, a T→C transition (Val¹⁶Ala), was shown to alter the MnSOD mitochondrial targeting sequence (53), and animal studies have linked reduction in MnSOD activity to increased DNA damage and higher incidence of cancer (54). Increased expression of MnSOD was found to suppress the malignant phenotype of human breast cancer cells in vitro (55).

Four epidemiologic studies have examined the association between smoking and breast cancer risk according to this SOD2 genotype (Table 5). These studies suggest that long-term smoking may increase risk in women homozygous for the Ala allele. Summary analyses produced meta-RR estimates for smoking of 0.8 (95% CI, 0.2-2.8) for the Val/Val genotype and 1.5 (95% CI, 1.1-2.1; fail-safe n = 3) for Ala/Ala.

The genes discussed above work in several ways to modulate the dose of carcinogens experienced by individuals and, consequently, the amount and type of DNA damage. Once DNA damage occurs, however, cells have ways of minimizing long-term damage that may result, including repair of the damage through several pathways. DNA repair capacity seems to play an important role in disease susceptibility (56).

XRCC1. X-ray repair cross-complementing group 1 (XRCC1) acts as the central scaffolding protein for POLB and ligase III in the base excision repair of oxidative damage, participates in the removal of nonbulky DNA adducts caused by exogenous agents, including several compounds in tobacco smoke, and interacts with poly(ADP-ribose) polymerase in the detection of single-strand breaks (56). Three polymorphisms resulting in nonconservative amino acid substitutions (Arg¹⁹⁴Trp, Arg²⁸⁰His, and Arg³⁹⁹Gln) have been identified in the XRCC1 gene (23). The Arg²⁸⁰His and Arg³⁹⁹Gln polymorphisms have recently been characterized as “possibly damaging” based on protein conservation analyses (Sorting Intolerant from Tolerant) that identify amino acids that are likely important for the function and structure of protein families (57).

Five epidemiologic studies have examined the association between smoking and breast cancer risk according to XRCC1 polymorphisms (Table 6). In an early report using data from the Carolina Breast Cancer Study (18), smoking duration of >20 years was positively associated with breast cancer risk among African American women homozygous for the Gln allele (OR, 2.9; 95% CI, 1.6-5.2), but no clear association was found among those with a wild-type allele, among women with or without Arg¹⁹⁴Trp polymorphism, or among White women. However, an updated analysis of Carolina Breast Cancer Study data using more cases subsequently showed a positive association with smoking among White women with the Gln/Gln genotype (1).

The latter findings are consistent with those of a recent case-control study nested in the American Cancer Society Cancer Prevention Study II (White women only; ref. 13), which showed women homozygous for the Gln allele (compared with those homozygous for Arg) had ORs of 2.8 (1.4-5.6) and 0.6 (0.3-1.3) for smokers and nonsmokers, respectively (P_interaction = 0.01). These findings are also consistent with those of a recent population-based case-control study in Finland (10) that examined the association according to Arg³⁹⁹Gln and Arg²⁸⁰His, showing smoking (>5 pack-years versus less) increased risk among carriers of the ³⁹⁹Gln allele (OR, 4.1; 95% CI, 1.7-12.0) but not clearly among carriers of ²⁸⁰His (OR, 2.0; 95% CI, 0.6-7.0). However, the most recent analysis of the Carolina Breast Cancer Study data showed positive associations between smoking and breast cancer risk among women with XRCC1 codon 194 Arg/Arg, 399 Arg/Arg, and 280 His/His genotypes (58).

Thus far, there has been little consistency among studies of smoking and breast cancer risk according to XRCC1 genotype. Summary analyses produced meta-RR estimates for smoking of 1.2 (95% CI, 1.0-1.5) for the 194 Arg/Arg genotype and 1.0 (95% CI, 0.7-1.5) for any Trp (Table 6). Summary analyses also produced meta-RR estimates for smoking of 1.1 (95% CI, 0.7-1.7) for the 399 Arg/Arg genotype and 1.1 (95% CI, 0.9-1.4) for any Gln.

XPD (ERCC2). Xeroderma pigmentosum D (XPD), also known as excision repair cross-complementing rodent repair deficiency, complementation group 2 (ERCC2), encodes a protein that plays a role in nucleotide excision repair of bulky DNA adducts (59), including those that may be caused by tobacco carcinogens. Several polymorphisms have been described in XPD, although relatively few epidemiologic studies have investigated XPD polymorphisms and cancer risk and the results of these studies have been inconsistent (60). Effect modification by a polymorphism in this gene (Lys⁷⁵¹Gln) has been examined in two recent population-based case-control studies of smoking and breast cancer risk, the Long Island Breast Cancer Study (15) and a study in Finland that also examined polymorphisms in XRCC1 (ref. 10; Table 6). This polymorphism has been shown to change the electronic configuration of the amino acid (61) and hence may have important consequences to DNA repair capacity (60). In both of these studies (10, 15), an increased breast cancer risk among smokers was limited to women who had both copies of the Gln allele. In the Finnish study, women who had ever smoked and were homozygous for the XPD Gln allele and also carried the XRCC1 ³⁹⁹Gln or the XRCC1 ²⁸⁰His allele had the highest breast cancer risk (10).

MGMT. O⁶-methylguanine DNA methyltransferase (MGMT) encodes a protein involved in the direct reversal repair of DNA damage at the O⁶-position of guanine, which can be caused by various exogenous alkylating agents, including tobacco-specific nitrosamines (62). Genetic alterations in the MGMT gene may impair the protein's ability to remove alkyl groups from this position and, through this mechanism, may possibly increase the risk of cancer (63). Three polymorphisms resulting in nonconservative amino acid substitutions (Leu⁸⁴Phe, Ile¹⁴³Val, and Lys¹⁷⁸Arg) have been identified in the MGMT gene (63). The results of epidemiologic studies of these polymorphisms in relation to cancer risk have been inconsistent (23), although the Leu⁸⁴Phe and Ile¹⁴³Val polymorphisms have recently been shown to modify chromosome aberration frequencies in human leukocytes exposed in vitro to the tobacco-specific nitrosamine carcinogen 4-(methylnitrosamino)-1-(3-pyridyl)-1-buyanone (64). Only one epidemiologic study has examined the association between smoking and breast cancer risk according to MGMT polymorphisms (23), a recent study from the Long Island Breast Cancer Study that showed heavy smoking (>31 pack-years) increased breast cancer risk particularly among women with the Leu⁸⁴Phe polymorphism. There was no effect modification according to Ile¹⁴³Val or Lys¹⁷⁸Arg in that study (Table 6).

BRCA1. Women carrying a deleterious mutation of the BRCA1 gene have an increased risk of developing breast cancer (65). Inherited mutations of BRCA1 play a role in ∼40% to 45% of hereditary breast cancers but in only 2% to 3% of all breast cancers (65). Multiple functions of BRCA1 may contribute to its tumor suppressor activity, including roles in cell cycle progression, specialized DNA repair processes, DNA damage-responsive cell cycle checkpoints, regulation of a set of specific transcriptional pathways, and apoptosis (65).

There have been four studies of smoking and breast cancer among BRCA1 mutation carriers (Table 6). A matched case-control study (186 cases and 186 controls) that first examined the association between smoking and breast cancer risk in women with BRCA1 and BRCA2 mutations showed an inverse association with smoking (66), although this finding was not confirmed in a later report from this group (67) using data from a greater number of subjects (1,097 cases and 1,097 controls). More recently, data from 316 female BRCA1 mutation carriers, including a subset of women who participated in the previous case-control studies (66, 67), were analyzed using a retrospective cohort design (176 cases and 140 noncases), showing a reduced risk of cancer among women who had ever smoked. An inverse association was dose-response with increasing pack-years of consumption. In this latter study, a reduced breast cancer risk was particularly prominent in BRCA1 mutation carriers who smoked and also had a 28 repeat allele for amplified in breast cancer 1 (AIB1) genotype. AIB1 is an estrogen receptor coactivator (68), although exactly how AIB1 CAG allele repeat lengths affect the function of this protein remains unclear (69). Given these associations, it is possible that BRCA1 and AIB1 genotypes and cigarette smoking will decrease breast cancer risk (e.g., through synergistic modulations of estrogen receptor activity and estrogen potency). However, smoking was not associated with breast cancer risk in a recent matched case-control study (348 cases and 348 controls) of BRCA1 mutation carriers (Table 6). To date, there seems to be no consistent association between smoking and breast cancer risk among BRCA1 mutation carriers.

The association between smoking and breast cancer risk remains controversial despite >100 epidemiologic studies conducted over the past three decades. The results of these studies overall suggest that smoking probably does not decrease risk and indeed suggest that there may be an increased risk with smoking, particularly heavy smoking of long duration (1). Some authors have suggested that any positive association is likely driven by confounding (or residual confounding) from alcohol consumption (70). Although confounding by alcohol may have occurred in some studies, two recent cohort studies (71, 72) found statistically significant positive associations between smoking of long duration and breast cancer risk among nondrinkers in their populations. The RR estimates among nondrinkers in these cohort studies were similar in magnitude to those among drinkers and in the entire cohort with statistical adjustment for alcohol consumption.

The results of studies of smoking and breast cancer stratified by genotype have also been largely inconsistent. Foremost among several possible explanations for this are inadequate sample size and lack of statistical power and precision. Our review of the available studies indicates that most were severely underpowered (Tables 1-6). Approximately 68% of the stratum-specific analyses shown in the tables had <80% power to detect an OR of at least 2.0. As with the studies reviewed here, small sample sizes have contributed to inconsistent findings more generally in studies of genetic variation and complex diseases (73). Another difficulty of comparing the results across studies is the lack of uniform methods of smoking characterization. The various smoking measures used in epidemiologic studies are often highly correlated, yet they may have distinct implications regarding disease etiology (1, 74).

Interpretation of the current literature is also complicated by the fact that many studies provide only a perfunctory rationale for selection of genes and gene polymorphisms. Whereas most studies have considered the role of selected genes in a given etiologic pathway and perhaps also the functional significance of a particular single nucleotide polymorphism, few have considered the known level of gene expression in breast tissue (75), the frequency of variant alleles and their differences among groups with known differences in disease incidence (73), or whether disease susceptibility may be conferred by a particular single nucleotide polymorphism, a haplotype, or by a specific combination of variants in several genes (76, 77). For example, two recent studies of breast cancer (that did not consider smoking) showed stronger positive associations with variants in carcinogen metabolism genes when considered jointly than individually (25, 78). Three studies reviewed here examined the association between smoking and breast cancer risk according to two genotypes (10, 34, 69), in each case examining different sets of genotypes in underpowered analyses. Thus, whether the association between smoking and breast cancer risk is modified by two or more interacting genotypes remains virtually unexplored.

Our summary analyses produced several statistically significant associations between smoking and breast cancer, and these results need to be viewed with caution. Meta-analyses, by virtue of their large sample sizes, can show relatively small statistically significant departures from null. Seemingly precise meta-RR estimates are not necessarily “true” or “noteworthy,” however, and the possibility of false-positive findings due to systematic error or chance requires careful consideration (79), particularly when effect sizes are modest. The relatively small number of relevant publications available for some genotypes also warrants caution given the possibility of publication bias (80). Our fail-safe n calculations indicate that the “file drawer problem,” the omission of negative unpublished results (81), may explain some of the statistically significant meta-RR estimates, such as those we observed in analyses according to GST and SOD2 genotypes.

These methodologic issues notwithstanding, some of the findings presented here suggest possible effect modification by genotype. For example, 14 epidemiologic studies to date have tended to show positive associations with long-term smoking among NAT2 slow acetylators, particularly among postmenopausal women, who are more likely than premenopausal women to be very long-term smokers. Considering that our fail-safe n estimates were in the 50 to 60 range, it seems unlikely that these findings can be explained by publication bias alone. Firozi et al. (82) showed that breast tissue from NAT2 slow acetylators had significantly higher levels of the diagonal radioactive zone (smoking-related) DNA adduct pattern than that from fast acetylators, supporting the hypothesis that women with NAT2 slow genotypes may suffer greater exposure to tobacco carcinogens and consequent higher risk of breast cancer. There was also a suggestion in two studies (42, 43) that long-term smoking increases breast cancer risk among women with CYP1A1-M1 variant genotype. That variant has also been linked to levels of a smoking-related benzo[a]pyrene-like adduct in breast tissue (82), an association that was particularly evident among women who also had the GSTM1-null genotype. These same genotypes, particularly when considered together, were also associated with higher levels of benzo[a]pyrene diol epoxide adduct in lung tissue (40) and leukocytes (83) of smokers.

That cigarette smoking can cause breast cancer in genetically susceptible women remains a reasonable hypothesis and an important public health concern. Studies that examined potentially effect-modifying effects of genotype suggest promising avenues for future studies. However, methodologic limitations, such as small sample sizes, varying definitions of smoking, and difficulties involving single nucleotide polymorphism selection, likely have contributed to the inconsistent findings. These methodologic issues should be addressed in future studies to help clarify the association between smoking and breast cancer.

We thank Dr. Heather Feigelson (American Cancer Society) for helpful comments on an earlier version of the article.