Smoking and Colorectal Cancer:  A Meta-analysis FREE

Although tobacco has been responsible for approximately 100 million deaths in the past century and 5.4 million in 2005 alone, there are still an estimated 1.3 billion smokers in the world. Currently, smoking prevalence has decreased in the United States and some other countries but has increased in less-developed world regions.¹

In the lung, where there is intimate exposure to inhaled tobacco smoke, about 80% of all primary cancers are attributable to smoking.² Smoking-related cancers are also frequent in the oropharynx and larynx, where there is direct contact with tobacco-related carcinogens. Smoking also increases the risk of cancer in organs such as the kidney, bladder, cervix, lower urinary tract, and pancreas, organs for which exposure to tobacco degradation products is indirect.³ With respect to the digestive tract, esophageal and gastric cancers have been strongly associated with tobacco, whereas the smoking–colorectal cancer (CRC) link remains controversial. Several large cohort studies linked smoking with CRC,^4- 6 and our group has recently reported a meta-analysis that revealed that smoking doubles the risk of colorectal polyps, known precursors for CRC.⁷ However, other studies have failed to detect a significant relationship between smoking and CRC.^8- 9 This inconsistency among studies could be partly explained by different study designs, population characteristics, and the heterogeneous treatment of the most likely confounders, such as diet, alcohol, physical activity, and body mass index (BMI, calculated as weight in kilograms divided by height in meters squared).¹⁰

Because smoking can potentially be controlled by individual and population-related measures, detecting a link between CRC and smoking could help reduce the burden of the world's third most common tumor, which currently causes more than 500 000 annual deaths worldwide.¹¹ In the United States alone, an estimate of approximately 50 000 deaths from CRC would have occurred in 2008.¹²

We conducted a meta-analysis with the following aims: (1) to review and summarize published data examining the link between smoking and CRC incidence and mortality; (2) to measure the smoking-CRC relationship according to different characteristics of the study populations, study designs, and CRC subsites; and (3) to study dose-response patterns of tobacco exposure on the risk of CRC.

We performed the following literature search to May 2008 using PubMed and EMBASE, without restrictions: ([Smoke or cigarette or tobacco or smoking] and Cancer and [Colon or Rectum or Colorectal or Colorectum or Colon rectum]) or (“Colorectal cancer” [MeSH Major Topic ] and “epidemiologic studies” [Mesh Terms ]) (Medical Subject Headings). We also identified the most cited articles on the topic using ISI Web of Knowledge (Science Citation Index Expanded; Journal Citation Report) and reviewed articles quoting them. In addition, we reviewed the references of all articles of interest and of the International Agency for Research on Cancer Monographs on tobacco smoke and involuntary smoking¹⁰ to identify additional relevant studies. Only reports fulfilling the following inclusion criteria were included in the meta-analysis.

When available, we used adjusted risk estimates and those based on population-based controls. Articles were reviewed and data were extracted and cross-checked independently by 3 investigators (E.B., S.I., and S.R.). Any disagreement was resolved by consensus among the 3.

An important study based on the British Doctors study¹³ was excluded from the meta-analysis because no measure of variability for the mortality rate was reported, but its results were used for other analyses.

The RR was used as a measure of the association between cigarette smoking and CRC. For case-control studies, the odds ratio (OR) was used as estimates of the RR because CRCs are sufficiently rare.¹⁴ When cohort studies reported only crude data and no information on person-years, we treated it as a control study using noncases as controls.

We used random rather than fixed-effects models to estimate pooled RRs in order to take into account the heterogeneity, however small, of the risk estimates and therefore to be more conservative.

Homogeneity of effects across studies was assessed using the χ² statistic and quantified by I², which represents the percentage of total variation across studies that is attributable to heterogeneity rather than chance.¹⁵ Subgroup analyses and meta-regression models were carried out to investigate potential sources of between-study heterogeneity.

When several risk estimates were present in a single study (ie, separate estimates for proximal and distal colon), we adjusted the pooled estimates for intrastudy (or within-study) correlation.¹⁶

In the dose-response analysis, we considered cigarettes per day, pack-years, and duration of smoking as explanatory variables. Because for many studies continuous exposures were reported as categorical data with a range, we assigned the mid-point of the range as the average exposure. For the highest open categories, we considered 60 cigarettes per day, 60 pack-years, and 60 years of duration as the maximum. In pooling dose-response data, we took into account correlation between RRs within the same study, using the method described by Greenland and Longnecker.¹⁷ Both linear and nonlinear models were fitted and evaluated on the logarithm of the RR. Nonlinear trends were evaluated using fractional polynomial curves of the second order.¹⁸

To consider differences among studies as a further source of random variability, an additional component of the variance was estimated and added in weighting each observation.¹⁹ The model minimizing Akaike information criterion (AIC), a penalized likelihood that takes into account the number of parameters estimated in the model, was used as a general rule in the model choice.

Publication bias was evaluated by funnel plots and quantified by the tests developed by Egger et al²⁰ and Begg et al.²¹ All analyses were performed with SAS software version 8.02 (SAS Institute Inc, Cary, North Carolina).

Detailed search steps are described in Figure 1. Briefly, from the initial literature search we identified and screened 1663 abstracts. Two hundreds forty-one articles were considered of interest and full text was retrieved for detailed evaluation. References of all 241 articles were reviewed, and 6 additional relevant studies were identified. One hundred twenty-six of these 247 articles were subsequently excluded from the meta-analysis (80 did not satisfy the inclusion criteria, 46 were based on the same study populations).

One-hundred six independent observational studies that met the inclusion criteria were included in the final analysis of incidence (Table 1 and Table 2 and eTable 1). Overall, the meta-analysis is based on a total of 39 779 incident cases.

Studies were published between 1969 and 2008; 31 were conducted in North America, 39 in Europe, 33 in Asia, and 3 in other areas. One hundred four were written in English, 1 in Korean, and 1 in French. Sixty-five studies (61%) provided at least 1 adjusted risk estimate, and 43 (66%) of them reported an adjusted estimate for at least 1 of the main lifestyle-related potential confounders (diet, BMI, alcohol consumption, and physical activity). Detailed information on adjustments is reported in Table 1 and Table 2.

Quantitative Data Synthesis . Initially, we considered all studies providing either adjusted risk estimates or only crude data from which we calculated unadjusted risk estimates. Although the pooled RR for ever vs never smokers was similar for adjusted and nonadjusted studies (adjusted RR, 1.18; 95% CI, 1.11-1.25; Table 3 and Figure 2; unadjusted RR, 1.11, 95% CI, 1.05-1.31; P = .45), unadjusted risk estimates were highly heterogeneous (χ²P <.01; I², 77%). Because adjusted risk estimates are generally more reliable and unbiased than unadjusted ones, we decided to limit the main analysis to studies providing adjusted risk estimates, which did not show evidence of heterogeneity (χ²P = .17; I², 28%), and to report an extended analysis based on all studies as supplementary material.

The pooled RRs for current vs never smokers and former vs never smokers were respectively 1.07 (95% CI, 0.99-1.16) and 1.17 (95% CI, 1.11-1.22; Table 3). Sex was not a significant source of heterogeneity, while the risk estimates were higher for rectal than for colon cancer among current smokers (P = .02) and ever smokers (borderline significant, P = .08). When we considered colon subsites, we found a nonsignificant increased risk for cancer of the proximal colon compared with the distal colon among ever and former smokers. When considering adjusted and unadjusted estimates together, cohort studies showed higher risk estimates compared with case-control studies (P = .07; eTable 2). We found higher risks in studies in which controls underwent colonoscopy than in studies in which they did not. This was statistically significant for ever smokers (eTable 2) and for former smokers (Table 3). Risk estimates were systematically higher in population-based than in hospital-based studies, but the difference was not statistically significant.

After evaluating dose-response patterns for cigarettes per day, pack-years, and duration of smoking for ever vs never smokers, we observed a linear increase in risk with increasing smoking consumption The risk increased by 7.8% (95% CI, 5.7%-10.0%) for every additional 10 cigarettes per day or by 4.4% (95% CI, 1.7%-7.2%) for every additional 10 pack-years (for example an individual who smoked 1 pack of cigarettes per day for 50 years or 2 packs per day for 25 years has a 24% increased risk of developing CRC compared with someone who never smoked). We also observed a nonlinear increase in risk with increasing smoking duration (Figure 3). The risk starts to increase after approximately 10 years of smoking and reaches statistical significance after 30 years.

From 19 cohort studies that reported information on person-years in smokers and nonsmokers, we could calculate absolute annual rates of CRC cases: 65.5 cases per 100 000 in smokers and 54.7 per 100 000 in nonsmokers, corresponding to an absolute risk increase of 10.8 cases per 100 000 (95% CI, 7.9-13.6).

The effect of smoking cessation was assessed in 7 studies. Among the 5 studies that reported a significant effect of smoking, 3 studies^22- 24did not show a clear trend in risk across smoking cessation categories while 2 studies^25- 26 reported a risk reduction with increasing length of cessation.

Four studies reported information on microsatellite instability (MSI) status of CRC. Two studies^27- 28 presented a stronger smoking-CRC association for high-MSI tumors than for low-MSI and microsatellite-stable tumors. Yang et al²⁸ reported an OR of 6.64 (P < .01) for high-MSI tumors vs 1.93 (P = .13) for low-MSI or microsatellite-stable tumors when comparing current with never smokers and an OR of 2.68 (P = .01) for high MSI vs 1.60 (P = .08) for low-MSI or microsatellite-stable tumors when comparing former to never smokers. Chia et al²⁷ reported an OR of 2.6 (95% CI, 1.7-4.1) for high-MSI vs 1.5 (95% CI, 1.2-2.0) for low-MSI or microsatellite-stable tumors for current vs never smokers and an OR of 1.8 (95% CI, 1.3-2.6) for high-MSI vs 1.2 (95% CI, 1.0-1.5) for low-MSI or microsatellite-stable tumors for former vs never smokers. Slattery et al²⁹ reported higher risk of developing MSI rather than microsatellite-stable tumors for smokers: OR, 2.3 (95% CI, 1.5-3.3) for current smokers and OR, 1.3 (95% CI, 1.0-1.8) for former smokers. Diergaarde et al³⁰ reported a similar risk for smokers and nonsmokers of developing high-MSI rather than microsatellite-stable tumors: OR, 0.8 (95% CI, 0.3-1.8) for ever vs never smokers. Because these 4 studies compared different MSI groups, we could not pool risk estimates.

Publication Bias and Sensitivity Analysis . Neither funnel plots nor Egger and Begg tests showed evidence of publication bias for ever smokers (Egger test, P = .92; Begg test, P = .63 for all estimates; Egger test, P = .42; Begg test, P = .84 for adjusted estimates; eFigure 1A, and for former smokers (Egger test, P = .20; Begg test, P = .40), although the Egger test suggested the presence of publication bias for current smokers (Egger test, P<.01; Begg test, P = .77). In a European hospital-based study,³¹ the adjusted risk estimate for current smokers (RR, 0.7; 95% CI, 0.6-0.8) was much lower than the pooled risk estimate. After excluding this single study, there was less evidence of publication bias (Egger, P = .39) and the pooled risk estimate reached statistical significance (RR, 1.10; 95% CI, 1.01-1.17); furthermore, the unexpected significant protective effect of current smoking in Europe disappeared (RR, 1.00; 95% CI, 0.87-1.13).

Because some studies reported risk estimates for current but not for former smokers or vice versa, in order to allow an unbiased comparison between the 2 smoking classes, we also computed the RR for current smokers (1.07, 95% CI, 1.00-1.14) and the RR for former smokers (1.18; 95% CI, 1.11-1.26; P < .01) from the 42 studies reporting both estimates.

Seventeen English-language cohort studies, published between 1990 and 2008, met the inclusion criteria and were included in the final analysis of mortality (Table 1): 7 were conducted in North America, 3 in Europe and 7 in Asia. In regard to the main potential confounders, all reported age- and sex-adjusted risk estimates and 10 (59%) reported an adjusted estimate for at least 1 of the main lifestyle-related potential confounders. Additional information on adjustments is reported in Table 1.

The pooled risk estimates are reported in Table 4: the RR for current vs never smokers, based on 14 studies, was 1.28 (95% CI, 1.15-1.42). The RR for former vs never smokers, based on 12 studies, was 1.23 (95% CI, 1.14-1.32). Six studies reported an estimate for ever vs never smokers. For 9 additional studies, we combined estimates for current and former smokers to obtain an estimate for ever smokers. The RR for ever vs never smokers, based on combining these 15 studies, was 1.25 (95% CI, 1.14-1.37; Table 4; Figure 4). We found some evidence of heterogeneity among the studies for ever (P = .02; I², 42%) and current smokers (P = .07; I², 41%) but not for former smokers (P = .56; I², 0%).

From 10 cohort studies that reported information on person-years in smokers and nonsmokers, we could calculate absolute annual rates of CRC deaths: 41.3 deaths per 100 000 in smokers and 35.3 per 100 000 in nonsmokers, corresponding to an absolute risk increase of 6.0 deaths per 100 000 person-years (95% CI, 4.2-7.6).

When stratifying for tumor site, risk estimates were statistically significant for both cancer of the colon and rectum in all 3 smoking classes and significantly higher for cancer of the rectum than of the colon, for current and ever smokers. There were not enough data to study proximal and distal colon sites separately. Sex, publication year, and geographical areas were not sources of heterogeneity.

We observed a linear increase in risk in CRC mortality with increasing number of cigarettes per day smoked: the risk increased by 7.4% (95% CI, 5.7%-9.2%) and 10.6% (95% CI, 8.7%-12.5%) for every additional 10 cigarettes per day for ever and current smokers, respectively. For duration of smoking, the risk increased linearly by 9.5% (95% CI, 5.5%-13.7%) for every additional 10 years of smoking (ie, an individual who smoked for 50 years has a 57.5% increased risk of developing CRC compared with a never smoker). We did not have sufficient data to perform the dose-response analysis for pack-years and to examine nonlinear trends as we did for incidence data.

The effect of smoking cessation on CRC mortality was assessed in 4 studies. In 2 studies smoking was not significantly associated with CRC mortality. Rohan et al³² reported a hazard ratio (HR) of 1.74 (95% CI, 0.91-3.33) for persons who stopped smoking 1 to 10 years before enrollment and 1.33 (95% CI, 0.70-2.57) for persons who stopped smoking 11 years or more before enrollment. Ozasa et al³³ reported that HRs were 2.05 (95% CI, 1.23-3.42) for those who stopped smoking in fewer than 5 years, 0.96 (95% CI, 0.55-1.68) for those who stopped between 5 and 14 years, and 1.27 (95% CI, 0.74-2.17) for those who stopped for least 15 years. In the Cancer Prevention Study II³⁴ and the Nurses' Health Study,³⁵ the 2 largest studies in the meta-analysis, an effect of smoking cessation reduced the risk of CRC mortality. In the first cohort, the risk of dying from CRC decreased with increasing time since cessation: RRs were 1.32 (95% CI, 1.19-1.47) for those who had stopped smoking in fewer than 10 years, 1.20 (95% CI, 1.08-1.35) for those who had stopped between 11 and 19 years, and 1.04 (95% CI, 0.94-1.16) for those who had stopped for at least 20 years (P for trend <.001). In the second cohort, a significant decrease of CRC risk in former smokers was observed after 20 years of cessation: the RR was 0.70 (95% CI, 0.53-0.93) with current smokers as the reference group.

Serial reports of the results of 2 large cohort studies—the British Doctors study^13,36- 37 and the US Veterans study^6,38- 40—were published over time allowing the possibility of studying changes in risk with increasing duration of follow-up (Figure 5). Although the CRC mortality risk was not or only modestly increased in the early reports of the US veteran study, it became significant (RR, 1.3; 95% CI, 1.23-1.42) in the 26-year report, similar to that reported in the 20 years' follow-up report of the British Doctors study³⁶ (RR, 1.35; 95% CI not available). After 50 years of follow-up the mortality rate ratio in the British Doctors study increased to 1.5 for ever vs never smokers.

Publication Bias . The funnel plot (eFigure 1B) and Egger and Begg tests did not show evidence of publication bias based on smoking status (Egger: current, P = .74; former, P = .46; and ever, P = .57; Begg: current, P = .53; former, P = .50; and ever, P = .68).

Through the past 2 decades, numerous publications have shown a consistent association between cigarette smoking and colorectal adenomatous polyps. In a recent meta-analysis,⁷ we reported a pooled risk estimate of 1.82 (95% CI, 1.65-2.00) for ever vs never smokers. In contrast, the association between cigarette smoking and CRC has been more controversial. In the present analysis, we found statistically significant excess risk of CRC for smokers vs nonsmokers, although the magnitude of the association is lower than that observed for adenomas: 1.18 (95% CI, 1.11-1.25) for CRC incidence and 1.25 (95% CI, 1.14-1.37) for CRC mortality.

Because the adenoma-CRC pathway accounts for a large proportion of CRCs, one should expect a similar smoking-related risk for adenomas and CRCs. Common explanations for this discrepancy are (1) the long latency period between smoking and CRC, which is not accounted for in many studies; (2) the effect of cigarette smoking may be stronger on nonprogressing adenomas; and (3) the inclusion of a high proportion of participants with adenomas in the unscreened control groups of most CRC case-control studies could have biased the estimates toward the null hypothesis of no association.^41- 42

The results of the present study support the first hypothesis, showing a statistically significant increase of CRC incidence after 30 years of smoking. In addition CRC mortality risk estimates increased with increasing follow-up duration in 2 large cohorts.^6,13 Although smoking was more strongly associated with high-risk than with low-risk adenomas in our meta-analysis, arguing against the second hypothesis, we found a stronger association between cigarette smoking and CRC in studies in which controls underwent colon examination, as observed for adenomas.⁷

As an alternative to the adenomatous polyps pathway, up to 15% of CRC can arise from serrated adenoma.^43- 45 Smoking seems to be more strongly associated with CRCs displaying MSI, the hallmark of the serrated pathway, than CRCs not displaying MSI. Three studies reported higher risk for microsatellite-unstable tumors than for microsatellite-stable tumors.^27- 29 One single study³⁰ reported similar risk estimates for stable and unstable tumors. Furthermore, MSI is a hallmark of hereditary nonpolyposis colorectal cancer syndrome, or Lynch syndrome, which it is thought to account for 1% to 5% of all CRCs,^46- 49 and smoking has been strongly associated with CRC in patients with hereditary nonpolyposis colorectal cancer syndrome.^50- 51

With respect to tumor location, we found higher smoking-related risk estimates for rectal and proximal colon cancer compared with distal colon cancer. This could be due to the differential anatomical location of serrated CRC. Although nonserrated neoplasia tend to have no site predilection,^7,52- 53 3 studies^54- 56 reported that serrated neoplasia arise more frequently in the proximal colon and in the rectum; 3 other studies^28,57- 58 reported that tumors involving MSI arise more frequently in the proximal colon than in the distal colon.

The incidence of CRC appears to be higher in former smokers than in current smokers, possibly due to a higher total exposure in former smokers. However, we could not evaluate this possibility because studies did not report data on pack-years and smoking duration separately for current and former smokers.

Risk estimates for CRC mortality were consistently higher than for incidence. Given the long latency period between smoking and CRC, this result could be due to longer follow-up, hence longer smoking exposure in mortality cohort studies. Also, mortality studies are easier to update through record linkage with national mortality registries than incidence studies. Another possible explanation is that smoking could cause more aggressive tumors. In a previous meta-analysis, we observed a stronger association between smoking and high-risk adenomas than between smoking and low-risk adenomas.⁷ Furthermore, some authors^59- 62 reported a significantly increased probability of diagnosing advanced-stage CRC smokers. However these findings could be due to different behavioral correlates of smoking, ie, tendency for smokers to delay seeking medical care.

In this meta-analysis we reviewed and summarized the extensive but controversial evidence on the association between smoking and CRC.¹⁰ The large number of studies and the consequent possibility of subgroup analyses permitted us to better understand the association in different subgroups. For incidence data, we could perform a meta-analysis based on studies that allowed us to adjust for confounding variables. For example, we were able to investigate the potential confounding of alcohol, a suggested risk factor for CRC⁶³; we found no difference between estimates adjusted for alcohol and estimates adjusted for other variables.

Furthermore, data from the British Doctors study,¹³ which was excluded from the meta-analysis because information to calculate uncertainty was absent, support our findings of a positive association between smoking and CRC mortality. This large prospective study found a mortality rate ratio of 1.5 after 50 years of follow-up for ever vs never smokers.

A possible limitation of our study is the heterogeneity of the studies in regard to adjustments of the estimates for potential confounders, such as diet, BMI, alcohol consumption, and physical activity, which limits the generalization of the results. The inadequate adjustment for various potential confounders, which could account for some of the small increases in risk that appear to be associated with smoking, explains why the International Agency for Research on Cancer did not consider the smoking-CRC evidence to be sufficient to establish causality.¹⁰ We tried to overcome this problem by providing subgroup analyses in more homogeneous subsets of studies. However, our decision to select only adjusted estimates for the main analysis of incidence, and therefore to exclude studies not targeted on the smoking-CRC relation, might have biased our conclusions.

Smoking has not been considered so far in the stratification of individuals for CRC screening.⁶⁴ However, several studies reported that CRC occurs earlier in smokers, particularly in those with heavy tobacco consumption,^62,65- 67 and our previous⁷ and present findings provide strong evidence of the detrimental effect of cigarette smoking on the development of adenomatous polyps and CRC. We believe that smoking represents an important factor to consider when deciding on the age at which CRC screening should begin, either by lowering the age in smokers or increasing the age in nonsmokers.

In conclusion, our results strongly suggest that cigarette smoking is significantly associated with both CRC incidence and mortality.