Cigarette Smoking Increases Risk of Barrett’s Esophagus: an Analysis of the Barrett’s and Esophageal Adenocarcinoma Consortium

Abstract

Background & Aims

Cigarette smoking has been implicated in the etiology of esophageal adenocarcinoma, but it is not clear if smoking is a risk factor for Barrett’s esophagus (BE). We investigated whether tobacco smoking and other factors increase risk for BE.

Methods

We analyzed data from 5 case-control studies included in the international Barrett’s and Esophageal Adenocarcinoma Consortium. We compared data from subjects with BE (n=1059) with those from subjects with gastroesophageal reflux disease (GERD controls, n=1332) and population-based controls (n=1143), using multivariable logistic regression models to test associations with cigarette smoking. We also tested whether cigarette smoking has synergistic effects with other exposures, which might further increase risk for BE.

Results

Subjects with BE were significantly more likely to have ever-smoked cigarettes than the population-based controls (odds ratio [OR]=1.67; 95% confidence interval [CI], 1.04–2.67) or GERD controls (OR=1.61; 95% CI, 1.33–1.96). Increasing pack-years of smoking increased the risk for BE. There was evidence for a synergy between ever-smoking and heartburn or regurgitation; the attributable proportion of disease among individuals who ever smoked and had heartburn or regurgitation was estimated to be 0.39 (0.25–0.52).

Conclusions

Cigarette smoking is a risk factor for BE. The association strengthened with increased exposure to smoking until ~ 20 pack-years, when it began to plateau. Smoking has synergistic effects with heartburn or regurgitation, indicating that there are various pathways by which tobacco smoking might contribute to the development of BE.

Introduction

Barrett’s esophagus is a columnar metaplasia of the distal esophagus associated with a 10- to 55-fold increased risk of esophageal adenocarcinoma ^1–7. Barrett’s esophagus ^8–11 and esophageal adenocarcinoma ^12–14 have been increasing in incidence, particularly in developed countries with predominantly Caucasian populations. For example, in the United States esophageal adenocarcinoma in whites has increased from 0.4 to more than 3 per 100,000 person-years during the last 35 years—a 650% increase ^{12, 15}. This increasing incidence is not solely due to changes in diagnostic practice, and has been attributed to temporal changes in exposure to risk factors ¹⁶.

The known risk factors for Barrett’s esophagus and esophageal adenocarcinoma are few and include gastroesophageal reflux ^{17, 18} and increasing body mass index (BMI) ^19–21. Cigarette smoking has also been implicated in the etiology of esophageal adenocarcinoma ²², but whether this is because smoking is a risk factor for early events in the carcinogenic pathway (i.e. Barrett’s esophagus) or for later events, such as the transformation of Barrett’s esophagus to cancer, is unclear given the conflicting findings of previous studies of Barrett’s esophagus risk factors, with some studies demonstrating a positive association between Barrett’s esophagus and cigarette smoking ^{18, 23–27} and others not ^28–32.

The inability to ascertain what, if any, relationship exists between Barrett’s esophagus and smoking has been due, in part, to imprecision rendered by limited numbers of subjects available for analysis in individual studies. This limitation has also reduced the ability to discern interactions between exposures; if tobacco smoking does increase risk of Barrett’s esophagus it could do so primarily through genotoxic mechanisms or by promoting GERD. Refining our understanding of the potential mechanism(s) of association is important with regard to the efficacy of preventative actions.

To better understand the relationship between Barrett’s esophagus and one of its few potentially modifiable risk factors, we assessed whether cigarette smoking was associated with Barrett’s esophagus, and the potential mechanism of association, by pooling, harmonizing, and analyzing individual patient data from five case-control studies in the international Barrett’s and Esophageal Adenocarcinoma Consortium (BEACON, http://beacon.tlvnet.net/).

Methodology

Study Population

The BEACON consortium was formed in 2005 with support from the U.S. National Cancer Institute. It is composed of investigators from around the world and brings together population-based case-control and cohort studies of esophageal adenocarcinoma and Barrett’s esophagus. The primary objectives of BEACON are to facilitate well-powered, combined investigations of risk factors in relation to these diseases, as well as helping the development of new studies of etiology, prevention and survival.

The five Barrett’s esophagus case-control studies included in this BEACON analysis, with abbreviated names shown in italics, were: FINBAR (Factors INfluencing the Barrett’s/Adenocarcinoma Relationship) study, based in Ireland ³³; Epidemiology and Incidence of Barrett’s Esophagus study nested within Kaiser Permanente Northern California (KPNC), USA ³⁴; Study of Reflux Disease, based in western Washington State, USA ³⁵; Study of Digestive Health, based in Brisbane, Australia ²⁶; and Epidemiologic Case-Control Study of Barrett’s Esophagus based at The University of North Carolina at Chapel Hill (UNC-Chapel Hill), USA. For comparison with Barrett’s esophagus cases, two control groups were available: gastroesophageal reflux disease (GERD) and population-based. There are advantages for each of these comparison groups. GERD controls represent the population undergoing endoscopy from which Barrett’s esophagus cases are diagnosed. Therefore, comparisons between these two groups are less affected by potential ascertainment bias than comparisons between Barrett’s esophagus cases and population-based controls, insofar as it inherently controls for known and unknown potentially confounding factors associated with being referred for and undergoing an endoscopic procedure. In addition, since most cases are identified in the course of investigating gastroesophageal reflux, the use of GERD controls, to some degree, inherently adjusts for the presence, although not severity, of symptomatic gastroesophageal reflux. The major advantage of the population-based control group is that it enables the assessment of gastroesophageal reflux as both an effect-measure modifier and independent risk factor, while also being representative of the local population from which the Barrett’s esophagus cases are referred and diagnosed. Studies which have conducted endoscopy on random samples of the general population provide more in-depth information on the relative advantages and disadvantages of each of these two control groups ^{36, 37}. All five studies contributed individual patient data to the GERD control group and four of the studies contributed individual patient data to the population-based control group. Study-specific definitions of the case and control groups are detailed in Table 1.

Table 1.

Recruitment, criteria and numbers for analysis in each of the Barrett’s esophagus case-control studies.

Study	Location	Barrett’s Esophagus Cases (n)	Barrett’s Esophagus Case Definition	GERD Controls (n)	GERD Control Definition	Population based Controls (n )	Population-based Control Definition
KPNC	Northern California, USA	187	Incident cases recruited from KPNC, October 2002 – September 2005. SIM & any length of macroscopic BE. Exclusions: gastric-type metaplasia only, indicative biopsy from squamocolumnar junction only.	172	GERD: prior to their index date: a GERD-related diagnosis code (ICD-9: 530.11 [reflux esophagitis] or 530.81 [gastroesophageal reflux]); a prescription ≥90 day supply of a H2RA or PPI in the previous year; recent esophagogastroduodenoscopy that was negative for esophageal columnar metaplasia (macroscopic or histologic) of any type. Frequency matched on age at the index date, sex, and geographic region to the distribution of BE cases. Exclusions: previous BE diagnosis.	185	KPNC members without an electronic diagnosis of BE at the time the BE cases were identified, frequency matched on age at index date, sex, and geographic region to the distribution of BE cases.
Study of Digestive Health	Brisbane, Queensland, Austraia	362	Incident cases of BE, or of dysplasia in previously diagnosed BE, recruited from two major private pathology laboratories and the single public pathology laboratory serving metropolitan Brisbane, Australia, February 2003 – June 2006. SIM & any length of macroscopic BE.	287	GERD: acute inflammatory changes on histology consistent with gastroesophageal reflux. Exclusions: Any other major pathology identified on endoscopy/histology.	562	Population-based controls, randomly selected from the Australian Electoral roll. Frequency matched on age and sex to a case series. Control participants had to reside within the same geographical region as the BE cases and GERD controls.
Study of Reflux Disease	Washington, USA	149	Incident cases recruited from 5 community gastroenterology clinics, October 1997 – September 2000. Endoscopy referral due to chronic GERD symptoms. SIM.	347	GERD: referred for endoscopy due to reflux symptoms, but negative for SIM. Frequency-matched on the month of biopsy and clinic to the distribution of BE cases.	172	Population-based controls, selected from geographic areas in close proximity to those of BE cases using a modified version of the Waksberg random-digit dialing method. Individually matched to BE cases on age (+/−3 years) and sex.
FINBAR	Northern Ireland and Republic of Ireland	187	Incident and prevalent cases recruited from the island of Ireland, March 2002 – December 2004. SIM & ≥3 cm of macroscopic BE. Exclusions: dysplasia.	217	Erosive esophagitis (incident and prevalent diagnoses):macroscopically diagnosed (grades 2–4 in the Savary Miller/Hetzel-Dent classification/grades B, C, or D in the LA classification), recruited from Northern Ireland, September 2004 – July 2005. Exclusions: BE, macroscopic or histologic evidence of infection, dysmotility, gastric outlet obstruction. Frequency matched on age and sex to the distribution of EA patients.	224	Population-based controls. Northern Ireland: selected at random from General Practice Master Index of each province. Republic of Ireland: selected at random from 4 general practices to reflect the urban/rural distribution of EA patients in the Republic of Ireland. Frequency matched on age and sex to the distribution of EA patients. Recruited from the island of Ireland between March 2002 – December 2000.
UNC-Chapel Hill	North Carolina, USA	174	Incident and prevalent cases recruited from an endoscopy clinic-based study between 2001–2006. SIM & any length of macroscopic BE.	309	GERD diagnosed by a physician: referred for endoscopy due to reflux symptoms, but negative for macroscopic BE & SIM. Sampled in 2:1 ratio to BE subjects.	0	N/A
		1,059		1,332		1,143

Abbreviations: BE, Barrett’s esopagus; EA, esophageal adenocarcinoma; GERD, gastroesophageal reflux disease; SIM, specialized intestinal metaplasia.

In total, the five studies provided 1,320 cases of Barrett’s esophagus, 1,659 GERD controls, and 1,434 population-based controls. For this analysis, and if a study provided such data, we excluded individuals who had ever-smoked pipe tobacco or cigars (156 Barrett’s esophagus cases, 132 GERD controls, 153 population-based controls), because comparing cigarette smokers with those who do not use other forms of tobacco provides a more accurate estimate of the effect of cigarette smoking. Ever-smoking of pipe tobacco or cigars was defined as meeting a study-specific low threshold exposure (a period of ≥6 months or ≥20 times over the life-course). Due to the relatively small number of non-white Barrett’s esophagus cases remaining (17 black, 31 Hispanic, 39 other, 18 missing), we restricted our analysis to white study participants. After exclusions, there remained 1,059 Barrett’s esophagus cases, 1,332 GERD controls, and 1,143 population-based controls for analysis. Data acquisition and data pooling for each study were approved by the Institutional Review Board or Research Ethics Committee of the institute(s) sponsoring the study.

Analytic Variables

The primary exposure variables were cigarette smoking status (ever vs. never) and total cigarette smoking exposure (pack-years; 0, <15, 15–29, 30–44, ≥45). Additional exposure variables included duration of cigarette smoking (<30 years, ≥30 years), cigarette smoking intensity (<1, 1, and >1 packs/day), age of cigarette smoking initiation (<17, ≥17 years), and duration of cigarette smoking cessation (<20 years, ≥20 years). Cigarette smoking intensity and cigarette smoking duration in the UNC-Chapel Hill study were ascertained in categories and were thus recoded to the median of the categories using the distributions of the other four studies combined. Ever-cigarette smoking was defined as either low threshold exposures (≥100 cigarettes, >=20 packs of cigarettes, 1 cigarette a day for six months of longer) or by asking whether the patient had ever-smoked. Covariates assessed for inclusion in regression models included: age; sex; BMI (weight divided by square of height [kg/m²]); education; alcohol; fat, and transfat consumption; calories per day; meat, vegetable, and fruit servings per day; fiber consumption; heartburn, and regurgitation (population-based control models only); esophagitis; Helicobacter pylori seropositivity; hiatal hernia; and medication use (non-steroidal anti-inflammatory drugs [NSAIDs], antacids, proton pump inhibitors [PPIs], H2-receptor antagonists [H2RAs]). A covariate was included in the fully adjusted models if it altered an estimate by >10% or it was considered a known confounder (age, sex, BMI, and education).

Statistical analysis

We used a two-step analytic approach. First, study-specific odds ratios (ORs) and 95% confidence intervals (CIs) for an exposure-outcome relationship were estimated from multivariable logistic regression models. Second, the study-specific ORs were combined using fixed-effects and random-effects meta-analytic models to generate summary ORs; both approaches gave similar estimates of association, thus we present only the random-effects models herein as such models are usually more conservative ³⁸. A study was excluded from the second-step of a specific variable’s analysis if the logistic regression model failed due to instability. The I² value and its 95% uncertainty interval were used to estimate the percentage of total variation across studies due to heterogeneity ³⁹. An I² statistic of 0% indicates no observed heterogeneity that cannot be attributed to chance, whereas larger values indicate increasing heterogeneity.

Exposure variables were assessed in relation to the outcome of Barrett’s esophagus using two comparison groups: GERD controls and population-based controls. Continuous variables were categorized to allow for nonlinear effects, for ease of interpretation, and to reduce the effect of any outliers; exceptions to this were the use of continuous variables for trends, product-terms, and spline models. Minimally adjusted models included the covariates age (years; <50, 50–59, 60–69, ≥70) and sex. Fully adjusted models also included BMI (<18.5, 18.5–24, 25–29, 30–34, 35–39, ≥40 kg/m²) and education (categorical: school only, tech/diploma, university; unavailable and so unadjusted for in UNC-Chapel Hill study). These models were also stratified by sex, BMI, and heartburn or regurgitation (population-based control comparisons only) to assess relationships (ORs) for effect-measure modification, with p values estimated via random effects meta-analysis of study-specific estimated effects of product-terms (e.g., ever-smoke x sex). Heartburn was generally described to the patient as having ever experienced burning pain or discomfort behind the breast bone while regurgitation was generally described as food or stomach fluid coming back up into the mouth accompanied with a sour-taste; KPNC excluded symptoms within 1 year prior to diagnosis of Barrett’s esophagus and FINBAR 5 years. In addition, FINBAR required symptoms to be frequent (more than 50 times per year/about once a week). Models of the additional exposures (cigarette smoking duration, intensity, initiation, and cessation) were also adjusted for total exposure (pack-years of cigarette smoking); because these variables contribute to total exposure, association testing without adjustment for total exposure could be misleading ^{40, 41}.

Spline models ⁴² were used to generate plots of the relationship between continuous pack-years of cigarette smoking and risk of Barrett’s esophagus, compared with each control group and adjusted for age, sex, BMI, and study using the pooled dataset of individual patient data. Restricted cubic spline models allow for easy visualization of non-linear relationships between an exposure and an outcome ^{43, 44}—in this case, cigarette smoking and Barrett’s esophagus. These models were plotted using a linear scale on the x-axis (pack-years of cigarette smoking) and a logarithmic (base 10) scale on the y-axis (OR).

To determine whether cigarette smoking biologically interacts with other exposures in relation to risk of Barrett’s esophagus, we tested for departure from additivity. Positive departure from additivity implies that the number of cases attributable to two exposures in combination is larger than the sum of the numbers of cases that would be caused by each exposure separately. The covariates tested for biological interaction with ever-cigarette smoking were BMI (<27.5, ≥ 27.5 kg/m²), heartburn and regurgitation (population-based control comparisons only), alcohol, H. pylori, and NSAIDs. For each combination of variables, we generated four exposure categories; using BMI as an example: A=never-smoker, low BMI; B=smoker, low BMI; C=never-smoker, high BMI; D=smoker, high BMI. These variables were modeled in the pooled dataset of individual patient data using logistic regression adjusted for age, sex, BMI, education, and study. Assuming that the OR approximates the relative risk, the output from these models was used to estimate three interaction statistics: interaction contrast ratio (ICR), attributable proportion (AP), and synergy index (S) ^{45, 46}. When the ICR and AP ≠ 0, and S ≠ 1 there is evidence for departure from additivity (biological interaction). ICR is the excess risk due to interaction relative to the risk without either exposure. AP is the proportion of disease attributable to interaction among individuals with both exposures. S is the ratio of the observed excess risk in individuals exposed to both factors relative to the expected excess risk assuming that both exposures are independent risk factors (i.e., under the assumption of no additive interaction). Confidence intervals for these metrics were estimated using the delta method ⁴⁵.

All analyses were performed using STATA software, version 11.1 (StataCorp LP, College Station, TX). All statistical tests were two-sided and p values less than 0.05 were considered to be statistically significant.

Results

Descriptors of cases and controls included in the analysis are shown in Table 2. The population-based control distributions were more similar to the cases in terms of age and sex than the GERD controls, and this is likely due to three of the four studies with population-based controls having matched on these variables to the Barrett’s esophagus case group; GERD controls were matched to the Barrett’s esophagus group on age and sex in only one study (Table 1). However, in other respects, such as BMI and alcohol, GERD controls had distributions more similar to the Barrett’s esophagus group, compared with the population-based control group.

Table 2.

Descriptors of Participants Selected for Analysis of Cigarette Smoking in Relation to Barrett’s Esophagus in the International BEACON Consortium

Variable	BE (n=1,059)	GERD Controls (n=1,332)		Population-based Controls (n=1,143)
	Mean (SD)/Percentage (95%CI)	Mean (SD)/Percentage (95%CI)	P value	Mean (SD)/Percentage (95%CI)	P value
Age (years)	59 (12)	54 (14)	<0.001	59 (12)	0.874
Sex (% male)	68 (66–71)	52 (50–55)	<0.001	64 (62–67)	0.039
Education (% college or university)	66 (63–69)	64 (61–67)	0.243	67 (64–69)	0.763
Body mass index (kg/m2)	28.4 (5.1)	28.3 (5.3)	0.759	27.6 (5.1)	<0.001
Heartburn (%)	83 (81–86)	n/a	n/a	47 (44–50)	<0.001
Regurgitation (%)	76 (74–79)	n/a	n/a	44 (42–47)	<0.001
Heartburn or Regurgitation (%)	91 (89–93)	n/a	n/a	57 (54–60)	<0.001
Alcohol (%)	77 (75–80)	76 (73–78)	0.328	82 (80–84)	0.005
Cigarette smoking (%)	63 (60–65)	50 (47–52)	<0.001	49 (46–51)	<0.001
Pack-years	31 (26)	26 (25)	<0.001	29 (29)	0.355
Cigarettes/day	22 (14)	20 (14)	0.007	21 (15)	0.169
Years Smoked	26 (14)	23 (14)	<0.001	25 (14)	0.330

The mean and standard deviation of each variable are provided, unless the variable is stated to be a percenatge (%) in which case the numbers in parentheses are the 95 percent confidence interval (95%CI) of the percentage. P values were determined using the t-test for continuous variables and the Pearson chi-squared test for categorical variables. P values in italics are statistically significant at alpha=0.05.

Table 3 shows the estimates of association between cigarette smoking variables and Barrett’s esophagus, compared with both GERD controls and population-based controls. Subjects with Barrett’s esophagus were significantly more likely to have ever-smoked cigarettes than both the population controls (OR=1.67) and the GERD controls (OR=1.61), although the GERD study-specific estimates appeared to be less heterogeneous (I²=11%, 95%UI:0–81%) than estimates from population-based control models (I²=82%, 95%UI:54–93%). Increasing pack-years of cigarette smoking was associated with an increasing OR for Barrett’s esophagus compared with both control groups (Table 3, Figure 1), albeit the risk relationship was not strictly linear in the categories used for assessment; the ORs for Barrett’s esophagus were approximately 1.5 for both <15 and 15–29 pack-years of smoking exposure groups, and approximately 2 for each of the higher exposure groups (30–44, and ≥45 pack-years of smoking), compared with each of the control groups and using never-smokers as the referent. The spline models, shown in Figure 2, are somewhat more indicative of a linear relationship—at least until approximately 20 pack-years of smoking—and this did not change when never-smokers were excluded. Conversely, the p value for trend for pack-years of smoking was statistically significant only when never-smokers were included for analysis (Table 3). Lastly, the additional cigarette smoking variables of duration, intensity, age of initiation, and duration of cessation were not associated with Barrett’s esophagus, after adjustment for total exposure (Table 3).

Table 3.

Fully adjusted odds ratios and 95% confidence intervals for the association between cigarette smoking and risk of Barrett’s esophagus.

	Barrett’s Esophagus
	n	Population Controls					n	GERD Controls
		n	OR	95% CI	I2 (95%UI)	N		n	OR	95% CI	I2 (95%UI)	N
Ever cigarette smoking
No	325	580		referent			382	643		referent
Yes	548	541	1.67	1.04 – 2.67	82 (54–93)	4	638	635	1.61	1.33 – 1.96	11 (0–81)	5
Pack-years of smoking
0 (never smokers)	325	580		referent			382	643		referent
<15	185	205	1.59	1.02 – 2.47	64 (0–88)	4	208	269	1.49	1.02 – 2.18	57 (0–84)	5
15 to <30	129	134	1.44	0.78 – 2.69	76 (33–91)	4	137	136	1.48	1.04 – 2.12	27 (0–71)	5
30 to <45	102	85	1.99	1.21 – 3.29	49 (0–83)	4	124	91	2.24	1.14 – 4.40	75 (38–90)	5
≥45	132	118	1.92	1.05 – 3.51	70 (12–89)	4	155	129	1.69	1.11 – 2.59	49 (0–81)	5
p for trend	873	1,122		0.057		4	1,006	1,268		0.009		5
p for trend excluding never smokers	548	542		0.193		4	619	625		0.170		5
Smoking duration*
<30 years	328	338		referent			352	395		referent
≥30 years	220	204	0.95	0.68 – 1.35	0 (0–81)	4	267	230	0.98	0.69 – 1.40	0 (0–76)	5
Smoking intensity*
<1 pack per day	207	230		referent			225	277		referent
1 pack per day	163	151	1.16	0.79 – 1.68	3 (0–85)	4	188	170	1.02	0.71 – 1.46	0 (0–78)	5
>1 pack per day	178	161	1.02	0.63 – 1.65	0 (0–53)	4	206	178	0.97	0.44 – 2.13	59 (0–85)	5
Age of smoking initiation*
<17 years	259	247		referent			258	239		referent
≥17 years	289	294	0.96	0.70 – 1.32	21 (0–88)	4	287	269	0.96	0.65 – 1.41	44 (0–81)	4
Smoking cessation*
<20 years	193	164		referent			224	232		referent
≥20 years	205	229	0.91	0.64 – 1.31	0 (0–76)	4	227	206	1.24	0.86 – 1.78	0 (0–58)	5

Results were adjusted for age (categorical: <50, 50–59, 60–69, ≥ 70), sex, body mass index (categorical: <25, 25–29, 30–34, 35–39 ≥ 40), and education (except UNC).

^*Also adjusted for pack-years of smoking (categorical: <15, 15–29, 30–44, ≥45). Abbreviations: GERD, gastroesophageal reflux disease; OR, odds ratio; UI, uncertainty interval.

Figure 1.

Forest plots of the relationship between increasing categories of cigarette smoking and Barrett’s esophagus compared with (A) population- based controls; and (B) gastroesophageal reflux (GERD) controls. Each study’s estimate is represented by the corresponding black square with the arms representing 95 percent confidence intervals. The grey-box overlaying each estimate represents the weight which it contributes to the pooled estimate. The pooled estimates are designated by the diamonds which follow each subgroup; the widths of the diamonds represent the 95 percent confidence intervals.

Figure 2.

Spline plots of the relationship between increasing categories of cigarette smoking and Barrett’s esophagus compared with (A) population-based controls; and (B) gastroesophageal reflux (GERD) controls. The solid line represents the estimate of the odds ratio while the broken lines either side represent 95 percent confidence intervals.

As shown in Figure 1, there were moderate-to-high levels of heterogeneity which were predominantly the product of the relatively lower estimates generated by the FINBAR study. When the FINBAR study was excluded, the summary ORs from the fully adjusted models slightly increased and the heterogeneity (I² values) decreased (Population-based controls: OR_ever-smoke=2.09 [95%CI:1.54–2.83, I2=44%]; OR_<15=1.93 [95%CI:1.36–2.74, I2=30%]; OR_15–29=1.75 [0.93–3.30, 68%]; OR_30–44=2.49 [1.70–3.65, 0%]; OR_≥45=2.57 [1.79–3.67, 0%]; GERD controls: OR_ever-smoke=1.75 [95%CI:1.43–2.15, I2=0%]; OR_<15=1.32 [0.95–1.84, 38%]; OR_15–29=1.62 [1.09–2.41,25%]; OR_30–44=2.87 [1.88–4.38, 19%]; OR_≥45=2.12 [1.50–3.00, 0%]).

The stratified models tested whether the effect of a single exposure in relation to Barrett’s esophagus was modified by another variable. When stratified by sex, the estimates for ever-smoking and categories of pack-years, in relation to Barrett’s esophagus, were slightly higher in men (OR_ever-smoke=1.81 [1.43–2.30, 0%]) than women (OR_ever-smoke=1.32 [0.91–1.92, 31%]), compared with GERD controls (Supplementary Table 1). Although ever-smoking stratified by sex was statistically significant (p=0.041), pack-years of cigarette smoking was not (p=0.5). Estimates of risk were not statistically different by sex when using population-based controls as the comparison group. Analyses stratified by BMI indicated that associations between cigarette smoking and Barrett’s esophagus may be stronger in those with a lower BMI (p=0.046), when using the population-based controls as the comparison group, while no pattern by BMI was discernable when compared with GERD controls (p=0.9; Supplementary Table 2). Analyses stratified by heartburn and regurgitation provided higher estimates for ever-smoking and pack-years of smoking in relation to Barrett’s esophagus in individuals without such symptoms (OR_ever-smoke=3.35 [1.55–7.26, 0%]) compared with individuals who reported symptoms (OR_ever-smoke=1.99 [1.50–2.65, 23%]) when using population-based controls as the referent, although these differences were not statistically significant (Supplementary Table 3).

Table 4 shows the results from the interaction models to test departures from additivity, which are considered as evidence for the existence of biologic interaction. Unlike effect-measure modification of ORs across strata of a second variable each with an independent referent group, interaction models simultaneously tested the effects of two exposures in relation to Barrett’s esophagus to assess whether there were synergistic effects. We found evidence for biologic interaction between ever-cigarette smoking and heartburn/regurgitation with an attributable proportion due to interaction amongst those exposed to both risk factors of 0.39 (95%CI: 0.25–0.52) (Table 4). Compared with the unexposed referent of population controls without heartburn/regurgitation who also never-smoked, the ORs for Barrett’s esophagus for each exposure category were 9.35 (95%CI: 6.08–14.39) for those exposed to heartburn/regurgitation only, 1.71 (1.04–2.80) for those exposed to smoking only, and 16.47 (10.73–25.29) for those exposed to both.

Table 4.

Interaction statistics for departure from additivity

Variables Tested for Interaction with Ever-smoking	Departure from Additivity
	ICR (95%CI) [null hypothesis=0]	AP (95%CI) [null hypothesis=0]	S (95%CI) [null hypothesis=1]
Barrett’s Esophagus vs. Population-based Controls
BMI	0.06 (−0.58, 0.69)	0.03 (−0.24, 0.29)	1.05 (0.64, 1.71)
Heartburn	2.75 (0.67, 4.82)	0.27 (0.10, 0.44)	1.43 (1.10, 1.86)
Regurgitation	4.58 (2.16, 7.00)	0.41 (0.27, 0.55)	1.82 (1.38, 2.40)
Heartburn or Regurgitation	6.41 (2.71, 10.10)	0.39 (0.25, 0.52)	1.71 (1.34, 2.18)
Alcohol	−0.24 (−0.98, 0.49)	−0.18 (−0.72, 0.35)	0.59 (0.17, 1.99)
Helicobacter pylori	0.16 (−0.79, 1.11)	0.06 (−0.32, 0.45)	1.12 (0.54, 2.35)
NSAID	0.31 (−0.16, 0.77)	0.20 (−0.09, 0.50)	2.55 (0.23, 28.62)
Barrett’s Esophagus vs. GERD Controls
BMI	−0.26 (−0.72, 0.20)	−0.20 (−0.55, 0.16)	0.54 (0.20, 1.51)
Alcohol	0.36 (−0.06, 0.78)	0.27 (−0.06, 0.60)	−
Helicobacter pylori	−0.07 (−0.78, 0.63)	−0.05 (−0.56, 0.46)	0.83 (0.18, 3.85)
NSAID	0.42 (−0.03, 0.86)	0.27 (0.00, 0.55)	5.22 (0.05, 561.36)

Abbreviations: AP, attributable proportion due to interaction; ICR, interaction contrast ratio; OR, odds ratio; S, synergy index. Italicized estimate and confidence interval represents statistical significance (p<0.05).

Discussion

The relationship between cigarette smoking and Barrett’s esophagus is unclear. Given the high prevalence of smoking, and its status as one of the few potentially-modifiable risk factors for Barrett’s esophagus, this relationship requires a more complete understanding. In this analysis of individual patient data from five studies within the international BEACON consortium, we found evidence for associations between ever-smoking and increasing pack-years with increased risk of Barrett’s esophagus. We did not find independent associations with related exposure variables, such as duration of smoking or the average number of cigarettes smoked per day, suggesting that the cumulative exposure to cigarette smoke is the most important exposure in this relationship. We also found tentative evidence that the relationship between cigarette smoking and Barrett’s esophagus may be stronger in men which could indicate sex differences in the role of smoking with respect to pathogenesis of Barrett’s esophagus. Lastly, evidence for biological interaction between heartburn/regurgitation and cigarette smoking suggests varied mechanistic effects of cigarette smoking in the development of Barrett’s esophagus.

Our understanding of the relationship between cigarette smoking and Barrett’s esophagus has been hampered by inconsistent data from studies too small to fully assess the issue; some studies have found evidence for an association, using population-based controls ^{23, 24}, endoscopy-negative controls ^{18, 25}, or GERD controls ^{18, 28–30}, while other studies have not found evidence for a relationship ^47–50. The analysis presented herein is much larger than any of these previous studies, and this larger sample size provided for greater statistical power and greater precision of risk estimates. In addition, the availability of GERD controls and population-based controls allowed for comparison to the source population undergoing endoscopy and the general population, respectively, with the latter also enabling assessment of heartburn/regurgitation as a potential effect-measure modifier and as a potential synergistic risk factor. A particular strength of the study is its use of pooled individual patient data through a large international consortium; this method provides more comparable statistical estimates than standard meta-analysis, which pool published odds ratios that differ in their variable definitions and the confounders included. Therefore, the results of this analysis are the strongest available data to date regarding cigarette smoking as a risk factor for Barrett’s esophagus.

Barrett’s esophagus is the recognized precursor lesion of esophageal adenocarcinoma, and, if cigarette smoking was a risk factor for Barrett’s esophagus, one might expect to observe an association between smoking and esophageal adenocarcinoma as well. Indeed, studies of this malignancy compared with population-based or hospital controls also provide evidence for an association with cigarette smoking ^50–55 including a recent pooled esophageal adenocarcinoma analysis from the international BEACON group ²². Given the concordance of these data, associations between cigarette smoking and Barrett’s esophagus, as well as cigarette smoking and esophageal adenocarcinoma, are likely to be real and, given the high prevalence of the exposure, may account for a large proportion (~40%) of esophageal adenocarcinomas ⁵⁶. It has not been known where smoking acts in the biological pathway. The current data suggest that smoking is associated with the risk of an early cancer precursor: Barrett’s esophagus.

Most of our primary exposure analyses had moderate to high levels of heterogeneity, an effect predominantly caused by the lower estimates of association from the FINBAR study ³³. Omission of this study reduced the heterogeneity and had minimal effects on the summary risk estimates attained, thus reinforcing the conclusions drawn. It is unknown why the associations between smoking and Barrett’s esophagus were lower in the Irish study population; the proportion of population-based controls that reported ever-smoking was higher (55%) than the other studies (45–47%), but this slightly higher rate is insufficient to mask the association evidenced in the other studies. In addition, the distribution of pack-years of cigarette smoking was similar across control groups and studies while provision of individual patient data enabled similar confounding structures to be constructed for study-specific models. FINBAR’s inclusion criteria did restrict recruitment of cases to those with long-segment Barrett’s esophagus (≥3 cm, Table 1); a criterion not employed by the other four studies included in this analysis. However, this is unlikely to have led to lower estimates of association given that a previous analysis of KPNC data evidenced a stronger association of cigarette smoking with long-segment Barrett’s esophagus (OR=1.72, 95%CI:1.12, 2.63) compared with that for short-segment Barrett’s esophagus (<3cm, OR=1.19, 95%CI:0.76, 1.85) ³¹. Thus, it remains unexplained why the FINBAR estimates of association were lower relative to the other studies included in this pooled analysis.

Analyses stratified by sex suggested that cigarette smoking may be a stronger risk factor for Barrett’s esophagus among men than among women. However, this relationship was only observed when assessing ever-cigarette smoking in Barrett’s esophagus cases compared with GERD controls; analyses of pack-years of cigarette smoking, and comparisons with population-based controls were null. Given the known genotoxic effects of tobacco smoke, evidence that effects of cigarette smoking are similar in men and women ⁵⁷, and the number of tests conducted, we believe this result represents a chance finding.

Interaction analyses indicated that heartburn/regurgitation symptoms and ever-smoking biologically interact in the risk of Barrett’s esophagus—the attributable proportion of disease amongst individuals exposed to these two factors was estimated to be 0.39 (95%CI:0.25–0.52). Biological interaction of these variables in this setting is plausible, given evidence that tobacco smoke may not only have direct genotoxic effects ⁵⁸ but may also induce transient lower esophageal sphincter relaxations (TLESRs) ^{59,60, 61}, increasing the likelihood, length and severity of gastroesophageal reflux—a major risk factor for Barrett’s esophagus ¹⁸ and the sequela, esophageal adenocarcinoma ¹⁷. Indeed, interaction between gastroesophageal reflux symptoms and smoking has been previously reported for Barrett’s esophagus with dysplasia ²⁶ and for esophageal adenocarcinoma ⁶².

There were several strengths of this analysis. First, the consortial approach enabled generation of the largest reported cohort of subjects with Barrett’s esophagus in the world’s literature upon which risk factor analysis has been performed. The large size of the pooled database enabled more precise estimates of association over previous studies, particularly in stratified analyses, spline models, and assessment of interaction. Second, although pooling and harmonization of data is a substantial undertaking and requires expertise, time, and resources, individual patient data allows for many benefits over meta-analysis of published estimates including building consistent models across studies, studying novel questions including interaction and using novel methods of analysis such as splines. Third, the availability of two control groups for comparison: population-based and GERD allows us to postulate where risk factors may be active in the pathogenesis of Barrett’s esophagus. This is important given that it is feasible that a significant proportion of the population-based control group may unknowingly have Barrett’s esophagus ⁶³, although such misclassification would bias results toward the null.

Limitations of this analysis include the moderate-to-high levels of heterogeneity for some analyses. Although constituents of tobacco smoke have changed over time ⁶⁴, the studies included in this analysis recruited incident cases and controls over a similar period (1997–2006). Regardless, constituents of tobacco smoke are likely to have differed geographically as is population susceptibility to genotoxic exposures. The unexplained heterogeneity does warrant a cautious interpretation of summary estimates, although associations were largely consistent in a majority of studies included and similar summary estimates with low heterogeneity were estimated when the study which was the source of the most heterogeneity was omitted from analysis. Another limitation is the possibility of recall bias, given the case-control design of the included studies, although the intensity and duration of smoking are usually recalled relatively reliably ⁶⁵. Lastly, we did not adjust for dietary variables in this analysis; although previous studies suggest that diet has minimal effects on relationships between smoking and Barrett’s esophagus, there remains the possibility of residual confounding through diet and other exposures.

In conclusion, cigarette smoking is a risk factor for Barrett’s esophagus, with adjusted odds ratios for multiple measures of association in the 1.5–2 range. The association appears to strengthen with increased exposure to cigarette smoking until approximately 20 pack-years, where it begins to plateau. If smoking is a causative agent of Barrett’s esophagus, it is an attractive modifiable risk factor, especially in high risk groups, such as elderly, obese males with GERD symptoms. Moreover, because the origins of BE are poorly understood, a better understanding of its risk factors and their biological interactions may allow inference as to the biological mechanisms involved in the nascent stages of Barrett’s esophagus. Indeed, the evidence we present for a biological interaction between smoking and heartburn/regurgitation suggest that cigarette smoking has multifaceted effects in the development of this precancerous metaplasia.