Intake of fruit and vegetables and risk of esophageal squamous cell carcinoma: A meta-analysis of observational studies

Authors

  • Jun Liu,

    1. Department of Radiation Oncology, Shanghai Chest Hospital, Shanghai Jiaotong University, Shanghai, People's Republic of China
    Search for more papers by this author
    • *J.L., J.W. and Y.L. contributed equally to this work.

  • Jian Wang,

    1. Department of Radiation Oncology, Zhongshan Hospital, Fudan University, Shanghai, People's Republic of China
    Search for more papers by this author
    • *J.L., J.W. and Y.L. contributed equally to this work.

  • Ye Leng,

    1. School of Life Science and Technology, Tongji University, Shanghai, People's Republic of China
    Search for more papers by this author
    • *J.L., J.W. and Y.L. contributed equally to this work.

  • Changxing Lv

    Corresponding author
    • Department of Radiation Oncology, Shanghai Chest Hospital, Shanghai Jiaotong University, Shanghai, People's Republic of China
    Search for more papers by this author

Correspondence to: Changxing Lv, Department of Radiation Oncology, Shanghai Chest Hospital, 241 Huaihai Xi Road, Shanghai Jiaotong University, Shanghai 200030, People's Republic of China, Tel./Fax: 086-021-22200000-3173, E-mail: lvchangxing001@126.com

Abstract

Quantification of the association between the intake of fruit and vegetables and risk of esophageal squamous cell carcinoma (ESCC) is controversial even though several studies have explored this association. We summarized the evidence from observational studies in categorical, linear and non-linear dose–response meta-analyses. Eligible studies published up to 31 July 2012 were retrieved via computer searches of MEDLINE and EMBASE as well as manual review of references. Random-effects models were used to calculate summary relative risks (SRRs) and the corresponding 95% confidence intervals (CIs). A total of 32 studies involving 10,037 cases of ESCC were included in this meta-analysis. The SRRs for the highest vs. lowest intake were 0.56 (95% CI: 0.45–0.69) for vegetable intake and 0.53 (95% CI: 0.44–0.64) for fruit intake (pheterogeneity<0.001 for both). Similar results were observed in a linear dose–response analysis. There was evidence of non-linear associations for intakes of fruit (pnon-linearity<0.001) and vegetables (pnon-linearity=0.041). There was no evidence of publication bias. These data support the hypothesis that intakes of vegetables and fruit may significantly reduce the risk of ESCC. Further investigation with prospective designs, validated questionnaires and good control of important confounders is warranted.

Abbreviations
BMI

body mass index

CI

confidence interval

EAC

esophageal adenocarcinoma

EC

esophageal cancer

ESCC

esophageal squamous cell carcinoma

FFQ

food frequency questionnaire

RR

Relative risk

SRR

summary relative risk

Esophageal cancer (EC) exists in two main histological types: esophageal squamous cell carcinoma (ESCC) and esophageal adenocarcinoma (EAC). These two entities are distinct with regard to etiological and pathological characteristics. Although a dramatic increment in the incidence rate of EAC has been reported in Western countries.[1, 2] ESCC remains the dominant histological type (ESCC ranks as the sixth most common cancer in the world and fourth in “developing” countries).[3-5] Importantly, ESCC occurs at a higher incidence in certain regions in the “Asian EC belt” that ranges from the Caucasian mountains, across northern Iran and through to northern China.[6] Epidemiological studies have revealed that tobacco smoking and alcohol use are the well-established risk factors for ESCC; >90% of cases can be attributed to these two factors in western countries.[7] In addition, mounting epidemiological evidence supports the important role of diet in the pathogenesis of ESCC.[8-10]

The intake of fruit and vegetables has long been associated with a decreased risk of various cancers, including EC. The suggested mechanisms for the major role of vegetables and fruit in the prevention of cancer include: modulation of DNA methylation; protection from and repair of DNA damage; promotion of apoptosis and induction of detoxifying phase-II enzymes.[11, 12] Based on a narrative review involving data from five cohort and 37 case–control studies, the Working Group from the World Cancer Research Fund/American Institute for Cancer Research (WCRF/AICR) 2007 report concluded that high consumption of vegetables and fruit “probably” protects against EC.[12] Unfortunately, in that report, no distinction was made between ESCC and EAC. In addition, the exact shape of the dose–risk relationship between the intake of fruit and vegetables and ESCC risk has not been clearly defined. Since that report was published, other relevant studies on this association have been published with mixed results.[13-22]

Therefore, to better characterize this issue, we conducted a comprehensive meta-analysis of the epidemiological literature using our own methods and criteria in the selection of studies, in the presentation of data and in our conclusions and interpretation of the evidence.

Material and Methods

Literature search

Two investigators (L.J. and W.J.) sought to identify the published English- language literature by using MEDLINE and EMBASE up to July 31, 2012 and by hand-searching the reference lists of the computer retrieved articles. For the databases search, MeSH terms or keywords were included: [esophag* AND (neoplasm OR carcinoma OR cancer)] combined with “nutrition OR diet OR lifestyle OR fruit OR vegetable.” A similar search was done using the word “oesophag*,” a common British spelling for esophagus. This meta-analysis was planned, conducted and reported according to the PRISMA statement.[23]

Definitions of exposure and outcome

In the present meta-analysis, we included the studies evaluating fruit or vegetable groups classified as “all” or “total.” Exposures presented as cooked vegetables, raw vegetables, green–yellow vegetables, green leaf vegetables, other vegetables, citrus fruit or other fruits were not considered as equivalent to “all” or “total” and thus were not included. Studies that reported “fresh vegetables” or “fresh fruit” were included according to the hypothesis that fresh vegetables or fruit accounts for a very high proportion of the total consumption.[24]

When outcomes were reported according to histological subtypes of EC in the included studies, we only extracted and pooled relative risks (RRs) for ESCC. For studies on total EC, we assumed that the majority of EC cases from non-Western countries were ESCC.[25] As the rise in the incidence of adenocarcinoma in Western countries mainly occurred in the most recent decades,[25] we excluded Western studies which initiated after 1990 and did not report on histological subtypes.

Study selection

Two researchers (L.J. and W.J.) independently reviewed all potentially relevant articles to determine whether an article met the general inclusion criteria, and disagreement was resolved by discussion between the researchers. To be included, studies had to: (i) use a case–control, nested case–control or cohort design; (ii) present data on incident cases of EC or mortality from EC and (iii) report associations in the form of RRs with the 95% confidence intervals (CIs) for total vegetables or total fruit at least adjusted or matched for age. Non-peer-reviewed articles, ecologic assessments, correlation studies, cross-sectional studies and mechanistic studies were not included. Studies were also excluded if they presented on several cancer sites combined, such as, upper aerodigestive tract cancers, cancers of oral cavity, pharynx and esophagus combined or cancers of esophagus and cardia combined. If data were duplicated in more than one study, the most recent or informative studies were included in this analysis.

Data extraction

Two researchers (L.J. and W.J.) independently extracted the following information: the name of the first author, study design, publication year, geographic locations, the number of cases and controls or participants, type of controls, the methods used for collection of data on exposure, exposure classification, duration of follow-up in cohort study, confounders adjusted for and the RR estimates with corresponding 95% CIs for the highest versus lowest level. From each study, we extracted the risk estimates adjusted for the greatest number of potential confounders.

Statistical methods

All statistical analyses were performed using STATA, version 11.0 (STATA, College Station, TX) and R-package statistical software (Version 2.11.0 beta). A two-tailed p value<0.05 was considered to be significant. We used the method of a random-effects model to calculate summary relative risks (SRRs) and 95% CIs for the highest versus lowest level, linear and non-linear dose–response analyses. This method of a random-effects model was developed by DerSimonian and Laird, which accounts for heterogeneity among studies.[26] For two studies[27, 28] which used a reference group other than the lowest category of fruit consumption, we estimated crude RRs (95% CIs) for different levels using exposure distributions.

We performed dose–response meta-analysis by using generalized least-squares trend estimation analysis.[29, 30] This requires that the number of cases and person-time or non-cases for at least three quantitative exposure categories is known. When this information was not available, we estimated the dose–response slopes by using variance-weighted least squares regression analysis.[29, 30] The medians for each category were assigned to each corresponding RR. If data were not reported, we assigned the median in each category by calculating the average of the lower and upper bound. If the highest category was open-ended, it was assumed that the open-ended interval length had the same length as the adjacent interval. When the lowest category was open-ended, the lowest boundary was considered as zero. The dose–response results in the forest plots are presented per 100 g/day increment in consumption of fruit and vegetables. In studies that reported intakes as frequency, we converted a serving size as 80 g according to study from Riboli et al.[24]

A potential non-linear dose–response relationship was checked by using fractional polynomial models.[31] The best-fitting second-order fractional polynomial regression model was defined as the one with the lowest deviance. A likelihood ratio test was used to assess the difference between the non-linear and linear models to test for non-linearity.[31]

Statistical heterogeneity among studies was assessed using both the Q statistic and the I2. If p value was <0.10, the heterogeneity was considered statistically significant. I2 takes values between 0% and 100%, and a value >50% is considered a measure of high heterogeneity.[32] Sources of heterogeneity were explored using subgroup analyses and random-effects meta-regression analysis, according to study design, geographic locations, the number of cases, type of Food Frequency Questionnaire (FFQ) and adjustment for confounders including smoking, alcohol use, total dietary energy intake and body mass index (BMI).

We also conducted sensitivity analysis to estimate the influence of each individual study on the summary results by repeating the random-effects meta-analyses after omitting one study at a time. Publication bias was assessed by using funnel plots and the further Begg's adjusted rank correlation and Egger's regression asymmetry test.[33, 34]

Results

Search results and study characteristics

The process of study selection is shown in Figure 1. We identified and screened 6,184 potentially relevant references. Of these, 95 were considered of potential value and the full text was retrieved for detailed evaluation. By studying reference lists, we identified 16 additional articles. Seventy-nine of these 111 articles were subsequently excluded from the meta-analysis for various reasons. Hence, 32 articles (five cohort and 27 case–control studies) involving 10,037 subjects with EC were used in this meta-analysis. The characteristics of these studies are presented in Tables 1 and 2. Seventeen studies were from Asia, four from North America, four from South America and seven from Europe.

Figure 1.

Flow diagram of systematic literature search on vegetables and fruit intake and the risk of ESCC.

Table 1. Characteristics of case–control studies of vegetables and fruit intake and esophageal squamous cell carcinoma risk
Author, yearCountryNumber of subjects enrolledDietary assessmentsExposure detailsRR (95% CI) (Highest vs. lowest)Adjusted or matched for
  1. Abbreviations: BMI: body mass index; EC: esophageal cancer; ESCC: esophageal squamous cell carcinoma; FFQ: food frequency questionnaire; NA: not available.

Hospital based case–control studies
Brown,1988 [47]USA207 EC, 422 ControlsFFQ-65 items, NAVegetable: Q3 vs. Q10.7(0.4–1.3)Age, sex, race, smoking, alcohol use
    Fruit: Q3 vs. Q10.5(0.3–0.9) 
De Stefani, 1990[48]Uruguay261 ESCC, 522 ControlsFFQ-10 itemsVegetable: <1, 1–3, 3–7, >7 times/wk0.56(0.3–1.0)Age, residence, smoking, alcohol use
   NAFruit: <1, 1–3, 3–7, >7 times/wk0.33(0.2–0.5) 
Negri, 1991[49]Italy119 EC, 6147 ControlsFFQ-37 itemsVegetable: Q3 vs. Q10.2(0.1–0.3)Age, sex, smoking, area, education,
   NAFruit: Q3 vs. Q10.3(0.2–0.4) 
Hu, 1994[50]China196 EC, 392 ControlsFFQ 32 itemsTotal fresh vegetable: Q4 vs. Q10.6(0.3–1.06)Age, sex, alcohol, smoking, family income
   NAFruit: Q4 vs. Q11.5(0.8–2.9) 
Hanock, 1994[51]Japan141 EC, 141 ControlsFFQFruit: <1, 1–2, 3–4, 5–7 /wk0.50 (0.18–1.39)Age, sex, domicile
   NA   
Rolon, 1995[52]Paraguay131 EC, 381 ControlsFFQ-50, validatedVegetable: Q4 vs. Q10.8(0.3–1.8)Age, sex, race, education, cigarettes, alcohol, fats, fish
Launoy, 1998[53]France208 ESCC, 399 ControlsFFQ validatedVegetable: 0–200, 200–300, 300–400, >400 g/d0.26 (0.13–0.55)Age, interviewer, smoking, aniseed aperitifs, hot Calvados, whisky, total alcohol,
    Fresh fruit: 0–60, 60–120, 120–180, >180 g/d0.59 (0.35–1.00)total energy intake and food groups
Takezaki, 2000[54]Japan284 ESCC, 11,936 ControlsFFQFruit: Occasionally or less, 3–4 times/week, Everyday0.7 (0.5-0.9)Age, year and season of visit, smoking and drinking
   NA   
Castelletto,2000[55]South America830 EC, 1,779 ControlsFFQ-50 items validatedVegetable: Never,1–3/week or month, daily0.62 (0.44–0.88)Age, hospital, residency, education, smoking, alcohol
    Fruit: Never,1–3/week or month, daily0.37 (0.27–0.51) 
Wolfgarten, 2001[27]German45 ESCC, 100 ControlsFFQFruit: <100,101–180, 181–265, >265 g/d0.1 (0.05–0.8)Age
   NA   
Phukan, 2001[56]India502 EC, 1,004 ControlsQuestionnaire; NAFruit: never, Occasionally0.3(0.08–4.2)Age, sex, education, income, chewing betel nut, tobacco, smoking and alcohol use
Onuk, 2002[28]Turkey44 ESCC, 100 ControlsQuestionnaire; NAVegetable: high vs. low0.10(0.04–0.22)Age, sex
    Fruit: high vs. low0.14(0.06–0.310 
Li, 2003[57]China1248 ESCC, 1248 ControlsFFQ-12 items, NAVegetable: <1,2–3, 4–7, >7 times/wk0.78(0.08–1.6)Age, sex, residence, income, occupation, smoking, alcohol
    Fruit: <1,2–3, 4–7, >7 times/wk0.08(0.01–0.2) 
Huang, 2004[58]Taiwan365 ESCC,522 ControlsFFQ validatedFresh vegetable: ≤1, >1 time/d0.5 (0.3–0.8)Age, education, ethnicity, source of hospital, smoking, alcohol drinking, areca nut chewing
    Fresh fruit: ≤1, >1 time/d0.6 (0.4–0.9) 
Yang, 2005[59]China185 EC, 185 ControlsQuestionnaire; NAVegetable >10 meals/week0.62 (0.32–1.17)Age, sex, alcohol, smoking, family history of esophageal cancer, occupation
    Fruit >1 time/week0.42 (0.19–0.89) 
Silvera, 2008[15]USA206 ESCC, 687 ControlsFFQ-104, validatedVegetable: per 1 serving/d0.88 (0.76–1.03)Age, sex, race, income, education, alcohol, smoking
    Fruit: per 1 serving/d0.86(0.73–1.01) 
Sapkota, 2008[60]Europe187 ESCC, 1288 ControlsFFQ, validatedVegetable: Q3 vs. Q10.69 (0.42–1.15)Age, country, sex, tobacco, education, BMI, alcohol consumption
    Fruit: Q3 vs. Q10.77 (0.43–1.38) 
Aune, 2009[17]Uruguay234 EC, 2032 ControlsFFQ-64 items,validatedVegetable: 76.2, 148.7, 261.9.Fruit: 52.5, 112.6, 241.1 g/d0.87 (0.55–1.37)0.59 (0.41–0.85)Age, sex, residence, income, interviewer,education, smoking status, alcohol, total meat, grains, fatty foods, mate drinking status, total energy intake and BMI.
Gao, 2011[19]China600 ESCC, 1514 ControlsQuestionnaire; NAFresh vegetable: <774, 774–999, 1000–1229, ≥1230 times/yr0.46 (0.35–0.62)Age, gender, geographic region
    Fresh fruit: <17, 17–64, 65–188, ≥189 times/yr0.53 (0.40–0.71) 
Hajizadeh, 2011[20]Iran47 ESCC, 96 ControlsFFQ-168 items ValidatedVegetable: 9.58–75.79 75.80–116.16 116.17–291.00 Servings/mo0.40(0.14–1.18)Age, sex, years of education, tobacco smoking, BMI, symptomatic gastroesophageal reflux and total energy intake
    Fruit: 2.00–38.57 38.58–78.00 78.01–199 Servings/mo0.66(0.23–1.87)0.13(0.04–0.45) 
Population based case–control studies
Gao, 1999[61]China81 EC, 234 ControlsFFQNAFruit: <1 times/mo, 1–3 times/mo, ≥1 times/wk0.75(0.36–1.55)Age, sex
Sharp, 2001[62]UK159 ESCC, 159 ControlsFFQFruit: 0–12, 12.01–18.04, 18.05–25.72, ≥25.73 times/wk0.64(0.25–1.67)Slimming diet, breakfast, salad, years smoking, regular use of aspirin, centre and temperature of tea/coffee
   NA   
Takezaki, 2001[63]China199 EC, 235 ControlsFFQ-152 items ValidatedVegetable:<1time/mo, 1–3 times/mo, 1–2 times/wk, >3 times/wk0.81(0.46–1.44)Age, sex, smoking, alcohol
    Fruit:<1time/mo, 1–3 times/mo, 1–2 times/wk, >3 times/wk0.91(0.48–1.73) 
Sun, 2003[64]China211 ESCC, 633 ControlsFFQVegetable: <700, 700–999, 800–1399, ≥1400 g/wk0.44 (0.21–0.95)Age, income, resident space and educational level
   NA   
Yokoyama, 2005[65]Japan234 ESCC, 634 ControlsFFQ validatedFruit: Not vs almost1.73(1.01–2.94)Age, sex, alcohol, smoking
Wu, 2006[66]China531 ESCC, 531 ControlsQuestionnaire, ValidatedVegetable: Q4 vs. Q11.37 (0.49–3.83)Age and gender, further adjusted for level of education, past economic status (group), smoking, alcohol drinking, BMI group, cancer family history, eating speed and food temperature;
     0.76 (0.15–3.72) 
    Fruit: Q4 vs. Q11.23 (0.51–2. 98) 
     1.17 (0.41–3.37) 
Sun, 2010[18]China250 EC, 750 ControlsQuestionnaire; NAVegetable: sometimes, always0.39(0.25–0.60)Age, sex, residence
    Fruit: sometimes, always0.65(0.45–0.94) 
Table 2. Characteristics of cohort studies of vegetables and fruits intake and esophageal squamous cell carcinoma risk
Author/yearCountryParticipants (N)Cases (n)Dietary assessmentsFollow-up, yExposure detailsRR (95% CI) (Highest vs. lowest)Adjustments
  1. Abbreviations: BMI: body mass index; EC: esophageal cancer; ESCC: esophageal squamous cell carcinoma; FFQ: food frequency questionnaire; NA: not available.

Tran, 2005[4]China29,584 Total NFFQ-9 items validated15Fresh vegetable:≤549, 549–732, 732–915, >915 times/yr1.02 (0.88–1.19)Age, sex
  1958 ESCC  Fresh fruit: ≤1,1–5,5–13,>13 times/yr0.80 (0.70–0.91) 
Freedman, 2007[13]USA490,802 Total NFFQ-124 items validated4.5Vegetable : 0.7, 1.15, 1.55, 2.08, 3.18 Serving/d/1000k0.57 (0.28–1.18)Age, sex, BMI, education, alcohol intake, cigarette, physical activity, total energy.
     Fruit: 0.40, 0.98, 1.46, 2.00, 3.25  
  103 ESCC  Serving/d/1000k0.46 (0.21–1.00) 
Fan, 2008[14]China18,244 Total N101 ECFFQ, NA15.5Vegetable: Q3 vs. Q10.71 (0.26–1.95)Age, year of interview, neighborhood of residence, education, BMI, smoking, drinks
     Fruit: Q3 vs. Q10.46 (0.25–0.88) 
Yamaji, 2008[16]Japan116 ESCCFFQ-138 items validated7.7Vegetable: 88, 165, 286 g/d0.68 (0.42–1.10)Age, area, cigarette smoking and alcohol drinking
     Fruit: 47, 137, 280 g/d0.65 (0.39–1.08) 
Steevens, 2011[21]The Netherlands120,852 Total N101 ESCCFFQ-150 items validated16.3Vegetable, 104, 146, 181, 222, 297 g/dFruit, 43, 107, 155, 215, 326 g/d0.61 (0.29–1.32)0.62 (0.32–1.22)Age, sex, cigarette smoking, alcohol, consumption of red meat, consumption of fish

Total vegetables

High versus low analyses

Twenty-four studies investigated the association between vegetable intake and ESCC risk. The summary RRs for the highest versus lowest analyses were 0.56 (95% CI: 0.45–0.69). There was high heterogeneity among studies (pheterogeneity<0.001, I2=75.8%; Fig. 2a).

Figure 2.

High versus low analysis of ESCC risk for the intake of vegetables (a) and fruit (b). Studies are sub-grouped according to study design. Circles represent study-specific RR; horizontal lines represent 95% CI; Diamonds represent SRRs.

We conducted subgroup and meta-regression analyses to explore the sources of heterogeneity (Table 3). Overall, there were inverse associations between high versus low intake of vegetables and ESCC risk in all strata, but the associations were not statistically significant in the population-based case–control (SRR: 0.57; 95% CI: 0.31–1.05) and cohort studies (SRR: 0.80, 95% CI: 0.61–1.07). When pooling estimates for only ESCC, the SRRs were statistically significant and were similar to those for total EC (PESCC vs. ESCC+EC=0.921). In addition, study design, adjustments for smoking and alcohol use and using a validated FFQ significantly altered the summary risk estimates for ESCC. Locations, the number of cases and confounders adjusted for BMI and energy intake did not significantly alter the summary risk estimates (Table 3).

Table 3. Subgroup analyses of vegetables and fruit intake and esophageal squamous cell carcinoma risk, high vs. low
Sub-groupsVegetablesFruit
Studies, nSRR (95% CI)I2 (%)pheterogeneitypdifferenceStudies, nSRR (95% CI)I2 (%)pheterogeneitypdifference
  1. Studies that provided risk estimates for esophageal squamous cell carcinoma only.

  2. Abbreviations: BMI: body mass index; EC: esophageal cancer; ESCC: esophageal squamous cell carcinoma; FFQ: food frequency questionnaire.

All (ESCC+EC)240.56(0.45–0.69)75.8<0.001 290.53(0.44–0.64)73.7<0.001 
Only ESCC[1]150.57(0.43–0.75)78.5<0.0010.921170.51(0.40–0.64)66.02<0.0010.801
Design
Cohort50.80(0.61–1.07)36.20.185 50.68(0.55–0.86)25.10.254 
Case–control190.52(0.41–0.65)64.60.0020.03240.51(0.41–0.63)71.5<0.0010.839
Sources of control
Population-based40.57(0.31–1.05)68.80.022 60.73(0.58–0.92)00.585 
Hospital-based150.51(0.40–0.65)66.0<0.0010.936180.44(0.34–0.57)74.9<0.0010.004
Geographic locations
Europe50.30(0.15–0.60)81.90.001 70.37(0.22–0.63)78.1<0.001 
USA30.65(0.43–0.98)00.910 30.40(0.18–0.92)66.90.049 
South America40.68(0.54–0.87)00.590 30.41(0.29–0.56)72.90.025 
Asia120.63(0.47–0.83)74.8<0.0010.936160.67(0.56–0.79)49.70.0130.086
Number of cases
<200120.52(0.37–0.73)66.50.001 150.47(0.33–0.67)67.1<0.001 
≥200120.59(0.44–0.78)79.2<0.0010.573140.57(0.47–0.70)70.7<0.0010.511
Type of FFQ
Validated120.70(0.55–0.88)57.80.006 120.61(0.49–0.77)67.1<0.001 
Not available120.45(0.34–0.61)66.5<0.0010.045170.47(0.36–0.61)73.4<0.0010.282
Adjustments
BMI, yes50.80(0.58–1.09)00.764 50.52(0.32–0.86)64.60.024 
no190.52(0.40–0.67)80.5<0.0010.166240.53(0.43–0.65)75.7<0.0010.991
Smoking, yes180.65(0.57–0.75)00.801 210.57(0.47–0.69)56.5<0.001 
no60.34(0.18–0.66)92.8<0.0010.02780.42(0.28–0.64)87.9<0.0010.335
Dietary energy intake, yes40.55(0.31–0.98)60.10.057 40.48(0.31–0.74)48.30.122 
no200.56(0.44–0.71)78.2<0.0010.989260.55(0.45–0.67)83.1<0.0010.636
Alcohol, yes180.65(0.57–0.75)00.801 200.57(0.47–0.70)69.8<0.001 
no60.34(0.18–0.66)92.8<0.0010.02790.44 (0.30–0.65)86.2<0.0010.374

Multivariate meta-regression analyses showed that study design and confounders adjusted by smoking and alcohol use were significant factors for the association. The between-study heterogeneity explained by the three variables was 39%. If the overall homogeneity and effect size were calculated by removing one study at a time, we confirmed the stability of the inverse association between vegetable intake and ESCC risk.

Dose–response analyses

Fifteen studies were included in the doseresponse analyses for total vegetable consumption (Supporting Information Figure S1a). The SRRs per 100 g/day increment in vegetable consumption were 0.84 (95% CI: 0.78–0.92) with evidence of severe heterogeneity (I2=82.0% and pheterogeneity<0.001). Study design did not alter this association. There was evidence of a non-linear association between vegetable intake and ESCC risk (p=0.041 for the best-fitting second-order fractional polynomial regression model, Fig. 3a), with a significant reduction in ESCC risk when increasing the intake up to about 160 g/d. Higher intakes were associated with a further decrease in risk.

Figure 3.

Relative risk and the corresponding 95% CI describing the non-linear association of the intake of vegetables (a) and fruit (b) and the risk of ESCC. The p values for the best-fitting second-order fractional polynomial regression model were 0.041 for vegetables and <0.001 for fruit intake, respectively.

Total fruit

High versus low analyses

A meta-analysis of 29 studies demonstrated an inverse association between total fruit intake and ESCC risk (SRR: 0.53, 95% CI: 0.44–0.64) (Fig. 2b). There was evidence of severe heterogeneity (I2=73.7% and pheterogeneity<0.001).

There were statistically inverse associations between high versus low fruit intake and ESCC risk in all strata. Summary estimates for only ESCC were similar with those for total EC (PESCC vs. ESCC+EC=0.801). Sources of controls and geographic locations significantly changed the summary estimates, whereas study design, type of FFQ, the number of cases and adjustments for confounders (BMI, energy intake, alcohol use and smoking) did not significantly modify the summary risk estimates for ESCC with fruit intake (Table 3).

Multivariate meta-regression analyses found that none of the variables described above were significant factors for the association between fruit intake and ESCC risk. When the overall homogeneity and effect size were calculated by removing one study at a time, we confirmed the stability of this inverse association.

Dose–response analyses

Nineteen studies were included in the dose–response analysis for fruit intake (Supporting Information Fig. S1b). The SRRs per 100 g/day increment in fruit intake were 0.61 (95% CI: 0.52–0.72), with evidence of severe heterogeneity (I2=89.7% and pheterogeneity<0.001). There was an evident non-linear association between fruit intake and ESCC risk (p<0.001 for the best-fitting second-order fractional polynomial regression model, Fig. 3b), with the most pronounced non-linear association observed for <20 g/day intake of fruits. Higher intake was associated with a further, but more modest decrease in risk.

Publication bias

There was no indication of publication bias for studies on the association between the intake of vegetables and fruit and ESCC risk. For studies focusing on high versus low vegetable intake, the p value was 0.980 for Begg's test and 0.832 for Egger's test (Fig. 4a). For studies on high versus low fruit intake, the p value was 0.268 for Begg's test and 0.128 for Egger's test (Fig. 4b).

Figure 4.

Begg's funnel plots of the log RRs versus the SEs of the log RRs in studies that evaluated the effect of vegetables (a) and fruit (b) intakes on the risk of ESCC.

Discussion

The results of this meta-analysis suggest that the intake of fruit and vegetables is associated with significant reductions in the risk of ESCC. We found, for the first time in a meta-analysis, evidence of a non-linear inverse association between the consumption of fruit and vegetables and ESCC risk.

We found a non-significant inverse association between vegetable intake and ESCC among studies with a population based on case–control and cohort designs. We assumed that this was due to low statistical power because there were only four population-based case–control studies and five cohort studies for vegetable intake, respectively. In addition, we found a significant inverse association among studies with a hospital-based case–control design. This result should be treated with caution because the hospital-based case–control design is subject to selection bias from the selection of cases and controls.

Based on non-linear dose–response analyses, we found that there might be a “threshold” level of ≈160 g/d for vegetable intake to identify statistically significant associations. However, we should be cautious about stating that there is a threshold because this may be related only to insufficient statistical power in those intake ranges used to identify significant associations. For fruit intake, the most pronounced reduction in ESCC risk was shown in an intake of <20 g/day. We should not over-emphasize this non-linear inverse association in those intake ranges because exposure assessment for fruit intake at very low levels (≈20 g/day) may be more prone to measurement error.

Several potential mechanisms might explain an inverse association between the intake of fruit and vegetables and ESCC risk. First, fruit and vegetables are good sources of various antioxidants, vitamins, minerals and other bioactive compounds (including vitamin C, lycopene, beta-cryptoxanthin, senium, beta-sitosterol and flavonols), which might induce the activity of detoxifying enzymes, reduce oxidative stress and inflammation and thereby play a major part in ESCC prevention.[35, 36] Vitamin C is an important nutrient for blocking intragastric nitrosation, a process that determines the development of ESCC.[37] In addition, fruit and vegetables are good sources of folate, which has been associated with a reduction in the risk of EC in several studies,[17, 38] but not in others.[39] Folate deficiency can induce poor incorporation of uracil into DNA. This can lead to the disruption of the integrity and repair of DNA, and induce altered expression of critical tumor suppressor genes and proto-oncogenes via the important role of folate in DNA methylation.[40, 41]

The meta-analysis described here has several strengths. This is the first comprehensive meta-analysis of a large number of studies on the association between the intake of vegetables and fruit and ESCC risk, including several prospective studies, including only outcome as ESCC (or mainly ESCC), using linear and non-linear meta-analytic methods and conducting several subgroup analyses. Additional strengths are extensive search of the literature and examination of the retrieved materials by at least two coauthors.

As a meta-analysis of observational studies, our findings have several limitations. Measurement errors are important in the assessment of dietary intake, which can lead to overestimation of the range of intake and underestimation of the magnitude of the relationship between dietary intake and cancer risk.[42, 43] Only 13 of the 32 studies used a validated FFQ to evaluate the consumption of fruit and vegetables. Indeed, whether or not using a validated FFQ significantly changes the association for vegetables intake (although this was not seen for fruit intake) is not known. One validation study showed that the consumption of vegetables and fruit assessed using a FFQ showed Spearman's correlation coefficients of 0.6 for fruit consumption and only 0.4 for vegetable consumption when compared with food records or food diaries.[21]

Incomparability of results between studies may also occur because definitions and categories of vegetables and fruit as well as analytical comparisons vary across studies. Studies from different regions, ethnicities and periods probably address very different exposures, and we therefore decided to consider only the studies that evaluated all types of fruit or vegetables to assess exposure that was as broad as possible. In addition, measurement error can occur if results based on different intake units are reported (such as portions per week or per month, grams per day and servings per year; and tertiles, quartiles or quintiles of consumption) without demarcating the cutoff points of exposure.

Severe heterogeneity was detected across studies. This phenomenon may reflect differences in study design, study populations, geographic locations, exposure measurement (e.g., face-to-face interviews vs. self-administered questionnaires), intake category (particularly in the open-ended high-intake category) and confounders adjusted for. Indeed, the meta-analysis models of cohort studies showed little variability, whereas significant heterogeneity was observed across case–control studies (pheterogeneity<0.001 for the intake of vegetables and fruit). A possible source of heterogeneity was explored using the methods of multivariable meta-regression analyses. We found that study design and confounders adjusted for smoking and alcohol use might partially (39%) account for the appreciable heterogeneity among studies with regard to vegetable intake. No variables were found to account for the severe heterogeneity across studies on fruit intake in the multivariable meta-regression analysis. In addition, analyses comparing high versus low intake are limited because true variability among the level and range of exposure among studies are not taken into account, and this may contribute to the severe heterogeneity observed.

Unmeasured or uncontrolled confounding inherited from observational studies is another concern. Individuals with higher intakes of fruit and vegetables are likely to have other healthy behaviors, e.g., a lower prevalence of tobacco smoking and overweight/obesity as well as lower intakes of alcohol.[44, 45] In nutritional studies, adjustment for total energy intake is important to account for the potential confounding due to dietary correlates.[46] Most of the studies included in this meta-analysis involved adjustment for smoking and alcohol use, but only a few studies employed adjustment for BMI and energy intake. Hence, we cannot fully exclude the possibility of residual confounding, although we found no significant differences between studies whether they were adjusted for BMI and energy intake or not.

As in any meta-analysis, the possibility of publication bias is of concern because small studies with negative results tend not to be published, although the results obtained from funnel plot analyses and statistical tests did not provide evidence for such bias.

We found significant reductions in the risk of ESCC associated with the consumption of fruit or vegetables. However, some caution is needed in interpreting the exact magnitude of the risk estimates due to the measurement errors of use of the dietary assessment methods, the high heterogeneity across studies and unmeasured confounding factors. Thus, given the high incidence of ESCC worldwide, particularly in non-western countries,[3] increasing intakes of fruit and vegetables in low consumers is of great importance for ESCC prevention from a public-health perspective.

In conclusion, our study suggests a significantly lower risk of ESCC associated with the intake of vegetables and fruit. However, the possibility that the association may be due to bias or confounding cannot be completely excluded (especially for vegetable intake). We, therefore, believe that these associations need further investigation in well-designed prospective studies with validated questionnaires and good control of important confounders.

Acknowledgements

Liu Jun, Wang Jian and Changxing Lv participated in the design of this article. Liu Jun, Wang Jian and Leng Ye participated in abstracting the data and performing statistical analysis. All authors read and approved the final article.

Ancillary