Perioperative chemo(radio)therapy versus primary surgery for resectable adenocarcinoma of the stomach, gastroesophageal junction, and lower esophagus

Abstract

Background

The outcome of patients with locally advanced gastroesophageal adenocarcinoma (adenocarcinoma of the esophagus, gastroesophageal (GE) junction, and stomach) is poor. There is conflicting evidence regarding the effects of perioperative chemotherapy on survival and other outcomes.

Objectives

To assess the effect of perioperative chemotherapy for gastroesophageal adenocarcinoma on survival and other clinically relevant outcomes in the overall population of participants in randomized controlled trials (RCTs) and in prespecified subgroups.

Search methods

We performed computerized searches in the Cochrane Central Register of Controlled Trials (CENTRAL), Database of Abstracts of Review of Effectiveness (DARE), the Cochrane Database of Systematic Reviews (CDSR) from The Cochrane Library, MEDLINE (1966 to May 2011), EMBASE (1980 to May 2011), and LILACS (Literatura Latinoamericana y del Caribe en Ciencias de la Salud), combining the Cochrane highly sensitive search strategy with specific search terms. Moreover, we handsearched several online databases, conference proceedings, and reference lists of retrieved papers.

Selection criteria

We included RCTs which randomized patients with gastroesophageal adenocarcinoma, in the absence of distant metastases, to receive either chemotherapy with or without radiotherapy followed by surgery, or surgery alone.

Data collection and analysis

Two independent review authors identified eligible trials. We solicited individual patient data (IPD) from all selected trials. We performed meta‐analyses based on intention‐to‐treat populations using the two‐stage method to combine IPD with aggregate data from RCTs for which IPD were unavailable. We combined data from all trials providing IPD in a Cox proportional hazards model to assess the effect of several covariables on overall survival.

Main results

We identified 14 RCTs with 2422 eligible patients. For eight RCTs with 1049 patients (43.3%), we were able to obtain IPD. Perioperative chemotherapy was associated with significantly longer overall survival (hazard ratio (HR) 0.81; 95% confidence interval (CI) 0.73 to 0.89). This corresponds to a relative survival increase of 19% or an absolute survival increase of 9% at five years. This survival advantage was consistent across most subgroups. There was a trend towards a more pronounced treatment effect for tumors of the GE junction compared to other sites, and for combined chemoradiotherapy as compared to chemotherapy in tumors of the esophagus and GE junction. Resection with negative margins was a strong predictor of survival. Multivariable analysis showed that tumor site, performance status, and age have an independent significant effect on survival. Moreover, there was a significant interaction of the effect of perioperative chemotherapy with age (larger treatment effect in younger patients). Perioperative chemotherapy also showed a significant effect on several secondary outcomes. It was associated with longer disease‐free survival, higher rates of R0 resection, and more favorable tumor stage upon resection, while there was no association with perioperative morbidity and mortality.

Authors' conclusions

Perioperative chemotherapy for resectable gastroesophageal adenocarcinoma increases survival compared to surgery alone. It should thus be offered to all eligible patients. There is a trend to a larger survival advantage for tumors of the GE junction as compared to other sites and for chemoradiotherapy as compared to chemotherapy in esophageal and GE junction tumors. Likewise, there is an interaction between age and treatment effect, with younger patients having a larger survival advantage, and no survival advantage for elderly patients.

Plain language summary

Chemotherapy before surgery in patients with adenocarcinoma of the esophagus, the gastroesophageal junction, and the stomach

This systematic review uses the data of individual patients from eight and published data from another six randomized controlled trials. We found that the administration of chemotherapy before surgery leads to longer survival in patients with adenocarcinoma of the esophagus, the junction between esophagus and stomach, and the stomach. The findings suggest that patients whose tumor is in the junction between esophagus and stomach and younger patients benefit most from the chemotherapy. Moreover, the addition of radiation to the chemotherapy seems to yield an additional advantage to patients, at least in tumors of the esophagus and the junction between esophagus and stomach. Chemotherapy before surgery does not increase the risk of suffering a complication during or after surgery.

Summary of findings

Summary of findings for the main comparison. Perioperative chemotherapy compared to primary surgery for resectable adenocarcinoma of the stomach, gastroesophageal junction, and lower esophagus.

Perioperative chemotherapy compared to primary surgery for resectable adenocarcinoma of the stomach, gastroesophageal junction, and lower esophagus
Patient or population: resectable adenocarcinoma of the stomach, gastroesophageal junction, and lower esophagus Settings: Intervention: perioperative chemotherapy Comparison: primary surgery
Outcomes	Illustrative comparative risks* (95% CI)		Relative effect (95% CI)	No of participants (studies)	Quality of the evidence (GRADE)	Comments
	Assumed risk	Corresponding risk
	Primary surgery	Perioperative chemotherapy
Overall survival among all patients	Moderate		HR 0.81 (0.73 to 0.89)	2422 (14 studies)	⊕⊕⊕⊕ high
Disease‐free survival (landmark time 6 months)	Moderate		HR 0.84 (0.69 to 1.01)	931 (7 studies)	⊕⊕⊕⊕ high
Overall survival by type of data ‐ Individual patient data	Moderate		HR 0.80 (0.66 to 0.97)	1049 (8 studies)	⊕⊕⊕⊕ high
Overall survival by type of data ‐ Aggregated data	Moderate		HR 0.81 (0.72 to 0.92)	1373 (6 studies)	⊕⊕⊕⊝ moderate1
Overall survival by tumor site ‐ Esophagus	Moderate		HR 0.87 (0.73 to 1.05)	473 (5 studies)	⊕⊕⊕⊕ high
Overall survival by tumor site ‐ GE junction	Moderate		HR 0.69 (0.54 to 0.87)	470 (6 studies)	⊕⊕⊕⊕ high
Overall survival by tumor site ‐ Stomach	Moderate		HR 0.94 (0.82 to 1.06)	828 (7 studies)	⊕⊕⊕⊕ high
*The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). CI: confidence interval; GE: gastroesophageal; HR: hazard ratio
GRADE Working Group grades of evidence High quality: Further research is very unlikely to change our confidence in the estimate of effect. Moderate quality: Further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate. Low quality: Further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate. Very low quality: We are very uncertain about the estimate.

¹Four out of the six studies included in this analysis have a high overall risk of bias.

Summary of findings 2. Perioperative chemotherapy compared to primary surgery for resectable adenocarcinoma of the stomach, gastroesophageal junction, and lower esophagus.

Perioperative chemotherapy compared to primary surgery for resectable adenocarcinoma of the stomach, gastroesophageal junction, and lower esophagus
Patient or population: resectable adenocarcinoma of the stomach, gastroesophageal junction, and lower esophagus Settings: Intervention: perioperative chemotherapy Comparison: primary surgery
Outcomes	Illustrative comparative risks* (95% CI)		Relative effect (95% CI)	No of participants (studies)	Quality of the evidence (GRADE)	Comments
	Assumed risk	Corresponding risk
	Primary surgery	Perioperative chemotherapy
Overall survival by chemo‐/radiotherapy ‐ Chemotherapy only	Moderate		HR 0.83 (0.75 to 0.91)	2033 (10 studies)	⊕⊕⊕⊕ high
Overall survival by chemo‐/radiotherapy ‐ Chemoradiotherapy	Moderate		HR 0.70 (0.50 to 0.99)	389 (4 studies)	⊕⊕⊕⊝ moderate1
Presence of tumor‐free resection margin	67 per 100	75 per 100 (67 to 81)	OR 1.42 (0.97 to 2.06)	1665 (10 studies)	⊕⊕⊕⊝ moderate1
Tumor stage at resection (T0 to T2 vs T3 to T4)	34 per 100	44 per 100 (34 to 54)	OR 1.53 (1.02 to 2.31)	1410 (8 studies)	⊕⊕⊕⊝ moderate1
Tumor stage at resection (N0 vs N+)	18 per 100	35 per 100 (25 to 47)	OR 2.43 (1.48 to 3.99)	1507 (8 studies)	⊕⊕⊝⊝ low1,2
Postoperative morbidity RD	30 per 100	31 per 100 (28 to 34)	See comment	1397 (9 studies)	⊕⊕⊕⊕ high	Risks were calculated from pooled risk differences: RD 0.01 (‐0.03 to 0.05)
Postoperative mortality RD	3 per 100	4 per 100 (2 to 5)	See comment	1606 (11 studies)	⊕⊕⊕⊝ moderate2	Risks were calculated from pooled risk differences:RD 0.00 (‐0.01 to 0.02)
*The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). CI: confidence interval; RR: risk ratio; OR: odds ratio; HR: hazard ratio; RD: risk difference
GRADE Working Group grades of evidence High quality: Further research is very unlikely to change our confidence in the estimate of effect. Moderate quality: Further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate. Low quality: Further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate. Very low quality: We are very uncertain about the estimate.

¹Heterogeneous trial results. 
 ²Funnel plot suggests publication bias.

Background

Description of the condition

The epidemiology of adenocarcinoma of the stomach, gastroesophageal (GE) junction, and esophagus ('gastroesophageal adenocarcinoma') has changed in recent years. Incidence and mortality figures for cancers of the distal stomach have decreased in most countries whereas corresponding figures for adenocarcinoma of the esophagus and gastroesophageal junction have risen (Ferlay 2010; Vial 2010). Combined, gastroesophageal adenocarcinoma ranks among the most common cancers worldwide with an estimated toll of approximately 1,000,000 deaths per year (DeMeester 2006; Ferlay 2010; Forman 2006; Gallo 2006).

Although differences in risk factors, gene expression, and tumor biology exist between adenocarcinoma of the stomach, gastroesophageal junction, and esophagus (Marsman 2005; Shah 2011), they are often regarded as one entity and conveniently treated alike in metastatic or non‐resectable stages. In fact, an analysis on a large number of patients showed that the degree of efficacy of chemotherapy does not differ for tumors of different origin in this setting (Chau 2009). Radical surgery is the only curative treatment modality for gastroesophageal adenocarcinoma. Depending on the exact tumor localization, distal or total gastrectomy, or esophagectomy with radical lymph node dissection, needs to be performed (Marsman 2005). Despite the fact that surgical techniques and perioperative management have substantially improved over the last decades (DeMeester 2006; Gallo 2006), five‐year survival after curatively intended resection is only 20% to 30% for patients with locally advanced disease (Hagen 2001; Siewert 1998).

Description of the intervention

Until recently, the standard treatment for gastroesophageal adenocarcinoma has been primary surgery; i.e. resection without prior tumor‐specific therapy. In light of poor survival rates (Hagen 2001; Siewert 1998), there has been a strong rationale to design new treatment modalities in order to achieve better outcomes for patients with non‐metastatic tumors, especially for those with locally advanced disease at diagnosis. Perioperative chemotherapy, defined as chemotherapy before and, optionally, after surgery, is one approach aiming to increase overall and disease‐free survival of patients. Based on promising results from phase II studies (Ajani 1995; Kelsen 1996; Ott 2003) it has been tested in several randomized controlled trials.

How the intervention might work

There are several proposed mechanisms for how perioperative chemotherapy might improve outcomes. A higher likelihood of tumor‐free resection margins due to preoperative down‐staging of the tumor, the elimination of micrometastases before and directly after surgery, and the rapid preoperative improvement of tumor‐related symptoms which leads to better tolerability of the upcoming large surgical intervention might all contribute to higher overall and disease‐free survival (Eguchi 2008; MAGIC 2006). One concern when administering perioperative chemotherapy is the potentially higher level of treatment‐related morbidity and mortality due to cytotoxic effects, which might be particularly hazardous during and directly after surgical procedures. Most phase II trials have shown that the applied chemotherapeutic regimens have acceptable morbidity and mortality (Ajani 1995; Kelsen 1996), but there have been reports of substantial adverse effects of certain regimens of cytotoxic drugs (Ajani 1993).

Why it is important to do this review

The evidence regarding the effect of perioperative chemotherapy on survival of patients with gastroesophageal adenocarcinoma is conflicting, with inconclusive results reported from trials (ACCORD 07 2011; CALGB 9781 2008; EORTC 40954 2010; FAMTX 2004; Feng 2008; Kobayashi 2000; MAGIC 2006; OE02 2009; RTOG 8911 2007; TROG‐AGITG 2005; Urba 2001; Walsh 2002; Wang 2000; Zhao 2006). Several recent systematic reviews aimed to summarize the available evidence.

A prior Cochrane Review, published in 2007, assessed perioperative chemotherapy for gastric adenocarcinoma by performing a meta‐analysis (Wu 2007). The authors did not find significant differences in survival between patients treated with perioperative chemotherapy and those directly operated. However, the review excluded studies in which patients had received additional postoperative chemotherapy and studies comprising patients with adenocarcinoma of the esophagus. Thus, the number of included trials was limited to only four. Furthermore, after the authors carried out their literature search in mid‐2005, the results of three large randomized controlled trials (RCTs) (ACCORD 07 2011; EORTC 40954 2010; MAGIC 2006), of which two showed significant survival benefits for perioperative chemotherapy, have been published. The review has meanwhile been withdrawn for methodological reasons.

A meta‐analysis published in 2004 included trials and previous meta‐analyses on patients with thoracic esophageal carcinoma of both histologies. It compared various neoadjuvant and adjuvant modalities, amongst which were perioperative chemotherapy and chemoradiotherapy, and found no survival advantage for perioperative chemotherapy alone and for perioperative chemoradiotherapy at one and two years postoperatively. A modest survival benefit for perioperative chemoradiotherapy was found at three years postoperatively. However, no subgroup analyses for the different histologies were conducted (Malthaner 2004). A Cochrane Review, which was updated in 2006, included 11 RCTs comparing perioperative chemotherapy with surgery alone for esophageal carcinoma of any histology. The authors concluded that the results suggested that perioperative chemotherapy may prolong survival, but judged the evidence to be inconclusive. Like the previous review, no subgroup analyses according to histological subtype were conducted (Vogt 2006). Another systematic review with meta‐analysis, relying on individual patient data (IPD) from trials including patients with carcinoma of the esophagus and gastroesophageal junction, was published as an abstract only (Thirion 2007). The authors reported a modest but significant benefit in terms of overall survival for perioperative chemotherapy. A subgroup analysis showed that this benefit was independent of histological subtype, and significant also in the subgroup of patients with adenocarcinoma, which comprised 46% of all included patients. The same group conducted an IPD meta‐analysis comparing perioperative chemoradiotherapy with surgery alone for esophageal carcinoma. Results were published as an abstract only (Thirion 2008). A survival benefit for chemoradiotherapy was found independently of histological subtype, but patients with adenocarcinoma constituted only 33% of the trial populations.

Lastly, a recently published update of a systematic review with meta‐analysis included RCTs comparing both perioperative chemotherapy and chemoradiotherapy with primary surgery in patients with esophageal cancer of both histologies (Sjoquist 2011). The analysis did not include patients with gastric adenocarcinoma. It showed a significant survival benefit for both perioperative modalities and in both histological subtypes with the exception of perioperative chemotherapy for squamous cell carcinoma, where the effect did not reach statistical significance.

None of the cited analyses included data for all three tumor sites of gastroesophageal adenocarcinoma (esophagus, GE junction, stomach). Moreover, none of the analyses, mostly due to non‐availability of IPD, were able to sufficiently assess in subgroup and multivariable analyses the extent to which certain covariables such as patient and tumor characteristics might alter the treatment effect of perioperative chemotherapy.

Therefore, we performed a new systematic comparison of perioperative chemotherapy with surgery alone for patients with locoregionally advanced resectable adenocarcinoma of the stomach, GE junction, and esophagus, including all three tumor sites and relying on IPD.

Objectives

The primary objective of this systematic review was to assess if perioperative chemotherapy leads to a longer overall survival as compared to surgery without prior tumor‐specific therapy in patients with locoregionally advanced resectable adenocarcinoma of the stomach, GE junction, and esophagus. Secondary objectives are to compare disease‐free survival, resectability, tumor stage upon resection, perioperative morbidity, and mortality, and to assess the safety and toxicity of perioperative chemotherapy as well as reasons for possible non‐administration of the postoperatively planned cycles of chemotherapy where foreseen in the study protocol.

Methods

Criteria for considering studies for this review

Types of studies

The review included only RCTs. Due to the specific intervention under study, blinding and placebo treatment are technically and/or ethically difficult because the unavoidable delay of surgery in a study arm where subjects receive a 'neoadjuvant' placebo treatment would indubitably lead to a significant worsening of their survival. Therefore, we did not consider blinding and placebo treatment as criteria for inclusion or exclusion.

Types of participants

To be included in the review, trials needed to be conducted on patients fulfilling the following criteria:

• histologically confirmed adenocarcinoma of the stomach, GE junction, or esophagus; for studies which include patients with both adenocarcinoma and other histological entities like squamous cell carcinoma, we sought to obtain IPD or aggregate measures (see below) relating to patients with adenocarcinoma only;

• previously untreated;

• locoregionally advanced (UICC stage Ib and higher for adenocarcinoma of the stomach, UICC stage II for adenocarcinoma of the esophagus (Sobin 2002));

• resectable based on staging exams;

• absence of distant or peritoneal metastases.

Types of interventions

The experimental intervention in the context of this systematic review was defined as surgery with curative intention combined with perioperative chemotherapy, defined as a treatment regimen with any kind of cytotoxic/antineoplastic drug or a combination of several of these drugs. To be regarded as perioperative, chemotherapy needed to be administered in a neoadjuvant (preoperative) setting and, optionally, in an additional adjuvant (postoperative) manner. We also included studies if patients received pre‐ or postoperative radiotherapy in addition to perioperative chemotherapy. We defined the control intervention as surgery with curative intention without any prior tumor‐specific therapy, and included patients undergoing any surgery with curative intention.

Types of outcome measures

Primary outcomes

The primary outcome was time to death measured in days from the date of randomization, based on an intention‐to‐treat analysis. If censored time‐to‐event data or adequate summary measures of time‐to‐event data were not available from the single trials, we used the vital status (alive or deceased) at the end of follow‐up.

Secondary outcomes

Secondary outcomes were:

• disease‐free survival time, defined as the time from a 'landmark' six months after randomization until recurrence (local or distant) or death of any cause (the landmark method accounts for the difference in the timing of surgery between the two treatment groups, for details see Data synthesis);

• presence of a tumor‐free resection margin, as assessed from the surgical specimen by a pathologist (dichotomous outcome yes/no);

• tumor stage at resection, as assessed from the surgical specimen according to the version of the UICC's TNM (T0‐4, N0‐3, M0‐1) classification (Sobin 2002) provided in the respective trial's IPD or aggregate data;

• safety of the perioperative chemotherapy regimen measured by toxicity according to the version of the National Cancer Institute Common Terminology Criteria for Adverse Events (CTCAE) provided in the respective trial's IPD or aggregate data;

• perioperative morbidity (measured by assessing how many of the following events occurred: anastomotic leakage, postoperative pneumonia, postoperative wound infection), and mortality (measured by assessing if a patient died during surgery or the consecutive hospital stay);

• the frequency and reason(s) for possible non‐administration of the postoperatively planned cycles of chemotherapy where foreseen in the study protocol (for the perioperative chemotherapy arms).

Search methods for identification of studies

Electronic searches

In September 2008, we performed a computerized literature search in:

• the Cochrane Central Register of Controlled Trials (CENTRAL), Database of Abstracts of Review of Effectiveness (DARE), the Cochrane Database of Systematic Reviews (CDSR) from The Cochrane Library (3rd Quarter 2008);

• MEDLINE (1950 to 2008 Sept week 2) (Appendix 1);

• EMBASE (1980 to week 38, 2008); and

• LILACS (Literatura Latinoamericana y del Caribe en Ciencias de la Salud) (up to September 2008).

We limited our search to studies in humans. There were no language restrictions for either searching or trial inclusion. We combined the Cochrane highly sensitive search strategy for identifying randomized trials in MEDLINE, sensitivity‐maximizing version, Ovid format (Higgins 2011) with specific search terms to identify randomized controlled trials in MEDLINE (see Appendix 1).

We adapted the MEDLINE search strategy for use in the other databases searched. Moreover, we searched the following online databases of ongoing trials:

• www.clinicaltrials.gov;

• www.centerwatch.com;

• www.cancer.gov/clinicaltrials;

• www.trialscentral.org;

• www.calgb.org;

• www.controlled‐trials.com;

• www.eortc.be;

• www.swog.org/Visitors/ClinicalTrials.asp;

• www.ctg.queensu.ca.

We extended our search using EMBASE, Ovid MEDLINE, and EBM Reviews ‐ Cochrane Central Register of Controlled Trials for studies published until 31 May 2011, which was the date when our database was closed for analysis (see Dealing with missing data).

Searching other resources

We handsearched the abstracts from 1995 to 2008 of the American Digestive Disease Week (DDW) published in Gastroenterology, the United European Gastroenterology Week (UEGW) published in Gut and the Annual Meetings of the American Society of Clinical Oncology (ASCO) published in the Journal of Clinical Oncology. We scanned reference lists of retrieved articles to identify further relevant trials. We contacted experts in the field about any unpublished or ongoing studies. We asked the authors of trial reports published only as abstracts or of ongoing studies to contribute IPD or completed papers.

Data collection and analysis

Selection of studies

Two independent review authors (UR, TS) extracted the data. They assessed title, keywords, and abstracts of all studies retrieved with the search strategy described above. If, based on this information, the authors believed studies met the defined inclusion criteria, they retrieved and further assessed the full text and made a final decision on whether to include a trial. Arbitration of a third author, which was foreseen in cases where one author believed a specific trial met the inclusion criteria for the review whereas the other author did not, was not required.

Data extraction and management

The review authors used a standardized data extraction form to compile and document relevant facts on general trial characteristics, trial quality, patients' characteristics, interventions, and outcomes as specified above. This data extraction was performed independently. The data extraction form compiled the following items.

• General information on the trial: title, authors, contact address, funding sources, language, publication status, year of publication, place(s), and year(s) of trial conduction.

• Trial design issues: inclusion and exclusion criteria, randomization/quasi‐randomization, concealment of treatment allocation (adequate/unclear/inadequate/not used), length of trial/follow‐up period.

• Baseline characteristics of participants: size of intervention and comparison group and for each group the distribution of age, sex, comorbidity (measured, if given as World Health Organization (WHO) performance status or American Society of Anesthesiologists (ASA) classification), tumor location (esophagus, gastroesophageal junction, stomach), and tumor stage (TNM and UICC stage).

• Characteristics of the intervention: used chemotherapeutic/antineoplastic drugs, regimens (frequency of chemotherapy, timing in relation to the date of surgery, application mode, cumulative dose of chemotherapy planned and administered both pre‐ and postoperatively).

• Frequency of different types of surgery (approach, extent) performed in the intervention and control groups.

• Loss to follow‐up in each group.

• Outcomes in each group: hazard ratios (HRs) and confidence intervals (CIs) both for overall and, if available, disease‐free survival; number of events (death, disease recurrence) if HRs are not given; number of resections with tumor‐free margins; tumor stage at resection (TNM and UICC stage); toxicity according to CTCAE (number of grade 3/4 adverse events); hospital mortality; morbidity as the number of the following events combined: anastomotic leakage, postoperative pneumonia, postoperative wound infection.

We pilot tested the data extraction form on five retrieved studies and slightly revised it. Two authors (UR, TS) performed data extraction independently. Consultation of a third author for arbitration, which was foreseen for cases where no consensus could be reached, was not required.

Individual patient data (IPD)

Data requests

For each trial, we requested IPD from the respective trialists. The solicited variables were as follows:

• age at randomization

• sex

• histological type

• allocated treatment arm (intervention/control)

• date of randomization

• comorbidity (WHO/ECOG performance status)

• site of tumor (esophagus/GE junction/stomach)

• pretreatment tumor stage (TNM T stage)

• pretreatment nodal stage (TNM N stage)

• chemotherapy regimen received (protocol, alternative, none)

• date of surgery

• surgical approach

• extent of resection

• reasons no surgery undertaken

• postoperative death

• non‐fatal postoperative complications (type)

• toxicity of preoperative treatment (type, grade) 

• tumor stage at resection (TNM yT stage)

• nodal status at resection (TNM yN stage)

• date of death or last visit

• vital status

• cause of death

• date of first recurrence

• date of progression

• nature of first failure (recurrence)

• lost to follow up (yes/no)

We requested data for all randomized patients in the trial (intention‐to‐treat population). We solicited trialists to provide the most complete and updated follow‐up data which were available, even if the follow‐up was longer than that used for the pertinent publication. We entered data in a dedicated database.

Quality control

We assessed the quality of the submitted IPD from the single trials in several ways.

• We compared the number of individual patient data sets with the intention‐to‐treat population reported in publications. We screened data sets for obvious duplicates or omissions (e.g. by checking patient IDs).

• We checked plausibility of the values supplied for each variable by looking for extreme outliers.

• We compared summary measures calculated from the data set with data reported in publications.

• We derived overall survival and disease‐free‐survival of the different treatment groups in each trial using the Kaplan‐Meier method and standard Cox regression analysis and compared with published survival estimates.

• We checked completeness and equality of follow‐up in the two trial arms by plotting a 'reverse' Kaplan‐Meier curve considering censored patients as patients who incurred the outcome (Stewart 1995). In addition, for the 'reverse' Kaplan‐Meier curves we evaluated the median follow‐up time.

We attempted to clarify and solve any detected inconsistencies with the respective trialists. 

Assessment of risk of bias in included studies

Both independent review authors assessed trial quality with regard to bias in the domains: selection bias, performance bias, detection bias, attrition bias, reporting bias, and other bias. The assessing author stated the level of bias ('low'/'high'/'unclear') that was assumed for each item. Based on this assessment, the author assigned an overall level of risk of existing bias to each trial ('low' if 'low' bias was assumed for all items, 'moderate' if 'high' or 'unclear' bias was assumed for one or two items, 'high' if 'high' or 'unclear' bias was assumed for at least three items). We used this bias level as a measurement of the quality of each trial in sensitivity and subgroup analyses (see below). In cases where the two review authors came to different conclusions regarding the risk of bias in the single domains, a third author acted as an arbiter and a consensus was reached. Statements, e.g. quotations from publications, supporting the judgment of the authors were to be given and presented in a 'Risk of bias' table for each trial.

Measures of treatment effect

We measured the effect of the intervention on overall and disease‐free survival with hazard ratios (HR). If possible the HR was based on IPD. If IPD were not available we calculated the HR (i) from the published reports, using methods described in Parmar et al (Parmar 1998) and Tierney et al (Tierney 2007), (ii) from binary mortality data. For the latter method a relative risk based on vital status (alive or deceased) at the end of follow‐up and sample sizes was calculated and imputed in the corresponding section in RevMan. Where it was feasible, we used various methods to indirectly estimate the trial HR, to check its reliability.

For each trial we estimated log HRs and the standard errors of log HR using the following methods (based on those reported by Parmar, Tierney, and Williamson; Parmar 1998; Tierney 2007; Williamson 2002), listed in order of preference:

1. HR and confidence interval calculated directly from IPD.

2. O‐E and variance of the log hazard ratio: log hazard ratio and its standard error estimated directly.

3. HR reported with confidence interval or log‐rank P value: standard error estimated from confidence interval or P value (confidence interval used if both available). This is the preferred indirect method since the HR is directly extracted and the standard error is estimated very accurately.

4. Adjusted HR reported with confidence interval or Cox proportional hazards P value: standard error estimated from confidence interval or P value (confidence interval used if both available). This will generally give an estimate close to the unadjusted HR, but different studies adjust for different factors, and the choice of adjustment factors could be data‐driven, leading to bias.

5. Numbers of events reported with log‐rank P value: HR estimated from numbers of events, standard error estimated from this estimated HR and P value. This gives a indirect estimate of the HR since all events are considered, but may not be close to the actual HR, particularly if the hazards are not proportional.

6. Kaplan‐Meier survival curve.

7. Actuarial rates at fixed follow‐up and log‐rank P value. This gives an estimate of the HR similar to that of method (3), but only events up to the fixed follow‐up time are considered.

For the effect of the intervention on disease‐free survival we calculated the HRs using these methods in their listed order of preference but for a landmark time of six months after date of randomization (for details see Data synthesis). This landmark analysis accounts for the difference in the timing of surgery between the two treatment groups.

We measured the intervention's effect on the presence of tumor‐free resection margins, a binary outcome, with an odds ratio (OR). We treated tumor stage upon resection as binary data by dichotomizing T stage (1/2 versus 3/4) and N stage (0 versus 1/2), and calculating ORs. We measured the intervention's toxicity by the total number of CTCAE grade 3/4 adverse events as well as of the single events. We compared postoperative mortality as well as postoperative morbidity, measured as the number of events specified above, by calculating risk differences.

Dealing with missing data

We performed analyses with the results of the intention‐to‐treat analysis if provided in the single studies. For missing data, we tried to contact the authors of the single studies and asked them for the specific values.

Our database was closed on 31 May 2011. Any data not available at that date, either because it was not provided by the trialists as IPD or because results of the respective trial had not been published as full manuscripts, were not included in our analyses.

Assessment of heterogeneity

We assessed heterogeneity clinically (by the judgment of the two independent review authors), as well as through the calculation of an I² statistic which is a measure for the percentage of the variability in effect estimates attributed to heterogeneity rather than sampling error. If heterogeneity between the effects found in single trials was shown to be too large, i.e. relevant clinical differences or an I² of above 0.5, we did not do a pooled analysis including all trials.

Assessment of reporting biases

To assess possible publication bias, if the number of included studies was sufficient, we created a funnel plot using the different outcomes and evaluated funnel asymmetry with Begg's and Egger's tests (Begg 1994; Egger 1997) with respect to continuous data or Peters' test (Peters 2006) with respect to binary data.

Data synthesis

For all outcomes, we combined IPD and aggregate data, according to availability from the single studies, using the two‐stage method (Riley 2007). This implies that from those studies for which IPD were available, we calculated the outcome measure, as defined above, from the provided data. For studies where IPD were not available, we used the aggregate outcome measure derived from the pertinent publication. If for a given outcome a summary measure was not available from either IPD or publications, the respective trial was not included in the analysis of that endpoint.

We performed data synthesis with results based on intention‐to‐treat analysis if available for the single studies. The estimated log HRs were combined using the generic inverse‐variance method, the result of which is presented as pooled HR with 95% confidence intervals on a logarithmic scale. The pooled HR represents the overall risk of an event for perioperative chemotherapy versus surgery alone. We calculated absolute effects on survival from the proportion of event‐free patients in the control group and the estimated HR (exp[ln(proportion event free)xHR]). We used a random‐effects model for all meta‐analyses. The usage of a random‐effects model was preferred to that of a fixed‐effect model because we assumed the existence of non‐explainable heterogeneity between the 'true' effects of the different treatment regimens implied in the studies. However, we re‐calculated all analyses with a fixed‐effect model in order to detect potential differences due to the methodological approach. In all tests of significance we calculated a two‐sided P value.

For the secondary outcome disease‐free survival, we only pooled the HRs of those trials which provided disease‐free survival calculated from a landmark time of six months from randomization, either obtained by IPD or by aggregated data. In this analysis, recurrence (local and distant) and death occurring within six months of randomization were regarded as events at this landmark time, thus defining the respective patient's disease‐free survival as zero. Likewise, analyzing IPD, patients discovered to be never disease‐free during the first six months after randomization (including patients with R2 resection, patients who were operated on without any resection performed, and patients who were not operated) were also regarded as events at this landmark time.

In the analyses of the secondary outcomes postoperative morbidity and mortality, the denominator included only patients who underwent surgery. In the analyses of the secondary outcome tumor stage at resection (pT stage and pN stage), the denominator was formed by the ITT population of the respective trials in order to treat patients who were not resected as 'treatment failures', as they were analyzed in the group of the more unfavorable pT and pN stage).

We performed analyses using Review Manager (RevMan 2011). We used the statistical software packages R and SAS (R 2010; SAS 2011) for additional analyses that could not be done with RevMan 2011.

Subgroup analysis and investigation of heterogeneity

For the primary outcome, we conducted subgroup analyses stratified for:

• tumor site (esophagus, GE junction, stomach); for stratifying patients according to tumor site, we used the definition from the single trial; either as variable in IPD databases or from a subgroup analysis in the respective publication;

• sequence of planned perioperative therapy in the intervention arm (preoperative only versus preoperative and postoperative combined);

• chemotherapeutic agents used in preoperative chemotherapy (platinum‐based non‐anthracycline regimens versus anthracycline‐based non‐platinum regimens versus regimens containing both platinum and anthracycline versus other regimens);

• regimens including radiotherapy versus chemotherapy‐only schemes

• performance status (PS 0 versus 1 versus 2 or higher);

• age upon randomization (< 65 years, 65 to 75 years, > 75 years);

• sex (male versus female).

Other subgroup analyses were defined based on exploratory analyses of the available data.

For trial‐specific subgroup analyses, such as those stratified according to sequence of perioperative therapy, chemotherapeutic agents and regimen, we calculated pooled HRs for every prespecified trial group. For patient‐specific subgroup analyses such as those stratified according to performance status, age, and sex, we undertook Cox regressions including the relevant treatment by subgroup interaction term within each trial and pooled the interaction coefficients across trials. For this PWT (pooling of within‐trial covariate interactions) approach we followed Fisher 2011 by assessing heterogeneity with I² statistics and reporting if the fixed‐effect and random‐effects results are consistent.

All patients were included in the analyses as originally randomized in the respective trial (i.e. according to the intention‐to‐treat principle), regardless of whether they were analyzed in the trial publication. In cases where specific data were missing, the respective patients were excluded from the analysis.

Investigation of important covariates

In a second step, we combined data from all trials providing IPD for the primary outcome 'overall survival' in a Cox proportional hazards model for clustered data. We used a shared frailty model for incorporating the trial, which represents a cluster, as random‐effects. By this, we examined how the association of treatment with survival was altered when covariates were accounted for. The covariates accounted for were treated as fixed‐effect, the trial as random‐effects. The following covariates were considered:

• treatment (chemotherapy only, chemoradiotherapy, surgery alone);

• tumor site (esophagus, GE junction, stomach);

• T stage (0/1/2 versus 3/4);

• N stage (0 versus 1/2/3);

• performance status (PS 0/1 versus 2 or higher);

• age upon randomization (as continuous variable);

• sex.

Sensitivity analysis

For all outcomes, we conducted sensitivity analyses based on the risk of bias assigned to studies as described in Assessment of risk of bias in included studies (low, moderate, high). Furthermore, we compared results from aggregate data for the primary endpoint and the secondary endpoint disease‐free survival with results from pooled IPD (Pignon 2001).

Results

Description of studies

Results of the search

After excluding duplicates, the prespecified electronic search of literature databases yielded 5848 results. An additional two trials, for which final efficacy results were already available, were identified by manual searches in conference proceedings and reference lists. Full papers of these trials (ACCORD 07 2011; EORTC 40954 2010) were published after our electronic search was conducted, but before our database was closed for analysis.

The extension of the search until the closure date of our database, 31May 2011 (see Electronic searches), yielded 1251 results after exclusion of duplicates. None of these were included in our analysis. One trial included only one patient with adenocarcinoma and was published only as conference abstract. One abstract was a duplicate of a fully published study included in our analysis (Walsh 2002). We assessed the full‐text paper of one trial. It showed that this trial formally met our inclusion criteria. However, since patients received chemotherapy for a very short period immediately prior to surgery (starting 72 and 68 hours and finishing 8 and 19.5 hours before surgery, respectively), we decided not to include this trial in our analysis as such an unconventional scheme is highly doubtful to influence our primary outcome, overall survival. Two other trials were not included because the full‐text publications were either unavailable or available only in Chinese for which a possibility of translation was not available at the time of the literature search, and because from the abstracts it remained unclear if the trials included any patients with adenocarcinoma. One trial was not included in our analysis as it was only available as a conference abstract and because it included only seven patients with adenocarcinoma (see Characteristics of excluded studies).

A summary of all searched, included, and excluded studies (PRISMA diagram) is given in Figure 1.

1.

Study flow diagram.

Included studies

From the search results, we identified 14 RCTs (total number of patients in ITT populations analyzed for overall survival: n = 3034) meeting our inclusion criteria (ACCORD 07 2011; CALGB 9781 2008; EORTC 40954 2010; FAMTX 2004; Feng 2008; Kobayashi 2000; MAGIC 2006; OE02 2009; RTOG 8911 2007; TROG‐AGITG 2005; Urba 2001; Walsh 2002; Wang 2000; Zhao 2006). Five of the trials (total n = 1657) included both patients with adenocarcinoma and squamous cell carcinoma of the esophagus (CALGB 9781 2008; OE02 2009; RTOG 8911 2007; TROG‐AGITG 2005; Urba 2001). These trials comprised n = 612 patients with squamous cell carcinoma and n = 1045 patients with adenocarcinoma. Thus, the total number of patients with adenocarcinoma in all included trials was n = 2422. One trial was a three‐armed RCT comparing two different preoperative chemotherapy regimens with one control group not receiving chemotherapy (Zhao 2006). For this trial (total n = 54) we combined the chemotherapy treatment arms into one (n = 34) and compared this combined treatment arm to the control group (n = 20). Consequently, our final analysis was based on data from n = 2422 patients. The 14 included RCTs were carried out between 1989 (first patient randomized) and 2004 (last patient randomized). The number of included patients per trial ranges from 42 (CALGB 9781 2008) to 533 (OE02 2009). Across all included trials, 2477 patients with adenocarcinoma had originally been randomized, out of which 25 (and an unknown number from OE02 2009) could not be included in our meta‐analysis either because they were not included in the trial analysis presented in the respective publication or because they were not contained in the provided IPD data sets (for details see Characteristics of included studies).

For all trials, data on the primary outcome, overall survival, were available either as HRs calculated from individual patient data (IPD) provided by the investigators (eight trials, n = 1049 patients), as HRs indirectly estimated from aggregate data presented in the original publications (three trials, n = 1207 patients), as HRs estimated from Kaplan‐Meier survival curves (one trial, n = 54 patients), or as HRs estimated from binary mortality data (two trials, n = 112 patients) (compare Methods). For four trials including patients with both adenocarcinoma and squamous cell carcinoma (CALGB 9781 2008; RTOG 8911 2007; TROG‐AGITG 2005; Urba 2001), HRs for the primary outcome were calculated for patients with adenocarcinoma based on IPD. For some secondary outcomes, data were not available from all included studies. Thus, respective analyses are based on a smaller number of studies (see Effects of interventions). 

Excluded studies

Mariette 2010 and van Hagen 2012 were published after closure of our database and were not included.

Risk of bias in included studies

Quality of single studies

Risk of bias was low in three studies, moderate in seven studies, and high in four studies (see Figure 2 and Figure 3).

2.

'Risk of bias' summary: review authors' judgments about each risk of bias item for each included study.

3.

'Risk of bias' graph: review authors' judgments about each risk of bias item presented as percentages across all included studies.

IPD quality 

• Number of patients in IPD data sets compared with the intention‐to‐treat population reported in publications.

The number of patients contained in IPD data sets were in all but one trial identical with the intention‐to‐treat population reported in the corresponding publications. Only in the data set belonging to Urba 2001 was there one additional patient in the surgery alone treatment group. The first author of Urba 2001 clarified this as a coding error in the data set the publication was based on.

• Extreme outliers?

We could not identify any extreme and implausible outliers in the IPD provided.

• Any deviations in summary measures calculated from IPD compared with data reported in publications.

We compared the number of patients in the different strata of tumor site, resection margin, T stage, N stage, performance status, median or mean age, age range, and sex between IPD and published data for the single studies:

Urba 2001: In the publication, data were only presented for the whole trial population (patients with both squamous cell carcinoma and adenocarcinoma); no separate data were available for patients with adenocarcinoma.

Walsh 2002: In the IPD 23 patients in the perioperative chemo and 44 patients in the surgery alone group had nodal metastasis; in the publication 23 patients in the perioperative chemo and 45 patients in the surgery alone group had nodal metastasis.

FAMTX 2004: In the IPD 16 patients in the perioperative chemo and 19 patients in the surgery alone group had R0 resection; in the publication 18 patients in the perioperative chemo and 19 patients in the surgery alone group had R0 resection.

TROG‐AGITG 2005: In the publication, data were only presented for the whole trial population (patients with both squamous cell carcinoma and adenocarcinoma); no separate data were available for patients with adenocarcinoma.

RTOG 8911 2007: In the publication, data were only presented for the whole trial population (patients with both squamous cell carcinoma and adenocarcinoma); no separate data were available for patients with adenocarcinoma.

CALGB 9781 2008: In the publication, data were only presented for the whole trial population (patients with both squamous cell carcinoma and adenocarcinoma); no separate data were available for patients with adenocarcinoma.

EORTC 40954 2010: No differences between IPD and published data.

ACCORD 07 2011: No difference between IPD and published data with respect to tumor site, resection margin, performance status, mean age, age range, and sex. The aggregated data of T stage and N stage upon resection could not be directly compared between IPD and aggregate data, as the figures given in the publication were based on a different denominator.

• Deviations in overall survival and disease‐free‐survival of the different treatment groups in each trial (derived from IPD using Kaplan‐Meier method and standard Cox regression analysis) as compared with published survival estimates.

In FAMTX 2004, an estimate of the hazard ratio and its standard error resulted in 1.52 (95% confidence interval 0.84 to 2.74) in favor of surgery alone using method 5 described in Measures of treatment effect. The hazard ratio estimate based on the corresponding database was 1.40 (95% confidence interval 0.78 to 2.53). This discrepancy in the hazard ratios resulted from a longer follow‐up of the patients in the IPD data set compared to the data set used for analysis in the pertinent publication.

The legend in TROG‐AGITG 2005, Figure 4 B and D, was not correct. The label of the patients at risk would need to be swapped (confirmed by the authors).

In ACCORD 07 2011 the estimated HRs and their 95% confidence intervals for the three tumor sites, esophagus, gastroesophageal junction, and stomach, based on the IPD differed from those given in Figure A1 in the corresponding publication. In a discussion with the authors it has been figured out that the discrepancies are due to the fact that the IPD has a longer follow‐up than the data the publication was based on.

• Completeness and equality of follow‐up in the two trial arms were checked by plotting a 'reverse' Kaplan‐Meier curve considering censored patients as patients who incurred the outcome (Stewart 1995). In addition, for the 'reverse' Kaplan‐Meier curves the median follow‐up time was evaluated.

For each trial with IPD the cumulative accrual rate was compared between the two treatment arms. No relevant differences occurred. By checking the completeness and equality of follow‐up in the two treatment arms by plotting the reverse Kaplan‐Meier curves (not shown here), again no relevant differences became obvious. Table 3 summarizes the median potential follow‐up time for each trial with IPD.

1. Median potential follow‐up time* in years for each trial with IPD.

Trial ID	Median potential follow‐up time in years (95% confidence interval)
Urba 2001	9.9 (6.3 to 10.3)
Walsh 2002	13.6 (12.2 to 16.1)
FAMTX 2004	9.7 (8.3 to 10.2)
TROG‐AGITG 2005	5.2 (4.4 to 6.1)
RTOG 8911 2007	8.9 (8.1 to 9.9)
CALGB 9781 2008	6.1 (5.5 to 6.3)
EORTC 40954 2010	4.4 (4.0 to 4.9)
ACCORD 07 2011	5.7 (5.1 to 6.6)

*Follow‐up time calculated according to Schemper 1996. 
 
 IPD: individual patient data

Effects of interventions

See: Table 1; Table 2

Primary outcome (overall survival)

Overall survival was reported by all 14 included studies. The meta‐analysis (Analysis 1.1) yielded a pooled hazard ratio (HR) of 0.81 (95% confidence interval (CI) 0.73 to 0.89, P < 0.0001) for patients who received preoperative chemotherapy or chemoradiotherapy as compared to those who underwent surgery alone. This corresponds to a relative increase in survival of 19%. The simple (non‐stratified) overall survival curves of perioperative chemotherapy plus surgery versus surgery alone showed an absolute improvement in survival of 9% at five years, increasing survival from 23% for patients undergoing primary surgery to 32% for patients receiving perioperative chemotherapy. The results of the single studies showed differences in the magnitude of the treatment effect, but the I² was 10% and the test for heterogeneity was not significant. HRs ranged from 0.45 for CALGB 9781 2008 to 1.40 for FAMTX 2004. The latter, together with Kobayashi 2000, were the only studies showing a trend towards longer overall survival in the surgery alone arm. The corresponding funnel plot (Figure 4) suggests possible reporting bias, as there are two small positive studies, but no small negative studies published. Nevertheless, Begg's and Egger's test revealed no evidence for funnel plot asymmetry (Begg's test P = 0.96, Egger's test P = 0.68, trim and fill method HR 0.82 with 95% CI 0.73 to 0.91).

1.1. Analysis.

Comparison 1 Overall survival, Outcome 1 Hazard ratio plot for overall survival.

4.

Funnel plot of comparison: 1 Overall survival, outcome: 1.1 Hazard ratio plot for overall survival.

The pooled HR of 0.81 is mirrored in the cumulative survival curve, which shows a sustained survival advantage for patients in the preoperative chemotherapy arms, starting at about 18 months after treatment onset, and lasting as long as 10 years (Figure 5).

5.

Simple (non‐stratified) overall survival curves of perioperative chemotherapy plus surgery versus surgery alone (perioperative chemo: 372 events, 525 total; surgery alone: 405 events, 524 total; hazard ratio 0.80, 95% CI 0.69 to 0.93).

Subgroup analyses

Tumor site

Definitions of tumor sites differed between trials, but they were usually stratified into esophagus, cardia/gastroesophageal (GE) junction, and stomach. One trial (TROG‐AGITG 2005) included only patients with esophageal tumors. Five trials (CALGB 9781 2008; OE02 2009; RTOG 8911 2007; Urba 2001; Walsh 2002) included patients with esophageal and GE junction tumors. One trial (Wang 2000) included only patients with gastroesophageal junction tumors. Four studies (FAMTX 2004; Feng 2008; Kobayashi 2000; Wang 2000; Zhao 2006) included only patients with gastric tumors. One trial (EORTC 40954 2010) included patients with gastric and GE junction tumors. Two studies (ACCORD 07 2011; MAGIC 2006) included patients with all tumor sites. For OE02 2009, site‐specific results were not presented in the publication, and IPD was not provided. For CALGB 9781 2008 and Urba 2001, site‐specific results were available neither from publications nor from provided IPD, and thus these studies could not be included in the subgroup analyses for tumor site, which could be performed for the three tumor sites: esophagus, GE junction, and stomach, based on data from five, six, and seven studies, respectively (Analysis 1.3). The results showed a more pronounced survival advantage for perioperative chemotherapy in patients with tumors of the GE junction (pooled HR 0.69, 95% CI 0.54 to 0.87) compared to the esophagus (pooled HR 0.87, 95% CI 0.73 to 1.05), and stomach (pooled HR 0.94, 95% CI 0.82 to 1.06), but the interaction test did not reach statistical significance (P = 0.08).

1.3. Analysis.

Comparison 1 Overall survival, Outcome 3 Hazard ratio plot for overall survival by tumor site.

In this subgroup analysis the effect of trial characteristics and patient characteristics are mixed as some trials include only certain tumor sites. Therefore, we investigated the interaction between treatment and tumor site in those trials where this was possible, i.e. IPD trials (Analysis 1.4). The heterogeneity coefficient of I² = 4% was not significant.

1.4. Analysis.

Comparison 1 Overall survival, Outcome 4 Interaction treatment‐tumor site (only IPD).

Sequence of planned perioperative therapy in the intervention arm

There were no relevant differences in the effect of preoperative chemotherapy on overall survival between the nine trials where only preoperative chemotherapy was stipulated and the five trials where both pre‐ and postoperative chemotherapy were foreseen (Analysis 1.6). However, it must be noted that even in the trials where postoperative chemotherapy was planned (ACCORD 07 2011; Feng 2008; Kobayashi 2000; MAGIC 2006; RTOG 8911 2007), only between 22% (ACCORD 07 2011) and 42% (MAGIC 2006) of patients received the full number of planned postoperative cycles (see below).

1.6. Analysis.

Comparison 1 Overall survival, Outcome 6 Hazard ratio plot for overall survival by timing of regimen.

Chemotherapeutic agents used in preoperative chemotherapy

Ten of the 14 included trials relied on a platinum‐based non‐anthracycline regimen. Therefore, this preplanned subgroup analysis was of limited value, as the other subgroups (anthracycline‐based non‐platinum regimens, regimens containing both platinum and anthracycline, and other regimens), consisted of one, one, and two trials. For platinum‐based non‐anthracycline regimens, the pooled HR was 0.80 (95% CI 0.72 to 0.89) and thus almost identical to the pooled HR of all included trials (Analysis 1.7).

1.7. Analysis.

Comparison 1 Overall survival, Outcome 7 Hazard ratio plot for overall survival by chemotherapeutic agents.

Regimens including radiotherapy versus chemotherapy only schemes

In 10 trials (ACCORD 07 2011; EORTC 40954 2010; FAMTX 2004; Feng 2008; Kobayashi 2000; MAGIC 2006; OE02 2009; RTOG 8911 2007; Wang 2000; Zhao 2006), the intervention arm stipulated a regimen which included chemotherapy only. In four trials (CALGB 9781 2008; TROG‐AGITG 2005; Urba 2001; Walsh 2002), the intervention was chemoradiotherapy (Analysis 1.5). All of the latter comprised exclusively patients with tumors of the esophagus or GE junction. Pooled subgroup analyses showed a significant benefit in terms of overall survival for both modalities. The magnitude of the effect was stronger for chemoradiotherapy, but the confidence interval of the pooled HR was wider due to the lower absolute number of participants in this subgroup. The interaction test was not significant. Of note, there was considerably higher heterogeneity among results of trials using chemoradiotherapy with the corresponding I² value just at the border of what we considered meaningful for a pooled analysis.

1.5. Analysis.

Comparison 1 Overall survival, Outcome 5 Hazard ratio plot for overall survival by chemo‐/radiotherapy.

Performance status

No publication provided sufficient information regarding overall survival stratified by performance status upon randomization. Two IPD data sets (Urba 2001; Walsh 2002) did not contain information on performance status. Five (ACCORD 07 2011; CALGB 9781 2008; EORTC 40954 2010; FAMTX 2004; TROG‐AGITG 2005) out of the six remaining IPD data sets measured performance status according to ECOG/WHO, whereas RTOG 8911 2007 used the Karnofsky score. This was converted to the ECOG/WHO classification according to Verger 1992.

In all trials with data on performance status, only five patients with a performance status of 2 or higher were included. Thus, this small group was not included in the subgroup analysis. The treatment by performance status interaction effect estimates from each trial were consistent (I² = 0%, P = 0.90, τ² = 0; pooled interaction effect under REM and FEM 0.34 with 95% CI ‐0.03 to 0.71; Analysis 1.9). The within‐trial HRs for overall survival in each subgroup are presented in Analysis 1.8 .

1.9. Analysis.

Comparison 1 Overall survival, Outcome 9 Interaction treatment‐performance status (only IPD).

1.8. Analysis.

Comparison 1 Overall survival, Outcome 8 Hazard ratio plot for overall survival by performance status (only IPD).

Age

No publication provided sufficient information regarding overall survival stratified by age upon randomization. This information was however available in all IPD data sets. Only 16 patients with an age of more than 75 years were randomized in any of the included trials, which precluded analysis in that subgroup. There was no evidence of subgroup differences with respect to the treatment by age interaction effect estimates. The latter were consistent over all trials (I² = 9%, P = 0.36, τ² = 0.02; pooled interaction effect under REM ‐0.07 with 95% CI ‐0.40 to 0.26; pooled interaction effect under FEM ‐0.06 with 95% CI ‐0.37 to 0.24; Analysis 1.11). The within‐trial HRs for overall survival are presented in Analysis 1.10.

1.11. Analysis.

Comparison 1 Overall survival, Outcome 11 Interaction treatment‐age (only IPD).

1.10. Analysis.

Comparison 1 Overall survival, Outcome 10 Hazard ratio plot for overall survival by age (only IPD).

Sex

No publication provided sufficient information regarding overall survival stratified by sex. This information was however available in all IPD data sets. About 83% of participants in the trials were male. Pooling of within‐trial covariate interactions revealed no subgroup difference between males and females (I² = 0%, P = 0.45, τ² = 0; pooled interaction effect under REM and FEM ‐0.18 with 95% CI ‐0.59 to 0.23; Analysis 1.13). The within‐trial HRs for overall survival are presented in Analysis 1.12.

1.13. Analysis.

Comparison 1 Overall survival, Outcome 13 Interaction treatment‐sex (only IPD).

1.12. Analysis.

Comparison 1 Overall survival, Outcome 12 Hazard ratio plot for overall survival by sex (only IPD).

Pretreatment tumor stage

Post hoc subgroup analyses according to pretreatment T stage (T0/1/2 versus T3/4) and pretreatment N stage (N0 versus N1/2/3) were done for all IPD data sets with available information (Analysis 1.14; Analysis 1.16). Three (CALGB 9781 2008; EORTC 40954 2010; RTOG 8911 2007) out of eight trials with IPD provided information on pretreatment T stage but only one (RTOG 8911 2007) had included enough patients with pretreatment stage T3/4 to investigate an interaction between treatment and T stage (Analysis 1.15). Four trials provided information on pretreatment N stage (CALGB 9781 2008; EORTC 40954 2010; RTOG 8911 2007; TROG‐AGITG 2005). Pooling of within‐trial covariate interactions revealed no subgroup difference between N0 and N1/2/3 (I² = 0%, P = 0.58, τ² = 0; pooled interaction effect under REM and FEM ‐0.02 with 95% CI ‐0.63 to 0.59; Analysis 1.17).

1.14. Analysis.

Comparison 1 Overall survival, Outcome 14 Hazard ratio plot for overall survival by pretreatment T class (only IPD).

1.16. Analysis.

Comparison 1 Overall survival, Outcome 16 Hazard ratio plot for overall survival by pretreatment N class (only IPD).

1.15. Analysis.

Comparison 1 Overall survival, Outcome 15 Interaction treatment‐T class (only IPD).

1.17. Analysis.

Comparison 1 Overall survival, Outcome 17 Interaction treatment‐N class (only IPD).

Resection status

In order to assess the influence of resection status (R0/1/2) on survival, we conducted a pooled survival analysis with patients from both arms of all five trials for which respective data were available (ACCORD 07 2011; EORTC 40954 2010; FAMTX 2004; RTOG 8911 2007; TROG‐AGITG 2005). A subgroup analysis according to resection status comparing overall survival for patients who received perioperative chemotherapy and those who underwent primary surgery was not appropriate, as the distribution of resection status was significantly different between treatment arms, probably because it was directly influenced by the treatment a patient received. For the same reason, the variable resection status was not included in the Cox model (see below). The Kaplan‐Meier curves show that resection status is a strong predictor for overall survival, with patients with R1 or R2 resection having a substantially poorer prognosis than patients with R0 resection (Figure 6). Five‐year survival probabilities were 37% (95% CI 33% to 42%) in the R0 group, 7% (95% CI 0% to 15%) in the R1 group, and 0% in the R2 group. The corresponding HRs were 0.31 (95% CI 0.22 to 0.44) for R0 versus R1 and 0.18 (95% CI 0.25 to 0.13) for R0 versus R2.

6.

Overall survival curves by type of resection: 719 patients with R0, R1 or R2, 99 patients not resected or missing (R0: 390 events, 611 total; R1: 43 events, 46 total; R2: 60 events, 62 total).

Sensitivity analyses

A sensitivity analysis which compared trials from which only aggregate data were available with those from which IPD were available revealed no differences, with both pooled HRs being almost identical (Analysis 1.2). A meta‐analysis excluding trials with high risk of bias yielded a HR of 0.82 (95% CI 0.72 to 0.93), which is virtually identical to the result of the meta‐analysis of all trials. A meta‐analysis including only the three trials with low risk of bias (424 patients) yielded a HR of 0.88 (95% CI 0.60 to 1.28). A meta‐analysis excluding the two trials from which only data on ‘vital status at end of follow‐up’ and no time‐to‐event data were available yielded a HR of 0.81 (95% CI 0.72 to 0.92).

1.2. Analysis.

Comparison 1 Overall survival, Outcome 2 Hazard ratio plot for overall survival by type of data.

Investigation of important covariates

A Cox proportional hazards model without any covariates but the treatment arm chemo(radio)therapy versus surgery alone as fixed‐effect and trial as random‐effects applied to all eight IPD data sets summarized in a single database (with 1049 patients) resulted in a HR of 0.80 (95% CI 0.69 to 0.93; see also Figure 5).

Two (CALGB 9781 2008; Urba 2001) out of eight trials with available IPD provided no information on exact tumor site (gastroesophageal junction or esophagus). For five (ACCORD 07 2011; FAMTX 2004; TROG‐AGITG 2005; Urba 2001; Walsh 2002) out of eight trials with available IPD no information on pretreatment T stage was available, for four (ACCORD 07 2011; FAMTX 2004; Urba 2001; Walsh 2002) trials no information on pretreatment N stage was available. For two trials (Urba 2001; Walsh 2002) information on performance status was missing. As the imputation of a covariate for a whole trial data set (not just a few patients in a trial) was not reasonable, no imputation methods for missing data were used and analyses were only conducted as complete case analyses.

In a first step, we analyzed the influence of each single prespecified covariate on the treatment effect of chemo(radio)therapy versus surgery alone by a Cox proportional hazards model with treatment and covariate as fixed‐effect and the trial as random‐effects based on all eligible patients (Table 4). The magnitude of the effect of perioperative chemo(radio)therapy on survival remained largely unchanged when adjusting for the separate covariates.

2. Adjusted hazard ratio (HR) for the treatment effect (death for perioperative chemotherapy versus surgery alone) in a Cox proportional hazards model.

Adjustment	Number of patients	HR (95% CI))	P value
For tumor site	931 (326 esophagus, 352 GE junction, 253 stomach)	0.82 (0.70 to 0.96)	0.0117
For cT stage	417 (152 T0/1/2, 265 T3/4)	0.83 (0.65 to 1.06)	0.1258
For cN stage	453 (250 N0, 203 N1/2/3)	0.90 (0.71 to 1.13)	0.3361
For performance status	846 (595 PS 0, 251 PS 1/2/higher)	0.86 (0.73 to 1.01)	0.0651
For age	1049 (median 62, minimum 23, maximum 80)	0.81 (0.70 to 0.93)	0.0031
For gender	1049 (870 male, 179 female)	0.80 (0.69 to 0.92)	0.0015

CI: confidence interval; GE: gastroesophageal; PS: performance status

In a second step, we did a complete case multivariable analysis with respect to the three treatment arms chemotherapy only, chemoradiotherapy, and surgery alone and all prespecified covariates except pretreatment T and N stage. For Cox proportional hazards regression analyses, we first tested the assumption of proportional hazards by plotting log(‐log(survival probability)) versus log of survival time stratified by treatment and by including time‐dependent covariables in the whole model. However, we detected no major violations of the proportional hazards assumption. Pretreatment T and N stage were not included as covariates because otherwise the complete case sample size would have dropped to 248. Only three patients had a performance status 2 or higher. Therefore, the Cox model deviated from the planned analysis by dichotomizing the performance status into 0 versus 1, 2, or higher. The results of the multivariable analysis based on 804 patients taking into account treatment arm, tumor site, performance class, age, and gender as covariates are summarized in Table 5.

3. Multivariable analysis: Cox proportional hazards model based on five trials* with 804 patients. Regression coefficients (β) with standard errors (SE) and hazard ratios (HR) for death with 95% confidence intervals (CI).

Variable	Number of patients	Univariable HR (95% CI)	Multivariable β (SE)	Multivariable HR (95% CI)	Multivariable P value
Treatment arm (chemotherapy only vs surgery alone)	324 chemotherapy only, 80 chemoradiotherapy, 400 surgery alone	0.85 (0.70 to 1.03)	‐0.16 (0.09)	0.86 (0.71 to 1.03)	0.0972
Treatment arm (chemoradiotherapy vs surgery alone)**		1.02 (0.72 to 1.45)	‐0.01 (0.18)	0.99 (0.69 to 1.40)	0.9351
Tumor site (esophagus vs stomach)	321 esophagus, 308 GE junction, 175 stomach	2.28 (1.60 to 3.26)	0.80 (0.18)	2.23 (1.55 to 3.18)	< 0.0001
Tumor site (GE junction vs stomach)		1.82 (1.35 to 2.46)	0.56 (0.15)	1.75 (1.29 to 2.37)	0.0003
Performance status (0 vs 1/2)	565 PS 0, 239 PS 1/2	0.74 (0.62 to 0.89)	‐0.30 (0.09)	0.74 (0.62 to 0.89)	0.0010
Age (continuous variable)	Median 61, Minimum 23, Maximum 81	1.01 (1.00 to 1.01)	0.01 (0.00)	1.01 (1.00 to 1.03)	0.0124
Gender (male vs female)	667 male, 137 female	1.26 (0.99 to 1.60)	0.11 (0.12)	1.12 (0.87 to 1.43)	0.3757

*ACCORD 07 2011; EORTC 40954 2010; FAMTX 2004; RTOG 8911 2007; TROG‐AGITG 2005

**The group 'chemoradiotherapy' includes 80 patients from one single trial (TROG‐AGITG 2005)

GE: gastroesophageal; PS: performance status

After adjusting for the various covariates, the effect of perioperative chemotherapy on survival decreased slightly in magnitude and was no longer statistically significant. There seemed to be no effect of perioperative chemoradiotherapy on survival, but we were able to include only a single trial, which was negative, in the model (TROG‐AGITG 2005) whereas three positive trials (CALGB 9781 2008; Urba 2001; Walsh 2002) could not be included due to unavailability of the variables required for the model. Tumor site, performance status, and age had a significant effect on survival even after multivariable adjustment.

We investigated interactions between treatment and one covariate at a time in the multivariable Cox regression model. An interaction between treatment and age was significant (multivariable P = 0.0168, not shown in table). This suggests that age may play a role in whether perioperative chemotherapy is more effective than primary surgery. Including an interaction term between treatment and age in the multivariable analysis revealed a trend of an increasing HR with increasing age, with the confidence interval including 1 at age 60, and with the point estimate of the HR being larger than 1 at age 70. While younger patients had a more pronounced survival benefit for perioperative chemotherapy, the treatment effect seemed to decrease or even disappear in older patients (Figure 7).

7.

Hazard ratio with 95% confidence interval for perioperative chemo(radio)therapy versus surgery alone for age in 5‐year increments.

Although an interaction between treatment and tumor site was not significant (multivariable P = 0.1346, not shown in table) we report the tumor site‐specific HRs for the treatment effect: esophagus HR 0.95 (95% CI 0.74 to 1.21), GE junction HR 0.73 (95% CI 0.55 to 0.95), stomach HR 1.15 (95% CI 0.76 to 1.73). The HR for patients with a resectable adenocarcinoma of the esophagus and of GE junction are less extreme in comparison to the forest plot (Analysis 1.3) whereas the HR for patients with a resectable adenocarcinoma of the stomach is reversed but does not reach statistical significance.

Secondary outcomes

Disease‐free survival

As we chose to calculate disease‐free survival according to the 'landmark' method (see Methods) and, with the exception of ACCORD 07 2011, no publication included the respective results, we could conduct this analysis exclusively for studies for which IPD were provided. However, the information in one IPD data set (Walsh 2002) was not sufficient to use the landmark method as data on disease‐free survival were not provided. Thus, our landmark analysis of disease‐free survival is based on seven studies comprising 931 patients (ACCORD 07 2011; CALGB 9781 2008; EORTC 40954 2010; FAMTX 2004; RTOG 8911 2007; TROG‐AGITG 2005; Urba 2001). The results (Analysis 2.1) resemble those of the pooled analysis of overall survival, with very small differences between HRs for overall and disease‐free survival. The pooled HR for disease‐free survival is 0.84 (95% CI 0.69 to 1.01) and thus slightly higher than for overall survival, and no statistical significance is reached (P = 0.06). Given the relatively small number of included trials, the corresponding funnel plot (Figure 8) is difficult to interpret, but does not suggest publication bias (Begg's test P = 0.88, Egger's test P = 0.92, trim and fill method added no studies).

2.1. Analysis.

Comparison 2 Disease‐free survival (landmark time 6 months), Outcome 1 Hazard ratio plot for disease‐free survival (landmark time 6 months).

8.

Funnel plot of comparison: 2 Disease‐free survival (landmark time 6 months), outcome: 2.1 Hazard ratio plot for disease‐free survival (landmark time 6 months).

As only studies with IPD were able to provide data for the meta‐analysis of disease‐free survival we could not perform a sensitivity analysis comparing trials with and without available IPD. None of the seven trials used in the meta‐analysis had a high risk of bias. When excluding the three trials with moderate risk of bias, the magnitude of the effect found in the meta‐analysis remained largely unchanged with a HR of 0.85 (95% CI 0.57 to 1.27).

Presence of tumor‐free resection margin

Data on the presence of a tumor‐free resection margin were available from 10 studies (Analysis 3.1). The meta‐analysis showed a higher likelihood of tumor‐free resection margin in patients who had received perioperative chemotherapy, but the association was just not statistically significant (odds ratio (OR) 1.29; 95% CI 1.04 to 1.60). There was heterogeneity between the single studies (I² = 55%). The corresponding funnel plot (Figure 9) did not suggest relevant publication bias (Peters' test P = 0.75, trim and fill method: OR 1.22; 95% CI 0.82 to 1.82).

3.1. Analysis.

Comparison 3 Presence of tumor‐free resection margin, Outcome 1 Odds ratio plot for tumor‐free resection margin.

9.

Funnel plot of comparison: 3 Presence of tumor‐free resection margin, outcome: 3.1 Odds ratio plot for tumor‐free resection margin.

The sensitivity analysis showed an OR of 1.53 (95% CI 1.05 to 2.23) when excluding two trials with high risk of bias, and of 1.69 (95% CI 0.98 to 2.93) when excluding another five trials with moderate risk of bias.

Tumor stage at resection

Data on T stage upon resection were available from seven studies (Analysis 4.1). The meta‐analysis showed that a T0/1/2 stage was significantly more frequent in patients who had received perioperative chemotherapy than in those who underwent surgery alone (OR 1.53; 95% CI 1.02 to 2.31). There was heterogeneity between results of the single studies, but all except two studies (Kobayashi 2000; RTOG 8911 2007) showed a higher frequency of T0/1/2 stages in the perioperative chemotherapy group. The corresponding funnel plot (Figure 10) did not suggest relevant publication bias (Peters' test P = 0.45, trim and fill method added no studies).

4.1. Analysis.

Comparison 4 Tumor stage at resection, Outcome 1 Odds ratio plot for tumor stage T0/T1/T2.

10.

Funnel plot of comparison: 4 Tumor stage at resection, outcome: 4.1 Odds ratio plot for tumor stage T0/T1/T2.

The sensitivity analysis showed an OR of 1.69 (95% CI 1.08 to 2.66) when excluding one trial with high risk of bias, and of 1.69 (95% CI 1.13 to 2.52) when excluding another three trials with moderate risk of bias.

Data on N stage upon resection were available from eight studies (Analysis 4.2). The meta‐analysis showed that a N0 stage was significantly more frequent in patients who had received perioperative chemotherapy (OR 2.00; 95% CI 1.58 to 2.55). There was considerable heterogeneity between results of the single studies (I² = 71%), with an OR as high as 23.4 in Urba 2001. All but one trial showed a higher frequency of an N0 stage in the intervention group. The corresponding funnel plot (Figure 11) suggested publication bias as the trim and fill method added three studies in the opposite direction of EORTC 40954 2010, Urba 2001 and Walsh 2002, and resulted in a considerably lower OR (Peters' test P = 0.04, trim and fill method: OR 1.62; 95% CI 0.95 to 2.74).

4.2. Analysis.

Comparison 4 Tumor stage at resection, Outcome 2 Odds ratio plot for nodal status N0.

11.

Funnel plot of comparison: 4 Tumor stage at resection, outcome: 4.2 Odds ratio plot for nodal status N0.

The sensitivity analysis showed an OR of 2.38 (95% CI 1.51 to 3.74) when excluding five trials with moderate risk of bias.

Postoperative morbidity

Data on postoperative morbidity were available from nine studies (Analysis 6.1). There was considerable variation in absolute reported morbidity between the single studies with rates in the single arms ranging from 0% (Wang 2000, both arms) to 63% (Walsh 2002, primary surgery arm). The meta‐analysis did not show any significant differences in postoperative morbidity between trial arms (risk difference 0.01; 95% CI ‐0.03 to 0.05). There was no relevant heterogeneity between results of the single studies and all confidence intervals included equity. The corresponding funnel plot (Figure 12) did not suggest relevant publication bias (Peters' test P = 0.93, trim and fill method added no studies).

6.1. Analysis.

Comparison 6 Postoperative morbidity, Outcome 1 Risk difference plot for postoperative morbidity.

12.

Funnel plot of comparison: 6 Postoperative morbidity, outcome: 6.1 Risk difference plot for postoperative morbidity.

The sensitivity analysis showed an absolute risk difference of 0.02 (95% CI ‐0.03 to 0.07) when excluding three trials with high risk of bias, and of 0.04 (95% CI ‐0.07 to 0.15) when excluding another three trials with moderate risk of bias.

Postoperative mortality

Data on postoperative mortality were available from 11 studies (Analysis 7.1). Overall, mortality rates ranged between 0% (Urba 2001; Wang 2000; Zhao 2006, both arms) and 7.4% (FAMTX 2004, perioperative chemotherapy arm) in the single trial arms. The meta‐analysis did not show any significant differences in postoperative mortality between trial arms (risk difference 0.00; 95% CI ‐0.01 to 0.02). There was no relevant heterogeneity between results of the single studies, and all confidence intervals included equity. The corresponding funnel plot (Figure 13) suggested a relevant publication bias, although the trim and fill method added only one trial with no relevant change in the overall risk difference (Peters' test P < 0.001, trim and fill method: RD 0.00 with 95% CI ‐0.01 to 0.02).

7.1. Analysis.

Comparison 7 Postoperative mortality, Outcome 1 Risk difference plot for postoperative mortality.

13.

Funnel plot of comparison: 7 Postoperative mortality, outcome: 7.1 Risk difference plot for postoperative mortality.

The sensitivity analysis showed an absolute risk difference of 0.00 (95% CI ‐0.01 to 0.02) when excluding three trials with high risk of bias, and of 0.02 (95% CI ‐0.02 to 0.05) when excluding another five trials with moderate risk of bias.

Safety

Data on the incidence of grade 3/4 toxicities during preoperative chemotherapy were available from six studies (Analysis 5.1). The incidence ranged between zero (Kobayashi 2000) and 1.96 (RTOG 8911 2007) events per patient.

5.1. Analysis.

Comparison 5 Safety of perioperative chemotherapy regimen, Outcome 1 Number of grade 3/4 toxicities per patient.

Number of grade 3/4 toxicities per patient
Study
ACCORD 07 2011	62/109 ˜ 0.57
CALGB 9781 2008	In publication: population with mixed histology, no stratified data available; in IPD: no data on toxicity available
EORTC 40954 2010	No data available
FAMTX 2004	41/27 ˜ 1.52
Feng 2008	No data on grade of toxicity available
Kobayashi 2000	"No complications were noted in the operative period related to preoperative dosage of 5'DFUR"
MAGIC 2006	330/237 who started treatment ˜1.39
OE02 2009	No data available
RTOG 8911 2007	225/115 ˜ 1.96
TROG‐AGITG 2005	In publication: population with mixed histology, no stratified data available; in IPD: no data on toxicity available
Urba 2001	In publication: population with mixed histology, no stratified data available; in IPD: no data on toxicity available
Walsh 2002	9/58 ˜ 0.16
Wang 2000	No data available
Zhao 2006	No data available

A formal sensitivity analysis according to risk of bias was not possible. One trial with a high risk of bias (Kobayashi 2000) reported that no toxicity occurred during the preoperative chemotherapy, whereas all trials with moderate or low risk of bias from which data were available reported some degree of toxicity. The two trials with low risk of bias (ACCORD 07 2011; FAMTX 2004) reported an incidence of 0.42 and 1.52 events per patient, respectively.

Non‐administration of postoperative therapy

Five trials (ACCORD 07 2011; Feng 2008; Kobayashi 2000; MAGIC 2006; RTOG 8911 2007) stipulated postoperative continuation of chemotherapy in their intervention arm. For only three of those trials (ACCORD 07 2011; MAGIC 2006; RTOG 8911 2007), sufficient information was available regarding frequency of non‐administration of planned postoperative therapy. In the ACCORD 07 2011 trial, 25/113 patients (22.1%) received all four postoperative chemotherapy cycles as planned. In the MAGIC 2006 trial, 104/250 (41.6%) patients received all three postoperative chemotherapy cycles as planned. In the RTOG 8911 2007 trial, 28/115 (24.3%) patients received both postoperative chemotherapy cycles as planned.

Fixed‐effect models

We repeated all analyses using fixed‐effect models. The only appreciable differences noted were that the 95% CI for disease‐free survival narrowed to 0.72 to 0.97 and reached statistical significance at P = 0.02, that the OR for tumor‐free margin changed to 1.29 (95% CI 1.04 to 1.60), and that the OR for postoperative N0 versus N1/2/3 stage changed to 2.00 (95% CI 1.58 to 2.55).

Discussion

The present meta‐analysis uses IPD to assess potential therapeutic benefits of perioperative chemotherapy for gastroesophageal adenocarcinoma. Unlike previous meta‐analyses (Sjoquist 2011; Thirion 2007; Wu 2007), it includes all tumor locations, i.e. esophagus, gastroesophageal (GE) junction, and stomach. Due to the availability of individual patient data (IPD) from several trials in esophageal cancer, it was possible to include patients with adenocarcinoma from trials with mixed histology, while excluding patients with squamous cell carcinoma, which is thought to have a different biological behavior (Chau 2009).

The principal and clinically most important finding of our meta‐analysis is that perioperative chemotherapy, with or without radiotherapy, for gastroesophageal adenocarcinoma was associated with prolonged overall survival. The cumulative survival curve, based on IPD from more than 1000 patients, showed a sustained benefit for patients who received perioperative chemotherapy, starting at about 18 months after surgery and lasting as long as 10 years. The magnitude of this benefit was roughly 9%. For example, five‐year survival was 23% in the surgery alone group and 32% in the perioperative chemotherapy group. Given the very poor overall long‐term prognosis of the disease, this difference must be considered as clinically highly relevant. A very similar association was found for disease‐free survival, which strengthens the validity of our results. Therefore, perioperative chemotherapy should be offered to all patients with locoregionally advanced gastroesophageal adenocarcinoma unless medical contraindications are present.

Several potential mechanisms of how perioperative chemo(radio)therapy could lead to prolonged survival in this tumor entity have been discussed (Eguchi 2008; MAGIC 2006). Some of these hypotheses are supported by our results. The odds of complete resection in patients treated with perioperative chemotherapy were about 1.4 times higher than in untreated patients, with borderline statistical significance in the meta‐analysis. A pooled analysis of patients from both treatment arms of trials for which respective data were available showed that complete resection is in fact a very strong predictor of survival. Regarding down‐staging of T and N stages and histopathologic response, there were insufficient direct data available from the single trials. We were however able to compare T and N stages upon resection between treatment arms and found a considerably higher incidence of more advanced stages in patients who were not pretreated with perioperative chemotherapy. As there is no reason to assume relevant baseline differences in T and N stage in the included randomized controlled trials (RCTs), this finding is an indirect proof of down‐staging. No data were available to us on the effect of perioperative chemotherapy on cancer symptoms, and on potential improvement of patients' performance status and quality of life, which in itself seems to be positively associated with survival and better surgical outcome (Chau 2004; Djarv 2010; McCulloch 2003), during preoperative treatment.

There were no tangible differences in perioperative morbidity and mortality between treated and untreated patients, which largely excludes relevant detrimental effects of the treatment on the incidence of perioperative complications. Thus, fears that pretreating patients with chemotherapy before resection of gastroesophageal malignancies jeopardizes their perioperative safety do not seem to be justified. Specifically, we could not find any evidence for excessive postoperative mortality after preoperative chemoradiotherapy from the three trials for which such data were available. Specific regimens or schemes might be more favorable with regard to perioperative outcomes than others, an assumption which could not be further assessed in our data due to the heterogeneity of applied treatment protocols in the single trials. Information on how well perioperative chemotherapy itself was tolerated was available only from few trials and showed relevant heterogeneity with reported incidences of grade 3/4 toxicities ranging from as low as one in 10 patients to as high as two per patient.

A strength of our IPD‐based analytical approach is that we were able to assess survival effects of perioperative chemotherapy in various clinically relevant prespecified subgroups both according to patient and treatment characteristics. As a note of caution, it must be kept in mind that these subgroup analyses were not planned for in most individual trials. Due to small numbers of patients in the single subgroups, some of the analyses lack the statistical power to detect associations of small magnitude. Likewise, direct comparisons of treatment effects between single subgroups need to be prudently appreciated, as there might be substantial confounding. Probably the most important finding was that the beneficial effect of perioperative chemotherapy on overall survival was largely consistent across most subgroups, which supports the broad application of this modality in patients with gastroesophageal adenocarcinoma. Of note, however, age showed an interaction with the effect of perioperative chemotherapy in the multivariable analysis, with the effect on overall survival being much more pronounced in patients of younger age and with no apparent effect in older patients. Although it cannot be shown with our data, toxicity might play an important role in this interaction. The finding supports the common clinical practice of critically evaluating the decision for perioperative chemotherapy above a certain age. Obviously, chronological age is a mere proxy for biological age and a person's overall health status and there should not be any strict age limits regarding eligibility for perioperative chemotherapy.

With respect to tumor site, the treatment effect seems to be largest for GE junction tumors, followed by esophageal and gastric tumors. Other than expecting potential bias inherent to the analytical approach and explained above, these results need to be interpreted with additional caution because distinction of the precise tumor site between distal esophagus, GE junction, and proximal stomach, usually done according to the Siewert classification (Siewert 1987), is sometimes difficult in the pre‐therapeutic setting especially when endoscopic conditions are sub‐optimal. In the different trials included in our meta‐analysis, inclusion criteria and exact classification of tumor site varied and were not always specified in detail in the respective publications, which might introduce additional heterogeneity

Regarding which specific perioperative chemotherapy regimen is most effective, our meta‐analysis can only provide limited information. The vast majority of included RCTs (ACCORD 07 2011; CALGB 9781 2008; EORTC 40954 2010; Feng 2008; OE02 2009; RTOG 8911 2007; TROG‐AGITG 2005; Urba 2001; Walsh 2002; Zhao 2006) used platinum‐based non‐anthracycline regimens, usually a combination of 5‐FU and cisplatin in varying intensity. Their results were very homogeneous. With the exception of one trial (TROG‐AGITG 2005), all of these RCTs yielded a hazard ratio (HR) smaller than one, which speaks in favor of the efficacy of this combination. One large trial (MAGIC 2006) added epirubicin to the 5‐FU/cisplatin doublet, which yielded a slightly lower HR. However, such an indirect comparison is insufficient to state that this triplet regimen is indeed more effective. The two studies preoperatively using only a fluorouracil‐derivate (Kobayashi 2000; Wang 2000) showed mixed results but were also small. The only trial which used an anthracycline‐based non‐platinum regimen, i.e. FAMTX, showed a trend towards poorer survival in the intervention arm and was thus stopped early (FAMTX 2004). On these grounds, the regimen has been discarded for further use in gastroesophageal adenocarcinoma. Recently, taxane‐based schemes have gained popularity also in gastroesophageal adenocarcinoma. One of the included RCTs (Feng 2008) used a combination of docetaxel, cisplatin, and 5‐FU and yielded a HR of 0.79. However, its statistical power was too small to reach statistical significance. An ongoing phase three trial (NCT01216644) compares the ECF scheme that was used in the MAGIC trial to the docetaxel‐based FLOT scheme with pathologic complete response as primary endpoint. To what extent the addition of postoperative chemotherapy to preoperative treatment is beneficial must remain open. Five of the 14 included trials (ACCORD 07 2011; Feng 2008; Kobayashi 2000; MAGIC 2006; RTOG 8911 2007) foresaw postoperative chemotherapy. The subgroup comparison with trials in which patients were treated only preoperatively showed no relevant differences. It must be noted that only a minority of patients in fact received the planned postoperative treatment to its full extent. Therefore, its role remains debatable. However, since two large positive trials (ACCORD 07 2011; MAGIC 2006) had used a treatment scheme including postoperative chemotherapy, it remains common clinical practice to offer postoperative continuation of chemotherapy to eligible patients.

In the subgroup analysis of perioperative schemes with or without inclusion of radiotherapy, the treatment effect seems to be larger for combined chemoradiotherapy. However, all included trials using chemoradiotherapy comprised exclusively patients with esophageal or GE junction tumors, which inevitably leads to substantial confounding. In our multivariable model, chemoradiotherapy is not associated with improved overall survival compared to surgery alone. This result is however based on a single negative trial (TROG‐AGITG 2005) and disregards three positive trials, from which the variables required for the model were unavailable (CALGB 9781 2008; Urba 2001; Walsh 2002), which makes its interpretation difficult. After closure of our database, two other RCTs comparing preoperative chemoradiotherapy followed by surgery with surgery alone were published. The CROSS trial showed a significant survival advantage in the chemoradiotherapy arm, with a HR of 0.73 for the subgroup of patients with adenocarcinoma, which made up for 75% of patients in the trial (van Hagen 2012). The FFCD 9901 trial, which has so far only been published in abstract form (Mariette 2010), found no survival advantage for patients in the chemoradiotherapy arm. However, it did not include GE junction tumors, 75% of patients had squamous cell carcinoma, and two‐thirds of patients had a node‐negative tumor stage. Therefore, the interpretation of its results with regard to the setting of our meta‐analysis, locoregionally advanced gastroesophageal adenocarcinoma, is difficult.

To demonstrate superiority of either perioperative chemoradiotherapy or chemotherapy was not within the scope of this meta‐analysis. To date, there are only two randomized trials directly comparing the two modalities (Burmeister 2011; Stahl 2009). Both trials stipulated cisplatin‐based chemotherapy schemes, included esophageal and GE junction tumors, and closed early due to slow recruitment. In a German phase three trial (Stahl 2009) the targeted statistical power for analysis of the primary outcome overall survival was not achieved. Nonetheless, patients in the chemoradiotherapy arm of the trial had a survival advantage of 20% at three years with borderline statistical significance (P = 0.07). The Australian trial (Burmeister 2011) was a randomized phase two trial which primarily assessed response and tolerability. No significant survival differences were found. Of note, both trials showed a markedly higher likelihood of pathological response in the chemoradiotherapy groups. A recent meta‐analysis included both trials and used specific techniques for an indirect comparison between perioperative chemotherapy and chemoradiotherapy for esophageal carcinoma of both histologic subtypes (Sjoquist 2011). It showed a combined HR of 0.88 which was borderline significant (P = 0.07). In summary, the available evidence suggests a benefit from adding radiotherapy to perioperative chemotherapy for adenocarcinoma of the esophagus and GE junction. However, due to limitations in the evidence level, no clear conclusion can be made and further RCTs directly comparing the two modalities are strongly required. To what extent perioperative chemoradiotherapy might be an appropriate treatment alternative for gastric adenocarcinoma remains an open question, as there is no higher‐level evidence. A number of phase two trials showed promising pathologic response rates and adequate tolerability (Ajani 2004; Ajani 2005; Ajani 2006), but as of today the concept has not been investigated in phase three trials.

In a multivariable analysis, we aimed to assess the independent influence of perioperative chemotherapy and other covariates on overall survival. Interestingly, in the final model, the effect of perioperative chemotherapy is slightly diminished in its magnitude and loses statistical significance. Other factors, specifically tumor site, age, and performance status, retain a significant effect on overall survival. From a clinical viewpoint, this is very plausible and intuitive. It should however not deter from recommending perioperative chemotherapy to eligible patients, as the achievable survival benefit is of a highly relevant magnitude in a disease with a still unacceptably poor oncological outcome.

A pooled analysis including all patients for whom respective information was available showed that patients with R1 and R2 resection have a substantially shorter survival than those with complete resection. This emphasizes that microscopically complete resection should always be the goal when planning surgery in a patient with gastroesophageal adenocarcinoma. For patients in whom complete resection seems doubtful based on preoperative staging, the indication for resection must be critically evaluated. Given the non‐negligible postoperative morbidity and the common early postoperative decrease in quality of life (Djarv 2012; Munene 2012), palliative resection should be reserved for patients with substantial tumor symptoms which can be expected to improve after surgery.

Perioperative chemotherapy competes with exclusively adjuvant, i.e. postoperative, treatment modalities for gastroesophageal adenocarcinoma. Both adjuvant chemotherapy and chemoradiotherapy have been evaluated in randomized trials. Based on positive results from a large phase three trial (Macdonald 2001), adjuvant chemoradiotherapy for adenocarcinoma of the stomach and GE junction has gained wide popularity in the United States. In Europe, the modality has not been widely adopted, as the trial included a large proportion of patients who had received only D1 lymphadenectomy, which is deemed insufficient according to European standards. The ongoing Dutch‐Swedish multicenter CRITICS trial evaluates if the addition of postoperative chemoradiotherapy to perioperative chemotherapy confers a survival benefit in a setting where a high rate of D2 lymphadenectomies is expected (Dikken 2011). Adjuvant chemotherapy is the standard of care for advanced gastric adenocarcinoma in Asia, where preoperative treatment is uncommon. A recent IPD meta‐analysis including 17 RCTs comparing postoperative chemotherapy with surgery alone for gastric adenocarcinoma showed a 7.4% benefit in 10‐year overall survival for patients receiving postoperative chemotherapy (Paoletti 2010). A sufficiently powered randomized direct comparison between preoperative and postoperative chemotherapy has not been conducted, and thus superiority of one or the other modality cannot be deducted. One factor often hampering postoperative modalities in clinical practice is that only a small percentage of eligible patients receive the planned therapy, mostly because of poor performance status or low compliance in the postoperative period during which adjuvant treatment should commence. In a prematurely halted RCT comparing preoperative with postoperative taxane‐based chemotherapy for gastric adenocarcinoma, 97 per cent of patients randomized for preoperative treatment received chemotherapy, whereas this was the case only for 66 per cent of patients randomized for postoperative treatment (Biffi 2010). The extent of this problem is emphasized by our findings, which show that in the three trials stipulating postoperative chemotherapy in their protocol, only 22 to 42 per cent of patients actually received all planned cycles. Conversely, one argument often raised against preoperative treatment, namely the fear of elevated postoperative morbidity, is not supported by our results, which do not show a higher risk of complications in pretreated patients.

The strength of our meta‐analysis is its rigorous methodology. By adhering to Cochrane standards, we are confident that all available studies meeting our inclusion criteria have been identified and included in the analyses. We included only RCTs in our meta‐analysis, which results in the highest possible evidence level. IPD meta‐analysis is considered to have numerous advantages over meta‐analysis of aggregate data (Riley 2010). It allows for uniform inclusion criteria across trials, to verify data consistency, and to use updated follow‐up data from the single trials. Most importantly, in our specific case the availability of IPD enabled us to include patients with adenocarcinoma from trials with carcinomas of mixed histology in our meta‐analysis of the primary and secondary outcomes as well as in the subgroup analyses (CALGB 9781 2008; RTOG 8911 2007; Urba 2001). We solicited IPD from all trialists, and repeatedly followed up our requests in case of no or evasive reply. By doing so, we were able to receive IPD from about 60 per cent of trials. Unfortunately, we were unable to obtain IPD from the two largest trials (MAGIC 2006; OE02 2009), and thus IPD were available for only about 45 per cent of patients. This rather low yield seems at first disappointing, but is a problem not unfamiliar to IPD meta‐analysts. Instead of limiting our analyses to trials which had provided IPD, we chose to combine IPD and aggregate data with the so‐called two‐stage method (Riley 2007). We believe that this approach draws the most comprehensive picture by incorporating all available evidence and lending sufficient statistical power to the analyses. A sensitivity analysis showed that there were no tangible differences when analyzing IPD and aggregate data separately, which supports the use of the two‐stage method in this case.

Including trials on patients with adenocarcinoma of three anatomically different locations, esophagus, GE junction, and stomach, into one joint meta‐analysis bears a risk of heterogeneity. This risk is augmented by the fact that surgical approaches are different between stomach, GE junction, and esophageal tumors, and that particularly for GE junction and esophageal tumors, surgical techniques differed vastly between trials and even between participating centers or single patients within specific trials. Similarly, the definitions used for stratifying tumors into different locations (esophagus, GE junction, stomach) differed between trials. Therefore, although statistical heterogeneity was rather low for most outcomes and subgroup analyses, there might be relevant clinical heterogeneity. In fact, our results suggest treatment effects differ between the three tumor locations. Nevertheless, given that gastroesophageal adenocarcinoma is frequently treated according to the same chemotherapeutical protocol irrespective of the exact tumor location, we deem it a valid approach to include all tumor locations into one joint meta‐analysis. This approach is supported by the finding that tumor site does not predict survival outcome for patients with gastroesophageal adenocarcinoma treated with chemotherapy in the palliative setting (Chau 2009). The subgroup analysis we conducted helps in identifying patient populations in which treatment effects are particularly pronounced or diminished.

Authors' conclusions

Implications for practice.

Perioperative chemotherapy and chemoradiotherapy appear to lead to prolonged survival in patients with gastroesophageal adenocarcinoma. It should thus be considered standard of care and offered to all eligible patients. There is a trend to a larger survival advantage for tumors of the gastroesophageal junction as compared to other sites and for chemoradiotherapy as compared to chemotherapy. Likewise, there is an interaction between age and treatment effect, with younger patients experiencing a larger survival advantage and no survival advantage for elderly patients. Fears of elevated postoperative morbidity and mortality following perioperative chemotherapy do not seem to be justified.

Implications for research.

The results of our subgroup analyses, such as the more pronounced survival advantage found for tumors of the gastroesophageal junction as compared to other sites, in younger patients compared to older ones, and for chemoradiotherapy compared to chemotherapy, need to be validated in prospective trials.

Appendices

Appendix 1. MEDLINE search strategy

1     exp drug therapy/ 
 2     chemotherap$.tw. 
 3     modality therap$.tw. 
 4     combined modality therapy/ or neoadjuvant therapy/ or photochemotherapy/ 
 5     Antineoplastic Combined Chemotherapy Protocols/ 
 6     exp Antineoplastic Agents/ 
 7     antineoplastic$.tw. 
 8     (neoadjuvant adj5 chemo$).tw. 
 9     (adjuvant adj5 chemo$).tw. 
 10     ((preoperative or postoperative or perioperative) adj5 chemo$).tw. 
 11     ((pre‐operative or post‐operative or peri‐operative) adj5 chemo$).tw. 
 12     or/1‐11 
 13     exp Adenocarcinoma/ 
 14     adenocarcinoma$.tw. 
 15     13 or 14 
 16     exp Esophagus/ 
 17     Esophagogastric Junction/ 
 18     (gastroesophag$ adj3 junction$).tw. 
 19     (gastro‐esophag$ adj3 junction$).tw. 
 20     oesophago‐gastric junction$.tw. 
 21     esophago‐gastric junction$.tw. 
 22     (gastrooesophag$ adj3 junction$).tw. 
 23     (gastro‐oesophag$ adj3 junction$).tw. 
 24     oesophagogastric junction$.tw. 
 25     esophagogastric junction$.tw. 
 26     Stomach/ 
 27     or/16‐26 
 28     15 and 27 
 29     exp Stomach Neoplasms/ 
 30     exp Esophageal Neoplasms/ 
 31     (esophag$ adj5 neoplas$).tw. 
 32     (oesophag$ adj5 neoplas$).tw. 
 33     (esophag$ adj5 cancer$).tw. 
 34     (oesophag$ adj5 cancer$).tw. 
 35     (esophag$ adj5 carcin$).tw. 
 36     (oesophag$ adj5 carcin$).tw. 
 37     (esophag$ adj5 tumo$).tw. 
 38     (oesophag$ adj5 tumo$).tw. 
 39     (esophag$ adj5 metasta$).tw. 
 40     (oesophag$ adj5 metasta$).tw. 
 41     (esophag$ adj5 malig$).tw. 
 42     (oesophag$ adj5 malig$).tw. 
 43     (esophag$ adj5 adenocarcinoma$).tw. 
 44     (oesophag$ adj5 adenocarcinoma$).tw. 
 45     (stomach adj5 neoplas$).tw. 
 46     (stomach adj5 adenocarcinoma$).tw. 
 47     (stomach adj5 cancer$).tw. 
 48     (stomach adj5 carcin$).tw. 
 49     (stomach adj5 tumo$).tw. 
 50     (stomach adj5 metasta$).tw. 
 51     (stomach adj5 malig$).tw. 
 52     (gastric adj5 neoplas$).tw. 
 53     (gastric adj5 adenocarcinoma$).tw. 
 54     (gastric adj5 cancer$).tw. 
 55     (gastric adj5 carcin$).tw. 
 56     (gastric adj5 tumo$).tw. 
 57     (gastric adj5 metasta$).tw. 
 58     (gastric adj5 malig$).tw. 
 59     or/29‐58 
 60     28 or 59 
 61     resect$.tw. 
 62     surg$.tw. 
 63     surgery/ 
 64     opera$.tw. 
 65     Esophagectomy/ 
 66     Gastrectomy/ 
 67     Gastrectomy.tw. 
 68     Esophagectomy.tw. 
 69     or/61‐68 
 70     60 and 69 
 71     12 and 70 
 72     randomized controlled trial.pt. 
 73     controlled clinical trial.pt. 
 74     randomized.ab. 
 75     placebo.ab. 
 76     drug therapy.fs. 
 77     randomly.ab. 
 78     trial.ab. 
 79     groups.ab. 
 80     or/72‐79 
 81     humans.sh. 
 82     80 and 81 
 83     71 and 82

Data and analyses

Comparison 1. Overall survival.

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1 Hazard ratio plot for overall survival	14	2422	Hazard Ratio (Random, 95% CI)	0.81 [0.73, 0.89]
2 Hazard ratio plot for overall survival by type of data	14		Hazard Ratio (Random, 95% CI)	Subtotals only
2.1 Individual patient data	8	1049	Hazard Ratio (Random, 95% CI)	0.80 [0.66, 0.97]
2.2 Aggregated data	6	1373	Hazard Ratio (Random, 95% CI)	0.81 [0.72, 0.92]
3 Hazard ratio plot for overall survival by tumor site	11		Hazard Ratio (Random, 95% CI)	Subtotals only
3.1 Esophagus	5	473	Hazard Ratio (Random, 95% CI)	0.87 [0.73, 1.05]
3.2 GE junction	6	470	Hazard Ratio (Random, 95% CI)	0.69 [0.54, 0.87]
3.3 Stomach	7	828	Hazard Ratio (Random, 95% CI)	0.94 [0.82, 1.06]
4 Interaction treatment‐tumor site (only IPD)	4	717	Interaction coefficient (Random, 95% CI)	0.00 [‐0.38, 0.38]
5 Hazard ratio plot for overall survival by chemo‐/radiotherapy	14		Hazard Ratio (Random, 95% CI)	Subtotals only
5.1 Chemotherapy only	10	2033	Hazard Ratio (Random, 95% CI)	0.83 [0.75, 0.91]
5.2 Chemoradiotherapy	4	389	Hazard Ratio (Random, 95% CI)	0.70 [0.50, 0.99]
6 Hazard ratio plot for overall survival by timing of regimen	14		Hazard Ratio (Random, 95% CI)	Subtotals only
6.1 Preoperative regimen	9	1236	Hazard Ratio (Random, 95% CI)	0.81 [0.68, 0.95]
6.2 Preoperative and postoperative combined regimen	5	1186	Hazard Ratio (Random, 95% CI)	0.80 [0.70, 0.91]
7 Hazard ratio plot for overall survival by chemotherapeutic agents	14		Hazard Ratio (Random, 95% CI)	Subtotals only
7.1 Non‐platinum and non‐anthracycline regimen	2	231	Hazard Ratio (Random, 95% CI)	0.89 [0.64, 1.23]
7.2 Platinum‐based non‐anthracycline regimen	10	1632	Hazard Ratio (Random, 95% CI)	0.80 [0.72, 0.89]
7.3 Anthracycline‐based non‐platinum regimen	1	56	Hazard Ratio (Random, 95% CI)	1.40 [0.78, 2.53]
7.4 Platinum‐ and anthracycline‐based regimen	1	503	Hazard Ratio (Random, 95% CI)	0.75 [0.60, 0.93]
8 Hazard ratio plot for overall survival by performance status (only IPD)	6		Hazard Ratio (Random, 95% CI)	Totals not selected
8.1 Performance status 0	6		Hazard Ratio (Random, 95% CI)	0.0 [0.0, 0.0]
8.2 Performance status 1	6		Hazard Ratio (Random, 95% CI)	0.0 [0.0, 0.0]
9 Interaction treatment‐performance status (only IPD)	6	841	Interaction coefficient (Random, 95% CI)	0.34 [‐0.03, 0.71]
10 Hazard ratio plot for overall survival by age (only IPD)	8		Hazard Ratio (Random, 95% CI)	Totals not selected
10.1 < 65 years	8		Hazard Ratio (Random, 95% CI)	0.0 [0.0, 0.0]
10.2 65 to 75 years	8		Hazard Ratio (Random, 95% CI)	0.0 [0.0, 0.0]
11 Interaction treatment‐age (only IPD)	8	1032	Interaction coefficient (Random, 95% CI)	‐0.07 [‐0.40, 0.26]
12 Hazard ratio plot for overall survival by sex (only IPD)	8		Hazard Ratio (Random, 95% CI)	Totals not selected
12.1 Male	8		Hazard Ratio (Random, 95% CI)	0.0 [0.0, 0.0]
12.2 Female	8		Hazard Ratio (Random, 95% CI)	0.0 [0.0, 0.0]
13 Interaction treatment‐sex (only IPD)	6	849	Interaction coefficient (Random, 95% CI)	‐0.18 [‐0.59, 0.23]
14 Hazard ratio plot for overall survival by pretreatment T class (only IPD)	3		Hazard Ratio (Random, 95% CI)	Totals not selected
14.1 T stage 0/1/2	3		Hazard Ratio (Random, 95% CI)	0.0 [0.0, 0.0]
14.2 T stage 3/4	3		Hazard Ratio (Random, 95% CI)	0.0 [0.0, 0.0]
15 Interaction treatment‐T class (only IPD)	1	231	Interaction coefficient (Random, 95% CI)	0.02 [‐0.57, 0.61]
16 Hazard ratio plot for overall survival by pretreatment N class (only IPD)	4		Hazard Ratio (Random, 95% CI)	Totals not selected
16.1 N stage 0	4		Hazard Ratio (Random, 95% CI)	0.0 [0.0, 0.0]
16.2 N stage 1/2/3	4		Hazard Ratio (Random, 95% CI)	0.0 [0.0, 0.0]
17 Interaction treatment‐N class (only IPD)	4	453	Interaction coefficient (Random, 95% CI)	‐0.02 [‐0.63, 0.59]

Comparison 2. Disease‐free survival (landmark time 6 months).

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1 Hazard ratio plot for disease‐free survival (landmark time 6 months)	7	931	Hazard Ratio (Random, 95% CI)	0.84 [0.69, 1.01]

Comparison 3. Presence of tumor‐free resection margin.

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1 Odds ratio plot for tumor‐free resection margin	10	1665	Odds Ratio (M‐H, Random, 95% CI)	1.42 [0.97, 2.06]

Comparison 4. Tumor stage at resection.

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1 Odds ratio plot for tumor stage T0/T1/T2	7	1410	Odds Ratio (M‐H, Random, 95% CI)	1.53 [1.02, 2.31]
2 Odds ratio plot for nodal status N0	8	1507	Odds Ratio (M‐H, Random, 95% CI)	2.43 [1.48, 3.99]

Comparison 5. Safety of perioperative chemotherapy regimen.

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1 Number of grade 3/4 toxicities per patient			Other data	No numeric data

Comparison 6. Postoperative morbidity.

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1 Risk difference plot for postoperative morbidity	9	1397	Risk Difference (M‐H, Random, 95% CI)	0.01 [‐0.03, 0.05]

Comparison 7. Postoperative mortality.

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1 Risk difference plot for postoperative mortality	11	1606	Risk Difference (M‐H, Random, 95% CI)	0.00 [‐0.01, 0.02]

Characteristics of studies

Characteristics of included studies [ordered by study ID]

ACCORD 07 2011.

Methods	RCT
Participants	Histologically proven adenocarcinoma of the lower third of the esophagus or esophagogastric junction or (beginning March 1998) stomach (no exact definition given on stratification of tumor location); UICC stage II or greater; suitable for curative resection; WHO performance status 0 or 1; age 18 to 75 yrs; 1995‐2003; France (multi‐center) 224 patients (187 male, 37 female) Median potential follow‐up calculated from IPD, see Risk of bias in included studies Patients randomized/analyzed in this meta‐analysis: 224/224
Interventions	Cisplatin, 5‐fluorouracil pre‐ and postoperative vs surgery alone Intervention group: 2 to 3 cycles (cisplatin 100 mg/m² on day 1 or 2; 5‐fluorouracil 4000 mg/m² cumulative dose over 5 days, then 22 days break) preop., surgery 4 to 6 weeks after last chemotherapy dose, 3 to 4 cycles (see above) postop. 4 to 6 weeks after surgery for patients who had R0 resection, no progression or major toxicity during preop. therapy and at least T3 or N+ tumor in histopathology Surgery: local surgeon decided procedure depending on tumor site and local practice; recommended: D2 lymphadenectomy
Outcomes	OS, DFS, presence of tumor‐free resection margin, tumor stage at resection, safety of chemotherapy, perioperative morbidity, non‐administration of postoperative therapy
Notes	IPD
Risk of bias
Bias	Authors' judgement	Support for judgement
Selection bias	Low risk	Randomization done via a telephone call to the Department of Biostatistics and Epidemiology at the Institute Gustave Roussy, stratified by grade, tumor site, center
Attrition bias	Low risk	Some patients in both groups did not receive allocated treatment, but exclusively due to complications/disease progression. All were included in final analysis.
Performance bias	Low risk	No obvious differences in type of surgery performed
Detection bias	Low risk	Follow‐up of both groups occurred every 6 months for 5 years
Bias due to missing data	Low risk	No missing data
Reporting bias	Low risk	Nothing that would suggest selective reporting/IPD available
Other bias	Low risk	Nothing that would suggest relevant other bias
Overall risk of bias	Low risk	Low

CALGB 9781 2008.

Methods	RCT
Participants	Histologically documented squamous cell or adenocarcinoma of the thoracic esophagus (below 20 cm) or GE junction (< 2 cm distal spread into gastric cardia), surgically resectable (T1‐3, NX), including regional thoracic lymph node (N1) metastases, supraclavicular lymph node metastasis < 1.5 cm, lymph node metastases to levels 15 to 20 < 1.5 cm; no age limit; 1997 to 2000; USA (multi‐center) 42 patients with adenocarcinoma (41 male, 1 female) out of 56 total patients Median potential follow‐up calculated from IPD, see Risk of bias in included studies Patients with adenocarcinoma randomized/analyzed in this meta‐analysis: 42/42
Interventions	Cisplatin, 5‐fluorouracil with concurrent radiotherapy preoperative vs surgery alone Intervention group: 1 cycle (cisplatin 200 mg/m² cumulative dose on days 1 and 29, 5‐fluorouracil 8000 mg/m² cumulative dose on days 1 to 4 and 29 to 32, radiotherapy (1.8 Gy/5 d/wk) begun within 24 hours of the chemotherapy administration, continued for 5.5 weeks, final 5.4 Gy given as a boost (total dose 50.4 Gy)) Surgery: left chest or right chest and abdomen (Ivor‐Lewis) recommended for mid‐esophageal and GE junction tumours, transhiatal esophagectomy discouraged, all technically accessible lymph nodes to be removed
Outcomes	OS, DFS, safety of chemotherapy, perioperative morbidity
Notes	IPD
Risk of bias
Bias	Authors' judgement	Support for judgement
Selection bias	Low risk	Central telephone registration
Attrition bias	Low risk	All randomized patients included in analysis of primary outcome
Performance bias	Unclear risk	Insufficient information given
Detection bias	Low risk	No relevant differences in follow‐up between groups
Bias due to missing data	Low risk	No missing data for primary outcome
Reporting bias	Low risk	Nothing that would suggest selective reporting/IPD available
Other bias	Low risk	Nothing that would suggest relevant other bias
Overall risk of bias	Unclear risk	Moderate

EORTC 40954 2010.

Methods	RCT
Participants	Histologically proven adenocarcinoma of the stomach (including AEG Type II and III according to the Siewert classification), cT3/4 NX M0/M1lymph, no signs of peritoneal carcinomatosis in staging laparoscopy; WHO performance status 0 to 1; age 18 to 70 years; 1999 to 2004, several European countries and Egypt (multi‐center) 144 patients (100 male, 44 female) Median potential follow‐up calculated from IPD, see Risk of bias in included studies Patients randomized/analyzed in this meta‐analysis: 144/144
Interventions	Cisplatin, 5‐fluorouracil, folinic acid preoperative vs surgery alone Intervention group: 1 cycle (cisplatin 150 mg/m² cumulative dose on days 1, 15 and 29; 5‐fluorouracil 12,000 mg/m² cumulative dose on days 1, 8, 15, 22, 29 and 36; folinic acid 3000 mg/m² cumulative dose on days 1, 8, 15, 22, 29 and 36); restaging, if no progression or toxicity 1 more cycle as described above restarting on day 50; surgery on days 57 to 63 of the second cycle Surgery: subtotal or total gastrectomy with extension depending on tumor location, either D1 lymphadenectomy (for perigastric nodes at lesser and greater curvature; 7 patients) or, preferably, D2 lymphadenectomy (for regional lymph nodes outside the perigastric area; 130 patients); resection could be extended to other organs or locations to achieve complete removal of the primary tumor or suspicious lymph nodes.
Outcomes	OS, DFS, presence of tumor‐free resection margin, tumor stage at resection, safety of chemotherapy, perioperative morbidity
Notes	IPD
Risk of bias
Bias	Authors' judgement	Support for judgement
Selection bias	Low risk	Computerized randomization
Attrition bias	Low risk	All patients included in analysis of primary endpoint
Performance bias	Low risk	Identical surgical approach stipulated for both groups, no relevant differences in rate of D2‐lymphadenectomy
Detection bias	Low risk	Meticulous follow‐up schedule, no significant difference in median follow‐up between groups
Bias due to missing data	Low risk	Analysis of primary outcome based on all patients
Reporting bias	Low risk	Nothing that would suggest selective reporting/IPD available
Other bias	Low risk	Nothing that would suggest relevant other bias
Overall risk of bias	Low risk	Low

FAMTX 2004.

Methods	RCT
Participants	Histologically proven adenocarcinoma of the stomach (not cardia, i.e. no tumour growth above Z‐line); > cT1; resectable with no evidence of distant metastases; WHO performance status 0 to 2; < 75 years; 1993 to 1996; Netherlands (multi‐center) 56 patients (32 male, 24 female) Median potential follow‐up calculated from IPD, see Risk of bias in included studies Patients randomized/analyzed in this meta‐analysis: 59/56
Interventions	5‐fluorouracil, leucovorin, doxorubicin, methotrexate preoperative vs surgery alone (D1 lymphadenectomy) Intervention group: 2 cycles (methotrexate 1500 mg/m² on day 2; 5‐fluorouracil 1500 mg/m² on day 2; leucovorin 240 or 480 mg (depending on MTX level) cumulative dose on days 3 to 4; doxorubicin 30 mg/m² on day 15; 13 days break); re‐staging; in case of response or stable disease another 2 cycles (see above); surgery: resection with D1 lymphadenectomy, with staging biopsy of para‐aortic lymph nodes
Outcomes	OS, DFS, presence of tumor‐free resection margin, tumor stage at resection, safety of chemotherapy, perioperative morbidity
Notes	IPD
Risk of bias
Bias	Authors' judgement	Support for judgement
Selection bias	Low risk	"...randomization in blocks of six and with stratification according to the centre was conducted." "The participating centres registered pts. by means of telephone calls to the central office of the trial..."
Attrition bias	Low risk	2 patients in intervention group and 1 patient in control group were excluded due to non‐fulfilment of inclusion criteria; thus there is no high likelihood of bias
Performance bias	Low risk	Same surgical approach stipulated for both groups
Detection bias	Low risk	No reason to assume differences in follow‐up among groups
Bias due to missing data	Low risk	No missing data
Reporting bias	Low risk	Nothing that would suggest selective reporting/IPD available
Other bias	Low risk	Nothing that would suggest relevant other bias
Overall risk of bias	Low risk	Low

Feng 2008.

Methods	RCT
Participants	Histologically proven adenocarcinoma of the stomach (only Bormann type IV); no distant metastases; Karnofsky index at least 80%; no age limit; 2002 to 2004; China (single‐center) 55 patients (36 male, 19 female) Median follow‐up not given in publication Patients randomized/analyzed in this meta‐analysis: 55/52 (for primary outcome)
Interventions	Docetaxel, cisplatin, 5‐fluorouracil pre‐ and postoperative vs only postoperative Intervention group: 3 cycles (docetaxel 75 mg/m² on day 1; cisplatin: 90 mg/m² cumulative dose on days 1 to 3; 5‐fluorouracil 2500 mg/m² cumulative dose on days 1 to 5, then 16 days break) preop.; surgery 12 to 14 days after end of last cycle; 3 cycles (see above) postop. starting 6 to 12 weeks after surgery Control group: immediate surgery; 6 cycles (docetaxel 75 mg/m² on day 1; cisplatin: 90 mg/m² cumulative dose on days 1 to 3; 5‐fluorouracil 2500 mg/m² cumulative dose on days 1 to 5, then 16 days break) postop. Surgery: no details given
Outcomes	OS, presence of tumor‐free resection margin, tumor stage at resection (only nodal status), safety of chemotherapy, perioperative morbidity
Notes	Aggregate data from publication
Risk of bias
Bias	Authors' judgement	Support for judgement
Selection bias	Unclear risk	No details on randomization given in manuscript
Attrition bias	Low risk	2/29 patients in intervention and 1/26 patient in control group were lost to follow‐up; thus there is no high likelihood of bias
Performance bias	Unclear risk	No details on stipulated surgical approach given in manuscript
Detection bias	Low risk	Patients in both groups were followed up the same way
Bias due to missing data	Low risk	Analysis of outcomes based on data from all patients who were followed up
Reporting bias	Unclear risk	No definition of predefined endpoints available
Other bias	Low risk	Nothing that would suggest relevant other bias
Overall risk of bias	High risk	High

Kobayashi 2000.

Methods	RCT
Participants	Resectable gastric cancer, excluding early gastric cancer and "complicated cases"; < 75 years; 1990 to 1993; Japan (multi‐center) 171 patients (120 male, 51 female) Median follow‐up not given in publication Patients randomized/analyzed in this meta‐analysis: 178/171
Interventions	Oral 5´‐deoxy‐5‐fluorouridine (5´‐DFUR) preoperative and 5´‐deoxy‐5‐fluorouridine (5´‐DFUR) combined with mitomycin C postoperative vs only 5´‐deoxy‐5‐fluorouridine (5´‐DFUR) combined with mitomycin C postoperative Intervention group: 5´‐DFUR 6100 mg/m² cumulative dose on days 1 to 10; surgery on day 11; mitomycin C 12 mg/m² cumulative dose on days 12 to 13; 5´‐DFUR 600 mg/day from day 19 for 2 years Control group: surgery on day 1; mitomycin C 12 mg/m² cumulative dose on days 2 to 3; 5´‐DFUR 600 mg/day from day 8 for 2 years Surgery: no details given
Outcomes	OS, presence of tumor‐free resection margin, tumor stage at resection (only T stage), safety of chemotherapy
Notes	Aggregate data from publication
Risk of bias
Bias	Authors' judgement	Support for judgement
Selection bias	Unclear risk	No clear description of allocation concealment, random sequence generation etc.
Attrition bias	Unclear risk	7 patients were excluded from further follow‐up after randomization due to remote metastasis and transfer to other hospitals; unclear, how those were distributed among groups
Performance bias	Unclear risk	No information on stipulated surgical approach given
Detection bias	Unclear risk	No information on follow‐up given
Bias due to missing data	Unclear risk	No information on potentially missing data given
Reporting bias	Unclear risk	No definition of predefined endpoints available
Other bias	Low risk	Nothing that would suggest relevant other bias
Overall risk of bias	High risk	High

MAGIC 2006.

Methods	RCT
Participants	Histologically proven adenocarcinoma of stomach or lower third of the esophagus (no exact definition given on stratification of tumor location); UICC Stage II or higher; no distant metastases or locally advanced inoperable disease; WHO performance status 0 or 1; no age limit; 1994 to 2002; UK, Netherlands, New Zealand, Germany, Singapore, Brazil (multi‐center) 503 patients (396 male, 107 female) Median follow‐up: 49 months in perioperative‐chemotherapy group and 47 months in surgery group Patients randomized/analyzed in this meta‐analysis: 503/503
Interventions	Epirubicin, cisplatin, 5‐fluorouracil pre‐ and postoperative vs surgery alone Intervention group: 3 cycles (epirubicin 50 mg/m² on day 1; cisplatin: 60 mg/m² on day 1; 5‐fluorouracil 4200 mg/m² cumulative dose on days 1 to 21) preop.; surgery 3 to 6 weeks after last chemotherapy dose; 3 cycles (see above) postop. starting 6 to 12 weeks after surgery Surgery: frozen section of preaortic, infracolic node, if positive management at discretion of surgeon; total or subtotal gastrectomy including greater and lesser omenta and any other organs involved by extension of the primary growth; resection lines at least 3 cm from edge of the macroscopic tumor; surgeon decided extent of lymph node dissection; nodes along lesser and greater curvatures and at origin of left gastric artery to be included; nodal sampling for histologic examination of other groups recommended; in esophagectomy thoracic approach not stipulated; object of nodal dissection to remove periesophageal nodes; separate sampling of subcarinal and celiac axis lymph nodes recommended
Outcomes	OS, DFS, presence of tumor‐free resection margin, tumor stage at resection, safety of chemotherapy, perioperative morbidity, non‐administration of postoperative therapy
Notes	Aggregate data from publication
Risk of bias
Bias	Authors' judgement	Support for judgement
Selection bias	Low risk	"Treatment was allocated with use of the minimization method according to the following stratification factors: age, tumor site, WHO perform. status, surgeon." "Patients...randomly assigned...by means of a telephone call to the MRC Clinical Trials Unit."
Attrition bias	Low risk	No attrition reported
Performance bias	Low risk	No relevant differences in surgical procedures between trial arms
Detection bias	Unclear risk	"The numbers of surviving patients with less than two years of follow‐up were 17 in the perioperative‐chemotherapy group and 35 in the surgery group."
Bias due to missing data	Low risk	Missing data on safety of preoperative chemotherapy for 4 patients, otherwise no missing data reported
Reporting bias	Unclear risk	No definition of predefined endpoints available
Other bias	Low risk	Nothing that would suggest relevant other bias
Overall risk of bias	Unclear risk	Moderate

OE02 2009.

Methods	RCT
Participants	Histologically confirmed squamous‐, adeno‐, or undifferentiated carcinoma of the esophagus (excluding 'postcricoid' tumors) and the cardia (no exact definition given on stratification of tumor location); no cervical lymph node involvement or metastases; resectable; no contraindication to surgery or chemotherapy; 1992 to 1998; Europe (multi‐center) 533 patients (sex distribution not available) with adenocarcinoma out of 802 total patients Median follow‐up for surviving patients: 6.1 years in the surgical resection alone group and 5.9 years in the CS group Patients randomized/analyzed in this meta‐analysis: of 871 patients randomized in the trial, only 802 were available for analysis; there are no data available as to how many of the excluded patients had adenocarcinoma
Interventions	Cisplatin, 5‐fluorouracil preoperative vs surgery alone Intervention group: 2 cycles (cisplatin 80 mg/m² on day 1; 5‐fluorouracil 4000 mg/m² cumulative dose over days 1 to 4; 17 days break); surgery within 3 to 5 weeks after completion of chemotherapy Preoperative radiotherapy (25 Gy in 5 fractions over 1 week, 32.5 Gy in 10 fractions over 2 weeks, or a biologically equivalent dose) permitted on a center basis and then to be used in both groups (given to 9% of patients in both groups) Surgery: local surgeon decided procedure for patients in both treatment groups in accordance with site of tumour and local practice
Outcomes	OS
Notes	Aggregate data from publication
Risk of bias
Bias	Authors' judgement	Support for judgement
Selection bias	Low risk	“Randomisation was done by telephone call to the Cancer Division of the MRC CTU. We randomly assigned patients by minimization, with criteria of surgeon, site of tumour, WHO PS, and histology.”
Attrition bias	Low risk	69 patients (34 in intervention group, 35 in control group) from China excluded from follow‐up due to logistic problems; however it is unlikely that this introduces systematic bias
Performance bias	Unclear risk	"The surgical procedure was selected by the surgeon according to tumor site and local practice." Thus, systematic differences cannot be excluded.
Detection bias	Low risk	No relevant differences among groups reported for follow‐up
Bias due to missing data	Low risk	Analysis of primary outcome based on all patients
Reporting bias	Unclear risk	No definition of predefined endpoints available
Other bias	Low risk	Nothing that would suggest relevant other bias
Overall risk of bias	Unclear risk	Moderate

RTOG 8911 2007.

Methods	RCT
Participants	Histologically confirmed squamous cell or adenocarcinoma of the thoracic esophagus (upper border > 18 cm from incisor teeth) or gastroesophageal junction (no exact definition given on stratification of tumor location); stage I‐III excluding T4 tumors; absence of supraclavicular or distant metastases; fit for surgery; at least 18 years; 1990‐1995; USA (multi‐center) 236 patients (213 male, 23 female) with adenocarcinoma out of 443 total patients Median potential follow‐up calculated from IPD, see Risk of bias in included studies Patients with adenocarcinoma randomized/analyzed in this meta‐analysis: 244/236
Interventions	Cisplatin, 5‐fluorouracil pre‐ and postoperative vs surgery alone Intervention group: 3 cycles (cisplatin 100 mg/m² on day 1; 5‐fluorouracil 1000 mg/m² cumulative dose on days 1 to 5, 23 days break); operation 2 to 4 weeks after the end of the last cycle; in case of stable or responsive disease upon surgery 2 postoperative cycles (see above, except cisplatin dose reduced to 75 mg/m²) starting 2 to 6 weeks after surgery Surgery: acceptable procedures were Ivor–Lewis esophagogastrectomy with high intrathoracic anastomosis above the level of the azygos vein, subtotal thoracic esophagectomy by left thoracoabdominal incision with anastomosis above aortic arch, or complete thoracic esophagectomy; restoration of continuity by gastric–esophageal anastomosis or colonic interposition with cervical anastomosis acceptable; transhiatal esophagectomy acceptable only for lesions located below carina; proximal and distal margins to be at least 2 cm below gross tumor; removal of all accessible lymph nodes strongly recommended; strongly recommended that participating surgeons perform at least 4 esophageal resections yearly
Outcomes	OS, DFS, presence of tumor‐free resection margin, safety of chemotherapy, perioperative morbidity, non‐administration of postoperative therapy
Notes	IPD
Risk of bias
Bias	Authors' judgement	Support for judgement
Selection bias	Low risk	"The randomization scheme described by Zelen was used with two stratification variables..."
Attrition bias	Unclear risk	23 randomized patients (17 in intervention group, 6 in control group) dropped out due to ineligibility, and another 4 eligible patients (3 in intervention group, 1 in control group) dropped out due to no follow‐up data
Performance bias	Low risk	Surgical approach clearly stipulated for both groups
Detection bias	Low risk	No reason to assume differences in follow‐up among groups
Bias due to missing data	Low risk	No missing data for patients included in the analyses
Reporting bias	Low risk	Nothing that would suggest selective reporting/IPD available
Other bias	Low risk	Nothing that would suggest relevant other bias
Overall risk of bias	Unclear risk	Moderate

TROG‐AGITG 2005.

Methods	RCT
Participants	Histologically proven invasive cancer of thoracic oesophagus; cT1‐3 cN0‐1 disease; no involvement of cervical esophagus or celiac nodes; ECOG performance status 0 or 1; no age limit; 1994‐2000; Australia, New Zealand, Singapore (multi‐center) 158 participants with adenocarcinoma (147 male, 11 female) out of 256 total patients Median potential follow‐up calculated from IPD, see Risk of bias in included studies Patients with adenocarcinoma randomized/analyzed in this meta‐analysis: 158/158
Interventions	Cisplatin, 5‐fluorouracil with concurrent radiotherapy preop. vs surgery alone 1 cycle (cisplatin 80 mg/m² on day 1; 5‐fluorouracil 3200 mg/m² cumulative dose on days 1 to 4) with 35 Gy radiotherapy in 15 fractions over 3 weeks, starting concurrently with chemotherapy; surgery 3 to 6 weeks after completion of radiotherapy; postoperative radiotherapy permitted for patients with residual disease after surgery if indicated clinically for patients assigned to surgery alone Surgery: total removal of tumour and regional lymph nodes as deemed necessary by operating surgeon; no particular operative approach stipulated; radical lymphadenectomy not mandatory; no minimum number of lymph nodes to be dissected
Outcomes	OS, DFS, presence of tumor‐free resection margin, tumor stage at resection (only nodal status)
Notes	IPD
Risk of bias
Bias	Authors' judgement	Support for judgement
Selection bias	Low risk	"The random sequence was generated by use of minimization...and blocks of four were used." "...stratified...and randomly assigned...by central telephone randomization done by the trial coordinator at the NHMRC Clinical Trials Centre." "The allocation sequence was concealed to all central staff. Research staff and investigators were blinded to treatment assignment before, but not after, randomization."
Attrition bias	Low risk	Some patients in both groups did not receive allocated treatment, but mostly due to complications/disease progression. One patient in the intervention group refused surgery after preoperative treatment. All patients were included in final analysis.
Performance bias	Unclear risk	Insufficient information given in the paper to make a valid judgment
Detection bias	Low risk	No relevant differences in median follow‐up between groups
Bias due to missing data	Low risk	No missing data
Reporting bias	Low risk	Nothing that would suggest selective reporting/IPD available
Other bias	Low risk	Nothing that would suggest relevant other bias
Overall risk of bias	Unclear risk	Moderate

Urba 2001.

Methods	RCT
Participants	Histologically confirmed squamous cell, adenocarcinoma, or mixed adenosquamous carcinoma of the esophagus or GE junction (no exact definition given on stratification of tumor location), limited to esophagus and regional lymph nodes (including celiac nodes); Karnofsky index >= 60%; <= 75 years of age; 1989 to 1994; USA (single‐center) 76 patients with adenocarcinoma (68 male, 9 female) out of 100 total patients Median potential follow‐up calculated from IPD, see Risk of bias in included studies Patients with adenocarcinoma randomized/analyzed in this meta‐analysis: 76/76 (1 patient was initially classified as squamous cell carcinoma, but then re‐classified as adenocarcinoma and thus analyzed in our meta‐analysis)
Interventions	Cisplatin, 5‐fluorouracil, vinblastine with concurrent radiotherapy preoperative vs surgery alone Intervention group: 1 cycle (cisplatin 200 mg/m² cumulative dose on days 1 through 5 and 17 through 21, 5‐fluorouracil 6300 mg/m² cumulative on days 1 through 21, vinblastin 8 mg/m² on days 1 through 4 and 17 through 20, radiotherapy in fractions of 1.5 Gy twice a day, on days 1 through 5, 8 through 12, and 15 through 19, to a total dose of 45 Gy) Surgery: esophagectomy using the transhiatal approach without thoracotomy, with anastomosis between cervical esophagus and gastric fundus above level of clavicles, performed whenever possible; pyloromyotomy, and feeding jejunostomy performed routinely; esophagus replaced with either colon or jejunum in patients with history of prior gastric resection and inadequate remaining length of stomach; accessible intra‐abdominal, paraesophageal, and subcarinal lymph nodes sampled; for localized tumors of the cardia, stomach was divided as to maintain a 4 cm to 6 cm margin beyond gross palpable tumor; for tumors of cervicothoracic esophagus, concomitant laryngopharyngectomy and permanent tracheostomy were required; for intrathoracic esophageal tumors that could not be resected safely using the transhiatal approach, a thoracotomy and transthoracic esophagectomy was carried out; regardless of the method of esophagectomy; cervical esophageal anastomosis was performed whenever possible
Outcomes	OS, DFS, presence of tumor‐free resection margin, tumor stage at resection, safety of chemotherapy, perioperative morbidity
Notes	IPD
Risk of bias
Bias	Authors' judgement	Support for judgement
Selection bias	Low risk	Randomization
Attrition bias	Low risk	All randomized patients included in analysis of primary endpoint
Performance bias	Low risk	Identical surgical approach stipulated for both groups
Detection bias	Unclear risk	No details on follow‐up given
Bias due to missing data	Unclear risk	Missing data for 1 patient
Reporting bias	Low risk	Nothing that would suggest selective reporting/IPD available
Other bias	Low risk	Nothing that would suggest relevant other bias
Overall risk of bias	Unclear risk	Moderate

Walsh 2002.

Methods	RCT
Participants	Adenocarcinoma of the esophagus or GE junction (no exact definition given on stratification of tumor location), excluding carcinomas of the cervical esophagus requiring laryngectomy; no distant metastases; ECOG PS < 3; <= 75 years of age; 1990 to 1995; Ireland (multi‐center) 113 patients (83 male, 30 female) Median potential follow‐up calculated from IPD, see Risk of bias in included studies Patients randomized/analyzed in this meta‐analysis: 113/113 Surgery: 5 operative approaches: tumors of the cardia resected through approach involving abdomen and left side of the chest; in selected patients total gastrectomy and distal esophagectomy performed through abdominal approach; tumors of lower third of esophagus resected by Lewis–Tanner operation; tumors of middle third by 3‐stage operation in which the esophagus was mobilized in the right side of the chest and the anastomosis performed in the neck; in patients with poor respiratory reserve transhiatal approach used
Interventions	Cisplatin, 5‐fluorouracil with concurrent radiotherapy preoperative vs surgery alone Intervention group: 2 cycles (cisplatin 75 mg/m² cumulative dose on day 7, 5‐fluorouracil 75 mg/kg of body weight cumulative on days 1 through 5; repeated after 42 days); radiotherapy in fractions of 2.67 Gy once a day, on days 1 through 5, 8 through 12, and 15 through 19; total dose 45 Gy; surgery 8 weeks after the beginning of treatment
Outcomes	OS, safety of chemotherapy, tumor stage at resection, perioperative morbidity
Notes	IPD
Risk of bias
Bias	Authors' judgement	Support for judgement
Selection bias	Unclear risk	No details on randomization are given
Attrition bias	Low risk	All randomized patients included in analysis of primary endpoint
Performance bias	Low risk	Identical surgical approach stipulated for both groups
Detection bias	Low risk	No relevant differences in follow‐up between the 2 groups
Bias due to missing data	Low risk	No missing data for primary outcome
Reporting bias	Low risk	Nothing that would suggest selective reporting/IPD available
Other bias	Low risk	Nothing that would suggest relevant other bias
Overall risk of bias	Unclear risk	Moderate

Wang 2000.

Methods	RCT
Participants	Endoscopic biopsy proven primary gastric cardia cancer (no details on definition given), previously untreated, potentially resectable; recruitment 1987‐1988, follow‐up period unclear; China (single‐center) 60 patients (50 male, 10 female) Median follow‐up not given in publication Patients randomized/analyzed in this meta‐analysis: 60/60
Interventions	FPLC (fluorouracil polyphase liposome composita pro orale, consisting of 5‐fluorouracil, oleic acid, ginseng polysaccharides, bean phospholipid and cholesterol) preoperative vs surgery alone Intervention group: 5‐fluorouracil 2000 mg cumulative dose on days 1 to 12; surgery on days 22 to 26 Surgery: no details given
Outcomes	5‐year survival, perioperative morbidity
Notes	Aggregate data from publication
Risk of bias
Bias	Authors' judgement	Support for judgement
Selection bias	Unclear risk	No information on randomization
Attrition bias	Low risk	All randomized patients included in analysis of primary outcome
Performance bias	Unclear risk	Insufficient information given
Detection bias	Unclear risk	Insufficient information given
Bias due to missing data	Low risk	No missing data for primary outcome
Reporting bias	Unclear risk	No definition of predefined endpoints available
Other bias	Low risk	Nothing that would suggest relevant other bias
Overall risk of bias	High risk	High

Zhao 2006.

Methods	RCT
Participants	Malignant gastric neoplasm, Karnofsky’s scale > 90, able to "endure chemotherapy and operation", age ≤ 70 years, 2001‐2005, China (bi‐center) 54 patients (37 male and 17 female) Median follow‐up not given in publication Patients randomized/analyzed in this meta‐analysis: 60/54
Interventions	5’‐deoxy‐5‐fluorouridine (5’‐DFUR) preoperative vs cisplatin/5‐FU preoperative vs surgery alone Group 1: 5’‐deoxy‐5‐fluorouridine (5’‐DFUR) 2400 mg to 6000 mg cumulative dose for 3 to 5 days, surgery on days 6 to 7 Group 2: 5‐fluorouracil 1500 mg to 2500 mg cumulative dose + cisplatin 600 mg to 1000 mg/d cumulative dose for 3 to 5 d, surgery on days 6 to 7 Group 3: immediate surgery Surgery: no details given
Outcomes	OS, presence of tumor‐free resection margin, tumor stage at resection, perioperative morbidity, perioperative mortality
Notes	Aggregate data from publication
Risk of bias
Bias	Authors' judgement	Support for judgement
Selection bias	Unclear risk	No details given on randomization
Attrition bias	High risk	15/60 patients not followed up; no information given on reasons and groups these patients were in
Performance bias	Unclear risk	No details given on operative procedure
Detection bias	Unclear risk	No details on follow‐up given
Bias due to missing data	Unclear risk	No details given on what data analysis of primary outcome is based
Reporting bias	Unclear risk	No definition of predefined endpoints available
Other bias	Low risk	Nothing that would suggest relevant other bias
Overall risk of bias	High risk	High

AEG: adenocarcinoma of the esophagogastric junction 
 CS: cisplatin 
 d: days 
 DFS: disease‐free survival 
 ECOG: Eastern Co‐operative Oncology Group 
 GE: gastroesophageal 
 IPD: individual patient data 
 MTX: methotrexate 
 OS: overall survival 
 RCT: randomized controlled trial 
 UICC: Union for International Cancer Control 
 WHO: World Health Organization 
 wk: week

Characteristics of excluded studies [ordered by study ID]

Study	Reason for exclusion
Bokhyan 2009	Published only as conference abstract, only 1 patient in the trial had adenocarcinoma
Furlong 2010	Duplicate with Walsh 2002
Jin 2008	Full‐text paper available only in Chinese, no translation available, unclear how many patients in the trial had adenocarcinoma
Jin 2011	No full‐text paper available; from the abstract it remained unclear how many patients in the trial had adenocarcinoma
Mariette 2010	Results published after closure of our database
Pokataev 2008	Published only as conference abstract; only 7 patients in the trial had adenocarcinoma
van Hagen 2012	Results published after closure of our database

Differences between protocol and review

There are several differences between the originally published protocol (Slanger 2009) and the final review. The most obvious difference is that the original protocol did not foresee collection and analysis of IPD. Throughout the review of the retrieved papers, we became aware that data for many of our predefined outcomes and subgroups were not readily available in the publications of the respective trials. Most importantly, for several trials which included patients with both squamous cell and adenocarcinoma of the esophagus, we could not extract sufficient data regarding our primary outcome, overall survival. Therefore, we decided to attempt to collect and analyze IPD from all single trials. This also enabled us to foresee additional analyses using a Cox proportional hazards model. While closely examining the provided IPD, we discovered that a valid analysis of disease‐free survival is difficult because the time between randomization and operation differs between trial arms. Therefore, we decided to apply a landmark method for analyzing disease‐free survival. Several smaller modifications regarding 'Risk of bias' assessment and quality control were also done throughout the conduct of the review.

Contributions of authors

The 'GE Adenocarcinoma Meta‐analysis Group' consists of (in alphabetical order): Bryan Burmeister (University of Queensland, Princess Alexandra Hospital, Brisbane, Australia), David Kelsen (Memorial Sloan‐Kettering Cancer Center, New York, NY, USA), Donna Niedzwiecki (Duke University Medical Center, Durham, NC, USA), Christoph Schuhmacher (Klinikum rechts der Isar, Chirurgische Klinik der TU München, München, Germany), Susan Urba (Cancer and Leukemia Group B Statistical Center, Division of Hematology/Oncology, University of Michigan Medical Center, Ann Arbor, MI, USA), Cornelis van de Velde (Department of Surgery, Leiden University Medical Centre, Leiden, The Netherlands), Thomas N Walsh (Royal College of Surgeons, Department of Surgery, Connolly Hospital, Blanchardstown, Dublin, Ireland), Marc Ychou (Centre Régional de Lutte Contre le Cancer, Montpellier, France).

All group members provided IPD from their respective trials and contributed to its correct analysis and interpretation.

All authors conceived of the review. Ulrich Ronellenfitsch (UR) is its guarantor.

UR, Katrin Jensen (KJ) and Tracy E Slanger (TES) designed and wrote the final version of the review which was then approved by all other authors.

TES and UR acted as independent reviewers and data extractors.

KJ and Meinhard Kieser (MK) performed the meta‐analysis, the Cox regression, and all other statistical calculations.

Sources of support

Internal sources

• University Medical Centre Mannheim, University of Heidelberg, Department of Surgery, Germany.

• University Medical Centre Mannheim, University of Heidelberg, 3rd Medical Department, Germany.

• University Hospital Heidelberg, Institute of Medical Biometry and Informatics, Germany.

External sources

• Federal Ministry of Education and Research (BMBF), Germany.

  Research grant within the program 'Clinical Studies/Systematic Reviews'

Declarations of interest

RH has participated in several trials assessing the effect of various chemotherapeutic agents on upper GI tract cancer.

New