Participants

The median age of the participants in the population of studies included in the analysis of comparisons 1 and 2 was in the range of 56 to 67 years. The proportion of participants with metastatic disease was between 62% (Cullinan 1985) and 100% (Bouche 2004; Koizumi 2008; Yamamura 1998). When comparing the different arms of one study, a difference in the proportion of participants with advanced versus metastatic disease greater than 15% between study arms was identified in only one study (Popov 2002), with a larger number of metastatic participants in the combination chemotherapy arm (90% versus 73%). Performance status was well‐balanced in all studies with no differences greater than 15% between study arms. The percentage of participants with ECOG 2+3 was in the range of 0% to 48%. Thirteen studies, which included 3182 participants (Boku 2009; Hironaka 2016; Koizumi 2008; Koizumi 2014; Komatsu 2011; Lu 2014; Narahara 2011; Nishikawa 2012; Ohtsu 2003; Shirao 2013; Wang 2013; Wu 2015; Yamamura 1998), were conducted in Asia.

Regarding comparisons 3 and4, the median age of participants included in these 11 studies was between 58 and 65 years. Between 46% (Kikuchi 1990) and 90% (Cascinu 2011; Kim 2001) of participants had metastatic disease, the percentage of participants with ECOG 2+3 was between 6% (Cascinu 2011) and 88% (Kikuchi 1990). The percentage of participants with advanced versus metastatic disease was well‐balanced in all studies included in these two comparisons.

The participants in comparison 5 had a median age between 58 (Dank 2008) and 70 (Komatsu 2011) years in the different study arms, with the majority of participants having metastatic disease.

The median age of the participants in comparison 6 was between 55 (Roth 2007; Van Cutsem 2006) and 70 years (Al‐Batran 2013). The percentage of participants with metastatic disease was between 69% (Al‐Batran 2013; Ochenduszko 2015) and 98% (Thuss‐Patience 2005), with the largest study having 97% of participants with metastatic disease and a median age of 55 years (Van Cutsem 2006). Most participants in these studies had a performance status of 0 or 1.

Regarding comparisons 7 and8, the median age of participants included in these studies was between 55 (Ocvirk 2012) and 65 years (Yamada 2015). Between 62% (Kim 2014) and 100% (Li 2016) had metastatic disease and most participants had a performance status of 0 or 1. Only in Popov 2008 the percentage of participants with ECOG 2‐3 was 29%.

The median age of the participants included in comparisons 9 and 10 was between 54 (Ajani 2005; Huang 2013; Li 2015) and 76 years (Boku 2009), between 86% (Roth 2007) and 95% to 100% of participants (Ajani 2005; Ajani 2010; Chen 2015) had metastatic disease. Most participants in these studies had a performance status of 0 or 1.

Groups of participants were well‐balanced regarding the most important prognostic factors as specified above in all studies included into comparisons 3 to 8.

Interventions

Participants were individually randomised into two or more treatment groups. The effect of the intervention was measured and analysed on the basis of single measurements for each outcome for each participant. In 11 studies, more than two groups with different interventions were compared (Boku 2009; Bouche 2004; Cullinan 1985; Cullinan 1994; Hironaka 2016; Loehrer 1994; Lutz 2007; Nishikawa 2012; Ohtsu 2003; Roth 2007).

In the studies by Bouche 2004, Lutz 2007, Ohtsu 2003, and Hironaka 2016 the single‐agent therapy arm was compared with the pooled results of both combination chemotherapy arms. In Loehrer 1994, Cullinan 1994, and Boku 2009 , the combination‐therapy arm was compared with the pooled results of both single‐agent arms. In Nishikawa 2012, the pooled results of the sequential therapy arms were compared with the pooled results of the concurrent therapy arms. All these studies were included in the comparison of single‐agent versus combination chemotherapy.

Furthermore, the irinotecan‐containing combination chemotherapy was compared to the pooled results of the non‐irinotecan‐containing combination chemotherapies and included in comparison 5 in 10 studies (Boku 2009; Bouche 2004; Dank 2008; Komatsu 2011; Li 2016; Moehler 2005; Moehler 2010; Narahara 2011; Roy 2012; Sugimoto 2014), and the pooled results of the docetaxel‐containing chemotherapies were compared with the non‐docetaxel‐containing combination chemotherapy and included in comparison 6 in eight studies (Al‐Batran 2013; Koizumi 2014; Ochenduszko 2015; Ridwelski 2008; Roth 2007; Thuss‐Patience 2005; Van Cutsem 2006; Wang 2016).

All three studies included in the analysis of comparison 1 used combination therapy regimens in the chemotherapy arm (Murad 1993: FAMTX; Pyrhönen 1995: FEMTX; Scheithauer 1996: 5‐FU/LV/Epirubicin).

Regarding comparison 2, most studies used 5‐FU in the single‐agent arm. In six of 17 studies (Barone 1998; Colucci 1995; Cullinan 1985; Cullinan 1994; De Lisi 1986; Loehrer 1994), 5‐FU was given as a bolus in doses of approximately 500 mg/m² days one to five every four weeks. A continuous infusion regimen was used in two studies, with either 2600 mg/m² every two weeks (Popov 2002), 800 mg/m² per day (Ohtsu 2003) on days one to three every four weeks or a bolus of 400 mg/m² 5‐FU, followed by 600 mg/m² as a two‐hour continuous infusion on days one and two every two weeks (Bouche 2004). One study (Levi 1986), applied doxorubicin 60 mg/m² every four weeks in the single‐agent arm. In the study by Loehrer 1994, the results from two single‐agent arms (5‐FU bolus 500 mg/m² days one to five and epirubicin 90 mg/m² day one every four weeks) were combined in the analysis. In addition, in the studies by Bouche 2004, Lutz 2007, Nishikawa 2012, and Ohtsu 2003 , the results of two combination therapy arms (LV5FU2/cisplatin and LV5FU2/irinotecan, D‐FU/FA and HD‐FU/FA/cisplatin, 5‐FU/paclitaxel and S‐1/paclitaxel, 5‐FU/cisplatin, and tegafur/mitomycin C) were combined in the analysis. Nishikawa 2012 used 5‐FU in doses of 800 mg/m² in days one to five every four weeks in one group or daily S‐1 in doses of 80 mg/m² for four weeks and a two‐week rest in the single‐agent arm until progression. This therapy was followed by paclitaxel (80 mg/m² on days one, eight and 15 every four weeks). A similar regimen of S‐1 was used in Narahara 2011 and Komatsu 2011. Boku 2009 used the same 5‐FU regimen in one arm or lower doses of S‐1 (40 mg/m² for four weeks and a two‐week rest) in the single‐agent arm. Results of two arms were combined in the single‐agent arm in Boku 2009; Cullinan 1994 and Nishikawa 2012. Wang 2013 used S‐1 according to body‐surface area with 40 mg twice daily for participants with a body surface area greater than 1.25 and lower than 1.5 for two out of four weeks. Koizumi 2008 and Hironaka 2016 as well used the oral prodrug S‐1.

In six of 23 studies, combination chemotherapy arms did contain an anthracycline (epirubicin 60 mg/m² every three weeks or 90 mg/m² every four weeks or doxorubicin 40 mg/m² every four to seven weeks) in addition to 5‐FU. Non 5‐FU‐based combination chemotherapy regimens, which instead included etoposide, irinotecan and cisplatin, S‐1 and irinotecan, S‐1 and paclitaxel or an anthracycline and cisplatin were applied in six studies (Barone 1998; Boku 2009; Komatsu 2011; Narahara 2011; Popov 2002; Wang 2013).

In comparisons 3 and4, regimens containing 5‐FU, an anthracycline and cisplatin were mainly FAP (5‐FU bolus 300 mg/m² either days one, eight, 15, 22 or days one to five, adriamycin 25 to 40 mg/m² and cisplatin 60 to 100 mg/m² once every three to five weeks) (Cullinan 1994; GITSG 1988; Kikuchi 1990), and PELF (cisplatin 40 mg/m² days one and five, epirubicin 30 mg/m² days one and five, leucovorin 200 mg/m² and 5‐FU bolus 300 mg/m² days one to four every eight weeks) (Cocconi 1994) in studies published before 1995 (Kim 2001; Ross 2002; Webb 1997). used mostly ECF (epirubicin 50 mg/m² and cisplatin 60 mg/m² once every three weeks, with 5‐FU as a continuous intravenous infusion of 200 mg/m² for up to six months) or LdCF (pegylated liposomal doxorubicin 20 mg/m² and cisplatin 50 mg/m² once every two weeks, with 5‐FU 400 mg/m² bolus followed by 600 mg/m² as 22‐hour continuous infusion on days one and two every two weeks) in Cascinu 2011.

In comparison 5, irinotecan was given in the following studies to substitute either cisplatin (Bouche 2004; Dank 2008; Moehler 2010), etoposide (Moehler 2005) or 5‐FU (Roy 2012) or paclitaxel (Sugimoto 2014). In contrast, irinotecan was given in addition to the treatment in the other arm in studies by Bouche 2004; Komatsu 2011; and Narahara 2011; or as other comparisons (Li 2016). Irinotecan was given weekly at 80 mg/m² for six weeks every 50 days in combination with infusional 5‐FU 2000 mg/m² and FA 500 mg/m² (Dank 2008; Moehler 2005). Moehler 2010 used irinotecan 250 mg/m² on day one in combination with capecitabine 1000 mg/m² orally days one to 14 every 22 days and Roy 2012 used irinotecan 250 mg/m² on day one in combination with docetaxel 60 mg/m² as infusion on day one every 22 days. Bouche 2004 used irinotecan at 180 mg/m² in combination with infusional 5‐FU every two weeks. Komatsu 2011 and Narahara 2011 used doses of 75 mg/m² and 80 mg/m² on days one and 15 every four weeks or six weeks in combination with oral S‐1 (initial doses of 40 mg/m² to 60 mg/m² twice daily on days one to 14 every four weeks or 80 mg/m²/day on days one to 21 every six weeks). In subsequent cycles, doses were varied according to the most severe adverse events during the preceding cycle.

In comparison 6, docetaxel was given to substitute either epirubicin and cisplatin (Roth 2007; Thuss‐Patience 2005) or 5‐FU and leucovorin (Ridwelski 2008). On the other hand, it was given in addition to the treatment in the other study arm in the studies by Wang 2016, Koizumi 2014, Van Cutsem 2006 and Al‐Batran 2013. The largest studies included in comparison 6's meta‐analysis of overall survival used the three‐drug regimen DCF (docetaxel 75 mg/m² intravenously day one, cisplatin 75 mg/m² intravenously day one, 5‐FU 750 mg/m² as a 24‐hour infusion) on days one to five every three weeks (Van Cutsem 2006), and the two‐drug regimen of docetaxel 75 mg/m² intravenously day one, in combination with cisplatin 75 mg/m² intravenously on day one every three weeks (Ridwelski 2008). In Koizumi 2014, docetaxel (40mg/m² intravenously on day one) was given with S‐1 (tailored to body surface area; days one to 14) every 21 days. In Roth 2007, the DCF regimen was used as described previously. Sadighi 2006 and colleagues used a modification of DCF with reduced doses of docetaxel and cisplatin (both at 60 mg/m²) every three weeks. Thuss‐Patience 2005 applied docetaxel 75 mg/mg/m² intravenously on day one in combination with 5‐FU 200 mg/m² /day over 24 hours on days one to 21 every three weeks. Al‐Batran 2013 used docetaxel 50 mg/m² intravenously on day one in combination with oxaliplatin 85 mg/m² and leucovorin 200 mg/m² followed by 5‐FU 2600 mg/m² as a 24‐hour continuous infusion.

Five studies (Kang 2009; Li 2016; Ocvirk 2012; Ochenduszko 2015; Van Cutsem 2015) are eligible for comparison 7. In Kang 2009 and Ocvirk 2012, the oral 5‐FU prodrug capecitabine (1000 mg/m² or 825 mg/m² twice daily on days one to 14 of a 21‐day regimen) was compared with 5‐FU (both in combination with cisplatin). In comparison 8, oxaliplatin was given once at 85 mg/m² in two weeks and compared with cisplatin 50 mg/m² in Al Batran 2008 and Popov 2008. Both agents were combined with FU/ leucovorin in Al Batran 2008 and Popov 2008. In Kim 2014 a combination of weekly docetaxel (35mg/m²) on days one and eight every three weeks, in combination with either cisplatin (60 mg/m²) or oxaliplatin (120 mg/m²) on day one was administered. In Yamada 2015, S‐1 (twice daily for the first three weeks of a five‐week cycle) plus (cisplatin 60 mg/m² on day eight) was compared against S‐1 (twice daily for first two weeks of a three‐week cycle) plus oxaliplatin (100 mg/m² infused for two hours on day one). For comparison 9, three studies (Ajani 2005; Roth 2007; Van Cutsem 2015) are available. Fluorouracil 200 mg/m²/day was given as a 24‐hour continuous infusion in or in doses of 750 mg/m²/day on days one to five every three weeks in Roth 2007. It was combined with docetaxel and cisplatin in Ajani 2005 and Roth 2007, and docetaxel and oxaliplatin in Van Cutsem 2015.

In comparison 10, S‐1 50 mg/m² was given orally in two daily doses on days one to 21 of a four‐week cycle (Ajani 2010), or in a lower dosage of 40 mg/m² orally in two daily doses on days one to 28 of a six‐week cycle (Boku 2009) and compared with continuous infusions of 5‐FU in doses of 1000 mg/m²/24 hours as 120‐hour infusion (Ajani 2010) or 800 mg/m²/day on days one to five, respectively. Of note, S‐1 was combined with 75 mg/m² of cisplatin every three weeks, and 5‐FU with 100 mg/m² of cisplatin every three weeks in the study by Ajani 2010. Huang 2013 compared the combination of weekly paclitaxel (60 mg/m²) on days one eight, and 15 every four‐week cycle and S‐1 (80 mg/m² to 120 mg/m²), dependent on the body‐surface area for two out of four weeks, with the same regimen of paclitaxel, but with 5‐FU (500 mg/m², in combination with leucovorin 20 mg/m² on days one to five every four weeks).

Outcomes

Median survival and response rates were the outcomes most commonly described in the included studies. The newer studies reported progression‐free survival instead of time to progression. Toxicity was not always classified according to WHO or NCI‐CTC and was reported in different ways (per number of participants, per number of cycles and only the worst toxicity per participant). A comparison across studies was therefore not possible. For this reason, the numbers of participants who discontinued treatment due to toxicity as well as the numbers of treatment‐related deaths were analysed. Information about second‐line therapies and secondary resectability was unavailable for most studies. Details are listed in the Characteristics of included studies table.

Overall survival

A total of three studies (N = 184) reported overall survival (Murad 1993; Pyrhönen 1995; Scheithauer 1996). The overall hazard ratio (HR) of 0.37 (95% confidence interval (CI) 0.24 to 0.55, moderate‐quality evidence) (Analysis 1.1) in favour of the chemotherapy arms demonstrates a convincing benefit over best supportive care (BSC) alone, which can be interpreted as an improvement in median survival from 4.3 months (weighted average in BSC) to 11 months (with chemotherapy). Cochrane's Q test (P = 0.19) as well as the index of homogeneity according to Thompson (I² = 39.7) demonstrate a significant statistical heterogeneity among the results of these studies, based on the differences in the chemotherapy regimens studied. A sensitivity analysis including only studies with adequate allocation concealment (Pyrhönen 1995; Scheithauer 1996), does not change the overall HR of 0.37 (95% CI 0.19 to 0.70).

Secondary outcomes

Overall survival

Twenty‐three studies including 4447 participants were summarised in this meta‐analysis (Barone 1998; Boku 2009; Bouche 2004; Colucci 1995; Cullinan 1985; Cullinan 1994; De Lisi 1986; Hironaka 2016; Koizumi 2008; Koizumi 2014; Komatsu 2011; Levi 1986; Loehrer 1994; Lu 2014; Lutz 2007; Narahara 2011; Nishikawa 2012; Ohtsu 2003; Popov 2002; Shirao 2013; Wang 2013; Wu 2015; Yamamura 1998). The overall HR in favour of combination chemotherapy (HR 0.84, 95% CI 0.79 to 0.89, moderate quality) provides evidence for a statistically significant, but modest survival benefit of combination versus single‐agent chemotherapy in the studied regimens (Analysis 2.1). Cochrane's Q test for heterogeneity showed non‐significant heterogeneity (I² = 0%, P = 0.48), indicating that results of the different studies were consistent in their findings. As the chemotherapy regimens in studies published before 2000 might not have the same efficacy as modern regimens, the median survival difference between single and combination chemotherapy was calculated separately for studies published before the year 2000 and thereafter. For studies published before 2000, the weighted median survival was approximately 7.3 with combination therapy and 6.4 months with single‐agent therapy. In studies published after 2000 (Boku 2009; Bouche 2004; Hironaka 2016; Koizumi 2008; Koizumi 2014; Komatsu 2011; Lu 2014; Lutz 2007; Narahara 2011; Nishikawa 2012; Ohtsu 2003; Popov 2002; Wang 2013, Wu 2015), median survival was 11.6 months with combination therapy, as compared to 10.5 months with single‐agent therapy.

To evaluate the influence of second‐line therapy, which was previously specified as a possible cause of heterogeneity, a second analysis was performed excluding the studies by Boku 2009, Koizumi 2008, Koizumi 2014, Narahara 2011, Ohtsu 2003, and Wang 2013 in which more than 50% of participants received a second‐line therapy. Exclusion of theses studies had no influence on heterogeneity (I² = 0% and P = 0.56) and the overall HR of 0.82 (95% CI 0.75 to 0.90) in favour of combination chemotherapy. Sensitivity analysis excluding those studies which were conducted in Asia (Boku 2009; Hironaka 2016; Koizumi 2008; Koizumi 2014 ; Komatsu 2011; Lu 2014; Narahara 2011; Nishikawa 2012; Ohtsu 2003; Shirao 2013; Wang 2013; Wu 2015; Yamamura 1998) resulted in a HR of 0.77 (95% CI 0.68 to 0.87) (I² = 0%, P = 0.57), with no appreciable change in heterogeneity as compared to the original analysis. When only those studies with adequate allocation concealment were included in the analysis (Barone 1998; Boku 2009; Bouche 2004; De Lisi 1986; Koizumi 2008; Levi 1986; Loehrer 1994; Lu 2014; Lutz 2007; Narahara 2011; Nishikawa 2012; Ohtsu 2003; Shirao 2013; Wang 2013; Wu 2015; Yamamura 1998), the resulting overall HR was 0.83 (95% CI 0.77 to 0.89) (I² = 24% and P = 0.19). For these reasons, the results of this comparison can be considered to be highly robust.

Secondary outcome measures

Overall survival

This meta‐analysis was based on 579 participants in four randomised studies (Cascinu 2011; Kim 2001; KRGGC 1992; Ross 2002). The data from the largest study in this comparison (published by Ross 2002) which was included in this analysis were provided by the authors and include people with gastric and GE‐junction adenocarcinoma only (the original publication included people with squamous cell carcinoma of the oesophagus as well). The resulting HR for overall survival of 0.74 (95% CI 0.61 to 0.89, moderate‐quality evidence) demonstrates a statistically significant benefit with very low heterogeneity (I² = 0%, P = 0.63) in overall survival in favour of the three‐drug combination, with a weighted average survival of 9.9 and 8.6 months, respectively. Allocation concealment was adequate in all three studies included in this comparison and heterogeneity was non‐significant (I² = 0%; P = 0.63) (Analysis 3.1).

Overall survival

Ten studies (N = 2135) were summarised in this meta‐analysis (Boku 2009; Bouche 2004; Dank 2008; Komatsu 2011; Li 2016; Moehler 2005; Moehler 2010; Narahara 2011; Roy 2012; Sugimoto 2014). To avoid double‐counting the irinotecan‐treated population in the study by Bouche 2004, we approximated a within‐study hazard ratio of 0.72 (95% CI 0.30 to 1.72) for the study's irinotecan‐treated population (N = 45) and non‐irinotecan‐treated population (N = 89). Overall, the pooled hazard ratio of irinotecan‐containing regimens compared to non‐irinotecan containing regimens was 0.87 (95% CI 0.80 to 0.95, high‐quality evidence), with low heterogeneity (I² = 0%, P = 0.86) (Analysis 5.1), and corresponds to pooled median survival times of 11.3 and 9.7 months, with a small, but significant benefit for the irinotecan‐containing regimens. When only those studies with information about adequate allocation concealment are included in the analysis (Boku 2009; Bouche 2004; Moehler 2005; Moehler 2010; Narahara 2011; Roy 2012) the resulting overall HR was 0.84 (95% CI 0.76 to 0.93) (P = 0.84 and I² = 0%) when all studies were pooled. For these reasons, the results of this comparison can be considered as robust. Further sensitivity analyses were not performed due to the low number of studies.

Next, the HR for overall survival was separately investigated for studies with substitutive (i.e. studies in which another chemotherapy was substituted by irinotecan), additive (i.e. studies in which irinotecan was added to other chemotherapies), and other comparisons of irinotecan and non‐irinotecan‐containing regimens (Analysis 5.1). The summary of the six studies (826 participants) with substitutive comparisons (Bouche 2004; Dank 2008; Moehler 2005; Moehler 2010; Roy 2012; Sugimoto 2014) results in a HR of 0.87 (95% CI 0.75 to 1.00, high‐quality evidence), with low heterogeneity between study results (I² = 0%, P = 0.94).

The meta‐analysis of the three studies (Bouche 2004; Komatsu 2011; Narahara 2011), including a total of 500 participants, where irinotecan was given in addition to the treatment in the non‐irinotecan‐containing arm shows a non‐significant benefit for participants treated with irinotecan (HR 0.88, 95% CI 0.76 to 1.03, low‐quality evidence) and small heterogeneity between treatment effects of individual studies (I² = 6%, P = 0.34). This result corresponds to a pooled median survival duration of 11.9 months with compared to 10.9 months without irinotecan. Bouche 2004 and Narahara 2011 stated a higher benefit of four and two months, respectively. However, Komatsu 2011 showed a disadvantage of three months for participants treated with irinotecan and S‐1 compared to participants with S‐1 monotherapy.

The meta‐analysis of two studies (Boku 2009; Li 2016), which could neither be classified as substitutive nor additive, including a total of 809 participants, revealed a slight benefit for participants treated with irinotecan (HR 0.87, 95% CI 0.76 to 1.00, very low‐quality evidence). However, the pooled result needs to be interpreted with caution in light of the considerable clinical differences between studies included under 'Other Comparisons' as well as their significant statistical heterogeneity (I² = 65%, P = 0.04). For instance, Li 2016 allowed participants to switch to second‐line therapy after failure of first‐line. The study by Boku 2009 alone, which contributed the majority of participants (n = 704), showed a HR of 0.84, 95% CI 0.73 to 0.95 and a corresponding benefit in median survival of around one month in favour of participants treated with irinotecan.

Secondary outcome measures

Overall survival

Eight studies (N = 2001) were summarised in this meta‐analysis (Al‐Batran 2013; Koizumi 2014; Ochenduszko 2015; Ridwelski 2008; Roth 2007; Thuss‐Patience 2005; Van Cutsem 2006; Wang 2016). The resulting HR for overall survival was estimated separately for studies with substitutive (i.e. studies in which another chemotherapy was substituted by docetaxel), additive (i.e. studies in which docetaxel was added to other chemotherapies), and other comparisons (Analysis 6.1).

The summary of the three studies (479 participants) with substitutive comparisons (Ridwelski 2008; Roth 2007; Thuss‐Patience 2005) slightly favours the non‐docetaxel‐containing regimens (HR 1.05, 95% CI 0.87 to 1.27, moderate‐quality evidence), but does not reach statistical significance. The index of homogeneity according to Thompson (I² = 0%) shows a low statistical heterogeneity among the results of these studies (Analysis 6.1). Because necessary data for calculation of the HR in the studies by Sadighi 2006 were missing, this study could not be included in the meta‐analysis of this comparison. Differences in pooled median survival between the docetaxel‐containing regimens (9.2 months) and the non‐docetaxel‐containing regimens (9.4 months) are neither statistically significant nor clinically relevant. All publications describe an adequate allocation concealment and all studies were conducted in Europe.

The meta‐analysis of the four studies (N = 1466) (Al‐Batran 2013; Koizumi 2014; Van Cutsem 2006; Wang 2016), including a total of 1466 participants, where docetaxel was given in addition to the treatment in the non‐docetaxel‐containing arm favours the docetaxel‐containing regimens (HR 0.80, 95% CI 0.71 to 0.91, moderate‐quality evidence) with little heterogeneity between treatment effects of individual studies (I² = 0%, P = 0.82). This result corresponds to pooled median survival times of 12.3 and 10.6 months, with a small benefit for the docetaxel‐containing regimens. Al‐Batran 2013 was conducted in Germany, Koizumi 2014 in Japan and Korea, Wang 2016 in China, and the fourth study (Van Cutsem 2006) was an international study that recruited participants in America, Europe, and Asia. Due to the small number of studies in the primary analysis, we did not perform sensitivity analyses.

The single study belonging to other comparisons (Ochenduszko 2015) with 56 participants demonstrated a non‐statistically significant (P = 0.43) advantage of docetaxel‐containing regimens compared to control (HR 0.80, 95% CI 0.46 to 1.39, very low‐quality evidence). This study permitted second‐line therapy with irinotecan monotherapy.

Secondary outcome measures

Overall survival

The results for this comparison are based on 732 randomised participants in five studies (Kang 2009; Li 2016; Ocvirk 2012; Ochenduszko 2015; Van Cutsem 2015). The HR for overall survival of 0.94 (95% CI 0.79 to 1.11, moderate‐quality evidence) favours the oral regimen, but does not reach statistical significance (Analysis 7.1). The index of heterogeneity for this comparison was small (I² = 12%, P = 0.34). The corresponding pooled median survival are 10.8 and 10.9 months, respectively for the capecitabine and 5‐FU‐containing arms respectively. When the studies which permitted second‐line therapy were excluded as part of sensitivity analysis (Li 2016; Ochenduszko 2015), the pooled HR for overall survival remained stable at 0.93 (95% CI 0.77 to 1.14).

Secondary outcome measures

Overall survival

Data were available on 1105 participants in five studies (Al Batran 2008; Hironaka 2016; Kim 2014; Popov 2008; Yamada 2015). The HR for overall survival of 0.81 (95% CI 0.67 to 0.98, low‐quality evidence) show a small, advantage for the regimen with oxaliplatin (Analysis 8.1), with a moderate heterogeneity index (I² = 38%, P = 0.17). This survival benefit is also reflected in the pooled median overall survival time of 14.0 months versus 11.3 months, respectively.

Secondary outcome measures

Overall survival

The results for this comparison are based on 482 randomised participants in three studies (Ajani 2005; Roth 2007; Van Cutsem 2015). The HR for overall survival of 0.86 (95% CI 0.71 to 1.06, very low‐quality evidence) favours the regimen with fluoropyrimidine (I² = 0%, P = 0.47), but does not reach statistical significance (Analysis 9.1). The corresponding pooled median survival times are 11.7 versus 10.0 months with a small benefit for participants treated with fluoropyrimidines. Allocation concealment was adequate in the studies included in this comparison.

Secondary outcome measures

Overall survival

This meta‐analysis was based on 1793 participants in four randomised studies (Ajani 2010; Boku 2009; Chen 2015; Li 2015). The resulting HR for overall survival of 0.91 (95% CI 0.83 to 1.00, high‐quality evidence) demonstrates a borderline statistically significant benefit with low heterogeneity (I² = 0%, P = 0.50) with overall survival in favour of S‐1 (Analysis 10.1). The corresponding pooled median survival times were 9.6 and 9.1 months with a clinically negligible benefit for participants treated with S‐1. Allocation concealment was adequate in Ajani 2010 and Boku 2009.

Secondary outcome measures