Bayesian network meta-analysis and TSA will be carried out in this study.

We registered on the international prospective register of systematic reviews (PROSPERO) to publish our study protocol. The protocol of network meta-analysis is performed according to the preferred reporting items for systematic review and meta-analysis protocol (PRISMA-P) recommendation, and the PRISMA extension statement for reporting of systematic reviews incorporating network meta-analyses of healthcare interventions.21 22

A systematic search will be performed using PubMed, EMBASE, the Cochrane Central Register of Controlled Trials (CENTRAL), Chinese Biomedical Literature Database (CBM) and Chinese Journal Full-text Database. The search strategy will be developed by LG and JHT, who have over 10 years of experience as information specialists. The references of included articles and reviews will be tracked to identify other relevant studies. We will not contact authors for detailed information of primary studies.

Search terms will be: inguinal hernia, groin hernia, inguinal hernioplasty, mesh, random* and others. Full details of the search strategy regarding PubMed, EMBASE and CENTRAL are:

PubMed: ((‘Inguinal hernia[Title/Abstract] OR ‘groin hernia[Title/Abstract] OR ‘inguinal hernioplasty[Title/Abstract] OR ‘Hernia, Inguinal[Mesh]) AND (mesh[Title/Abstract])) AND (Random*[All Fields] OR ‘randomized controlled trial*[All Fields] OR ‘randomized trial*[All Fields] OR Randomized Controlled Trial[ptyp] OR ‘Randomized Controlled Trials as Topic[Mesh])

EMBASE: (‘inguinal hernia’/exp/mj OR ‘inguinal hernia’ OR ‘groin hernia’/exp/mj OR ‘groin hernia’ OR ‘inguinal hernioplasty’) AND (mesh) AND (random* OR[controlled clinical trial]/lim OR(randomized controlled trial]/lim) NOT [MEDLINE]/lim

CENTRAL: #1 ‘Inguinal hernia’ or ‘groin hernia’ or ‘inguinal hernioplasty’:ti, ab, kw (Word variations have been searched)

#2 MeSH descriptor:[Hernia, Inguinal] explode all trees

#3 #1 OR #2

#4 mesh:ti, ab, kw (Word variations have been searched)

#5 random*

#6 #3 AND #4 AND #5.

Type of patients: adults (aged 18 years or older) with inguinal hernia, who scheduled for open inguinal hernia repair. The open mesh repair includes all kinds of techniques, such as Stoppa, Lichtenstein and mesh plug. Patients will be excluded if the hernia is inoperable with open inguinal hernia repair, or if the hernia repair technique changed to another one (eg, laparoscopic access methods).

Type of designs: truly random or quasi-random controlled trials; systematic reviews or meta-analyses will also be included to track their references.

Type of interventions: different mesh fixation methods (or fixation vs no fixation) in open inguinal hernia repair, including, but not limited to, sutures, tacks or staples, self-fixing meshes and fibrin or other glues.

Type of outcomes: the primary outcomes are incidence of chronic groin pain and risk of hernia recurrence. The secondary outcomes include complications, operative time, length of hospital stay and postoperative pain. The end points are defined in table 1.

Other criteria: we will include RCTs reported in the English and Chinese languages. There will be no limitations on year of publication, publication status, duration of study follow-up or period of study conduct.

Literature search records will be imported into ENDNOTE X6 literature management software, while a standard data abstraction form will be created using Microsoft Excel 2013 (Microsoft Corp, Redmond, WA, http://www.microsoft.com) to collect data of interest. A pilot test will be performed for literature selection and data extraction, and a ‘cheat sheet’ with detailed definitions and examples will be developed to ensure high inter-rater reliability among the reviewers.

Two independent reviewers will examine the title and abstract of studies found in the search, to identify related studies according to eligibility criteria. Thus, full-text versions of all potentially relevant studies will be obtained. Excluded trials and the reasons for their exclusion will be listed and examined by a third reviewer.

To extract the data, a rigorous process will be used. First, a draft data extraction form will be conducted. Subsequently, a random sample of five included RCTs will be pilot tested, and the k statistic will be calculated. The form will be revised, as necessary, to confirm the final data extraction form. Finally, two reviewers will independently extract the data of interest, and conflicts will be resolved by a third reviewer.

We will extract all data of interest from each included RCT, including patient characteristics, study characteristics and outcomes. Data extraction item details can be found in table 1. We will consider the following factors as effect modifiers: median age, intervention, size and type of mesh, number and type of fixation, follow-up, and sample size.

The risk of bias of included RCTs will be evaluated according to the Cochrane Handbook V.5.1.0,23 including method of random sequence generation (selection bias), allocation concealment (selection bias), blinding (performance bias and detection bias), incomplete outcome data (detection bias), selective reporting (detection bias) and other bias. We will evaluate methodological quality as low, high or unclear risk of bias. The risk of bias assessment will be completed by two independent reviewers, and conflicts will be resolved by a third reviewer.

We will not contact authors to obtain missing information of primary studies. If binary outcomes are missing, we will perform an available-case analysis, but we will assess the impact of ‘best-best’, ‘best-worst’, ‘worst-best’ and ‘worst-worst’ scenario analyses.24 Regarding the continuous outcomes, we will impute the mean from median and SD for SE, interquartile range, or p values, according to the Cochrane Handbook for Systematic Reviews of Interventions. If such studies are included, we will perform a sensitivity analysis to assess their impact.

We will perform pairwise meta-analysis, using STATA V.12.0 software (Stata Corporation, College Station, Texas, USA). Pooled ORs with 95% CI will be calculated for dichotomous outcomes, mean differences (MDs) with 95% CI for continue outcomes. Heterogeneity of treatment effects across trials will be assessed by c² and I² statistics. If the p value ≥0.1 and I² is ≤50%, it suggests that there is no statistical heterogeneity, and the Mantel-Haenszel fixed effects model will be used for meta-analysis. If the p value <0.1 and I² is >50%, we will explore sources of heterogeneity by subgroup analysis and meta-regression. If there is no clinical heterogeneity, the Mantel-Haenszel random effects model will be used to perform meta-analysis.23 Reporting bias will be examined using the Begg's and Egger's funnel plot method.25 26 In addition, the contour-enhanced funnel plot will be used as an aid to distinguish asymmetry due to publication bias from that due to other factors.27

TSA18 will be performed to reduce the risk of random errors. We will add trials according to the year of publication, and if more than one trial is published in a year, the trials will be alphabetically added according to the last name of the first author.28 TSA will be performed for dichotomous outcomes as well as for continuous outcomes, to control the risks of random errors due to sparse data and multiplicity.29 30 We will also adapt a relative risk reduction of 20%, an α (type I error) of 5%, a β (type II error) of 20% and the diversity of the meta-analysis.19 30

A network plot will be drawn to describe and present the geometry of the treatment network of comparisons across trials to ensure if a network meta-analysis is feasible. Trials will be excluded if the trials are not connected by treatments. Network geometry will use nodes to represent different interventions and edges to represent the head-to-head comparisons between interventions. The size of nodes and thickness of edges are associated with sample sizes of intervention and numbers of included trials, respectively.

A Bayesian network meta-analysis will be performed using WinBUGS 1.4.3 software (MRC Biostatistics Unit, Cambridge, UK). The random and fixed effect models with vague priors for multiarm trials developed by Ade et al31 will be used. The pooled estimation and the probability of which treatment is the best will be obtained using the Markov Chains Monte Carlo method. Three Markov Chains will be run simultaneously with different arbitrarily chosen initial values. We will first generate 50 000 simulations for each chain, and these simulations will then be discarded as the ‘burn-in’ period. Then posterior summaries will be based on 100 000 subsequent simulations. The model convergence will be assessed by trace plots and Brooks-Gelman-Rubin plots.32 The statistical heterogeneity in the entire network will be assessed on the bias of the magnitude of heterogeneity variance parameter (I² or τ²) estimated from the network meta-analysis models using R-3.2.2 software (R Foundation for Statistical Computing, Vienna, Austria). The results of dichotomous outcomes will be reported as posterior medians of OR with 95% credible intervals (CrIs), and medians of MD with 95% CrI for continue outcomes. If a loop connecting three arms exists, inconsistency between direct and indirect comparisons will be evaluated using a node splitting method.33 The choices between fixed and random effect models, consistency and inconsistency models, will be made by comparing the deviance information criteria (DIC) for each model.34 35 The model with the lowest DIC will be preferred (differences >3 are considered significant).

Clinical decisions about the choice of treatments can be recommended based on the probability results of ranking when the differences in effect size of different treatments are small.36 The surface under the cumulative ranking area (SUCRA) will be calculated to summarise and report the probability values. SUCRA values are expressed as percentages—SUCRA value will be 100% for the best treatment, while SUCRA value will be 0% for the worst treatment.37

In order to explore the sources of heterogeneity or inconsistency in the entire network, we will perform network meta-regression or subgroup analysis. Network meta-regression will be conducted using random effects network meta-regression models to examine potential effect moderators such as size of mesh, number of tack, follow-up and sample size.

If we include enough trials per comparison, a sensitivity analysis will be conducted. We will conduct a sensitivity analysis excluding trials that are missing relative data, and we will conduct another sensitivity analysis excluding trials with a total sample size of <50 randomised patients.

The quality of evidence will be classified by the GRADE group into four levels—high quality, moderate quality, low quality and very low quality.38 This process will be performed using GRADE pro 3.6 software (http://www.gradeworkinggroup.org/).

Furthermore, a comparison-adjusted funnel plot will be conducted to identify whether there will be a small sample effect among intervention networks, using STATA V.12.0 software (Stata Corporation, College Station, Texas, USA).

Ethical approval and patient consent are not required since this is a meta-analysis based on published studies.

This protocol has been registered on the international prospective register of systematic reviews (PROSPERO).39 The procedures of network meta-analysis will be conducted according to the PRISMA extension statement for reporting of systematic reviews incorporating network meta-analyses of healthcare interventions. The results of this network meta-analysis and TSA will be submitted to a peer-reviewed journal for publication.