We will conduct a systematic review with meta-analysis according to reporting standards [
43]. Literature searchers, identification of eligible studies, data extraction, and bias assessment will be undertaken independently by at least two researchers.
PICO tables and eligibility criteria
The PICO criteria were agreed by the review researchers and defined as follows. Participants of eligible trials and studies will be adults and children with incidence of end-stage kidney disease starting long-term dialysis treatment. The intervention will consider any type of PD and its variants (continuous ambulatory PD (CAPD), automated PD (APD)/continuous cyclic PD (CCPD), (nocturnal) intermittent PD (IPD/NIPD), tidal PD (TPD), or continuous flow PD (CFPD)). The comparison will consider HD and its variants (hemofiltration, hemodiafiltration, acid-free biofiltration). We will exclude studies evaluating combined HD and PD strategies, or where the hemodialysis comprises intensive dialysis (i.e., greater than 3.5 times per week, or greater than 6 h per treatment [
44]). The primary outcome will be death from any cause.
Eligible studies and trials will include published or unpublished reports in any language that assess associations between PD and HD with the outcome of interest. We will include randomized controlled trials and quasi-RCTs and prospectively or retrospectively recruited longitudinal cohort studies. We will exclude studies published before 1995. Narrative reviews and health technology assessments related to the topic will be retained to investigate their references for further eligible studies.
Literature search
We will search manually for additional studies by cross-checking the reference lists of all included primary studies and lists of relevant systematic reviews. In addition, study authors and experts will be contacted for additional studies. The search strategy will be developed by the research team in collaboration with an experienced librarian and checked by a referee according to the Peer Review of Electronic Search Strategies (PRESS) guidelines. The search strategy is shown in
Appendix. Search results will be managed using Endnote (Clarivate Analytics, Philadelphia, PA).
Study selection
The title and abstract of each article will be screened and assessed against predefined inclusion criteria by two independent reviewers. Full texts of all potentially relevant articles will then be assessed for inclusion by two reviewers independently. Disagreements will be resolved through discussion and consensus or consulting a third person. The corresponding authors of eligible articles will be contacted for clarification where necessary. We will record the reasons for exclusion and report the study selection process using the PRISMA flow diagram. A list of excluded studies will be provided.
A standardized data extraction sheet will be designed and tested. Two reviewers will independently extract data from the included studies. Any disagreements will be resolved through discussion and consensus or by involving a third reviewer. Where necessary, studies will be translated before assessment and data extraction.
The following data will be extracted: study characteristics (design, sample size, duration of follow-up, number of participants randomized/included in the analysis); participant characteristics—demographics (age, sex), relevant medical conditions, and cause of ESKD; presence and extent of adjustment for co-variates (age, sex, diabetes mellitus); sub-modality of PD; and sub-modality of HD death from any-cause.
In case outcome data are missing, we will contact study authors and request the data. For prospective and retrospective studies, the most adjusted values for effect size will be extracted.
Risk of bias assessment
We will use the Cochrane tool to assess study risk of bias in randomized and quasi-randomized trials. For each assessment, we will provide support for judgment. For non-randomized studies, the Newcastle-Ottawa Scale will be used. Items will be rated as low, high, or unclear risk of bias. The following domains will be assessed: representativeness of exposed cohort, ascertainment of exposure, statistical methods, outcomes of interest defined a priori (outcomes reporting bias), assessment of outcomes, and follow-up times for outcomes and attrition [
45]. Any disagreements will be resolved through discussion and consensus. If necessary, we will involve a third reviewer.
Data analysis
For prospective and retrospective cohort studies, we will summarize the adjusted risk ratios (relative risk, hazard ratio, odds ratio) for PD versus HD as reported by the studies or calculated for dichotomous outcomes using DerSimonian and Laird random effects meta-analysis. A summary risk estimate will be reported together with a 95% confidence interval. When individual studies report results separately for multiple subgroups of patients, we will extract results for each cohort to include in the meta-analysis. The results for each cohort within a study will be combined using fixed effect meta-analysis before being entered into the overall meta-analytical model. Results for observational studies and trials will be summarized separately.
Clinical and statistical heterogeneity between studies will be assessed by two reviewers. We will evaluate for heterogeneity using the I2 statistic and consider the I2 thresholds of < 25%, 25–49%, 50–75% and > 75% to represent low, moderate, high, and very high heterogeneity, respectively. Given the likelihood of clinical or statistical heterogeneity, we will apply a random-effect model. Analyses will be conducted using Stata IC 14/15 (Statacorp, College Station, TX).
Potential sources of statistical heterogeneity will be evaluated through subgroup analyses. If possible, we will undertake subgroup analyses according to age (children, adults), duration of follow-up (6 months, 1 year, 2 years), era of study (> 2000, 2000–2010, > 2010), and the type of country of study according to its economy and capital markets (advanced, developing [
46]). Effect modification by age, gender, and diabetes will be ascertained by meta-regression of study-level summary data and, depending on those results, explored in subgroups according to cut points suggested by the visual inspections of fitted models. Where possible, we will conduct the following analyses to determine if results are sensitive to the influence of fixed-effect model versus random-effect model assumptions; the inclusion of studies at high risk of bias (the overall risk will be considered high if any of the domains of the Cochrane Risk of Bias tool are judged to be at high risk of bias for RCTs and if the comparability of cohorts is not enhanced by design or analyses that adjust, stratify, or match for age and diabetes); the inclusion of publications that include deaths up to 90 days (including the interim or short-term HD patients who have very high mortality due to elements unrelated to dialysis); and studies using an as-treated framework (“did the exposure that the patient actually receive affect mortality?”) (e.g., [
47‐
50]), as opposed to an intention-to-treat framework (“did exposure that the patient initially receive affect mortality, irrespective of subsequent changes that occurred along the way?”) [
51].
Level of evidence
The confidence that may be placed in the summary estimates will be evaluated using the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) tool [
52]. The following domains will be considered: risk of bias/study limitations, directness of evidence (generalizability), consistency of prognostic estimates among studies, and precision (width of confidence interval and impact on clinical significance). The quality of the body of evidence will be assessed by two reviewers independently. The GRADE system specifies four levels of certainty, namely, high quality (where further research is very unlikely to change our confidence in the estimates of effect), moderate (where further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate), low quality (where 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), and very low quality (where any estimate of effect is very uncertain) evidence.