ReviewReproducible research practices are underused in systematic reviews of biomedical interventions
Introduction
Biomedical researchers are increasingly encouraged to use reproducible research practices, which allow others to recreate the findings of studies, given the original data [1], [2], [3]. Such practices include providing a detailed description of the data collected and used for analysis, along with descriptive metadata, clearly reporting the analysis methods and results, and sharing the data set and statistical code used to perform analyses [1], [4]. There are many benefits of performing such practices in the context of systematic reviews (SRs) of studies. For example, users can check for possible data-entry errors when summary statistics for each included study are reported in sufficient detail. Transparent reporting of meta-analyses also makes it possible for others to reanalyze published meta-analyses using different inclusion criteria or statistical methods or to perform additional analyses that address secondary research questions [5]. For example, readers may reanalyze a published meta-analysis by restricting it to the subset of studies conducted in the setting where they work. In addition, sharing of data sets and statistical analysis code allows other researchers to cumulatively add new data that are published, thus keeping meta-analytic effect estimates up-to-date [6], [7].
The limited data on use of reproducible research practices in SRs come from studies that have recorded how well SRs adhere to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement. The PRISMA statement includes an item recommending that for all outcomes considered, systematic reviewers report, for each study, “simple summary data for each intervention group and effect estimates and confidence intervals” [8]. However, not all studies evaluating PRISMA adherence have provided data on adherence to this item, opting to present a total score summed across all PRISMA items instead [9]. Furthermore, many studies that have identified low adherence to this item have assessed SRs in a single clinical specialty (e.g., [10], [11], [12], [13], [14]), which limits generalizability of the findings. To our knowledge, no study has quantified how often systematic reviewers report the data needed to recreate all meta-analytic effect estimates in an SR (including subgroup meta-analytic effects and sensitivity analyses) nor investigated whether completeness of reporting varies by type of outcome (i.e., primary or other).
Efforts to increase transparent reporting of SR articles have existed for many years (e.g., the PRISMA statement was disseminated in top medical journals in 2009); however, little attention has been given to the sharing of data collected as part of SRs [15]. For example, since 2015, the BMJ encourages authors of all research articles to link their articles to the raw data from their studies but requires data sharing on request as a minimum for clinical trials only [16]. No study has investigated how often sharing of data sets and statistical analysis code is done by authors of SRs.
We investigated how often research practices that facilitate reproducibility of analyses were used in a cross-sectional sample of SRs of therapeutic interventions. We also explored whether the use of such reproducible research practices was associated with whether an SR was a Cochrane review and with the systematic reviewers' reported use of the PRISMA statement.
Section snippets
Methods
We conducted this project in accordance with a study protocol, which is available on the Open Science Framework (RRID:SCR_003238): https://osf.io/523bq/. This study was conducted concurrently with another project evaluating the application and interpretation of statistical methods in SRs. The results of the other project will be described elsewhere.
General characteristics of SRs
We evaluated 110 SRs of therapeutic interventions. Nearly all (97/110 [88%]) were published in late 2013, and 68/110 (62%) were led by systematic reviewers based in China, UK, Canada, or USA (Table 1). Most SRs (78/110 [71%]) were not Cochrane reviews. The SRs focused on interventions for one of 19 different conditions, with diseases of the digestive system, diseases of the circulatory system, infectious and parasitic diseases, and neoplasms being the most common. The interventions studied were
Discussion
The use of reproducible research practices in SRs of therapeutic interventions was suboptimal in our sample. Systematic reviewers reported the data needed to recreate all meta-analytic effect estimates in the SR, including subgroup meta-analytic effects and sensitivity analyses, in only 65% of SRs. This percentage was higher in Cochrane than in non-Cochrane SRs (94% vs. 54%). In contrast, the data needed to recreate the index (primary or first reported) meta-analysis were available in nearly
Conclusions
Reproducible research practices in SRs of therapeutic interventions are suboptimal. Strategies are needed to facilitate the provision of detailed descriptions of data gathered and data used for analysis, transparent reporting of the analysis method and results, and sharing of data sets and statistical analysis code so that others can recreate the findings or perform secondary analyses.
Acknowledgments
The authors thank Sean Harrison (University of Bristol) for assistance with data analysis.
Authors' contributions: All authors declare to meet the ICMJE conditions for authorship. M.J.P., D.G.A., and D.M. conceived the study design. L.S. and J.E.M. provided input into the study design. M.J.P., D.G.A., L.S., J.E.M., and D.M. selected items for inclusion in the data-collection form. M.J.P., N.A., D.W., and F.Y. collected data. M.J.P. undertook the statistical analyses. M.J.P. wrote the first draft
References (39)
- et al.
Meta-analysis in clinical trials
Control Clin Trials
(1986) - et al.
Impact of an online writing aid tool for writing a randomized trial report: the COBWEB (Consort-based WEB tool) randomized controlled trial
BMC Med
(2015) - et al.
Disclosure of data and statistical commands should accompany completely reported studies
J Clin Epidemiol
(2016) - et al.
Reproducible research practices and transparency across the biomedical literature
PLoS Biol
(2016) - et al.
Scientific standards. Promoting an open research culture
Science
(2015) - et al.
An open investigation of the reproducibility of cancer biology research
Elife
(2014) - et al.
What does research reproducibility mean?
Sci Transl Med
(2016) - et al.
On the reproducibility of meta-analyses: six practical recommendations
BMC Psychol
(2016) - et al.
Living systematic reviews: an emerging opportunity to narrow the evidence-practice gap
PLoS Med
(2014) - et al.
Live cumulative network meta-analysis: protocol for second-line treatments in advanced non-small-cell lung cancer with wild-type or unknown status for epidermal growth factor receptor
BMJ Open
(2016)
Preferred reporting items for systematic reviews and meta-analyses: the PRISMA statement
PLoS Med
Relation of completeness of reporting of health research to journals' endorsement of reporting guidelines: systematic review
BMJ
Endorsement of PRISMA statement and quality of systematic reviews and meta-analyses published in nursing journals: a cross-sectional study
BMJ Open
Reporting quality of systematic reviews and meta-analyses of otorhinolaryngologic articles based on the PRISMA statement
PLoS One
Reporting quality of systematic reviews/meta-analyses of acupuncture
PLoS One
Association of study quality with completeness of reporting: have completeness of reporting and quality of systematic reviews and meta-analyses in major radiology journals changed since publication of the PRISMA statement?
Radiology
Reporting and methodological quality of systematic reviews in the orthopaedic literature
J Bone Joint Surg Am
Time to consider sharing data extracted from trials included in systematic reviews
Syst Rev
The BMJ requires data sharing on request for all trials
BMJ
Cited by (0)
Funding: There was no direct funding for this study. M.J.P. is supported by an Australian National Health and Medical Research Council (NHMRC) Early Career Fellowship (1088535). J.E.M. is supported by an NHMRC Australian Public Health Fellowship (1072366). F.C.-L. is supported by the Generalitat Valenciana (PROMETEOII/2015/021). A.C.T. is funded by a Tier 2 Canada Research Chair in Knowledge Synthesis. D.M. is supported in part by a University Research Chair, University of Ottawa. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the article.
Conflicts of interest: The authors have read the journal's policy and the authors of this article have the following competing interests: A.C.T. is an associate editor for Journal of Clinical Epidemiology but had no involvement in the peer review process or decision for publication. M.J.P. and J.E.M. are affiliates of Cochrane Australia. M.J.P. is a co-convenor of the Cochrane Bias Methods Group. J.E.M. is a co-convenor of the Cochrane Statistical Methods Group. A.C.T. is an author of two of the systematic reviews included in this study but was not involved in eligibility assessment or data collection. D.G.A. is a senior investigator of National Institute for Health Research.
Data availability: The study protocol, data-collection form, and the raw data and statistical analysis code for this study are available on the Open Science Framework: https://osf.io/523bq/.