Evaluating characteristics of PROSPERO records as predictors of eventual publication of non-Cochrane systematic reviews: a meta-epidemiological study protocol
Systematic Reviews volume 7, Article number: 43 (2018)
Epidemiology and the reporting characteristics of systematic reviews (SRs) and meta-analyses (MAs) are well known. However, no study has analyzed the influence of protocol features on the probability that a study’s results will be finally reported, thereby indirectly assessing the reporting bias of International Prospective Register of Systematic Reviews (PROSPERO) registration records.
The objective of this study is to explore which factors are associated with a higher probability that results derived from a non-Cochrane PROSPERO registration record for a systematic review will be finally reported as an original article in a scientific journal.
The PROSPERO repository will be web scraped to automatically and iteratively obtain all completed non-Cochrane registration records stored from February 2011 to December 2017. Downloaded records will be screened, and those with less than 90% fulfilled or are duplicated (i.e., those sharing titles and reviewers) will be excluded. Manual and human-supervised automatic methods will be used for data extraction, depending on the data source (fields of PROSPERO registration records, bibliometric databases, etc.). Records will be classified into published, discontinued, and abandoned review subgroups. All articles derived from published reviews will be obtained through multiple parallel searches using the full protocol “title” and/or “list reviewers” in MEDLINE/PubMed databases and Google Scholar. Reviewer, author, article, and journal metadata will be obtained using different sources. R and Python programming and analysis languages will be used to describe the datasets; perform text mining, machine learning, and deep learning analyses; and visualize the data. We will report the study according to the recommendations for meta-epidemiological studies adapted from the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement for SRs and MAs.
This meta-epidemiological study will explore, for the first time, characteristics of PROSPERO records that may be associated with the publication of a completed systematic review. The evidence may help to improve review workflow performance in terms of research topic selection, decision-making regarding team selection, planning relationships with funding sources, implementing literature search strategies, and efficient data extraction and analysis. We expect to make our results, datasets, and R and Python code scripts publicly available during the third quarter of 2018.
Meta-epidemiological research is designed to evaluate non-clinical aspects of primary and secondary studies, especially their methodological quality . Epidemiology and the reporting characteristics of systematic reviews (SRs) have been previously described by a research group using a cross-sectional random sample of 300 published reviews at two different points in time: in 2004, by Moher et al. , and in 2014, by Page et al. . In both cases, several characteristics of the published reviews were analyzed, taking into account the potential differences between Cochrane and non-Cochrane reviews, focus of study (therapeutic, diagnosis, epidemiology, other), self-reported use of Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA), and differences in SR features between 2004 and 2014. Those authors estimate that more than 8000 SRs are indexed in MEDLINE annually. This type of research may give us a better picture of the overall state of this synthesis of documents, using a representative sample of finally published reviews. Furthermore, some studies have even looked at discrepancies between protocols and the final published reviews [4, 5]. For example, Tricco et al. found that a third of reviews changed or did not specify the primary outcome when the results of International Prospective Register of Systematic Reviews (PROSPERO) registration records were published . These studies highlighted some new issues related to the usefulness of preparing a priori protocols for developing SRs and meta-analyses (MAs) if there are no controls governing the development and reporting of workflow (repository curators, journal editors and reviewers, etc.).
Recently, Tsujimoto et al. have found that 26% of non-Cochrane records registered during the first year after PROSPERO creation remained unpublished . We recently performed a scoping review (data not published) to scrape protocols from the PROSPERO repository by iteratively running a custom Python script from November 25 to December 1, 2017, from RecordIDFootnote 1 “00001” to “80000,” due to the lack of a clear pattern for assigning RecordID to every published registration record and PROSPERO’s limitations for massively mining data using query tools on the user’s web interface . As a result, after a rapid curating process, 20,272 full non-Cochrane registration records were obtained (25.3%), of which only 1042 reviews were finally associated with an article communicating results in a scientific journal (5.1%). These numbers are very impressive and make us question whether there is a new and underestimated source of publication bias in the field of SRs and MAs. The magnitude of the problem may be even greater if we consider that by 2015, the majority of the authors of SRs (78%) published in high-impact journals were neglecting to register the protocols they used . Our group has previously demonstrated that certain factors—such as the number of authors with conflicts of interest (COIs), sources of funding, number of authors, and the architecture of author-affiliation networks—may influence the methodological quality of SRs regarding psoriasis [10, 11]. However, no study has analyzed the influence of PROSPERO registration record features on the probability that the results will be finally reported, thereby indirectly assessing the reporting bias of the registration records of SRs and MAs that have completed all PROSPERO stages.
Our hypothesis is that there may be some features related to non-Cochrane registration records that may help predict which results will be finally published in a scientific journal.
In terms of the abovementioned discrepancies between the number of completed registration records at PROSPERO and the low rate of finally published results, we will conduct a meta-epidemiological study with the following objectives:
To develop and assess different predictive models of results reporting using deep learning strategies based on data and metadata extracted from both published registration records collected in the PROSPERO repository and derived original articles published as SRs or meta-analyses in scientific journals
To evaluate the influence of discrepancies between the methodological quality of registration records and SRs on the publication bias found in the PROSPERO repository using AMSTAR, AMSTAR-2, and a modified version of AMSTAR (without reporting items, in the case of protocol assessment)
Web scraping will be performed using a custom Python script and Chrome’s Web Scraper website data extraction tool (http://webscraper.io/) to automatically and iteratively extract the raw data of all completed non-Cochrane registration records stored in the PROSPERO repository (https://www.crd.york.ac.uk/prospero/) from February 2011 to December 2017. We will not perform PROSPERO registration record sampling. Rather, our objective is to obtain the entire universe of non-Cochrane PROSPERO registration records—from the first document to the last one registered just before the date of the web scraping—not a representative sample of them. The search specificity for non-Cochrane PROSPERO registration records is based on Python script designed to recognize only the format of these records, which differs from registration records for Cochrane and non-human studies. These cannot be scraped using our script due to the structural differences in PROSPERO forms between them. Thus, the sensitivity and specificity for the web scraping are 100%.
The extracted data will be stored locally as .csv files, where rows will represent protocols and columns protocol sections.
Eligibility and screening
Registration records with less than 90% of their sections fulfilled or those that are duplicated (i.e., those sharing titles and reviewers) will be dropped from the dataset. Included registration records must have achieved “all completed stages” status (“preliminary searches,” “piloting of the study selection process,” “formal screening of search results against eligibility criteria,” “data extraction,” “risk of bias [quality] assessment,” and “data analysis”). Finally, these completed registration records will be classified into three groups: (a) published reviews, if at least one publication with results associated with a registration record is available; (b) discontinued reviews, if the authors explain why the results have not been finally published; and (c) abandoned reviews, if the results were not published and the authors never explained why. An R script will automatically perform the screening process. After that, the results will be subjected to human verification.
Table 1 displays all the fields that will be extracted from non-Cochrane registration records. All articles derived from published reviews will be obtained by multiple parallel searches using the full registration record “title” and/or “list reviewers” in MEDLINE/PubMed databases and Google Scholar. We will only contact authors to request the final reports associated with completed protocols that will be found after bibliographic database searches. Reviewer, author, article, and journal metadata will be obtained using different sources (e.g., SCOPUS, Web of Science, Google Scholar). Table 2 shows the variables that will be included from among reviewer, author, article, and journal metadata. Data from the selected PROSPERO registration records will be automatically extracted using specific regex R syntax. Human verification of the R script process will be done on a sample of records (20%) by two independent researchers. Data and metadata from authors, journals, and articles associated with full PROSPERO registration records will be extracted manually by several researchers using a specifically designed AppSheet form. Given the expected large amount of manual work, the published articles will be assigned to two different teams. Each member of the same team will mine only certain number of variables; thus, the analysis of every group of variables will be performed by two different researchers. Regression analysis will be conducted using the identification of the research member involved in the data extraction as a fixed factor in the mixed effects logistic regression model.
Data analysis and reporting
Qualitative variables will be summarized by level frequency (number, %) or displayed using several types of graphs (mosaic plots, density/histogram plots, etc.). Quantitative variables will be summarized using the mean (standard deviation) or median (interquartile range) for non-normally distributed variables. We will look for features of PROSPERO registration records (Table 1) and metadata (Table 2) that can predict publication (or non-publication) using mixed effects logistic regression models. Results will be displayed using odds ratios (OR). The 95% confidence intervals will be estimated by simple random sampling with replacement for bootstrapping for every variable included in the model.
We plan to look for publications in a random 10% sample of abandoned PROSPERO records following the same method used for the completed and published PROSPERO records. By doing this, we can check the robustness of our approach as well as the ratio of return information by authors after results are published.
R and Python programming and analysis languages will be used to describe the datasets; perform text mining, machine learning, and deep learning analyses; and visualize the data. We will implement actions aimed at making every effort to improve the transparency, reproducibility, and efficiency of our work. First, by publishing this a priori protocol of our meta-epidemiological study, we will avoid publication bias and issues related to analytical flexibility (outcome switching and P-hacking). Second, our raw datasets and R and Python code scripts will be made publicly available in the GitHub repository, using version control and R packages such as R Markdown, knitR, and packrat. Third, we will report the study according to a previously adapted recommendation of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement for meta-epidemiological studies, proposed in 2017 by Murad and Wang . Those authors adapted the items used in the PRISMA statement for systematic reviews and meta-analyses to improve the transparency of this type of research. The items were categorized according to their title (1), abstract (1), introduction (2), methods (10), results (6), discussion (3), and funding (1).
This meta-epidemiological study will explore for the first time, using a large sample of studies, the factors that may be associated with the success of an a priori systematic review registration record, defined as one entailing a high probability that a study’s results will finally be published as an original article in a scientific journal. The findings of this study will be useful in the future for improving protocol design when researchers using non-Cochrane protocols decide to develop a systematic review and meta-analysis. The evidence may also help to improve the performance of review workflow, including selecting a better topic of research, making better decisions regarding team selection, planning relationships with research funding sources, implementing literature search strategies, and performing more efficient data extraction and analysis. Our results, datasets, and R and Python code scripts are expected to be made publicly available during the third quarter of 2018.
By replacing XXXXX with the RecordID at https://www.crd.york.ac.uk/prospero/display_record.php?RecordID=XXXXX.
A Measurement Tool to Assess the Methodological Quality of Systematic Reviews
Preferred Reporting Items for Systematic Reviews and Meta-Analyses
Murad MH, Wang Z. Guidelines for reporting meta-epidemiological methodology research. BMJ Evid. Based Med. 2017;22:139-142.
Moher D, Tetzlaff J, Tricco AC, Sampson M, Altman DG. Epidemiology and reporting characteristics of systematic reviews. PLoS Med. 2007;4(3):e78.
Page MJ, Shamseer L, Altman DG, Tetzlaff J, Sampson M, Tricco AC, Catalá-López F, Li L, Reid EK, Sarkis-Onofre R, Moher D. Epidemiology and reporting characteristics of systematic reviews of biomedical research: a cross-sectional study. PLoS Med. 2016;13(5):e1002028.
Silagy CA, Middleton P, Hopewell S. Publishing protocols of systematic reviews: comparing what was done to what was planned. JAMA. 2002;287(21):2831–4.
Kirkham JJ, Altman DG, Williamson PR. Bias due to changes in specified outcomes during the systematic review process. PLoS One. 2010;5(3):e9810.
Tricco AC, Cogo E, Page MJ, Polisena J, Booth A, Dwan K, MacDonald H, Clifford TJ, Stewart LA, Straus SE, Moher D. A third of systematic reviews changed or did not specify the primary outcome: a PROSPERO register study. J Clin Epidemiol. 2016;79:46–54.
Tsujimoto H, Tsujimoto Y, Kataoka Y. Unpublished systematic reviews and financial support: a meta-epidemiological study. BMC Res Notes. 2017;10:703.
PROSPERO. York, England: Centre for Reviews and Dissemination, University of York; [updated December 2017]. Available at HYPERLINK https://www.crd.york.ac.uk/prospero/.
Tsujimoto Y, Tsujimoto H, Kataoka Y, Kimachi M, Shimizu S, Ikenoue T, Fukuma S, Yamamoto Y, Fukuhara S. Majority of systematic reviews published in high-impact journals neglected to register the protocols: a meta-epidemiological study. J Clin Epidemiol. 2017;84:54–60.
Gómez-García F, Ruano J, Aguilar-Luque M, Gay-Mimbrera J, Maestre-Lopez B, Sanz-Cabanillas JL, Carmona-Fernández PJ, González-Padilla M, Vélez García-Nieto A, Isla-Tejera B. Systematic reviews and meta-analyses on psoriasis: role of funding sources, conflict of interest and bibliometric indices as predictors of methodological quality. Br J Dermatol. 2017; https://doi.org/10.1111/bjd.15380.
Sanz-Cabanillas JL, Ruano J, Gomez-Garcia F, Alcalde-Mellado P, Gay-Mimbrera J, Aguilar-Luque M, Maestre-Lopez B, Gonzalez-Padilla M, Carmona-Fernandez PJ, Garcia-Nieto AV, Isla-Tejera B. Author-paper affiliation network architecture influences the methodological quality of systematic reviews and meta-analyses of psoriasis. PLoS One. 2017;12(4):e0175419.
Murad MH, Wang Z. Guidelines for reporting meta-epidemiological methodology research. Evid Based Met. 2017;22:139–42.
Wicherts JM. Peer review quality and transparency of the peer-review process in open access and subscription journals. PLoS One. 2016;11(1):e0147913.
We wish to acknowledge the assistance received from Dr. Camilla Hansen, Research Coordinator, Cochrane Bias Methods Group, University of Southern Denmark and Odense University Hospital, Odense, Denmark, in answering our questions about some aspects of how to evaluate the bias risk of PROSPERO protocols during the preparation phase of our protocol. The authors are grateful to all the contributors and to the advisory group of the PROSPERO registry. The authors are also grateful to the reviewers for their helpful comments and suggestions. We would also like to thank Editage (www.editage.com) for the English language editing.
This work is supported, in part, by project ICI1400136 to JR, integrated into the National Plan of R+D+I 2008–2011 and co-financed by the ISCIII-Subdireccion General de Evaluacion and European Regional Development Fund (ERDF); by project PIN-0316-2017-Consejería de Salud-Junta de Andalucía to JR; and by grants PP13/009 (JR) and PP17/001 (JG-M) of Plan Propio de movilidad para investigadores del Instituto Maimonides de Investigacion Biomedica de Cordoba (IMIBIC).
Availability of data and materials
Ethics approval and consent to participate
Consent for publication
The authors declare that they have no competing interests.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
About this article
Cite this article
Ruano, J., Gómez-García, F., Gay-Mimbrera, J. et al. Evaluating characteristics of PROSPERO records as predictors of eventual publication of non-Cochrane systematic reviews: a meta-epidemiological study protocol. Syst Rev 7, 43 (2018). https://doi.org/10.1186/s13643-018-0709-6