Lack of Preregistered Analysis Plans Allows Unacceptable Data Mining for and Selective Reporting of Consensus in Delphi Studies
Published in: Journal of Clinical Epidemiology, Volume 99 (July 2018), Pages 96-105. doi:10.1016/j.jclinepi.2018.03.007
Posted on RAND.org on April 17, 2018
To empirically demonstrate how undisclosed analytic flexibility provides substantial latitude for data mining and selective reporting of consensus in Delphi processes.
Study Design and Setting
Pooling data across eight online modified-Delphi panels, we first calculated the percentage of items reaching consensus according to descriptive analysis procedures commonly used in health research but selected post hoc in this article. We then examined the variability of items reaching consensus across panels.
Pooling all panel data, the percentage of items reaching consensus ranged from 0% to 84%, depending on the analysis procedure. Comparing data across panels, variability in the percentage of items reaching consensus for each analysis procedure ranged from 0 (i.e., all panels had the same percentage of items reaching consensus for a given analysis procedure) to 83 (i.e., panels had a range of 11% to 94% of items reaching consensus for a given analysis procedure). Of 200 total panel-by-analysis-procedure configurations, four configurations (2%) had all items and 64 (32%) had no items reaching consensus.
Undisclosed analytic flexibility makes it unacceptably easy to data mine for and selectively report consensus in Delphi processes. As a solution, we recommend prospective, complete registration of preanalysis plans for consensus-oriented Delphi processes in health research.