Automated detection of over- and under-dispersion in baseline tables in randomised controlled trials.

Background : Papers describing the results of a randomised trial should include a baseline table that compares the characteristics of randomised groups. Researchers who fraudulently generate trials often unwittingly create baseline tables that are implausibly similar (under-dispersed) or have large differences between groups (over-dispersed). I aimed to create an automated algorithm to screen for under- and over-dispersion in the baseline tables of randomised trials. Methods : Using a cross-sectional study I examined 2,245 randomised controlled trials published in health and medical journals on PubMed Central . I estimated the probability that a trial's baseline summary statistics were under- or over-dispersed using a Bayesian model that examined the distribution of t-statistics for the between-group differences, and compared this with an expected distribution without dispersion. I used a simulation study to test the ability of the model to find under- or over-dispersion and compared its performance with an existing test of dispersion based on a uniform test of p-values. My model combined categorical and continuous summary statistics, whereas the uniform test used only continuous statistics. Results : The algorithm had a relatively good accuracy for extracting the data from baseline tables, matching well on the size of the tables and sample size. Using t-statistics in the Bayesian model out-performed the uniform test of p-values, which had many false positives for skewed, categorical and rounded data that were not under- or over-dispersed. For trials published on PubMed Central , some tables appeared under- or over-dispersed because they had an atypical presentation or had reporting errors. Some trials flagged as under-dispersed had groups with strikingly similar summary statistics. Conclusions : Automated screening for fraud of all submitted trials is challenging due to the widely varying presentation of baseline tables. The Bayesian model could be useful in targeted checks of suspected trials or authors.