Diagnosis checking of statistical analysis in RCTs indexed in PubMed

Statistical analysis is essential for reporting of the results of randomized controlled trials (RCTs), as well as evaluating their effectiveness. However, the validity of a statistical analysis also depends on whether the assumptions of that analysis are valid.

To review all RCTs published in journals indexed in PubMed during December 2014 to provide a complete picture of how RCTs handle assumptions of statistical analysis.

We reviewed all RCTs published in December 2014 that appeared in journals indexed in PubMed using the Cochrane highly sensitive search strategy. The 2014 impact factors of the journals were used as proxies for their quality. The type of statistical analysis used and whether the assumptions of the analysis were tested were reviewed.

In total, 451 papers were included. Of the 278 papers that reported a crude analysis for the primary outcomes, 31 (27·2%) reported whether the outcome was normally distributed. Of the 172 papers that reported an adjusted analysis for the primary outcomes, diagnosis checking was rarely conducted, with only 20%, 8·6% and 7% checked for generalized linear model, Cox proportional hazard model and multilevel model, respectively. Study characteristics (study type, drug trial, funding sources, journal type and endorsement of CONSORT guidelines) were not associated with the reporting of diagnosis checking.

The diagnosis of statistical analyses in RCTs published in PubMed-indexed journals was usually absent. Journals should provide guidelines about the reporting of a diagnosis of assumptions.