Background: Prospective trial registration is an important tool for reducing reporting bias. Registration prior to enrolment of the first participant helps ensure transparency by publicly documenting key characteristics and outcomes prior to data collection or analysis. It has been proposed that systematic reviews should only include prospectively registered studies but it is currently unclear whether risk of bias is generally higher for retrospectively registered studies.
Objectives: To compare study characteristics and overall risk of bias for prospectively and retrospectively registered studies.
Methods: We included all 1753 studies registered on the ANZCTR in 2016. Registrations before enrolment of the first participant were defined as prospective; registrations after were defined as retrospective. We analysed whether timing of registration was related to allocation method (randomised versus non-randomised), blinding and target sample size. We also combined these criteria to assess overall risk of bias: studies that were randomised, blinded and had an above-median sample size were classified as low risk of bias.
Results: As shown in the Figure, prospectively registered trials were more likely to be randomly allocated (OR=1.29, 95% Confidence Interval (CI)=1.01-1.64), and to blind their participants to treatment allocation (OR=1.59, 95% CI=1.27-1.98). There was no statistically significant difference in target sample size (Median (Interquartile Range) prospective=72 (157); Median (Interquartile Range) retrospective=70(129), p=.39). Of the studies classified as low risk of bias, 41% were retrospectively registered and 59% were prospectively registered, this difference was statistically significant (OR=1.53, 95%CI=1.18-1.98).
Conclusion: While on average prospectively registered studies were less likely to be biased, 4 out of 10 of all studies that had an overall low risk of bias were retrospectively registered. Excluding retrospectively registered studies from systematic reviews would result in the exclusion of large numbers of low risk-of-bias studies that would otherwise add important information. Other potential solution will be presented.