Abstract
Nonrandomized studies are increasingly used to evaluate interventions where randomization is not feasible or desired such as with policy reforms or practice change. A common practice is to statistically compare baseline characteristics between the control and intervention group to determine imbalances and confounders for model adjustment. This practice, however, has been shown to be inappropriate since false positives and negatives are not controlled. Moreover, the use of this practice to select confounders for model adjustment can introduce rather than protect against bias. The goals of this assessment were 1) to assess current publishing guidelines regarding baseline testing and 2) to elaborate recommendations. The guidelines from 16 high-impact journals were assessed. The journals did not provide direct guidance and referred authors to one or more of the following guidelines: ICMJE (International Committee of Medical Journal Editors), STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) and the Equator Network, including TREND (Transparent Reporting of Evaluations with Nonrandomized Designs). ICMJE provided no specific guidance and referred to STROBE. While STROBE did correctly recommend choosing confounders at study design stage; no guidance on specific analytical methods were given. Finally, TREND actually promoted baseline testing. Experts recommend that adjustment variables should be chosen at the design stage based on clinical knowledge. Sensitivity analyses, such as the use of doubly-robust methods, are also recommended. In conclusion, reporting guidelines need to be updated to offer more appropriate methods. Journal editors have the power to promote good research by explicitly discouraging baseline testing in nonrandomized studies.