How to demonstrate similarity by using noninferiority and equivalence statistical testing in radiology research

Abstract

Demonstrating similarity between compared groups--that is, equivalence or noninferiority of the outcome of one group to the outcome of another group--requires a different analytic approach than determining the difference between groups--that is, superiority of one group over another. Neither a statistically significant difference between groups (P < .05) nor a lack of significant difference (P ≥ .05) from conventional statistical tests provides answers about equivalence/noninferiority. Statistical testing of equivalence/noninferiority generally uses a confidence interval, where equivalence/noninferiority is claimed when the confidence interval of the difference in outcome between compared groups is within a predetermined equivalence/noninferiority margin that represents a clinically or scientifically acceptable range of differences and is typically described by Δ. The equivalence/noninferiority margin should be justified both clinically and statistically, considering the loss in the main outcome and the compensatory gain, and be chosen conservatively to avoid making a false claim of equivalence/noninferiority for an inferior outcome. Sample size estimation needs to be specified for equivalence/noninferiority design, considering Δ in addition to other general factors. The need for equivalence/noninferiority research studies is expected to increase in radiology, and a good understanding of the fundamental principles of the methodology will be helpful for conducting as well as for interpreting such studies.