Comparison of statistical methods of handling missing binary outcome data in randomized controlled trials of efficacy studies

摘要: The presence of some missing outcomes in randomized studies often complicates the estimation measures effect, even well designed controlled trials. process may be complicated further when efficacy rates are close to 0% or 100% as standard binomial model is susceptible non-convergence. main objective this study was compare performance multiple imputation (MI) and Complete Case analysis for dealing with binary modeling a risk difference. Firstly, however, regression COPY method Cheung’s modified Ordinary Least Squares (OLS) were examined using simulation processes their appropriateness difference modeling. It found that number copies (for method) required minimize non-convergence coincided gave most biased estimates true while increasing made problems bias worse; method, there convergence unbiased effect size. Simulation methods used complete case (CC) several models handling outcome data over wide range environments value assumptions. When at random (MAR) completely (MCAR), MI analyses included treatment group membership calculations yielded differences. CC good, better, than MAR MCAR, coverage 95% many situations – but neither nor produced not (MNAR). concluded equally good terms producing situations, applying intention treat principle (ITT) which requires all patients primary RCT, should adopted first choice, accompanied by secondary sensitivity purposes (i.e. investigate extent any likely bias).