Comparative Simulation Study On Missing Data Handling Using Pattern Mixture Models

Background:In clinical trials,missing data can arise for a variety of reasons,including lost to follow-up,non-adherence,adverse events and using rescue medication.Missing data can jeopardize the key advantage of randomized clinical trials,reduce the power of clinical trials and become a potential source of bias,thereby degrading interpretation of the results and their reliability.Therefore,it is important to use proper methods to handling missing dataWidely used single imputation methods assume that,for those with missing value,the predicted value is the same as the imputation value with probability one.For example,Last Observation Carried Forward(LOCF),which is based on the assumption that all subsequent,scheduled,but missing evaluations are equal to the last measured value of the endpoint,underestimates the variance of values,distorts the distribution of data and the relationship between data.Mixed model for repeated measurements(MMRM)and multiple imputation(MI)assume data missing at random(MAR),which is unlikely to be entirely true.When data is missing not at random(MNAR),these methods will introduce even more bias.Pattern mixture models(PMM)specify the relationship of distributions for the missingness and for the outcome under MNAR assumptions,thus making the interpretation of assumptions to be theoretically straightforward and clinically plausible,and allowing for analyzing missing data with multiple assumptions in a sensitivity analysis.However,it is yet still inconclusive that which of the aforementioned methods have better statistical performance.Furthermore,the statistical performance of PMM still needs comprehensive and detailed comparative simulation studies,since it receives attention much later,and then much less implementation as well as comparative study with other methods.Objective:This study aims to evaluate the statistical performance of commonly used approaches and pattern mixture models in missing data handling under different missing mechanisms(single missing mechanism and mixed missing mechanisms)in order to provide basis for missing data handling in clinical trialsMethod:Monte Carlo method was used to simulate the completed datasets and then generate corresponding missing datasets under various situations,including missing mechanisms,missing rates,correlation coefficient and effect changing pattern.The performance of the different methods were evaluated with type ? error,power,estimated bias of treatment effect in each group and between groups,the coverage probabilities and the width of 95%confidence interval(95%CI)compared with the performance of the completed datasets analysis.Results:Both MMRM and MI controlled type ? error well and slightly reduced power compared with completed dataset analysis.In most scenarios,Type ? error of LOCF was not well controlled and the variability of power was large.PMM controlled type? error below the specific level and had lower power compared with MMRM and MI.In most scenarios,MMRM and MI showed smaller bias.However,under MNAR,the bias increased with overestimated treatment effect of experimental group and the coverage probabilities of 95%CI between groups slightly declined.LOCF performed unsteadily,which was worsened under the MAR and MNAR.When dealing with data under single MNAR or substantial amount of MNAR,PMM demonstrated smaller bias and the coverage probabilities of 95%CI were higher.In most scenarios,PMM underestimated the treatment effect of test group as well as between groups.In all scenarios we considered,MI and PMM had the largest width of 95%CI,followed by MMRM and LOCF.Basically,the higher correlation coefficient and the smaller the missing missing rates or difference leads to better statistical performance.LOCF and PMM were affected considerably by effect changing patternConclusion:Either MMRM or MI approaches can be considered as the primary statistical method according to actual conditions,since they performed equally well and were regularly affected by other factors under MCAR and MAR missing mechanism.LOCF underestimated the variability and hence improved precision because of its specific imputation method,but it should be applied with caution since its biggest disadvantage was the weak robustness as well as the weak control of type I error.When some data are MNAR or mixed missing mechanisms,the statistical performance of MMRM and MI become weaker.Conducting sensitivity analyses by PMM is recommended to test the robustness of results given the missing data,but more attention should be taken since PMM tend to have over-conservative estimation.