Clinical Trials

Clinical Trials: A Practical Guide ■❚❙❘This approach is simple and easy to implement, but it is based on the assumptionthat had the participant remained in the trial until completion, the participant’smeasurements would have remained exactly the same as at the time of droppingout. A number of problems have been documented with this approach [7].Time trends in the data, when combined with differential dropout rates betweengroups, can introduce severe bias. This method also ignores the fact that, even ifa participant’s disease state remains constant, measurements of this state areunlikely to stay exactly the same, introducing a spurious lack of random variabilityinto the analysis.Other alternatives to LOCF are occasionally adopted: carrying forward theworst case, carrying forward the best case, or carrying forward the baselinevalue [10,11]. These methods have drawbacks similar to those of LOCF and wetherefore recommend that these strategies be used predominantly to conducta sensitivity analysis.Multiple imputationLOCF and other strategies where previously observed values (last, worst, or bestobserved) are carried forward to fill in the missing values belong to a class of methodstermed imputation methods [12]. They can be labeled nonparametric or data-basedimputation methods, as opposed to parametric or model-based imputation [13].Parametric imputation involves replacing missing values with predictions from astatistical model; in nonparametric imputation, the missing value is replaced withan observed value. Provided that an appropriate statistical model is used (one thattakes account of the observed predictors and responses that are related to themissingness mechanism) then parametric imputation can give rise to unbiasedparameter estimates under MAR. However, the fact that only a single imputationof each missing value is made, and that there is no allowance for imprecision insuch imputations in the analysis, introduces spurious precision into the estimates.Multiple imputation can be used to take appropriate account of uncertainty in theimputed values [8]. Provided that an appropriate model is used for imputation,this approach can result in valid unbiased estimates under MAR. Horton andLipsitz present the multiple imputation method as a three-stage process [14].• Firstly, sets of ‘plausible’ values for missing observations are created.Each of these sets is used to fill in the missing value and createa ‘completed’ dataset.• Secondly, each of these datasets is analyzed.• Finally, the results are combined, taking account of imprecision in parameterestimates from each ‘completed’ dataset and variation between datasets.343