Incomplete Data


Health Economics

  • It is common for the data in clinical studies and cost-effectiveness analyses to be incomplete. This may or may not be a significant problem. Examples of types of missing data include single missing items (e.g. failure to record a survey result for one item in an EQ-5D schedule), missing whole questionnaires, and missing data due to drop-out. Whether these omissions matter will depend partly on whether they are 'missing completely at random', in which case the sample remains representative, 'missing at random', in which case they can be imputed, or whether they are not randomly missing (sometimes termed 'non-ignorable non-response'). Ways of coping (which can hardly be commended) have included ' last observation carried forward ', ' complete case analysis ' (i.e. using only complete cases with no imputed values, with the risk of bias if the sample with omissions is not representative); ' unconditional mean imputation ' (i.e. replacing missing data with the mean value of the data in the sample with omissions), with again evident risk of bias. Better methods include 'regression imputation', 'stochastic regression imputation', and 'multiple imputation' in which missing values are replaced with plausible alternatives in a process that takes account of the uncertainty about the right value to impute.