Handling Dropouts by Copulas

The author shows how the copulas can be used to analyze repeated measurements with missing data. The main problem in practice is that there does not exist any theoretical joint distribution describing the repeated measurements, and then the copula function is a good tool to create the joint distribution. As the measurements are, in general, dependent, the conditional distribution of measurement in moment k, conditioned by the past measurements, might be highly informative for forecasting missing values. Imputations can be performed using conditional models. A simple method for imputation is presented using normal copula. The explicit formula to impute dropouts has been derived in the case of compound symmetry. The result can be easily applied to non-normal marginals using normalizing transformation.