Imputation of missing data via penalization techniques

The aim of this master thesis is to give the user an estimate of uncertainty over missing data imputation. The full factorization approach is compared to the state-of-the-art approach of full conditional. The special feature in both algorithms is the penalization techniques. Both algorithms are used with different types of missing data like MAR, MCAR and NMAR. Simulated datasets were conducted with copulas. Simulations were varied in rate of missing observations, refitting times and the use, or not, of LASSO regression for last fit. Results are given in terms of accuracy of predicted values, pooled variance estimates and errors which occurred during programming and runtime. The full factorization approach showed advantages over full conditional especially if one looks on ratios of 10:1 observations to covariables. In cases were covariables were in higher numbers than observations, full conditional and full factorization nearly covered same results when ridge regression was used for all fits. Generally lasso regression did not improve accuracy of imputation results. This result can be generalized for all missing types used and simulations conducted in this master thesis. Imputed observations showed paths which are similar to MCMC bayesian statistics. The imputation steps alternated and converged to certain values. Results were stable when different datasets and seeds for random numbers were processed. Run time for both approaches was high due to errors that occurred in different R packages and functions which are costly in terms of CPU usage.