Efficiency of Multiple Imputation for Retrieving Data with Missing Not at Random (MNAR) Patterns Under Objective Bayesian and Frequentist Approaches (306518)*Sofia De los Ángeles Bartels Gómez, University of Costa Rica
Keywords: Multiple Imputation, simulations, Missing Not at Random
Using survey data from a psychosocial study, this research pretends to assess which paradigm, the Objective Bayesian or the Frequentist, is more efficient in retrieving the real data values, when working with Missing Not at Random (MNAR) patterns, and, using Multiple Imputation to estimate the missing data. Several simulations scenarios are created, using as a reference, the real scenario of the study in which missing data appeared. In one extreme, the scenario presents MCAR patterns. In the other extreme, the scenario shows the situation in which there are three variables (one categorical and two continuous) that are highly correlated with the presence of missing values in the variable of interest. The scenarios are simulated in two conditions: 20% and 40% missing data in the variable of interest. Conclusions are drawn in terms of which method provides more accurate results under which particular settings, considering also aspects of cost-effectiveness in relation to computational time and other practical issues. Recommendations regarding the handling of missing values will be given to Social Scientists who frequently face these problems in their surveys.