Evaluating Imputation Methods for Missing Data in a MCI Dataset

Missing data is a recurrent problem in experimental studies, mostly in clinical and sociodemographic longitudinal studies due to the dropout and the negative of some subjects to answer or perform some tests. To address this problem different strategies have been designed to deal with missing values, but incorrect treatment of missing data can result in the database being biased in one or more parameters, compromising the viability of the database and future studies. To solve this problem different imputation techniques have been developed over the last decades. However, there are no regulations or clear guidelines to deal with these situations. In this study, we will analyze and impute a real, incomplete database for the early detection of MCI, where the loss of values on 3 main variables is strongly correlated with the years of studies. The imputation will follow two strategies: assuming that those people would have got a bad scoring if they had taken the test, defining a ceiling score, and a multiple imputation by fully conditional specification. To determine if any kind of bias in mean and variance has been introduced during the imputation, the original database was compared with the imputed databases. Taking a p-value = 0.1 threshold, the database imputed by the multiple imputation method is the one that best preserved the information of the original database, making it the more appropriate imputation method for this MCI database.