Imputation Methods for Incomplete Dependent Variables in Finance

Missing observations in dependent variables is a common feature of many financial applications. Standard ad hoc missing value imputation methods invariably fail to deliver efficient and unbiased parameter estimates. A number of recently developed classical and Bayesian iterative methods are evaluated for the treatment of missing dependent variables when the independent variables are completely observed. These methods are compared by simulation to commonly applied alternative missing data methodologies in the finance literature. The methods are then applied to a system of simultaneous equations modelling the maturity, secured status, and pricing of U.S. bank revolving loan contracts. Two of the four dependent variables in this application are characterised by severe missingness. The system of equations approach allows us to also exploit the additional information contained in the interdependencies among these features. The results indicate that proper treatment of missingness can be important for many financial applications.