Imputation methods for incomplete panel data with applications to latent growth curves

Abstract

<p>Researchers who conduct longitudinal studies face the problem of incomplete data. In this investigation we study the efficiency of three methods for data imputation---Expectation Maximization (EM), multiple imputations (MI), and full information maximum likelihood (FIML)---as they apply to three patterns of missing data: missing completely at random (MCAR), missing at random (MAR), and nonignorable (NI) missing. The study unfolded in two phases, in the first we analyzed real data and in the second we included simulations to assess robustness of estimates against violations of multivariate normal (MVN) distribution. We used latent growth curves model to evaluate the performance of imputation methods;With real data, the results showed that estimates using MI and FIML were unbiased and reduced standard errors with respect to the listwise deletion (LD) method under MVN distribution. These findings were confirmed using simulated data. All imputation methods (EM, MI, and FIML) performed well for the mechanism MCAR and MAR, in particular under MVN distribution. MI showed less bias compared with estimates from the LD method. For the NI mechanism all the methods yielded biased estimates with real and simulated data. These findings indicate that even under the worst assumptions the use of imputation methods is highly recommended for recovering incomplete observations. Inferences are more reliable using imputed data than using listwise deletion.</p>