Evaluating the Performance of Different Multiple Imputation Methods When Imputing Missingness in Time-Series-Cross-Sectional Data

Abstract

This thesis evaluates the performance of different multiple imputation methods in imputing country-level proportions of key indicators that are missing in time-series-cross-sectional (TSCS) data. When imputing the country-level proportions missing in TSCS data due to questions not asked in the survey, we found that Amelia and Multiple Imputation by Chain Equation for two-level panel data (mice.2l.pan) performed best among seven methods being evaluated for both methods converged fast, produced reasonable and stable imputations and had small out-of-sample root mean squared error (RMSE) less than ±5% for proportions imputed and 95% coverage rate (CR_95) very close to 95%. In addition, we found that including incomplete auxiliary variables that are correlated with targeted incomplete variables improved the imputation performance regardless of the missing rate of the auxiliary variables. However, including the cluster means had little impact on the imputation performance. The goal of the thesis is to produce empirical evidence on the performance of different multiple imputation methods in imputing missingness in TSCS data.