Recovering Lost Information: Imputing Data to Generate Synthetic Complete Dataset

• In Event: Challenges in Learning from Panel Data

Abstract

Missing data is a common issue in social sciences and presents serious threats to unbiased inference. The common practice of case wise deletion discards incomplete observations, throwing away useful information and biasing inference. We introduce a powerful, yet easy to implement, imputation method to use this information for unbiased inference. That is synthesizing a complete and reasonable dataset based on an incomplete real dataset, then learning from model output of the complete dataset. We test this method on famous missing datasets in comparative politics, American politics and international relations as well as simulated missingness in datasets. We find in all cases the Synthetic Complete Datasets (SCD) method can provide researchers valuable insight. We finally provide some easy-to-use recommendations based on the SCD method for researchers.

Authors

• Jeff Gill

• Ali Amini, Center for Data Science - American University