141 – Tree-Based Methods for Missing Data and Evaluation of Missingness Mechanisms
Several Scenarios for Influential Observations and Methods for Their Treatment
Stephen Kaputa
U.S. Census Bureau
Mary Mulry
U.S. Census Bureau
Broderick E. Oliver
U.S. Census Bureau
In survey data, an observation is considered influential if it is reported correctly and its weighted contribution has an excessive effect on a key estimate, such as an estimate of total or change. Influential observations occur infrequently in economic surveys but have a detrimental effect on the key estimate when they do appear. In previous research with data from the U.S. Monthly Retail Trade Survey (MRTS), two methods, Clark Winsorization and weighted M-estimation, have shown potential to detect and adjust influential observations. This paper discusses results of the application of an improved simulation methodology that generates more realistic population data. The analyses consider several scenarios for the occurrence of influential observations in the MRTS and assess the performance of the two methods in detecting influential values under the different scenarios.
