A Comparison of two Robust Estimation Methods for Business Surveys

Abstract

SummaryTwo alternative robust estimation methods often employed by National Statistical Institutes in business surveys are two-sided M-estimation and one-sided Winsorisation, which can be regarded as an approximate implementation of one-sided M-estimation. We review these methods and evaluate their performance in a simulation of a repeated rotating business survey based on data from the Retail Sales Inquiry conducted by the UK Office for National Statistics. One-and two-sided M -estimation are found to have very similar performance, with a slight edge for the former for positive variables. Both methods considerably improve both level and movement estimators. Approaches for setting tuning parameters are evaluated for both methods and this is a more important issue than the difference between the two approaches. M-estimation works best when tuning parameters are estimated using historical data but is serviceable even when only live data is available. Confidence interval coverage is much improved by the use of a bootstrap percentile confidence interval.

Stratification of Domains Using Composite Estimation to Measure the Revenue Level of Small Businesses in Poland

“…Others use detection techniques to decide whether an observation is a potential outlier and should be moved to a special post-stratum, where its contribution to the survey estimates is downweighted (Chambers et al, 2014). If the number of outliers is small, estimation in the post-stratum can be difficult (Clark, Kokic, Smith, 2017). The approach proposed in this article is an attempt to cope with this problem.…”

Setting M-Estimation Parameters for Detection and Treatment of Influential Values

“…This is not equivalent to finding the optimal estimator in terms of achieving the minimum MSE over all possible estimates. In the latter case, one would attempt to guarantee a global solution for the problem at hand in the current data set, as recently studied by Martinoz et al (2015) and Clark et al (2017). These papers offer methodology for treating influential values in business surveys, but produce global solutions that adjust both influential and non-influential observations.…”