Bias-variance trade off on the use of person non-response weights in inequality estimates

Presentation at the Q2022 - Bias-variance trade off on the use of non-response weights in inequality estimates

Author

👨 Josep Espasa Reig - Data Scientist @ LIS - Cross-National Data Center in Luxembourg

Introduction

👋 Welcome to the repository of my presentation for the Q2022 - European Conference on Quality in Official Statistics conference in Vilnius!

Contents

Here you will find a short summary of the presentation and the paper in PDF format.

Summary / TL;DR:

❓ Should inequality researchers use person-level weights instead of household-level ones if these are available?

In recent years, some National Statistical Offices have started producing weights at individual-level. Compared to those at household-level, these contain an extra adjustment by the non-response propensity of individuals. This adjustment might imply more variance in weights and thus larger Standard Errors for estimates. If the variables used to correct for individual non-response are not associated with those used for inequality analyses, then there wouldn't be any benefit using the individual-level weights.

The paper uses LIS datasets for Germany and US. When using person-level weights (instead of household-level ones):

• Inequality estimates increase slightly
• The gap in the coverage ratio with National Accounts gets reduced
• There is an increase in variance and therefore a reduction in the effective sample size
• The points above are especially true for the last decade of German surveys and for indicators such as the ‘Poverty Gap’ and ‘Poverty Headcount’

Researchers should therefore use person-level weights whenever available, as they are most likely reducing bias. They should, however, be aware that this adjustment comes at a cost of decreased precision in estimates.

Paper in PDF

• 📝 See the paper in PDF format here.