Note on Percentile Methodology Used in LIS Key Figures
Beginning with the Winter data release of December 2025, LIS adopts Definition 4 from Hyndman & Fan (1996) for the computation of the Key Figures and all percentile-based estimates (median, percentile ratios, poverty lines based on the median, decile shares, etc.). Under this definition, a percentile is obtained by locating its position in the cumulative distribution and linearly interpolating between the two adjacent ordered values, producing smooth and stable estimates.
We are adopting this method because it:
- works consistently across R and Stata,
- relies on cumulative weights, which ensures stability when computing shares or averages within distributional groups. Such estimates can be highly sensitive on small subsamples and when many repeated values or ties are observed in the microdata.
As a result, previously published figures have been revised to align with this unified approach. Even indicators not directly based on percentiles may be affected, as an interquartile rule is applied to cap top and bottom coded disposable household incomes in every dataset. Only 36 statistics showed a difference larger than 1 % between old and new estimate (=0.23 % of all concerned statistics). Nevertheless, even with a harmonized methodology, minor precision differences across software environments may still occur, but these remain rare.
To ensure consistent replication, users can already apply this methodology in LISSY:
- R users may rely on the ‘lissyrtools’ package, which implements Definition 4 by default, when using the relevant functions;
- Stata users may use the ‘percentils’ command developed by Philippe Van Kerm, which is pre-installed and ready to use in LISSY.
Reference scripts illustrating these procedures are available in the Replication Code section of the LIS website here.
Please note that DART has not yet been updated to this new methodology; its revision is planned for an upcoming release.
For further questions, please contact usersupport@lisdatacenter.org.