Case studies
Before I propose to move a program's estimation to new code, I show that the new code gets the agency's own published numbers. Each study below takes a national survey's public microdata and reproduces its official estimates and standard errors in Python, with code anyone can rerun.
Reproduces the report's sampling-error appendix: estimates, standard errors and design effects.
For: Health ministries and DHS implementing partners.
Read on svylab.com → U.S. healthReproduces two NCHS data briefs at published precision, including their Korn–Graubard intervals. Agrees with R's survey package to machine precision.
For: U.S. health research teams.
Read on svylab.com → Official statisticsImplements Statistics Canada's Poisson bootstrap and validates it against the agency's own R code.
For: Statistics offices that publish replicate-weight files.
Read on svylab.com → Labour force · PanelShows what a rotating panel changes: the weight for levels, and half the standard error for quarter-to-quarter change.
For: Labour-force programs.
Read on svylab.com → Time useTurns a 10-million-row diary into person-level estimates and reproduces every table in the fact sheet.
For: Teams with large, messy source files.
Read on svylab.com → Living standards · PanelReproduces both reports, under the cross-sectional and the longitudinal weights, plus the change tables.
For: Panel surveys and LSMS partners.
Read on svylab.com → ModernizationAHRQ's R workshop, rebuilt step by step in Python.
For: Agencies moving off legacy code.
Read on svylab.com →Two more were done independently by survey statisticians in Brazil and Mexico, on their own national data (all case studies). Across 37 estimators, svy matches R's survey package to six decimals (the comparison).
Discuss a modernization engagement →