Measurement error and the Labour Force Survey
Felix Ritchie and Damian Whittard
Presentation by
Felix Ritchie
Professor of Applied Economics
DirectorBristol Centre for Economics and Finance
Date
Measurement error
• Problem: observed value = actual + error
directly unobservable
• Statistical problemo Linear statistics biased if mean error ≠0o Model estimates biased if error ≠0
• Solutionso Data collection: follow-up/verification surveyso Marginal analysis: instrument variableso Linear statistics: bigger confidence intervals?
• Can we do better by understanding the cause of error?
Aims of the research
• Does measurement error in LFS earnings data exist? How? Why? Does it matter?
• What we can learn about household responses generally?
• Can we improve our response strategies?
UK Labour Force Survey
• Major personal survey in UK
• 60,000 individuals surveyed each quartero Five-quarter longitudinal follow-up
• Voluntary: response rates 20%-40%
• Information ono Personal characteristics (ethnicity, education etc)o Labour market activityo Health
• Significant userso ONS: labour market statistics by personal characteristicso Government: BEIS, DWP, Low Pay Commission particularlyo Academia
• Collected/accessible Europe-wide on similar basis
Why study the LFS?
• Labour Force Surveyo very widely used for analysis and policyo long-running and well understoodo very standard surveyo common-ish across Europeo riddled with measurement error…
Strategy:
• What does the wage distribution look like?
• Do observed wage values differ from their true values?o Where do they differ? o How do they differ?o Can we suggest why they differ?o Is it important?
• Method:o Focus on low earningso Triangulate with other surveyso Test responses against psychological modelso Evaluate ad hoc and graphical analyses
Why low earnings?
• Compressed wages
• Variety of wage payment schedules
• Minimum wageo known true valueo anchor
What do wage distributions look like?Weekly wage: £300 £400 £600£326 £383 £575
Salary: £17k £20k £30K
Rounding of wages in ASHEAbsolute wage
Highest factor, in pence 5 10 25 50Expected frequency 20% 10% 4% 2%
Observed frequency
All size bands 35% 22% 15% 10%
0‐9 employees 51% 35% 32% 24%
10‐49 employees 48% 32% 24% 18%
50‐249 employees 39% 25% 17% 12%
250+ employees 30% 17% 9% 7%
Hourly wages, within £1 of the minimum wage
What determines rounding?
• size of firm (smaller => more rounding)
• public/private (private sector => more rounding)
• unionised workplace (=> less rounding)
• some industries (retail less rounding, hospitality/cleaning more)
LFS and ASHE compared
Rounding of wages in LFSASHE LFS
Highest factor, in pence 10 50 10 50Expected frequency 10% 2% 10% 2%
Observed frequency
All size bands 22% 10% 55% 31%
0‐9 employees 35% 24% 64% 42%
10‐49 employees
32% 18% 57% 32%
50‐249 employees
25% 12% 49% 25%
250+ employees
17% 7% 43% 21%
Hourly wages, within £1 of the minimum wage
What determines rounding in LFS?
• Size of firm (smaller => more rounding)
• Being in a low-paying sector (esp retail, hospitality, cleaning, food)
• Source of datao proxy response – much more roundingo payslips use – much less
• er…
• That’s it
ASHE vs LFS on rounding
• Firm size effect in both – genuine? – but larger in LFS
• LFS – dominated by use of documentation (or not)
• Otherwise, random in LFS
Can we be more specific?
Can we be more specific?
Can we be more specific?
Can we be more specific?
Can we be more specific?
Can we be more specific?
Can we be more specific?
• Yes
• Can compareo known distributions from ASHEo Minimum wages
o Outcomes very predictable
• Results same acrosso time (same values in different years)o scale (same penny value at different pound values)o surveys (similar findings in BHPS)o concept (hourly wages vs weekly vs salary)
Summary of LFS wage error
• Distinctive (triangulation with ASHE)
• Predictable (anchored to MW and other focus points)
• Consistent (across time, scale, surveys)
• Psychologically well-foundedo (apart from the rounded derived wage)
Does it matter?
• In theory, yes
Does it matter?
• In practice…
• Impact on simple measures substantialo Ritchie et al (2016, 2017, 2018) – large impact on non-compliance estimateso LPC definition of “low pay worker” as “minimum wage +5p”
Does it matter?
• In practice…
• Fry and Ritchie (2012) – fiddled values around NMW to reflect ASHE distribution
• Le Roux et al (2013) – adjusted for rounding in compliance models
• For this papero estimation with wage as dependent variable + adjustment factorso IV estimation with wage explanatory
• Impact:
•o None worth talking about
Conclusion
• Measurement error in LFS wages easily found and explained
• Simple descriptives: substantial impacto ‘hard’ boundaries (eg min wage) being crossedo Error is not mean zero
o Can be managed
• Marginal analyses: practical impact smallo precision less important
Next steps
• more modelling with RHS wage variables
• comparison with German/Polish/other data
• ..?
Thank you