A comparison of sample and register based survey:
the case of labour market data
De Gregorio C., Filipponi D., Martini A., Rocchetti I.
Contents
Survey(LFS) – ADMIN
Strategic issue
Previous ESS research
Long term innovation process
Our purposes
Answers and new questions
Innovation leverage in several fields
Microdata LFS vs. ADMIN
Integration: labour input measurement
Definition of employment, Regular vs Irregular
First: employment status comparison
ADMIN wrt:
LFS reference week, Employed and Self-employed.
Our purposes
Managing inconsistencies between LFS and ADMIN
Measuring Regular and Irregular employment
Assessing Accuracy of LFS and ADMIN (Assumed error models, MSE’s derivation and computation, No
considered benchmark )
Estimating ADMIN Over-coverage (precision)
Estimating ADMIN Under-coverage (irregular)
Estimating LFS Under-coverage (understatement)
Our model: LFS sample
“True” status
REGULAR
IRREGULAR
NOT EMPLOYED
“ADMIN employed” status
“LFS employed” status
Inconsistencies
EMPLOYED Not employed TOTAL
EMPLOYED
Not employed
TOTAL POP >15
LFS
ADMIN
REGULAR IRREGULAR
Our model
• Hypotheses (to simplify)– If LFS employed then employed– If True Regular then ADMIN employed– No LFS Non-response or substitution bias– ADMIN exhaustive and with no error– No problems with record linkage
• Key estimates– Probability of being truly employed if “ADMIN employed”– Rate & number of LFS false negatives– Probability of being truly employed if “LFS not employed”
• Assume it’s OK!
• Compare LFS and ADMIN MSE• Error model for LFS employment status (z) given
the true employment status (y)
kkkkkk yyyz )1(
)1()1( ,01,10 kkkkk yyx
• Error model for ADMIN employment status (x)
),1Pr( kk and
and
),1Pr( ,10,10 kk ADMIN under-coverage (irregular employment)
)1Pr( ,01,01 kk ADMIN over-coverage (false employment signal)
MSE by domain
N
Y
N
Y
N
N
n
fNY
n
NYtMSE S 1
1
1)1()ˆ( 222
LFS
ADMIN
210010101011010 )()1(1)ˆ( YNYNYtMSE A
>95% of total MSE - given “true” employment, population and sample size
Linear locus of “low impact” on MSE
• LFS MSE: depends on the probability of under-coverage• ADMIN MSE : balance of two opposite errors
DOMAIN λ θ01 θ10 LFS ADMIN
TOTAL 5.1 6.2 9.5 46.7 5.1 2.4
NW 3.6 3.9 6.8 51.3 3.6 3.1
NE 3.4 6.5 6.1 52.7 3.4 0.3
CE 4.1 5.3 10.3 49.3 4.1 4.8
SO 8.8 7.9 14.2 38.4 8.8 1.5
ITA 5.0 6.3 8.7 45.4 5.0 1.1
EU 4.8 2.8 21.1 70.3 4.9 19.9
XEU 6.6 4.3 15.6 63.9 6.7 13.2
M 4.5 8.8 9.2 57.5 4.5 2.7
F 6.1 4.6 9.9 36.8 6.1 2.0
NU
TS 1
CIT
IZ.
SEX
Empl. rate
CV
• LFS & ADMIN both have errors• LFS has sampling and under-coverage errors• Apparently ADMIN performs better, as the
sources of errors tend to compensate• ADMIN worsens in the domains with higher
irregularity rates• ADMIN produces higher errors at micro-level• For analysis purposes, survey and ADMIN data
should be integrated further• An efficient usage of exhaustive ADMIN data
should count on survey based estimates of actual employment status
To conclude