L. Di Consiglio, S. Loriga, A. Martini, R. Ranaldi - The Italian LFS sampling design: recent and future developments
1. 9th
Workshop on Labour Force Survey Methodology – Rome, 15-16 May 2014
The Italian LFS
sampling design:
recent and future
developments
9th
Workshop on Labour Force Survey Methodology
Rome, 15-16 May 2014
Loredana Di Consiglio
Silvia Loriga
Alessandro Martini
Rita Ranaldi
2. 9th
Workshop on Labour Force Survey Methodology – Rome, 15-16 May 2014
Two-stages with
stratification of the PSUs (municipalities) in the
first stage
rotation of the FSUs (households) in the last
stage
IT-LFS sampling design /1
3. 9th
Workshop on Labour Force Survey Methodology – Rome, 15-16 May 2014
At the first stage …
Municipalities are stratified at NUTS III level according to
resident population and they are divided into two groups:
SR (self-representative) municipalities:
have larger demographic size (over a given threshold)
each represents one stratum
always selected in the sample
NSR (non self-representative) municipalities:
have smaller demographic size
are stratified in groups (strata) having almost the same total
population
only one municipality for each stratum is selected in the sample,
with probabilities proportional to its demographic size
IT-LFS sampling design /2
4. 9th
Workshop on Labour Force Survey Methodology – Rome, 15-16 May 2014
At the second stage …
Households are randomly selected from the population
registers of municipalities drawn at the first stage
FSUs are rotated according to a 2-(2)-2 rotation scheme.
Households are interviewed during two consecutive
quarters. After a two-quarters break, they are again
interviewed twice in the corresponding two quarters of
the following year
IT-LFS sampling design /3
6. 9th
Workshop on Labour Force Survey Methodology – Rome, 15-16 May 2014
Leaving unchanged the general structure of the sample,
in 2012 sampling design has been revised for the
following main reasons:
the previous sample was designed in 2001-2002,
considering the target variables estimated at that time by the
quarterly LFS and the frame information for stratification
referred to 2002
several changes occurred in the boundaries of the
administrative units such as municipalities and provinces
to further improve the monthly representativeness of the
sample, considering the high relevance of monthly LFS
estimates
budget constraints forced to reduce the sample size
Sampling design revision /1
7. 9th
Workshop on Labour Force Survey Methodology – Rome, 15-16 May 2014
Every quarterEvery quarter
71,536 theoretical71,536 theoretical
sample householdssample households
(average sampling(average sampling
rate: 1/350)rate: 1/350)
Sample size after revision
allocatedallocated
in more than 1,000in more than 1,000
sample municipalitiessample municipalities
(about 1 out of 7)(about 1 out of 7)
8. 9th
Workshop on Labour Force Survey Methodology – Rome, 15-16 May 2014
The following methodological and operational
constraints have been taken into account:
Eurostat precision requirements (Reg.577/98), but also additional
constraints for national purposes have been considered
the unemployment figures considered as target variables for the
evaluation of precision requirements are referred to the pre-crisis period
the information on non responses has been considered when distributing
the sample units among the territorial units
the monthly distribution of the sample guarantees that each month is
representative of the whole national territory
the new selected PSUs have to overlap as more as possible with the
previous PSUs in order to minimize the impact on the fieldwork (and on the
final estimates)
a random rotation of the PSUs has to be applied every year to maintain
the sample updated over time
Sampling design revision /2
9. 9th
Workshop on Labour Force Survey Methodology – Rome, 15-16 May 2014
Distribution of the sample over space
Because of the national precision requirement about unemployment
estimates in NUTS 3 domains, the distribution of the sample is not
proportional to the demographic size of the domains
NUTS 3 domains
ITALY
MIN MEAN MAX
Resident households
(N)
24,779 231,841 1,769,720 25,502,535
Unemployment rate %
(2004-2007)
2.56 7.29 18.50 7.16
Sample size (n) 192 650 3,408 71,536
Inclusion probabilities
(n/N%)
0.12 0.39 3.94 0.28
Minimum, mean value and maximum of resident population, unemployment rate,
quarterly sample size and inclusion probabilities in NUTS 3 domains
The sample deviates from the optimal sample we should have
obtained considering just Eurostat NUTS 2 constraints. In any case,
Eurostat constraints are satisfied.
10. 9th
Workshop on Labour Force Survey Methodology – Rome, 15-16 May 2014
The quarterly sample size is uniformly distributed among the
13 weeks, each stratum is observed at least in 3 weeks per
quarter and the monthly representativeness of the
sample is guaranteed
The largest PSUs are in the sample all the 13 weeks of the
quarter
Other PSUs (among them also some chief towns at NUTS3
level) are in the sample just 3 weeks per quarter, assigning
them reference weeks that are triplet of weeks in which the
distance between them is 4 weeks: 1-5-9 or 2-6-10 or 5-9-13
and so on
Distribution of the sample over time /1
11. 9th
Workshop on Labour Force Survey Methodology – Rome, 15-16 May 2014
The months are not fixed, but they are composed by a
number of weeks that is variable (4 or 5) and depends on the
number of Thursdays falling in each solar month
Distribution of the sample over time /2
Possible
combinations
Weeks
1 2 3 4 5 6 7 8 9 10 11 12 13
4-4-5
4-5-4
5-4-4
Month 1
Month 2
Month 3
Week 5 may be included into months 1 or 2 and week 9 may
be included into months 2 or 3
Some PSUs, to which the weeks 5 or 9 have been assigned,
may fall into different months
In the sample revision we guaranteed that the chief towns
in NUTS 3 domain, observed just 3 weeks per quarter, are
not to be observed neither in week 5 neither in week 9
12. 9th
Workshop on Labour Force Survey Methodology – Rome, 15-16 May 2014
Aim: to minimize the impact on the fieldwork (changing
all the municipalities, or the majority of them, would have
meant to recruit and to train a lot of new interviewers,
with evident effects on the fieldwork and risks on the
quality of the final estimates)
Method: use of Permanent Random Numbers (PRN)
applying the method suggested by Ernst (2004)
Results: 831 municipalities, about 75% of the PSUs
selected according to the new design, overlapped with
the previous PSUs
Maximum overlapping of old and new PSUs
13. 9th
Workshop on Labour Force Survey Methodology – Rome, 15-16 May 2014
Aim: to maintain the sample updated over time and to
reduce the statistical burden, in particular for the
households living in municipalities with a small number
of residents
Method: Probabilistic rotation by applying permanent
random numbers (PRN) and constant shift method
(Brewer et al. 1972, Ohlsson 1995)
Results: in 2014, 143 municipalities have been rotated,
about 13% of the PSUs that were sample in 2013,
almost all the municipalities with less than 1,000
inhabitants and nearly three out of four municipalities of
1,001-2,000 inhabitants
Rotation of PSUs of small demographic size
14. 9th
Workshop on Labour Force Survey Methodology – Rome, 15-16 May 2014
The new sample has been gradually introduced starting
from the first wave of 2012Q3. For 5 quarters, until 2013Q3,
old and new sampling designs were overlapped
The estimation and analysis of variance procedures have
been reviewed assuming that the two different sub-
samples are independent
The comparability of the accuracy between the two
designs is not simple for the wide variations in the estimates
due to the current economic situation and to usual seasonal
effects as well observed in this period of 15 months
The analysis was conducted using regression models that fit
sampling errors, in order to obtain estimates of sampling
errors independently by the observed phenomena, even with
an approximate evaluations of the errors
Accuracy evaluation of the two sample designs /1
15. 9th
Workshop on Labour Force Survey Methodology – Rome, 15-16 May 2014
Accuracy evaluation of the two sample designs /2
Coefficient of Variation- IT-LFS Whole Country
0
5
10
15
20
25
30
35
40
5000 35000 65000 95000 150000 220000 340000 550000 850000 1300000 2000000
Old Sampling Design New Sampling Design
Graph 1 - Comparison of regression model of sampling
errors for the new and old sample design of IT-LFS
16. 9th
Workshop on Labour Force Survey Methodology – Rome, 15-16 May 2014
Accuracy evaluation of the two sample designs /3
Graph 2 - Difference of coefficient of variation between old
and new IT-LFS sample design by Nuts II level
17. 9th
Workshop on Labour Force Survey Methodology – Rome, 15-16 May 2014
Istat has undertaken its process of renewal with the:
transition to CAPI mode of several PAPI surveys
integration of the Trips and Holidays Survey as module into HBS
introduction of web in the surveys on PHD and on high school
graduates
Carrying out several CAPI sample
households surveys together with
the new Population Rolling Census
makes necessary to develop a
coordinated approach to obtain
harmonized sampling designs
and to optimize the distribution of
the sample over space and time
Future perspectives
18. 9th
Workshop on Labour Force Survey Methodology – Rome, 15-16 May 2014
Thank you for your attention