ESSnet on Small Area Estimation - CROS

advertisement
ESSnet on Small Area
Estimation
Stefano Falorsi
Istat
ESSnet on Small Area Estimation
When and How SAE?
When to use SAE methods:
Whenever direct estimators which are based only
on sampling units observed for each small area are
not reliable (small sample size or sometimes even
no observed units)
i.e. CVs or other measures of sampling variability
of direct estimates are considered to be too high
for the target indicator at area level
How to use SAE methods:
Exploiting relationship among variables by means
of explicit or implicit modeling
SAE: Borrowing Strength from?









Partners of ESSnet SAE
project
Istituto Nazionale di Statistica
Institut National de la Statistique France et des Etudes Economiques
Statistisches Bundesamt
Centraal Bureau voor de Statistiek
Statistisk Sentralbyrå
Główny Urząd Statystyczny
Instituto Nacional de Estadística de España
Office for National Statistics
Swiss Federal Statistical Office
contacts :
Stefano Falorsi : coordinator of the project
stfalors@istat.it
saessnet@istat.it
Denisa Florescu : Eurostat
denisa.florescu@ec.europa.eu
webpage : http://www.ess-net-portal.eu
(ISTAT)
(INSEE)
(DESTATIS)
(CBS)
(SSB)
(GUS)
(INE)
(ONS)
(FSO)
Aims of the ESSnet on SAE
The activity of the ESSnet - SAE was directed to
analyze methods and previous experiences on
SAE, focusing on the application of SAE methods
for social surveys.
The aim was the promotion of the use of SAE
methods in the production of statistical
information, through the standardization of SAE
estimation process.
The activities of the project were grouped into
work packages in order to ensure the best
interaction and mixing of know-how among the
partners.
Outcomes of the ESSnet on SAE
The project was planned in phases which are a series of
theoretical and application activities in order to
facilitate and promote the use of small area techniques
in the production of statistical information
The results of the ESSnet-SAE project will be helpful
to detect the best practises
to define the guidelines to be followed for the applications of
SAE
 to give some advices to software tools users.


Dissemination of results was another important step of the
project, therefore a web-site is available in order to
share information and the results among the partners
be a forum platform in order to boost the communication among
NSIs willing to apply SAE methods.


Description of the Project
The project was composed by 7 work-packages:

WP1 - Project management
WP2 (GUS,Destatis, INSEE, INE, ONS) State of the
art

The WP2 was aimed to provide a comprehensive overview of
small area estimation in the ESS social surveys with respect to
implementation, needs and expectations.
Last ten years SAE literature and other SAE project outcomes
have been reviewed.
Moreover a survey on NSI’s SAE experiences has been carried
out.
Description of the Project
WP3 (INE, Istat, CBS, INE, GUS, ONS) Quality
assessment

This work-package aimed to review and develop suitable
criteria to assess the quality of SAE methods. In details
methods for:
top-down assessment (from larger to smaller domains)
determination of accuracy thresholds
choosing among different set of auxiliary information
model diagnostic (model fitting, error analysis, etc.)
overall quality assessment by means of internal validation to
check for bias of SAE estimates or by means of external
validation using already available information (for instance
Census data)





Description of the Project

WP4 – (Istat, CBS, GUS) Software tools
This WP was aimed to provide recommendations for
standardization and certification of ESS software tools for
SAE. For this reason, a study was devoted to analyze the real
capacity of routines to be applied to large scale surveys
(characterized by a large number of records and domains).
New software routines have been developed in order to
provide an integrated set of software tools to produce small
area estimates for large scale social surveys.
The need of sharing as much as possible the outputs of this
WP implied the use of free licensed software and this was the
main reason for the choice of the open source software R to
develop new software routines.
Description of the Project

WP4 – (Istat, CBS, GUS) Software tools
The main deliverable of the work-package is a collection of R functions to
perform SAE estimation, model selection and model diagnostic.
The R functions will include :
 Estimation :
o unit level : Synthetic, EBLUP (uncorrelated random effects), EBLUP (correlated
random effects)
o area level : Synthetic, EBLUP (uncorrelated random effects) unit level (EBLUP
type)
o logistic mixed model
Model choice :
o conditional AIC and cross validation (unit and area level linear mixed model) unit
level : Synthetic, EBLUP (uncorrelated random effects), EBLUP (correlated random effects)
 Model diagnostic :
o Bias diagnostic, Goodness of fit diagnostic, Coverage diagnostic, Calibration
diagnostic
Description of the Project
 WP5 (CBS – FSO, GUS. INSEE, INE, Istat, SSB) Case
studies
The activities of this WP focused on applying significant SAE
methods acknowledged in WP2 and relevant tools for model
diagnostic, model selection, quality assessment identified in WP3.
The case studies will also be the ground for training the software or
routines developed in WP4.
The NSIs’ case studies focus on:
- CBS:
Crime survey
- FSO: New Census strategy: simulation study from previous census
- GUS: Labour Force survey
- INE: New Census strategy: simulation study from previous census
- INSEE: Labour Force survey
- ISTAT: Health survey
- SSB: Register of deaths: modelling mortality rates
Description of the Project

WP6 (Istat – CBS, FSO, GUS, INE, ONS, SSB)
Guidelines
This work-package aimed to summarize the activities and the
results produced in the previous WPs and to provide
practical guidelines in ESS social surveys context. The
guidelines contain:
-a description of the process that should be followed when
applying SAE methods;
- standard SAE methods and their main extensions;
- main tools to assess the quality of the estimates;
- recommendations about the use software and routines to
be used to produce small area estimates.
Main WP6 Outcomes: Guidelines
Main WP6 Outcomes: Guidelines
Description of the Project
WP7 - (Istat - CBS - Destatis, FSO, GUS,
Insee, INE, SSB, ONS) Transfer of knowledge
and know-how

The WP7 aimed to transfer knowledge and know-how to
non participating NSIs and to disseminate the results
through a course and three trainings on the job.
A website devoted to the ESSnet-SAE project was designed
within the ESSnet portal.
All the final reports and software tools are available at the
ESSnet portal http://www.essnet-portal.eu/.
Future work?
Issues on SAE not exploited in the project
but of great interest
The ESSnet-SAE covered only univariate crosssectional methods. The following items were
let aside:

strategies for SAE at the design stage
time correlation for repeated surveys
multivariate responses
SAE for business surveys (now research activity
in WP6 of BLUE-ETS and one of the topic covered
by the MEMOBUST project)




Future Istat’s work
Relevant Istat’s works in progress in this field are:
 is under study an extension of web SMART, SMall ARea
estimation Tool (http://smart.istat.it/smart/) to include variables
related to income and poverty. The tool is available only for LFS ILO
variables: each user may perform an online application of SAE
estimates for municipalities or their aggregations. The level of
aggregation is defined interactively by the user. The tool do not
produce official statistics but only domain estimates useful for
research or policy.
 Istat is conducting a study on
“Small area estimation of poverty rate at different territorial
dissaggregations”
 The empirical performances of different model based SAE
estimators will be compared using EU-SILC data: ELL method
(Elbers et al. 2005), EB (Molina and Rao 2010), EBLUP unit level
with spatial correlation, furthermore Modified Direct and Bias
corrected projection estimators will be considered too;
 The work will compare the results of the above estimators for
different domain levels: NUTS3, LAU1, LAU2.
Download