PowerPoint pateikties šablonas (anglų k.)

Download Report

Transcript PowerPoint pateikties šablonas (anglų k.)

ПРОБЛЕМЫ ПЕРЕСЧЁТА
КВЕД 2005 – КВЕД 2010
Bronislava Kaminskienė
www.stat.gov.lt
OUTLINE




Background
Re-coding from NACE 1.1 to NACE 2
Why do we need to backcast
How to backcast time series
 Using micro records
 Using macro records
 Seasonal adjustment issues
 Conclusions
www.stat.gov.lt
NACE 2 Implementation
 Statistics Lithuania omitted to publish NAACE 2 based estimates
 annual surveys : from1998 reference year on
 sub-annual surveys:
 first data in 2000
 last ones by 1998
 N.B. For many sample surveys this meant redesigning the survey and/or draw a new sample on a
NACE 2 basis.
www.stat.gov.lt
TRANSITIONS FROM OLD TO NEW
One-to-one
NACE 1.1
NACE 2
216 industries
www.stat.gov.lt
TRANSITION FROM OLD TO NEW
Many-to-one
NACE 1.1
NACE 2
69 industries
www.stat.gov.lt
TRANSITION FROM OLD TO NEW
One-to-many
NACE 1.1
NACE 2
www.stat.gov.lt
TRANSITION FROM OLD TO NEW
Many-to-many
NACE 1.1
NACE 2
693 industries
www.stat.gov.lt
Why do we need to backcast time series on NACE 2 basis?

For annual series:
 to provide historical growth rates
 For sub-annual series:
 to enable seasonal adjustment
www.stat.gov.lt
Alternative backcasting methods

Using re-coded micro records:
 domain estimation
 can be costly
 high CV-s
 applied successfully to LFS (they had domain estimation before as well)
 Using concordances at macro level:
 applied for most series at Statistics Canada
www.stat.gov.lt
Example of backcasting using micro records
 Double coded all records in sample in 2005 and 2008
 Used hot deck imputation for NACE code:
 recipient matched to donors on class of worker, province, sex, age, education
 only re-imputed if code changed
 donor was more likely from deck closer in time
 After imputation obtained domain estimates
 Quality of historical series excellent.
www.stat.gov.lt
BACKCASTING AT MACRO LEVEL USING CONCORDANCES
Assume at year t frame is double coded according to both NACE 1.1 and
NACE 2.
Yt : the population total at year t of a variable
(e.g. shipment)
Y ht : the total in industry class h, h=1, . . .,H (NACE 1.1)
Y gt : the total in industry class g, g=1, . . .,G (NACE2)
www.stat.gov.lt
CALCULATING CONCORDANCES
where:
shipment in i-th establishment in
industry h, i=1,. . .,Nh
shipment in i-th establishment in
industry g, i=1,. . .,Ng
www.stat.gov.lt
CALCULATING CONCORDANCES
For year t we can calculate concordance
coefficients
Then
can be obtained as a weighted
sum of
Typically
is zero for most industries h.
www.stat.gov.lt
CLASSIFYING CONCORDANCES
1. One-to-one mapping
If for a given g, chg equals 1 for only one industry h
and equals 0 for the rest.
2. Many-to-one mapping
For a given g, chg takes the value 1 or zero only.
3. One-to-many mapping
For a given g, 0 < chg < 1 for only one industry h
and it is zero for the rest.
4. Many-to-many mapping
For a given g, 0 < chg < 1 for at least two industries h.
www.stat.gov.lt
MACRO LEVEL BACKCASTING
Concordance coefficients
were calculated based on
variable Yt on the frame (value added) and applied to variable Xt (e.g.
operating revenue).
Then for annual estimates:
error due to using
to obtain
instead of
Furthermore, these coefficients are also applied to
previous years: t - 1, t - 2, …, t - n, introducing further error:
error due to using concordance from year t
instead of year t - n
www.stat.gov.lt
MACRO LEVEL BACKCASTING
For quarterly estimates:
k = 1, . . ., 4
Error due to using
to obtain
Sampling error of
Error due to using the same chg for all
quarters
www.stat.gov.lt
MACRO LEVEL BACKCASTING
Error is reduced somewhat by benchmarking
the quarterly estimates
to annual totals yielding
satisfying:
www.stat.gov.lt
CONSEQUENCES OF USING MACRO BACKCASTING
 Four types of errors contributed to:
 erratic intra-annual movements in quarterly data;
 historical pattern not similar to present for some.
 Seasonal adjustment quality suffers.
 Could evaluate by comparing concordance based 2008 NACE 2 estimates with true NACE2
estimates from redesigned survey.
www.stat.gov.lt
MACRO LEVEL BACKCASTING
 Calculated concordance coefficients for years: 2005, 2006,….2010 (resistance rules).
 Calculated separate concordance coefficients for 3 variables
 Dropped coefficients below 0.001 and re-scaled.
www.stat.gov.lt
Quality of NACE2 estimates
 None of the four types of errors are present.
 Annual NACE 2figures should be correct, unless there was some miscoding.
 “Strange” growth rates in some industries is evidence of miscoding.
 Overall quality good
www.stat.gov.lt
Possible remedies to problems in backcasting
 Use interpolated monthly concordances consistent with the yearly ones to eliminate
December to January jump.
 Do multivariate benchmarking to yearly totals forcing production to be positive.
 Use micro approach, that is: transfer the double codes to the microfile and produce
historical domain estimates for NACE 2. Post-stratify to known industry totals (or
benchmark). Large CV-s ???
www.stat.gov.lt
Correcting the historical estimates
 Could macro convert NACE 1.1 estimates to NACE 2 and compare to true NACE 2
based.
 Correction factors could be calculated for the history of the series based on average
monthly discrepancies to improve seasonal pattern.
www.stat.gov.lt
Issues when seasonally adjusting NACE 2 series
 Expect more volatile series than before.
 One measure of NACE 2 conversion quality: % of series suitable for seasonal adjustment
before and after.
 Apply shorter seasonal moving averages to pick up new pattern faster.
 Revisit series after three years and adjust historical estimates to be more in line with
recent seasonal pattern.
www.stat.gov.lt
Conclusions
 Two approaches for backcasting
 Micro approach
 costly, not always feasible
 suitable for some series
 resulting domain estimates can have high CV-s
 Macro approach
 can introduce four types of errors
 best if concordances are based on
 the variable to be estimated
 separate concordances per year
 separate concordances per month
 historical trends, seasonality could be distorted
www.stat.gov.lt
Conclusions
 Some corrective action can be taken:
 interpolate monthly concordances;
 use multivariate benchmarking;
 if both synthetic and true NACE 2 series exist for several years:
 apply correction factors based on discrepancy;
 revisit NACE 2 series after three years and modify the historical seasonality.
www.stat.gov.lt