Transcript Annex B5.13 Business Sampling

Business Sampling
25th April 2012
Peter Linde
Survey and Methods
Statistics Denmark
Methods
Methods:
•
•
•
•
•
Support whole Statistics Denmark
A part of business statistics, where most work are
Steering committee – four Directors
6 persons working med sampling, SA, imputation, …
Optimizing the 30 business surveys one of the tasks
… inspection every 3-4 year
… when new sources
… increasing sampling error
Business register
•
•
•
•
•
•
•
•
•
Business register is the frame for all business Statistics
It is updated by business, Statistics Denmark and official authorities
Statistics Denmark is e.g. responsible for sector code
When at statistics discover a problem it contact the unit working
with Business register
Business in general has a low interest in sector code, but
sometimes at interest not to be a concrete sector because of the
price for insurance
A business can cover several sectors – is this updated?
Then Business register is updated with active business. The under
or over cover is very small – that’s is enough to make statistics
Problems with sector must be solve in the estimation
Metadata about change over time
Sampling and estimation business surveys
Principles for optimizing samples
•
Least possible mean error
•
99% participation - mandatory
•
No bias – simple random sampling inside strata
a) Panels must not give bias
b) Two steps when calibrating weights
1) Against the selection population
2) Against the known population when published
•
Burden for small companies
•
Medium size companies participate at most 3/5 years
•
Adjust for cut off
Two steps when calibrating weights
Two steps when calibrating weights
• Step one: The known population when sample is
selected
• If e.g. sector is not correct information is sent to
business register
• All other surveys also send back if something is wrong
• In addition, companies and other offices can update
the business register
• Step two: When the statistics is published a locked
version is used for all statistics with the accepted
updates
• Adjust for cut-off by ancillary information
Panels must not give bias (1)
• Panels are used to optimize estimation of changes and to
reduce sampling costs
• If a company is changing to a smaller size strata - with a
smaller selection probability - a proportional numbers of
these companies are given free
– and opposite when companies grow
Else an underestimation will arise when the economy grows
- and opposite an overestimation when the economy is decreasing