Recent work on revisions in the UK

1
Recent work on revisions in the UK

• Robin Youll
• Director
• Short Term Output Indicators Division
• Office for National Statistics
• United Kingdom

2
Overview

• Scene setting
• Why are we concerned with revisions?
• Will cover three recent developments in the UK
• Revisions triangles as a tool for identifying
  causes of revision
• News v . Noise debate
• Linking revised and historical series

3
Why are we concerned with revisions?

• .after all they are a necessary part of the
  statistical process
• Two reasons to be interested
• 1. Systematic revisions
• Bias
• but beware of time dependency
• difference between tendency to revise up and
  always revising up
• Can benchmark annually, but may undermine purpose
  of STS?
• Variance
• increasing over time? news v noise (see later)
• 2. Dating the cycle
• false/missed turning points policy regret
• Credibility/reputation of NSIs

4
Development 1.Taking a longitudinal view of
revisions the power of revisions triangles

• Typically, revisions analysis looks at point to
  point revisions.
• e.g. mean revision between time t and t12.
• Revisions triangles can be used to view the
  history of particular revisions (see-SLIDE)
• The longitudinal view can help us to understand
  the causes of revision.
• e.g. is there a tendency for first estimates (t)
  to be close to previous estimates (t-1)?
• In the UK we now monitor revisions using this
  longitudinal approach.
• Main reason for each significant revision is
  recorded according to a typology (late data,
  seasonal adjustment, revisions to trend
  component, benchmarking, error correction, etc.)
• The has allowed the identification of systematic
  revisions arising from the compilation process
  (rather than being data driven)

5
Revisions Triangle (real time database)
6
Recreation - New Data
Recreation New Data affecting SA
Education - Industry Review
Recreation - New Data
Government and other services 2002 Q2
Education - ACAs
Health Social Work - New Data Methodology
7
Development 2. News v. Noise .

• So what can we do to reduce revisions?
• Our approach depends on our belief about the
  underlying process which causes them.
• Two schools of thought on this sometimes called
  the News v. Noise debate
• Broadly, later vintages become more accurate by
  either
• Eliminating noise
• Incorporating news

8
News v. Noise Noise

• 1. Noise hypothesis
• Here, preliminary estimates are simply noisy
  versions of the truth
• and later vintages of data become more accurate
  by eliminating measurement errors
• Solution apply filtering techniques to
  preliminary data to extract the truth from the
  noisy series.

9
Noise the theory

• 1. Noise Hypothesis
• Here preliminary estimate is equal to a later
  vintage ( vintages later) plus a measurement
  error ( ).
• Revisions are uncorrelated with but
  correlated with
• i.e. there is information in the error term that
  isnt in the current vintage.
• Test using the regression
• Ho
• Also the variance of different vintages should
  fall over time, so that
• for all

10
News v. Noise News

• 2. News hypotheses
• assumes early vintages do not reflect the fully
  available information at that time.
• Gains are potentially available by using this
  information (e.g. external surveys).
• Solution find and incorporate extra information
  available at the time the preliminary estimate is
  made

11
News the theory

• 2. News, or efficient forecast hypothesis
• Here later vintages ( vintages later) equal
  earlier vintage plus a measurement error (
  ).
• Revisions ( ) are correlated
  with but uncorrelated with
• i.e. there isnt any information in the error
  term that isnt already incorporated in the
  current vintage.
• Test using the regression
• Ho
• Also the variance of different vintages should
  rise over time, so that
• for all

12
News v Noise results in the UK

• Standard Deviation of different data vintages
  (1993q1-2002q4)
• vintage GDP ISP IIP
• 25 days 0.251 0.211 0.856
• 8 weeks 0.267 0.211 0.902
• 12 weeks 0.251 0.229 0.902
• 1 year 0.296 0.266 0.913
• 2 years 0.309 0.290 0.814
• So, in the UK filtering is appropriate to the ISP
  and GDP (since they conform to the News
  hypothesis, and
• seeking alternative data sources is appropriate
  to the IIP, which conforms more closely to the
  Noise hypothesis.

13
(No Transcript)
14
Development 3.Linking revised and historical
series

• ONS has fairly structured policy on when to
  publish revisions.
• This can often delay publication of known
  revisions for years, because
• Logistics of taking on revisions (particularly
  for estimates linked to national accounts, e.g.
  ISP)
• Minor revisions irritate users
• But, approach to linking revised to historical
  series can distort growth rates spanning the
  link period
• especially if growth into the link period has
  been significantly revised
• An example

15
(No Transcript)
16
(No Transcript)
17
(No Transcript)
18
Revisions policy linking on revisions

• For Monthly series this effect is most noticeable
  for 3-month on 3-month growth rates based on
  index levels which span the link period
• For Quarterly series, quarter on same quarter a
  year ago growth rates are most affected.
• Solution adopted in the UK
• Identify the oldest revision to growth greater
  than some value (say 0.2 points), and link growth
  from that period on (Nov 05 in our example).

19
(No Transcript)
20
(No Transcript)
21
Recent work on revisions in the UK

• Robin Youll
• Director
• Short Term Output Indicators Division
• Office for National Statistics
• United Kingdom

22
Example of News and Noise Hypotheses

• Examples of News v Noise.xls

23
Example of News Hypothesis
24
Example of Noise Hypothesis