Revisions Analysis at Statistics Canada October 10, 2008
Statistics Canada • Statistique Canada 2003-4 CSNA RP Toolkit Why Study Revisions? Statistical users and providers are both concerned about data quality. Main dimensions of quality: relevance, coherence, timeliness, accuracy and reliability. Preliminary estimates are provided primarily for timeliness, while final estimates are deemed to be accurate or the most accurate possible. As well as being timely, preliminary estimates should be reliable or are close approximates of final estimates. Revision studies aim to measure the reliability of the preliminary (and intermediate) estimates. The focus is on revision patterns and their change over time. Revision studies are in international and Statistics Canada’s data quality assessment frameworks. Usually preliminary estimates are revised a number of times and each round of revision is expected to improve the accuracy of the final estimates. Reliability of progressive vintages of the estimates would be expected to incrementally improved but this is not predictable in practice. A final estimate is deemed accurate if it is close to the true value of what’s being measured. Since the true value is never known, accuracy cannot be directly measured, especially for complex series like GDP which uses many source data. It is indirectly measure by assessing source data and methods. Estimates are considered 'final' when they have gone through the regular revision process over a number of normal updates as well as historical revisions. Preliminary (or intermediate) estimates are reliable if they closely approximate the final estimates. Reliability can therefore be directly assessed from revision characterisitcs (bias, dispersion, changes of revision patterns over time and so on). Revision studies aim to inform users of the reliability of the preliminary, or intermediate, estimates so they can use them with confidence. 3 Statistics Canada • Statistique Canada 17/11/2018
Statistics Canada • Statistique Canada 2003-4 CSNA RP Toolkit TYPES OF REVISION Revisions can be classified by reasons or by timing Reasons Incorporation of better source data Routine recalculation Improvements Correction of errors Timing Current revisions Annual revisions Historical (or comprehensive, major or benchmark) revisions More details: Carson, C.S., Khawaja, S. and Morrison, T.K. Revisions Policy for Official Statistics: A Matter of Governance. Sixteenth Meeting of the International Monetary Fund Committee on Balance of Payments Statistics, Washington D.C., December 1-5, 2003. BOPCOM-03/3A, pp 5-8. Reasons for revisions include the following: Incorporation of more complete or otherwise better source data including those with more complete or otherwise better reporting and those more closely match the concepts, as well as replacement with source data of judgement or of values derived largely by statistical techniques. Routine recalculation, for example incorporating updated seasonal factors and updating of the base period. Improvements. Changes in statistical methods, concepts, definitions, and classifications Correction of errors in source data and computation Timing-wise, revisions of a statistical series can be made at various times after a first release and can be classified as follows: Current revisions are made shortly after the first release and affect the current weekly, monthly or quarterly data Annual revisions are made after data for all the months or quarters of a year become available Less frequent revisions, often four or years apart, are “historical” revisions (also called “comprehensive”, “major” or benchmark”) In general, current revisions are mostly associated with the incorporation of better source data, while annual revisions would additionally take in routine recalculations. Changes in concepts, sources and methods usually bring on “Historical” revisions. “Historical revisions are often carried back, or backcast, for a large number of reference periods. Revisions to correct errors have no predictable timing but would be rare as normally errors are not expected to occur. 7 Statistics Canada • Statistique Canada 17/11/2018
FORMS OF REVISION STUDIES 2003-4 CSNA RP Toolkit FORMS OF REVISION STUDIES Revision studies at Statistics Canada have taken the following forms: Study on levels or growth rates Comparison between ranges of time periods (e.g. decades) Component profiles and impact of sub-aggregates Time profiles (different vintages) International comparisons Review of source data and estimation methods 8 Statistics Canada • Statistique Canada 17/11/2018
QUARTERLY GDP REVISION SCHEDULE RP0304 QUARTERLY GDP REVISION SCHEDULE The GDP revision cycle normally spans over four years so that estimates for the four quarters of Year t are finalised in the first quarter of Year t+4. For more details of the revision schedule, see 'The 1999-2002 Revisions of the National Economic and Financial Accounts' on Page 35 of Canadian Economic Accounts Quarterly Review, First Quarter 2003, Statistics Canada Catalogue No. 13-010-XIE. Prel. - Preliminary or first published Income and Expenditure Accounts estimates are released 60 days after the end of the reference period. Rev1 to Rev6 - First to sixth revision to the preliminary estimates. Benchmarks from the preliminary input-output tables are adopted in Rev6 estimates. Final - Final estimates, for which benchmarks from revised input-output estimates are adopted. These are the fourth to seventh revision of, respectively, the fourth, third, second and first reference quarters' preliminary estimates. Final estimates may be subject to further revision at rebasing, as well as at historical revisions. Historical revisions provide an occasion to improve estimation methods, eliminate statistical breaks resulting from more limited revisions and introduce conceptual changes into the system. Rebase since 1981: base years 1986, 1992 (third quarter 1997) and 1997 (first quarter 2001); Chain Fisher estimates were also introduced in the first quarter of 2001. Historical revisions since 1981: 1986 and 1997 Statistics Canada • Statistique Canada 17/11/2018 5
QUARTERLY GDP REVISION SCHEDULE RP0304 QUARTERLY GDP REVISION SCHEDULE REVISIONS IN THE PRELIMINARY YEAR PICK UP LATE SURVEY RETURNS AND RE-ESTIMATIONS AND UP DATES TO SEASONAL FACTORS The GDP revision cycle normally spans over four years so that estimates for the four quarters of Year t are finalised in the first quarter of Year t+4. For more details of the revision schedule, see 'The 1999-2002 Revisions of the National Economic and Financial Accounts' on Page 35 of Canadian Economic Accounts Quarterly Review, First Quarter 2003, Statistics Canada Catalogue No. 13-010-XIE. Prel. - Preliminary or first published Income and Expenditure Accounts estimates are released 60 days after the end of the reference period. Rev1 to Rev6 - First to sixth revision to the preliminary estimates. Benchmarks from the preliminary input-output tables are adopted in Rev6 estimates. Final - Final estimates, for which benchmarks from revised input-output estimates are adopted. These are the fourth to seventh revision of, respectively, the fourth, third, second and first reference quarters' preliminary estimates. Final estimates may be subject to further revision at rebasing, as well as at historical revisions. Historical revisions provide an occasion to improve estimation methods, eliminate statistical breaks resulting from more limited revisions and introduce conceptual changes into the system. Rebase since 1981: base years 1986, 1992 (third quarter 1997) and 1997 (first quarter 2001); Chain Fisher estimates were also introduced in the first quarter of 2001. Historical revisions since 1981: 1986 and 1997 Statistics Canada • Statistique Canada 17/11/2018 5
QUARTERLY GDP REVISION SCHEDULE RP0304 QUARTERLY GDP REVISION SCHEDULE REVISIONS IN YEAR T+1 ARE LEVEL REVISIONS RELATED TO BENCHMARK LEVELS FOR THREE YEARS EARLIER – FEW NEW GROWTH RATE INDICATORS ARE AVAILABLE BUT SEASONAL FACTORS ARE UPDATED The GDP revision cycle normally spans over four years so that estimates for the four quarters of Year t are finalised in the first quarter of Year t+4. For more details of the revision schedule, see 'The 1999-2002 Revisions of the National Economic and Financial Accounts' on Page 35 of Canadian Economic Accounts Quarterly Review, First Quarter 2003, Statistics Canada Catalogue No. 13-010-XIE. Prel. - Preliminary or first published Income and Expenditure Accounts estimates are released 60 days after the end of the reference period. Rev1 to Rev6 - First to sixth revision to the preliminary estimates. Benchmarks from the preliminary input-output tables are adopted in Rev6 estimates. Final - Final estimates, for which benchmarks from revised input-output estimates are adopted. These are the fourth to seventh revision of, respectively, the fourth, third, second and first reference quarters' preliminary estimates. Final estimates may be subject to further revision at rebasing, as well as at historical revisions. Historical revisions provide an occasion to improve estimation methods, eliminate statistical breaks resulting from more limited revisions and introduce conceptual changes into the system. Rebase since 1981: base years 1986, 1992 (third quarter 1997) and 1997 (first quarter 2001); Chain Fisher estimates were also introduced in the first quarter of 2001. Historical revisions since 1981: 1986 and 1997 Statistics Canada • Statistique Canada 17/11/2018 5
QUARTERLY GDP REVISION SCHEDULE RP0304 QUARTERLY GDP REVISION SCHEDULE REVISIONS TWO YEARS LATER INCLUDE EARLY BENCHMARK INFORMATION FOR WAGES AND SALARIES FROM T4 AND THE CAPITAL EXPENDITURE SURVEY The GDP revision cycle normally spans over four years so that estimates for the four quarters of Year t are finalised in the first quarter of Year t+4. For more details of the revision schedule, see 'The 1999-2002 Revisions of the National Economic and Financial Accounts' on Page 35 of Canadian Economic Accounts Quarterly Review, First Quarter 2003, Statistics Canada Catalogue No. 13-010-XIE. Prel. - Preliminary or first published Income and Expenditure Accounts estimates are released 60 days after the end of the reference period. Rev1 to Rev6 - First to sixth revision to the preliminary estimates. Benchmarks from the preliminary input-output tables are adopted in Rev6 estimates. Final - Final estimates, for which benchmarks from revised input-output estimates are adopted. These are the fourth to seventh revision of, respectively, the fourth, third, second and first reference quarters' preliminary estimates. Final estimates may be subject to further revision at rebasing, as well as at historical revisions. Historical revisions provide an occasion to improve estimation methods, eliminate statistical breaks resulting from more limited revisions and introduce conceptual changes into the system. Rebase since 1981: base years 1986, 1992 (third quarter 1997) and 1997 (first quarter 2001); Chain Fisher estimates were also introduced in the first quarter of 2001. Historical revisions since 1981: 1986 and 1997 Statistics Canada • Statistique Canada 17/11/2018 5
QUARTERLY GDP REVISION SCHEDULE RP0304 QUARTERLY GDP REVISION SCHEDULE REVISIONS IN YEAR THREE REFLECT A FULL BENCHMARKING TO INPUT OUTPUT TABLES USING A FULL SET OF ANNUAL SURVEYS COVERING THE ENTIRE ECONOMY The GDP revision cycle normally spans over four years so that estimates for the four quarters of Year t are finalised in the first quarter of Year t+4. For more details of the revision schedule, see 'The 1999-2002 Revisions of the National Economic and Financial Accounts' on Page 35 of Canadian Economic Accounts Quarterly Review, First Quarter 2003, Statistics Canada Catalogue No. 13-010-XIE. Prel. - Preliminary or first published Income and Expenditure Accounts estimates are released 60 days after the end of the reference period. Rev1 to Rev6 - First to sixth revision to the preliminary estimates. Benchmarks from the preliminary input-output tables are adopted in Rev6 estimates. Final - Final estimates, for which benchmarks from revised input-output estimates are adopted. These are the fourth to seventh revision of, respectively, the fourth, third, second and first reference quarters' preliminary estimates. Final estimates may be subject to further revision at rebasing, as well as at historical revisions. Historical revisions provide an occasion to improve estimation methods, eliminate statistical breaks resulting from more limited revisions and introduce conceptual changes into the system. Rebase since 1981: base years 1986, 1992 (third quarter 1997) and 1997 (first quarter 2001); Chain Fisher estimates were also introduced in the first quarter of 2001. Historical revisions since 1981: 1986 and 1997 Statistics Canada • Statistique Canada 17/11/2018 5
QUARTERLY GDP REVISION SCHEDULE RP0304 QUARTERLY GDP REVISION SCHEDULE REVISIONS IN YEAR FOUR ARE CONSIDERED « FINAL » - DETAILLED SOURCE DATA ON HEALTH AND EDUCATION SECTORS BECOME AVAILABLE The GDP revision cycle normally spans over four years so that estimates for the four quarters of Year t are finalised in the first quarter of Year t+4. For more details of the revision schedule, see 'The 1999-2002 Revisions of the National Economic and Financial Accounts' on Page 35 of Canadian Economic Accounts Quarterly Review, First Quarter 2003, Statistics Canada Catalogue No. 13-010-XIE. Prel. - Preliminary or first published Income and Expenditure Accounts estimates are released 60 days after the end of the reference period. Rev1 to Rev6 - First to sixth revision to the preliminary estimates. Benchmarks from the preliminary input-output tables are adopted in Rev6 estimates. Final - Final estimates, for which benchmarks from revised input-output estimates are adopted. These are the fourth to seventh revision of, respectively, the fourth, third, second and first reference quarters' preliminary estimates. Final estimates may be subject to further revision at rebasing, as well as at historical revisions. Historical revisions provide an occasion to improve estimation methods, eliminate statistical breaks resulting from more limited revisions and introduce conceptual changes into the system. Rebase since 1981: base years 1986, 1992 (third quarter 1997) and 1997 (first quarter 2001); Chain Fisher estimates were also introduced in the first quarter of 2001. Historical revisions since 1981: 1986 and 1997 Statistics Canada • Statistique Canada 17/11/2018 5
Statistical Events related to Revision Patterns Third Quarter 1969 Shift from 90-day lag for release to 60 day lag First Quarter 1986 Adoption of GDP over GNP as feature estimate Full integration of quarterly accounts with annual Input Output Tables Significant Statistical Revisions Third Quater 1997 Conceptual Revision to SNA 1993 First Quarter 2000 Adoption of Fisher Formula for calculation of GDP First Quarter 2001 Capitalization of Software Statistics Canada • Statistique Canada 17/11/2018
Statistical Events related to Revision Patterns Other Significant events Various rebase exercises to Laspeyres-based real GDP Re-design of business survey strategies (1981 to 1989) Introduction of Provincial IO Program and overhaul of Annual Survey System (1999 to 2002) Agressive adoption of administrative data to reduce response burden (2003 – 2007) Next Historical Revision 2012/2013 Statistics Canada • Statistique Canada 17/11/2018
Overview of revisions analyses Pre and post timeliness change 1969 Post adoption of SNA 1993 – in 1997 Recent revision patterns – 2001 to date Statistics Canada • Statistique Canada 17/11/2018
Pre and Post Timeliness Change Statistics Canada • Statistique Canada 17/11/2018
Revision Study of seasonally adjusted quarterly real GDP, 1981 to 2003 RP0304 Revision Study of seasonally adjusted quarterly real GDP, 1981 to 2003 After adoption of SNA 1993 GDP growth rate - constant price, seasonally adjusted, quarter-on-quarter Total revisions, incremental and cumulative revisions at each vintage (preliminary; 1Q, 1Y, 2Y, 3Y and 4Y later) Revisions due to base-year, concepts-sources-method changes; historical revision Revisions in the 1980s versus those in the 1990s Revisions and the economic cycle Seasonally adjusted real quarterly GDP growth rate at constant price is the series most often related to ‘growth in the economy’ by statistical users. The use of this series, as opposed to GDP level or current price growth rate, has the advantages of eliminate the trend problem whereby revisions to GDP tend to grow over time because economic growth and inflation raise the GDP level at current price. It also reduces the need to account for the impact on revisions of the changes base year, and concepts, sources and methods. Subsequent modules of this project will assess the impact of these changes. GDP revisions are studied for a variety of aspects and perspectives. Section 2 of the report on module 1 provides details on the issues involved. The report is also in the share-file location mentioned in the notes on Slide #1. Statistics Canada • Statistique Canada 17/11/2018 4
GROWTH RATES - DIFFERENT VINTAGES RP0304 GROWTH RATES - DIFFERENT VINTAGES The 1981 to 1999 GDP growth rate estimates from the 5 vintages followed a similar pattern in general, although there were variations from one vintage to another. The seemingly least vintage-variable period from 1995 fourth quarter to 1996 fourth quarter was in fact due to the suspension of annual revisions while historical revisions were being compiled. Statistics Canada • Statistique Canada 17/11/2018 12
PRELIMINARY & FINAL ESTIMATES & TOTAL REVISIONS 2003-4 CSNA RP Toolkit PRELIMINARY & FINAL ESTIMATES & TOTAL REVISIONS Quarterly GDP Growth Rates and Total Revisions, 1981 to 1999 Seasonally adjusted, constant price, quarter-on-quarter, % point Graphs can give a quick picture of estimates and their revisions. They can provide useful information even when there are only a few observations available, particualrly at an early stage of developing revision analysis or during transition periods when there have been many or major changes to concepts, sources and methods. The absolute size of revisions decreased from 1980s to 1990s but part of the reduction was due to the lower growth rates. There was a positive linear relationship between revisions and growth rates (positive growth rates tended to be accompanied by positive revisions, high-magnitude revisions came with high-magnitude growth rates; and vice-versa). This linear relationship is shown in slide no. 15. Statistics Canada • Statistique Canada 17/11/2018 11
Statistical Analysis BIAS & DISPERSION 2003-4 CSNA RP Toolkit Quarterly GDP Growth Rates and Total Revisions, 1981 to 1999 Seasonally adjusted, constant price, quarter-on-quarter, % point Mean revisions (or bias) were low at 0.06, 0.1 and 0.08 percentage points respectively for the 1980s, 1990s, and the two decades combined. When the outlier was excluded the mean revisions were uniform at 0.1 percentage point; however the deviation from the mean as measured by the standard deviation was much reduced from 0.49 in the 1980s to 0.39 in the 1990s. This improvement in revision dispersion was also picked up by the mean absolute revision, which lowered from 0.42 (or 0.40 when the outlier was excluded) to 0.33 percentage point. Moving from the 1980s to the 1990s, revisions grew smaller in absolute values, the average was 0.42 percentage points in the earlier period versus 0.33 percentage points in the later. However this improvement of 21.4% should be interpreted in relation to the change in the average absolute GDP growth rate. The last row in the table shows that there was, indeed, a fall in this relative measure (the relative dispersion) but the improvement evident from this was a much smaller 4.4%, from 0.405 in the 1980s to 0.387 in the 1990s. Therefore, the reduction in the revisions was due mainly to the drop in the GDP growth rates themselves. Also, when the outlier was excluded, the relative dispersion was steady from 0.373 to 0.387, compared to 0.379 for the 2 decades combined. Statistics Canada • Statistique Canada 17/11/2018 13
INCREMENTAL REVISIONS @ DIFFERENT VINTAGES RP0304 INCREMENTAL REVISIONS @ DIFFERENT VINTAGES 2003-4 CSNA RP Toolkit (Incremental revisions from these four graphs sum to the total revision for each quarter.) For 1981 to 1999, 62 of the 76 (or over 80%) of the revisions from the preliminary to estimates at the update one quarter later were no more than 0.1 percentage point in magnitude, with mean revision of 0.0013 and mean absolute of 0.093 percentage point. Additional revisions, from the update one quarter later to the update one year after the preliminary estimates (and from one year to two years later), generally had higher magnitudes with 66% (58%) no more than 0.1 percentage point. For these additional revisions, the mean revisions were 0.016 and -0.012, while the mean absolute revisions were 0.1882 and 0.2976 percentage point respectively. While these measures indicate that revision values increased at progressive vintages, all the values were low. Statistics Canada • Statistique Canada 17/11/2018 12
DISTRIBUTION OF THE REVISIONS 2003-4 CSNA RP Toolkit DISTRIBUTION OF THE REVISIONS “NORMAL”? For the 1981 to 1999 revisions, 67%, 96% and 100% of the revisions were respectively within one, two and three standard deviations of the mean; corresponding percentages in a normal distribution are 68%, 95% and 99.7%. A normal distribution is symmetrical with the mean, median, and mode all coinciding in value. The 1981 to 1999 revisions had only a small negative skew with a mean of 0.10 and a median of 0.16. The set of revisions, numbering seventy-five, is sufficiently large to conduct statistical inference based on a normal or a t-statistic. This graph shows that while the 1981 to 1999 revisions have the general shape of a normal distribution, it is not an exact fit. Therefore additional precaution should be exercised when interpreting z, t, and f-test results. Statistics Canada • Statistique Canada 17/11/2018 14
ECONOMIC EXPANSION AND CONTRACTION RP0304 ECONOMIC EXPANSION AND CONTRACTION Another measure of the difference between revisions in the two types of periods is the shares of negative and positive revisions. Over 1981 to 1999, these shares were 22% and 56% for negative revisions in the expansionary and contracting periods respectively, while 78% and 44% were the corresponding shares of positive revisions. The differences between these shares are noticeable. Statistics Canada • Statistique Canada 17/11/2018 20
REGRESSION AND CORRELATION 2003-4 CSNA RP Toolkit REGRESSION AND CORRELATION The regression line Y = -0.04 + 0.2X is fitted to the points on this scattergram. The line passes through the y-intercept -0.04 and the point (0.705, 0.101) where 0.705 is the mean of the final GDP growth rate estimates and 0.101 is the mean of the revisions. The coefficient of determination equal 0.156 which means that 15.6% of the variation in the revisions is due to its regression linear relationship with the economic cycle. The correlation coefficient = 0.395 indicate that the relationship is a positive one. Statistics Canada • Statistique Canada 17/11/2018 15
IMPACT OF SUB-AGGREGATES EXPENDITURE SIDE 2003-4 CSNA RP Toolkit IMPACT OF SUB-AGGREGATES EXPENDITURE SIDE Impact of a component on the total aggregate is defined to be the product of the mean of the absolute values of the revisions of the component (m(Abs))and the component’s share of the total aggregate in value terms (share). Personal consumption (HFCE) has the largest share at 57.11% and the mean(Abs) of its growth rates revisions is 0.21 percentage point, then its impact on GDP growth rate is 0.1217. This impact of the component, which constitutes the largest % share of GDP, is selected as the standard impact value. Similarly exports of goods and government fixed capital expenditure (GFKE) respectively have impact values of 0.3554 (39.19% share, m(Abs) of its growth rates revisions is 0.89 percentage point) and 0.0448 (2.45% share, m(Abs) of its growth rates revisions is 1.83 percentage point) respectively, then the isoquant of the former lies above the standard impact isoquant while the isoquant of the latter lies below it. Therefore, a review of quality would look into exports first. More details in Toolkit Section C.3 and in Oller, L-E. and Hansson, K-G. December 2002. ‘Revisions of Swedish National Accounts 1980-1998 and An International Comparison’. Development and Improvement of Economic Statistics. SOU 2002:118 Translation in abridged form. 2003. Stockholm: SS. PP. 148 to 150. Statistics Canada • Statistique Canada 17/11/2018 18
Revisions of quarterly real GDP(SAAR), 1981-2003 RP0304 Revisions of quarterly real GDP(SAAR), 1981-2003 KEY FINDINGS Marginally significant negative bias (0.1 percentage point) The estimates were revised upwards more often than otherwise (52% of the time versus 38%) A persistent run of positive revisions in 2nd half of 1990s (13 out of 15 revisions) Evidence of lower dispersion measures. Mean absolute revision was reduced by 21.4% from 1980s to 1990s, but relative dispersion only dropped by 4.4%. Preliminary IO benchmarks @ ‘3-years-later’ were the source of the largest shares of total revisions (over 50% on average) Base-year and CSM changes had minor effect on revisions Marginally significant negative bias of 0.1 percentage point for the whole period, as well as for each of the two decades. Statisticians are ‘conservative’, so the preliminary estimates tended to be lower than the final - a world-wide phenomenon, e.g. International comparison by Sweden in 2002 for annual constant price GDP growth rates: Sweden (0.4), UK (0.3), US (0.1) & Canada (0). From 1996 second quarter, a run of positive revisions – most likely due to adopting data of expanded coverage from PIPES (Module 2 will check) An improvement in the reliability of the esimates is seen in the sizeable drop in the dispersion of the revisions from the 1980s to the 1990s However some of the reduction was due to the positive linear relationship between growth rates and their revisions, see Slides 10 and 11. Looking at revisions from different vintages, benchmarks from preliminary IO tables were found to be the source of the largest shares of total revisions As expected, base-year, concepts, sources & methods changes did not cause much revisions to this GDP series, although their effects on revisions to the levels estimates were quite large. The net Chain Fisher effect (net of the usual base-year change effect, from 1992$ to 1997$) was mostly small, 1997$ and chain fisher estimates of this GDP series were in the main not very different. Statistics Canada • Statistique Canada 17/11/2018 7
Revisions of quarterly real GDP(SAAR), 1981-2003 RP0304 Revisions of quarterly real GDP(SAAR), 1981-2003 KEY FINDINGS Preliminary and intermediate GDP growth rates estimates correctly indicated the sign (direction) of economic growth 95% of the time Accelerations or decelerations were correctly identified two thirds of the time Near-trend rates were correctly measured by different vintages 66% to 89% of the time. Troughs were correctly indicated by all vintages, only one peak was missed by the preliminary series and one was missed by the two-year-later series. The final real GDP growth rate sereis is special in that it is seen to portray the economic cycle. Therefore, it is desirable for the preliminary (and internediate) estimates to be able to identify changes in the cycle adequately. Canadian series are reliable and compares well with other countries. Similar measures for US NIPA were 98% (sign), 75% (AC/DC) aand 75% (near trend), and the US series had more misses of peaks and troughs than the Canadian data (Survey of Current Business December 1998). For Swedish NA they were 97% (sign), 82% (AC/DC); trend and peak/trough were not measured in their 2002 study. Trend estimates based on 24-quarter moving average (a regression estimate) are derived for each vintage of GDP gorwth rate estimates. Trend estimates of the final series are used as a standard. For each vintage, near-trend rates are values within 50% of the corresponding trend estimates of the final series. The percentage-correct measure is the number of near-trend rates as a proportion of total number of trend estimates. The above definition was adopted to facilitate comparison with the US. A simpler and more intuitive near-trend rate definition might be «within 0.2 percentage point of the corresponding trend in the final series», in which case the measures were 79% to 98%. Statistics Canada • Statistique Canada 17/11/2018 8
Recent Revision Patterns Statistics Canada • Statistique Canada 17/11/2018
Recent Revision Analysis Statistics Canada • Statistique Canada 17/11/2018
Recent Revision Analysis Statistics Canada • Statistique Canada 17/11/2018
Recent Revision Analysis Statistics Canada • Statistique Canada 17/11/2018
Conclusions and questions for discussion Revisions Analysis is essential to a statistical process for determining areas for improvement Revision statistics need to be interpreted in the context of changes to the statistical system Source data improvements Conceptual changes Economic changes How should STC improve communications of revisions? Link, break or re-estimate? Historical revisions – « Big Bang » or Staged Questions? Statistics Canada • Statistique Canada 17/11/2018