Correcting for Coverage Errors in Enterprise Surveys A Register-based Approach Anders Wallgren, Britt Wallgren, Statistics Sweden

Transcription

1 Proceedings of Q2008 European Conference on Quality in Official Statistics Correcting for Coverage Errors in Enterprise Surveys A Register-based Approach Anders Wallgren, Britt Wallgren, Statistics Sweden 1. Introduction Integration of microdata from many sources leads to new methodological issues. Those who compile the National Accounts receive macrodata from multiple sources, and when they analyse this macrodata they are actually working with macro editing of multiple sources. During this editing work they find many inconsistencies and probable errors, but it is difficult to find the causes behind these errors and to develop methods to reduce these errors if not the microdata behind these macrodata are integrated and analysed. During 2006 and 2007 we have been in charge of a project at Statistics Sweden consisting of persons from the National Accounts, the Business Register and different business surveys. Our aim has been to develop methods for integration of microdata from many sources, both sample surveys and data from administrative registers. We have also developed and tested methods to find and reduce errors and inconsistencies in data from different sources. We have found coverage errors in the Business Register and many enterprise surveys and can explain which consequences they have for the National Accounts. We have also developed methods to reduce these coverage errors and inconsistencies with register-statistical methods that includes that a special calendar year version of the Business Register is created. The project has delivered a final report (in Swedish) with a number of proposals to improve the Business Register and the system of enterprise surveys that deliver data to the yearly version of the National Accounts (NA). 2. The old method The frame used for Structural Business Statistics (SBS) has been created during November each year. Administrative data for 2004 are received during Chart 1. Different frames and register populations used for yearly NA for 2004 Quartly frames Yearly frame for SBS Deliveries of administrative data for 2004

2 The November frame consists of all enterprises registered as active during November according to Statistics Sweden s Business Register. This register has up to now been based on an exact copy of the Tax authorities administrative Business Register that is updated every week. During 2004 and 2005 large amounts of administrative data from different sources regarding 2004 have been delivered to Statistics Sweden. These sources include: monthly VAT-declarations, monthly wage sum reports, monthly external trade (Extrastat), yearly income statements from employers (wage sums) and yearly income declarations from enterprises and self-employed. 3. The new idea At the end of 2005 all administrative sources regarding 2004 have been delivered. It is then possible to create a register with all enterprises that have been active in at least one administrative source regarding This calendar year register can t be used as sampling frame as it is created a long time after 2004, but it can be used as register population for all administrative sources regarding According to the new idea, that has been tested in our project, this calendar year register should be used as standardised population, see Wallgren & Wallgren (2007), for all sample surveys and register-based surveys regarding The calendar year register can be created during the beginning of year t+2 (here year 2006) and earlier sample surveys can be calibrated against this register to produce revised estimates that are consistent with this register population. Also, all sets of administrative data should use this register population. Then the same sector codes and industrial activity codes for all enterprises will be used in all surveys. Finally, the National Accounts will get consistent estimates regarding year t (2004) during spring year t+2 (2006). At this time the work with the definitive yearly accounts for year t (2004) starts according to the traditional time table. Chart 2. Calendar year register for 2004 is created during January 2006 Quartly frames Calendar year register 2004 Yearly frame based on ALL sources for SBS 2004 All enterprises active during Deliveries of administrative data for 2004 National Accounts 2004 data delivery In our project, a calendar year register for 2004 was created in January We then compared statistics based on the November frame for 2004 with statistics based on administrative data for the enterprises in the calendar year register. Both

3 overcoverage and undercoverage errors and different kinds of inconsistencies were found in the statics based on the November frame. 4. Coverage errors and inconsistencies with the old method The November frame 2004 contained legal units, classified as active according to the tax authorities. The same identity numbers are used for tax administration and in Statistics Sweden s Business register. The activity codes from the tax authorities are also used by Statistics Sweden. In our project it was found, that the statistical quality of this administrative variable was low. A substantial overcoverage was found: legal units, classified as active had not reported any economic activity in any administrative source regarding Also undercoverage was found: legal units were classified as has never been active, units were classified as not active and units were missing completely in the Business Register system in spite of the fact that they had reported economic activity in at least one administrative source. Chart 3. Coverage errors, as number of legal units, in the November frame 2004 November frame 2004 Calendar year register (CYR) "Active" according to Nov. frame Overcoverage Not in CYR "Active" according to November frame In CYR "Has never been active" acc. to Nov. frame "Not active" according to Nov. frame In CYR Missing completely in Business Register Total undercoverage Nov Jan It should be noted that from administrative point of view, the activity classification is not bad. Enterprises report zero and pay zero tax which is correct. For these enterprises it is convenient to remain in the tax authorities register in case that they will become active again in the future. The undercoverage consists of many small companies that are new and/or only exist during a short time span. From statistical point of view, the overcoverage and undercoverage will give rise to quality problems for the National Accounts. We have also found that these coverage errors are highly selective, and that different economic variables are affected in different ways, which means that inconsistencies will arise. In Charts 4 7 below different kinds of coverage errors and inconsistencies are described.

4 4.1 Undercoverage errors Chart 4. Errors due to undercoverage in the November frame 2004 Wage sum Turnover SEK Millions SEK Millions Active according to the November frame Has never been active according to the November frame Not active according to the November frame Missing in the November frame Calendar Year Population, total Errors due to undercoverage in the November frame Errors due to undercoverage in the November frame, % In Chart 4, the overall errors due to undercoverage are shown. The error in the November frame wage sum estimate is only 0.65% but the error in turnover is more substantial, 2.13%. This is explained by the fact that many small enterprises outside the present Business Register have no employees. However, if wage sums and turnover are estimated by industry, we find substantial errors in both wage sums and turnover for some industries. These errors are described in Chart 5 below. The errors shown in Chart 5 are highly selective as different industries have quite different errors. We can also see that undercoverage errors by industry are quite different for the two variables wage sums and turnover. The consequences of this are that wage sums and turnover in the Structural Business Statistics survey are inconsistent with each other. Chart 5. Undercoverage in November frame 2004, non-financial sector by industry Wage sums, yearly income statements Turnover, VAT-register Total Undercoverage Total Undercoverage Non financial sector SEK SEK SEK SEK Millions Millions Percent Millions Millions Percent and 11 and except Total: Chart 6. The effect of undercoverage for wage sums in the National Accounts 2004 All wage sums have been adjusted to the definitions used by the NA Sector Wage sum, millions Total wage sum according to the best administrative source Wage sums from different surveys Enterprises and self employed 1, 2 and Government Municipalities Non-profit institutions serving households Sum of these surveys: Difference best source and sector surveys, SEK millions:

5 In Chart 6 above, the wage sum corresponds to the Calendar Year Population estimate in Chart 4. Due to undercoverage in the different sector surveys, at the National Accounts they find that millions SEK are missing. They therefore have to adjust wage sums by sector but they have had very little statistical support for these adjustments. The need for this kind of adjustments will be reduced if calendar year registers are used in the future. 4.2 Overcoverage errors The overcoverage has been interpreted as nonresponse in the Structural Business Statistics (SBS) and we have also found that over- and undercoverage affect different industries differently, so that these two errors will not cancel out. Chart 7. Overcoverage in SBS based on November frame 2004, by industry Number of legal units SBS, Structural Business Statistics: Value of production Total Overcoverage Total Overcoverage Non financial sector SEK SEK Industry, NACE: Percent millions millions Percent and and Total: The overall overcoverage error in Chart 7 is 7.7% of the legal units, but only 0.6% of the total production value. This means that the overcoverage consists of mainly small enterprises. But the error is selective, some industries as 05 (fishing) and (health establishments) have serious errors. 5. Gaps and overlap with the old method Aggregated data by sector and industry are delivered by many surveys to the National Accounts. Have all statistical units/enterprises in the calendar year population been included? Have some statistical units been included in more than one survey? The calendar year register created after all administrative sources have been delivered to Statistics Sweden can be used to answer these questions. In our project we used the calendar year register for 2004 to see if there were gaps or overlaps between different surveys. We found that approximately financial enterprises had not been included in any survey and that there were approximately 100 enterprises that were included in both the Structural Business Statistics survey and the Financial sector survey. It is to be noted that these enterprises did not

6 belong to the undercoverage mentioned above. In Chart 8 below the enterprises with known NACE are included. The shaded cells in the table indicate were we found many enterprises that had been completely forgotten or had been included in more than one survey. Chart 8. Number of legal units in the Calendar Year Register for 2004 by sector and industry Industry, NACE: Sector: Financial intermediation Total 1 Non-financial enterprises Financial enterprises Government Municipalities Self-employed Non-profit organisations Total: Similar problems are illustrated in Chart 9. The energy industry is surveyed both by a special energy survey but also by the Structural Business Statistics survey. The two populations differ considerably which will result in inconsistent energy statistics delivered to the National Accounts. Chart 9. Inconsistent frame populations Income and expenditure for Swedish energy enterprises are surveyed, both within Structural Business Statistics (SBS) and Energy Statistics (EN). Consistency editing has shown that different legal units are included in the two frame populations. Some legal units are included in both, others in only one of the surveys. Income 1998 as per: Income 1999 as per: Costs 1998 as per: Costs 1999 as per: SBS EN SBS EN SBS EN SBS EN Only SBS 32% 0% Only SBS 16% 0% Only SBS 32% 0% Only SBS 18% 0% Both 68% 76% Both 84% 96% Both 68% 84% Both 82% 96% Only EN 0% 24% Only EN 0% 4% Only EN 0% 16% Only EN 0% 4% Total 100% 100% Total 100% 100% Total 100% 100% Total 100% 100% 6. Conclusions We have found that a calendar year register with all enterprises active in at least one administrative source can be a very important tool to improve the quality of enterprise statistics. In this paper we have not discussed the methods that should be used when such a register is created. Object editing, defined in Wallgren & Wallgren (2007), must be used as one enterprise can use different identity numbers for taxation and because enterprise units change over time. References Statistics Sweden (2007), Register-based economic statistics based on a standardised register population A calendar year version of the Business Register with consistent microdata designed for the yearly National Accounts (in Swedish), Background Facts Economic Statistics 2007:6 Wallgren, A. and Wallgren, B. (2007), Register-based Statistics Administrative Data for Statistical Purposes, Chichester: Wiley.