Register-based National Accounts Anders Wallgren, Britt Wallgren Statistics Sweden and Örebro University, e-mail: ba.statistik@telia.com Abstract Register-based censuses have been discussed for many years and many countries have changed or are planning to change from a traditional population census to a completely or partially register-based census. We suggest that NSI:s also should start planning for a change towards register-based National Accounts. In this paper we will compare with the register-based census and if there are some experiences from that transition that can be used when we discuss a transition to register-based National Accounts. Statistics Canada distinguishes between the following four different kinds of survey: 1. Census, 2. Sample survey, 3. Register survey and 4. Macro data survey The transition from a traditional census to a register-based census means that NSI:s move from methodology 1 to methodology 3 above or from 1 to a combination of 2 and 3 in case of a partially register-based census. Today, National Account statistics are created with methodology 4 and we suggest a partial transition towards methodology 3 above. This means that statistical registers with enterprise and trade data should be created that can be used by the yearly National Accounts in the first hand. There will be substantial gains regarding quality. The production system is also improved by more efficient use of administrative data and this will make it possible to reduce response burden and costs. Keywords: Statistical Registers, Coverage errors, Coherence 1. Survey Design for Yearly National Accounts Statistics Canada (2009) distinguishes between the following four different kinds of survey in their Quality Guidelines, and the four methodologies or survey designs can be described in the following way (our terms): 1. A census, which attempts to collect micro data from all members of a population. 2. A sample survey, in which micro data are collected from a random sample of population members. 3. A register survey, or collection of micro data based on administrative records covering all members of a population. 4. A macro data survey, in which data are estimated, modelled, or otherwise derived with macro data from existing statistical data sources. National Accounts are today derived in this way.
Eurostat (2009) also mentions these together with two other kinds of survey (our terms): 5. A survey involving multiple data sources. 6. A complex sample survey, often with non-probabilistic design. The transition from a traditional census to a register-based census means that NSI:s move from methodology 1 to methodology 3 above or from 1 to a combination of 2 and 3 in case of a partially register-based census. Today, National Account statistics are created with methodology 4 and we suggest a partial transition towards methodology 3 above. This means that statistical registers with enterprise and trade data should be created that can be used by the yearly National Accounts in the first hand. In this way the survey design of the National Accounts will be changed. It is necessary to distinguish between the survey design of the National Accounts (design 4 above) with the designs of the surveys that are used by the National Accounts. These surveys can be designed in different ways. The municipality census with economic data from all municipalities is one example of design 1 above. The Structural Business Statistics survey (SBS) where questionnaires are used by Statistic Sweden for one part of the enterprise population while a combination of administrative data and sample survey data are used for the rest of the population is an example of survey design 5. In Section 5 we discuss the surveys that are used by the National Accounts; in Sections 2-4 we discuss the design of the yearly National Accounts survey itself. 2. What can we learn from the Register-based Census? In all Nordic countries the register-based population census is based on a system of statistical registers. During a period of 10-30 years these registers were developed with the intention to replace the costly traditional census. This process is described in UNECE (2007). The Population Register, the Income, Education and Employment Registers are the main registers on persons and to this system also the Business Register and registers on dwellings and housing conditions are added. All registers on persons belonging to the census system are perfectly consistent and coherent. Exactly the same population by age, sex and region is used in all these registers. In the Netherlands the census is based on registers but as some registers are still missing in their system they also use sample surveys. E.g. data from the labour Force Survey from three years are combined into one set of data, and the sampling weights are calibrated to produce estimates that are consistent with the register-based estimates. Three important lessons can be learned from the experiences with the register-based population and housing census: A system of register surveys and sample surveys is created. Within the system micro data can be combined by micro integration. Micro integration means that different sets of micro data are combined by record linkage, and that micro data in this combined set is processed so that consistent and coherent estimates can be produced.
The method used, that results in consistent and coherent estimates, consists of creating a standardised population and a number of standardised variables that are used by all surveys in the system. To develop such a system requires long-term efforts and coordination between many units at the NSI. Sustained efforts from top management are necessary for the development and implementation of the system. 3. An outline of a Register Survey for Yearly National Accounts As for the register-based census there must be a common population that is used for all registers and sample surveys in the system with enterprise and trade statistics used by the yearly National Accounts. As many economic variables are flow variables, this population should consist of all enterprises that were active during at least some part of the calendar year. At Statistics Sweden we have created such calendar versions of the Business Register during the last years. This version can be created about 12 months after the year in question has ended. In Chart 1 the work processes behind two different versions of the Business Register are compared. The traditional frame population that is used for yearly enterprise surveys is created by Statistics Sweden during November of the reference year. About one year later, the calendar year version of the Business Register can be created. Chart 1. Traditional Frame Population for SBS and Calendar Year Register Population 2007 Quartly frames Calendar year register 2007 Yearly frame based on ALL sources for SBS 2007 All enterprises active during 2007 2007 2008 2009 Deliveries of administrative data for 2007 Data delivery to National Accounts 2007 All important administrative sources and sample surveys are then micro integrated with the Calendar year register as the basis. It is then possible to compare all sources at the micro level. These comparisons will reveal inconsistencies and errors. After this error finding work, follows a process that will transform data so that all sources are consistent and coherent. Finally, consistent and coherent estimates are produced that can be used by the yearly National Accounts. The Calendar year register will serve as the standardised population for the system of enterprise and trade surveys that deliver data to the yearly National Accounts. This register should contain standardised classification variables as economic activity, sector and also region for the establishment version. In this way all surveys can be adjusted to give estimates for the same population and the same domains.
4. Register-based National Accounts what do we gain? There will be substantial gains regarding quality. The production system is also improved by more efficient use of administrative data and this will make it possible to reduce response burden and costs, this is discussed in section 5. The quality components comparability and coherence can be improved when data from different surveys are micro integrated. Coverage errors will be made visible when the calendar version of the Business Register is used as the basis for comparisons. The coverage errors are caused by the existence of a large number of small enterprises and self employed that are difficult to cover with traditional frames. Frames are created with the Business Register, but frames must always be created before the data collection and at that time the information on small enterprises is not up to date. So all frames are created at least one year earlier than the calendar version and consequently the frames have larger coverage errors. There are two issues that require long-term methodological work. The first concerns the fact that different sources use different statistical or administrative units. The micro integration will then consist of the creation of complex statistical units so that different sources can be combined. The second methodological issue concerns the variable economic activity, NACE. Today this variable is used in a way that generates large inconsistencies between different surveys. The Calendar year register offers new possibilities for consistent estimates by economic activity for all surveys that have been micro integrated with the Calendar year register as a common base. Estimation methods for multi-valued variables such as NACE are discussed in Wallgren and Wallgren (2007, Chapter 9). Consistent and coherent enterprise statistics based on a Calendar year register are discussed in op. cit. p 220. 4.1 Coverage errors Undercoverage and overcoverage reduce accuracy in all surveys. These coverage errors will be different in different surveys and thus also give rise to lack of comparability and coherence in the system of enterprise and trade surveys used by the National Accounts. As the National Accounts use estimates by sector and economic activity, differences regarding these classification variables between different surveys will yield coverage errors at the domain level that can be substantial. According to our experience, these coverage errors can be very selective and give rise to serious quality problems. In Chart 2, the coverage errors measured as number of units are illustrated. Undercoverage here is 25 % and overcoverage is 8 %, but if the errors instead are measured as per cent of population totals of economic variables, undercoverage is 1 % - 2 % and overcoverage about 0.5 %. However, if measured as per cent of domain totals by industry, undercoverage can be 10 % - 20 % and overcoverage up to 5 %. Examples of these kinds of errors are reported in Wallgren and Wallgren (2008). We have seen no support for assumptions that undercoverage and overcoverage cancel each
other out or that all industries are affected in the same way. Instead, some industries are affected by undercoverage and some other industries are affected by overcoverage and also different variables show different patterns. Chart 2. Coverage errors, as number of legal units, in the November frame 2007 November frame 2007 Calendar year register (CYR) 2007 92 375 "Active" according to November frame Overcoverage Not in CYR 856 146 "Active" according to November frame In CYR 856 146 43 240 94 760 "Has never been active" acc. to Nov. frame "Not active" according to Nov. frame In CYR 150 738 Missing completely in Business Register 288 183 Total undercoverage 2008 288 183 15 Nov. 2007 15 Jan. 2009 The coverage errors mentioned above are easy to find and the work with correcting for these errors can start as soon as the errors have become visible. The first measure should be to improve the Business Register; the coverage errors in all surveys that are based on that base register will then be reduced from the beginning and later revisions will be smaller. Due to the information in Chart 2, the Swedish Business Register was redesigned during 2010; in January 2012 we will evaluate the change. The next measure should be to adjust all sample surveys by calibrating the sampling weights according to the methods presented in Särndal and Lundström (2005). Auxiliary variables from different administrative sources should be included in the Calendar year register for this purpose. Revised estimates can be produced during January one year after the end of the reference year. Also surveys based on administrative sources may need corrections. Available administrative sources can be used to complete the object set so that it corresponds to the Calendar year register. These sources are also used to generate variable values for the missing parts. The monthly and quarterly surveys used for the Short Term Statistics survey and quarterly National Accounts can of course not use the Calendar year register when they are published for the first time. But afterwards the coverage errors can be analysed with the Calendar year register and the quarterly estimates can be revised. Coverage errors have perhaps not received the same attention as the other non-sampling errors nonresponse and measurement errors. If administrative data are used to create statistical registers with good coverage then routine comparisons with these registers will give valuable information on the size of the coverage errors and this information can be used when surveys are redesigned.
4.2 Coordination between surveys With the Calendar year register the lack of coordination between different business surveys will also become clear. The Structural Business Statistics survey (SBS), the Farm survey, the Energy sector survey and the Financial sector survey may overlap each other and also some enterprises may have been forgotten by all these surveys. Analyses with the Calendar year register will reveal this. Those who work with the National Accounts (NA) prefer to use the source they judge to be the best one for each part of the accounts. E.g. the Farm Structure Survey (FSS) is used for the agricultural part of the NA, not the Structural Business Statistics survey (SBS). The FSS is based on the farm population in the Farm Register with 72 609 units during 2007. But the Calendar year register has 160 367 units in NACE 1 during 2007 and statistics for this domain was reported by the SBS, but that information was not used by the NA. This means that about 88 000 units in the enterprise population for 2007 were not included in the National Accounts for 2007. After analysing these 88 000 units it was found that they should have been included in other parts of the NA, but not in the agricultural part, this is discussed in Wallgren and Wallgren (2010). In Chart 3 there are about 10 800 financial enterprises either belonging to NACE 65-67 or the financial sector. Only less than 1 000 of these are included in the financial survey used by the National Accounts. Even if these are the largest and most important ones, there are about 10 000 legal units that were not included in the NA. Chart 3. Number of legal units in the Calendar year register 2007 Economic activity, NACE 2002 Sector 01 02-64 65-67, financial intermediation 70-99 Non-financial enterprises 160 367 503 643 7 583 437 832 Financial enterprises 1 8 3 251 22 Government 13 7 8 556 Municipalities 0 7 5 547 Non-profit organisations 143 1 220 1 027 27 996 The Swedish National Accounts apply some kind of corrections for undercoverage. E.g. a high quality estimate of total gross wages is 1 042 SEK billions according to the best administrative source for 2004. When the estimates for all sectors are summated, where the best possible source is used for each sector, then this sum is only 1 023 SEK billions. The sector estimates are then adjusted so that their sum will be 1 042 SEK billions. A related issue is highlighted when the Calendar year register is analysed in this way. How are units classified by sector and economic activity? Inconsistencies found can indicate that groups of units should be reclassified and that more efforts should be spent on coding. Again, the coverage errors mentioned above are easy to find. The inconsistent surveys should be redesigned so that future coverage errors will be smaller, and remaining errors should be corrected in the same way as mentioned in section 4.1 above. In Sweden, the statistical system is decentralised. Statistics Sweden is not responsible for agricultural or financial statistics this means that different authorities responsible for different parts of official statistics must cooperate to reduce coverage errors.
4.3 Coherence Consistency regarding populations and variables are necessary for the coherence of estimates from different surveys. The yearly National Accounts (NA) are based on a large number of estimates from different surveys. Those who work with the NA know that the estimates they get have errors and are inconsistent and that their work consists of macrointegration of these inconsistent sources. They transform all these inconsistent estimates into consistent accounts. In Chart 4 estimates for the same variable (employment) by economic activity can be compared for four different surveys. In this example the inconsistencies are visible. The differences between columns (1) and (2) are caused by different choice of statistical units and that main economic activity for enterprises and establishments can differ within the same enterprise. As the National Accounts are based on surveys with different kinds of statistical units this kind of inconsistencies give problems, but when surveys describe different variables it is not clear which estimates are inconsistent. The differences between columns (2) and (3) are mainly caused by differences between populations. The Business Register has coverage errors as described above but the Employment Register has not these errors. The differences between columns (3) and (4) are mainly caused by different measurement methods. Employment in the Employment Register is a model-based derived variable based on administrative variables and the Labour Force Survey (LFS) is a sample survey with telephone interviews. The surveys used for the National Accounts are inconsistent in this manner, but as e.g. one survey reports production, a second gross fixed capital formation, a third gross wages and a fourth hours worked, we know that there are errors, but we don t know where. Chart 4. Number of Employees November 2004 according to four different surveys Employees November 2004, thousands Business Register Employment Register LFS Economic activity Enterprises Establishments Persons main job's first NACE Error NACE 2002 (1) (2) (1)-(2) (3) (2)-(3) (4) margin (3)-(4) A Agriculture and forestry 34 36-1 37-1 26 5 10 B Fishing 1 1 0 0 0 0 1 0 C Mining and quarrying 9 8 1 7 0 5 2 3 D Manufacturing 679 629 51 710-81 635 23 75 E Electricity, gas and water 21 22-1 28-6 29 5-1 F Construction 197 209-12 215-6 199 14 17 G Wholesale and retail trade 456 453 3 484-31 456 20 28 H Hotels and restaurants 89 93-4 99-5 106 10-8 I Transport, communication 240 242-2 243-1 236 15 7 J Financial intermediation 83 77 6 85-8 78 9 7 K Real estate, business activities 457 524-67 457 67 470 20-13 L Government 139 215-77 239-24 230 15 9 M Education 382 408-27 431-23 462 20-30 N Health and social work 836 684 152 675 8 675 24 0 O Other service activities 142 163-21 175-12 168 13 8 Unknown activity 0 0 0 38-37 4 34 Total 3 763 3 763 0 3 924-160 3 778 145 We suggest that the method used by the yearly National Accounts today macrointegration of estimates from many enterprise and trade surveys is replaced with micro-
integration of data from the same surveys based on the Calendar year register. Our reason for this is that micro-integration will make it possible to detect and correct inconsistencies. All estimates will describe the same enterprise population and the same domains. 5. Improving the Production System Laitila, Wallgren and Wallgren (2011) mention that quality of an administrative source is defined as the usability of the source for the production of statistics at the NSI. Apart from using an administrative source directly for producing statistics it may be possible to use a source to improve some part or some parts of the production system at the NSI. This aspect they call the production process quality of the source. The base registers (black rectangles in Chart 5) and the links between them constitute the basis of the Production System in a NSI that utilises administrative data to full extent. All sample surveys use one of them as sampling frame and all register surveys use one of them as register population and also use the links in the system to integrate data from different sources. The base registers are discussed in Wallgren and Wallgren (2007). Chart 5. A Production System with full access to Administrative Data Sampling of persons or households Sampling of activities Other registers on persons Population Register Activity Register Other registers on activities Other registers on real estate Property Register Business Register Other registers on enterprises Sampling of real estates or buildings Sampling of enterprises At the time of quality assessment of an administrative source or register, its potentials for supporting current statistics production can be measured. As we here primarily discuss enterprise and trade statistics, a list of such parts of the system that could be improved by an administrative source can be given. Can the source be used to improve the Business Register? improve the Structural Business Statistics survey (SBS)? replace SBS-questionnaires to some extent? improve or replace other enterprise surveys? improve Intrastat? replace Intrastat-questionnaires to some extent? improve or replace other trade surveys? This way of using administrative sources is today often overlooked. The reason for this is that this kind of analysis requires micro integration of many sources with the Calendar
year register and this micro integration is a new way of working with data. To be successful it is required that this work is undertaken by persons that combine the following requirements: Understanding of the system of surveys: the Business Register, the enterprise surveys and the National Accounts. Understanding of the administrative sources used the tax forms and rules that define the administrative variables. A capability to work with large sets of data and do statistical analysis with the purpose to find errors and inconsistencies and do the necessary corrections. To make our vision of Register-based National Accounts come true it will be necessary to organise a new unit that is responsible for the Calendar year register and the micro integration of the enterprise and trade surveys and all relevant administrative sources that are used for the National Accounts. This unit will find errors and inconsistencies and the surveys should then be redesigned to improve the production system. This work will be the concern of many surveys and will take time and must be supervised and encouraged by top management. The quality of the National Accounts should be improved by this and it is quite possible that the need for revisions will be reduced. The production system will also be improved and due to more efficient use of available data it will be possible to reduce some enterprise surveys. References Eurostat (2009): ESS Standard for Quality Reports. Methodologies and working papers. Statistics Canada (2009): Quality Guidelines. Fifth Edition October 2009. Statistics Netherlands (2004): The Dutch Virtual Census of 2001 Analysis and Methodology. Laitila T., Wallgren A., Wallgren B. (2011): Quality Assessment of Administrative Data. Research and Development Methodology Reports from Statistics Sweden 2011:1. Särndal C-E., Lundström S. (2005): Estimation in Surveys with Nonresponse. John Wiley & Sons, Ltd. Wallgren A., Wallgren B. (2006): Register-based Economic Statistics on Enterprises Editing Issues. Working paper for the UNECE Work session on statistical data editing in Bonn 25-27 September 2006. Wallgren A., Wallgren B. (2007): Register-based Statistics Administrative Data for Statistical Purposes. John Wiley & Sons, Ltd. Wallgren A., Wallgren B. (2008): Correction for Coverage Errors in Enterprise Surveys a Register-based Approach. European Conference on Quality in Official Statistics, Rome 8-11 July 2008. Wallgren A., Wallgren B. (2010): Correction Using administrative registers for agricultural statistics. In Benedetti, Bee, Espa and Piersimoni (eds): Agricultural Survey Methods. John Wiley & Sons, Ltd. UNECE (2007): Register-based statistics in the Nordic countries. United Nations publication, ISBN 978-92-1-116963-8.