Institut National de la Statistique et des Etudes Economiques INSEE

Département Répertoires, Infrastructures et Statistiques Structurelles
Division élaboration des statistiques annuelles d'entreprise

Direction générale INSEE
Département Répertoires, Infrastructures et Statistiques Structurelles
Division élaboration des statistiques annuelles d'entreprise
88, avenue verdier, bureau 3A044

92120 Montrouge

Structural business statistics (SBS) describes the structure, conduct and performance of economic activities, down to the most detailed activity level (several hundred economic sectors). SBS covers all activities of the business economy with the exception of agricultural activities, public administration and (largely) non-market services such as education and health. Main characteristics (variables) of the SBS data category:

• "Business demographic" variables (e.g. Number of active enterprises)

• "Output related" variables (e.g. Net turnover, Value added)

• "Input related" variables: labour input (e.g. Number of employees and self-employed persons, Hours worked by employees); goods and services input (e.g. Purchases of goods and services); capital input (e.g. Gross investments)

Business services statistics (BS) collection contains harmonised statistics on business services. From 2008 onwards BS become part of the regular mandatory annual data collection of SBS. The BS’s data requirement includes variable “Turnover” broken down by products and by type of residence of client. 

The annual regional statistics collection includes three characteristics due by NUTS-2 country region and detailed on NACE Rev 2division level (2-digits).

Statistical Classification of Economic Activities in the European Community (NACE):  NACE Rev.2 is used from 2008 onwards. Key data were double reported in NACE Rev.1.1 and NACE Rev.2 only for 2008. From 2002 to 2007 NACE Rev. 1.1 was used and until 2001 NACE Rev.1

The regional breakdown of the EU Member States is based on the Nomenclature of Territorial Units for Statistics (NUTS). 

The product breakdown is based on the Classification of Products by Activity (CPA) as stated in the Regulation establishing CPA 2008 and its amending Commission Regulation (EU) No 1209/2014  (from reference year 2015 onwards).

Starting reference year 2021 onwards SBS cover the economic activities of market producers within the NACE Rev. 2 Sections B to N, P to R and Divisions S95 and S96.
Until 2007 the SBS coverage was limited to Sections C to K of NACE Rev.1.1 and from the reference year 2008 to 2020 data was available for Sections B to N and Division S95 of NACE Rev.2.

From 2013, as the first reference year, to 2020 information is published on NACE codes K6411, K6419 and K65 and its breakdown.

From 2008 reference year data collection BS covers NACE Rev 2 codes: J62, N78, J582, J631, M731, M691, M692, M702, M712, M732, M7111, and M7112.

SBS constitutes an important and integrated part of the new European Business Statistics Regulation N° 2152/2019

Data requirements, simplifications and technical definitions are defined in Commission Implementing Regulation (EU) 2020/1197

Since the reference year 2017, the French SBS results are produced in enterprise, and are fully complient with the statistical unit Regulation N° 696/93.

However, the data transmitted by the French National Bank for class 64.11, 64.19 and division 65 are in legal units.

The data collection units are mainly the legal units. For the biggest groups, which are manually profiled, some data are collected at the enterprise level. For the others, automatic profiling is carried out using legal unit data only.

The French SBS system Esane covers all market and economically consistent units, without any threshold of size, excluding financial units (save for class 64.20 and division 66) and agriculture.

The statistics produced by the Esane process are supplemented by results produced by the French National Bank for class 64.11, 64.19 and division 65.

Only activities conducted on the French territory including overseas departments are considered.

Branches of foreign units – that is local units not constituting separate legal entities but dependent on foreign enterprises and classified as quasi-corporations in accordance with ESA –, were, until 2014, with some exceptions, automatically considered as out of scope. They are now potentially in the scope, on the condition that they have declared a fiscal form to the French tax administration for the reference year or the previous year and secondly that they have at least one employee.

1996-2021

Not applicable.

• Number of enterprises and number of local units are expressed in units.

• Monetary data are expressed in millions of €.

• Employment variables are expressed in units.

• Per head values are expressed in thousands of € per head. 

Ratios are expressed in percentages.

2021
Data refer to calendar year

Starting with reference year 2021 two new regulations currently form the legal basis of SBS:

• Regulation (EU) 2019/2152 repealing 10 legal acts in the field of business statistics (EBS Regulation), and
• Commission Implementing Regulation (EU) 2020/1197 laying down technical specifications and arrangements (general implementing act) pursuant to Regulation (EU) 2019/2152

Year 1995 was the first year for the implementation of the Council Regulation No 58/97 (SBS Regulation).

The Council Regulation No 58/97 has been amended three times: by Council Regulation No 410/98, Commission Regulation No 1614/2002 and European Parliament and Council Regulation No 2056/2002. As a new amendment of the basic Regulation it was decided to recast the Regulation No 58/97 in order to obtain a new "clean" legal text.

Not applicable.

Statistical confidentiality is defined in Act no. 51-711 of 7 June 1951 (amended) on legal obligation, coordination and confidentiality in the field of statistics. Economic or financial information may not be communicated to anyone for twenty-five years, unless special dispensation is granted by the statistical confidentiality committee, given that the information will not be used for tax inspection or punitive measures.

For business statistics, cells concerning less than three enterprises and cells in which a single enterprise represents over 85% of the total value are protected since they fall under the primary confidentiality rules. Secondary confidential cells are also protected. However, the dissemination of lists extracted from the register of enterprises and establishments may mention economic activity, number of employees category and turnover size bracket.

Primary suppression:

• Minimum frequency: n=3 (safety margin=10%)

• (n,k)-dominance rule with n=1 and k=85%

• 6 response variables are used and suppression patterns are then used for all disseminated variables (use of shadow variable).

• To take into account the dominance rule for variable with negative values, we make primary suppression first with the variable and then with its absolute value. If there is primary secret in one of the two case, then we consider there is secret

Secondary suppression:

• If possible, use of modular approach
• The software Tau-Argus is used to handle primary and secondary suppression
• A specific file contains the hierarchical structure 

• Secondary cell suppression for linked tables is made simultaneously

• For “special” European aggregates asked by Eurostat, it is not possible to handle directly primary and secondary protection within Tau-Argus taking into account all other “standard” aggregates. Consequently for these aggregates simple ad hoc methods are used to handle confidentiality issues. 
• There are different levels of detail for explanatory variables depending on the table to protect. That prevents the data protector from automating secondary suppression resolution (because we deal with microdata and batch mode in Tau-Argus – we cannot use “recode” option in batch mode).

• For variables with negative values, the secondary secret is computed with an ad hoc variable for the cost, resulting from a calculation allowing to respect the total of the variable of interest, the order of the values ​​and that there are no negative values anymore.

To reduce the number of confidential cells, we have used a minimization algorithm (Modular) and the cost function for secondary suppression is equal to the response variable (except in case of negative values).

Business data are published on our website between August and September N+2.

Data are announced in a general calendar of INSEE publications: https://www.insee.fr/fr/information/1302152

 https://www.insee.fr/fr/information/1302152

We release each year aggregated business data on our website.

Moreover, we provide to the public authorities and a limited set of registered main users information which is not publicly available .

Annual.

There are no specific press releases directly linked to SBS data

Electronic publications are available on our website:

- Synthetic data named "chiffres clés" are disseminated in about fifty data with possibly a graph. These data are available in French and in English, but only for one year, as they are replaced with data from the following year.

- More detailed data: main aggregates, profit and loss account, balance sheet (liabilities, gross assets, accounting depreciation) , employment,  investment, rates and branches data are released:

https://www.insee.fr/fr/statistiques/6482871 (link to 2020 data)

No online database available

non-anomnymized microdata are available on restricted access to researchers only, after examination by the French Secret Committee, which vouches for the respect by the researchers of the statistical confidentiality.

The data are sent to Eurostat, to be used in European aggregates and to be released also as national data

Methodology is available on the Website:

Documentation about the statistical process Esane, definitions of variables:

https://www.insee.fr/fr/information/4226820

https://www.insee.fr/fr/statistiques/7651565#documentation-sommaire

https://www.insee.fr/fr/information/3056089

Documentation about the survey:

http://www.insee.fr/fr/metadonnees/source/s1269

Methodology is available in English:

https://www.insee.fr/en/metadonnees/source/serie/s1188

https://www.insee.fr/en/metadonnees/source/serie/s1269

A national quality report on 2021 Esane results will be available by the end of 2024 . This report will be available in French.

A continuous quality control process is performed on the Esane system (French SBS system). Each year, the different sub-processes – the handling of restructuring, the data collection of the survey, the data editing process of the survey data, the data editing process of the administrative data, the consistency monitoring process between the different sources, and the final output editing step – are reviewed with a view toward ongoing improvement. The documentation and the training related to each sub-process are adjusted accordingly.

In addition to the quality report mentionned in §10.7 and to the process review mentioned in §11.1, a quality approach involving our main internal users and coordinated by the Quality Unit of the French NSI is conducted.

Quality is evaluated as medium.

We have regular consultations with some of our main users: National accounts, technical ministries for quantifying policy measures, professional organizations. 

The national data are more detailed (level of nomenclature, and additional variables) compared to SBS. The data available in the fiscal source are implemented with more details than asked for by SBS; results are also given at the NAF level (French version of the NACE), which is more detailed.

We have not organised a punctual or a regular survey related to the public users' satisfaction. 

The users are very interested by the results, especially at detailed levels. 

The satisfaction of our main internal users is regularly assessed, by regular meetings and through the quality approach previously mentioned.

An internal mission of the General Inspectorate has been conducted to estimate the needs satisfaction degree of early users.

 Some variables are unvalaible or have been estimated :

• 220201 : approximated with number of paid hours, excluding some parts of  financial section. Those for which Banque de France is responsible for.

• 260106 : not available.

• 240103 : partial data only, as it come from ESA survey which does not cover the all SBS scope.

• 240105 :

  • from EICEC survey for section B and C ;

  • for section F, it is approximated as we don't have the information in ESA survey ;

  • not available for sections D and E.

• 260107 : not available.

• 260105 and  260102 : provision estimated as exact data is not available. Not available the previous years.

• 260104 : provision estimated as exact data is not available. Includes investments on existiing buildings.

• 260103 : no data.

As in all statistical process involving a survey, the results of the Esane system can be affected by sampling errors, coverage errors, non-response and measurement errors.

However, as we will see in the following paragraphs, in principle, none of these potential sources of errors leads to a significant bias on the results.

The French business register has no under coverage problems, and the use of external administrative information allows us to deal with the problem of over coverage.

A complex data editing process – based on selective editing combined with micro-editing for the data editing carried out on a continuous flow basis during the production process supplemented by a final output editing process – aims to reduce as much as possible the significant measurement errors.

Imputation procedures – imputation by historical information, failing that by class imputation means or medians – and reweighting procedures are used to deal with missing units/items, both in survey and administrative data. We try also to minimise the non-response in the survey, thanks to the mandatory character of the survey and to a strict follow-up procedure.

Finally, the use of a complex adjustment procedure – involving the non-response adjustment steps already mentioned, a winsorization procedure and a calibration on margins based on administrative data – and calculating aggregates by simple sum using jointly survey and administrative data tends to reduce as much as possible the non-response bias and to improve the accuracy of the statistics on the main variables of interest.

The change in method from a specific estimator to a simple sum leads to zero variance. However the new estimators are biased (although we try to reduce this bias - see above).

Data transmitted are produced by combining both administrative data and survey data, such as the variables ENT, EMPL, TOVT and AV.
We therefore have exhaustive data for these variables (and so the variables SPL_CVG, SPL_RT, SPL_SZ, WRP_RT and  NW_RPRT are not applicable).

A few variables are only provided from the survey.

Starting with the reference year 2016, the survey is drawn by stratified simple sample random sampling of enterprises. Except for the biggest groups, which are manually profiled and for which data are directly collected at the enterprise level, data are still collected on legal units, with the following rule: when an enterprise is selected, then all legal units within this enterprise will be surveyed (so, from the legal unit point of view, the sample design can be seen as a two-stage cluster sampling). Automatic algorithms allow then to obtain consolidated data at the enterprise level, based on the data collected at the legal unit level. A Neyman allocation under local precision constraints is applied (in the sampling frame of enterprises) in order to optimise the accuracy of the estimation of the total turnover, both at the global level and by activities (with enterprises concepts).

Few months after the drawing of the sample, the delineation of the enterprises are updated thanks to the application of the weight share method (See Fizzala EESW 2019)

Sample survey is combined with administrative sources.

The use of a complex adjustment procedure – involving the non-response adjustment steps detailed in §13.3, a winsorization procedure to deal with outliers and a calibration on margins based on administrative data – and the calculation of aggregates (by simple sum) using jointly survey and administrative data tends to reduce as much as possible the non-response bias and to improve the accuracy of the statistics on the main variables of interest.

The change in method from a specific estimator to a simple sum leads to zero variance. However the new estimators are biased (although we try to reduce this bias - see above).

Methods for taking into account the unit non-responses:

 - In the survey, the unit non­ response is taken into account by two different ways:

• in the take all part of the sample, an imputation procedure is used, mainly based on past responses of the business. If the firm is new in the sample, mean imputation procedures are used, except for categorical data, imputed through random hot-deck method. The adjustement cells used for mean or hot-deck imputations are shaped according to industry and employment size classes. For big businesses, informations from their public reports are also used;

• in the take‑some part of the sample, the « RHG model » weighting procedure is used. The RHGs are defined with the CHAID algorithm applied to various auxiliary variables explaining the response mechanism. The minimum size of each RHG is fixed to 100, in order to ensure estimation of response probability robustness. RHG should also contain at least 10 % of respondents, to avoid a too strong dispersion of the weights after non-response adjustments.

- Unit non‑response is treated by an imputation procedure for the administrative data. This imputation procedure (presented at the EESW 2013) is mostly based on the use of historical information. When this is not possible, the imputed values are implemented by strata medians.

- Moreover, a calibration procedure is applied on the take‑some part of the sample, using turnover and number of enterprise – available in administrative data – as auxiliary variables.

For minimising the unit non-response:

1. The survey is mandatory
2. When an enterprise does not send the questionnaire, there are several follow-up: official letters and then a direct contact for big non-responding firms.

Weighted unit non-response rate:

The sampling weights of legal units was used for weighting non-response. Groups 531, 559, 721, 722, 750 and 813 are not in the scope of the survey. Consequently, some variables only from the survey are not produced for these sectors. 

Bias:

Bias resulting from non-response : the estimates are unbiased under the RHG model (for the take‑some part of the sample) and under the assumptions of the imputation models (for the take‑all part of the sample and for the administrative data).

The bias of the estimate: the estimate has zero variance, but it is biased, albeit with a limited effect on the overall accuracy of the estimate.

The impact of imputation:

Not important. Imputation of administrative data concerns only small or very small businesses and represent a very limited share of the aggregated fiscal turnover. Imputations of survey responses for businesses in the take-all strata mainly use past responses of the firm.

Coverage errors - Frame:

Almost none, because of the statistical inference that is made

Out-of-scope units:

The business register has no under coverage problems. External information (VAT files, fiscal data, registers of compulsory liquidations) is used to identify and discard out-of-scope units, which do not participate in the ex-post treatments and in the estimates.

For the statistical survey, data collection spreads from February to December N+1.

For the administrative data: we receive different files between April N+1 and January N+2, considered for SBS data of reference year N.

Preliminary data (built from the administrative source) are sent to Eurostat in October N+1.

First dissemination in France is in January N+2 (only in legal unit and for internal use by the national accounts). Definitive dissemination – both in legal units for internal users and in enterprises – is in July N+2.

Data with 2021 as reference year have been delivered on June 30th 2023

The same statistical concepts are applied in the entire national territory.

Series are comparable from 1996 to 2007.

There was a change of system for the production of results 2008.

Moreover, the results of 2010 and years after include a new category of small enterprises created in 2009 (“autoentrepreneurs” in French); this category, if not very important considering macro-economic aggregates (as turnover), is composed of numerous small units.  

Moreover, the results from 2012 include for the first time some categories of small units without employees, kinds of French civil property companies and private management leasers, that are composed of numerous and in general small units, for most of them in sections F and L.

In addition, in 2015, the variable that defines the productive character of the statistical units belonging to the field has changed, which improves the way we define the SBS statistical population. Furthermore, branches of foreign units – that is local units not constituting separate legal entities but dependent on foreign enterprises and classified as quasi-corporations in accordance with ESA – which until 2014 were, with some exceptions, automatically considered as out of scope, are now potentially part of the scope, on the condition that they have declared a fiscal form to the French tax adminstration for the 2015 year or the previous year, and secondly that they have at least one employee.

Since the reference year 2017, the French SBS results are produced in enterprise, which leads to a major break in time series compared to the previous years, when the SBS results were produced mainly in legal units, except for some of the biggest groups which were manually profiled.

Finally, in 2021, agregate calculus method was modified: we changed from a composite method to a classical additive method. Thus, data are different from those calculated the previous years. 

• Employment:

From the year 2020, the number of employees provided in the Structural Business Statistics is estimated from a new source, which significantly improves the accuracy of the estimates. This change is being made in coordination with Business Demography Statistics as well as with the Business Register.

From the year 2021, we improve the number of employees (EMPL) as we managed to calculate an average number of employees over the year. Indeed, this information is now available within the Eurostat planning. Average number is calculated for emplyees only. For self-employed, we improved the estimation, but we still use the number of self-employed persons at 31/12 instead of average.

From 2021, SBS and BD use the same business register to delineate their respective scope, so the consistency between SBS and BD has been improved.

Nevertheless, the difference remains significant. We plan to reduce this gap by working on micro enterprises and the defintion of active unit

The aggregates are always consistent with their main sub-aggregates (e.g. the total for different size classes, NACE, CPA, NUTS...). The modality FRZ (extra region) of NUTS is not taken into account for Eurostat validation checks. That explains messages appearing as "error" which are not relevant.

Esane was especially implemented to reduce businesses response burden. The use of administrative fiscal data enabled to drastically reduce the size of surveys questionnaire. An optimization of sampling design was carried on that enabled to reduce the size of the sample by half in take-some strata while keeping the same precision for estimates. Finally, a new coordination method has been introduced in 2013 : it enables coordination of all business surveys together, so as to smooth response burden between small and medium businesses while respecting each survey' sampling design. The method is described in more details in "Sampling coordination of business surveys conducted by Insee" a paper by Fabien Guggemos and Olivier Sautory, presented at the fourth International Conference on Establishment Survey (Montreal, 2012).

No revision policy has been put into place. However, revisions occur at least once a year when provisional data published in November is replaced by definitive data. Revisions also occur when methodological or scope changes are made. We estimate these revisions to approximately represent 0.2 point of the value added growth rate.

The data used for compiling the preliminary results comes from a first delivery of the fiscal files (arriving in Insee in July N+1). 

Definitive data are elaborated with both supplementary files from the fiscal source and data coming from the surveys.

The sources are statistical survey combined with administrative source. The sample survey is conducted on a sample of legal units, and the administrative information is exhaustive.

Annual data collection.

• The survey

It was initially an exclusively postal survey by questionnaire. From the reference year 2012, it was possible to answer via internet by filling out a form identical to the paper questionnaire, with a specific file format. Web questionnaire is available since the reference year 2018.

The number of units surveyed is about 160 000.

Dashboards are used to follow-up non-response.

Enterprises are contacted again by mail twice if necessary and the fourth time with the threat of a financial sanction.

Specific e-mail are sent to very large companies if they did not answer the first time.

• Administrative source

We use an extraction from the fiscal database, transmitted by fiscal administration

Checks consist in:

• completeness checks (data integrity rules)
• validity checks (internal consistency)
• plausibility checks 

The editing system is based on selective editing.

For a description, see:

https://www.unece.org/fileadmin/DAM/stats/documents/ece/ces/ge.44/2009/wp.41.e.pdf

https://www.unece.org/fileadmin/DAM/stats/documents/ece/ces/ge.44/2012/25_France.pdf

Moreover, this selective editing procedure carried out on a continuous flow basis during the production process is supplemented by a final output editing process.

Finally, Edit tool controls are computed before sending data to Eurostat (format and file structure checks, Intra-dataset checks, Inter-dataset checks, times series checks).

Inference (grossing-up):

Statistical combined estimates are used, described in detail in:

https://doi.org/10.1515/jos-2015-0036

Imputation is used to treat item non‑response for all take‑some units in the survey. For the take‑all units as well as the administrative data, imputation is used to treat both item and unit non‑response.

Imputation by historical data – using the previous reference period – is carried out as far as possible, otherwise, the mean, median or ratio imputation by class method is used.

In case the enterprise has a fiscal year different from the calendar year, the data of this fiscal year are affected to the calendar year including the most important part of the fiscal year

In 2021, some activity sectors have a strong evolution due to rebound in economic activity, after Covid period.