The creation of an experimental longitudinal database of business micro-data

Abstract

In recent years there has been an increasing need to extend the research potential of micro-level business data by allowing dynamic analyses via panel datasets. One option, using official statistics to build a firm level data panel, is to exploit archival data. However, an endemic problem with this approach is the management of firms transformational events: mergers, splits and spin-offs. This paper considers a way to address this problem. It describes the operative procedures utilized in building an experimental panel starting from firm-level databases available from the Italian National Institute of Statistics (Istat).

Keywords

Panel building event of transformation firm-level data business register event register

1. Introduction

There is growing interest on the part of researchers, policy makers, and citizens for increasingly complex databases, in particular, those including dynamic relationships. For this reason, there is a need to widen statistical information and database development in order to extend the research potential of official statistical data. Panel data are required more and more because they allow dynamic analysis and control for unobserved heterogeneity via panel econometric methods. Hence, there is a need for official statistics to meet these requirements. Therefore, the availability of longitudinal data is a key issue, and this paper focuses on this problem.

In order to provide the answers to this problem, namely, making the best use of data collection resources, widening available statistical information, and, at the same time, reducing or containing the response burden, governmental statistical agencies need to redesign the process of providing data [1, 2]. A strategic approach, in the reengineering context, consists of combining different data sources [3]. In this study, we focus on the need to widen statistical information and increase the availability of longitudinal data.

Historically, the creation of standard panel data has usually involved the need to conduct a panel survey, implying the requirement to sample a number of firms in the initial year, then resample them in subsequent years, in the process trying to maintain sample representativeness with respect to the target population. Among the problems associated with this kind of survey are the costs and time needed to collect data, the impossibility of collecting data prior to the survey’s starting year, and the sample attrition which may cause a loss in information. This attrition may be caused by firm mortality, a non-response, or firm transformation events. Transformation events consist of mergers, demergers, split-offs, and spin-offs. But, of course, in many countries, including Italy, structural panel data are not collected directly (i.e., with a business survey originally planned to have a longitudinal dimension); hence, this kind of panel information is often lacking.

There are various ways to address this problem, including the use of archival data collected in surveys or through administrative sources in order to form both follow-back and catch-up panels. Following Kessler and Greenberg [5], a follow-back or retrospective panel consists of the selection of a cross-sectional sample in the present and then utilizing archival data to locate the same observational units at an earlier time point in order to create a longitudinal dimension of the sample itself. In contrast, a catch-up panel is formed by selecting a past sample from an archival source and then identifying the same observational units in the present in order to re-observe them. However, in the present case and in contrast, what will be considered is the creation of a catch-up panel developed on the basis of archive data availability only, without the need to collect data or re-interview the firms. In our case, the catch-up panel involves the selection of one or more cross-sectional datasets from an archival source (for instance, a survey on enterprises’ structural characteristics [4]) at some time in the past (beginning in year $t_{0}$ ). The observation units in the first cross-sectional survey dataset are then identified in the following year (year $t_{1}$ ) and so on until the survey data are available for each year until the last year (year $t_{p}$ ). Effectively, this process involves tracking each observed firm over time, thus simulating an actual longitudinal survey. This particular kind of catch-up panel can be compared to a secondary analysis of an existing set of panel data. The idea of a catch-up panel is particularly attractive in those instances in which the researcher manages to isolate a source of baseline archival data which is especially rich in information [5].

While the panels built in this way are less expensive than a standard panel, yet can make use of past data, the problem of natural attrition is not solved. Furthermore, an additional difficulty is to distinguish between the actual death of a firm and its transformation into another when the observations are recorded many years after the time at which the defining event occurs. Many attempts have been made to create longitudinal databases starting from archive data pertaining to the most recent year; including both in Italy [6, 7] and elsewhere, such as specifically for American households [8] and firms [9].

In the case of the present study, an in-depth analysis has been carried out in order to taken into account the effect of transformation events, the aim being to insure the longitudinal consistency of the panel and avoid the information loss that would otherwise occur. Fundamentally, the objective here is to describe how different sources can be combined to obtain longitudinal data and supply a best practice to those who want to build a similar longitudinal database. The study describes the operative procedures utilized in building an experimental panel starting from firm-level databases available from the Italian National Institute of Statistics (Istat). In particular, we are going to present the conceptual framework, methodology, and procedures used to create a longitudinal database containing the main economic variables for Italian firms.

The main feature of this approach is its treatment of M&As (mergers and acquisitions), or in other words, the transformation events (i.e., all events, such as mergers, demergers, and split-offs, which transform firms and may make their continuity over time questionable).

The paper is laid out as follows. Section 2 of our paper describes the panel construction sources and concepts, Section 3 takes into account different approaches for building this kind of longitudinal database and discusses the advantages and disadvantages of each. Section 4 focuses on the applied methodologies. In Section 5 some statistical analyses of the information on quality are supplied. Section 6 concludes.

2. Experimental Panel database main facts and building blocks

2.1 Main facts

The firm transformation event, such as a merger or split-off, is an emerging phenomenon that dates largely from the Nineties. In the last decades, particularly, the number of such events has increased constantly at the world level. Notice, for instance, in Fig. 1 the number of cross-border M&As [10].

Figure 1.

Number of cross-border M&As – 1990–2013.

As seen in Fig. 1, the M&A activity presents several waves which form the substantial ascending movement. The last wave of merger and spilt-off activities started in 2003 and is still continuing, notwithstanding a moment of slowdown due to the last crisis [11]. In Italy, as well, there is an increase in the number of events prior to the last financial crisis (Fig. 2).

Figure 2.

Number of firm transformation events – 2005–2103.

Figure 3.

Number of firm transformation events by firm size class – 2005–2103 (Index numbers; base $=$ 2005).

The increment in M&A activities is a feature mainly of larger firms (50 persons employed and over) rather than the smallest (under 50 persons employed). In Fig. 3, we can see that the increment is substantially positive for both groups until 2009. The trend then presents a decrease after 2009 for both groups even as the trend for larger firms dominates, over the entire period, the others. We need to bear in mind that larger enterprises are mostly incorporated, while the smaller enterprises are more commonly privately held.

Figure 4 presents the pattern of firms’ transformation events broken down by type of event. Three different types of events are considered: Transfer (transfer of a part of one or more enterprises), Cessation (firms that are completely absorbed in a new enterprise), and Others (other events). This last type is not represented here.

Figure 4.

Number of event of firm transformation by type of event – 2005–2103 (Index numbers; base $=$ 2005).

It is interesting to note that in a period of positive economic conditions, the greater increasing rate is shown by the Transfer type, while, in negative economic conditions, the Cessation events prevail. If we consider the size (in terms of persons employed) of firms involved in transformation events (Fig. 5), we can better understand the importance of including event management in the process of panel creation.

Figure 5.

Persons employed by the firms involved in transformation events – 2005–2103.

Notwithstanding that the total number of persons employed by firms involved in transformation events varies greatly from year to year, it remains over 1 million for every year starting in 2005. Furthermore, even if there appears to be a decrease after 2009, the firms affected after that year seem to be larger. Notice, as a result, that even if there are only small gains in representativity, in terms of firms involved in the panel, there is a greater gain in representativity in terms of economic variables.

2.2 Building blocks

The first action in building the Experimental Panel (EP) is to identify all firms included in the Italian Statistical Business Register (ASIA) for all years in a given period. The main purpose of the panel is to represent medium-large Italian firms (enterprises with 20 persons employed and over) for all industrial and service sectors (excluding monetary and financial intermediation) in the period from 2008 until 2011. In order to determine the longitudinally persistent core units, it is necessary to drawn all the links between firms included in ASIA from 2008 to 2011. The Experimental Panel is based mainly on the cross-sectional business register with the integration of administrative micro-data for ensuring the matching of economic items over time.

The cross-sectional data that are included in the EP show a widespread overlap over time and a relevant longitudinal component. Three different data sources are included. The first source is the Istat ASIA, and the second is the Events Register, a database in which all information on enterprises changes and transformations are registered. The Event Register is created on the basis of the ASIA Register and coordinated with it. The last source is the annual financial reports of incorporated firms collected by Chambers of Commerce (Balance sheet dataset, BIL).

When an historical complex archive of data is available (for instance, administrative data sources available for a long time period) this mass of information may be exploited and organized longitudinally in order to follow single firms and understand their economic and behavioural changes [12]. One method for creating a panel consists of merely linking the firms’ codes year after year. However, this solution ignores the problem of mergers and split-offs, which dramatically increased before the last crisis (above all for larger firms). The result is a longitudinal database incorporating a greater attrition, which affects the representativity of economic variables.

An alternative method is to create a longitudinal database by considering also transformation events. Transformation business events registered in the Event Register, such as mergers and split-offs, have been considered in the panel, following a backward-looking perspective. Detailed criteria are defined in order to take business transformations into account (in any case, newly formed firms are included, so the panel can be used to analyse the entry and exit dynamics of firms). The enterprise transformations are classified into different groups according to transformation type, such as legal form modification, transfer of property, and similar events. The single enterprise status changes or may be transformed over time. In EP, all firms involved in business transformation events are considered, and they can be clustered by means of appropriate codification. The codification permits the longitudinal data of all relevant firms to be included in the panel and permits the events’ management. Changes may affect:

•
the dimensional structure (number of persons employed),
•
the economic activity,
•
the legal form (firm name, fiscal code).

Figure 6.
Clustering events scheme.

Moreover, many events can imply the death of an enterprise. More detailed descriptions of the three databases are included in Appendix A.
3. The choice of the correct approach

We can consider three different approaches for preparing panel data from an archival source: cross section, i.e., a series of cross-sections bound together [13]; a simple codes panel; and, finally, a panel which considers firms’ transformation events [14]. The first approach consists simply of taking the same database from different years and comparing them, for instance, by aggregating the firms by geographical area or by size class. In this case, we do not need a punctual identification of firms by means of an ASIA code. The second approach is based on the link of every single unit by firms’ identification codes (ASIA code). In this last case, if a merger occurs, at least one code is lost (firm). Here, we consider also longitudinalization through clusters of firms involved in transformation events; therefore, we consider all transformation events clustering the involved firms by means of a special code: the code identifying group (CIG), building an economic super-entity which may traced back to the firm(s) existing in the starting year. By means of this procedure, two targets can be achieved: assessing the economic continuity of a single firm and maximizing the representativity of firms’ panel data.

The simple code panel (COD), in which the transformation events are not considered, has the merit of simplicity, shortening preparation time and allowing simple conceptualization. This approach of not taking transformation events into account does not guarantee the actual continuity (both legal and economic) of firms, generating an important misrepresentation with effects on analytical conclusions. The cross-section approach can be considered here as a benchmark methodology against which to compare the others. The cross-section approach is the one which better fits firms’ actual legal and economic developments; but, in this case, we cannot, via an economic point of view, correctly link longitudinally these entities.

In Fig. 6, we present the conceptual scheme which we follow in order to cluster firms involved in an illustrative series of events during a 4-year period. The codes in bold are the new longitudinal codes created for clustering the firms involved in transformation events, while the codes in italics are the usual ASIA codes. Firm clusters are created each year. The relevant quantitative (economic) data will be treated opportunely [15]; for instance, they may be summed up, while, for qualitative data, some prevalence criteria will be adopted. In this case, we obtain four clusters representing the virtual firms spanning the considered time period and representing both the six starting firms and the seven firms present in the last period. If we had not utilized the clustering approach and a code panel had been built instead, we would have had only two firms in the code panel itself (H and M) with a loss of information.

Analysing the effects of different choices, we can implement a longitudinal analysis of the data. We consider three coordinates (or proprieties) which our longitudinal data should have:

•
the application criteria of simplicity (easy to treat with respect to the available databases),
•
the time coverage of the database (we can obtain this coverage as a result of applying different methodologies),
•
the consistency with actual facts we register.

Each approach seems to have strengths and weaknesses with respect to the above-mentioned factors. Cross-section presents greater consistency and treatment simplicity with respects the others two approaches, but it, obviously, does not have any longitudinal analytical power.

The code panel is easy to implement, but it fails to express variable levels consistent with the real ones, and it does not grant a strong longitudinality and coverage to the firms included in the set of interest.

The panel built by means of clustering firms involved in transformation events gives more longitudinal reliability, but sometimes it fails in creating a consistent framework with the actual facts; moreover, it may be very difficult to implement. Figure 7 supplies an exemplified synthesis (in terms of the tree dimensions described above) of the considerations discussed previously.

Figure 7.
Outlines of the three different longitudinal approaches’ properties.

The values for the three factors reported in Fig. 7 are, obviously, merely indicative and subjective. The reflections reported here constitute the basis for the actual building of a longitudinal dataset by means of the grouping of firms involved in transformation events and following the scheme presented above in Fig. 6. In particular, we first built a “grid” of firm clusters by using the event register, and then we apply the “grid” to the whole ASIA register. Finally, we link the list of enterprises obtained in this way to the economic data coming from financial reports on limited enterprises provided to Chambers of Commerce over a 4-year span (from 2008 to 2011).

The methodology described in this paper is aimed only at creating a longitudinal scheme of firms by means of events and register data treatment. We considered firms included in Istat Business register, ASIA, according to the definition in the Eurostat Regulation (Council Regulation (EEC) No 696/93) of economically “active” (i.e., it has produced sales or it has employed personnel).
4. Experimental Panel construction procedures

4.1 Introduction

The panel creation consists of different steps. The first step (clusterization) creates the longitudinal scheme of the firms involved in transformation events according to the scheme shown in Fig. 6. In the second step (linkage), the longitudinal scheme is linked to ASIA and Balance sheet databases year by year. The third step creates, under some continuity principles, the firms’ longitudinal trajectories and detects the events implying firms’ mergers and split-offs. The fourth step consists of managing particular set of firms and events. All elaborations are performed by utilizing the SAS package V9.3.

Table 1
Change of legal form nomenclature

4001	Transfer by means of transformation in a new firm
4002	Cessation by means of transformation in a new firm
4003	Transfer by means of transformation in a existing firm
4004	Cessation by means of transformation in a existing firm
4005	Cessation by means of a legal form change
4101	Birth from transformation of a changing firm
4102	Birth from transformation of a ceasing firm
4103	Acquisition from transformation of a changing firm
4104	Acquisition from transformation of a ceasing firm
4105	Birth from legal form change

Table 2

Transfer of property nomenclature

4011	Cessation by means of succession on death
4012	Cessation by means of takeover
4013	Transfer by means of business/enterprise lease
4111	Birth by means of succession on death
4112	Birth by means of property transfer
4113	Birth by means of business/enterprise lease

4.2 Clusterization

The sources utilized for this step are: ASIA, the Balance sheet database, and the Events Register. In the Events Register, the enterprises’ transformations events are classified into two groups: the first related to legal form modifications, and the second related to transfer of property (codes presented in Tables 1 and 2). Due to the various events which can occur, a firm code may refer to an entity which is notably different from year to year. The economic indicator can be strongly influenced by this issue. In order to link the firm/entity over time, the Events Register provides the following information:

•
Code 1, origin firm code (for instance, the seller),
•
Code 2, destination firm code (for instance, the buyer),
•
Event Code, according to the above-mentioned scheme,
•
Event Date, which provides month and year,
•
Quality Code.

The combination of the origin and destination and the event codes enables a complete registration of all possible transformation events which may occur. It is also possible to use this information for panel construction purposes.

The clusterization process entails the need to trace firms’ codes involved in transformation events all along the panel time in order to avoid:

a)
a decrease in the rate of participation in the panel (above all for larger enterprises),
b)
a bias due to false links which actually group very different firms.

We follow a longitudinal approach aiming to minimize the loss of information (unlike the firms’ codes approach) and create longitudinal links to detect the firms involved in transformation events. It is necessary to consider four fields:

a)
Event’s date,
b)
Longitudinal code,
c)
Firm’s code,
d)
Event’s code.

The idea is to create, forward in time, clusters of firms involved in events; the clusters will have information in the fields described above and will be marked by longitudinal code. The cluster, by applying the appropriate continuity criterions, will detect the firms that exist permanently all along the panel, although transformed. The implementation of a longitudinalized database entails the application of the following rules:

1)
A cluster is created if interacting firms have longitudinal codes (CIG) equal to zero (namely, when there is not any other event registered for these firms).
2)
When two longitudinal codes (different from zero) interact, the greatest (namely, the code that represents the larger number of firms) or older prevails. In this case, a code disappears, and the surviving code represents a larger number of firms.

The operating method used to create the longitudinal codes (CIG) starts by using the file containing the Event Registers merged for various years. In particular, the file can be presented as in Table 3.

Table 3
CIG code creation

Year Event year Code1 Code2 Cig

1 1 051 001 1

1 1 052 002 2

1 1 006 003 3

2 2 057 006 4

2 2 059 008 5

3 3 002 010 6

3 3 062 001 7

3 3 051 012 1

4 4 001 015 8

4 4 057 016 4

4 4 059 018 5

4 4 051 017 1

The CIG code column is added by our information system. A CIG code is given to all events involving two firms in a consecutive way. When a firm involved previously in an event matches again with another firm in the following years, the newly involved firm takes the CIG of the first one (look at CIGs 1, 4, and 5 in the example showed in Table 3).

The second operative step is to split off the file into two other files: one containing the firms labelled with code1, and the other one containing the firms labelled with code 2. After we append the files, the event database described above shall be transformed in the database showed in Table 4.

Finally, we obtain identification codes for the firms and a code identifying the cluster of firms involved in one or more transformation events. This procedure separates the codes of acquiring firms from those of acquired firms.

Table 4
Firm code verticalization

Year Code1 Code2 Cig

1 $\cdot$ 001 1

1 $\cdot$ 002 2

1 $\cdot$ 003 3

1 051 $\cdot$ 1

1 052 $\cdot$ 2

1 006 $\cdot$ 3

Year Code Cig

1 001 1

1 051 1

1 002 2

1 052 2

1 003 3

1 006 3

A firm, as we have seen in the previous example, can match another firm in one year and others firms in the following years. Moreover, a firm can have more than one event in the same year with one or more other firms. We called the first case Alpha, and the second case Beta.
4.2.1 Alpha case

Year	Event year	Code1	Code2	Cig
1	1	051	001	1
1	1	052	002	2
1	1	006	003	3
2	2	057	006	4
2	2	059	008	5
3	3	002	010	6
3	3	062	001	7
3	3	051	012	1
4	4	001	015	8
4	4	057	016	4
4	4	059	018	5
4	4	051	017	1

Year	Code1	Code2	Cig
1	$\cdot$	001	1
1	$\cdot$	002	2
1	$\cdot$	003	3
1	051	$\cdot$	1
1	052	$\cdot$	2
1	006	$\cdot$	3

Year	Code	Cig
1	001	1
1	051	1
1	002	2
1	052	2
1	003	3
1	006	3

In this case, we have to consider the firm’s matches in each year which are included in the panel. Hence, we match, using CIG and codes, firms involved in transformation events in year $t_{0}$ with firms involved in transformation events in year $t_{1}$ (Table 5). We detect the firms present in both years and then force the CIG of the first firm onto the second firm (involved in an event in year $t_{0}$ ). This procedure is then replicated sequentially for each year of the panel.

Table 5
Alpha case (codes from Table 4)

Once we check for all codes matching forward, we need to check backward to change all of the CIGs which, in that period, had more than 1 event. Let’s take, for instance, the case1 exemplified in Table 6. We merge first by using CIG, so we detect all codes that have had events with firms with the same CIG when passing from time $t_{0}$ to time $t_{1}$ . Contemporaneously, we give to all firms the time $t_{0}$ CIG code. In the case illustrated in Table 5, we can track over time the events which have affected different firms over four years. The solid line arrows indicate firms which have had events in different (distant) years and for which the CIG codes have already been changed and made equal to the first one. We can identify two groups shadowed in grey and white (but others can be identified). The dotted line arrows indicate, as an example, firms and groups linked by events in a distant year but for which the change of CIG remains to be done. The groups are coloured in bright grey.

Table 6

A real case

cig t ${}_{0}$	Code t ${}_{0}$	Cig t ${}_{1}$	Code t ${}_{1}$
68	4859	68	4859
68	619053	68	628967
5503	628967
5503	8907

cig t ${}_{0}$	Code t ${}_{0}$	Cig t ${}_{1}$	Code t ${}_{1}$
68	4859	68	4859
68	619053	68	628967
68	628967	68	619053
68	8907	68	8907

Initial

Modified

Thus, Table 6 Firm 4859 had an event in year $t_{0}$ with Firm 619053, giving it the CIG 68. Then, Firm 4859 had an event with Firm 628967 in year $t_{1}$ , giving it, in the same way, the CIG code number 68. Now we have to go back and match the code 628967 with other firms involved in events in the year $t_{0}$ in order to ascertain if it is registered under another CIG for that year. In this case, the code 628967 in the year $t_{0}$ is registered under the CIG number 5503, signalling an event with another firm. For both firms, it is necessary, now, to force the CIG code 68 in the year $t_{0}$ .

4.2.2 Betas 1 and 2 cases

In these cases, the same firms are involved in more than one event in the same year. These cases may be classified under two typologies: events involving the same firms’ set (Beta 1) and events involving different firms’ set (Beta 2).

In Table 7, both cases are described. In the Beta 1 case, we transform the original codification (CIG+ASIA code) by means of the auxiliary field $c$ . The original CIG is transformed into the new CIG by forcing it with $c$ . Then we eliminate the duplicated classification code. The CIG code is transformed to 009 for Firms 001 and 003 because it is the first event.

In the Beta 2 case, we operate in the same manner. For instance, the CIG code 012 is changed to 009 for Firms 004 and 005 because of a preceding event involving Firm 001 belonging to CIG 009. The transformed classification codes present also, in this case, duplications which have to be eliminated.

Table 7
Cases Beta 1 and Beta 2

4.3 Linkage with business register

Once we obtain the clusterization scheme, we have to link the firm code with the Business Register files concerning the different years included in the panel. In Table 8, the procedure used to link the file containing the clusterized events (G) and ASIA is described.

Table 8
Linkage with ASIA

First, the file G is linked with ASIA at time $t_{1}$ . Not all firms included in G are also present in ASIA because many firms were born in the years after $t_{1}$ . After the linkage, only firms which are present in year $t_{1}$ are considered (file W). File W will be appended to file $N_{e}$ $<$ yeart ${}_{1}$ $>$ a file containing all ASIA firms in year $t_{1}$ which are not involved in events.

Also, a CIG code will be created for the firms in $N_{e}$ $<$ yeart ${}_{1}$ $>$ in order to identify firms having events and firms not having events. The procedure is then replicated for the following years.

We obtain a number of files equal to the number of the years included in the panel. These files are derived from ASIA and present the additional information given by CIG (clusterization code). Now, the problem is to merge these files in order to build an actual panel.

Some CIGs can be linked from one year to another because the firms in the cluster present events in both years (shadowed area in Table 9). There are two firms sets which cannot be linked because they did not have events in other years. These two sets should be linked to the corresponding firms in the other file (year) in order to have a complete longitudinal dataset.

In other words, we have to link single-year clusters with the cluster firm codes in the following and precedent years. It is necessary to carry forward (or backward) the clusters’ codes which disappear in the following year or are born in an intermediate year. These clusters represent, for instance, firms which have had only one event in the period. This procedure can be performed only if the whole set of Business Registers (or Balance sheets) is available online and contains all of the information for each year of the reference period.

Table 9

Linking year after year

So, a longitudinalized file is obtained, and it contains the CIG codes for the reference period. Having, finally, obtained the firms’ codes for firms present in the period and having taken into account the transformation events, we have fulfilled the necessary conditions to link the Balance sheet dataset. This final file already contains the firm codes coming from ASIA, so it is easily linkable to the financial data contained in the Balance sheet dataset. Not all firms present in the ASIA Register can be linked to economic data because only a part of them (the corporations) are obliged to file financial statements.

Table 10

Final panel structure

4.4 Managing the experimental panel

Now it is necessary to perform a further step to control and manage the panel described above. We need a grid which allows the tracking of the longitudinal path of every firm in relation to the others. In particular, we need a flag to point out the presence of a firm in a given year and a further two flags to identify cases of merger and split-off. In addition, it will be necessary to detect exactly which firm(s) split-off or merge in the case of a multi-firm cluster.

A data structure having these characteristics is created after having performed some complex data elaborations. An exemplification of this database is presented in Table 10. The table presents real data2 taken from the final prototypical database created for the period 2008–2011. In Table 10, code08 represents the code for a firm involved in the events of 2008, p08 signals the presence of a firm in the year 2008, s0809 signals that split-off event involved the firm between 2008 and 2009, while m0809 signals that a merger event involved the firm between 2008 and 2009.

In Table 10, we can note many explicative cases related to five clusters identified by shadowed areas. The first cluster (number 68) consists of four firms that became two in 2009 due to the acquisition of the second firm by the first one and the fourth by the third. These events are marked by the flag m0809 (m is for merge) that pairs, on the one hand, the first and the second firms (flag code 1) and, on the other hand, the third and fourth firms (flag code 2).

The cluster 25721 presents two cases of split-off. Code 23259795 splits off the code 26415252 in 2010, and the code 18924855 splits off codes 26412894 and 26412896 in 2011. These events are described by flags s0910 and s1011 (s is for split-off). Flag code 1 (for s1011) marks codes 18924855, 26412894, and 26412896, while flag code 2 (for s0910) marks codes 23259795 and 26415252.

The building of an experimental panel with the properties described above allows us to manage and define a longitudinal database matching the various requests coming both from researchers and scholars.

If, for instance, the query is related to the extraction of all firms which disappear and, at the same time, are born with another code, due to the occurrence of a specific event, it is possible to retrieve the relevant firm record and extract the required codes by summing and processing information included in the final panel structure. We can easily obtain the scheme presented in Table 11.

Table 11
Scheme of an elaboration for firms which completely slit-off or merge

cig	code08	code09	code10	code11	p08	p09	p10	p11
100	2300				1
100		2999	2999	2999		1	1	1
100	4010				1
200	1100	1100	1100		1	1	1	1
200			3300	3300			1	1
300	2100	2100			1	1
300	2200				1
300		2400	2400	2400		1	1	1
300			2500				1

This is an effective approach for detecting firms which only apparently disappear. The treatment of these cases may be implemented by applying different continuity criterions.

5. Experimental Panel description

The relevant period in which to implement the experimental panel includes the years from 2008 to 2001. Approximately 200,000 firms were involved in transformation events well distributed over those years (Table 12).

Table 12
Distribution of firms involved in transformation events by year (All firms)

Event year	Frequency	Percent	Cumulative frequency	Cumulative percent
2008	54,047	27.59	54,047	27.59
2009	52,127	26.61	106,174	54.20
2010	46,696	23.84	152,870	78.04
2011	43,019	21.96	195,889	100

Table 13

Firm distribution involved in events by event type (All firms)

Event code	Frequency	Percent	Cumulative frequency	Cumulative percent
4001	21,455	10.95	21,455	10.95
4002	47,965	24.49	69,420	35.44
4003	17,514	8.94	86,934	44.38
4004	45,127	23.04	132,061	67.42
4005	30,714	15.68	162,775	83.10
4011	32,247	16.46	195,022	99.56
4012	6	0	195,028	99.56
4013	392	0.2	195,420	99.76
Others	469	0.23	195,889	100

The CIG number is around 90,000, while the average number of firms associated with a CIG is approximately 2.2. For more than 90 percent of the cases, the CIG dimension is equal to 2 (Fig. 8).

Figure 8.

CIG distribution by size (number of firms involved).

Table 13 presents firms’ distribution by event type. We note two peaks for the distribution corresponding to cessation events (codes 4002 and 4004). These kinds of events represent approximately half of all firms involved in events. Conversely, the transfer events share is around 20 percent.

We also perform the same elaboration for corporations only. In this case, we take into account only the events involving corporations on both sides of an event relationship (i.e., in the Event Register, both code 1 and code 2 represent corporations). Table 14 shows the distribution of corporations involved in events by event year. Also, in this case, the firms are distributed equally.

Table 14

Distribution of firms involved in events by year (corporations only)

Event year	Frequency	Percent	Cumulative frequency	Cumulative percent
2008	4,344	27.33	4,344	27.33
2009	3,988	25.09	8,332	52.42
2010	3,888	24.46	12,220	76.88
2011	3,675	23.12	15,895	100

Table 15

Firms distribution by event type (corporations only)

Event code	Frequency	Percent	Cumulative frequency	Cumulative percent
4001	6,979	43.91	6,979	43.91
4002	149	0.94	7,128	44.84
4003	7,132	44.87	14,260	89.71
4004	994	6.25	15,254	95.97
4012	6	0.04	15,260	96.01
4013	297	1.87	15,557	97.87
Others	338	2.13	15,895	100

Figure 9.

CIG distribution by size (corporations only).

Figure 9 shows the CIG distribution by size in terms of firms involved. Also, in this case, around 90 percent of CIGs present a size equal to 2 forms and an average size of 2.3 firms. There are approximately 7,000 CIGs.

The events distribution by event type for corporations is diametrically opposed to the one calculated for all firms (Table 15). In this case, the transfer events (codes 4001 and 4003) prevail over the cessation events.

An elaboration was carried out in order to detect the coverage rate of the experimental panel with respect to the target population data coming from the ASIA register and SBS3 estimations. The target population is represented by firms with 20 persons employed and over in Italy. Table 16 presents the coverage rates of the experimental panel with respect to the target population for some important economic variables, such as turnover, intermediate costs, personnel costs, value added, persons employed, and number of firms.

Table 16

Coverage rate of the experimental panel with respect to the target population – Years 2008–2011 (percentages)

Year	Firms	Turnover	Intermediate costs	Personnel costs	Value added	Persons employed
2008	66.8	82.4	81.9	83.0	85.0	81.0
2009	70.6	84.4	83.6	84.6	85.9	83.0
2010	62.3	88.4	87.8	88.2	89.2	84.0
2011	67.8	89.7	89.5	89.1	89.6	86.7

Sources: ASIA, BIL, SBS estimations.

In general, the rate of coverage of the panel is adequate (approximately 2/3 of the total population of firms). The lower rate of coverage is due substantially to the under coverage of financial statements with respect to the corporations obliged to file them. The coverage rate improves when using economic variables which present values for 80 percent up to 90 percent of firms. This makes us confident of a good representativity of the experimental panel. In Table 17, the coverage rate for the codes panel (a panel created matching the firms’ codes over time without taking in account the transformation events) is presented.

Table 17

Coverage rate of a codes panel with respect to the target population – Years 2008–2011 (percentages)

Year	Firms	Turnover	Intermediate costs	Personnel costs	Value added	Persons employed
2008	66.5	75.6	74.9	78.4	80.1	77.5
2009	70.2	77.6	76.7	79.6	80.5	79.2
2010	61.8	83.6	83.0	84.6	85.4	81.5
2011	67.3	87.6	87.3	86.5	86.9	84.4

Sources: ASIA, BIL, SBS estimations.

The coverage rates (with respect to firms with 20 persons employed and over) for all variables are higher for the experimental panel than they are for the codes panel. It should also be noted that there is a progressive decrease in these gaps from 2008 to 2011. While the number of covered firms increase from codes panel to experimental panel is very small in relative terms, the coverage improvement in terms of economic variables is relevant.

6. Final remarks

The availability of a longitudinalised grid of events represented by the experimental panel is very important in order to exploit the additional information coming from the Event Register. While all firms are linked by a code which detects when a firm belongs to a group containing all of the firms which have had at least one event in common during the period, it is possible to imagine the use of different treatments to trace the economic variable trends correctly. We can, for instance, aggregate the economic variables in different ways in order to analyse specific economic facts.

Hence, we can sum the characteristics of the firms belonging to a group or adopt a prevalence criterion. We can assess other, more advanced, criteria for the aggregation of data. We can consider, for instance, the merging (or splitting-off) of firms as elements of a virtual “parent” super-firm which includes these firms in a manner similar to “divisions” or “kind-of activity-units” (KAU).4 Then it will be possible to operate both on “parent” firms and on firms’ “divisions”.

All of these kinds of approaches, as well as the extraction of particular events which may be relevant for a particular economic analysis, can be performed thanks to the implementation of an experimental panel which takes into account transformation events. The use of the experimental panel allows:

the consistency of longitudinal data to be improved by tracing in a better way the individual stories;

an important action for data recovery to be performed for firms which change ASIA codes. This action can be demonstrated by the representativity comparison between the usual codes panel and the experimental panel.

The experimental panel covers the period spanning 2008 to 2011, but is easily expandable to more extended periods (at least from 2000 up to the present time). The procedures have, also, been easily expandable to cover all kinds of firms (other than corporations). Here, we highlight some results of the panel including unlimited firms. However, we have to observe that, in this case, not all economic information is available.

Footnotes

The codes are randomly modified.

Data transmitted to Eurostat according to SBS (structural Business Statistics) Regulation (EC) No 295/2008.

The definition of KAU is contained in Regulation (EC) No 696/1993.

Acknowledgments

The author is grateful to Fabrizio Solari for the help in developing SAS macros and to Mirella Morrone for her clarifications about ASIA Event Register. The usual disclaimers apply.

Appendix

Glossary and references

References

Black

Taylor

and Zaber

. Empirical evidence in the study of labor markets: Opportunities and challenges for a new household survey. Journal of Economic and Social Measurement. 2015; 40(1-4): 97-122.

Moffitt

Schoeni

Brown

Chase-Lansdale

Couper

Diez-Roux

, et al. Assessing the need for a new nationally representative household panel survey in the United States. Journal of Economic and Social Measurement. 2015; 40(1-4): 1-26.

Biffignandi

Nascia

and Zeli

. Combining survey and administrative data in Italian business surveys. Journal of Economic and Social Measurement. 2016; 41(1): 67-83.

Zeli

. A mixed-mode survey using mail and web questionnaires: the SCI case – First analysis. Istat Working paper. 2013; 1.

Kessler

Greenberg

. Linear Panel Analysis: Models of quantitative Change. Cambridge: Academic Press; 1981.

Oropallo

Inglese

. The Development of an Integrated and Systematized Information System for Economic and Policy Impact Analysis. Austrian Journal of Statistics. 2004; 33(1-2): 211-236.

Oropallo

. Nuove basi dati panel di microdati d’impresa: unità di analisi, integrazione delle fonti e imputazione. Istat workshop: Micro dati per l’analisi della performance delle imprese: fonti, metodologie, fruibilità, evidenze. 2015; Rome, 21

{}^{\rm st}

September.

Rivera Drew

Flood

Warren

. Making full use of the longitudinal design of the Current Population Survey: Methods for linking records across 16 months. Journal of Economic and Social Measurement. 2014; 39(3):121-144.

Jarmin

Miranda

. The Longitudinal Business Database. Center for Economic Studies. U.S. Census Bureau. 2002; Working Papers 02-17.

10.

Unctad. World Investment Report. 2015.

11.

Duysters

Cloodt

Schoenmakers

Jacob

. Internationalisation Efforts of Chinese and Indian Companies: An Empirical Perspective. Tijdschrift voor Economische en Sociale Geografie. 2015; 106(2): 169-186.

12.

Frees

. Longitudinal and Panel Data: analysis and application in the social sciences. Cambridge: University Press; 2004.

13.

Grazzi

Sanzo

Secchi

Zeli

. The building process of a new integrated system of business micro-data 1989–2004. Journal of Economic and Social Measurement. 2013; 38(4): 291-324.

14.

Nardecchia

Sanzo

Zeli

. La costruzione di un panel retrospettivo di micro-dati per le imprese italiane con 20 addetti ed oltre dal 1998 al 2004. Documenti Istat. 2010; 7.

15.

Zeli

Nardecchia

, From BR to SBS surveys. Timing and events. presented at Eurostat Workshop: Survey Frames and Feedback: How the Business Register works with Survey Areas in the Statistical Office. Dublin, 2012; 24

{}^{\rm th}

–25

{}^{\rm th}

April.

Year	Event year	Code1	Code2	Cig
1	1	051	001	1
1	1	052	002	2
1	1	006	003	3
2	2	057	006	4
2	2	059	008	5
3	3	002	010	6
3	3	062	001	7
3	3	051	012	1
4	4	001	015	8
4	4	057	016	4
4	4	059	018	5
4	4	051	017	1

Year	Event year	Code1	Code2	Cig
1	1	051	001	1
1	1	052	002	2
1	1	006	003	3
2	2	057	006	4
2	2	059	008	5
3	3	002	010	6
3	3	062	001	7
3	3	051	012	1
4	4	001	015	8
4	4	057	016	4
4	4	059	018	5
4	4	051	017	1

The creation of an experimental longitudinal database of business micro-data

Abstract

Keywords

1. Introduction

2. Experimental Panel database main facts and building blocks

2.1 Main facts

4.1 Introduction

Table 1 Change of legal form nomenclature

Table 5 Alpha case (codes from Table 4)

Table 7 Cases Beta 1 and Beta 2

Table 8 Linkage with ASIA

Table 11 Scheme of an elaboration for firms which completely slit-off or merge

Table 12 Distribution of firms involved in transformation events by year (All firms)

Footnotes

Acknowledgments

Appendix

Glossary and references

References

Table 1
Change of legal form nomenclature

Table 5
Alpha case (codes from Table 4)

Table 7
Cases Beta 1 and Beta 2

Table 8
Linkage with ASIA

Table 11
Scheme of an elaboration for firms which completely slit-off or merge

Table 12
Distribution of firms involved in transformation events by year (All firms)

Year	Event year	Code1	Code2	Cig
1	1	051	001	1
1	1	052	002	2
1	1	006	003	3
2	2	057	006	4
2	2	059	008	5
3	3	002	010	6
3	3	062	001	7
3	3	051	012	1
4	4	001	015	8
4	4	057	016	4
4	4	059	018	5
4	4	051	017	1