Labour Market Areas for territorial policies: Tools for a European approach

3.1The implemented algorithm

In this section the rule based iterative agglomerative algorithm implemented in the package LabourMarketAreas is briefly presented; see [18] for its full description, it is based on the experience outlined in [17]. The algorithm in the package LabourMarketAreas belongs to the set of methodologies identified by [13] as the most suitable ones at European level. The input of the algorithm is the census/register based commuting matrix of flows surveyed at basic geographical units, called communities. Given such Origin-Destination trips to work, the algorithm creates a partition of the territory such that all areas satisfy the so called validity rule. Such rule depends on area size, the number of commuters living in the area, and on self-containment, the proportion of the commuters not crossing area borders. In particular, the validity rule sets a trade-off between area size and its level of self-containment. The rationale is the smaller the area the higher the self-containment to be considered adequate. Lower self-containment values are instead acceptable for larger areas. These cut-off levels are defined by users through four parameters: minSZ, tarSZ, minSC and tarSC. minSZ is the area acceptable minimum size whereas tarSZ is the size for an area to be considered large. minSC is the acceptable minimum level of self-containment for large areas whereas tarSC is the acceptable minimum level of self-containment for small areas. Values of self-containment range by definition from 0 to 1, however possible values for the minimum minSC lie between 0.6 and 0.7 and for the target, tarSC, between 0.7 and 0.85.

The algorithm starts by considering each community as an area that is checked against the validity rule. At each iteration, the furthest from validity rule area is dissolved into its constituent communities. Based on the strength of the link to other areas, as defined in [4], one selected community is attached to a new area. When this new agglomeration increases the validity value w.r.t. its previous level, the assignment is performed. Otherwise, the community is put into a reserve list to be assigned after convergence, see [18]. Convergence occurs when all areas satisfy the validity rule.

3.2LMA delineation process

The implemented process consists of four subsequent stages. Table 1 lists them and the corresponding tools in the package; the full documentation is present in the package description [15]. In the first stage, the user selects the parameters and applies the rule-based algorithm (Section 3.1). The main output consists of a territorial partition characterised by some core figures such as size of initial communities in terms of commuters living in; size of LMAs and number of commuters working there; flows between LMAs. Moreover, information regarding the path to the algorithm convergence is provided. Then, LMAs are named after the community having the maximum incoming flows. When community geographical coordinates are available in geospatial vector format, tools to deploy this information at LMA level are included in the package in order to visualise them.

Quality assessment (second stage) is performed by analysing statistical indicators provided by the package. Such statistics describe areas, flows and the entire partition; they allow testing the compliance of the solution to LMA principles. Examples of area level indicators include the level of self-containment, the number of attracting communities and commuters living in the area measuring, respectively, autonomy, relevance and homogeneity characteristics.

As the algorithm is based solely on commuting flows, some areas may include communities which are not spatially contiguous. Based on geospatial information the tools implemented in the R package allow complying with the contiguity principle (third stage). Table 1 lists the tools which identify non-contiguous areas and re-assign communities accordingly. This fine tuning procedure is based on both visualisation and strength of links between geographical entities.

Sensitivity analysis (fourth stage) can be performed by setting different groups of parameters in the algorithm. In this way the multivariate optimisation problem of finding the most appropriate partition can be addressed.

This package has been used by Istat to produce the Italian 2011 LMAs, see [19]. The commuting to work data stem from the 2011 Italian Population Census. Roughly 19 millions commuters (as defined by the census) create nearly 600 thousands commuting flows between the 8092 Italian municipalities. In terms of commuters living in such municipalities the smallest community has 4, the largest (Roma) has nearly 900 thousands. The self-containment was set equal to 0.75 (tarSC) and 0.6 (minSC) for size area of 1000 (minSZ) and 10000 (tarSZ). After fine tuning, the resulting partition comprises 611 LMAs; the package takes about two hours to reach convergence. These LMAs range from small size areas in mountainous regions to large ones in developed areas (e.g. the LMAs of Roma and Milano have well over one million commuters living in).

Other National Statistical Institutes (NSIs) have used the package to produce LMAs (e.g. Denmark) or are investigating its potentials, see [20] for example. In the ESS, each country determines, after sensitivity analysis, the combination of parameters better representing both the geography and the labour market.

4.1Register-based indicators

The statistical registers commonly used within ESS are exhaustive data sources. Frame-SBS, Istat statistical data warehouse at firm-level, is a business register yearly produced since 2011. Such register allows for computing key economic account variables, i.e. structural business statistics, at a very refined level of detail, thus facilitating the dissemination of better focused data. A recent Istat project further extended Frame-SBS register by developing additional indicators. The new indicators aim at measuring different socio-economic dimensions like labour input, performance, competitiveness, etc. Territorial characteristics have also received a special attention during the Frame-SBS extension project. The interaction between enterprises and the territory is measured mainly through their establishments. Both FRAME-SBS and establishment registers have the end of the year as reference date. They are derived using a mixture of administrative and survey data. By considering the local unit information which is registered at municipality (i.e. community) level, many indicators reflecting the delocalisation strategies of the enterprises may be derived using a bottom-up approach. Basic examples involve number of employees and/or local units and/or headquarters of enterprises by LMA.

Further indicators may be derived by mixing different economic or territorial dimensions with the number of employees. Entrepreneurship density, E𝑑𝑒𝑛𝑠, and attractiveness rate, A𝑟𝑎𝑡𝑒, are examples of this latter approach. E𝑑𝑒𝑛𝑠 is given by the ratio of employees working in private businesses and the total number of LMA residents. It represents an entrepreneurship indicator, linking characteristics of the workplace and the workforce. A𝑟𝑎𝑡𝑒 is an indicator reflecting how much a LMA (territory) is involved in the delocalisation strategies of enterprises based in other LMAs. Its similarity to the commuting flows concept is obvious. Indeed, instead of involving commuters, A𝑟𝑎𝑡𝑒 might be sought as a supply side synthetic indicator of economic factors like workforce, capacity, raw material, policies, etc. In Fig. 1, a comparison of the E𝑑𝑒𝑛𝑠 and A𝑟𝑎𝑡𝑒 with their median values is shown.

Figure 1.

Combined index of E𝑑𝑒𝑛𝑠 and A𝑟𝑎𝑡𝑒. Frame-SBS, reference year 2012. HH (LL) = both E𝑑𝑒𝑛𝑠 and A𝑟𝑎𝑡𝑒 higher (lower) than their median values; HL (LH) =E𝑑𝑒𝑛𝑠 greater (lower) than its median value and A𝑟𝑎𝑡𝑒 lower (higher) than its median value

Figure 2.

ST-EBLUP estimates of unemployment and employment rate at LMA level in Italy and efficiency gain of the estimator over the direct one (year 2014).

More complex approaches may be envisaged when combining the LMAs and the principal economic activity information. For example, an analysis of homogeneous groups of LMAs was performed in [19]. The homogeneity was achieved by iteratively applying statistical techniques like correspondence and cluster analysis on number of employees by NACE Rev.2 group.

Finally, care needs to be taken for economic performance indicators at LMA level. In fact, key economic variables (as value-added at factor costs, turnover, etc.) are not registered at local unit level in Frame-SBS. Consequently, before applying any bottom-up procedure, a preliminary data treatment is necessary. Indeed, when using different territorial partitions (regions/LMAs) and different analysis units (firms/ local units) a critical issue to be faced by official statistics is consistency of disseminated indicators. Different economic theories, based on well-known working assumptions, may be applied to estimate economic indicators at establishment level in order to compute the LMA one. Istat is currently testing such hypotheses (e.g. based on weighting schemes, see [19]) to release these indicators yearly.

4.2Small-area estimation of Labour Force Survey figures

The most important large scale surveys carried out by NSIs are repeated surveys. They are intended to produce the estimates for several parameters, such as overall counts, means, ratios of the whole population, as well as for population sub-sets. Generally, sample sizes are not large enough to guarantee, for all the sub-populations/domains of interest, reliable direct estimates (i.e. based solely on sample information from the sub-population itself). In such cases small area estimation (SAE) methods can be applied to overcome the problem. These methods borrow strength from the information collected in other domains, resulting in increasing the effective sample size for each small area. In this section Istat method to compute small area annual estimates of labour status for all Italian LMAs based on LFS is briefly described.

The LFS is the main source of information on the Italian labour market and is aimed to produce official estimates of employment, unemployment and inactivity rates over different territorial domains. It is a quarterly two stage municipality/household survey with partial overlap of sampling units according to a precise rotation scheme. For each quarterly sample about 1350 municipalities and 200,000 individuals are involved. Direct estimates are disseminated monthly, quarterly and annually. Only annually reliable estimates can be produced for the Italian Provinces (NUTS 3). Regional (NUTS 2), macro-regional (NUTS 1) and national estimates are, instead, computed and disseminated quarterly. LMAs are unplanned domains cutting across provinces (NUTS 3) which are the LFS finest planned domains whose direct estimates are released annually. Only about 450 LMAs out of 611 are covered by the LFS.

Given the repeated nature of the LFS, an EBLUP estimator based on unit level mixed model borrowing strength both from other areas and from other survey cycles has been considered to compute the estimates at LMA level [21]. The resulting estimator allows to increase the efficiency of the results by exploiting the previous survey occasions data and by enabling for a within area variation. The random effects of the model are introduced to take into account the spatial autocorrelation among LMAs as well as the temporal trend of the target variables. The spatial trend is introduced assuming that between area correlation matrix depends on the distances among the areas [22]. In particular, the distance among the areas based on the traveling time between the centroids [23] is used to specify the correlation among LMAs. For the temporal random effects, an auto-regressive AR(1) process is considered in order to take into account the high value of time correlation coefficient, whose estimate is around 0.75. The use of the time information allows for relevant efficiency gains of the estimates. The SAE estimates of unemployment and employment rate for each LMA coupled with their efficiency gains w.r.t. direct estimates are presented in Fig. 2.

LMA yearly SAE estimates are obtained by averaging the quarter’s count estimates. However, the aggregation over a planned larger domain (e.g. NUTS 2) may differ from the regional direct estimates. In order to make them consistent a benchmarking procedure is applied to SAE estimates yearly released by Istat.