Análisis "Cluster" de los flujos laborales andaluces

Resumen: Este trabajo consiste en la aplicación de una metodología de formación de grupos o clusters a la información disponible sobre las colocaciones registradas en las oficinas públicas de empleo andaluzas entre enero de 2007 y diciembre de 2010. Dicha metodología permite dividir al mercado de trabajo en grupos o mercados locales mediante el agrupamiento de segmentos laborales definidos según municipio, ocupación y sector de actividad. Los desempleados tienen mayores opciones de encontrar empleo si buscan dentro de su grupo, ya que dicho grupo se define sobre las experiencias exitosas de colocación de otros buscadores con similares características. Nuestra herramienta informativa pretende mejorar el emparejamiento y la intermediación laboral. Abstract: In this paper we propose an empirical tool in order to account for the role of heterogeneities in the labour matching process, and we then make use of it in an application to the Andalusian labour market, which relies on a database of individual microdata of considerable size. The nature of our data, with information on vacancies, unemployed workers and job placements, links up our work directly with the search and matching theoretical models. In those models an improvement in the labour information allows to increase the matches per period given a certain number of unemployed workers and job vacancies; if this occurs, the Beveridge curve shifts towards the origin with the consequent reduction of unemployed workers and job vacancies. Our work is not meant to extend or evaluate matching models, but instead it tries to handle empirically important elements involved in these models, as heterogeneities and segmentation. Our data for the empirical application refer to the matches registered in the Andalusian Public Employment Agency (ServicioAndaluz de Empleo, SAE) in the four years 2007 to 2010. The available information allows us to make a detailed division of workers and jobs into groups, with the combination of various characteristics, and yet have enough data in each group to be statistically representative. Three characteristics have been considered for both (workers and jobs): location, defined by municipality (770 different municipalities in our data); skills, defined by group of occupation (482 different occupations); plus sector of economic activity (56 different sectors). In the case of workers we generate worker groups, and for the jobs we have "job groups". When there is a job placement, the union of both groups, which do not necessarily have to be of the same type, forms a joint group. During the four years analysed, there was a flow of more than 16 million registered matches, but all the values of the full set of characteristics (the three of the worker and the three of the job) are known in only just over 9 million matches, distributed between 2,848,977 different joint groups, 456,109 different worker groups and 261,167 different job groups. To ensure the representativeness of each group analysed we have decided to consider the entire period as a single time interval. Moreover, in order to manage some of our subsequent calculations in the cluster analysis, we have been forced to reduce the large amount of information available by selecting a sample of 1,542 common groups that appear in the 10,000 joint groups with the most matches; they are common in the sense that each one appears as a worker group and as a job group in our selected sample. For these 1,542 groups, there are 1,837,123 matches distributed between 69,653 different joint groups. These common groups can be successively grouped into clusters starting from a similarity (or lower distance) measure based on the matches that occur into the period. Thus, we consider that two worker groups are more similar the more they resemble in the way they match with job groups, and that two job groups are more similar the more they resemble in the way they match with worker groups. The clustering process is developed as follows: we define "similarity" (lower "distance") between two worker groups as the superposition of their matching distributions respect to the job groups, and "similarity" between two job groups as the superposition of their matching distributions respect to the worker groups. We have found that the degree of similarity between two worker groups corresponds very significantly with the degree of similarity between those two groups when they are acting as job groups, so we can define the "proximity" between any two groups as the average of the two similarities previously defined. We use a hierarchical method of clustering, with groups gradually fusing to form increasingly larger groups or clusters. This method starts by merging the two groups with the highest similarity into a new group or cluster; the similarity of this new group with the rest of the groups is then recalculated, and the next two groups with the highest similarity are merged together. This process continues until we obtain a single cluster for the entire labour market. The grouping process can be visualised with a graphical display called dendrogram or tree diagram. The process can be stopped when a specified number of clusters is reached or when the last similarity calculated falls below a specified level. This clustering process lets us obtain a better overview of the structure of the labour market. To display the results that may be obtained by applying our methodology, we have chosen to analyse two scenarios of grouping. The first scenario is the result of stopping the clustering process when we have only 16 clusters two clusters per province on average; in this case, we try to show a better overview of the Andalusianlabour market. In a second scenario, the clustering algorithm is stopped when we have a total of 128 clusters 16 clusters per province on average, so that we can analyse the Andalusianlabour market structure in greater detail. For this second scenario, as an illustrative example, the clusters where the provinces of Malaga or Seville dominate are described. The main results obtained are the following: In general, the clustering process generates clusters of large, medium and small size. Small clusters merge with medium and large ones insofar we continue the grouping process. Most clusters are formed connecting neighbouring municipalities; this suggests that workers prefer in general to change their group of occupation or their sector of activity instead of their original location. If we consider the scenario of 128 clusters, we observe that in 87.2% of the matches the worker has been employed within her own cluster. In these "intra-cluster" matches, 61.5% of the workers have found employment outside their initial group, so in general Andalusian workers (those who get a job) are seeking for a job in their own group but also in their own cluster (or local labour market captured by our methodology). On one hand, the most frequent occupations observed in the clusters are: other qualified workers in agriculture, agricultural labourers, bricklayers, construction labourers, cleaning staff, shop assistants and waiting staff. On the other hand, the most observed sectors of activity are: agriculture, construction, catering, other entrepreneurial activities, retail sales and Public Administration. In the big clusters (those with more matches) workers aged between 30 and 44 years old stand out, and also those with secondary (general education) or primary education. Additionally, in all clusters there is a clear predominance of fixed-term contracts. The main conclusion of this paper is that worker mobility, geographical or occupational, and the availability of relevant information are important requirements for effective labour matching, and constitute a prominent element that should be taken into account to guide the design of active labour market policies. The empirical tool proposed in this paper may be useful in this regard, by helping jobseekers and firms looking for workers to follow successful paths of matching previously used by others agents with similar characteristics. The clustering methodology allows past information on matches to be processed in order to generate a \'roadmap\' of possible routes to different labour market clusters, which can also include the probability of success in each route. The versatility of the methodology proposed makes it possible to enrich the information provided from this perspective and to take into consideration other variables of interest, such as the best search channels for each cluster. Another extension may consist in the comparison among different time periods in order to detect emergent employment sectors. Finally, further research is required to test the practical usefulness of this methodology in labour intermediation.