La représentation relationnelle des données statistiques : application au traitement des données de panel

The statistical languages and softwares used in économies represent numerical data in multidimensional tables by means ofthe usual conventions of classical algebra. This logical model, when associated with numerical computations, is not very effective in reorganizing major sets of heterogeneous and complex data, e.g. in harmonizing surveys, monitoring panels, managing multiple nomenclatures, constructing time séries, recognizing hierarchical structures, etc. The theory of relational algebra makes it possible to define an alternative logical model of représentation of multidimensional data which relies on the index notation ofthe statistician. The data are represented in theform of sets ofn-tuples, each élément being identified by a single identification "key" which corresponds to the index values for an observation. This model makes it possible, through the use of simple mathematical rules, to formalize cumbersome and apparently complex opérations, such as the handling of surveys over a long period or the use of a single nomenclature to deal with data obtained from differing institutional contexts. The reorganisation ofthe data is carried out by means of the so-called "relational algebra" opérations derived from set theory. A "vectorisation" fonction makes it possible to generate classical data structures that are recognized by statistical languages and softwares. The application of this theoretical approach raises some practical difficultés with some relational data base management softwares, and this entails analysing the minimum functionalities needed to manage effectively the major statistical data bases. Market softwares appear to hâve been developed in a high-flow transactional framework, in which a substantial number of users make regular day to day changes from their Personal computers. However, the statistician needs a System that is able to extract and reorganize a large flow of factual informations, and to manage the hierarchical relations in thèse major sets. This paper attempts to define a "benchmark" whose fonction is to give an account ofthe behaviour of a System concerning limiting relational calculation opérations. This theoretical analy sis is completed by a description of the findings obtained when modelling feed consumption. By means of this example, we are able to make more explicit the methods used in order to carry out matricial calculation with relational algebra opera- tors. To conclude, the simple relational model, as it was defined and used in the seventies is suf fuient at the présent time to move beyond proprietary hardware and software, and to define solutions to deal with présent and future Computing configurations in research laboratory environments.Les logiciels statistiques représentent les données numériques dans des tables avec les conventions usuelles de l'algèbre classique. Ce modèle logique est peu efficace pour restructurer de grands ensembles de données hétérogènes et complexes : harmonisation d'enquêtes, suivi de panels, gestion de nomenclatures, construction de séries chronologiques, reconnaissance de structures hiérarchisées, etc. La théorie de l'algèbre relationnelle permet de définir un modèle alternatif de représentation de ces données, qui s'appuie sur la notation indiciaire usuelle. La réorganisation des données s'exprime au moyen d'opérations ensemblistes dites de "calcul relationnel". Une fonction de "vectorisation" permet de générer les tables reconnues par les logiciels statistiques. L'application de ce modèle théorique pose des problèmes avec les logiciels du commerce, ce qui conduit à analyser les fonctionnalités minimales requises pour gérer efficacement les grandes bases statistiques, et à proposer un test d'évaluation des performances.Poupa Jean-Claude. La représentation relationnelle des données statistiques : application au traitement des données de panel. In: Cahiers d'Economie et sociologie rurales, N°26, 1er trimestre 1993. pp. 63-96