A Measure of Trust: The Italian Regional Divide in a Latent Class Approach

The NUTS classification (nomenclature of territorial units for statistics) is a hierarchical system of Eurostat to subdivide a country in territorial areas spatially homogeneous and comparable among different countries (EUROSTAT 2013). In Europe, NUTS lists 98 regions at level 1, 276 regions at level 2 and 1342 regions at level 3; in particular, NUTS2 corresponds to the classical Italian regional partitions in 21 areas. EVS is a large-scale, cross-national, and longitudinal survey research program on basic human values that provides insights into the ideas, beliefs, preferences, attitudes, values and opinions of the European citizens.

The use of administrative boundaries is open to criticism because social phenomena are not confined by political boundaries. However, in this paper, as will be clear later, we compute a regional measure of trust through a proper statistical aggregation of the trust measures obtained at micro level for each household. This approach, that natively takes into account the territorial dispersion and the specific characteristics of the sampled households, generates less distortion than the macro measures usually proposed in literature. In fact, the regional classification (NUTS2) is a very wide aggregation and it would be particularly useful to have information with greater spatial details. Unfortunately, the ISTAT (Italian National Institute for Statistics) survey (so other surveys) does not allow a proper statistical representation of trust for smaller territories as provinces, local labor systems or municipalities. It should be added that ISTAT knows province and municipality of residence of the sampled households; however, the Institute's confidentiality policies require the deletion of this information to protect the privacy of the respondents. In any case, for the ADL survey, provinces and municipalities of residence do not represent domains of analysis from which to make proper inference on the reference population. Furthermore, it should be specified that the ISTAT database is not publicly accessible but made available as result of positive evaluation of a project, and this could increase the benefits of its use in this paper, because this peculiarity limits the use of the ADL survey in research.

The ADL survey is the only official source that contains information on trust of the Italian households. At regional level, the only alternative to ISTAT data on trust is the wave 2010 of Survey of Household Income and Wealth (SHIW) by Bank of Italy, but we note that questions on trust have been erased in the subsequent surveys 2012 and 2014 (Banca d’Italia 2012, 2015; Albanese et al. 2013).

Italy is divided into 21 NUTS2 areas: that is 20 regions with Trentino Alto Adige usually broken down into two “province autonome” Trento and Bolzano. The northern regions are 8: Trentino Alto Adige (that is, Bolzano and Trento), Veneto, Friuli Venezia Giulia, Emilia Romagna (North East); Piemonte, Valle d’Aosta, Lombardia, Liguria (North West). The central regions are 4: Toscana, Umbria, Marche, Lazio (Centre). The southern regions are 6: Abruzzo, Molise, Campania, Puglia, Basilicata, Calabria; finally, the 2 major islands: Sicilia and Sardegna. The 12 central and northern regions are named “Centro-Nord” and the remaining 8 regions denominated “Mezzogiorno”.

Also, we have tried to eliminate, in turn, some items and to repeat the analysis but the slight improvement in statistical estimates did not justify the elimination of these items; also the further increase in the number of classes does not lead to improvements. These difficulties are not surprising; the cases (households) are numerous, the items are numerous and the responses distributed over a very large scale with values between 0 and 10. In this regard, it is worth mentioning an our attempt to reduce the complexity of the estimation problem by reducing the response scale of the six items of TI in a new 1-4 scale similar to the TO scale; but, once again, the result was not better and the clustering has not been improved. In short, the 8-cluster model for the six TI items on the 0-10 scale is perhaps not the ideal model, but certainly the best model possible for the micro data used in this paper, whereas the model for TO is much more satisfactory from a statistical point of view.

There are several techniques for treating sampling weights in models of latent classes. The prevailing method is to construct a weighted observed frequency table and then to analyze it as if it were an unweighted table (pseudo-maximum-likelihood estimation method). An alternative method is to ignore temporarily sampling weights in estimations, but to correct in a second step the model parameters using the weights (two-step approach). Here, taking into account that the weights used in ADL are case weights (used to adjust for cases that are under-represented or over-represented in the sample compared to the reference population), we follow a third approach of Vermunt and Magidson (2007) considered statistically better in literature for latent class models such as those used in this paper. It extends the procedure of ML estimation with sampling weights in log-linear analysis, where the unweighted observed frequencies are used as data and the inverse of the cell-specific weights are included as a term in the log-linear model (Agresti 2013). It is, therefore, an ML approach whose application to LC models leads to particular cases of the Haberman log-linear model with cell weights for frequency tables derived by indirect observations (Vermunt and Magidson 2007).

It is known that Sardegna is an atypical region (from the point of view of the economic and social characteristics) among those in Mezzogiorno, to which it belongs only by statistical convention albeit geographically located around the center of Italy (ISTAT 2017).

The time interval 1998-2013 is determined by the availability of official statistical data with the intent to give a general representation of the medium-long term growth of the Italian regions. Obviously, the use of a narrower range has some repercussions on the position of the regions in relation to different trends of the regional economies, but the changes are substantially limited and do not affect the positive relationship that exists between trust and economic growth.

The map of the territorial gaps is very similar among the indicators commonly associated with trust measures (ISTAT 2016, 2017). For example, the voter turnout (that has important connections with political trust) is much lower in the southern regions than in the northern ones, as well as judgment towards politicians and political institutions (BANCA D’ITALIA 2014; Zmerli and Van der Meer 2016).