STATISTICA & APPLICAZIONI - 2019 - 2

The Big Data expands the range of sources that have the potential to be used for Official Statistics and represents an effective reply to the declining response rates and the rising costs of conducting surveys, offering, in the meanwhile, potentially more timeliness and granular statistics. The use of these non-survey data sources generates a paradigm shift: from designed data to data-oriented or data-driven statistics. Therefore, it is necessary to determine under which conditions these sources make valid inference on the finite target population. Several statistical and quality frameworks on Big Data have this objective. Nevertheless, they are defined according to a general perspective. The paper aims to concretize these frameworks going into detail about the statistical tools to apply in each phase of the data generating process. Our proposed approach relies on combining information from multiple data sources with standard or innovative procedures and makes an integrated and coordinated use of the methods. A real example of the use of Big Data in Official Statistics shows how to create the conditions to define a process for obtaining accurate and consistent estimates.

The main objective of this article is to show that, based on empirical experience, the introduction of a Centralized Data Collection approach (CDC), such as the one introduced in the Italian Statistical Institute from the year 2016, has a positive effect on the Total Survey Error (TSE) of current surveys. This will present the features of the new CDC approach and the innovations introduced in terms of generalized tools, services and procedural solutions applied. In order to enhance the above effects, the attention will be focused on three case studies that represent as many examples of product and process innovations introduced by CDC in different survey domains. The introduction of a CDC setup, on the base of empirical experience, involved positive effects on TSE. Moreover, it provides a pecialized approach to the management of cross-cutting services, produced significant increasing of response rates, improving processes efficiency and some dimensions of data quality (e.g. timeliness).

The mixed-mode design are adopted by NSIs both to contrast declining response and coverage rates and to reduce the cost of the surveys. However, mixed-mode design introduces several issues, such as a possible increase in the total survey error due to extra measurement error introduced by additional data collection modes and a selection effect due to differences in the population coverage and in nonresponse processes. Selection and measurement effects, components of the mode effect, are generally confounded, this involving a complication of the inferential process that can be facilitated with experimental survey designs. The work shows the main results of the analyses carried out to assess the impact on final estimates of the switching from single to mixed-mode of the Aspect of Daily Life household survey. The analyses are developed in an experimental setting in which two parallel survey designs with two samples independently selected were carried out and are mainly focused on the evaluation of the overall quality of the two realized samples of respondents in terms of bias, of the total mode effect and the impact on the multivariate distributions of the survey variables.

Hybrid calibration refers to an approach where techniques of classical calibration and more recent model-assisted calibration are combined for a joint calibration methodology. The classical calibration does not assume a model but uses the original auxiliary data as aggregates, whereas in model calibration, unit-level predictions from a model are used as pseudo auxiliary information. By combining these approaches we introduce hybrid methods, where aggregate data from different levels of the population are supplied to the model-free component and unit-level data are incorporated into the model-assisted component. The choice of the model depends on the type of the target variable. We use here linear and logistic mixed models. In the estimation for population subgroups or domains, the classical calibration fails when domain sample sizes become small. Our hybrid calibration methods were more accurate in small domains. In our studies, the basic model-assisted calibration was usually the best in accuracy, but the method requires population-level information on auxiliary variables in the model. The basic hybrid calibration method overcomes this restriction by including a model-free calibration component in the model-assisted calibration procedure. A new two-level hybrid calibration technique provides a further extension applicable for hierarchically structured populations. In this method, calibration in the model-free part is performed at a higher regional level, instead of the domain level. In our simulation experiments, the two-level hybrid calibration performed well: its accuracy and design bias were comparable to model calibration. The most stable weight distributions were obtained by the two-level method and Ha´jek type estimators developed in the paper.

Nowadays all universities carry out surveys to measure student satisfaction of teaching quality. In this work we study which are the main determinants of student satisfaction of courses, analyzing teaching evaluations filled in by students. Survey questions are designed to assess general satisfaction of students based on the following constructs: teaching, course organization and infrastructure. Structural Equation Modeling, and in particular Partial Least Squares - Path Modeling, is used to examine the relationships between the latent constructs, with the aim of evaluating student satisfaction. Data consists of teaching evaluations collected in a large public university in the academic year 2017/2018. The main result is that teacher characteristics matter and they are the main drivers of student overall satisfactions. The ability of teachers to stimulate student interest, together with the clarity of their explanations, are key drivers to increase student satisfaction. Moreover, organization of teaching and infrastructures are positively correlated with expected satisfaction, but their effect is (much) smaller.