This doctoral thesis investigates territorial disparities in Italy by means of Bayesian statistical inference, with a focus on municipality-level vulnerability in two applied domains, namely education and assistance to violence against women. The empirical work consists of a total of three case studies, motivated by the need for granular and uncertainty-aware evidence to support territorial planning and policy-making. On the methodology side, the thesis adopts hierarchical areal models including conditional autoregressive (CAR) priors, considering the municipal partitioning of Italy as an irregular discrete spatial domain. Inference is carried out primarily through the Integrated Nested Laplace Approximation (INLA), which leverages on the sparse precision structure of conditional autoregressive models and enables efficiency model estimation. Particular attention is devoted to the issue of spatial confounding in the interpretation of regression outcomes. On the application side, the first original research work covered herein consists in an exploratory analysis of the Italian school system data, facilitated by an open and reproducible data infrastructure integrated in the SchoolDataIT R package, which downloads and harmonises information from multiple data sources and is accessible on CRAN. The second research work studies the relationship between the student's abilities in the second year of high school and the infrastructural endowment in all Italian municipalities. Municipal student scores are obtained by averaging standardised and spatially homogeneous indicators of student outcomes provided by the Invalsi Institute for two subjects, Italian and Mathematics. To explain spatial variability in Invalsi scores, we employ a bivariate intrinsic conditional autoregressive latent model, defined alternatively at the higher aggregation levels of provinces or catchment areas of infrastructural poles. We find that alongside a significant association with infrastructural indicators chosen as explanatory variables, spatially structured latent effects are still necessary to explain the different student outcomes across municipalities. The third research work analyses the access to local anti-violence centers in Apulia in 2021--2024 by spatio-temporal Poisson regression, comparing standard conditional autoregressive models, namely intrinsic CAR, proper CAR, Leroux proper CAR and scaled Besag-York-Mollié model, and proposing a joint penalised complexity prior construction for all the parameters of models employed, extending the application field of penalised complexity priors in areal modellling. Results suggest the presence of spatial constraints in seeking help by anti-violence centers, most a decline of accesses as distance from anti-violence centers increase; in addition, lower education levels may also contribute to under-reporting in disadvantage areas, while higher economic development may be associated with lower incidence of reported violence. Overall, this thesis shows how Bayesian spatial modelling can be useful to quantify uncertainty, stabilise inference in sparse territorial setting, and suggest policy-relevant insights on territorial inequalities in Italy.
Questa tesi di dottorato indaga le disparità territoriali in Italia mediante inferenza statistica bayesiana, con particolare attenzione alla vulnerabilità a livello comunale in due ambiti applicativi: l’istruzione e il supporto alle donne vittime di violenza. Il lavoro empirico si articola in tre casi di studio, motivati dalla necessità di disporre di evidenze granulari e consapevoli dell’incertezza a supporto della pianificazione territoriale e delle politiche pubbliche. Sul piano metodologico, la tesi adotta modelli areali gerarchici che includono prior autoregressive condizionali (CAR), considerando la partizione comunale italiana come dominio spaziale discreto irregolare. L’inferenza è condotta principalmente tramite l’Integrated Nested Laplace Approximation (INLA), che sfrutta la struttura sparsa della matrice di precisione dei modelli autoregressivi condizionali e consente una stima efficiente dei modelli. Particolare attenzione è dedicata al problema del confondimento spaziale nell’interpretazione dei risultati di regressione. Sul piano applicativo, il primo lavoro di ricerca originale incluso nella tesi consiste in un’analisi esplorativa dei dati del sistema scolastico italiano, resa possibile da un’infrastruttura dati aperta e riproducibile integrata nel pacchetto R SchoolDataIT, che scarica e armonizza informazioni provenienti da fonti multiple ed è accessibile su CRAN. Il secondo lavoro di ricerca studia la relazione tra le competenze degli studenti del secondo anno della scuola secondaria di secondo grado e la dotazione infrastrutturale in tutti i comuni italiani. I punteggi comunali degli studenti sono ottenuti mediando indicatori standardizzati e spazialmente omogenei degli esiti scolastici forniti dall’Istituto Invalsi per due materie, Italiano e Matematica. Per spiegare la variabilità spaziale dei punteggi Invalsi, impieghiamo un modello latente bivariato intrinseco autoregressivo condizionale, definito alternativamente ai livelli di aggregazione superiori delle province o dei bacini di utenza dei poli infrastrutturali. I risultati mostrano che, accanto a un’associazione significativa con gli indicatori infrastrutturali selezionati come variabili esplicative, effetti latenti spazialmente strutturati restano necessari per spiegare le differenze negli esiti scolastici tra comuni. Il terzo lavoro di ricerca analizza gli accessi ai centri antiviolenza locali in Puglia nel periodo 2021--2024 mediante regressione di Poisson spazio-temporale, confrontando modelli autoregressivi condizionali standard — intrinsic CAR, proper CAR, Leroux proper CAR e modello Besag-York-Mollié scalato — e proponendo una costruzione congiunta di prior di complessità penalizzata per tutti i parametri dei modelli impiegati, estendendo il campo applicativo delle prior di complessità penalizzata nella modellazione areale. I risultati suggeriscono la presenza di vincoli spaziali nella ricerca di aiuto presso i centri antiviolenza, in particolare una diminuzione degli accessi all’aumentare della distanza dai centri; inoltre, livelli di istruzione più bassi possono contribuire a fenomeni di sotto-segnalazione nelle aree svantaggiate, mentre un maggiore sviluppo economico può essere associato a una minore incidenza di violenza denunciata. Nel complesso, questa tesi mostra come la modellazione spaziale bayesiana possa essere utile per quantificare l’incertezza, stabilizzare l’inferenza in contesti territoriali con dati sparsi e suggerire evidenze rilevanti per le politiche sulle disuguaglianze territoriali in Italia.
Stochastic Spatial Areal Models for Mapping Territorial Vulnerability / Modelli Stocastici Areali per la Mappatura della Vulnerabilità Territoriale
CEFALO, LEONARDO
2026
Abstract
This doctoral thesis investigates territorial disparities in Italy by means of Bayesian statistical inference, with a focus on municipality-level vulnerability in two applied domains, namely education and assistance to violence against women. The empirical work consists of a total of three case studies, motivated by the need for granular and uncertainty-aware evidence to support territorial planning and policy-making. On the methodology side, the thesis adopts hierarchical areal models including conditional autoregressive (CAR) priors, considering the municipal partitioning of Italy as an irregular discrete spatial domain. Inference is carried out primarily through the Integrated Nested Laplace Approximation (INLA), which leverages on the sparse precision structure of conditional autoregressive models and enables efficiency model estimation. Particular attention is devoted to the issue of spatial confounding in the interpretation of regression outcomes. On the application side, the first original research work covered herein consists in an exploratory analysis of the Italian school system data, facilitated by an open and reproducible data infrastructure integrated in the SchoolDataIT R package, which downloads and harmonises information from multiple data sources and is accessible on CRAN. The second research work studies the relationship between the student's abilities in the second year of high school and the infrastructural endowment in all Italian municipalities. Municipal student scores are obtained by averaging standardised and spatially homogeneous indicators of student outcomes provided by the Invalsi Institute for two subjects, Italian and Mathematics. To explain spatial variability in Invalsi scores, we employ a bivariate intrinsic conditional autoregressive latent model, defined alternatively at the higher aggregation levels of provinces or catchment areas of infrastructural poles. We find that alongside a significant association with infrastructural indicators chosen as explanatory variables, spatially structured latent effects are still necessary to explain the different student outcomes across municipalities. The third research work analyses the access to local anti-violence centers in Apulia in 2021--2024 by spatio-temporal Poisson regression, comparing standard conditional autoregressive models, namely intrinsic CAR, proper CAR, Leroux proper CAR and scaled Besag-York-Mollié model, and proposing a joint penalised complexity prior construction for all the parameters of models employed, extending the application field of penalised complexity priors in areal modellling. Results suggest the presence of spatial constraints in seeking help by anti-violence centers, most a decline of accesses as distance from anti-violence centers increase; in addition, lower education levels may also contribute to under-reporting in disadvantage areas, while higher economic development may be associated with lower incidence of reported violence. Overall, this thesis shows how Bayesian spatial modelling can be useful to quantify uncertainty, stabilise inference in sparse territorial setting, and suggest policy-relevant insights on territorial inequalities in Italy.| File | Dimensione | Formato | |
|---|---|---|---|
|
Doctoral_Thesis_LC_A_signed_signed.pdf
accesso aperto
Licenza:
Tutti i diritti riservati
Dimensione
26.55 MB
Formato
Adobe PDF
|
26.55 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/378561
URN:NBN:IT:UNIBA-378561