Tesi etd-06302025-133844 |
Link copiato negli appunti
Tipo di tesi
Tesi di laurea magistrale
Autore
SALMASO, FILIPPO
URN
etd-06302025-133844
Titolo
On the Estimation of Heterogeneous Functional Treatment Effects
Dipartimento
INFORMATICA
Corso di studi
DATA SCIENCE AND BUSINESS INFORMATICS
Relatori
relatore Prof. Guidotti, Riccardo
relatore Prof.ssa Chiaromonte, Francesca
relatore Dott. Testa, Lorenzo
relatore Prof.ssa Chiaromonte, Francesca
relatore Dott. Testa, Lorenzo
Parole chiave
- casual inference
- conditional average treatment effect
- doubly-robust
- functional data analysis
- meta-learners
Data inizio appello
18/07/2025
Consultabilità
Non consultabile
Data di rilascio
18/07/2028
Riassunto
This work addresses the estimation of the Functional Conditional Average Treatment Effect (F-CATE )
– that is, the causal effect of a treatment on a functional outcome, conditional on covariates – in the
framework of Functional Data Analysis (FDA). The objective is to uncover how treatment effects evolve
over time and vary across levels of observed characteristics, a central challenge in modern causal inference
with functional responses.
To tackle this problem, we propose a novel class of functional meta-learners, specifically adapting
R- and DR-learners to the functional setting, and emphasizing doubly robust (DR) approaches that
combine outcome regression and propensity modeling to obtain unbiased estimators, even under partial
mis-specification. Our methodology is grounded in a rigorous theoretical analysis, providing convergence
guarantees and mathematical justification for the robustness properties of the estimators.
We conduct extensive simulation studies to assess performance under various nuisance function mis-
specification scenarios, illustrating the double robustness of the DR-learner: accurate recovery of F-CATE
is achieved as long as either the propensity model or the outcome model is correctly specified.
Finally, we apply our estimators to real-world data from the Survey of Health, Ageing and Retirement
in Europe (SHARE), evaluating the causal effect of chronic medical conditions on the quality-of-life
trajectories of older adults. The analysis controls for a rich set of covariates and provides visualizations
of treatment effects across different subpopulations. Results demonstrate the practical value of functional
meta-learners in capturing treatment heterogeneity over time.
– that is, the causal effect of a treatment on a functional outcome, conditional on covariates – in the
framework of Functional Data Analysis (FDA). The objective is to uncover how treatment effects evolve
over time and vary across levels of observed characteristics, a central challenge in modern causal inference
with functional responses.
To tackle this problem, we propose a novel class of functional meta-learners, specifically adapting
R- and DR-learners to the functional setting, and emphasizing doubly robust (DR) approaches that
combine outcome regression and propensity modeling to obtain unbiased estimators, even under partial
mis-specification. Our methodology is grounded in a rigorous theoretical analysis, providing convergence
guarantees and mathematical justification for the robustness properties of the estimators.
We conduct extensive simulation studies to assess performance under various nuisance function mis-
specification scenarios, illustrating the double robustness of the DR-learner: accurate recovery of F-CATE
is achieved as long as either the propensity model or the outcome model is correctly specified.
Finally, we apply our estimators to real-world data from the Survey of Health, Ageing and Retirement
in Europe (SHARE), evaluating the causal effect of chronic medical conditions on the quality-of-life
trajectories of older adults. The analysis controls for a rich set of covariates and provides visualizations
of treatment effects across different subpopulations. Results demonstrate the practical value of functional
meta-learners in capturing treatment heterogeneity over time.
File
Nome file | Dimensione |
---|---|
La tesi non è consultabile. |