What I work on, who I work with, and where to start reading.
How do we learn, while people use an app, take a course or follow a treatment, which option works best for whom, and when? I develop reinforcement-learning and multi-armed bandit methods that are statistically considerate: they optimise outcomes while still allowing valid scientific conclusions, often within a Bayesian framework (e.g. Thompson sampling).
Applications: biostatistics (just-in-time adaptive interventions in mobile health, dynamic treatment regimes, SMARTs); education (adaptive experiments in online courses); finance (sequential decision-making under uncertainty).
Team: Bibhas Chakraborty and former co-supervised PhD students Xueqing Liu and Xinru Wang (Duke-NUS, Singapore); Joseph Jay Williams (University of Toronto); the Google team behind PEARL, including Amy Armento Lee (PhD student, University of Cambridge) and Hulya Emir-Farinas; Caroline A. Figueroa and Adrian Aguilera (DIAMANTE study); my PhD student Simone Cuonzo (Sapienza).
Software: the R package countts for Thompson sampling with count outcomes.
Where to start: Reinforcement Learning in Modern Biostatistics: Constructing Optimal Adaptive Interventions (International Statistical Review, 2025).
How sure are we? I work on quantifying uncertainty in a principled way, from Bayesian inference to conformal prediction: their interplay and how to get honest coverage guarantees, highest-density regions for complex and multimodal data, and calibration testing with e-values. Recent work brings conformal ideas into sequential decision-making (conformal bandits) and into small-area estimation.
Applications: finance (volatility models, tail risk, decisions under uncertainty); official statistics (small-area estimation and accuracy of register-based statistics with ISTAT; disaster-induced production losses in agriculture with FAO).
Team: Brunero Liseo (Sapienza University of Rome); my PhD students Simone Cuonzo and Emanuele Masillo (Sapienza); Davide Carmelo Di Leonforte; colleagues at ISTAT and at the FAO Statistics Division.
Where to start: The interplay between Bayesian inference and conformal prediction (Philosophical Transactions of the Royal Society A, 2026).
Adaptive designs, such as response-adaptive randomisation and bandit-driven experiments, assign more participants to the options that seem to work better. This is more ethical and often more efficient, but it changes how the data must be analysed: classical tests and estimators can become biased or lose error control. I develop tests with guaranteed error control for adaptively collected data, and study the trade-off between learning and helping participants.
Applications: biostatistics (response-adaptive clinical trials in rare diseases, such as StratosPHere 2 in pulmonary arterial hypertension); education (adaptive experiments in online courses).
Team: Sofia S. Villar and Rajenki Das ( MRC Biostatistics Unit, University of Cambridge); Mark Toshner (Department of Medicine, University of Cambridge, and Royal Papworth Hospital); Joseph Jay Williams (University of Toronto), Anna Rafferty and Audrey Durand.
Where to start: Finite-sample and asymptotic validity of a novel design-based test statistic for response-adaptive randomisation (Biometrics, 2025).
Education is where many of my methods meet real data. With the Adaptive Experimentation Accelerator, winner of the 2023 XPRIZE Digital Learning Challenge, we run bandit-driven adaptive experiments in real online courses; a Sapienza research grant (PI, 2025) now brings these methods into our own statistics courses. With ISTAT, I develop scalable methods to measure the accuracy of education statistics built from multisource registers. And with colleagues at Sapienza, we model students’ careers, for example to understand late graduations.
Team: Joseph Jay Williams (Intelligent Adaptive Interventions Lab, University of Toronto), Anna Rafferty, Audrey Durand, and colleagues at Carnegie Mellon University and North Carolina State University; Piero Demetrio Falorsi, Stefano Falorsi, Diego Chianella and Giorgio Alleva ( ISTAT); Brunero Liseo, Andrea Tancredi and Edoardo Pandolfo (Sapienza).
Where to start: Platform-based Adaptive Experimental Research in Education: Lessons Learned from Winning a Digital Learning Challenge (LAK25, 2025) and Scalable Generalised Accuracy Estimation for Multisource Register-based Official Statistics (JRSS-A, 2026).
Athletes’ biological passports are long, multivariate records of blood and urine markers. We build Bayesian copula-based models that look at these markers jointly over time, to flag suspicious profiles more reliably than one-marker-at-a-time tests.
Funding: World Anti-Doping Agency, WADA Research Grant (Co-PI; PI: Brunero Liseo); Sapienza research grant (PI).
Team: Brunero Liseo (Sapienza); my PhD students Emanuele Masillo and Eva Pigi Vyltanioti; Dimitris Karlis (Athens University of Economics and Business); NADO Italia and the World Anti-Doping Agency.
Where to start: A Copula-based Bayesian Framework for Doping Detection (Computational Statistics, 2025).
Interested in a PhD or a research visit on one of these topics? Get in touch at nina.deliu [at] uniroma1 [dot] it with a short CV and the project that interests you.
👶 Heads-up: I am on maternity leave until late February 2027, so replies may be slower than usual.