Back to search

Boston University

Methods for multi-domain prediction with applications to air pollution and ultrafine particulate matter

Abstract

dc:description.abstract

Many research and prediction settings involve data collected from multiple domains, such as different studies, sites, or other related data sources. Differences in distributions across domains can make model generalization to new domains challenging. Motivated by multi-site air pollution monitoring with between-site heterogeneity, particularly ultrafine particle (UFP) monitoring near Boston Logan International Airport, this dissertation develops new methods for prediction in heterogeneous multi-domain settings. First, we consider the multi-study ensemble framework, where individual models are fit to each domain and their predictions combined (ensemble approach), rather than fitting a single model to the combined data (merged approach). The ensemble approach can improve generalization when domain heterogeneity is high but, in practice, it is often unclear when it will outperform the merged approach. We propose a method that blends the predictions of both approaches using a domain-aware weight. The resulting method removes the need to choose between the two approaches, provides robust performance from low to high domain heterogeneity, and in some settings improves prediction beyond what either approach achieves alone. Next, we extend the multi-study ensemble framework to address negative transfer from misaligned domains, where systematic differences in outcome distributions can degrade prediction. Misaligned domains may still contain useful predictive information; for example, an air pollution monitor located farther from a primary emission source may record lower pollutant levels yet still be informative. We propose a rank-based ensemble strategy that modifies how misaligned domains contribute to the final ensemble. By applying rank-based transformations to the predictions, the method can preserve informative relative ordering while avoiding harmful differences on the original outcome scale. Finally, we investigate whether chemical transport model (CTM) outputs, which provide process-based estimates of pollutants, can improve UFP prediction. Their utility in statistical modeling of UFP remains largely unexplored. CTM outputs showed relatively weak association with monitored UFP and limited utility for prediction within observed sites, but appeared more useful at new, unobserved sites. We propose two CTM-guided methods that use CTM information to characterize relationships among sites and improve prediction at unobserved locations. These results highlight the potential value of CTMs for generalization in UFP modeling.

Degree

thesis:*
Name thesis:degree_name
Doctor of Philosophy
Level thesis:degree_level
doctoral
Discipline thesis:degree_discipline
Biostatistics
Grantor
Boston University
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Kojis, Daniel
Advisor dc:contributor.advisor
  • Patil, Prasad

Subjects

dc:subject × 1

Rights

Language dc:language.iso
en_US

Identifiers

dc:identifier.*
Handle dc:identifier.uri
https://hdl.handle.net/2144/53578

Chain of custody

source
Harvested from
Boston University
Base URL
open.bu.edu/oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
related terms
citation

Kojis, Daniel. Methods for multi-domain prediction with applications to air pollution and ultrafine particulate matter. doctoral thesis, Boston University, 2026. https://hdl.handle.net/2144/53578