Research
Working papers and preprints
RKHS Inference for Heterogeneous and Distributional Partial Effects
Working paper
We develop an RKHS framework for jointly capturing heterogeneity across treatment levels, covariates, and the outcome distribution, with inference for infinite-dimensional partial effect estimands.
Debiased Counterfactual Generation via Flow Matching from Observations
We use flow matching to estimate a “deconfounding flow” which transports samples from the observational distribution to the counterfactual distribution, under observed confounding. The estimation target is typically better behaved than estimating a generative model for the counterfactual distribution from scratch, and is especially well suited to mild confounding and high-dimensional outcomes.
Selected publications
Identifying Metric Structures of Deep Latent Variable Models
Latent variables in generative models are known to be identifiable only up to certain indeterminacies, even under strong model restrictions. In this paper, we show that all such indeterminacies are metric isometries under the pullback metric of the generator, which supports interpreting distances, angles and volumes in the latent space more reliable than their raw coordinates.
Distinguishing Cause from Effect with Causal Velocity Models
Functional causal discovery methods distinguish between cause and effect in bivariate systems by positing restricted structural models for the causal direction that are assumed to be misspecified for the anti-causal. Model classes are often selected because they can be estimated without making distributional assumptions on the noise. This paper leverages the dynamics of the effect distribution when perturbing the cause in a bivariate system to propose a new class of models that can be estimated using score matching without specifying a noise distribution.
Propensity Score Alignment of Unpaired Multimodal Data
Many different measurements (“modalities”) are possible for biological samples, each carrying potentially shared and unique information. Unfortunately, many such measurements are destructive and thus only one modality can be observed for each sample. Multimodal learning, which aims to harness the unique information from different modalities, typically requires paired samples. Motivated by potential outcomes as a missing data problem, we use the propensity score to estimate a common latent space for unpaired multimodal data when interventional samples are available, from which a distance can be computed to re-align different modalities.
Indeterminacy in Generative Models: Characterization and Strong Identifiability
We provide a precise description of latent variable indeterminacy (non-identifiability) in generative models in terms of measure-preserving transports in the latent space, thereby characterizing a series of identifiability results based on restrictions on the generator and base distribution space as restricting the admissible transports. Based on this, we find two ways to obtain strong identifiability, i.e., exact uniqueness of the latent variable, by fixing the base distributions either in a sufficient number of environments, or by using triangular monotonic increasing maps.