Steven Wilkins-Reeves

Logo

PhD candiate studying latnet variable models, networks and experimentation under interference

View My GitHub Profile

I am a PhD candidate in the Department of Statistics at the University of Washington. I have been fortunate to be advised by Professors Yen-Chi Chen, and Tyler McCormick as well as collaborate with Kwun Chuen Gary Chan. Previously I recieved my MS in Statistics at the University of Toronto and BSc in Mathematical Physics at Queen’s University. My email is stevewr (at) uw (dot) edu, and you can find my CV here.

Research

My research is motivated by the applications and theory of latent variable models, particularly with applications to statistical network analysis and neurodegenerative disease research. My theoretical interest lie in understanding the limits of deconvolution and identification of geometric properties of latent variable models. Recently, I have also been interested in causality under interference.

Estimating Curvature of Latent Variable Models

Models of Network formation often rely on assumptions that connections are drawn according to some model where nodes have positions on a latent space, and edge probabilities vary as the distances on the space. This paper considers the problem of estimating the curvature of the latent space and establishes asyptotically normal estimates of the curvature of the underlying space. This allows one to treat the estimates of the curvature of the underlying space as a well-behaving random variable, leading to downstream applications such as changepoint detection and tests of constant curvature of the latent space. We apply this to datasets in cybersecurity and collaboration networks and study their geometry.

Paper, code

Nonparametric Data Harmonization

In neuropsychological research collections of tests are used in order to measure particular cognitive traits. Over the versions of tests may be changed, leading to challenges in longitudinal analysis over prolonged time periods. We present a model for outcome imputation which posits a nonparametric latent distribution and converts the scores via matching the quantiles of the corresponding measurements. This allows a user to impute the missing outcomes on the desired scale, and allows for longer term study of cognitive decline.

Paper, code

Rate Optimal Deconvolution of the Semiparametric Rasch Model and High Dimensional Goodness-of-Fit Tests

We consider the semiparametric Rasch model, the canonical nonparametric item response theory model. This paper develops minimax opimal rate estimation in the problem of deconvolution of the latent trait, the first result of this form in item response theory.

Paper (Preprint forthcoming)

Teaching

I have been a teaching assistant for the following courses.

University of Washington

University of Toronto