Recent content on Herb Susmann
I had the pleasure of presenting our work on non-overlap treatment effect bounds at the American Causal Inference Conference held in Salt Lake City. I was part of a really nice invited session on positivity violations, with other talks given by Paul Zivich, Harsh Parikh, and with a discussion by Caleb Miles.
My joint work with Jeffrey Näf and Junhyung Park on Causal Distributional Random Forests was accepted at AISTATS 2026.
Our paper on shrinkage estimation for estimating large sets of causal parameters is now available at The Annals of Applied Statistics: Asymptotically efficient data-adaptive penalized shrinkage estimation with application to causal inference. You can also download a PDF of the article. This is joint work with Yiting Li, Mara A. McAdams-DeMarco, Wenbo Wu, and Iván Díaz.
I have a new preprint on ArXiv with NYU Grossman PhD student Antonio D’Alessandro: The Counterfactual Combine: A Causal Framework for Player Evaluation. We show how evaluating the performance of healthcare providers is fundamentally the same problem as evaluating player performance in sports, and adapt causal inference methods developed for healthcare provider profiling to the sports…
I presented our work on non-overlap treatment effect bounds at the Online Causal Inference Seminar, an international seminar on causal inference.
My paper with Adrian Raftery Bayesian Projection of Extant Refugee and Asylum Seeker Populations is now published in Demography. We propose a Bayesian framework for producing probabilistic projections of refugee and asylum seeker populations by country of origin.
I presented joint work with Adrian Raftery on “Projections of Refugee and Asylum Seeker Populations by Country of Origin and Destination” at the Wittgenstein Centre Conference 2025 on “Demographic Perspectives on Migration in the 21st Century”. Part of the work we presented has now been published in Demography. The event took place at the Austrian Academy of Sciences in Vienna. Ceiling…
I had the pleasure of visiting Ludwig-Maximilians-Universität München to collaborate with Michael Schomaker under the auspices of the Center for Advanced Studies. While I was there, I gave a joint talk with my post-doc advisor Iván Díazon Causal Inference Based on Machine Learning for Complex Longitudinal Exposures.
I was happy to visit UMass Amherst to give a talk for their Statistics and Data Science Seminar Series on our non-overlap bounds project, which is with Alec McClean and Iván Díaz.
I had the pleasure of giving an invited talk today in my home department, the Department of Population Health, at NYU Grossman School of Medicine as part of their Emerging Leaders in Research lecture series. I presented recent joint work with Alec McClean and Iván Díaz on non-overlap treatment effect bounds. An annotated version of the slides from today’s talk are also available.
Our paper Doubly robust nonparametric efficient estimation for healthcare provider evaluation is now published in JRSS-A. We propose doubly robust estimators based on Targeted Minimum Loss-Based Estimation (TMLE) for indirect standardization parameters relevant to healthcare provider profiling. This was joint work with Yiting Li, Mara A. McAdams-DeMarco, Iván Díaz, and Wenbo Wu.
I have some new work out on ArXiv today discussing a non-parametric method for estimating time-smoothed effect curves. This method makes it practical to estimate the sequential effect of interventions on time-varying outcomes. This is joint work with Nick Williams, Richard Liu, Jessica Young, and Iván Díaz: Computationally and statistically efficient estimation of time-smoothed counterfactual…
I have a new preprint out on joint work with Alec McClean and Iván Díaz on a novel method for handling overlap (positivity) violations via the use of partial identification bounds: Non-overlap Average Treatment Effect Bounds.
My paper on non-parametric bounds for treatment effects under outcome left-censoring was just published in the Journal of Applied Statistics: Environmental Statistics and Data Science. Environmental data often exhibit left-censoring, because measurements are subject to a lower limit of detection. For example, instruments might not be able to measure the concentration of a chemical in a sample…
This post gathers some notes on the history of various topics with a focus on the history of semi-parametric efficiency theory and related estimators. Some of these are more fleshed out than others. I intend to add to this as I learn more. Bibtex is provided at the bottom. Statistical Functionals and von Mises calculus The concept of a statistical functional has origins in von Mises (1947). There…
I visited JSM for the first time this year to present our work on generalizations of the Average Treatment Effect on the Treated to longitudinal settings as part of the invited session “Advanced Strategies for Longitudinal Causal Inference and Treatment Switching”. A preprint is available on ArXiv: Longitudinal Generalizations of the Average Treatment Effect on the Treated for…
Antoine Chambaz and I had our work on Quantile Super Learner published in Computational Statistics & Data Analysis. We establish theoretical guarantees for Super Learner ensembles (also known as model stacking) built to estimate conditional quantiles, and apply the method to some real-world applications relevant to solar forecasting.
We have a new preprint out looking at how to apply shrinkage estimation to estimating large sets of causal parameters: Asymptotically Efficient Data-adaptive Penalized Shrinkage Estimation with Application to Causal Inference. The main idea is to define a new causal parameter of interest that is the solution of an optimization problem that balances fidelity to the parameter of interest and a…
Leontine Alkema and I had our article on flexible modeling of demographic transitions using B-splines published in JRSS-C. This approach is now used in the Family Planning Estimation Tool for estimation and projection of family planning indicators; more details are available in a preprint led by Leontine.
There was a lot of statistical demography activity at PAA this year, centered around three consecutive sessions on Saturday morning. I gave a talk on my work with Leontine Alkema on using Bayesian shrinkage priors to handle shocks in demographic and health indicators; the slides are available online. Adrian Raftery also presented our work on projecting refugee and asylum seeker populations, which…
I had the pleasure of visiting ENAR for the first time to give a talk on our work on generalizations of the Average Treatment Effect on the Treated to longitudinal settings. Slides and speaker notes for my talk can be found here: Longitudinal Generalizations of the Average Treatment Effect on the Treated for Multi-valued and Continuous Treatments
Our paper Probabilistic prediction of arrivals and hospitalizations in emergency departments in Île-de-France was published in the International Journal of Medical Informatics. We used a combination of ensemble learning and Adaptive Conformal Inference to predict emergency department arrivals and hospitalizations in and around Paris. This paper builds on some previous theoretical work establishing…
Two new preprints out today related to our work on Bayesian hierarchical models for demographic and health indicators. The first preprint, joint with Leontine Alkema and Evan Ray, introduces a data model that incorporates Bayesian shrinkage priors as a way to flexibly handle data sources with reporting issues. We have also used shrinkage priors for modeling indicators with large shocks (for…
Jeffrey Näf and I have a new preprint available on ArXiv, Causal-DRF: Conditional Kernel Treatment Effect Estimation using Distributional Random Forest. We took some of Jeff’s previous work on building distributional random forest’s and adapted it to the case of estimating kernel treatment effects. Jeff contributed the heavy-duty theory and I worked on the coding and simulation studies…
This post gathers resources that may be helpful in learning the semi-parametric statistical theory that is relevant to statistical methods development for causal inference. I intend to continually update this post. BibTex is provided at the bottom. Articles All of Edward Kennedy’s expository writing on the subject is excellent; I recommend starting with the following two articles: Kennedy,…
I and Leontine Alkema have a new preprint available, Flexibly Modeling Shocks to Demographic and Health Indicators with Bayesian Shrinkage Priors. We discuss how to build models for estimating and projecting demographic and health indicators that exhibit large short-term fluctuations, which we call shocks, using Bayesian shrinkage priors.
We now have a new preprint available, Doubly Robust Nonparametric Efficient Estimation for Provider Evaluation, co-authored with Yiting Li, Mara A. McAdams-DeMarco, Iván Díaz, and Wenbo Wu. We develop doubly robust estimators for causal parameters relevant to evaluating the performance of healthcare providers.
We recently released an updated preprint of our work on longitudinal generalizations of the Average Treatment Effect on the Treated based on longitudinal modified treatment policies. The updated preprint is titled Longitudinal Generalizations of the Average Treatment Effect on the Treated for Multi-valued and Continuous Treatments and is co-authored by Nicholas T. Williams, Kara E. Rudolph, and…
I recently had an article published in COMPUTO, a journal of the French Royal Statistical Society, titled AdaptiveConformal: An R Package for Adaptive Conformal Inference. The article was cowritten with my collaborators Antoine Chambaz and Julie Josse. COMPUTO has an interesting publication process that focuses on reproducibility. We wrote the manuscript as an R Quarto document, and we were…
Chiesa di San Simeon Piccolo, Venice, Italy. I had a wonderful time attending ISBA 2024 in last week, held at the Ca' Foscari University of Venice. I had the opportunity to present at the invited session “Novel and Flexible Bayesian Approaches for Causal Inference in Complex Settings”, which was organized by Mike Daniels and Maria Josefsson and chaired by Arman Oganisian. The slides…
Eugène Boudin, 1863: Beach Scene at Trouville. Courtesy National Gallery of Art, Washington. Setting up and running simulation studies are a ubiquitous task in applied statistics. In this post, I’ll write up a small simulation study to show how I usually approach setting them up in R. This post will assume a certain level of familiarity with causal inference. The simulation study will compare a…
Fitting a Gaussian Process and its derivative in Stan.
Conditioning a Gaussian Process on derivative observations, with code in R.
A few years ago, The New York Times published an article about several Major League Baseball players who use the sitcom Friends to improve their English. Friends seems to be a very popular tool for learning English: there’s even an ESL program developed around it, and you can easily find advice on how to use Friends as a language learning tool. In my own language learning I’ve relied heavily on…
Drawing from a Gaussian Process and its derivative.
Gun dealers in the U.S. are required to conduct instant background checks before selling weapons to individuals. The FBI provides data for the number of these background checks performed by month/year, which serves as a proxy for the total number of gun sales in the U.S. I brought the data into R for a quick and dirty analysis, with the intent of finding spikes in background checks around major…
Note: this post has been updated with more recent data. I often use random walk/autoregressive models in my research as a component in time-series analysis, and I wanted to get some more experience fitting them to data. FiveThirtyEight publishes several polling datasets, including polling for the 2020 Democratic presidential primary. I used Stan to fit a Bayesian random walk model to the polling…
Autoregressive (AR) processes are a popular choice for modeling time-varying processes. AR processes are typically written down as a set of conditional distributions, but if we do some algebra we can show how they can also be written as a Gaussian process. One reason having a Guassian process representation is useful is because it makes it more clear how an AR process can be incorporated into…
Write flexible Stan models by using the R formula interface.
We developed a novel interface for reporting results to participants of exposure studies.
New York Times columnist Nick Kristof covers the Detox Me Action Kit project.
Analyzing tweets from @dog_rates.
Calculating an optimal route between all the Boston area breweries using a Traveling Salesman Problem solver.
Calculating the probability of starting with a complete word in Bananagrams.
We built a crystal radio sets at a workshop for teens.
It turns out it is possible to write a minimal raytracer in Bash.
For my undergraduate degree, I studied Mathematics at the State University of New York at Geneseo. I then worked from 2014-2018 as a software developer at Silent Spring Institute in Newton, Massachusetts, where I worked on methods for making complex scientific information accessible to lay audiences. From 2018-2022 I pursued a PhD in Biostatistics in the Department of Biostatistics & Epidemiology…
import {Runtime, Inspector} from 'https://cdn.jsdelivr.net/npm/@observablehq/runtime@4/dist/runtime.js'; import define from 'https://api.observablehq.com/@herbps10/b-splines.js?v=3'; new Runtime().module(define, Inspector.into('#observablehq-66aca9e3')); .observablehq--inspect { display: none; } Source code: Observable notebook.
import {Runtime, Inspector} from 'https://cdn.jsdelivr.net/npm/@observablehq/runtime@4/dist/runtime.js'; import define from 'https://api.observablehq.com/@herbps10/dirichlet-distribution.js?v=3'; new Runtime().module(define, Inspector.into('#observablehq-763d3248')); .observablehq--inspect { display: none; } Source code: Observable notebook.
import {Runtime, Inspector} from 'https://cdn.jsdelivr.net/npm/@observablehq/runtime@4/dist/runtime.js'; import define from 'https://api.observablehq.com/@herbps10/gaussian-processes.js?v=3'; new Runtime().module(define, Inspector.into('#observablehq-870dca1f')); .observablehq--inspect { display: none; } Source code: Observable notebook.