Lucas Resende

Robust statistics is about methods that remain reliable when data deviates from ideal assumptions — for instance, when observations have heavy tails or contain outliers.

• Trimmed Sample Means for Robust Uniform Mean Estimation and Regression
  with Roberto Imbuzeiro, AoS, 2025
  In this work we use the trimmed mean to obtain robust estimators for uniform mean and regression, obtaining optimal performance in terms of moments.
  [View on AoS] [View on arXiv] [Poster]
• Robust High-Dimensional Gaussian and Bootstrap Approximations for Trimmed Sample Means
  solo paper, preprint, 2025
  The literature on robust statistics focus on high probability concentration bounds. In this work I explore the high-dimensional Gaussian and bootstrap approximations properties of the trimmed mean. The key takeway is that it yields good approximations under heavy tails, while the empirical mean only support Gaussian approximations under light tails.
  [View on arXiv] [Poster]
• Improved Concentration for Mean Estimators via Shrinkage
  with Antônio Catão and Paulo Orenstein, 2025
  We study a family of adaptive shrinkage-based methods for robust mean estimation that achieve strong statistical guarantees. This project investigates their theoretical properties and practical performance across a range of applications.
  [View on arXiv]

Models with fixed effects often face the incidental parameter problem — biases that appear when the number of nuisance parameters grows with the sample size. My research develops methods for valid and numerically stable inference in such settings, extending classical econometric approaches.

• Statistical Inference in Large Multi-way Networks
  with Guillaume Lecué, Lionel Wilner, Philippe Choné, 2025
  We propose a scalable estimator for weighted networks that eliminates the incidental parameter problem. By generalizing tetrad methods to count data via Poisson models, our approach efficiently identifies both extensive and intensive margins in multi-way settings.
  [View on arXiv] [Github] [Python Package]
• Bootstrap Inference for Fixed Effects in Network Formation
  with Guillaume Lecué, Philippe Choné, ongoing
  While the standard network formation literature typically treats node-specific fixed effects as nuisance parameters to be bypassed or differenced out, this project develops a bootstrap-based approach to directly estimate and conduct valid inference on them. This semi-parametric method allows researchers to recover and analyze individual heterogeneity that is otherwise lost in traditional incidental parameter solutions.
  

I collaborate across public health, economics, and industry to apply modern machine learning and statistical methods to real-world problems.

• Measuring the impact of primary care gatekeeping via foundation models
  with Meilame Tayebjee, Guillaume Lecué and Philippe Choné, ongoing
  We combine GPT embeddings trained on the entire French national health database with off-policy evaluation to measure how primary care gatekeeping shapes patient outcomes.
  [Graph4Health's page]
• Epidemiological Modeling for COVID-19 School Reopening
  with COMORBUSS team, 2021
  Developed models to assess the effects of reopening strategies under varying vaccination and masking scenarios, informing public health decisions.
  [Technical Note for Policymakers] [Outreach on Piauí Magazine] [PLOS One paper]
• Deep Hashing via Householder Quantization
  with Lucas Schwengber, Paulo Orenstein and Roberto Imbuzeiro, 2023
  Motivated by large-scale similarity search in streaming platforms, this project led to a new deep-hashing quantization method that improves fast retrieval of similar items.
  [View on arXiv]