Survey Statistics: Individualism Doesn’t Work (Even when Weighted)
Multilevel regression‑poststratification (MRP) aims to estimate population means, but most machine‑learning pipelines still optimize an individual‑level loss, which measures error for each respondent. The true MRP objective is a population‑level loss, focusing on the aggregate mean rather than individual predictions. Recent studies show that models minimizing individual‑level loss—even when weighted to the population—can be poorly ordered relative to the population loss. The aggregation step, not weighting, drives this discrepancy, and using post‑stratified estimates can lead to over‑fitting without proper validation.
SparseNUTS: Preconditioning Hierarchical Models in HMC with a Sparse “Laplace Approximation” At the Marginal Mode
Researchers led by Cole Monnahan released SparseNUTS, an R package that preconditions Hamiltonian Monte Carlo using a sparse Laplace approximation at the marginal mode of hierarchical models. By leveraging the sparse precision matrix from TMB or lme4, the method replaces...
New Course on Generative AI for Behavioral Science
Northwestern University introduced a graduate seminar titled "Generative AI for Social Science," merging computer science and communications students to examine how large language models can simulate human behavior. The course surveys emerging applications, methodological challenges, and metascientific concerns, culminating in...
“The Idea of Israel” . . . More Generally, The Idea of X, for Different Values of X
Ilan Pappe’s 2014 book *The Idea of Israel* chronicles the 1990s “post‑Zionist” surge in Israeli academia, arts and media that challenged the dominant patriotic narrative. The book argues that this brief period of critical scholarship was later curbed by a...
Bayesian Inferences and Frequentist Evaluations
Researchers Forster, Novelli, and Welch applied four frequentist and two Bayesian sequential designs to the COVID‑disrupted UK DISC clinical trial. All six approaches confirmed the trial’s original finding of treatment superiority but suggested different optimal points for restarting patient recruitment....
Ethics Corner: “As a Statistical Consultant, if You’re a Co-Author on a Substantive Paper, Is It Your Duty to...
Statistical consultants who become co‑authors must balance honesty with scope. They should fully disclose the analyses they performed and any limitations, but are not obligated to fix every statistical flaw beyond their contract. If a manuscript contains questionable methods, the...
From Junk Science (Largely Non-Political) to Junk Medical Treatments (Mostly Associated with the Far-Right): A Financial Connection
Paul Krugman highlights a growing financial link between the multi‑billion‑dollar wellness industry and right‑wing extremist movements. He notes that U.S. spending on wellness reaches roughly $500 billion annually, with nutritional supplements alone accounting for about $70 billion, while regulators like the FDA...