[Paper Review] What are the most important statistical ideas of the past 50 years?

Andrew Gelman, Aki Vehtari|arXiv (Cornell University)|Nov 30, 2020

Statistical Methods and Inference156 references4 citations

TL;DR

This paper identifies and reviews eight transformative statistical ideas from the past 50 years—counterfactual causal inference, bootstrapping, overparameterized models with regularization, Bayesian multilevel modeling, generic computation algorithms, adaptive decision analysis, robust inference, and exploratory data analysis—highlighting their conceptual foundations, computational evolution, and impact on modern data science and applied research.

ABSTRACT

We review the most important statistical ideas of the past half century, which we categorize as: counterfactual causal inference, bootstrapping and simulation-based inference, overparameterized models and regularization, Bayesian multilevel models, generic computation algorithms, adaptive decision analysis, robust inference, and exploratory data analysis. We discuss key contributions in these subfields, how they relate to modern computing and big data, and how they might be developed and extended in future decades. The goal of this article is to provoke thought and discussion regarding the larger themes of research in statistics and data science.

Motivation & Objective

• To identify and synthesize the most important statistical ideas that have emerged in the past 50 years, based on their conceptual novelty and impact on modern statistics and data science.
• To examine how these ideas evolved from pre-1970 foundations into distinct, widely applicable frameworks that now underpin contemporary statistical practice.
• To explore the interplay between these statistical advances and the rise of big data, high-dimensional modeling, and modern computational power.
• To stimulate discussion on future research directions at the intersection of computation, causal inference, and model interpretability in statistics and data science.

Proposed method

• Categorizes key statistical developments into eight core ideas based on conceptual and historical analysis of literature and expert input.
• Traces the evolution of each idea from its pre-1970 antecedents to its modern formulation, emphasizing conceptual shifts rather than chronological order.
• Uses case studies and representative works (e.g., Rubin’s potential outcomes, Efron’s bootstrap, Gelman’s multilevel modeling) to illustrate methodological advances.
• Emphasizes the role of computation in enabling simulation-based inference, regularization, and scalable modeling, especially in high-dimensional and complex data settings.
• Integrates insights from diverse fields—econometrics, epidemiology, computer science, psychology—to show cross-disciplinary convergence.
• Proposes future research directions by linking current methods to emerging challenges in model interpretability, robustness, and validation of inferential procedures.

Experimental results

Research questions

• RQ1 What are the most influential statistical ideas of the past 50 years, and how do they differ from earlier statistical thinking?
• RQ2 How have computational advances enabled the development and adoption of simulation-based and regularization-based methods like bootstrapping and overparameterized modeling?
• RQ3 In what ways has counterfactual reasoning transformed causal inference in observational studies across disciplines?
• RQ4 How do Bayesian multilevel models improve inference in hierarchical and complex data structures compared to traditional approaches?
• RQ5 What future directions should statistical research pursue at the intersection of computation, causality, and model interpretability?

Key findings

• Counterfactual causal inference, formalized through potential outcomes and graphical models, has enabled rigorous identification of causal effects under structural assumptions, transforming fields from epidemiology to policy analysis.
• The bootstrap and simulation-based inference have become foundational by replacing analytical approximations with computationally feasible resampling, especially in complex or non-normal settings.
• Overparameterized models with regularization (e.g., ridge, lasso) have enabled effective inference in high-dimensional data, balancing model fit and generalization.
• Bayesian multilevel models have significantly improved hierarchical and clustered data analysis by borrowing strength across groups and enabling partial pooling.
• Generic computation algorithms, such as Hamiltonian Monte Carlo and variational inference, have made Bayesian inference scalable and practical for complex models.
• Exploratory data analysis and robust inference have gained renewed importance in the era of big data, particularly for detecting model misfit and ensuring reliability in high-dimensional and messy data.