“There are two cultures in the use of statistical modeling to reach conclusions from data.”
Leo Breiman’s essay described a divide between researchers who begin with a stochastic model of how data are generated and those who treat the mechanism as unknown and optimize predictive accuracy. He argued that statistics had invested too heavily in the first culture and could gain scientific value from algorithmic methods such as decision trees and ensembles. The provocation helped bring predictive evaluation into closer conversation with statistical explanation. That context prevents a milestone from becoming a slogan and keeps attention on the concrete problem the authors were trying to solve.
Prediction and explanation answer different questions. A model can forecast accurately without revealing a trustworthy causal mechanism, while an interpretable model can express assumptions clearly yet predict poorly outside them. Breiman urged researchers to test performance on unseen data and remain open to methods that capture complex structure. The quote does not require choosing one culture forever; it asks analysts to be explicit about the goal and the evidence supporting it. That reading preserves both ambition and humility. It recognizes genuine progress while refusing to stretch the evidence beyond the conditions under which it was produced.
Machine learning made Breiman’s challenge central. Cross-validation, held-out tests, random forests, and benchmark comparisons reward empirical performance, while high-stakes fields still need causal reasoning and interpretable assumptions. Teams should decide whether they are predicting risk, estimating an intervention’s effect, explaining a historical pattern, or supporting a decision. Using the wrong kind of evidence can create confidence without answering the question people actually asked. For contemporary teams, this means documenting the task, comparison baseline, operating conditions, affected users, and signals that would justify revising or withdrawing a deployment.
Predictive accuracy can hide instability, leakage, and structural inequity. Explanatory stories can likewise disguise assumptions as facts. Responsible analysis combines rigorous out-of-sample testing with domain knowledge, uncertainty, sensitivity analysis, and attention to who is represented. When a model guides action, evaluation should include the feedback created by that action rather than treating the data-generating process as fixed. A durable implementation keeps human agency visible through consent, explanation, meaningful alternatives, and accountability that cannot be delegated to the model. Breiman’s two cultures remain in productive tension. The strongest work often draws from both: empirical discipline about prediction and conceptual discipline about what a model can legitimately explain. Responsible progress preserves the insight while widening the evaluation to include reliability, fairness, rights, and consequences outside the benchmark. A mature field carries breakthroughs forward without erasing their boundaries, using history to sharpen experiments and to recognize the new responsibilities created by wider deployment.
One culture starts from an assumed data-generating model; the other evaluates algorithms mainly through prediction on new data.
Breiman argued that statistical practice should engage more seriously with the second approach.
A system may forecast an outcome accurately without identifying what would change that outcome.
Intervention decisions require causal assumptions and evidence beyond ordinary predictive performance.
Model choice should begin with the actual task: forecasting, explanation, causal estimation, or decision support.
Validation, uncertainty, domain knowledge, and post-deployment effects all matter when outputs guide action.
Explore more "Quotes of The Day"
Continue exploring how writers, thinkers, artists, and advocates have described the choices that shape an everyday life.