Science and Statistics

Box's 1976 R.A. Fisher Memorial Lecture argues that scientific progress comes from a motivated iteration between theory and practice — not from pure theory alone, nor from undirected data collection. He uses Fisher's decade at Rothamsted Experimental Station (1919–1927) to illustrate how a scientist who stays in close contact with real problems can drive rapid theoretical advance.

The paper lays out principles of good scientific method. Science is a feedback loop: tentative theory confronts practice, discrepancies induce modified theory, and the cycle repeats. This requires flexibility — the willingness to seek out and exploit error, especially one's own. It requires parsimony: since all models are wrong, the scientist should seek economical descriptions rather than elaborate ones. And it requires selective worrying: being alert to what is importantly wrong, not distracted by irrelevancies ("mice when there are tigers abroad").

On mathematics, Box draws a sharp line: pure mathematics deals in conditional statements ("given A, does B follow?") with no claim on real-world truth. Applied mathematics makes assumptions known to be false but useful nonetheless. A statistical technique cannot be known to work unless it is used; major advances come only through the theory-practice iteration.

The bulk of the paper traces Fisher's Rothamsted work through episodes: analyzing decades of Broadbalk wheat-yield records (leading to orthogonal polynomials, residual analysis, and distributed-lag models), a faulty potato-fertilizer experiment (producing the first complete ANOVA table and the principle of randomization), the tea-tasting lady (illustrating experimental design), and collaborations on cotton-yarn strength (extreme-value theory) and bacterial counts (nonlinear design). Each shows theory arising from, and returning to, practical problems. Fisher also introduced the sponsor/critic dual role in data analysis: the analyst first sponsors a model to select procedures, then switches to critic to apply diagnostic checks on the residuals.

Box coins two pathologies that arise when the feedback loop breaks. Cookbookery forces every problem into one or two routine techniques without thought to objectives or assumptions. Mathematistry develops theory for its own sake, never touching practice, often redefining the problem to fit a tractable mathematical form rather than solving it. Both sever the iteration that drives genuine advance.

The paper closes with a call to train statisticians as scientists — integrating teaching with project work where practical experience is gained — rather than producing theorem-provers or cookbook-users in isolation. Fisher, Box argues, was great precisely because he was simultaneously an applied statistician, a mathematical statistician, a data analyst, and a designer of investigations. "It is surely because he was all of these that he was much more than the sum of the parts."