The Predictability of Random Paths

The apparatus in the South Kensington Museum, the one I constructed to demonstrate the laws of error, reliably shows the same pattern. Each shot, each bead, falls through the pins, seemingly by chance, yet coalescing into the familiar bell shape at the bottom. The individual path is unpredictable, but the aggregate, the distribution, is not. One observes this principle in many domains.

Consider the recent bulletin, concerning the efficacy of Aboriginal burning practices. A Cohen’s d of 4.44 is reported, a figure of striking magnitude. This value, we are told, is derived not from direct paired measurements on site, but from simulated uniform distributions over published fire-return-interval ranges. The ranges themselves, 3-7 years for managed sites and 15-40 years for unmanaged, are the primary data. The large d value, then, describes the separation of these two distributions. It is, in effect, a measure of the non-overlapping gap between the two ranges.

This method of inference, moving from the distribution of observed ranges to a calculated effect size, is a powerful one. Yet, how often do such critical methodological nuances, such precise descriptions of the data’s provenance, remain tucked away? The headline figure, the d=4.44, will be cited. The qualification, that it arises from a simulated distribution of ranges rather than raw, paired-site measurements, may be overlooked. It is a footnote, a parenthetical, a detailed amendment. The observed separation of fire-return intervals is the core finding. The simulation is the means by which a common statistical metric is applied to that separation.

One recalls the meticulous records of fingerprint patterns, gathered over decades. The initial observations, the distinct arch, loop, and whorl, were clear. But the system of classification, the means to make these individual observations yield a useful population-level insight, that took time to develop and disseminate. The substance of the discovery is in the pattern, the correlation, the distribution. The precise method of its elucidation, however, often remains in the fine print. The utility of the insight, its application, often precedes the full appreciation of the methodological rigour that underpins it. The critical step, in this instance, is the recognition of two distinct, non-overlapping distributions of fire-return intervals. The Cohen’s d quantifies this separation.