It is the 1950s in the South African goldfields, and a mining engineer must estimate the ore grade at an unsampled point from a scatter of costly boreholes — and simple interpolation ignores that nearby samples are more informative than distant ones in a way the spatial data themselves reveal. Build a regression on geography: model the correlation between samples as a function of the distance between them, then predict each new location as a weighted average of the boreholes, with weights and an honest error variance from that structure. Get it wrong and you sink a shaft where the gold isn't, or walk away from a rich seam your interpolation smoothed into mediocrity — spatial prediction lives or dies on modeling the correlation, not ignoring it.
Chose Partial pooling with a hierarchical model — wrong. Bayesian computation and model checking was the one that fit.
Gelman's hierarchical Bayesian workflow, developed from the 1990s onward, can incorporate spatial covariance structure directly into a joint posterior, treating ore grade at unsampled locations as a latent variable with a spatially correlated prior, a genuinely correct modern way to handle this exact problem. His applied emphasis on checking model assumptions carefully is useful discipline for validating a chosen covariance function. But his primary body of work centers on hierarchical modeling in political science and public health broadly rather than geostatistics specifically, and kriging's classical formalization by Krige and Matheron predates and sits somewhat outside the Bayesian workflow he is best known for. Andrew Gelman would be starting close to scratch on this problem's specific statistical content, however formidable the surrounding general expertise may be.
Kalman's 1960 filter recasts optimal estimation from noisy, correlated observations as recursive state-space inference, and kriging can be reformulated as a spatial analogue of Kalman filtering under certain covariance structures, both being instances of Gauss-Markov optimal linear prediction from correlated data. His rigorous treatment of how to combine an old estimate with new noisy information transfers conceptually to combining scattered borehole readings into a spatial prediction. But his own applications were aerospace guidance and control, not mining or spatial statistics, and he did not build the specific distance-based covariance framework kriging depends on, leaving him a mathematically adjacent contributor rather than this problem's specific method. That leaves Rudolf Kalman as a capable, well-informed generalist reaching into this problem from an adjacent tradition, rather than someone who built the specific correction firsthand.
Battle #13 · 8/9/2026, 5:02:27 PM · this result is deterministic: the same two personas on this problem always resolve the same way.