Cross-validation can give many equally good answers, and the one you pick changes what you see.

Geographically weighted principal component analysis, or GWPCA, finds local patterns in complex map data. Its results depend heavily on a setting called bandwidthbandwidthA setting that controls how many nearby points are used when analyzing each location; a small bandwidth uses very local data, a large one uses more distant data., which controls how large each local neighborhood is. Researchers pick this bandwidth using cross-validation, a test that scores settings by how well they predict held-out data. The problem is that cross-validation often turns up multiple good bandwidth scores instead of one clear winner. The authors studied this ambiguity by grouping all bandwidths that score similarly well into an equivalent set. They found these tied settings yield very different data patterns and trade-offs. In simulations, this tied set covered roughly a quarter of all tested settings. The true best bandwidth usually sat inside this set, but it differed from the single chosen minimum.

Why it matters

Geographically weighted PCA is used to find spatial patterns in data like voting behavior or soil chemistry. If the method's key tuning parameter has multiple plausible values that produce different results, conclusions drawn from a single choice may be misleading. This paper shows that ambiguity is the rule, not the exception, and offers a way to report it honestly.

Questions answered