Estimand Drift Under Spatial Block Hold-Out
Year: 2026
Type: article
Abstract: Which population does a block hold-out error average over? On a terrain graph, the hold-out is a block cluster sample trimmed by a held-out-side buffer (unscored strips beside training), with closed-form inclusion probabilities for most nodes. For a predictor fixed across draws (random block permutations), error drift is exactly that sample’s design variance. In simulated radio coverage, a constant predictor at the imputed level has, at valid (non-imputed) nodes, a between-draw standard deviation averaging 8.02 dB over sixteen city–quadrant cells (95% jackknife interval [7.28;8.75]), 3.68 times ([2.30;5.06]) the 2.18 dB spread across cell means. In four first-quadrant cells, its precision is that of a simple random sample of 1.1 to 2.6 nodes. Trained graph neural network and multilayer perceptron errors drift too: their between-draw variance component exceeds the between-seed one, at the lower 95% limit, in three of four pairs. The bias against the domain mean splits exactly into design and ratio terms of either sign, both small beside the drift. Errors should be reported by population, as distributions over draws, beside a constant scored on the same blocks. Jointly scoring a block K-fold with the buffer cuts the constant predictor’s between-draw standard deviation to 0.76 dB.
Source: Mathematics
DOI: 10.3390/math14224129
Authors: Luis F. C. Seelig, Ronaldo M. Salles
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7225739341