Facial soft tissue thickness means have long been used as a proxy to estimate the soft tissue envelope, over the skull, in craniofacial identification. However, estimation errors of these statistics are not well understood, making casework selection of the best performing estimation models impossible and overarching method accuracies controversial. To redress this situation, residuals between predicted and ground truth values were calculated in two experiments: (1) for 27 suites of means drawn from 10 recently published studies, all examining the same 10 landmarks (N ≥ 3051), and tested against six independent raw datasets of contemporary living adults (N = 797); and (2) pairwise tests of the above six, and five other, raw datasets (N = 1063). In total, 380 out-of-sample tests of 416 arithmetic means were conducted across 11 independent samples. Experiment 1 produced an overarching mean absolute percentage error (MAE) of 29% and a standard error of the estimate (S(est)) of 2.7 mm. Experiment 2 yielded MAE of 32% and S(est) of 2.8 mm. In any instance, MAE was always ≥20% of the ground truth value. The overarching 95% limits of the error, for contemporary samples, was large (11.4 mm). CT-derived means from South Korean males and Black South African females routinely performed well across the test samples and produced the smallest errors of any tests (but did so for Black American male reference samples). Sample-specific statistics thereby performed poorly despite discipline esteem. These results—and the practice of publishing means without prior model validation—demand major reforms in the field.