Score gains on g-loaded tests: No g

64 test-retest studies, N = 26,990

Pooled test-retest, coaching and training samples on cognitive-ability batteries; mixed ages including school-age

Meta-analysis using the method of correlated vectors - correlating each subtest's g-loading with the score gain observed on that subtest. Graded C because the method is contested (it is sensitive to artifact corrections and to the number of subtests) and because it pools primary studies of mixed quality. The direction of the result, however, has been reproduced across independent datasets and is corroborated by an entirely different design (Lievens et al. 2007, which tests criterion validity directly rather than g-loadings).

The correlation between how strongly a subtest loads on g and how much scores rise on that subtest with practice, coaching or training is approximately -1.00 after artifact corrections. The subtests that gain most are the ones that measure the construct least. The authors' word for the gains is "hollow", and they report that practice and coaching actively REDUCE a test's g-loadedness - the test measures less of what it was built to measure after preparation than before. For this archive this is the mechanism behind the whole test-preparation verdict: a coached score is not a proportionally smaller version of a real gain, it is a gain in a different thing.

Low relevance in the usual sense - the finding is about what changes rather than about who differs. It does bear on the archive's premise - it is a direct measurement-level demonstration that training moves test performance without moving g, matching Ritchie, Bates & Deary (2015) on schooling.

The method of correlated vectors has been criticised (notably by Ashton & Lee and by Dolan), so the exact magnitude should not be over-read; the near-perfect negative direction has been reproduced across several independent gain datasets and is convergent with criterion-validity evidence obtained by a different route.