Marginal and sequential biases in human random sequence generation are uncorrelated

AbstractHuman-generated random sequences depart from randomness in two logically distinct ways: unequal use of the response options (marginal bias) and dependence of each choice on the preceding choice (sequential bias). Accounts in which both biases reflect a single latent dimension imply that individuals who deviate strongly in one respect should deviate strongly in the other. We test this implication directly in an open dataset of 851 adults who each produced 200 random digit choices. Marginal bias was measured with Cohen’s w and sequential bias with Cramer’s V on the lag-1 transition table. Both measures were internally consistent (Cronbach’s α = .83 for each), yet their rank correlation across participants was near zero (Spearman’s ρ = .004, bootstrap 95% CI [−.066, .074]). A surrogate control confirmed that the geometry of the measures induces no correlation across the observed range of behavior. The design had 80% power to detect |ρ| ≥ .096. The two biases are therefore best viewed as distinct dimensions of non-random behavior, not as expressions of a single latent trait.