A watershed model of individual differences in fluid intelligence

Abstract

Fluid intelligence is a crucial cognitive ability that predicts key life outcomes across the lifespan. Strong empirical links exist between fluid intelligence and processing speed on the one hand, and white matter integrity and processing speed on the other. We propose a watershed model that integrates these three explanatory levels in a principled manner in a single statistical model, with processing speed and white matter figuring as intermediate endophenotypes. We fit this model in a large (N=555) adult lifespan cohort from the Cambridge Centre for Ageing and Neuroscience (Cam-CAN) using multiple measures of processing speed, white matter health and fluid intelligence. The model fit the data well, outperforming competing models and providing evidence for a many-to-one mapping between white matter integrity, processing speed and fluid intelligence. The model can be naturally extended to integrate other cognitive domains, endophenotypes and genotypes.

Keywords: Cognitive ageing; Fluid intelligence; Processing speed; Structural Equation Modelling; Watershed model; White matter.

Fig. 1

A watershed model of psychopathology (adapted from Cannon and Keller, 2006). Point “1” represents the most complex phenotype, such as schizophrenia (or fluid intelligence for our purposes). Points “2a–2d” represent endophenotypes, such as lower-level behavioural consequences; points “3a–3g” represent the neural antecedents of those behavioural phenotypes. Points 4a–4d represent hypothetical genetic influences (not measured here).

Fig. 2

Behavioural measurements. Simple RT (a), choice RT (b), AudioVisual RT (C) and Cattell (d) (fictional example item shown).

Fig. 3

A single-factor confirmatory factor analysis for Cattell fits data well (left). Linear and polynomial fit of age-related differences in fluid intelligence (right).

Fig. 4

MIMIC model for Processing Speed and Fluid intelligence. Below age-related trajectories for each processing speed measure ranging from strong CRTspeed, r=−0.64) to absent (AVspeed, r=0.03, n.s..). Residual covariances between PS variables are allowed but not shown for simplicity.

Fig. 5

A) All ten white matter tracts used in our analysis, based on the JHU Atlas. B) Differential ageing of the ten tracts, correlations ranging from −0.71 (Forceps Minor) to −.10 (Ventral Cingulum, or CINGHipp).

Fig. 6

Full watershed model. Significant parameters are shown in green and red, R-squared is represented as the degree of shading of the variables. Residual covariances between processing speed variables and white matter tracts are allowed, but not shown for simplicity. See Supplementary Table 1 for the full covariance matrix, and Supplementary Table 2 for the unstandardised parameter estimates and se's.