An approach to determine the net predictive multivariate association pattern between cardiorespiratory fitness and a high-resolution physical activity spectrum in the presence of multicollinear covariates

Multivariate pattern analysis was recently extended with covariate projections to solve the challenging task of modelling and interpreting associations in the presence of linear dependent multivariate covariates. Within a joint model, this approach allows quantification of the net association pattern between the outcome and the explanatory variables and between the individual covariates and these variables. The aim of this paper is to apply this methodology to establish the net multivariate association pattern between cardiorespiratory fitness (CRF) and a high-resolution linear dependent physical activity (PA) intensity descriptor derived from accelerometry in children and to validate the crucial sub-regions in the PA spectrum predicting CRF. We applied the Andersen test as a measure of CRF, a PA descriptor with 23 PA intensities, and included age, sex, and three nearly linear dependent measures of adiposity as covariates. Net predictive association patterns are calculated for unadjusted and adjusted data. The explained variance in CRF was reduced from 25% for unadjusted data to 7.2% for adjusted data, but the association pattern was robust and dominated by vigorous PA across models. Models of sub-regions of the PA spectrum validated the association pattern for the full model and the crucial influence of vigorous PA.