|
Abstract:
|
Most nutritional epidemiology studies examining diet-disease trends use unsupervised dimension-reduction methods, like principal component analysis (PCA), to create diet patterns. The supervised dimension-reduction method described in Chun and Keles (2010), sparse partial least squares (SPLS), was found to create more interpretable diet patterns by selecting fewer components and predictors via a hard-thresholding approach while conserving predictive ability for continuous outcomes. We propose incorporating SPLS in the Cox proportional hazards model to study the relationship between correlated data and right-censored survival outcomes while imposing fewer model parameters than existing methods. We compare sparse and non-sparse PLS methods for survival data in their ability to create interpretable and predictive components, via simulations. Across various event rates and varied predictor covariance structures, the proposed method selected the fewest number of irrelevant predictors while maintaining a similar number of components and predictive ability. By using variable selection, this method can construct more interpretable dietary patterns related to time-to-disease-events.
|
