Abstract:
|
Multifactor dimensionality reduction (MDR) was developed by Ritchie et al in 2001 to identify high order gene-gene or gene-environment interactions for large dimensional data; different combinations of variable values are labeled as high or low risk to characterize each subject in terms of an optimal gene-gene or gene-environment interaction. Extensions of MDR have been pursued, such as quantitative MDR (Gui et al 2013), aggregated MDR (Dai et al 2013), and aggregated quantitative MDR (Crouch 2016). Our work considers a situation where multiple interactions of a particular order may be considered simultaneously but not necessarily collapsed into a single risk score as in aggregated or aggregated quantitative MDR. Rather, we obtain P (>1) risk scores and use them to predict the continuous outcome for each subject. There are multiple ways to define the P risk scores, and part of our work entails selecting a best set of P risk scores. A simulation study shows that a best set of 2 risk scores can predict an outcome better than competing models based on a single risk score.
|