In practice, for a given data set, where nonlinear associations or high-order interactions are suspected, complex approaches like random forests can at least serve as a benchmark predictor: If a linear or other parametric model with a limited number and degree of interaction terms can reach the (cross validated or test sample) prediction accuracy of the more complex model, the extra complexity may be uncalled for and the simpler, interpretable model should be given preference. If, however, the prediction accuracy cannot be reached with the simpler model, and, for example, the high importance of a variable in a random forest is not reflected by its respective parameters in the simpler model, relevant nonlinear or interaction effects may be missing in the simpler model and it may not be suited to grasp the complexity of the underlying process.
