Abstract:
|
In much practical work, model selection is an important part of the work flow. While traditional practices of model selection (such as information criteria or pre-testing) are by now well studied, in recent years alternative approaches have emerged. Two distinct suggestions for mitigating the model selection problem are using penalized estimation methods, or using model averaging. In this work, we compare the most popular penalized and frequentist model averaging estimators. In addition, a simulation study is conducted using setups from both strands of the literature and the performances are also investigated in an application. We find that no method uniformly dominates the other, but that the model averaging estimators behave well even in circumstances in which they are expected to be inferior.
|