[R-sig-ME] Using LRT with alpha = .2 in model reduction guided by PCA

[Maarten Jung]

Hi all,

Bates et al. (2015) [1] suggest an iterative reduction of degenerate
maximal models. First they use PCA to obtain the appropriate dimensionality
of the variance-covariance matrix of the random-effect structure. Given
this reference point, they suggest to drop all non-signi cant variance
components and correlation parameters from the model. But each reduction
can lead to a signi cant loss in goodness of fi t indicated by LRTs - in
which case the parameter should stay in the model.

Is it, in this PCA context, a good idea to choose an alpha level of 0.2 as
model-selection criterion as suggested by Matuschek et al. (2017) [2] to
balance the Type I error rate and power?

[1] https://arxiv.org/abs/1506.04967
[2] https://arxiv.org/abs/1511.01864


Best,
Maarten

	[[alternative HTML version deleted]]