Abstract: When a model for a statistical analysis is not given before the analysis, but is the result of a model search endeavor, the uncertainty about the model that is used for inference has consequences for hypothesis testing and for the construction of confidence intervals for the model parameters of interest. Ignoring this uncertainty leads to overoptimistic results, implying that computed p-values are too small and that confidence intervals are too narrow for the intended coverage.
I will explain how to use confidence distributions to obtain valid inference after model selection for the parameters of interest. Under some assumptions, uniformly most powerful post-selection confidence curves are obtained.
This is joint work with Andrea Garcia-Angulo.
Gerda Claeskens is a professor at the Research Centre for Operations Research and Statistics (ORSTAT), KU Leuven. Her research interests are Model selection and model averaging; Post-selection inference; Nonparametric regression.
LSA Statistics Seminar