SMS scnews item created by Andrew Mathas at Thu 13 Feb 2014 1054
Expiry: 14 Feb 2014
Calendar1: 14 Feb 2014 1100-1200
CalLoc1: AGR Seminar
CalTitle1: AGR Seminar: The Harmful Effect of Preliminary Model Selection on Confidence Intervals
Auth: firstname.lastname@example.org (assumed)
The Harmful Effect of Preliminary Model Selection on Confidence Intervals
A/Prof Paul Kabaila (La Trobe University)
La Trobe University
It is very common in applied statistics to carry out a preliminary statistical (i.e. data-based) model selection by, for example, using preliminary hypothesis tests or minimizing a criterion such as the Akaike Information Criterion (AIC). This is often followed by the construction, using the same data, of a confidence interval, with nominal coverage 1 - alpha, for the parameter of interest based on the assumption that the selected model had been given to us a priori (as the true model). This assumption is false and it typically leads to confidence intervals with minimum coverage probabilities far below 1 - alpha, making them completely inadequate. In practice, a wide variety of forms of statistical model selection have been proposed, for a variety of models. It is important that all of these forms of model selection are carefully analyzed with respect to their effect on subsequently-constructed confidence intervals. As pointed out by Kabaila (2009), preliminary statistical model selection is sometimes motivated by a desire to utilize uncertain prior information. Uncertain prior information may result from previous experience with similar data sets and/or expert opinion and scientific background. This brings us to the second part of the talk: the construction of confidence intervals that utilize uncertain prior information, without the intermediate step of model selection. A branch of this type of work was initiated by the eminent statisticians Charles Stein and John Pratt. Another branch (dealing with different models, different kinds of uncertain prior information and different desiderata) is due to Paul Kabaila and PhD students Jarrod Tuck, Khageswor Giri, David Farchione, Dilshani Tissera and Waruni Abeysekera. Reference: Kabaila, P. (2009). The coverage properties of confidence regions after model selection. International Statistical Review, 77, 405-414.
Dr Andriy Olenko
If you would like to attend this seminar in our access grid room then
please book the access grid room referring
to this scnews item. Please liaise with the host institution to make any
necessary arrangements and then send an email to email@example.com
to let the CSOs know of any special requirements for the seminar.