Criticality of predictors in multiple regression
Razia Azen*
Department of Educational Psychology, University of Wisconsin-Milwaukee, USA
David V. Budescu
Department of Psychology, University of Illinois at Urbana-Champaign, USA
Benjamin Reiser
Department of Statistics, University of Haifa, Israel
A new method is proposed for comparing all predictors in a multiple regression
model. This method generates a measure of predictor criticality, which is distinct
from and has several advantages over traditional indices of predictor importance.
Using the bootstrapping (resampling with replacement) procedure, a large
number of samples are obtained from a given data set which contains one response
variable and p predictors. For each sample, all 2
p
2
1 subset regression models are
®tted and the best subset model is selected. Thus, the (multinomial) distribution
of the probability that each of the 2
p
2
1 subsets is `the best’ model for the data set
is obtained.
A predictor’s criticality is de®ned as a function of the probabilities associated
with the models that include the predictor. That is, a predictor which is included in
a large number of probable models is critical to the identi®cation of the best-®tting
regression model and, therefore, to the prediction of the response variable.
The procedure can be applied to ®xed and random regression models and can use
any measure of goodness of ®t (e.g., adjusted R
2
, C
p
, AIC) for identifying the best
model. Several criticality measures can be de®ned by using different combinations
of the probabilities of the best-®tting models, and asymptotic con®dence intervals
for each variable’s criticality can be derived. The procedure is illustrated with
several examples.
1. Introduction
Multiple regression (MR) models are used to predict a single criterion variable from several
predictors. Consider the MR model
Y 5 Xb 1 «,
where Y is an n 3 1 data vector (the criterion), X is an n 3 ( p 1 1) full-rank data matrix
(the predictors); « is an n 3 1 vector of unobservable `error’ terms and b is a ( p 1 1) 3 1
vector of parameters, estimated from the data by
Ã
b
5 (X
9
X)
21
X
9
Y. Here p represents the
British Journal of Mathematical and Statistical Psychology (2001), 54, 201±225 Printed in Great Britain
© 2001 The British Psychological Society
201
* Requests for reprints should be addressed to Dr Razia Azen, Department of Educational Psychology, University of
Wisconsin-Milwaukee, PO Box 413, Milwaukee, WI 53201, USA.