Selection of variables in exploratory factor analysis: An empirical comparison of a stepwise and traditional approach

Abstract

Stepwise variable selection, exploratory factor analysis, goodness-of-fit, varimax rotation, statistical bias, selection accuracy, pattern accuracy,

Using Machine Learning Methods to Predict Experimental High Throughput Screening Data

“…Then, we selected the 209 atomic descriptors, among the 825 originally considered, whose values provided the highest correlation with the normalized average biological activity of the library compounds in order to compare them to the 209 molecular descriptors calculated by MOE [17]. Afterwards, we carried out the stepwise variable selection procedure [18] based on the linear regression, and including both forward and backward variable selection, using the MATLAB package (Version R2008a) in order to select the best atomic and molecular descriptors to be included in the combined data set of explanatory variables.…”

“…This indicates that the differences between the predictions provided by the molecular and atomic descriptors were not significant. Then, we proceeded by selecting the "best combined variables" among the molecular and atomic descriptors using stepwise variable selection [18]. This technique combines the advantages of the forward and backward selection procedures: at each step, a single explanatory variable may be added (forward selection) or deleted (backward elimination) from the data set.…”

“…The combined data set of 139 descriptors retained by stepwise selection [18] will be considered in this section as a basis for our simulation study intended to compare the performances of the CART, NN, SVM, LDA, PLS and kNN methods in the context of HTS. At each step, each sample was divided into the training and test subsets to build the model for prediction (for more details on the sample content, see Tables 1, 2 and 3 and Section 3).…”

“…The rationale of this approach is as follows: a single variable (i.e., descriptor) should be added to, or respectively deleted from, the data set of explanatory variables if the method's sensitivity increases after its addition, or respectively decreases after its deletion. The combined data set of 139 descriptors determined via stepwise selection [18] was used as the initial data set of explanatory variables. Then, in turn, the 134 remaining molecular descriptors (209 -75 = 134) and 145 atomic descriptors (209 -64 = 145) were tested for the addition to a new "optimal set of predictors".…”

Variable Neighborhood Search Heuristics for Selecting a Subset of Variables in Principal Component Analysis

“…Not surprisingly, the general problem of variable selection in multivariate analysis has been acknowledged for more than 40 years (Beale, Kendall, & Mann, 1967) and continues to be of tremendous importance (Duarte Silva, 2001;Fueda, Iizuka, & Mori, 2009). Especially noteworthy is the development of variable selection methods for the following multivariate statistical models: (1) multiple linear regression (Brusco, Steinley, & Cradit, 2009;Furnival & Wilson, 1974;Miller, 2002), (2) principal component analysis (Jolliffe, 1972Krzanowski, 1987;Tanaka & Mori, 1997), (3) factor analysis (Hogarty, Kromrey, Ferron, & Hines, 2004;Kano & Harada, 2000), (4) discriminant analysis (McCabe, 1975;McKay & Campbell, 1982a, 1982b, and (5) cluster analysis (Brusco & Cradit, 2001;Steinley & Brusco, 2008). 707 nurse scheduling (Burke, Curtois, Post, Qu, & Veltman, 2008), and supply-chain management (Lejeune, 2006) problems.…”