A Review of Statistical Methods for Determination of Relative Importance of Correlated Predictors and Identification of Drivers of Consumer Liking

Abstract

This article attempts to deliver the following message to the researchers and practitioners in the sensory field. (1) Theoretically, drivers of consumer liking is based on relative importance of explanatory variables in a linear model. The problem is complicated when the variables involve linear dependence, which is the common situation in sensory and consumer data. (2) The commonly used methodologies, e.g., conjoint analysis, preference mapping and Kano's model, have serious limitations for determination of relative importance of correlated attributes and identification of drivers of consumer liking. (3) The conventional statistics, e.g., correlation coefficient, standard regression coefficient and P values of tests for regression parameters, etc., are inadequate and invalid measures of relative importance of correlated attributes. (4) There are three state-of-the-art methods for determination of relative importance of correlated attributes. They are the Lindeman, Merenda and Gold's method, Breiman's Random Forest and Johnson's relative weight. This article also provides statistical background and almost exhaustive main references on the topic of relative importance of variables scattered in various academic journals in different fields. The information will help the sensometricians and researchers with more statistical knowledge to embrace the mainstream of the research on the topic and to pursue advanced methods for drivers of consumer liking. PRACTICAL APPLICATIONSThis article reviews some new methods for determination of relative importance of correlated explanatory variables to response variable in a regression model. The methods can be used for identification of drivers of consumer liking. The article also provides the sources of the corresponding computer packages and codes implementing the new methods. The packages and codes are freely available and easy to use. The R packages "relaimpo" for the LMG method, "randomForest" and "party" for the original and modified Breiman's Random Forest method are available at http://cran.r-project.org. The R or S-Plus code "johnson" for Johnson's relative weight is available from the online supplementary Appendix S1 of this article. Journal of Sensory Studies ISSN 0887-8250RELATIVE IMPORTANCE AND DRIVERS OF LIKING J. BI

An upper-branch Brewer–Dobson circulation index for attribution of stratospheric variability and improved ozone and temperature trend analysis

“…11 we calculate the relative importance of each regressor without (blue) and with (red) the UBDC index by decomposing R 2 (see Bi, 2012, for a comprehensive review of this technique), which depends on the order in which regressors are considered, unless the regressors are orthogonal, which is usually not the case for the 3 decades we consider here (see e.g. Chiodo et al, 2014).…”

Exploring the Drivers of Purchase Intent and Consumer Satisfaction of Chicken Eggs Using Principal Component Analysis and the Kano Model

“…Despite the fact that Kano's method does not measure attribute relative importance within a Kano category, it remains the best approach for classification in the Kano model as it maintains validity and high reliability for assessment of existing as well as nonexistent attributes and also has no technical limitations regarding the number of attributes for analysis compared with other approaches (Mikulić and Prebežac 2011;Bi 2012;Chen 2012).…”

Dynamic Importance Analysis of Components with Known Failure Contribution of Complex Systems

“…Comparison of traditional methods is shown in Table 1; although several papers [24,27] have reviewed traditional importance measures, it is seen that these papers cover various categories with different emphases. In this paper, the dynamic characteristics of the traditional methods are compared.…”

“…In addition, many researchers focus on the importance measures of -out-of-system [25,26]. Moreover, the most comprehensive research of current work in importance analysis was conducted in [27,28]. This paper aims to seek an efficient method for evaluating component importance of complex system as a multistate system.…”