Incorporating biometric data in models of consumer choice

ABSTRACT

The use of neuro-physiological data in models of consumer choice is gaining popularity. This article presents some of the benefits of using psycho-physiological data in analyzing consumer valuation and choice. Eye-tracking, facial expressions, and electroencephalography (EEG) data were used to construct three non-conventional choice models, namely, eye-tracking, emotion and brain model. The predictive performance of the non-conventional models was compared to a baseline model, which was based entirely on conventional data. While the emotion and brain models proved to be as good as conventional data in explaining and predicting consumer choice, the eye-tracking model generated superior predictions. Moreover, we document a significant increase in predictive power when biometric data from different sources were combined into a mixed model. Finally, we utilize a machine learning technique to sparse the data and enhance out-of-sample prediction, thus showcasing the compatibility of biometric data with well-established statistical and econometric methods.

KEYWORDS:

• Biometric data
• choice
• Lasso
• predictive power

JEL CLASSIFICATION:

• C25
• D03

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ Fehr and Rangel (2011) provide an extensive review of the foundation and recent developments in Neuroeconomics as a tool to analyze consumer choice.

² The FFT transforms the EEG signal into the frequency domain by examining how similar the raw EEG is to sine waves consisting of certain pure frequencies.

³ In line with the relevant literature, the following definitions have been adopted for the article: True Positive occurs when the model predicts purchase and the subject purchases the product; True Negative occurs when the model predicts non-purchase and the subject does not purchase the product; False Positive occurs when the model predicts purchase but the subject does not purchase the product; False Negative occurs when the model predicts non-purchase and the subject purchases the product.

⁴ The cutoff values and the corresponding tpr and fpr are made-up for illustration purposes to explain the logic behind the ROC analysis.

⁵ Note that if a ROC Curve goes below the 45 degree line, then it means that the model predicts the non-purchase situation better than the purchase situation. In that case, one can flip the figure (or reassign the non-purchase event as the event of interest) and re-conduct the ROC Curve analysis.

⁶ Generally, this formula can be expressed with a kernel (DeLong, DeLong, and Clarke-Pearson 1988):

$\mathrm{\Psi }(\beta )=\frac{1}{mn}{\sum }_{i\in m}{\sum }_{j\in n}\mathrm{\Gamma }\left({\beta }^{\prime }{X}_i>{\beta }^{\prime }{X}_j\right)$ where $\mathrm{\Gamma }\left({\beta }^{\prime }{X}_i>{\beta }^{\prime }{X}_j\right)=\left\{\begin{array}{cc}1& {\beta }^{\prime }{X}_i>{\beta }^{\prime }{X}_j\\ \frac{1}{2}& {\beta }^{\prime }{X}_i={\beta }^{\prime }{X}_j\\ 0& {\beta }^{\prime }{X}_i<{\beta }^{\prime }{X}_j\end{array}\right.$

Then, $\frac{1}{2}E(\mathrm{\Psi }(\beta ))=Pr({\beta }^{\prime }{X}_i>{\beta }^{\prime }{X}_j)+\frac{1}{2}Pr({\beta }^{\prime }{X}_i={\beta }^{\prime }{X}_j)$. Since for a continuous distribution $Pr(\beta X_i^{\prime }={\beta }^{\prime }{X}_j)=0$, then we end up with the formula in Equation (5).

⁷ We also conducted this process using 1000 different combinations of training and test sets and the results were not different.

⁸ We employed R software and relevant packages in our empirical analyses (Team 2014; Robin et al. 2011; Hlavac 2014; Zeileis and Hothorn 2002).