Evoked emotions predict food choice

Abstract

In the current study we show that non-verbal food-evoked emotion scores significantly improve food choice prediction over merely liking scores. Previous research has shown that liking measures correlate with choice. However, liking is no strong predictor for food choice in real life environments. Therefore, the focus within recent studies shifted towards using emotion-profiling methods that successfully can discriminate between products that are equally liked. However, it is unclear how well scores from emotion-profiling methods predict actual food choice and/or consumption. To test this, we proposed to decompose emotion scores into valence and arousal scores using Principal Component Analysis (PCA) and apply Multinomial Logit Models (MLM) to estimate food choice using liking, valence, and arousal as possible predictors. For this analysis, we used an existing data set comprised of liking and food-evoked emotions scores from 123 participants, who rated 7 unlabeled breakfast drinks. Liking scores were measured using a 100-mm visual analogue scale, while food-evoked emotions were measured using 2 existing emotion-profiling methods: a verbal and a non-verbal method (EsSense Profile and PrEmo, respectively). After 7 days, participants were asked to choose 1 breakfast drink from the experiment to consume during breakfast in a simulated restaurant environment. Cross validation showed that we were able to correctly predict individualized food choice (1 out of 7 products) for over 50% of the participants. This number increased to nearly 80% when looking at the top 2 candidates. Model comparisons showed that evoked emotions better predict food choice than perceived liking alone. However, the strongest predictive strength was achieved by the combination of evoked emotions and liking. Furthermore we showed that non-verbal food-evoked emotion scores more accurately predict food choice than verbal food-evoked emotions scores.

Figure 1. Schematic overview of the Leave One Out Cross Validation (LOOCV).

This schematic overview shows how the LOOCV was implemented for unbiased prediction of individualized product choices. The complete data set is indicated in blue. Data operations with the independent individual are given in green, while data operations with the data from all remaining individuals are given in orange. The data operations within the plane that is bordered in black were repeated for every individual.

Figure 2. PCA biplots of EsSense and PrEmo emotion data.

The figure shows a biplot a PCA performed on the EsSense and PrEmo emotions. Every data point represents all rated emotions per method that were rated by a participant on a single product. The data points are colored based on empirical choice; chosen products are colored green and not chosen products are colored red. In blue we plotted the loadings of all emotion variables.

Figure 3. Correlations between the independent variables.

The figure shows the Pearson correlations between the independent variables that were used in the analysis. The figure indicates high correlations between liking and the valence components from the emotion measurement methods. Furthermore, there is a weak correlation between the arousal components of the emotion measurement methods.

Figure 4. Result of the LOOCV predictions.

This figure shows the prediction outcomes of the LOOCV using different subsets of the predictors Liking, Premo PC1 and Essense PC1. For the prediction, the multinomial logit model was used to create a distribution of p-values that represented the chances for every product to be chosen by each individual. These p-values were transformed in ranks; the product with the highest predicted chance of being chosen received rank 1 and the product with the lowest predicted chance of being chosen received rank 7. In the figure we show the percentage of final product choices per predicted rank (e.g. product choice for 54% of the participants, was (correctly) predicted as rank 1 by the model when using Liking & Premo PC1 as predictors). The dashed line indicates how the model would perform on chance level (14.3%). Note that an improvement in prediction performance would reflect a distribution change from right to left in the plot.