Analysis of variance for identifying temporal drivers of liking

- We use analysis of variance to reveal key drivers of temporal liking.
- A two-way model identifies positive and negative drivers and if they are global or product-specific.
- Imputing missing values and weighting the liking data by duration of dominance was studied.
- We used Tukey's one-degree-of-freedom for non-additivity in unreplicated data sets.

The data analysis workshop at the 2014 Sensometrics meeting involved the extension of temporal dominance of sensation (TDS) to temporal liking, using data collected on six fresh cheeses. One objective of the workshop was to explore data analysis techniques that would reveal which TDS attributes drove temporal liking. A simple two-way analysis of variance with interaction was performed to address the objective. TDS data were collected from a group of consumers and from a trained TDS panel. The consumers also rated their liking of the six cheeses continuously over time during consumption. The consumer TDS and temporal liking data were merged and average liking while dominant (LWD) ratings were calculated for each consumer, product and dominant attribute. The trained panel's average TDS curves were merged with the consumers' average temporal liking ratings and average LWD values were calculated for each product and attribute. For both data sets, the LWD values were arranged into a two-way table of products-by-attributes. Not all attributes were dominant for all products, so the two-way tables have many missing values. Three analyses were performed. The first was conducted on the product-by-attribute data with no changes or adjustments. The second was conducted on the product-by-attribute data after the missing values had been imputed. The third was conducted on the product-by-attribute data that had been weighted based on the amount of time a particular attribute was dominant during consumption. The results of the analyses revealed that garlic and fresh herbs were positive drivers of temporal liking, cooked herbs and cream were negative drivers. The weighted analyses appear to be slightly more discriminating than the other two approaches. Also, the consumer TDS data were more discriminating than the trained panel TDS data.