Powerful visualization of product-attribute associations for temporal data

Abstract

Temporal Check-All-That-Apply (TCATA) extends classical Check-All-That-Apply (CATA) by adding a temporal dimension to the evaluation. Because TCATA extends CATA, an obvious visualization of product-attribute associations over time is to treat product x time combinations as individual observations and then use classical Correspondence Analysis (CA) to visualize the associations. Often the CA results and visualization emphasize the chronological features. However, this approach could lead to misinterpretations as time is not just a feature but also a confound. Because of time, all products might show convergence to, e.g., off flavor, which is produced only by a few observations that provide a relative but not an absolute peak in this attribute.Therefore, we suggest alternative CA approaches to analyze TCATA data that emphasize (Canonical CA, CanCA) or remove (Escofier’s Conditional CA, ConCA) temporal effects. Generally, CanCA was designed to analyze CA data in the presence of row and column covariates; it is related to canonical correlation analysis. When there is only one set of covariates (e.g., row), CanCA is more akin to redundancy analysis. Here, we use external row information – time and product – to emphasize the overall temporal profile applying to all products. CanCA nicely displays the main product differences within the attribute space. CanCA better emphasizes than CA the unique properties of each product over time. Escofier’s conditional CA (ConCA) removes confounding effects such as time. ConCA provides two features for TCATA: (1) effects adjusted for time and (2) more appropriate measures of strength of association that can be used with CA for better visualization.We exemplify the proposed methods by means of data from a study on orange squashes. The relevance of off flavor is (correctly) found to be largely de-emphasized compared to standard CA: CanCA shows off flavor as an average effect because of time and ConCA shows off flavor does not contribute to the overall effect. Together CanCA and ConCA facilitate a richer, more detailed, and potentially more accurate interpretation of the data. The approaches can be equally used for Temporal Dominance of Sensations (TDS) data.