Constructing composite scores for contemporaneous behaviors: A comparison of four approaches

Abstract

Composite scores, where the results of two or more measures are combined, are commonly used in many fields, including ethology. Composite scores can simplify the analysis and interpretation of data while capturing the salient features of the underlying latent variable(s) approximated by the score. Here we outline four approaches for constructing composite scores in ethological studies: ad hoc (AH) assignment, discriminant analysis (DA), principal components analysis (PCA), and partial least squares (PLS). We give examples of each using previously published data from a study of responses of lone star ticks (Amblyomma americanum) to several deterrent phytochemicals. In most cases, researchers construct AH composite scores by subjectively assigning weights and signs to the behavioral components; unity weighting constrains weights to −1 or 1 on standardized variables. Because the weights and signs of the coefficients are subjectively assigned, AH scores may generate a spurious result. DA can be used to construct composite behavioral scores when there are clearly defined treatments or preference tests using distinct stimuli. The DA score created consists of orthogonal variables that capture the variability in the behavioral measures most closely aligned with the differences among treatment or stimuli variables. This approach assumes that subjects discriminate treatment or stimuli differences, but may not manifest clear overt behavior that they are able to do so; it reduces dimensionality, usually to a single axis, representing the underlying latent variable of interest. The PCA approach is similar to DA except that the composite score is created independently of treatment or stimuli variables. Thus, this method can be used to investigate possible relationships between a composite score and any relevant independent variable, perhaps measured asynchronously with the behaviors. PLS is a multivariate method related to DA and PCA and is also used to create latent orthogonal variables. However, these new variables are constructed to maximize correlation with one or more continuous independent variables. Creation of a composite score requires the researcher to consider not only the method used to create it, but, at an earlier stage in the research, which behaviors should be components and how best to measure them.