(PS8-C61) Extending the Genetic Algorithm to Scale Construction: Application to Triarchic Trait Measurement

Background: The genetic algorithm (GA) is a metaheuristic optimization technique that has been used to construct short-form scales. Based on evolutionary principles, the GA selects and evaluates different combinations of items from a full-length scale to find a set that most closely approximates full-length scores using the fewest number of items. However, the GA has not yet been adapted for use in selecting items to index an external referent (e.g., a factor analytic dimension). The current study sought to adapt the GA to create new scales for indexing dimensions corresponding to the trait constructs of the triarchic model of psychopathy (Patrick et al., 2009) – boldness, meanness, and disinhibition – from a factor analysis of established indicators of these traits. The candidate item pool for these scales was the 178-item Elemental Psychopathy Assessment (EPA; Lynam et al., 2011). In addition to parameters of scale length and optimal score estimation, the GA was adapted to balance item polarity within each scale and constrain correlations across them.

Method: A latent variable model of validated scale measures of the triarchic traits was specified, and regression-estimated factor scores were extracted for each trait. Based on this analysis and prior work (e.g., Drislane & Patrick, 2017), the new Boldness and Meanness scales were expected to correlate between .20-.30, Meanness/Disinhibition .40-.50, and Boldness/Disinhibition -.05-.05. In addition to number of items and variance explained in target factor scores, the GA was adapted to balance positively and negatively worded items within each scale and maintain inter-scale correlations within the specified ranges. Participants (N=811, 56% male) were split 60/40 into training and testing samples, and the adapted GA was run on data for the training sample (n=486).

Results: From the original 178 EPA items, the GA produced a 10-item Boldness scale (5 negatively worded; α=.81), a 14-item Meanness scale (6 negatively worded, α=.81), and a 15-item Disinhibition scale (6 negatively worded, α=.80). Each scale correlated strongly with its target factor score (rs = .93, .90, and .91, respectively). Correlations among the scales were: Boldness/Meanness .30, Meanness/Disinhibition .69, and Boldness/Disinhibition .06. After removing 5 overlapping items between Meanness/Disinhibition scales, intercorrelations were .32, .42, and -.01, respectively.

Conclusions: The GA was successfully adapted to create new triarchic trait scales, referenced to factor scores from a latent variable model of these traits, using items drawn from a separate pool. The resultant brief Boldness, Meanness, and Disinhibition scales showed good internal consistency and a relative balance of positively and negatively worded items. Although the new scales were more highly correlated than expected, assigning a stronger weight to this parameter in the GA could be expected to reduce intercorrelations. A next step in this project will be to use Item Response Theory methods to further refine the scales. Overall, the GA shows promise as a data-driven approach to developing scales, and can be extended to the assessment of other constructs and further adapted to consider other important parameters (e.g., model fit indices).