4 - (SYM 102) Optimizing Scale Construction Using the Genetic Algorithm: Operationalizing the Triarchic Model of Psychopathy

Recently, scholars have employed metaheuristic optimization techniques such as genetic algorithms (GAs) to create abbreviated scales. GAs mimic natural selection to optimize efficiency, or explain variance in the full-length score with as few items as possible. Beyond scale abbreviation, however, GAs can be adapted to optimize scale construction, a domain that is largely untapped in the field of computational psychometrics. The current study modified a GA to 1) approximate a referent using items from a separate candidate item pool and 2) consider parameters of item polarity and discriminant validity (in addition to efficiency) when evaluating each item set in the “survival of the fittest.” To illustrate these modifications, the current study specified a latent variable model based on the triarchic model of psychopathy, including latent traits of boldness, meanness, and disinhibition (Patrick et al., 2009). Factor scores for each trait were used as referents for this modified GA to create scales indexing each dimension, using the 178-item Elemental Psychopathy Assessment (EPA; Lynam et al., 2011) as a separate candidate item pool for the new scales. Participants (N=811, 56% male) were split 60/40 into training and testing samples, and the modified GA was run on the training sample (n=486) six times due to the non-deterministic nature of GAs. Items chosen at least once across the six iterations and not chosen for more than one scale were retained, resulting in a 19-item Boldness scale, a 19-item Meanness scale, and a 22-item Disinhibition scale (ordinal omegas = .90, .91, and .90, respectively). Sum scores for these new scales correlated strongly with their target factor scores (rs = .91, .88, and .88, respectively). These new EPA-triarchic scales were then entered into the original latent variable model as indicators of their respective latent traits and evinced the strongest loadings (λs ≥ .95) compared to other well-established indicators. Next, the scales will be further refined using item-response theory (IRT) methods, and the results and final scales will be reviewed at the time of presentation. Overall, the current study demonstrates how computational methods including GAs can be leveraged to balance data-driven and theoretical approaches to scale development. Implications and future directions will be discussed, including the extension of GAs and other optimization algorithms to incorporate other parameters (e.g., model fit indices) and integrate multimodal data in the operationalization of latent constructs.