Fitting Problems: Evaluating Model Fit in Behavior Genetic Models

Abstract

Dissertation under the direction of Professor Joseph Lee Rodgers In behavior genetics, like many fields, researchers must decide whether their models adequately explain their data – whether their models “fit” at some satisfactory level. Well-fitting models are compelling, whereas poorly-fitting models are not (Rodgers & Rowe, 2002). Oftentimes, researchers evaluate model fit by employing “universal” rules of thumb (e.g., Hu and Bentler, 1999). However, these rules are not universal, and are – in fact – model specific (Kang et al., 2016). Accordingly, I focused on developing fit criteria emulating Hu and Bentler (1999) for classic univariate models (ACE; CE; AE) by fitting simulated twin data to correctly- and incorrectly-specified models. Ideal criteria should consistently accept correct models and reject incorrect models. Classic ACE models were indistinguishable and virtually all fit indices were non-informative because (or especially when) they are saturated models. For non-ACE models, criteria were informative. Nevertheless, every metric employed, except TLI differed markedly across models and/or conditions. Universal solutions remain elusive, but promising approaches include nested model comparisons, increasing degrees of freedom, and ruthless skepticism.