The problem of model overspecification4 merits a separate look in this chapter. In the present example, many studies include multiple indicators of abuse in the same model (e.g., Cusick et al., 2012; Ehrensaft et al., 2003; Fergusson & Lynskey, 1997; Haapasalo & Moilanen, 2004; Herrera & McCloskey, 2003; Maschi, 2006; McDaniels, 1998; Pollock et al., 2006; Sousa et al., 2011; Zingraff et al., 1993). While this type of modeling may be important for disentangling effects of sexual abuse from physical abuse or neglect, for example, to test the authors’ own research question, we believe that this practice is problematic for testing our own research question. Many victims experience multiple forms of abuse or trauma, so holding physical abuse constant, for example, in a model looking at the association between sexual abuse victimization and violent behavior narrows the interpretation of the coefficient and reduces statistical power. Imagine that physical abuse is strongly associated with violence and explains 20% of the variation in violent offending. Imagine that sexual abuse is also associated with violence and explains 20% of the variation in violent offending. If participants in the study who were sexually abused were also physically abused in almost all cases (for the purpose of the example), then the partial coefficient, reflecting the unique association between sexual abuse and violent offending may be near zero, and the magnitude of the estimated association between physical abuse and violent offending would also be reduced. Even though, in combination, the two variables may add substantially to the predictive ability of the overall model, the reported partial coefficients may not even imply this. There are solutions for assessing the problem, but these solutions are rarely applied in the published works in this area.
Controlling for prior violence is another example of overspecification. Although we have cited studies controlling for prior violence above as interpretable as controlling for child effects, it is also important to point out that controlling for prior offending or prior violence also controls out any effects that abuse already exerted on earlier offending. Many studies included controls of this kind (e.g., McDaniels, 1998; McNulty & Bellair, 2003; Rebellon, 2002; Rebellon & van Gundy, 2005; Teague, Mazerolle, Legosz, & Sanderson, 2008). Other studies control for analogous constructs such as prior prison stays (Pollock et al., 2006), “aggressive coping” (Spaccarelli et al., 1995), incarceration for violent offense and prior incarceration (Steiner & Wooldredge, 2009), and adolescent and adult “problem behavior syndrome” (Swinford et al., 2000). In these studies, it is unclear whether the timing of the abuse was taken into account. This type of control could be seen as wholly inappropriate if the abuse occurred before the measure of prior wave violence, and a downwardly biased effect on later violence would be nearly guaranteed. The best models might control for an early childhood indicator of antisociality, but by the time we can measure antisocial tendencies, it is possible that abuse has caused them.