Model Specification

Though we did not present a full review of studies on the association between abuse and violent behavior, believing that the association has already been clearly established, we do wish to return to that topic and consider the impact of three potential confounding factors that we have highlighted in this book.

Control for Nonviolent Offending

One way to disentangle the association between abuse and violent offending from the association between abuse and antisociality more generally is to employ a control for nonviolent offending. There are studies where this has been done. Four studies reported predominantly significant associations between physical abuse and violent offending behavior, controlling for aggression (Benda, 2005), previous wave offending (Rebellon & van Gundy, 2005), “problem behavior syndrome” (Swinford et al., 2000), and conduct disorder (Ehrensaft et al., 2003). Benda (2005) found the same supportive pattern in an analysis of the effect of sexual abuse. When Ehrensaft et al. (2003) looked at sex abuse, the effect on violence was in the right direction but not statistically significant with the control variable.

Other findings can also be interpreted as supportive. Fagan (2005) was interested in minor and serious partner violence in early adulthood. Half of the multivariate coefficients she presents are significant, all are in the predicted direction, and the others remain substantial in almost all cases (she controlled for prevalence of prior offending). Pears and Capaldi (2001) found that physical abuse was associated with violence in a multivariate model controlling for antisocial behavior. In one paper, the pattern of findings was wholly consistent the differential etiology hypothesis. Mersky and Reynolds (2007) comment that when they followed up to see if their models for nonviolent offending (“any nonviolent”) were robust, they found that the association between maltreatment and violent offending was statistically significant but that maltreatment was no longer significantly associated with exclusively nonviolent offending. They interpret the finding the way we might: saying that the association between maltreatment and nonviolent offending was confounded due their common association with violent offending, and they see this finding as part of the “prevailing evidence” (p. 254). These studies provide fairly strong evidence in favor of the differential etiology thesis.

There are 3 other studies where associations between abuse and violence were not statistically significant when indicators of nonviolent offending were included as control variables, but we weigh these findings less heavily. Due to model overspecification, we do not see their findings as a strong contradiction to the differential etiology hypothesis. We will elaborate.

The main problem is that control variables that may overlap with violent behavior (such as prior arrest) are included in statistical models (e.g., Cusick et al., 2012; McDaniels, 1998). Pollock et al. (2006) control for prior prison stays, which would certainly be a case of overspecification4 because prison sentences are often given because of violent offending or a violent history. Others have controlled for prior wave violence, which would be very conservative and this practice changes the interpretation of the coefficients in the model. In most cases, abuse experience had probably occurred before the “prior violence” was measured so the lion’s share of the influence of abuse may have been exerted by that time. From a mathematical standpoint, any covariation between the measure of abuse and the measure of “prior violence” would not be a part of the partial coefficient estimates testing the impact of abuse on the current measure of violence. One might also think of it this way: we would not expect abuse experienced several years ago to exert much impact on the change in violent behavior in just the past year.

The strongest case against the differential etiology hypothesis was actually presented in our own analysis, designed to test it directly. Using Add Health data, we found that the significant correlation between physical abuse and violence dissipated and was no longer statistically significant in a model controlling for nonviolent offending (Savage et al., 2014).

< Prev   CONTENTS   Source   Next >