| Abstract: | Sociologists and criminologists have become increasingly concerned with nonlinear relationships and interaction effects. For example, some recent studies suggest that the positive relationship between neighborhood disadvantage and violent crime is nonlinear with an accelerating slope, whereas other research indicates a nonlinear decelerating slope. The present paper considers the possibility that this inconsistency in findings is partially caused by lack of attention to an important methodological concern. Specifically, we argue that researchers have not been sensitive to the ways in which logarithmic transformation of the dependent variable can bias tests for nonlinearity and statistical interaction. We illustrate this point using demographic and violent crime data for urban neighborhoods, and we propose an alternative procedure to log transformation that involves the use of weighted least‐squares regression, heteroscedasticity consistent standard errors, and diagnostics for influential observations. |
