A simple regression of treatment on outcome may not recover the causal effect due to two potential forms of bias—two-way causality between treatment and outcome and omitted variables that influence both treatment and outcome. A variety of econometric and statistical methods are available for estimating causal effects when these forms of bias are present. This review discusses one such method, instrumental variables (IV) regression. To set the stage, we discuss why a randomized experiment provides an estimate of the causal effect of the treatment on the outcome of interest as measured by a quantity called the population average treatment effect (PATE). We next elaborate on why analyses of nonexperimental data may lead to asymptotically biased estimates of the PATE. After this, we discuss the conditions under which estimates from an IV regression uncover the local average treatment effect (LATE), a constituent component of the PATE. We next turn to two illustrative examples. The first, on whether pretrial incarceration affects case outcomes, illustrates the use of IV regression to address omitted variable bias. We also use this as an opportunity to elaborate on the proper interpretation of the LATE. The second example illustrates the use of IV regression to identify the causal mechanisms underlying the PATEs estimated by randomized experiments. We close by discussing other statistical and econometric methods to address endogeneity in criminological research.
Jha et al. (Fri,) studied this question.