The two instrumental variables employed to estimate a rate of return (the dependent variable, expressed in earnings) of schooling (the exogenous independent variable, operationalized as the highest grade of school completed) were birth date and compulsory schooling laws. In the example, this quote from Angrist and Krueger references, the hypothesized cause of an effect on earnings is schooling. Originally developed to address problems other than selection bias, instrumental variables “solve the omitted variables problem by using only part of the variability” in the hypothesized cause “– specifically, a part that is uncorrelated with the omitted variables – to estimate the relationship between” the hypothesized cause and the effect of interest ( Angrist and Krueger, 2001: 73).
In other words, the new – instrumental – variable is related to the dependent variable only indirectly, through its relationship to the independent variable. One way to do so is to introduce into the model a variable which prior knowledge suggests (1) has no independent relationship with the outcome variable of interest, (2) does have a (at minimum, associative) relationship with the independent variable, and (3) is not correlated with other (unobservable, omitted) variables that may directly effect the outcome (dependent) variable. When propensity score matching's strong assumption of unconfoundedness is not justified by the data available, steps to reduce selection bias must account for unobservables (see Caliendo and Kopeinig, 2008: 35). McDonald, in International Encyclopedia of Education (Third Edition), 2010 Instrumental Variables
0 kommentar(er)
