Specification Bias in Causal Models with Fallible Indicators

Abstract

In the bivariate case, measurement error in the independent variable produces an attenuated estimate of the true regression coefficient. In the multivariate case, the bias which results from specifying, incorrectly, a model with no measurement error will produce biased estimates which are predictable in neither their direction nor magnitude. This paper demonstrates some of these biases in a causal model of educational attainment.