Although there has been growing recognition of concerns associated with endogeneity and omitted variable bias, the way in which the inclusion of control variables influences the interpretation of statistical findings has not received sufficient attention. In this paper, we discuss what we refer to as the Construct Uncertainty Principle. This principle represents a paradox, where researchers can either know the construct being examined but not fully know the unbiased effect of that construct, or where researchers can have high confidence in the parameter estimate provided in an analysis (e.g., the regression coefficient) but not fully know the construct validity of the residual predictor used in that analysis. Using a simulation based on actual correlation matrices, we show that semi-partialled constructs have an equivalent reliability less than 0.70 roughly 50% of the time, and actually falls below an equivalent reliability of 0.50 roughly 25% of the time. The problem is larger when more control variables are used. Overall, this paper demonstrates that construct validity issues need to be much more carefully considered when using multivariate analysis and the interpretation of analyses merits equivalent attention to that given to endogeneity concerns. Recommendations for helping to address the concerns revealed in this paper are provided.
