Abstract
Assuming without loss of generality that conditional expectation is the conditional attribute of interest, we show why the properties of nonexperimental data give rise to a different interpretation of the error term in stochastic mean regression and in turn allow possible misspecification for conditional expectation.