Tests for Over-identifying Restrictions in Partially Identified Linear Structural Equations

Cragg and Donald (1996) have pointed out that the asymptotic size of tests for overidentifying restrictions can be much smaller than the asymptotic nominal size when the structural equation is partially identified. This may lead to misleading inference if the critical values are obtained from a chi-square distribution. To overcome this problem we derive the exact asymptotic distribution of the Byron test statistic. This allows the calculation of asymptotic critical values and p-values corrected for possible failure of identification.