The results of Sections 3-7 of the main text treat the limiting random variables (g g γ) as observed and consider the problem of testing H 0 : m = 0,In this appendix, we show that under mild assumptions, our results for the limit problem (2) imply asymptotic results along sequences of models satisfying (1). We first introduce a useful invariance condition for the weight function a and then prove results concerning the asymptotic size and power of CLC tests.We previously wrote the weight functions a of CLC tests as functions of D alone, since, in the limit problem, the parameter γ is fixed and known. In this appendix, however, it is helpful to instead write a(D γ). Likewise, since the estimatorμ D used in plug-in tests may depend on γ, we will write it asμ D (D γ).
Postmultiplication-Invariant Weight FunctionsOur weak convergence assumption (1), together with the continuous mapping theorem, implies that D T → d D for D normally distributed, where we assume that D is full rank almost surely for all (θ γ) ∈ Θ × Γ . In many applications, such convergence will only hold if we choose an appropriate normalization when defining g T , which may seem like an obstacle to applying our approach. In the linear IV model, for instance, the appropriate definition for g T will depend on the strength of identification.EXAMPLE I-Weak IV (Continued): In Section 2, we assumed that the instruments were weak, with π T = c √ T, and showed that g T = √ TΩ This apparent dependence on normalization is not typically a problem, however, since many CLC tests are invariant to renormalization of (g T g T γ). In