Page 345 - Probability and Statistical Inference
P. 345

6. Sufficiency, Completeness, and Ancillarity  322

                           U, and hence by the Theorem 6.3.2, T itself is a minimal sufficient statistic for
                                                                         n
                           θ. But recall that E (X ) = (1 + n )θ and E [(n – 1) ∑ (X -X )] = θ for all
                                                       -1
                                                                      –1
                                                               θ
                                           θ
                                                                         i=1
                                                                             i
                                                                               n:1
                                             n:1
                           θ > 0. Let us denote h(t) = (1 + n ) x  – (n – 1)         for all t
                                                                     –1
                                                       –1 –1
                                                           n:1
                                            +
                                                                         –1 –1
                           ∈  T = (θ,  ∞) ×  ℜ , so that  E [h(T)] =  E [(1 +  n )   X  – (n – 1) –1
                                                       θ          θ            n:1
                                          = 0 for all θ > 0. But obviously h(t) is not identically zero
                           for all t ∈ T ≡ (θ, ∞) × ℜ . Hence, the minimal sufficient statistic T can not be
                                               +
                           complete. !
                              In the Examples 6.6.5-6.6.6, one can easily find other nontrivial real val-
                           ued functions h(t) such that E [h(T)] ≡ 0 for all θ, but h(t) is not identically
                                                    θ
                           zero. We leave these as exercises.
                              Theorem 6.6.1 Suppose that a statistic T = (T , ..., T ) is complete. Let U
                                                                     1
                                                                          k
                           = (U , ..., U ) be another statistic with U = f(T) where g: T → U is one-to one.
                               1
                                    k
                           Then, U is complete.
                              Proof For simplicity, let us pretend that T, U have continuous distributions
                           and that T, U and θ are all real valued. Let the pdf of T be denoted by g(t; θ).
                           Since f(.) is one-to-one, we can write:
                           Now, assume that U is not complete. Then, there is a function a(U) such that
                           E [a(U)] ≡ 0 but P {u : a(u) ≠ 0} > 0 for all θ ∈ Θ. From (6.6.7) we can then
                            θ
                                          θ
                           claim that E [h(T)] ≡ 0 for all θ ∈ Θ where h ≡ a f stands for the composition
                                     θ
                                                                    o
                           mapping. But, this function h(.) is non-zero with positive probability which
                           contradicts the assumed completeness property of T. !
                              Now, we state a remarkably general result (Theorem 6.6.2) in the case of
                           the exponential family of distributions. One may refer to Lehmann (1986, pp.
                           142-143) for a proof of this result.
                              Theorem 6.6.2 (Completeness of the Minimal Sufficient Statistic in
                           the Exponential Family) Suppose that X , ..., X  are iid with the common
                                                                     n
                                                               1
                           pmf or the pdf belonging to the k-parameter exponential family defined by


                           with some appropriate forms for p(x) ≥ 0, q(θ) ≥ 0, θ  and R (x), i = 1, ...,
                                                                                i
                                                                          i
                           k. Suppose that the regulatory conditions stated in (3.8.5) hold. Denote
   340   341   342   343   344   345   346   347   348   349   350