TUTORIAL 1 ________________________________________________________________________ THEORY IN SUMMARY Sufficiency Let X ( X 1 , X 2 ,..., X n ) denote a random sample of size n from a distribution with ~ probability density function (pdf) f ( x | ) , and F denote the family of pdf's ~ ~ defined as follows. F { f ( x | ) : } . Then the statistic T ( X ) T ( X 1 , X 2 ,..., X n ) is ~ ~ ~ called a sufficient statistic for or for the family F if the conditional distribution of X ~ ~ given T ( X ) t is independent of for all the values of t that the conditional probability ~ ~ is defined. Comments: To prove sufficiency we often make use of the Neyman-Fisher theorem ( T(X) is ~ a sufficient statistic if and only if f ( x | ) g[T ( X ) | ] h( X ) ). ~ ~ ~ The definition of sufficiency is mainly used to prove that a statistic is NOT sufficient. Intuitively, a sufficient statistic incorporates all the necessary sample information about the unknown parameters. The data vector is always a sufficient statistic If * () is a 1-1 and onto function of and T is a sufficient statistic for , then T is a sufficient statistic for * as well. 1 Completeness Using the above notation and defining a function g : R n R , then F is a complete family if for every function g defined as before the relationship E [ g ( X )] 0, ~ ~ implies g ( X ) 0 for every X R n that the probability P( X ) is not zero. ~ ~ ~ Comments: This definition implies that the only unbiased estimator of 0 is 0 itself. For an intuitive explanation of completeness, look at Papaioannou-Ferentinos: Mathematical Statistics (in Greek, p.30) It is possible that the function g (X ) is non-zero in values where the probability P( X ) is zero. ~ Order statistics Let X ( X 1 , X 2 ,..., X n ) denote a random sample of size n from a continuous ~ distribution with probability density function (pdf) f X ( x | ) , and cumulative ~ ~ density function (cdf) FX . Then the pdf and cdf of the n-th order statistic (i.e., the maximum) Y X (n ) is given by: FY ( y) [ FX ( y)] n and f Y ( y) n[ FX ( y)] n1 f X ( y) Similarly for the pdf and cdf of the first order statistic (i.e., the minimum) Y X (1) : FY ( y) 1 [1 FX ( y)] n and f Y ( y) n[1 FX ( y)] n1 f X ( y) and finally the pdf of the k-th order statistic Y X (k ) 2 f Y ( y) n! [ FX ( y )] k 1[1 FX ( y )] nk f X ( y ) (k 1)!(n k )! Sketch of the proof: For Y X (n ) we obtain that FY (Y y) P[X ( n ) y] P[X1 y X n y] [FX ( y)] n and then we take the derivative. We use a similar argument for the minimum order statistic. 1. EXERCISES X ( X 1 , X 2 ,..., X n ) Let ~ denote a random sample of size n from a U (, ), 0 distribution. Prove that the statistic T ( X ) max{ X (1) , X ( n ) } is a ~ sufficient statistic for . Find other sufficient statistics for . 2. Show that the first order statistic of a random sample of size n from the distribution having pdf f ( x | ) exp[ ( x )], x , , zero elsewhere is a complete and sufficient statistic for . 3. Roussas: Statistical Inference (in Greek), Volume I, page 69, Exercise 1.7 (i), (iii) 4. Let X ( X 1 , X 2 ,..., X n ) denote a random sample of size n from a U(0, ), 0 ~ distribution. Find sufficient statistics for . 3