This preview shows pages 1–4. Sign up to view the full content.
This preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full Document
Unformatted text preview: Binomial Distribution: Y 1 , . . . , Y n ind Bin ( n i , i ), i = n i , ( , y ) = X i log n i y i + y i log i + ( n i y i ) log(1 i ) ( ; y ) = X i log n i y i + y i log i n i + ( n i y i ) log n i i n i D * ( y ; ) = 2 X i log n i y i + y i log y i n i + ( n i y i ) log n i y i n i  log n i y i + y i log i n i + ( n i y i ) log n i i n i D * ( y ; ) = 2 X i y i log y i i + ( n i y i ) log n i y i n i i In general, for a sample from an E.D. family density D * ( y ; ) = X i 2 w i n y i ( i i ) b ( i ) b ( i ) o / = D ( y ; ) / where D ( y ; ) is the (unscaled) deviance, which is a function of the data only (it is free of ), and i = i ( y i ), i = i ( i ). 101 Asymptotic Distribution of the Deviance and X 2 Statistics: From the asymptotic chisquareness of 2 log its tempting to conclude D * ( y , ) a 2 ( n p ) This is not necessarily true! Wilks result is based on asymptotics under which n while t 1 , t 2 stay fixed. However, in some cases the saturated model requires estimation of an increasing number of parameters as n , so that standard theory does not apply. Therefore, we need to be careful in using the chisquare approximation to the distribution of the deviance, and in using the deviance as a true goodnessoffit statistic. Similar comments apply to Pearsons generalized X 2 statistic: X 2 ( y ; ) = X i ( y i i ) 2 v ( i ) /w i Under some grouped data situations, both X 2 and D * are asymp totically chisquare. We assume here that data have been grouped as far as possible and that X 2 and D * are computed as sums of independent contributions over g groups: D * ( y ; ) = 2 g X i =1 { i ( y i ; y i ) i ( i ; y i ) } X 2 ( y ; ) = g X i =1 ( y i i ) 2 v ( i ) /w i 102 Fixedcells Asymptotics: Classical assumptions in asymptotic theory for grouped data imply i. a fixed number of groups ii. increasing sample sizes n i , i = 1 , . . . , g , such that n i /n i where i > 0, i = 1 , . . . , g , are fixed proportions iii. a fixed number of cells k in each group (think of a discrete outcome variable in each group with k possible values) iv. a fixed number of parameters being estimated in the current model Under these assumptions and certain regularity conditions, then both X 2 and D * are asymptotically chisquare under the null hypothesis that the current model holds with X 2 , D * a 2 ( t 1 p ) where t 1 = g ( k 1) = the number of independent parameters estimated under the full model This result holds for a more general class of goodnessoffit statis tic, the powerdivergence family , within which X 2 and D * are special cases (see F&T, sec. 3.4)....
View
Full
Document
This note was uploaded on 11/13/2011 for the course STAT 8620 taught by Professor Hall during the Fall '11 term at University of Florida.
 Fall '11
 Hall
 Binomial

Click to edit the document details