an introduction to generalized linear models - GDM@FUDAN ...
an introduction to generalized linear models - GDM@FUDAN ...
an introduction to generalized linear models - GDM@FUDAN ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
|An| = det A are all positive.<br />
4. The r<strong>an</strong>k ofthe matrix A is also called the degrees offreedom ofthe<br />
quadratic form Q = y T Ay.<br />
5. Suppose Y1, ..., Yn are independent r<strong>an</strong>dom variables each with the Normal<br />
distribution N(0,σ2 ). Let Q = �n 2<br />
i=1 Yi <strong>an</strong>d let Q1, ..., Qk be quadratic<br />
forms in the Yi’s such that<br />
Q = Q1 + ... + Qk<br />
where Qi has mi degrees offreedom (i =1,... ,k). Then<br />
Q1, ..., Qk are independent r<strong>an</strong>dom variables <strong>an</strong>d<br />
Q1/σ 2 ∼ χ 2 (m1), Q2/σ 2 ∼ χ 2 (m2), ···<strong>an</strong>d Qk/σ 2 ∼ χ 2 (mk),<br />
if<strong>an</strong>d only if,<br />
m1 + m2 + ... + mk = n.<br />
This is Cochr<strong>an</strong>’s theorem; for a proof see, for example, Hogg <strong>an</strong>d Craig<br />
(1995). A similar result holds for non-central distributions; see Chapter 3<br />
ofRao (1973).<br />
6. A consequence ofCochr<strong>an</strong>’s theorem is that the difference oftwo independent<br />
r<strong>an</strong>dom variables, X 2 1 ∼ χ 2 (m) <strong>an</strong>d X 2 2 ∼ χ 2 (k), also has a chi-squared<br />
distribution<br />
provided that X 2 ≥ 0 <strong>an</strong>d m>k.<br />
1.6 Estimation<br />
1.6.1 Maximum likelihood estimation<br />
X 2 = X 2 1 − X 2 2 ∼ χ 2 (m − k)<br />
Let y =[Y1, ..., Yn] T denote a r<strong>an</strong>dom vec<strong>to</strong>r <strong>an</strong>d let the joint probability<br />
density function of the Yi’s be<br />
f(y; θ)<br />
which depends on the vec<strong>to</strong>r ofparameters θ =[θ1, ..., θp] T .<br />
The likelihood function L(θ; y) is algebraically the same as the joint<br />
probability density function f(y; θ) but the ch<strong>an</strong>ge in notation reflects a shift<br />
ofemphasis from the r<strong>an</strong>dom variables y, with θ fixed, <strong>to</strong> the parameters θ<br />
with y fixed. Since L is defined in terms ofthe r<strong>an</strong>dom vec<strong>to</strong>r y, it is itselfa<br />
r<strong>an</strong>dom variable. Let Ω denote the set ofall possible values ofthe parameter<br />
vec<strong>to</strong>r θ; Ω is called the parameter space. The maximum likelihood<br />
estima<strong>to</strong>r of θ is the value � θ which maximizes the likelihood function, that<br />
is<br />
L( � θ; y) ≥ L(θ; y) for all θ in Ω.<br />
Equivalently, � θ is the value which maximizes the log-likelihood function<br />
© 2002 by Chapm<strong>an</strong> & Hall/CRC