STAT 3690 Lecture 09
Hello, dear friend, you can consult us at any time if you have any questions, add WeChat: daixieit
STAT 3690 Lecture 09
2022
Sampling distributions of and s (J&W Sec 4.4)
* Sample variance S2 = (n - 1)- 1
- ,n( - µ)/σ ~ N (0, 1)
- (n - 1)S2 /σ 2 ~ χ2 (n - 1)
- ,n( - µ)/S ~ t(n - 1)
(Xi - )2
- x1 , . . . , xn MVNp (u, О), n > p
- S ll , i.e., ML ll ML
- ,nО - 1/2( - u) ~ MVNp (一, A)
- (n - 1)S = n ML ~ Wp (n - 1, О)
- n( - u)T S- 1 ( - u) ~ Hotelling’s T2 (p, n - 1)
- Def: Wp ( О, n) is the distribution of Yi Yi(T) with Y1 , . . . , Yn MVNp (一, О)
* A generalization of χ2 -distribution: Wp ( О, n) = χ2 (n) if p = О = 1
- Propoties
* 〇〇T > 0 and w ~ Wp ( О, n) ÷ 〇w〇T ~ Wp (〇О〇T , n)
* wi Wp ( О, ni ) ÷ w1 + w2 ~ Wp ( О, n1 + n2 )
* w1 ll w2 , w1 + w2 ~ Wp ( О, n) and w1 ~ Wp ( О, n1 ) ÷ w2 ~ Wp ( О, n - n1 )
* w ~ Wp ( О, n) and a e Rp ÷
~ χ2 (n)
* w ~ Wp ( О, n), a e Rp and n > p ÷
a(a)T(T) ~ χ2 (n - p + 1)
* w ~ Wp ( О, n) ÷
tr(О - 1 w) ~ χ2 (np)
● Hotelling’s T2 distribution
- A generalization of (Student’s) t-distribution
- If x ~ MVNp (一, A) and w ~ Wp (A, n), then
xT w- 1 x ~ T2 (p, n)
- Y ~ T2 (p, n) 兮 Y ~ F (p, n - p + 1)
● Wilk’s lambda distribution
- Wilks’s lambda is to Hotelling’s T2 as F distribution is to Student’s t in univariate statistics.
- Given independent w1 ~ Wp ( О, n1 ) and w2 ~ Wp ( О, n2 ) with n1 > p,
Λ = ~ Λ(p, n1 , n2 )
- Resort to approximations for computation: {(p - n2 + 1)/2 - n1 } ln Λ(p, n1 , n2 ) s χ2 (n2p)
Hypothesis testing
● Model: x ~ f9 * e {f9 : 9 e 入}
- 9 * : parameters of interest, fixed and unknown
- 入: the parameter space
● Hypotheses H0 : 9* e 入0 v.s. H1 : 9* e 入1
- 入0 n 入1 = 0
- 入0 u 入1 = 入
● Rejection/critical region R
- Reject H0 if x e R
● Level α : sup9∈o0 β(9 ) s α
- Power function: β(9) = Pr9 (x e R)
- When 9 * e 入0 , Pr(type I error) = β(9 * ) s sup9∈o0 β(9 ) s α
* Type I error: H0 is incorrectly rejected
- When 9 * e 入1 , Pr(type II error) = 1 - β(9 * )
* Type II error: H0 is incorrectly accepted
● p-value: alternative to rejection region
- Impossible to be well-defined in some cases
- p = p(z) is defined such that sup9∈o0 Pr9 {p(z) e [0, α)} s α for all α e [0, 1]
* R = {z : p(z) e [0, α)}
● Necessary components in reporting a testing result
1. Hypotheses
2. Name of approach
3. Value of test statistic
4. Rejection region/p-value
5. Conclusion: e.g., at the α level, we reject/do not reject H0 , i.e., we believe. . .
Likelihood ratio test (LRT)
● Minimize the type II error rate subject to a capped type I error rate (under certain classical circumstances)
● Test statistic
- 0 : ML estimator for 9 e 入0
- : ML estimator for 9 e 入
● Rejection region R = {z : λ(z) s c}
- z is the realization of x
- c e R is chosen such that
sup Pr(λ(x) s c) = α.
9∈o0 9
* Have to know the null distribution of λ(x), i.e., the distribution of λ(x) with 9 e 入0
● p-value
p(z) = sup Pr{λ(x) s λ(z)}
9∈o0 9
● Null distribution of λ(x)
- Use the accurate distribution of λ(x) if it is known; otherwise see below for an approximation.
- As n → o,
-2 ln λ(x) ~ χ2 (ν)
* ν: the difference in numbers of free parameters between H0 and H1
* Leading to an (asymptotic) rejection region {z : -2 ln λ(z) > χ1 -α } ' χ1 -α is the (1 - α)- quantile of χ2 (ν).
2022-03-17