Lecture 9 Two-Sample Test. Fall 2013 Prof. Yao Xie, H. Milton Stewart School of Industrial Systems & Engineering Georgia Tech

Transcription

1 Lecture 9 Two-Sample Test Fall 2013 Prof. Yao Xie, yao.xie@isye.gatech.edu H. Milton Stewart School of Industrial Systems & Engineering Georgia Tech

2 Computer exam 1 18 Histogram 14 Frequency More Bin mean std 6.02 median 90

3 Midterm 2 Cover Confidence interval One sided and two sided confidence intervals Hypothesis testing Two approaches Fixed significance level p- value Can bring a 1- page 1- sided cheat sheet Make- up lecture on Friday Nov. 8: tentatively noon- 1:20pm in the area in front of my office, Groseclose #339

4 Outline Test difference in the mean Known variance Unknown variance Test difference in sample proportion Test difference in variance

5 Motivating Example Safety of drinking water (Arizona Republic, May 27, 2001) Water sampled from 10 communities in Pheonix And 10 communities from rural Arizona Arsenic concentration (AC): determines water quality, ranges from 3 ppb to 48 ppb Is there a difference in AC between these two areas? If the difference is large enough?

6 Formulate into statistical method Answered by statistical methods Pheonix μ 1 rural Arizona μ 2 Whether or not there is a difference between in mean AC level, μ 1 and μ 2, in these two areas? Equivalent to: test whether μ 1 - μ 2 is different from 0?

7 In general: comparing two populations Comparing two population means is often the way used to prove one population is different or better than another Competing Companies / Products Treatment vs. No Treatment New method vs. Old method

8 Test difference in the mean

9 Test difference in mean, variance known Solve the following hypothesis test H 0 : µ 1 µ 2 = Δ H 1 : µ 1 µ 2 Δ Assumptions for two sample inference

10 Test statistics A reasonable estimator for μ 1 - μ 2 is Under H 0, its mean is Δ Its variance is Detection statistic σ 1 2 X 1 X 2 2 n 1 + σ 2 n 2 Z = X 1 X 2 Δ σ n 1 + σ 2 n 2

11 Detection for two sample difference For given significance level: Reject H 0 when Z > b Z = X 1 X 2 Δ σ 1 2 n 1 + σ 2 And decide threshold b for that given significance level 2 n 2

12 p-value Probability of observing sample difference even more extreme, under H 0 P( Z > z )= 1 Φ( z ) 0 0

13 Example: paint drying time

14 Solution test difference in mean drying time H 1 : µ 1 µ 2 > Δ Δ 0 H 0 : µ 1 = µ 2 H 0 : µ 1 µ 2 = Δ H 1 : µ 1 > µ 2 form test statistic Z = X 1 X 2 σ σ 2 2 n 1 n 2

15 Fixed significance level approach Reject H 0 when Calculate: α = 0.05 Z = X 1 X 2 σ σ 2 2 n 1 z 0.05 = 1.65 n 2 > z α x 1 x 2 σ n 1 + σ 2 n 2 > 1.65 Reject H 0

16 Calculate p-value Compute p- value: Value of the statistic from data p- value: P( Z > z 0 )= 1 Φ( z 0 )= 1 Φ( 2.52)= Reject H 0 since its value is less than 0.01

17 Outline Test difference in the mean Known variance Unknown variance Test difference in sample proportion Test difference in variance

18 Case 2: test difference in mean, variance unknown, true variance equal Solve the following hypothesis test Variances are equal but unknown, so we pool the samples to estimate the variance H 0 : µ 1 µ 2 = Δ H 1 : µ 1 µ 2 Δ S 2 = (n 1 1)S 2 + (n 1)S p n + n S 1 and S 2 are sample variances S p 2 (n 1 + n 2 2) σ 2 ~ χ n1 +n 2 2

19 Use the following as the test statistics X 1 X 2 Δ S p 1/ n 1 +1/ n 2 ~ t n1 +n 2 2 For the following hypothesis test H 0 : µ 1 µ 2 = Δ H 1 : µ 1 µ 2 Δ Reject H 0 when X Y (µ 1 µ 2 ) S p 1/ n 1 +1/ n 2 > t α /2 19

20 Example α = 0.05 n 1 = 10 x 1 = 28 S 1 2 = 4 n 2 = 10 x 2 = 26 S 2 2 = 5 Assume true variance equal Test Statistic: t = x-y S 1 p n1 + 1 n 2 S 2 p = S 2 1 (n 1 1) + S 2 1 (n 2 1) n 1 + n 2 2 = 4(9) + 5(9) 18 = 4.5

21 S p 2 = 4.5 Recall degrees of freedom here is n + m 2 = 18 Threshold: t 18,0.025 = t = /10 +1/10 = 2.11>2.101 Weakly reject H 0 Calculate p-value p value = P( T > 2.11)=2P(T > 2.11) = =0.0982

22 Outline Test difference in the mean Known variance Unknown variance Test difference in sample proportion Test difference in variance

23 Formulation Two binomial parameters of interests Two independent random samples are taken from 2 populations Estimation of sample proportion X ~ Bin(n 1, p 1 ), Y ~ Bin(n 2, p 2 ) ˆp 1 = X n 1, ˆp 2 = Y n 2 H 0 : p 1 = p 2 H 1 : p 1 p 2

24 Test statistics Z = p ( 1 1 p ) 1 Pooled estimate ˆp 1 ˆp 2 ( ) n 1 + p 2 1 p 2 Estimate the test statistic: ˆp 1 ˆp 2 n 2 ˆp = X 1 + X 2 n 1 + n 2 ˆp 1 ˆp " ( ) 1 n + 1 # $ n 1 2 % & '

25 Two-sided test Z = ˆp 1 ˆp 2 ( p 1 p ) 2 p ( 1 1 p ) 1 For two-sided test, ( ) n 1 + p 2 1 p 2 n 2 H 0 : p 1 = p 2 reject H 0 when H 1 : p 1 p 2 ˆp 1 ˆp ˆp 1 ˆp 2 " ( ) 1 n + 1 # $ n 1 2 % & ' > z α /2

26 Test statistics and one-sided test H 0 : p 1 = p 2 H 1 : p 1 < p 2 Reject H 0 when H 0 : p 1 = p 2 H 1 : p 1 > p 2 Reject H 0 when ˆp 1 ˆp 2 ˆp 1 ˆp 2 ˆp 1 ˆp " ( ) 1 n + 1 # $ n 1 2 % & ' < z α ˆp 1 ˆp " ( ) 1 n + 1 # $ n 1 2 % & ' > z α

27 Comparing 2 population proportions: Example A new drug is being compared to a standard using 200 clinical trials (100 patients for each group). For the new drug, 83 of 100 patients improved. For the standard, 72 of 100 improved. Is the new drug statistically superior? Standard drug X ~ Bin(100, p 1 ) New drug Y ~ Bin(100, p 2 )

28 Fixed significance level approach H : p = p H : p < p X 1 = 72, X 2 = 83 n 1 = n 2 = 100 ˆp 1 = 0.72, ˆp 2 = 0.83 z 0.05 = 1.65 ˆp 1 ˆp ˆp 1 ˆp 2 " ( ) 1 n + 1 # $ n 1 2 % & ' = < 1.65 Reject H 0

29 p-value p- value P(Z < ) = Less than α = 0.05, reject H 0 Reject H 0, with p- value

30 Outline Test difference in the mean Known variance Unknown variance Test difference in sample proportion Test difference in variance

31 Test difference in variance two independent normal populations means and variances of the two normals are unknown test whether or not two variances are the same H 0 : H 1 :

32 Test based on sample variance ratio Test statistics: ratio of two sample variances F = S S 2 Need to introduce F distribution Let W and Y be independent chi-square 1 2 a ba b random variables with u and v degrees of freedom, respectively. Then the ratio 1 2 F W 1 2 a b a b c a b d u (10-28) Y v is said to follow the F distribution with u degrees of freedom in the numerator v degrees of freedom in the denominator. It is usually abbreviated as F u,v. 1 2 a ba b 1 2 a b a b c a b d

33 F distribution A continuous distribution mean = we should reject H 0 when the statistic is large

34 Sample distribution Under H 0 the detection statistic 2 χ n1 1 F = S 2 1 S = (n 1 1)S /σ 1 / (n 1 1) 2 2 (n 2 1)S /σ 2 / (n 2 1) ( σ = σ ) 2 has are indepe F n1 1,n 2 1 d 2 χ n2 1 distribution 34

35 Form of test Null hypothesis: H 0 : Test statistic: F 0 S2 1 (10-31) S 2 2 Alternative Hypotheses H 1 : H 1 : H 1 : Rejection Criterion f 0 f 2,n 1 1,n 2 1 or f 0 f 1 2,n 1 1,n 2 1 f 0 f,n1 1,n 2 1 f 0 f 1, n1 1,n 2 1 f (x) 2 n 1 f (x) 2 n 1 f (x) 2 n 1 α /2 α /2 α α α /2, n 1 (a) 2 α /2, n 1 x 0 (b) 2 α, n 1 x α, n 1 Figure 10-6 The F distribution for the test of with critical region values for (a), (b) H 1 : 2 2 H 1 : 2 2 H 0 : 2 2, and (c) H 1 : (c) 35 x

36 Example: Semiconductor etch variability variability in oxide layer of semiconductor is a critical characteristic of the semiconductor two kind of semiconductors, sample standard deviation s 1 = 1.96 s 2 = 2.13 n 1 = n 2 = 16 α = 0.05 test: whether or not their variances are the same 36

37 1. Parameter of interest: The parameter of interest are the variances of oxide thickness 2 1 and 2 2. We will assume that oxide thickness is a normal random variable for both gas mixtures. 2. Null hypothesis: H 0 : Alternative hypothesis: H 1 : Test statistic: The test statistic is given by equation 10-31: f 0 s2 1 s Reject H 0 if : Because n 1 n 2 16 and 0.05, we will reject H 0 : if f 0 f 0.025,15, or if f 0 f 0.975,15,15 1 f 0.025,15,

38 7. Computations: Because s 2 1 (1.96) and s 2 2 (2.13) , the test statistic is f 0 s2 1 s Conclusions: Because f 0.975,15, f 0.025,15, , we cannot reject the null hypothesis H 0 : at the 0.05 level of significance. 38

39 p-value Observe test statistic more extreme than what we got Alternative Hypotheses H 1 : H 1 : H 1 : calculate using R command p <- pf(x,d1,d2) p- value 2Ρ F > f 0 ( ) or 2Ρ F < f 0 ( ) ( ) Ρ F > f 0 Ρ F < f 0 ( ), depends on f 0 fall in upper or lower tail 39

40 Back to semiconductor example computed value of the test statistic in this example is f P(F 15, ) a p- value 2(0.3785) calculate using R command 40