Goodness of Fit Goodness of fit - 2 classes

Transcription

1 Goodness of Fit Goodness of fit - 2 classes A B Do these data correspond reasonably to the proportions 3:1? We previously discussed options for testing p A = 0.75! Exact p-value Exact confidence interval Normal approximation

2 Goodness of fit - 3 classes AA AB BB Do these data correspond reasonably to the proportions 1:2:1? Multinomial distribution Imagine an urn with k types of balls. Let p i denote the proportion of type i. Draw n balls with replacement. Outcome: (n 1, n 2,..., n k ), with i n i = n, where n i is the no. balls drawn that were of type i. n! P(X 1 =n 1,..., X k =n k )= n 1! n k! pn 1 1 pn k k if 0 n i n, i n i =n Otherwise P(X 1 =n 1,..., X k =n k ) = 0.

3 Example Let (p 1, p 2, p 3 ) = (0.25, 0.50, 0.25) and n = 100. P(X 1 =35, X 2 =43, X 3 =22) = 100! 35! 43! 22! Rather brutal, numerically speaking. Take logs (and use a computer). Goodness of fit test We observe (n 1, n 2, n 3 ) Multinomial(n,p={p 1, p 2, p 3 }). We seek to test H 0 : p 1 = 0.25, p 2 = 0.5, p 3 = versus H a : H 0 is false. We need two things: A test statistic. The null distribution of the test statistic.

4 The likelihood-ratio test (LRT) Back to the first example: A n A B n B Test H 0 :(p A, p B ) = (π A, π B ) versus H a :(p A, p B ) (π A, π B ). MLE under H a : ˆp A = n A /n where n = n A + n B. Likelihood under H a : Likelihood under H 0 : L a = Pr(n A p A = ˆp A )= ( ) n n A ˆp n A A (1 ˆp A ) n n A L 0 = Pr(n A p A = π A )= ( ) n n A π n A A (1 π A) n n A Likelihood ratio test statistic: LRT = 2 ln (L a /L 0 ) Some clever people have shown that if H 0 is true, then LRT follows a χ 2 (df=1) distribution (approximately). Likelihood-ratio test for the example We observed n A = 78 and n B = 22. H 0 :(p A, p B )=(0.75,0.25) H a :(p A, p B ) (0.75,0.25) L a = Pr(n A =78 p A =0.78) = ( ) = L 0 = Pr(n A =78 p A =0.75) = ( ) = LRT = 2 ln (L a /L 0 ) = Using a χ 2 (df=1) distribution, we get a p-value of We therefore have no evidence against the null hypothesis.

5 Null distribution Obs LRT A little math... n=n A +n B, n 0 A = E[n A H 0 ] = n π A, n 0 B = E[n B H 0 ] = n π B. ( ) na ( Then L a /L 0 = na n nb 0 n A 0 B ) nb ) Or equivalently LRT = 2 n A ln( na n 0 A ) +2 n B ln( nb. n 0 B Why do this?

6 Generalization to more than two groups If we have k groups, then the likelihood ratio test statistic is LRT = 2 k i=1 n i ln ( ni n 0 i ) If H 0 is true, LRT χ 2 (df=k-1) The chi-square test There is an alternative technique. The test is called the chi-square test, and has the greater tradition in the literature. For two groups, calculate the following: X 2 = (n A n 0 A) 2 n 0 A + (n B n 0 B) 2 n 0 B If H 0 is true, then X 2 is a draw from a χ 2 (df=1) distribution (approximately).

7 Example In the first example we observed n A = 78 and n B = 22. Under the null hypothesis we have n 0 A = 75 and n 0 B = 25. We therefore get X 2 = (78-75) (22-25)2 25 = = This corresponds to a p-value of We therefore have no evidence against the hypothesis (p A, p B )=(0.75,0.25). Note: using the likelihood ratio test we got a p-value of Generalization to more than two groups As with the likelihood ratio test, there is a generalization to more than just two groups. If we have k groups, the chi-square test statistic we use is X 2 = k (n i n 0 i ) 2 i=1 χ 2 (df=k-1) n 0 i

8 Test statistics Let n 0 i denote the expected count in group i if H 0 is true. LRT statistic { } Pr(data p = MLE) LRT = 2 ln =... = 2 i Pr(data H 0 ) n i ln(n i /n 0 i ) χ 2 test statistic X 2 = (observed expected) 2 expected = i (n i n 0 i ) 2 n 0 i Null distribution of test statistic What values of LRT (or X 2 ) should we expect, if H 0 were true? The null distributions of these statistics may be obtained by: Brute-force analytic calculations Computer simulations Asymptotic approximations If the sample size n is large, we have LRT χ 2 (k 1) and X 2 χ 2 (k 1)

9 Recommendation For either the LRT or the χ 2 test: The null distribution is approximately χ 2 (k 1) if the sample size is large. The null distribution can be approximated by simulating data under the null hypothesis. If the sample size is sufficiently large that the expected count in each cell is 5, use the asymptotic approximation without worries. Otherwise, consider using computer simulations. Composite hypotheses Sometimes, we ask not p AA = 0.25, p AB = 0.5, p BB = 0.25 But rather something like: p AA = f 2, p AB = 2f(1 f), p BB =(1 f) 2 for some f. For example: Consider the genotypes, of a random sample of individuals, at a diallelic locus. Is the locus in Hardy-Weinberg equilibrium (as expected in the case of random mating)? Example data: AA AB BB

10 Another example ABO blood groups 3 alleles A, B, O. Phenotype A genotype AA or AO B genotype BB or BO AB genotype AB O genotype O Allele frequencies: f A, f B, f O (Note that f A + f B + f O = 1) Under Hardy-Weinberg equilibrium, we expect p A = f 2 A + 2f A f O p B = f 2 B + 2f B f O p AB = 2f A f B p O = f 2 O Example data: O A B AB LRT for example 1 Data: (n AA, n AB, n BB ) Multinomial(n,{p AA, p AB, p BB }) We seek to test whether the data conform reasonably to H 0 : p AA = f 2, p AB = 2f(1 f), p BB =(1 f) 2 for some f. General MLEs: ˆp AA = n AA /n, ˆp AB = n AB /n, ˆp BB = n BB /n MLE under H 0 : ˆf =(naa + n AB /2)/n p AA = ˆf 2, p AB = 2ˆf (1 ˆf), p BB =(1 ˆf) 2 LRT statistic: LRT = 2 ln { } Pr(nAA, n AB, n BB ˆp AA, ˆp AB, ˆp BB ) Pr(n AA, n AB, n BB p AA, p AB, p BB )

11 LRT for example 2 Data: (n O, n A, n B, n AB ) Multinomial(n,{p O, p A, p B, p AB }) We seek to test whether the data conform reasonably to H 0 : p A = f 2 A + 2f A f O,p B = f 2 B + 2f B f O,p AB = 2f A f B,p O = f 2 O for some f O, f A, f B, where f O + f A + f B = 1. General MLEs: ˆp O, ˆp A, ˆp B, ˆp AB, like before. MLE under H 0 : Requires numerical optimization Call them (ˆf O,ˆf A,ˆf B ) ( p O, p A, p B, p AB ) LRT statistic: LRT = 2 ln { } Pr(nO, n A, n B, n AB ˆp O, ˆp A, ˆp B, ˆp AB ) Pr(n O, n A, n B, n AB p O, p A, p B, p AB ) χ 2 test for these examples Obtain the MLE(s) under H 0. Calculate the corresponding cell probabilities. Turn these into (estimated) expected counts under H 0. Calculate X 2 = (observed expected) 2 expected

12 Null distribution for these cases Computer simulation (with one wrinkle) Simulate data under H 0 (plug in the MLEs for the observed data) Calculate the MLE with the simulated data Calculate the test statistic with the simulated data Repeat many times Asymptotic approximation Under H 0, if the sample size, n, is large, both the LRT statistic and the χ 2 statistic follow, approximately, a χ 2 distribution with k s 1 degrees of freedom, where s is the number of parameters estimated under H 0. Note that s = 1 for example 1, and s = 2 for example 2, and so df = 1 for both examples. Example 1 Example data: AA AB BB MLE: ˆf = (5 + 20/2) / 100 = 15% Expected counts: Test statistics: LRT statistic = 3.87 X 2 = 4.65 Asymptotic χ 2 (df = 1) approx n: P 4.9% P 3.1% 10,000 computer simulations: P 8.2% P 2.4%

13 Example 1 Est d null dist n of LRT statistic Observed 95th %ile = LRT Est d null dist n of chi!square statistic 95th %ile = 3.36 Observed X 2 Example 2 Example data: O A B AB MLE: ˆfO 62.8%, ˆf A 25.0%, ˆf B 12.2%. Expected counts: Test statistics: LRT statistic = 1.99 X 2 = 2.10 Asymptotic χ 2 (df = 1) approx n: P 16% P 15% 10,000 computer simulations: P 17% P 15%

14 Example 2 Est d null dist n of LRT statistic Observed 95th %ile = LRT Est d null dist n of chi!square statistic Observed 95th %ile = X 2 Example 3 Data on number of sperm bound to an egg: count Do these follow a Poisson distribution? MLE: ˆλ = sample average = ( ) / Expected counts n 0 i = n e ˆλ ˆλi / i!

15 Example observed expected X 2 = (obs exp) 2 exp =... = 42.8 LRT = 2 obs log(obs/exp) =... = 18.8 Compare to χ 2 (df = = 4) P-value = (χ 2 ) and (LRT). By simulation: p-value = 16/10,000 (χ 2 ) and 7/10,000 (LRT) Null simulation results Observed Simulated! 2 statistic Observed Simulated LRT statistic

16 A final note With these sorts of goodness-of-fit tests, we are often happy when our model does fit. In other words, we often prefer to fail to reject H 0. Such a conclusion, that the data fit the model reasonably well, should be phrased and considered with caution. We should think: how much power do I have to detect, with these limited data, a reasonable deviation from H 0? Contingency Tables

17 2 x 2 tables Apply a treatment A or B to 20 subjects each, and observe the reponse. Question: N Y A B Are the response rates for the two treatments the same? Sample 100 subjects and determine whether they are infected with viruses A and B. I-B NI-B I-A NI-A Question: Is infection with virus A independent of infection with virus B? Underlying probabilities Observed data B 0 1 A 0 n 00 n 01 n 0+ 1 n 10 n 11 n 1+ n +0 n +1 n Underlying probabilities B 0 1 A 0 p 00 p 01 p 0+ 1 p 10 p 11 p 1+ p +0 p +1 1 Model: (n 00, n 01, n 10, n 11 ) Multinomial(n,{p 00, p 01, p 10, p 11 }) or n 01 Binomial(n 0+, p 01 /p 0+ ) and n 11 Binomial(n 1+, p 11 /p 1+ )

18 Conditional probabilities Underlying probabilities B 0 1 A 0 p 00 p 01 p 0+ 1 p 10 p 11 p 1+ p +0 p +1 1 Conditional probabilities Pr(B = 1 A = 0) = p 01 /p 0+ Pr(B = 1 A = 1) = p 11 /p 1+ Pr(A = 1 B = 0) = p 10 /p +0 Pr(A = 1 B = 1) = p 11 /p +1 The questions in the two examples are the same! They both concern: p 01 /p 0+ = p 11 /p 1+ Equivalently: p ij = p i+ p +j for all i,j think Pr(A and B) = Pr(A) Pr(B). This is a composite hypothesis! 2 x 2 table A different view B 0 1 A 0 p 00 p 01 p 0+ p 00 p 01 p 10 p 11 1 p 10 p 11 p 1+ p +0 p +1 1 H 0 : p ij = p i+ p +j for all i,j H 0 : p ij = p i+ p +j for all i,j Degrees of freedom = = 1

19 Expected counts Observed data B 0 1 A 0 n 00 n 01 n 0+ 1 n 10 n 11 n 1+ n +0 n +1 n Expected counts B 0 1 A 0 e 00 e 01 n 0+ 1 e 10 e 11 n 1+ n +0 n +1 n To get the expected counts under the null hypothesis we: Estimate p 1+ and p +1 by n 1+ /n and n +1 /n, respectively. These are the MLEs under H 0! Turn these into estimates of the p ij. Multiply these by the total sample size, n. The expected counts The expected count (assuming H 0 ) for the 11 cell is the following: e 11 = n ˆp 11 = n ˆp 1+ ˆp +1 = n (n 1+ /n) (n +1 /n) =(n 1+ n +1 )/n The other cells are similar. We can then calculate a χ 2 or LRT statistic as before!

20 Example 1 Observed data N Y A B Expected counts N Y A B X 2 = ( ) ( ) (2 5.5) (9 5.5)2 5.5 = 6.14 LRT = 2 [18 log( ) log( 5.5 )] = 6.52 P-values (based on the asymptotic χ 2 (df = 1) approximation): 1.3% and 1.1%, respectively. Example 2 Observed data I-B NI-B I-A NI-A Expected counts I-B NI-B I-A NI-A X 2 = (9 5.2) ( ) (9 12.8) ( ) = 4.70 LRT = 2 [9 log( ) log( 58.2 )] = 4.37 P-values (based on the asymptotic χ 2 (df = 1) approximation): 3.0% and 3.7%, respectively.

21 Fisher s exact test Observed data N Y A B Assume the null hypothesis (independence) is true. Constrain the marginal counts to be as observed. What s the chance of getting this exact table? What s the chance of getting a table at least as extreme? Hypergeometric distribution Imagine an urn with K white balls and N K black balls. Draw n balls without replacement. Let x be the number of white balls in the sample. x follows a hypergeometric distribution (w/ parameters K, N, n). In urn white black sampled x n not sampled N n K N K N

22 Hypergeometric probabilities Suppose X Hypergeometric (N, K, n). No. of white balls in a sample of size n, drawn without replacement from an urn with K white and N K black. ( K N K ) Pr(X = x) = x)( n x ( N n) Example: In urn 0 1 sampled not N = 40, K = 29, n = 20 Pr(X = 18) = ( 29 )( ( ) ) 1.4% Back to Fisher s exact test Observed data N Y A B Assume the null hypothesis (independence) is true. Constrain the marginal counts to be as observed. Pr(observed table H 0 ) = Pr(X=18) X Hypergeometric (N=40, K=29, n=20)

23 Fisher s exact test 1. For all possible tables (with the observed marginal counts), calculate the relevant hypergeometric probability. 2. Use that probability as a statistic. 3. P-value (for Fisher s exact test of independence): The sum of the probabilities for all tables having a probability equal to or smaller than that observed. An illustration The observed data N Y A B All possible tables (with these marginals):

24 Fisher s exact test: example 1 Observed data N Y A B P-value 3.1% In R: fisher.test() Recall: χ 2 test: P-value = 1.3% LRT: P-value = 1.1% Fisher s exact test: example 2 Observed data I-B NI-B I-A NI-A P-value 4.4% Recall: χ 2 test: P-value = 3.0% LRT: P-value = 3.7%

25 Summary Testing for independence in a 2 x 2 table: A special case of testing a composite hypothesis in a onedimensional table. You can use either the LRT or χ 2 test, as before. You can also use Fisher s exact test. If Fisher s exact test is computationally feasible, do it! Paired data Sample 100 subjects and determine whether they are infected with viruses A and B. I-B NI-B I-A NI-A Underlying probabilities B 0 1 A 0 p 00 p 01 p 0+ 1 p 10 p 11 p 1+ p +0 p +1 1 Is the rate of infection of virus A the same as that of virus B? In other words: Is p 1+ = p +1? Equivalently, is p 10 = p 01?

26 McNemar s test H 0 :p 01 = p 10 Under H 0, e.g. if p 01 = p 10, the expected counts for cells 01 and 10 are both equal to (n 01 + n 10 )/2. The χ 2 test statistic reduces to X 2 = (n 01 n 10 ) 2 n 01 + n 10 For large sample sizes, this statistic has null distribution that is approximately a χ 2 (df = 1). For the example: X 2 = (20 9) 2 / 29 = 4.17 P = 4.1%. An exact test Condition on n 01 + n 10. Under H 0,n 01 n 01 + n 10 Binomial(n 01 + n 10, 1/2). In R, use the function binom.test. For the example, P = 6.1%.

27 Paired data Paired data I-B NI-B I-A NI-A P = 6.1% Unpaired data I NI A B P = 9.5% Taking appropriate account of the pairing is important! r x k tables Blood type Population A B AB O Florida Iowa Missouri Same distribution of blood types in each population?

28 Underlying probabilities Observed data 1 2 k 1 n 11 n 12 n 1k n 1+ 2 n 21 n 22 n 2k n r n r1 n r2 n rk n r+ n +1 n +2 n +k n Underlying probabilities 1 2 k 1 p 11 p 12 p 1k p 1+ 2 p 21 p 22 p 2k p r p r1 p r2 p rk p r+ p +1 p +2 p +k 1 H 0 : p ij = p i+ p +j for all i,j. Expected counts Observed data A B AB O F I M Expected counts A B AB O F I M Expected counts under H 0 : e ij = n i+ n +j /n for all i,j.

29 χ 2 and LRT statistics Observed data A B AB O F I M Expected counts A B AB O F I M X 2 statistic = (obs exp) 2 exp = = 5.64 LRT statistic = 2 obs ln(obs/exp) = = 5.55 Asymptotic approximation If the sample size is large, the null distribution of the χ 2 and likelihood ratio test statistics will approximately follow a χ 2 distribution with (r 1) (k 1) d.f. In the example, df = (3 1) (4 1) = 6 X 2 = 5.64 P = LRT = 5.55 P = 0.48.

30 Fisher s exact test Observed data 1 2 k 1 n 11 n 12 n 1k n 1+ 2 n 21 n 22 n 2k n r n r1 n r2 n rk n r+ n +1 n +2 n +k n Assume H 0 is true. Condition on the marginal counts Then Pr(table) 1/ ij n ij! Consider all possible tables with the observed marginal counts Calculate Pr(table) for each possible table. P-value = the sum of the probabilities for all tables having a probability equal to or smaller than that observed. Fisher s exact test: the example Observed P!value = 48% " log(n ij!) Since the number of possible tables can be very large, we often must resort to computer simulation.

31 Another example Survival in different treatment groups: Survive Treatment No Yes A 15 5 B 17 3 C D 17 3 E 16 4 Is the survival rate the same for all treatments? Results Observed Survive Treatment No Yes A 15 5 B 17 3 C D 17 3 E 16 4 Expected under H 0 Survive Treatment No Yes A 15 5 B 15 5 C 15 5 D 15 5 E 15 5 X 2 = 9.07 P = 5.9% (how many df?) LRT = 8.41 P = 7.8% Fisher s exact test: P = 8.7%

32 All pairwise comparisons N Y A 15 5 P=69% B 17 3 N Y A 15 5 P=19% C N Y A 15 5 P=69% D 17 3 N Y B 17 3 P=4.1% C N Y B 17 3 P=100% D 17 3 N Y B 17 3 P=100% E 16 4 N Y C P=9.6% E 16 4 N Y D 17 3 P=100% E 16 4 N Y A 15 5 E 16 4 P=100% N Y C D 17 3 P=4.1% Is this a good thing to do? Two-locus linkage in an intercross BB Bb bb AA Aa aa Are these two loci linked?

33 General test of independence Observed data BB Bb bb AA Aa aa Expected counts BB Bb bb AA Aa aa χ 2 test: X 2 = 10.4 P = 3.5% (df = 4) LRT test: LRT = 9.98 P = 4.1% Fisher s exact test: P = 4.6% A more specific test Observed data BB Bb bb AA Aa aa Underlying probabilities BB Bb bb AA 1 4 (1 θ)2 1 2 θ(1 θ) 1 4 θ2 Aa 1 2 θ(1 θ) 1 2 [θ2 +(1 θ) 2 1 ] 2θ(1 θ) 1 aa 4 θ2 1 2 θ(1 θ) 1 4 (1 θ)2 H 0 : θ = 1/2 versus H a : θ < 1/2 Use a likelihood ratio test! Obtain the general MLE of θ. Calculate the LRT statistic = 2 ln Compare this statistic to a χ 2 (df = 1). { Pr(data ˆθ) } Pr(data θ=1/2)

34 Results BB Bb bb AA Aa aa MLE: ˆθ = LRT statistic: LRT = 7.74 P = 0.54% (df = 1) Here we assume Mendelian segregation, and that deviation from H 0 is in a particular direction. If these assumptions are correct, we ll have greater power to detect linkage using this more specific approach.