Multiple Testing. Hoang Tran. Department of Statistics, Florida State University

Transcription

1 Multiple Testing Hoang Tran Department of Statistics, Florida State University

2 Large-Scale Testing Examples: Microarray data: testing differences in gene expression between two traits/conditions Microbiome data: which bacteria are differentially expressed between two traits/conditions? Prostate cancer data: n = 102 subjects (52 cases and 50 controls) and N = 6033 genes. How do we test for differences in gene expression?

3 Large-Scale Testing 1. For the jth gene: compute the two-sample t statistic comparing gene expression between cases and controls (t j ). 2. Test H 0j : gene j s expression levels are the same between the two groups at significance level α. 3. Repeat for all N = 6033 genes. What s the problem?

4 Multiple Comparisons Running 100 separate hypothesis tests at α = 0.05 will produce about 5 significant results even if each case is actually null. Examples: Efficacy of a drug in terms of reduction of disease symptoms. Two methods of teaching writing are used on students. Students in the groups are compared in terms of grammar, spelling, etc. Increased likelihood of type I errors

5 Bonferroni Correction Family-Wise Error Rate: the probability of rejecting any true null hypothesis. Test each individual hypothesis at α/n. Let J 0 be the indices of the true H 0j with J 0 = N 0. ( FWER = P p j α ) N J 0 J 0 P { p j α } N = N 0 α N N α N = α The Bonferroni Correction ensures FWER α.

6 Bonferroni Correction No requirement of independence of p i s, but it s perhaps too conservative. N = 6033 and α = 0.05: only reject when p j ! We want to control type I errors, but we also want to find interesting/significant genes.

7 Holm s Procedure Order the p-values: p (1) p (2) p (N) with H 0(j) the null hypotheses. Let j 0 be the smallest index j such that p (j) > α/(n j + 1) Reject all H 0(j) for j < j 0 and accept all with j j 0. Satisfies FWER α but not as conservative as Bonferroni (more rejections).

8 Stepdown Procedures Holm s procedure: look at most significant test first and continue rejecting hypotheses if p-values are small. An improvement: incorporate the dependence structure of individual tests.

9 Generic Stepdown Method K {1,..., N}, H K : intersection hypothesis that all H 0j with j K are true. T j is the jth test statistic. T (1) T (N) and H 0(j). Generic Stepdown Method 1. Let K 1 = {1,..., N}. If T (N) ĉ K1 (1 α) then accept all hypotheses and stop; otherwise, reject H 0(N) and continue. 2. Let K 2 be the indices of the hypotheses not previously rejected. If T (N 1) ĉ K2 (1 α) then accept all hypotheses and stop; otherwise, reject H 0(N 1) and continue.

10 Generic Stepdown Method How do we find the critical values ĉ K (1 α)? Under certain conditions, FWER α (Lehmann and Romano 2005). FWER P (max j J 0 T j > ĉ J0 (1 α)) α Critical values are the (1 α) quantile of max j K T j under H K. Not as conservative as Holm s procedure in general.

11 The Hypothesis of Homogeneity Consider testing for all ( N 2 ) pairs i < j. H i,j : µ i = µ j, i < j {H 1,2, H 2,3 } cannot be the set of all true hypotheses Previous methods allow acceptance of H 1,2 : µ 1 = µ 2 and H 2,3 but rejection of H 1,3

12 A Holm type approach Setup (N = 6): Normal random variables with common variance σ 2. X(1) X (N) and µ (j). ˆp (i),(j) : the p-value for testing µ (i) = µ (j). Procedure: 1. If ˆp (6),(1) α/ ( N 2 ), accept all hypotheses and terminate. Otherwise, reject µ (1) = µ (6) and continue. 2. Test the largest of X (6) X (2) and X (5) X (1) by comparing ˆp (6),(2) or ˆp (5),(1) with α/( ( N 2 ) 1). FWER is controlled at α.

13 An improvement Suppose we are at step 2 (µ (1) = µ (6) has been rejected). µ (1) = µ (2) or µ (2) = µ (6) must be false. Possible true hypotheses: ( 6 2) 5 = 10 < ( 6 2) 1 = 14. No violation of FWER and more rejections.

14 An improvement

15 FWER Summary Family-Wise Error Rate: the probability of rejecting any true null hypothesis. Bonferroni, Holm, and Generic Stepdown method all control FWER Holm and Generic Stepdown method have at least as much power as Bonferroni We can exploit the structure of pairwise tests to improve the Holm procedure s power for these situations

16 False-Discovery Rates FWER probably too conservative for very large N, such as N 20. For N in the thousands/millions, the issue is exacerbated. A more liberal criterion: False-Discovery Rates.

17 False-Discovery Rates

18 False-Discovery Rates The number of false discoveries (a) is unobservable. We want to minimize Fdp = a/r. Define FDR(D) = E(Fdp(D)). We can t observe Fdp but we can control FDR. Decision rule D controls FDR at q (0, 1) if FDR(D) q

19 Benjamini-Hochberg Procedure for FDR Control 1. For given q, let j max be the largest j such that p (j) j N q 2. Let D q be the rule that rejects H 0(j) for all j j max. If p-values are independent: FDR(D q ) = N 0 N q q. FDR is more generous than FWER.

20 Benjamini-Hochberg q-values p.adjust in R computes FDR (q-value). In practice, q < 0.10 is significant. Example: q = 0.05 means we expect 5% of significant tests to result in false positives.

21 Microbiome Example for FDR Goal: study the association of the microbiome with asthma exacerbations. n = 3122 samples and N = 268 taxa. Question: which taxa are differentially expressed between samples with/without asthma exacerbations? 498 exacerbators, 2624 non-exacerbators. These are (possibly overdispersed) count data so we fit negative binomial regressions for each taxa.

22 Microbiome Example for FDR

23 Bayesian Interpretation of FDR Each of the N cases is null with prior probability π 0 or non-null with probability π 1 = 1 π 0. Each z statistic has density f 0 (z) if null (i.e. N(0, 1)), f 1 (z) if non-null (unknown). F 0 (z) and F 1 (z) are cdf s with survival curves S 0 (z) = 1 F 0 (z) and S 1 (z) = 1 F 1 (z). Define S(z) = π 0 S 0 (z) + π 1 S 1 (z) and f(z) = π 0 f 0 (z) + π 1 f 1 (z).

24 Bayesian Interpretation of FDR Suppose z j > z 0 = 3. Then Fdr(z 0 ) P (case j is null z j z 0 ) = π 0 S 0 (z 0 )/S(z 0 ) S 0 (z 0 ) usually known (i.e. 1 Φ(z 0 )) Ŝ(z 0 ) = #{z j z 0 }/N Empirical Bayes estimate: Fdr(z 0 ) = π 0 S 0 (z 0 )/Ŝ(z 0)

25 Bayesian Interpretation of FDR p (j) (j/n)q from BH procedure beomes So then Fdr(z 0 ) π 0 q S 0 (z (j) ) Ŝ(z (j))q BH rejects cases for which the empirical Bayes posterior probability of nullness is too small.

26 A Note about False-Negative Rates The false negative proportion: Fnp = (N 1 b)/(n R) The expectation of Fnp is a measure of Type II error. Let A be the region in which a null hypothesis is accepted. Then 1 Fdr(A) estimates the Bayesian false negative rate.

27 Local FDR Instead of a tail-area probability, what about z j = z 0? Local FDR: fdr(z 0 ) = P (case j is null z j = z 0 ) = π 0 f 0 (z 0 )/f(z 0 ) π 0 is unknown but can be estimated (Efron 2010) or set to 1 (most cases are null) f(z) is unknown but can be estimated

28 Local FDR

29 Local FDR Conventionally interesting threshold: fdr(z) 0.2. Local and tail-area FDR: Fdr(z 0 ) = E[fdr(z) z z 0 ] Often Fdr(z 0 ) < fdr(z 0 ).

30 Local FDR Computing ˆf(z): A fourth-degree log polynomial Poisson regression fit to the histogram of z-values. See next figure.

31 Local FDR

32 Choosing the Null Distribution In large-scale testing we can examine hundreds/thousands/millions of z-values. The chosen null distribution is inappropriate. What if we empirically determine the null distribution? Use the R package locfdr (also computes local FDR).

33 Empirical Null Distribution

34 FDR Summary FDR: a more liberal criterion than FWER. Many practitioners prefer to control FDR; in gene studies it is often more important to discover interesting genes. The BH procedure for FDR rejects cases for which the empirical Bayes posterior probability of nullness is too small. Local FDR: investigate more than just the tail-area probability.