Download presentation
Presentation is loading. Please wait.
Published byEugenia Crawford Modified over 9 years ago
1
Ark nr.: 1 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk Rotation Tests - Computing exact adjusted p-values in multiresponse experiments Øyvind Langsrud, MATFORSK, Norwegian Food Research Institute.
2
Ark nr.: 2 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk Campylobacter experiment Three biological replicates (block variable) 312 FT-IR wavelengths as multiple responses polysaccharide region [1200-900 cm -1 ]
3
Ark nr.: 3 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk Analysis with MINITAB - first wavelength
4
Ark nr.: 4 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk ANALYSIS with 50-50 MANOVA
5
Ark nr.: 5 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk Adjusted means as curves
6
Ark nr.: 6 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk Effect of Day - 312 single response p-values Ordinary significance tests are not longer suitable A lot of significant results cased by random variation (since several tests/responses) The p-values need to be adjusted So that they are interpretable
7
Ark nr.: 7 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk Adjusted p-values So that experimentwise (or familywise) error rate is controlled Bonferroni correction (classical method) pAdj = #responses pRaw Conservative upper bound (in practice often too conservative ) Dependence among responses not investigated Modern methods Makes active use of dependence among responses Permutation tests Rotation tests
8
Ark nr.: 8 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk Assume a regression model (simplified model without constant term) Separate F-tests for each response Random variables: F 1, F 2 …, F q Observed values: f 1, f 2 …, f q Maximal F-value (= minimum p-value) obtained for response number k Raw p-value: Adjusted p-value:
9
Ark nr.: 9 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk Adjusting the minimum p-value by permutations For m =1,2 …. M permute data ( Y P (m) Y ) compute maximal F-statistic from these data Compute p-value as
10
Ark nr.: 10 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk How is dependence handled? Estimate of covariance matrix under H 0 : Estimate based on permuted data: The permutation test is a conditional test Conditioned on the covariance matrix estimate Conditioned on sufficient statistics for the unknown parameters Fisher's exact test for 2 2 contingency tables is the most famous conditional test
11
Ark nr.: 11 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk Conditional test under multivariate normality? Need distribution of Y conditioned on Y T Y Answer Y is distributed as RY obs where Y obs is the observed matrix and where R is an uniformly distributed orthogonal matrix (random rotation matrix) Relation to well-known tests t-test, F-tests, Hotelling T 2, Wilks’ are special cases of rotation testing But these test statistics do not depend on Y T Y Conditioning not needed Simulations not necessary
12
Ark nr.: 12 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk Adjusting the minimum p-value by rotations For m =1,2 …. M simulate rotated data ( Y R (m) Y ) where R (m) is a simulated random rotation matrix compute maximal F-statistic from these data Compute p-value as In practice: a much more efficient algorithm is applied
13
Ark nr.: 13 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk Adjusting the other p-values (permutations or rotations) Remove response with minimum p-value Adjust minimum p-value in new data set and so on Enforce monotonicity All calculations can be done simultaneously
14
Ark nr.: 14 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk Permutation test or rotation test Exact permutation testing The only assumption: independent observations Useless for few observations Useless for complex ANOVA and regression models Exact rotation testing Assumes multivariate normality Does not need as many observations as permutation testing Can be use for complex ANOVA and regression models F-test rotation test
15
Ark nr.: 15 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk Adjusted p-values (FWE) non-adjusted p-values (RAW) False significance at 1% level is expected in 1% of all the investigated responses If you have 5000 responses ….. “Classically ” adjusted p-values (FWE) False significance at 1% level is expected in not more that 1% of all experiments where the method is applied. The experimentwise (or familywise) error rate is controlled
16
Ark nr.: 16 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk False Discovery Rate (FDR) Adjusted p-values according to False Discovery Rate False significance at 1% level is expected in 1% of all cases (responses) reported as significant at 1% level. If you have 5000 responses and 200 are reported as significant at 1% level, one will expect two of these as false. “q-values” is proposed instead of “adjusted p-values”
17
Ark nr.: 17 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk Calculation of FDR adjusted p-values Several methods exist Most of them do not handle the dependence among the responses but OK if the “weak dependence requirement” is met New variant based on rotations (or alternatively permutations) handles any kind of dependence conservative compared to other methods since the method does not involve an estimate of the amount of responses with true null hypotheses
18
Ark nr.: 18 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk Adjusted p-values (first 30 wavelengths)
19
Ark nr.: 19 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk Adjusted p-values (30 most significant wavelengths)
20
Ark nr.: 20 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk p-values
21
Ark nr.: 21 | Forfatter: Øyvind Langsrud - a member of the Food Science Alliance | NLH - Matforsk - Akvaforsk Rotation Tests - Conclusion Simulation principle for computing exact Monte Carlo p-value for any test statistic. Based on multivariate normal distribution. Generalisation of classical tests. Related to permutation testing. Useful for computing adjusted p-values (F-tests) FWE, FDR General linear models (ANOVA and regression) Implemented in the 50-50 MANOVA program ( www.matforsk.no/ola )
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.