• Decision Theory ( DT ) risk minimisation approach. parameters: • However, this penalises all high K models even if sample size is large too. 16 Information Criterion • Akaike Information Criterion ( AIC ), penalising number of • AIC = − 2 ln ( L ) + 2 K • Corrected Akaike Information Criterion ( AICc ) • AICc = AIC + 2 K ( K + 1 ) n − K − 1 • Alternatively, there is the Bayesian Information Criterion ( BIC ): • BIC = − 2 ln ( L ) + Kln ( n )
parameters: • However, this penalises all high K models even if sample size is large too. 16 Information Criterion • Akaike Information Criterion ( AIC ), penalising number of • AIC = − 2 ln ( L ) + 2 K • Corrected Akaike Information Criterion ( AICc ) • AICc = AIC + 2 K ( K + 1 ) n − K − 1 • Alternatively, there is the Bayesian Information Criterion ( BIC ): • BIC = − 2 ln ( L ) + Kln ( n ) • Decision Theory ( DT ) risk minimisation approach.
• What if everything fits poorly? • Information criterion test relative goodness of fit instead of absolute • Parametric Bootstrapping/Posterior Predictive Simulation • If the model is reasonable then data simulated under should resemble the empirical data 17 Limitations
• What if everything fits poorly? • Information criterion test relative goodness of fit instead of absolute • Parametric Bootstrapping/Posterior Predictive Simulation • If the model is reasonable then data simulated under should resemble the empirical data 17 Limitations
• What if everything fits poorly? • Information criterion test relative goodness of fit instead of absolute • Parametric Bootstrapping/Posterior Predictive Simulation • If the model is reasonable then data simulated under should resemble the empirical data 17 Limitations
• What if everything fits poorly? • Information criterion test relative goodness of fit instead of absolute • Parametric Bootstrapping/Posterior Predictive Simulation • If the model is reasonable then data simulated under should resemble the empirical data 17 Limitations
Branch Support Testing
18 Bootstrapping in General The bootstrap (unknown) true value of estimate of empirical distribution of sample (unknown) true distribution Bootstrap replicates Distribution of estimates of parameters Slide from Joe Felsenstein
19 Bootstrapping Phylogenies The bootstrap for phylogenies sites Original Data ^ T sample same number T (1) Bootstrap sample of sites, with replacement #1 Bootstrap sample same number T (2) sample of sites, with replacement #2 (and so on) Slide from Joe Felsenstein
20 Bootstrapping Phylogenies
21 Bootstrapping Phylogenies The majority-rule consensus tree Trees: E B E B C D A D F A C F C B E B E B A C D A E D F F D F C A How many times each partition of species is found: AE | BCDF 4 B E ACE | BDF 3 0.8 ACEF | BD 1 A D AC | BDEF 1 0.6 AEF | BCD 1 0.6 ADEF | BC 2 F C ABCE | DF 3 Slide from Joe Felsenstein
22 Combining the results
• Randomly reweighing the sites in an alignments • Probability of a site being excluded 1 1 n n • Asymptotically approximately 0 36 • Goal to simulate an infinite population (number of alignment columns) 23 What is the bootstrap doing?
• Randomly reweighing the sites in an alignments n n • Asymptotically approximately 0 36 • Goal to simulate an infinite population (number of alignment columns) 23 What is the bootstrap doing? • Probability of a site being excluded 1 − 1
• Randomly reweighing the sites in an alignments n n • Goal to simulate an infinite population (number of alignment columns) 23 What is the bootstrap doing? • Probability of a site being excluded 1 − 1 • Asymptotically approximately 0 . 36
• Randomly reweighing the sites in an alignments n n • Goal to simulate an infinite population (number of alignment columns) 23 What is the bootstrap doing? • Probability of a site being excluded 1 − 1 • Asymptotically approximately 0 . 36
• Typically underestimates the true probabilities • i.e biased but conservative • Computationally demanding • Assumes independence of sites • Relies on good input data • Only answers to what extent does input data support a given part of the tree 24 Limitations
• Typically underestimates the true probabilities • i.e biased but conservative • Computationally demanding • Assumes independence of sites • Relies on good input data • Only answers to what extent does input data support a given part of the tree 24 Limitations
• Typically underestimates the true probabilities • i.e biased but conservative • Computationally demanding • Assumes independence of sites • Relies on good input data • Only answers to what extent does input data support a given part of the tree 24 Limitations
• Typically underestimates the true probabilities • i.e biased but conservative • Computationally demanding • Assumes independence of sites • Relies on good input data • Only answers to what extent does input data support a given part of the tree 24 Limitations
• Typically underestimates the true probabilities • i.e biased but conservative • Computationally demanding • Assumes independence of sites • Relies on good input data • Only answers to what extent does input data support a given part of the tree 24 Limitations
• Typically underestimates the true probabilities • i.e biased but conservative • Computationally demanding • Assumes independence of sites • Relies on good input data • Only answers to what extent does input data support a given part of the tree 24 Limitations
• Simulate data sets of this size assuming the estimate of the tree is the truth • Key for many more sophisticated tests. • Can be used to generate p -values, but non-trivial 25 Parametric Bootstraps
• Rapid Bootstraps ( RBS ) • Ultrafast Bootstraps ( UFBoot ) • Instead of re-doing the full ML inference just re-sample the site ln L values and sum 26 Alternative Approaches • Resampling estimated log-likelihoods ( RELL )
• Rapid Bootstraps ( RBS ) • Ultrafast Bootstraps ( UFBoot ) • Instead of re-doing the full ML inference just re-sample the site 26 Alternative Approaches • Resampling estimated log-likelihoods ( RELL ) ln ( L ) values and sum
• Ultrafast Bootstraps ( UFBoot ) • Instead of re-doing the full ML inference just re-sample the site 26 Alternative Approaches • Resampling estimated log-likelihoods ( RELL ) ln ( L ) values and sum • Rapid Bootstraps ( RBS )
26 • Instead of re-doing the full ML inference just re-sample the site Alternative Approaches • Resampling estimated log-likelihoods ( RELL ) ln ( L ) values and sum • Rapid Bootstraps ( RBS ) • Ultrafast Bootstraps ( UFBoot )
• Parametric aLRT : 2 of • Non-parametric SH-aLRT • aBayes : P X T c P T c 0 P X T i P T i with 27 P T 2 P T 1 P T 0 flat prior i • P T c 2 X • Comparing the 3 nearest NNIs based on RELL branch vs. closest NNIs for to a given branch: Likelihood Tests
• Non-parametric SH-aLRT • aBayes : P X T c P T c 0 P X T i P T i with 27 P T 2 P T 1 P T 0 flat prior i X 2 • Comparing the 3 nearest NNIs • P T c based on RELL branch vs. closest NNIs to a given branch: Likelihood Tests • Parametric aLRT : χ 2 of δ for
• aBayes : P X T c P T c 0 P X T i P T i with 27 P T 2 P T 1 P T 0 flat prior i 2 X • Comparing the 3 nearest NNIs • P T c based on RELL branch vs. closest NNIs to a given branch: Likelihood Tests • Parametric aLRT : χ 2 of δ for • Non-parametric SH-aLRT
P X T c P T c 0 P X T i P T i with 27 P T 2 P T 1 P T 0 flat prior i 2 X • Comparing the 3 nearest NNIs • P T c based on RELL branch vs. closest NNIs to a given branch: Likelihood Tests • Parametric aLRT : χ 2 of δ for • Non-parametric SH-aLRT • aBayes :
• Comparing the 3 nearest NNIs to a given branch: branch vs. closest NNIs based on RELL flat prior 27 Likelihood Tests • Parametric aLRT : χ 2 of δ for • Non-parametric SH-aLRT • aBayes : P ( X | T c ) P ( T c ) • P ( T c | X ) = ∑ 2 i = 0 P ( X || T i ) P ( T i ) with P ( T 0 ) = P ( T 1 ) = P ( T 2 )
Comparing Trees
28 https://github.com/mtholder/TreeTopoTestingTalks How to compare competing hypotheses?
29 https://github.com/mtholder/TreeTopoTestingTalks How to compare competing hypotheses?
30 Simplistic Comparison
k p k 1 p n • 4 sites favour the red tree, 2 favour the blue • n k • 4 out of 6 p 0 6875 • 40 out of 60 p 0 0124 • 400 out of 600 p 2 3 10 16 31 Qualitative Comparison
• 4 sites favour the red tree, 2 favour the blue • k • 4 out of 6 p 0 6875 • 40 out of 60 p 0 0124 • 400 out of 600 p 2 3 10 16 31 Qualitative Comparison ( n ) p k ( 1 − p ) n − k
• 4 sites favour the red tree, 2 favour the blue • k • 40 out of 60 p 0 0124 • 400 out of 600 p 2 3 10 16 31 Qualitative Comparison ( n ) p k ( 1 − p ) n − k • 4 out of 6 p = 0 . 6875
• 4 sites favour the red tree, 2 favour the blue • k • 400 out of 600 p 2 3 10 16 31 Qualitative Comparison ( n ) p k ( 1 − p ) n − k • 4 out of 6 p = 0 . 6875 • 40 out of 60 p = 0 . 0124
• 4 sites favour the red tree, 2 favour the blue • k 31 Qualitative Comparison ( n ) p k ( 1 − p ) n − k • 4 out of 6 p = 0 . 6875 • 40 out of 60 p = 0 . 0124 • 400 out of 600 p = 2 . 3 ∗ 10 − 16
• 2 15 22 • t 2 N 0 148 • therefore: p 0 888 under 5 d f 32 Quantiative Comparison • µ = ( − 5 . 2 + 3 . 1 + 0 . 9 + 6 . 6 + 0 . 3 − 0 . 2 ) / 6 = 0 . 916
• t 2 N 0 148 • therefore: p 0 888 under 5 d f 32 Quantiative Comparison • µ = ( − 5 . 2 + 3 . 1 + 0 . 9 + 6 . 6 + 0 . 3 − 0 . 2 ) / 6 = 0 . 916 • σ 2 = 15 . 22
0 888 under 5 d f • therefore: p 32 Quantiative Comparison • µ = ( − 5 . 2 + 3 . 1 + 0 . 9 + 6 . 6 + 0 . 3 − 0 . 2 ) / 6 = 0 . 916 • σ 2 = 15 . 22 √ • t = µ σ 2 ∗ N = 0 . 148
32 Quantiative Comparison • µ = ( − 5 . 2 + 3 . 1 + 0 . 9 + 6 . 6 + 0 . 3 − 0 . 2 ) / 6 = 0 . 916 • σ 2 = 15 . 22 √ • t = µ σ 2 ∗ N = 0 . 148 • therefore: p = 0 . 888 under 5 d . f .
T 1 T 2 T 1 T 2 2 to get a critical value for ? 33 • Expectation under null • Tree space is difficult. • Why can’t we just use 0 X X ln L T 2 X 2 ln L T 1 X • the data equally well. More robust approaches • Null: if no sampling error (infinite data) T 1 and T 2 would explain
T 1 T 2 2 to get a critical value for ? the data equally well. • Expectation under null X 0 • Why can’t we just use • Tree space is difficult. 33 More robust approaches • Null: if no sampling error (infinite data) T 1 and T 2 would explain • δ ( T 1 , T 2 | X ) = 2 [ ln L ( T 1 | X ) − ln L ( T 2 | X )]
2 to get a critical value for ? the data equally well. • Why can’t we just use • Tree space is difficult. 33 More robust approaches • Null: if no sampling error (infinite data) T 1 and T 2 would explain • δ ( T 1 , T 2 | X ) = 2 [ ln L ( T 1 | X ) − ln L ( T 2 | X )] • Expectation under null E [ δ ( T 1 , T 2 | X )] = 0
• Tree space is difficult. the data equally well. 33 More robust approaches • Null: if no sampling error (infinite data) T 1 and T 2 would explain • δ ( T 1 , T 2 | X ) = 2 [ ln L ( T 1 | X ) − ln L ( T 2 | X )] • Expectation under null E [ δ ( T 1 , T 2 | X )] = 0 • Why can’t we just use χ 2 to get a critical value for δ ?
• Tree space is difficult. the data equally well. 33 More robust approaches • Null: if no sampling error (infinite data) T 1 and T 2 would explain • δ ( T 1 , T 2 | X ) = 2 [ ln L ( T 1 | X ) − ln L ( T 2 | X )] • Expectation under null E [ δ ( T 1 , T 2 | X )] = 0 • Why can’t we just use χ 2 to get a critical value for δ ?
• Many avenues: • Non-parametric bootstrapping • Parametric bootstrapping • Related approaches. 34 Estimating variance of the null
T 1 T 2 T 1 T 2 T 1 T 2 0 • Can’t handle multiple comparisons. selection bias • Due to centring assumption can’t be used for optimal tree i.e. two-tail t -test • Test • Non-parametric Bootstrap to estimate Null variance 35 • First, winning sites test • H a 0 X ln L T 2 X ln L T 1 • H 0 Kishino-Hasegawa Test
T 1 T 2 T 1 T 2 • First, winning sites test • H a 0 • Non-parametric Bootstrap to estimate Null variance • Test two-tail t -test • Due to centring assumption can’t be used for optimal tree i.e. selection bias • Can’t handle multiple comparisons. 35 Kishino-Hasegawa Test • H 0 : [ ln L ( T 1 | X ) − ln L ( T 2 | X )] = E [ δ ( T 1 , T 2 )] = 0
T 1 T 2 • First, winning sites test • Non-parametric Bootstrap to estimate Null variance • Test two-tail t -test • Due to centring assumption can’t be used for optimal tree i.e. selection bias • Can’t handle multiple comparisons. 35 Kishino-Hasegawa Test • H 0 : [ ln L ( T 1 | X ) − ln L ( T 2 | X )] = E [ δ ( T 1 , T 2 )] = 0 • H a : E [ δ ( T 1 , T 2 )] ̸ = 0
T 1 T 2 • First, winning sites test • Non-parametric Bootstrap to estimate Null variance • Test two-tail t -test • Due to centring assumption can’t be used for optimal tree i.e. selection bias • Can’t handle multiple comparisons. 35 Kishino-Hasegawa Test • H 0 : [ ln L ( T 1 | X ) − ln L ( T 2 | X )] = E [ δ ( T 1 , T 2 )] = 0 • H a : E [ δ ( T 1 , T 2 )] ̸ = 0
• First, winning sites test • Non-parametric Bootstrap to estimate Null variance • Due to centring assumption can’t be used for optimal tree i.e. selection bias • Can’t handle multiple comparisons. 35 Kishino-Hasegawa Test • H 0 : [ ln L ( T 1 | X ) − ln L ( T 2 | X )] = E [ δ ( T 1 , T 2 )] = 0 • H a : E [ δ ( T 1 , T 2 )] ̸ = 0 • Test E [ δ ( T 1 , T 2 )] two-tail t -test
• First, winning sites test • Non-parametric Bootstrap to estimate Null variance • Due to centring assumption can’t be used for optimal tree i.e. selection bias • Can’t handle multiple comparisons. 35 Kishino-Hasegawa Test • H 0 : [ ln L ( T 1 | X ) − ln L ( T 2 | X )] = E [ δ ( T 1 , T 2 )] = 0 • H a : E [ δ ( T 1 , T 2 )] ̸ = 0 • Test E [ δ ( T 1 , T 2 )] two-tail t -test
• First, winning sites test • Non-parametric Bootstrap to estimate Null variance • Due to centring assumption can’t be used for optimal tree i.e. selection bias • Can’t handle multiple comparisons. 35 Kishino-Hasegawa Test • H 0 : [ ln L ( T 1 | X ) − ln L ( T 2 | X )] = E [ δ ( T 1 , T 2 )] = 0 • H a : E [ δ ( T 1 , T 2 )] ̸ = 0 • Test E [ δ ( T 1 , T 2 )] two-tail t -test
• Approximately Unbiased Test • Swofford–Olsen–Waddell–Hillis same idea but uses parametric • Compares candidate tree sets • H 0 all topologies equally good • Very conservative when the number of candidate trees is large • Can be corrected with weighted SH-test overcomes. • Achieves weighted by varying bootstrap size for each tree. • Better for larger comparisons, can have issues with P-space curvature. bootstraps instead. • Sensitive to model misspecification. 36 Alternative tests • Shimodaira-Hasegawa Test
• Approximately Unbiased Test • Swofford–Olsen–Waddell–Hillis same idea but uses parametric • Compares candidate tree sets • H 0 all topologies equally good • Very conservative when the number of candidate trees is large • Can be corrected with weighted SH-test overcomes. • Achieves weighted by varying bootstrap size for each tree. • Better for larger comparisons, can have issues with P-space curvature. bootstraps instead. • Sensitive to model misspecification. 36 Alternative tests • Shimodaira-Hasegawa Test
• Approximately Unbiased Test • Swofford–Olsen–Waddell–Hillis same idea but uses parametric • Compares candidate tree sets • Very conservative when the number of candidate trees is large • Can be corrected with weighted SH-test overcomes. • Achieves weighted by varying bootstrap size for each tree. • Better for larger comparisons, can have issues with P-space curvature. bootstraps instead. • Sensitive to model misspecification. 36 Alternative tests • Shimodaira-Hasegawa Test • H 0 = all topologies equally good
• Approximately Unbiased Test • Swofford–Olsen–Waddell–Hillis same idea but uses parametric • Compares candidate tree sets • Very conservative when the number of candidate trees is large • Can be corrected with weighted SH-test overcomes. • Achieves weighted by varying bootstrap size for each tree. • Better for larger comparisons, can have issues with P-space curvature. bootstraps instead. • Sensitive to model misspecification. 36 Alternative tests • Shimodaira-Hasegawa Test • H 0 = all topologies equally good
• Approximately Unbiased Test • Swofford–Olsen–Waddell–Hillis same idea but uses parametric • Compares candidate tree sets • Very conservative when the number of candidate trees is large • Can be corrected with weighted SH-test overcomes. • Achieves weighted by varying bootstrap size for each tree. • Better for larger comparisons, can have issues with P-space curvature. bootstraps instead. • Sensitive to model misspecification. 36 Alternative tests • Shimodaira-Hasegawa Test • H 0 = all topologies equally good
• Swofford–Olsen–Waddell–Hillis same idea but uses parametric • Compares candidate tree sets • Very conservative when the number of candidate trees is large • Can be corrected with weighted SH-test overcomes. • Achieves weighted by varying bootstrap size for each tree. • Better for larger comparisons, can have issues with P-space curvature. bootstraps instead. • Sensitive to model misspecification. 36 Alternative tests • Shimodaira-Hasegawa Test • H 0 = all topologies equally good • Approximately Unbiased Test
Recommend
More recommend