Combining data versus consensus methods

1
Combining data versus consensus methods
2
Multiple data sets for the same sets of taxa

• Two strategies
• Analyze each data set separately and then compare
  the trees (consense)
• Concatenate the data and conduct a single
  combined analysis (combine)

3
Argument for consensus

• If the same clades appear for multiple data sets
  we can be more confident
• The method is conservative

4
Is consensus conservative?Barrett et al. 1994.
Syst. Zool. 40486
5
Is consensus conservative?Barrett et al. 1994.
Syst. Zool. 40486
6
Argument for combining data

• Data partitions are arbitrary
• Better signalnoise ratio
• Can evaluate confidence in the combined data set
• Should look at the total evidence

7
Arguments against combined analysis

• Some data sets might have strong misleading
  signals (e.g., due to lab errors)
• How should one weight partitions?
• Different partitions might have tracked different
  histories

8
Combinational combined analysis

• First assess if the data conflict significantly
• If they do not combine
• If they do analyze separately

9
Tests of data set conflict

• Topology tests (Templeton, Kishino-Hasegawa
  Shimodaira-Hasegawa)
• One data partition versus trees from the other
  partition
• Incongruence Length Different (ILD) test
  Partition Homogeneity test
• Direct comparison of the partitions

10
Topology tests for conflict
? reject
What is wrong with this test?
11
Topology tests for conflict
Confidence interval for data set 1
Confidence interval for data set 2
x
x
Do these data sets conflict?
Does each data set reject the optimal tree from
the other data set?
12
But topology tests can be used more carefully

• Two data sets do conflict if
• Data set 1 rejects all tree that lack a certain
  clade
• Data set 2 rejects all tree that have that same
  clade
• Look at clades in the separate analyses that are
  well supported and contradict relationships in
  the other

13
ILD versus Topology tests

• ILD can quickly identify data set conflict, but
  do not localize the conflict
• Use selective deletion?
• Topology tests can often miss conflict
• When conflict is found it is easily interpretted

14
Option if you find conflict

• Conduct separate analyses only
• Delete taxa until conflict disappears - then
  combine
• Combine anyway

15
Conditional conditional combined analysis

• You believe that conflict reflects data
  partitions tracking different histories
• Keep the data separate and find ways to summarize
  the discrepancy
• You believe that conflict reflects artifactual
  signals (noise) in one or both data sets
• Combine anyway in the hope that the real signal
  will come to dominate