putting lots of things in order: r-values for ranking in large-scale...
TRANSCRIPT
Putting lots of things in order: r-values for ranking in large-scale inference
Michael Newton Nick Henderson
Statistics Day, CDC
a general, unresolved statistics problem
●
−4 −2 0 2 4 6 8scale
error-free measurement
●
−4 −2 0 2 4 6 8
●●
scale
measurement
●
−4 −2 0 2 4 6 8
●●
scale
measurement
estimate +/- 2 SE
●
●
●
●
●
●
●
●
●
●
−4 −2 0 2 4 6 8
●
●
●
●
●
●
●
●
●
●
multivariate
●
●
●
●
●
●
●
●
●
●
−4 −2 0 2 4 6 8
●●
●●
●●
●●
●●
● ●
●●
●●
●●
● ●
multivariate
incr
easin
g pa
ramet
er
what we want
●
●
●
●
●
●
●
●
●
●
−4 −2 0 2 4 6 8
● ●
●●
●●
● ●
●●
●●
●●
●●
●●
●●
multivariate
incr
easin
g es
timat
e
what we get
● ●● ● ●● ●●●● ●●● ●●●●● ● ● ●●●● ●●● ●● ●● ●● ●● ●● ●● ●●●● ● ● ●● ●●●● ● ●● ●●●● ●● ●● ● ●● ●● ● ● ●● ●● ● ●● ● ●● ●● ●● ●● ●●● ● ● ●● ●● ●● ●● ●●● ●● ●● ●● ●●● ● ●●● ● ●●● ●● ●●●● ●●●●● ● ●● ●●● ● ●● ●●● ● ●●● ●● ●●●● ● ●● ●● ●● ● ●●● ● ●●●●● ●● ● ●● ●●●● ● ● ●●●●●●●●● ●● ● ●● ●●●● ●●●●●● ● ●●● ●●● ● ●●●●● ● ●● ●●● ●●● ●● ● ●●●●● ●●● ●● ●●● ● ●●●● ● ●●● ●● ● ●● ●● ●● ●● ●● ●●● ● ●● ● ●●●●● ●● ● ● ●● ●●● ●●● ● ●● ●● ●● ● ●●●●● ●● ●● ●● ●● ● ●●●●●● ● ●●●●●● ●●● ●● ● ● ●● ●● ●● ●●●●●● ●●● ● ●● ●● ● ●●● ●● ●●●●● ● ●●●● ●●●●● ●●●●● ●● ●●● ● ●●● ● ●●●● ●●● ●● ●● ●● ●●●● ●● ●●●● ●● ●●● ●●● ●●●●● ● ●●● ● ●●● ●● ●● ●● ●●●●● ● ●● ●● ●● ● ●●● ●●●●●● ●●● ●● ●●●●● ● ●● ●●●●● ● ●●●●●● ●● ●●●●● ●● ●● ●● ● ●● ●●●● ●● ●● ●● ●● ● ●●●● ●●● ● ●● ●●●●● ● ● ● ●●● ●●●● ●●● ●●● ●●● ● ●● ●●● ● ● ● ●●● ●● ●●● ●● ●● ● ●● ●●● ● ●●● ●●● ●●● ●● ●● ●● ●●● ●●● ●●● ●● ●● ●● ● ●● ●●● ●● ●●● ●● ●●● ●● ●● ● ●●● ●● ● ●● ●●● ●●● ●● ●● ●●● ● ●●● ●●● ●●● ●● ●● ●● ●●●● ●● ●● ●●● ●● ● ●●●●●● ●● ●●●●●●● ●● ●● ●●●● ●● ●●● ●● ●● ●● ● ●● ● ● ●●● ● ●● ●● ● ●● ●● ●● ●●● ●●● ● ●●●● ●●● ●●● ●● ●● ●●● ●● ● ● ●● ●● ●● ●● ●●●● ● ● ●●● ●● ●● ● ●●●● ●●● ● ●●●● ●●●●● ● ●● ●●● ● ●● ●● ●●● ●● ● ●● ●● ●●● ●● ●● ●● ●●●●● ● ● ●● ●● ●● ●●● ●● ●●● ● ●● ●● ●●● ● ●●●●● ●●●●●● ●●● ●●● ●● ● ●● ●●●● ●● ●●● ●●● ●●●● ● ●● ● ●●● ●● ●●● ● ●● ●●●● ● ●● ● ●●● ● ●●●● ● ●●● ● ● ●● ●●● ● ●●●● ● ●● ● ●●● ● ●● ●●● ● ● ●● ●●● ●● ●● ● ●●● ● ●● ●●● ● ●● ● ●●● ● ●●●● ●● ●●● ●●● ●●● ●●
−4 −2 0 2 4 6 8
● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●● ● ●● ●● ●● ●● ●● ●● ●● ●●●● ●● ●● ●●●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●● ●● ● ●● ●●●● ●● ●● ●●●● ●●●● ●●●● ●● ●●● ● ●● ●● ●● ●● ●● ●● ●●● ●●● ●●●● ●● ●● ●● ●● ●● ●● ●● ●●● ● ●● ●●● ●●● ●● ●● ●● ●● ●● ●● ●● ●●●● ●●●●● ● ●●● ●●●●● ●● ●● ●● ●● ●● ●● ●● ●●●● ●● ●● ●●● ●●●● ● ●● ●● ●●●● ●● ●●●●●●●● ●● ●●● ● ●●●●● ●●● ●● ●● ●● ●●● ● ●● ●●●●● ● ●●●● ●● ●● ●● ●● ●● ●●●● ●● ●● ●● ●● ●● ●● ●●●● ●● ●●●●●●●●● ●●●● ● ●● ●● ●● ●● ●● ●●●● ●● ●● ●● ●● ●● ●● ●● ●●●●●●● ●●●● ●●● ●● ●● ●● ●●●●● ● ●● ●●●●●●●●● ●● ● ●● ●●● ● ●● ●●● ● ●● ●●● ●● ●● ●●● ●● ●●●● ●● ●●●● ●● ●●● ● ●●● ●● ● ●● ●● ●●●●●● ●● ●● ●● ●●●● ●●●●● ●●● ●●●●●●● ● ●● ●● ●●● ● ●● ●● ●● ●●●● ●●● ●●●● ●●●● ●●●● ●●●●●● ●● ●●●● ●●● ●●●●● ●●●●● ●●● ●●● ●● ●●●● ●●●●● ●●●●●● ● ●● ●●●● ●●●● ●●●● ●● ●● ●● ●●●●●● ●●● ● ●●●● ●●● ●●●● ●●● ●●●● ●●●●● ●● ●●● ●● ●● ●●●●●●●● ●● ●●● ● ●●● ● ●● ●● ●● ●● ●● ●● ●●● ●●●●●●●●●● ●● ● ●●●●● ●● ● ●● ●● ●●● ●●● ●●●●●● ●●●● ●●●●●●●●●● ●● ●● ●● ●●●● ●●●●●●● ●● ●●●●●●●●●● ●●●●● ●●● ●● ● ●● ●● ●● ●● ●● ●●●●●●● ●●●● ●● ● ●●● ● ●●● ●●●●● ●●●●● ●● ● ●●● ● ●●● ●●●● ●●●●● ●●●●● ●●●● ●● ● ●● ●●●●●●● ●●● ●●●●● ●● ● ●●●●●●●●● ● ●●●● ●●●● ●● ●● ●●●●● ●● ●●●● ●●●● ●●●● ●●● ●● ●●● ● ●● ●● ●● ●●●●●● ●● ●● ●● ●●● ● ●●●●●●●●● ●●● ●●● ● ●●●●●●● ●● ●●● ●● ●●●●● ●●●● ●●●●● ●●●●● ●● ●● ●●●● ● ●●● ● ●●●● ●● ●●●● ●●●●● ●● ● ●●●● ●● ●● ●●● ●● ●● ●● ●●●●●● ●● ● ●●● ●● ●● ●●●● ●●●●●●●●●● ●● ●●● ●● ●●●●● ●●●●●● ●● ● ●● ●●● ●●●● ●● ● ●● ●●● ●● ●●●● ●●●● ●● ●● ●● ●●● ●●●●● ● ●●● ●●●●●● ●●●● ●●●● ●● ●●●● ● ●●● ●● ●●● ●● ●●● ●● ●● ●● ● ●●● ●● ●● ●● ● ●●● ●●●●●●●●●● ● ●●●●● ● ●●●●● ● ●●● ●●●● ● ●●● ● ●● ●●● ●●●●●●● ●●● ● ●●●●●● ●●● ●●●●●●● ●●● ● ●●●● ●● ●●●●● ●● ●● ●● ●●●● ● ●●●●●● ●● ●●● ●●●● ●●●● ●●●●●● ● ●● ●● ●●● ●●●● ●● ●●●● ●● ●● ●● ●●●● ●●●● ● ●●●●● ●● ●● ●● ● ●●● ● ●●● ●● ● ●●● ● ●● ●●●●●●● ●● ●● ● ●●● ●●●●●●●●●●●● ●● ● ●●● ● ●●●● ●●●●●●● ●●●● ●● ●● ●● ●●●● ● ●●● ●●●● ●● ●●●● ●● ●●● ●●●●● ●● ● ●●● ●●●● ●●● ●●● ● ●●● ● ●●● ● ●●●●●● ●●● ●● ●●● ●●● ●● ●●●●●●●● ●●●● ●● ● ●●●●●●● ● ●●● ●● ●●●●● ●● ●● ●●● ●●●● ●●●●● ●●● ● ●●● ●● ●● ●● ●● ●●●●●● ●● ●● ●●●● ●● ●● ●● ●●●● ● ●●●●● ●● ●●●●●●●● ● ●● ●●●●● ●● ●●● ●●● ● ●●● ●● ●● ●●●● ●●●● ●●●● ●● ● ●●● ●● ● ●●● ●●●● ●●● ●●● ●●●● ●●●● ●● ●● ●●● ●●● ●● ●● ●● ●●● ●●● ● ●●● ●●●● ●●●● ●● ●●●● ●●●● ●● ●● ●●●● ● ●●● ● ●●● ●● ●● ●●●●● ●●●●● ● ●●●●●●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●●●●● ●● ●● ●● ● ●●●●● ● ●●●●● ●● ●● ●● ●● ●● ●● ●● ●● ●●●●●● ●● ●● ●● ●● ●● ●● ●● ●● ●●●● ●● ●● ●● ●● ●● ●● ● ●● ●●● ●● ● ●●●●●●● ●● ● ●●● ●● ●● ●● ●●●● ● ●●● ●● ●● ●● ●● ●● ●● ●● ● ●●● ●● ●●●● ●● ●●● ●●● ●● ●● ●● ●● ●● ●●●● ● ●●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●● ●●
large scale
incr
easin
g es
timat
e
• regression effect• variance effect
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
950 960 970 980 990 1000
0.5
1.0
1.5
2.0
2.5
x[950:1000]
y[950:1000]
rank of point estimate (from bottom)
stan
dard
err
or
• large-scale• not sparse• ranking/sorting/prioritizing• variance artifacts• agreement• empirical Bayes • r-values
outline
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●●
●
●●●
●
●
●
●
●
●●●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●●
●
●
●
● ●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ●
●
●
●
●
●●
●●
●
●●
●
●
●
●●
●●
●● ●
●
●●
●●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●●
●
●
●●
● ●
●
●●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●●
●
●●
●●●
●●
●
●
●
●
● ●
●
●
●
●●
●
●●
●
●
●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ●
●●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
● ●
●●
●●●●●●●
●
●
●●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ● ●
●●
●●
● ●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
● ●
●●●
●
● ●
●
●
●
●
●
●●
●● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
● ●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●●
●●●
●●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●● ●●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
● ●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●● ●
●
●●
●●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●●
●●●
● ●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●●●●
●
●●
●●●●
●
●●
●
●
●
●●
●●
●●
●
●
●●●
●
●
●●●●
●
●
●
●
●●●
●●●
●● ●
●●●
●
●
●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●●●●●
●
●● ●
●
●
●●●
●
●●
●
●●
●
●
●●●
● ●●
●
●
●
●
●
●
●● ●
●●
●
●
●
●
●
●●●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●
● ●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●●
●
●●●
●
●
●
●●
●
●
●●
●●●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
● ●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
● ●
●
●
●
●
●●●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●●●
●
●
●●
●●
●●
●
●
●●
●●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●
● ●●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
● ●
●
●
●●
●●●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●●
●
●●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●●
●
● ●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
● ●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ● ●
●
●
●
●
●
●●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●●●
●
●
●
●
●
●
●●● ●
●
●●●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●●●●●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
● ●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●
●
●●
●
● ●●●
●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
● ●
●●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●●●
●●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
● ●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●●
●
●● ●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
● ●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●●●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●● ●
●
●
● ●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
● ●
●
●
●●
●
●●●●
●
●
●●
●
●
●●
●
●●
●
●●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●
● ●
●
●
●
●
●
●
●
●
●
●●● ●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●●●
●●●●
●
●
●
●
●●
●
●●
● ●
●
●●●
●●
●●
●●●
●
●
●
●●●●
●●
●●
●
●●
●
●
●
●
●
●●
●
●●●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ● ●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●● ●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
● ●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●●
●
●
●
●●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
● ●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
● ●●
●
●●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●●
●●
●
●
●
● ●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
● ●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
● ●
●
●
●
●
●
●
●●
●
●
●
● ●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●●●
●●
●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●● ●●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
● ●
● ●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
● ●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
● ●
●
●● ●
●
●
●
●
●
● ●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●
●
● ●●
●●●●●
●●
●● ●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
● ●
●
●
●●
●
●● ●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
● ●
●
● ●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ●
●●
●
●
●
●
●● ●●● ●
●●
●●
●● ●
●
●●● ●
●
●
●
●● ●
●
●
●
●●●● ●● ●
●●● ●●● ●●●
●●●●●●●●●● ●
●
●●●
●●
●
●
●●● ●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●●
●
●
●
●●●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●●
●●●●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
● ●
●
●●●
●
●
●
●●
●●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●●●
●
●
●●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●●●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●●
●●
●
●
● ●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●● ●
●
●
●
●
●
●●
●
●
●
●●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
● ●●●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●●●
●●●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●●●
●●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●●
●
●
●
●0.01 0.02 0.05 0.10
0.05
0.10
0.15
0.20
0.25
0.30
0.35
Standard Error
Log
Odd
s
Example Type 2 Diabetes (T2D) GWAS (Morris et al. 2012, Nat Gen)
• case/control (22,669 / 58,119)
• lots of T2D associated loci, but of small effect
(3371 SNPs shown)
• ?how to rank order?
log
odds
ratio
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●●
●
●●●
●
●
●
●
●
●●●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●●
●
●
●
● ●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ●
●
●
●
●
●●
●●
●
●●
●
●
●
●●
●●
●● ●
●
●●
●●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●●
●
●
●●
● ●
●
●●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●●
●
●●
●●●
●●
●
●
●
●
● ●
●
●
●
●●
●
●●
●
●
●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ●
●●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
● ●
●●
●●●●●●●
●
●
●●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ● ●
●●
●●
● ●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
● ●
●●●
●
● ●
●
●
●
●
●
●●
●● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
● ●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●●
●●●
●●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●● ●●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
● ●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●● ●
●
●●
●●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●●
●●●
● ●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●●●●
●
●●
●●●●
●
●●
●
●
●
●●
●●
●●
●
●
●●●
●
●
●●●●
●
●
●
●
●●●
●●●
●● ●
●●●
●
●
●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●●●●●
●
●● ●
●
●
●●●
●
●●
●
●●
●
●
●●●
● ●●
●
●
●
●
●
●
●● ●
●●
●
●
●
●
●
●●●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●
● ●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●●
●
●●●
●
●
●
●●
●
●
●●
●●●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
● ●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
● ●
●
●
●
●
●●●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●●●
●
●
●●
●●
●●
●
●
●●
●●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●
● ●●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
● ●
●
●
●●
●●●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●●
●
●●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●●
●
● ●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
● ●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ● ●
●
●
●
●
●
●●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●●●
●
●
●
●
●
●
●●● ●
●
●●●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●●●●●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
● ●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●
●
●●
●
● ●●●
●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
● ●
●●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●●●
●●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
● ●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●●
●
●● ●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
● ●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●●●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●● ●
●
●
● ●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
● ●
●
●
●●
●
●●●●
●
●
●●
●
●
●●
●
●●
●
●●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●
● ●
●
●
●
●
●
●
●
●
●
●●● ●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●●●
●●●●
●
●
●
●
●●
●
●●
● ●
●
●●●
●●
●●
●●●
●
●
●
●●●●
●●
●●
●
●●
●
●
●
●
●
●●
●
●●●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ● ●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●● ●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
● ●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●●
●
●
●
●●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
● ●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
● ●●
●
●●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●●
●●
●
●
●
● ●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
● ●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
● ●
●
●
●
●
●
●
●●
●
●
●
● ●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●●●
●●
●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●● ●●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
● ●
● ●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
● ●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
● ●
●
●● ●
●
●
●
●
●
● ●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●
●
● ●●
●●●●●
●●
●● ●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
● ●
●
●
●●
●
●● ●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
● ●
●
● ●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ●
●●
●
●
●
●
●● ●●● ●
●●
●●
●● ●
●
●●● ●
●
●
●
●● ●
●
●
●
●●●● ●● ●
●●● ●●● ●●●
●●●●●●●●●● ●
●
●●●
●●
●
●
●●● ●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●●
●
●
●
●●●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●●
●●●●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
● ●
●
●●●
●
●
●
●●
●●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●●●
●
●
●●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●●●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●●
●●
●
●
● ●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●● ●
●
●
●
●
●
●●
●
●
●
●●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
● ●●●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●●●
●●●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●●●
●●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●●
●
●
●
●0.01 0.02 0.05 0.10
0.05
0.10
0.15
0.20
0.25
0.30
0.35
Standard Error
Log
Odd
s
Example Type 2 Diabetes (T2D) GWAS (Morris et al. 2012, Nat Gen)
• case/control (22,669 / 58,119)
• lots of T2D associated loci, but of small effect
(3371 SNPs shown)
• ?how to rank order?
log
odds(T2D|A)
odds(T2D |Ac)
log
odds
ratio
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●●
●
●●●
●
●
●
●
●
●●●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●●
●
●
●
● ●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ●
●
●
●
●
●●
●●
●
●●
●
●
●
●●
●●
●● ●
●
●●
●●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●●
●
●
●●
● ●
●
●●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●●
●
●●
●●●
●●
●
●
●
●
● ●
●
●
●
●●
●
●●
●
●
●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ●
●●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
● ●
●●
●●●●●●●
●
●
●●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ● ●
●●
●●
● ●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
● ●
●●●
●
● ●
●
●
●
●
●
●●
●● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
● ●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●●
●●●
●●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●● ●●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
● ●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●● ●
●
●●
●●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●●
●●●
● ●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●●●●
●
●●
●●●●
●
●●
●
●
●
●●
●●
●●
●
●
●●●
●
●
●●●●
●
●
●
●
●●●
●●●
●● ●
●●●
●
●
●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●●●●●
●
●● ●
●
●
●●●
●
●●
●
●●
●
●
●●●
● ●●
●
●
●
●
●
●
●● ●
●●
●
●
●
●
●
●●●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●
● ●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●●
●
●●●
●
●
●
●●
●
●
●●
●●●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
● ●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
● ●
●
●
●
●
●●●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●●●
●
●
●●
●●
●●
●
●
●●
●●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●
● ●●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
● ●
●
●
●●
●●●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●●
●
●●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●●
●
● ●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
● ●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ● ●
●
●
●
●
●
●●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●●●
●
●
●
●
●
●
●●● ●
●
●●●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●●●●●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
● ●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●
●
●●
●
● ●●●
●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
● ●
●●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●●●
●●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
● ●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●●
●
●● ●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
● ●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●●●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●● ●
●
●
● ●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
● ●
●
●
●●
●
●●●●
●
●
●●
●
●
●●
●
●●
●
●●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●
● ●
●
●
●
●
●
●
●
●
●
●●● ●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●●●
●●●●
●
●
●
●
●●
●
●●
● ●
●
●●●
●●
●●
●●●
●
●
●
●●●●
●●
●●
●
●●
●
●
●
●
●
●●
●
●●●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ● ●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●● ●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
● ●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●●
●
●
●
●●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
● ●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
● ●●
●
●●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●●
●●
●
●
●
● ●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
● ●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
● ●
●
●
●
●
●
●
●●
●
●
●
● ●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●●●
●●
●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●● ●●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
● ●
● ●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
● ●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
● ●
●
●● ●
●
●
●
●
●
● ●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●
●
● ●●
●●●●●
●●
●● ●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
● ●
●
●
●●
●
●● ●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
● ●
●
● ●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ●
●●
●
●
●
●
●● ●●● ●
●●
●●
●● ●
●
●●● ●
●
●
●
●● ●
●
●
●
●●●● ●● ●
●●● ●●● ●●●
●●●●●●●●●● ●
●
●●●
●●
●
●
●●● ●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●●
●
●
●
●●●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●●
●●●●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
● ●
●
●●●
●
●
●
●●
●●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●●●
●
●
●●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●●●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●●
●●
●
●
● ●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●● ●
●
●
●
●
●
●●
●
●
●
●●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
● ●●●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●●●
●●●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●●●
●●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●●
●
●
●
●0.01 0.02 0.05 0.10
0.05
0.10
0.15
0.20
0.25
0.30
0.35
Standard Error
Log
Odd
s
Example Type 2 Diabetes (T2D) GWAS (Morris et al. 2012, Nat Gen)
• case/control (22,669 / 58,119)
• lots of T2D associated loci, but of small effect
(3371 SNPs shown)
• ?how to rank order?
log
odds
ratio
10 20 50 100 200 500 1000
0.05
0.10
0.20
0.50
set size N
prop
ortio
n of
set d
etec
ted
by R
NA
i
Example gene-set enrichment, RNAi (Hao et al. 2013, PLoS Comp Bio)
• 984 human genes linked to influenza-virus replication
• functional content measured against Gene Ontology (5719 sets)
• ?how to rank order?
Example
• 461 NBA players (2013-2014)
• free throw percentage
• ?how to rank order?
●●●●●●●●●●●●●
● ●● ●●●●● ●●● ●●● ●● ●● ● ●●●● ● ● ●●●●● ● ●● ●● ● ●● ●●● ● ●● ●●●●● ● ●● ●● ● ●●● ●●●● ●●● ●● ● ●●●● ●●●●●●● ●● ●● ●●● ●● ●●●● ●● ● ●● ●●● ● ●● ●●● ● ●●● ●●● ●●●●● ● ●●● ●●● ● ●●●● ●● ●●●● ● ●●●● ●●●●●● ●● ● ●● ●● ● ●● ● ●● ● ●●● ● ●●● ● ●●● ●●● ●●●● ●●● ●● ● ●● ●● ●● ● ●● ● ●● ●● ●●●● ●● ●● ●● ●● ●● ● ●●● ●● ●●●●●●●●● ●●●●● ●● ● ●● ●●● ●● ●● ●● ●● ● ●● ●● ●●●● ●●● ●● ●● ●●●● ● ● ●● ●●●●●● ●● ●●●●●●● ● ●●● ●● ● ●●● ●● ●●● ●●●●● ● ●● ●●● ● ● ●●●●●●●● ●●● ●● ● ●●●● ●● ●● ●●●● ●● ● ● ●● ●● ● ●●●● ●● ●●● ●●●● ●●● ●● ● ●●●●● ●● ● ●●●●●●● ● ●●● ●● ●●● ●● ●●●● ●● ●●●● ● ●● ●●
●●●●●●●●●●● ●● ● ●●● ●● ●
●●●●
●●
●
●●
●
●●●● ●●●
1 5 10 50 100 500 1000
0.0
0.2
0.4
0.6
0.8
1.0
# Free Throw Attempts
Free
Thr
ow P
erce
ntag
e
simulation
Xi
�2i✓i
N�✓i,�
2i
�
f = N(0, 1)
signals noise levels
measured signals
g = Gam(a, b)
units i = 1, 2, . . . , B
simulation
(X1,�21), (X2,�
22), · · · , (XB ,�
2B)data pairs:
ranking statistic: R1, R2, . . . , RB
aim: highly rank units with largest signals
p-value lead units by p-value are enriched for those with small variance
σ
0 1 2 3
p (�i| p-valuei p0.1 )
p (�i)
p-value lead units by p-value are enriched for those with small variance
σ
0 1 2 3
p (�i| p-valuei p0.1 )
p (�i)
same for q-value!
rank by “local” maximum likelihood estimate
other approaches
• estimated log odds ratio• proportion of gene set on gene list • free throw percentage
lead units by MLE are enriched for those with large variance
p (�i|Xi � x0.1 )
σ
0 1 2 3
p (�i)
local MLE
lead units by posterior mean are enriched for those with small variance
σ
0 1 2 3
p (�i)
p {�i|E(✓i|Xi,�i) � e0.1 }
local PM
Problems and solutions
We’ve found a generic empirical bayes ranking/selection method
Improved ranking and selection 13
Hall, P. and H. Miller (2010). Modeling the variability of rankings. The Annals of Statis-tics 38 (5), 2652–2677.
Hao, L., Q. He, Z. Wang, M. Craven, M. A. Newton, and P. Ahlquist (2013). Limitedagreement of independent rnai screens for virus-required host genes owes more to false-negative than false-positive factors. PLoS computational biology 9 (9), e1003235.
Jost, J. and X. Li-Jost (1998). Calculus of variations, Volume 64. Cambridge UniversityPress.
Kass, R. E. and A. E. Raftery (1995). Bayes factors. Journal of the American StatisticalAssociation 90 (430), pp. 773–795.
Kendziorski, C., M. Newton, H. Lan, and M. Gould (2003). On parametric empiricalbayes methods for comparing multiple groups using replicated gene expression profiles.Statistics in medicine 22 (24), 3899–3914.
Laird, N. M. and T. A. Louis (1989). Empirical bayes ranking methods. Journal of Educa-tional and Behavioral Statistics 14 (1), 29–46.
Lehmann, E. (1986). Testing statistical hypotheses (2nd ed.). Wiley series in probabilityand mathematical statistics: Probability and mathematical statistics. Wiley.
Leng, N., J. A. Dawson, J. A. Thomson, V. Ruotti, A. I. Rissman, B. M. Smits, J. D.Haag, M. N. Gould, R. M. Stewart, and C. Kendziorski (2013). Ebseq: an empiricalbayes hierarchical model for inference in rna-seq experiments. Bioinformatics 29 (8),1035–1043.
Lin, R., T. A. Louis, S. M. Paddock, and G. Ridgeway (2006). Loss function based rankingin two-stage, hierarchical models. Bayesian Analysis 1 (4), 915–946.
McCarthy, D. J. and G. K. Smyth (2009). Testing significance relative to a fold-changethreshold is a treat. Bioinformatics 25 (6), 765–771.
Morris, A. P., B. F. Voight, T. M. Teslovich, T. Ferreira, A. V. Segre, V. Steinthorsdottir,R. J. Strawbridge, H. Khan, H. Grallert, A. Mahajan, et al. (2012). Large-scale associationanalysis provides insights into the genetic architecture and pathophysiology of type 2diabetes. Nature genetics 44 (9), 981–990.
Niemi, J. (2010). Evaluating individual player contributions in basketball. In JSM Proceed-ings, Statistical Computing Section, Alexandria, VA, pp. 4914–4923. American StatisticalAssociation.
Noma, H., S. Matsui, T. Omori, and T. Sato (2010). Bayesian ranking and selection methodsusing hierarchical mixture models in microarray studies. Biostatistics 11 (2), 281–289.
Normand, S.-L. T., M. E. Glickman, and C. A. Gatsonis (1997). Statistical methods forprofiling providers of medical care: issues and applications. Journal of the AmericanStatistical Association 92 (439), 803–814.
Paddock, S. M. and T. A. Louis (2011). Percentile-based empirical distribution functionestimates for performance evaluation of healthcare providers. Journal of the Royal Sta-tistical Society: Series C (Applied Statistics) 60 (4), 575–589.
12 Henderson and Newton
5.3. Theorem 3By continuity, V!{t!!(!
2),!2} = "!. WLOG assume that #1 < #2, and note that by defini-tion as cumulative probabilities, V!1
(x,!2) ! V!2(x,!2) for any arguments. If, contrary to
the assertion of the theorem, t!!1(!2) = t!!2
(!2) = t, at some !2, then
V!2(t, !2)" V!1
(t, !2) = "!2" "!1
.
However, from the assumed condition, we have
V!2(t, !2)" V!1
(t, !2) =
! !2
!1
$V!(t, !2)
$#d# >
! !2
!1
d"!
d#d# = "!2
" "!1,
and thus a contradiction.Remark: Since by definition $V!(x,!2)/$# > 0, the condition $V!(x,!2)/$# > d"!/d#is automatically satisfied whenever "! is decreasing in #.
Acknowledgements
This research was supported in part by two grants from the US National Institutes of Health:R21 HG006568 and T32 GM074904. The authors thank Christina Kendziorski for criticalcomments on an earlier draft. Additional details on data analyses, threshold functions, andcomputation are provided in a Supplementary Material document. The R package rvaluesis available at http://www.stat.wisc.edu/~newton/.
References
Berger, J. O. and J. Deely (1988). A bayesian approach to ranking and selection of relatedmeans with alternatives to analysis-of-variance methodology. Journal of the AmericanStatistical Association 83 (402), 364–373.
Brijs, T., D. Karlis, F. Van den Bossche, and G. Wets (2007). A bayesian model for rankinghazardous road sites. Journal of the Royal Statistical Society: Series A (Statistics inSociety) 170 (4), 1001–1017.
Coelho, C. A. and J. T. Mexia (2007). On the distribution of the product and ratio ofindependent generalized gamma-ratio random variables. Sankhya: The Indian Journalof Statistics , 221–255.
de los Campos, G., J. M. Hickey, R. Pong-Wong, H. D. Daetwyler, and M. P. Calus (2013).Whole-genome regression and prediction methods applied to plant and animal breeding.Genetics 193 (2), 327–345.
Efron, B. (2010). Large-scale inference: empirical Bayes methods for estimation, testing,and prediction, Volume 1. Cambridge University Press.
Gelman, A., P. N. Price, et al. (1999). All maps of parameter estimates are misleading.Statistics in Medicine 18 (23), 3221–3234.
Gibbons, J. D., I. Olkin, and M. Sobel (1979). An introduction to ranking and selection.The American Statistician 33 (4), 185–195.
helpful observation • a ranking method corresponds to
a family of threshold functions:
T = {t↵ : ↵ 2 (0, 1)}
• unit i ranked in top if ↵ Xi � t↵(�2i )
helpful observation • a ranking method corresponds to
a family of threshold functions:
T = {t↵ : ↵ 2 (0, 1)}
• each one is a function t↵(�2)X
�2
• unit i ranked in top if ↵ Xi � t↵(�2i )
helpful observation • a ranking method corresponds to
a family of threshold functions:
T = {t↵ : ↵ 2 (0, 1)}
• each one is a function t↵(�2)X
�2
• size constraint: P�Xi � t↵(�
2i ) = ↵ (marginal!!)
• unit i ranked in top if ↵ Xi � t↵(�2i )
a. MLE b. p−value
c. posterior mean d. maximal agreement
T2D example rank by sweeping through the family
X
�2 �2
f=N(0,1)
14 Henderson and NewtonTable 1. Threshold functions associated with various ranking criteria, nor-mal/normal modelcriteria ranking variable threshold function t!(!
2)MLE Xi u!
PV H0 : "i = 0 Xi/!i u!!PV H0 : "i = c (Xi ! c)/!i c+ u!!PM Xi/(!
2i + 1) u!(!
2 + 1)
PER P ("i " "|Xi,!2i ) u!
!
(!2 + 1)(2!2 + 1)
BF 1(Xi > 0)P (Xi|"
2
i,#i !=0)
P (Xi|"2
i,#i=0)
"
!2(!2 + 1)#
u! + log ("2+1)"2
$
max agreement r-value "!(!2 + 1)! u!
!
!2(!2 + 1)
Pyeon, D., M. A. Newton, P. F. Lambert, J. A. Den Boon, S. Sengupta, C. J. Marsit,C. D. Woodworth, J. P. Connor, T. H. Haugen, E. M. Smith, et al. (2007). Fundamen-tal di!erences in cell cycle deregulation in human papillomavirus–positive and humanpapillomavirus–negative head/neck and cervical cancers. Cancer research 67 (10), 4605–4619.
Shen, W. and T. A. Louis (1998). Triple-goal estimates in two-stage hierarchical models.Journal of the Royal Statistical Society: Series B (Statistical Methodology) 60 (2), 455–471.
Smyth, G. K. et al. (2004). Linear models and empirical bayes methods for assessingdi!erential expression in microarray experiments. Stat Appl Genet Mol Biol 3 (1), 3.
Storey, J. D. (2003). The positive false discovery rate: A bayesian interpretation and theq-value. The Annals of Statistics 31 (6), pp. 2013–2035.
Wright, D. L., H. S. Stern, and N. Cressie (2003). Loss functions for estimation of extremawith an application to disease mapping. Canadian Journal of Statistics 31 (3), 251–266.
Xie, M., K. Singh, and C.-H. Zhang (2009). Confidence intervals for population ranks in thepresence of ties and near ties. Journal of the American Statistical Association 104 (486),775–788.
nota bene
Even though the unit-level parameters are unobserved, their distribution may be well estimated.
Kiefer Wolfowitz, 1956
P (✓i � ✓↵) =
Z 1
✓↵
f(✓) d✓ = ↵
✓↵We can estimate such that:
Agreement
P�Xi � t↵(�
2i ), ✓i � ✓↵
reported in top fraction
truly in top fraction
maximal agreement
Improved ranking and selection 3
g(!2). The empirical Bayesian uses the full data set to estimate the prior distributions f(")and g(!2); we ignore the estimation error at this level, and focus on ranking units withinthe estimated population.
Relative to a single unit i, Xi might be the maximum likelihood estimator of "i, and !i
that estimator’s standard error. The independence assumption may be reasonable if somecare has been taken in this local analysis, for example, by variance-stabilizing transforma-tion. Typically, the variance !2
i is estimated rather than known exactly, and we examinethis case in Section 5. There are a number of important examples where Xi is either dis-crete or multivariate and we also take up these extensions in Section 5. We consider firstthe continuous model, involving prior distributions and sampling distributions all havingdensities with respect to Lebesgue measure. **some regularity** The canonical samplingmodel within this class has Xi|"i,!2
i ! Normal("i,!2i ).
We make some headway by associating each ranking/selection procedure with a familyT of thresholding functions T = {t! : # " (0, 1)}. Each t! is a function t!(!2) havingthe interpretation that unit i is reported to be in the top # fraction of units if and only ifXi # t!(!2
i ). This interpretation is supported by the size constraint, namely, that marginalto all parameters and data,
P!
Xi # t!(!2
i )"
= # for all # " (0, 1) . (1)
Table 1 reports threshold functions associated with a variety of ranking methods in the nor-mal observation model, and under the extra condition that the prior f(") is Normal(µ, $2).Figure 2 illustrates four of these families in the T2D case study. Notionally, the linear rank-ing of units is obtained by sweeping through the family T , beginning with the smallest # atthe top of the graph. Clearly, distinct families of threshold functions can produce distinctrankings of the units, with the family’s shape revealing how it trades o! observed signal Xi
with measurement variance !2i to prioritize the leading units.
2.2. Thresholds via direct optimizationTable 1 and Figure 2 introduce a family T ! = {t!!} that is optimal in the continuous modelin the sense that for all # " (0, 1):
P!
Xi # t!!(!2
i ) , "i # "!"
# P!
Xi # t!(!2
i ) , "i # "!"
(2)
for any other family T = {t!} which also satisfies the size constraint (1). Here "! is the# upper quantile of the prior; that is P ("i # "!) = #. In other words, T ! maximizesagreement: the joint probability that unit i is placed in the top # fraction and its drivingparameter "i is in the top # fraction of the population, for all #. We emphasize thatthe probabilities in (2) cover the joint distribution of Xi,!2
i , "i, which respects both thesampling distribution of data local to unit i and the fluctuations of unit-specific parameters.A calculus-of-variations argument provides direct optimization of the joint probability in (2),subject to the size constraint, model regularity, smoothness of the threshold functions.
Theorem 1. In the continuous model, a necessary condition for the function t!! to beoptimal as in (2), within the class of continuously di!erentiable threshold functions, is thatit satisfies:
P!
"i # "!|Xi = t!!(!2),!2
i = !2"
= c! for all !2. (3)
Theorem 1: Under certain smoothness conditions,
{t⇤↵} is optimal if for all �2
maximal agreement
can solve directly if f=N(0,1)
14 Henderson and NewtonTable 1. Threshold functions associated with various ranking criteria, nor-mal/normal modelcriteria ranking variable threshold function t!(!
2)MLE Xi u!
PV H0 : "i = 0 Xi/!i u!!PV H0 : "i = c (Xi ! c)/!i c+ u!!PM Xi/(!
2i + 1) u!(!
2 + 1)
PER P ("i " "|Xi,!2i ) u!
!
(!2 + 1)(2!2 + 1)
BF 1(Xi > 0)P (Xi|"
2
i,#i !=0)
P (Xi|"2
i,#i=0)
"
!2(!2 + 1)#
u! + log ("2+1)"2
$
max agreement r-value "!(!2 + 1)! u!
!
!2(!2 + 1)
Pyeon, D., M. A. Newton, P. F. Lambert, J. A. Den Boon, S. Sengupta, C. J. Marsit,C. D. Woodworth, J. P. Connor, T. H. Haugen, E. M. Smith, et al. (2007). Fundamen-tal di!erences in cell cycle deregulation in human papillomavirus–positive and humanpapillomavirus–negative head/neck and cervical cancers. Cancer research 67 (10), 4605–4619.
Shen, W. and T. A. Louis (1998). Triple-goal estimates in two-stage hierarchical models.Journal of the Royal Statistical Society: Series B (Statistical Methodology) 60 (2), 455–471.
Smyth, G. K. et al. (2004). Linear models and empirical bayes methods for assessingdi!erential expression in microarray experiments. Stat Appl Genet Mol Biol 3 (1), 3.
Storey, J. D. (2003). The positive false discovery rate: A bayesian interpretation and theq-value. The Annals of Statistics 31 (6), pp. 2013–2035.
Wright, D. L., H. S. Stern, and N. Cressie (2003). Loss functions for estimation of extremawith an application to disease mapping. Canadian Journal of Statistics 31 (3), 251–266.
Xie, M., K. Singh, and C.-H. Zhang (2009). Confidence intervals for population ranks in thepresence of ties and near ties. Journal of the American Statistical Association 104 (486),775–788.
maximal agreementa. MLE b. p−value
c. posterior mean d. maximal agreement
X
�2
maximal agreement
σ
0 1 2 3
p (�i)
p��i|Xi � t⇤0.1(�
2i )
pretty close, considering that we’re targeting agreement not the artifact
local tail probability V↵(Xi,�
2i ) = P
�✓i � ✓↵|Xi,�
2i
�
4 Henderson and Newton
All observations coincident with the graph of a given optimal threshold curve thus havea common posterior probability c! that their unit-specific parameters exceed the quantile!! associated with that curve. In the normal model for Xi and the normal prior f(!),the optimal threshold function (Figure 2d) is readily extracted from (3). Working on astandardized scale without loss of generality (µ = 0 and "2 = 1), the local posterior for !iis normal with mean Xi/(#2
i + 1) and variance #2i /(#
2i + 1). Thus,
t!!(#2) = !!(#
2 + 1)! u!
!
#2(#2 + 1), (4)
where u! is determined by the size constraint (1). Indeed u! is a!ected by the distributiong(#2), since it is defined implicitly through the constraint-induced equation:
1! $ =
"
"
0
"
#
!! ! u!
$
#2
1 + #2
%
g(#2) d#2 (5)
where " is the standard normal cumulative distribution.**hard to discover explicit thresholds this way****put in the list-conditional variance story an a second plot series (maybe supplemen-
tary) showing how the optimal thresholds are least sigma biased **Some comments on the threshold functions in Table 1 are warranted... **put in story
about posterior expected ranks as approximate threshold function **
2.3. Posterior tail probabilities and ranking variablesInsight into the structure of optimal thresholds comes by further examining their relation-ship to local posterior tail probabilities: V!(Xi,#2
i ) = P (!i " !!|Xi,#2i ).
Theorem 2. Suppose that for every $ # (0, 1) there exists %! such that
P&
V!(Xi,#2
i ) $ %!
'
= $, (6)
and furthermore suppose that V!(x,#2) is right-continuous and non-decreasing in x forevery fixed $ and #2. Then the family of thresholds
t!!(#2) = inf{x : V!(x,#
2) " %!} (7)
satisfies the size constraint (1) and is optimal in the sense of (2).
A family of threshold functions is a device to think about converting observations intorankings (i.e. by sweeping through the family). Indeed, the index $ associated with thethreshold curve on which data point (Xi,#2
i ) lands may be viewed as a ranking variable.Computation of the ranking variable amounts to solving the inversion Xi = t!(#2
i ) for $,which is well defined under suitable conditions. **maybe refer to optimal normal modelproof** For the thresholds (7) that are optimal for agreement between the inferred top listand the true top list, we have
Theorem 3. Suppose that V!(x,#2) is continuous in x for every $ and #2, di!erentiablein $ for every x and #2, and further that %! is di!erentiable in $. If &V!(x,#2)/&$ >d%!/d$ for every ($, x,#2), then for any $1 %= $2 it holds that for all #2, t!!1
(#2) %= t!!2(#2).
Theorem 2: Under certain conditions, the optimal family satisfies: {t⇤↵}
where,
P�V↵(Xi,�
2i ) � �↵
= ↵
next step
from thresholds back to ranking variables
ranking variables ri(Xi,�
2i ) = inf{↵ : Xi � t↵(�
2i )}a. MLE b. p−value
c. posterior mean d. maximal agreement
ranking variables ri(Xi,�
2i ) = inf{↵ : Xi � t↵(�
2i )}a. MLE b. p−value
c. posterior mean d. maximal agreement
Improved ranking and selection 5
Remarks...Like for all the families shown in Figure 2 and Table 1, the optimal thresholds do not
touch or cross under the conditions of Theorem 3, and they conform to our intuition abouthow ranking procedures might be constructed from threshold functions. Thus, we mayreasonably introduce a special ranking variable that inverts the optimal threshold. For theith unit, we define the r-value:
ri(Xi,!2
i ) = inf!
" : V!(Xi,!2
i ) ! #!
"
. (8)
Essentially, unit i is placed by its r-value at position r (a percentile, measured from thetop) if when ranking the units by P ($i ! $r|Xi,!2
i ), it also happens to land at positionr. Further, the top " fraction of units by r-value has higher overlap with the true top "fraction of units than could be obtained by any other ranking procedure, in the sense of (2).
It is worth recognizing that these findings go beyond what is already known about theuse of V!(Xi,!2
i ) to optimally rank units. **Louis/Lehmann** ; also you get di!erentresults when ranking by di!erent "... *** **example of where your in the top 5
3. Connections
3.1. Connection to Bayes ruleThe proposed ranking procedure is a kind of Bayes rule for a percentile when consideringmultiple loss functions and a distributional constraint. To see how, introduce a collectionof loss functions
L!(a, $i) = 1" 1 (a # ", $i ! $!)
where action a is a percentile value in (0, 1), " $ (0, 1) indexes the collection, and again$! = F!1(1"") is a quantile in the population of interest. Specifically, no ""loss occurs ifthe inferred upper percentile a and the actual upper percentile 1"F ($i) both are less than". The marginal (pre-posterior) Bayes risk of rule %(Xi,!2
i ) is
risk! = 1" P!
%(Xi,!2
i ) # " , $ ! $!"
, (9)
which is one minus the agreement (2). In the absence of other considerations, the Bayesrule for loss L! degerenates to %(Xi,!2
i ) = 0. Degeneration is avoided if we enforce on theestimated percentile the additional structure that it share with the true percentile 1"F ($i)the property of being uniformly distributed over the population of units. Such a constrainedBayes rule then minimizes the modified objective function:
risk! + &!P!
%(Xi,!2i ) # "
"
where &! is chosen to enforce the size constraint P!
%(Xi,!2i ) # "
"
= ".The constrained Bayes rule is computed conditionally, per observed (Xi,!2
i ), by mini-mizing the (modified) posterior expected loss (PEL)
PEL! = 1" P!
%(Xi,!2
i ) # " , $i ! $!#
#Xi,!2
i
"
+ &!1!
%(Xi,!2
i ) # ""
(10)
=
$
1" V!(Xi,!2i ) + &! if %(Xi,!2
i ) # "1 if %(Xi,!2
i ) > "
r-value
r-value
smallest such that unit i in top when ranking by: V↵(Xi,�
2i ) = P
�✓i � ✓↵|Xi,�
2i
�↵ ↵
general form
6 Henderson and Newton
without extra conditions there could be distinct !1 != !2 such that t!1("2
i ) = t!2("2
i ) forsome "2
i . For the thresholds (7) that are optimal for agreement between the inferred toplist and the true top list, we find the following.
Theorem 3. Suppose that V!(x,"2) is continuous in x for every ! and "2, di!erentiablein ! for every x and "2, and further that #! is di!erentiable in !. If $V!(x,"2)/$! >d#!/d! for every (!, x,"2), then for any !1 != !2 it holds that for all "2, t!!1
("2) != t!!2("2).
Like for all the families shown in Figure 3 and Table 1, the optimal thresholds do not touchor cross under the conditions of Theorem 3, and they conform to our intuition about howranking procedures might be constructed from threshold functions. Thus, we introduce aspecial ranking variable that inverts the optimal threshold. For the ith unit, we define ther-value:
r(Xi,"2
i ) = inf!
! : V!(Xi,"2
i ) " #!
"
. (8)
Essentially, unit i is placed by its r-value at position ! (a relative rank, measured from thetop) if when ranking the units by P (%i " %!|Xi,"2
i ), it also happens to land at position!. Further, the top ! fraction of units by r-value has higher overlap with the true top !fraction of units than could be obtained by any other ranking procedure, in the sense of (2).
It is worth recognizing that these findings go beyond what has been reported aboutthe use of the conditional tail probability V!(Xi,"2
i ) to rank units. Classical theory onoptimal selection establishes the role of this conditional tail probability in maximizing anexceedance probability within the selected sample (e.g., Lehmann, 1986, pages 117-118).Also, the conditional tail probability has been used for ranking (e.g., Normand et al., 1997;Niemi, 2010), and is closely related to a Bayes optimal ranking under a certain loss function(Lin et al., 2006). The critical di!erence with the proposed ranking is in the role of theindex !. Conceptually, we imagine ranking the units by V!(Xi,"2
i ) separately for all possibleindices ! (not just a pre-specified one); then the r-value for unit i is the smallest index !such that unit i is placed in the top ! fraction by that ranking. By aiming to maximizeagreement at all list sizes, the proposed method does not require a pre-specified exceedancelevel to generate its ranking.
2.4. More generalityThe r-value concept makes sense in various elaborations of the the measurement modelfrom Section 2.1. We retain univariate parameters of interest {%i} varying according toa distribution F , but we allow data Di on each unit to take more general forms than the(Xi,"2
i ) pair structure. We also retain the assumption of mutual independence among units,though extensions could be developed in cases where posterior computation is feasible. Inseeking units with largest %i, the critical quantity is the local exceedance probability:
V!(Di) = P (%i " %!|Di)
for ! # (0, 1) and for upper quantiles %! of the marginal distribution F : i.e., %! = F"1(1$!). Induced by the marginal distribution of Di, the tail probability V!(Di) has cumulativedistribution function H!(v), and from it we obtain the upper quantile: #! = H"1
! (1 $ !).Then by analogy to (8), the r-value is defined:
r(Di) = inf {! : V!(Di) " #!} .
6 Henderson and Newton
without extra conditions there could be distinct !1 != !2 such that t!1("2
i ) = t!2("2
i ) forsome "2
i . For the thresholds (7) that are optimal for agreement between the inferred toplist and the true top list, we find the following.
Theorem 3. Suppose that V!(x,"2) is continuous in x for every ! and "2, di!erentiablein ! for every x and "2, and further that #! is di!erentiable in !. If $V!(x,"2)/$! >d#!/d! for every (!, x,"2), then for any !1 != !2 it holds that for all "2, t!!1
("2) != t!!2("2).
Like for all the families shown in Figure 3 and Table 1, the optimal thresholds do not touchor cross under the conditions of Theorem 3, and they conform to our intuition about howranking procedures might be constructed from threshold functions. Thus, we introduce aspecial ranking variable that inverts the optimal threshold. For the ith unit, we define ther-value:
r(Xi,"2
i ) = inf!
! : V!(Xi,"2
i ) " #!
"
. (8)
Essentially, unit i is placed by its r-value at position ! (a relative rank, measured from thetop) if when ranking the units by P (%i " %!|Xi,"2
i ), it also happens to land at position!. Further, the top ! fraction of units by r-value has higher overlap with the true top !fraction of units than could be obtained by any other ranking procedure, in the sense of (2).
It is worth recognizing that these findings go beyond what has been reported aboutthe use of the conditional tail probability V!(Xi,"2
i ) to rank units. Classical theory onoptimal selection establishes the role of this conditional tail probability in maximizing anexceedance probability within the selected sample (e.g., Lehmann, 1986, pages 117-118).Also, the conditional tail probability has been used for ranking (e.g., Normand et al., 1997;Niemi, 2010), and is closely related to a Bayes optimal ranking under a certain loss function(Lin et al., 2006). The critical di!erence with the proposed ranking is in the role of theindex !. Conceptually, we imagine ranking the units by V!(Xi,"2
i ) separately for all possibleindices ! (not just a pre-specified one); then the r-value for unit i is the smallest index !such that unit i is placed in the top ! fraction by that ranking. By aiming to maximizeagreement at all list sizes, the proposed method does not require a pre-specified exceedancelevel to generate its ranking.
2.4. More generalityThe r-value concept makes sense in various elaborations of the the measurement modelfrom Section 2.1. We retain univariate parameters of interest {%i} varying according toa distribution F , but we allow data Di on each unit to take more general forms than the(Xi,"2
i ) pair structure. We also retain the assumption of mutual independence among units,though extensions could be developed in cases where posterior computation is feasible. Inseeking units with largest %i, the critical quantity is the local exceedance probability:
V!(Di) = P (%i " %!|Di)
for ! # (0, 1) and for upper quantiles %! of the marginal distribution F : i.e., %! = F"1(1$!). Induced by the marginal distribution of Di, the tail probability V!(Di) has cumulativedistribution function H!(v), and from it we obtain the upper quantile: #! = H"1
! (1 $ !).Then by analogy to (8), the r-value is defined:
r(Di) = inf {! : V!(Di) " #!} .
6 Henderson and Newton
without extra conditions there could be distinct !1 != !2 such that t!1("2
i ) = t!2("2
i ) forsome "2
i . For the thresholds (7) that are optimal for agreement between the inferred toplist and the true top list, we find the following.
Theorem 3. Suppose that V!(x,"2) is continuous in x for every ! and "2, di!erentiablein ! for every x and "2, and further that #! is di!erentiable in !. If $V!(x,"2)/$! >d#!/d! for every (!, x,"2), then for any !1 != !2 it holds that for all "2, t!!1
("2) != t!!2("2).
Like for all the families shown in Figure 3 and Table 1, the optimal thresholds do not touchor cross under the conditions of Theorem 3, and they conform to our intuition about howranking procedures might be constructed from threshold functions. Thus, we introduce aspecial ranking variable that inverts the optimal threshold. For the ith unit, we define ther-value:
r(Xi,"2
i ) = inf!
! : V!(Xi,"2
i ) " #!
"
. (8)
Essentially, unit i is placed by its r-value at position ! (a relative rank, measured from thetop) if when ranking the units by P (%i " %!|Xi,"2
i ), it also happens to land at position!. Further, the top ! fraction of units by r-value has higher overlap with the true top !fraction of units than could be obtained by any other ranking procedure, in the sense of (2).
It is worth recognizing that these findings go beyond what has been reported aboutthe use of the conditional tail probability V!(Xi,"2
i ) to rank units. Classical theory onoptimal selection establishes the role of this conditional tail probability in maximizing anexceedance probability within the selected sample (e.g., Lehmann, 1986, pages 117-118).Also, the conditional tail probability has been used for ranking (e.g., Normand et al., 1997;Niemi, 2010), and is closely related to a Bayes optimal ranking under a certain loss function(Lin et al., 2006). The critical di!erence with the proposed ranking is in the role of theindex !. Conceptually, we imagine ranking the units by V!(Xi,"2
i ) separately for all possibleindices ! (not just a pre-specified one); then the r-value for unit i is the smallest index !such that unit i is placed in the top ! fraction by that ranking. By aiming to maximizeagreement at all list sizes, the proposed method does not require a pre-specified exceedancelevel to generate its ranking.
2.4. More generalityThe r-value concept makes sense in various elaborations of the the measurement modelfrom Section 2.1. We retain univariate parameters of interest {%i} varying according toa distribution F , but we allow data Di on each unit to take more general forms than the(Xi,"2
i ) pair structure. We also retain the assumption of mutual independence among units,though extensions could be developed in cases where posterior computation is feasible. Inseeking units with largest %i, the critical quantity is the local exceedance probability:
V!(Di) = P (%i " %!|Di)
for ! # (0, 1) and for upper quantiles %! of the marginal distribution F : i.e., %! = F"1(1$!). Induced by the marginal distribution of Di, the tail probability V!(Di) has cumulativedistribution function H!(v), and from it we obtain the upper quantile: #! = H"1
! (1 $ !).Then by analogy to (8), the r-value is defined:
r(Di) = inf {! : V!(Di) " #!} .
data for unit i Di
local posterior tail probability
marginal upper quantile
r-value
next step
how to calculate r-values
NBA Binomial likelihoode.g., Beta prior/posteriors
Free Throw Ability
dens
ity
0.2 0.4 0.6 0.8 1.0
05
1015
2535
NBA Binomial likelihoode.g., Beta prior/posteriors
Free Throw Ability
dens
ity
0.2 0.4 0.6 0.8 1.0
05
1015
2535
↵
✓↵
e.g., NBA
α
exce
edan
ce p
roba
bilit
y
0.002 0.005 0.010 0.020 0.050 0.100 0.200 0.500 1.0000
0.05
0.1
0.2
0.4
0.8
1
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●
●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●
●●●●
●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●
●●●●●●●
●●●●●●●●●●●●●●●●●●
●●●●●●●●●
●●●●●●●●●●●●●●●●
●●●●●●
●●●●●●●●●
●●●●●●●●●●●●●●●
●
●
●
P( θi ≥ θα | Di )two examplesempirical quantileλαr−value
DRay.Allen = 105 116
DLeBron.James = 439 585
Top 10
player Free throw % r-value post. mean
qual. rank
FTP rank
PM rank
RV rank
Brian Roberts 125/133 94.0 0.002 91.3 1 17 1 1 Ryan Anderson 59/62 95.2 0.003 89.8 15 2 2 Danny Granger 63/67 94.0 0.005 89.3 16 3 3 Kyle Korver 87/94 92.6 0.008 89.2 19 4 4 Mike Harris 26/27 96.3 0.010 86.6 14 15 5 J.J. Redick 97/106 91.5 0.011 88.6 22 6 6 Ray Allen 105/116 90.5 0.016 88.0 25 8 7 Mike Muscala 14/14 100.0 0.017 84.4 7 34 8 Dirk Nowitzki 338/376 89.9 0.018 89.1 2 30 5 9 Trey Burke 102/113 90.3 0.018 87.7 28 9 10 !
Predictive accuracy
●
●
●● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
5 10 15 20 25
0.00
0.05
0.10
0.15
0.20
0.25
0.30
t = rank from top
E[ s
imila
rity_
t{ R
anks
(thet
a) ,
Ran
ks.h
at[m
idse
ason
] } |
com
plet
e se
ason
]
●
● ● ●
●
●
●
●●
●●
●●
●●
● ●●
●●
●●
●● ●
● ● ●
● ● ● ● ● ● ● ● ● ● ●●
●●
●●
● ● ● ● ● ● ● ● ● ● ● ●
r−valueposterior meanMLE
T2D
Improved ranking and selection 17
a. MLE b. p−value
c. posterior mean d. maximal agreement
Fig. 3. Threshold functions, T2D example, data and axes as in Fig 1: Calculations use an inverse-gamma model for !2. Forty two threshold functions are shown, ranging in " values from a smallpositive value (red) just including the first data point up to " = 0.10 (blue). (Most data points aretruncated by the plot, as in Fig 1; also, the grid is uniform on the scale of log2[! log2(")].) Unitsassociated with a smaller " (i.e., more red) are ranked more highly by the given ranking method.Two units landing on the same curve would be ranked in the same position.
0 1 2 3 4 50
12
34
56
σ2
x
0 1 2 3 4 5
01
23
45
6
σ2
x
�2 ⇠ Exp(1)
kick-up
N[ k]
MLE
O
O
O
O
O
O
O
O O
O
O
OO
O
O
O O
O
OO
O
O
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
O
O
OO
OO
OO
O
O
O O
O
O
O
OOO
O
OO
O
OOO
O OO
O
O
O OOO
O
O
OOO
O
O
O
O
OO
O
O
O
OO
OO O
OO
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O
OOO O
O
OO
O
OO
O
O
O
O
OO
OO
OO
OO
O
OO O
O
OO
OO
O
O
O
O
O
O
OOO
OO
O
OO
OO
O
O
O
OO
O
OO O
OO
O
O
O
O
O
O
OO
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OOO
OO
O
O
O
OO
OOO
O O
O
O
O
OOOO
OO
O
OO
O
O
O
O
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O OO
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O O
O
O
O
O
O
O
O
O
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
OOO
O
O
O
O
O
O
O O
OOO
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O
O
O
OOO
O
OO
O
O
OOO
O
O O
O
O
O
O
O
OO
O
O
O O
O
OOOO
O
O
O O
OO
O
O
OO
O
O
O
O
OO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OOO
O
O
O
O
OO
O
O
O
O
OO
O
O
OO
O
O
O
OO
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
OOO
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OO
O
O
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O
O
O
O O
OOO
O
OO
O
O
OO
O
O
OO
O
O
O
O
OO
O
O
OO
O
O
O
O
OO
O
O
OO
O
O O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OO
O
O
O
OO
O
O
O
O
O
O
O
O
O
OO
OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
OO
OO
O
O
O
O
O
O
O
O
O
O
O
O
OOO
O
O
OO
OO
O
O
O
O
O
OO
O
O
O
OO
O
O
O
O
O
OOO
O
O
OOO
O
O
O
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O O
O
O
O
OO
O
O
OO
O
O
O
O
O
OO
O
O O
OO
O
OO
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O OO
O
OOO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O OO
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O O
O
O
O
O
OO
OO
O
O
O
O
OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OOO
O
O
OO
OO O
O
O
O
O
O
O
O
O
O
O
O
O
O
O O
O
O
O
O
O
O
O OO
O
O
O
OO
O
O
O
OO
O
O O
O
OO
O
O
O
OO
O
OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O O
O
O
O
OO
O
O
O
O
O
O OO
O
O
OOO
OO
O
OO
O
OO
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
OO O
O
O
O O
OO
O
O
O
OO
O
O
OO
O
O
O
O
O
OO
O
O
OO
O
O
O
OO
O
OO
O
O
O
O
O O
O
OO
O
O
O
O
O
O
OO
O
OOO
O
O
O
O
O
OO
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OOOO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O O
O
OO
OO
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
OOO
O
OO
O
OO
O
O
O
O
O
O
O
O
OO
O
O
OO
OO
O
O
O
O
O
O
O
OO
OO
OOO
O
O
O
OO
O
OO
O
O
OO
O
O
OOO
O
O
OO
OO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O O
O
O
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O O
OO
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O OO
O
O
O
OO
OO
OO
O
O
OO
O
O
O
O
O
O
O
OO
O
OO
O
O
OO
O
O
O
O
O
O
O
OO
OOO
O
O
O
O
O
OO
O
O
O O
OO
O
O
O
O
O
O
O
O
O O
O
O
O
OO
OO
O
O
O
O
O
O
O
OOO
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OO
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OO
O
O
O
O
O
O
O
O
OO
O O
O
O
OO
O
O
O
O
OO
O
O
OOOO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
OOOO OO
O
O
O
O
OO
OO
O
O
O
OO
O
OO
O
OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
OOO
OO
OO
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OO
OO
O
O
O
O
OO
O OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
OO
O
OO
O
O
O
O
OO
O
O
O
OO
O
O
O
O
O
OO
O
O
O
OO
O
O
OO
O
O
O
OO
O O
O
OO
O
O
O
O
OO
O
O
O
O
O
O O
O
O
O
OOO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
OOOOO
O
O
O
O
O
OO
O OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
OO
O
O
O
OO
O
O
O
O
O
O
OO
O
OO
O
O
O
O
O
O
O
O
O
O
O OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
O
OO
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
O
O O
O
O
OO
O
O O
O
O
O
O
O
OOO
OO
O
O
O
O
O
O
O
OO
O
O
O
OO
O
O
O
O
O
O
O
OO
OO
O
O
O
O
OOO
O
O
OO OO
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O
OO
O
O
O
O
O
O
O
OOO
O
O
O
O
O
OO
O
OO
O
O
OO
OO
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
OO
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
OO
O
O
OOOO
O
O
O
O
OO
O
OO
O
O
O
O
O
O
O
OO
O
O
OO
O
O
OO
OOO O
O
OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
OO O
O
OOO O
O
O
OO
O
O
O
OO O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
OOO
O
O
OO
OO
O
O
O
O
O
OOO
O
O
O
O
O
OO
O
O
O
O
O
O
O O
O
O
O
O
OO
O
O
OOO
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
OO
O
O
O
O O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
OOO
O
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O
OO
O
O
O
OO
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O O
O
O
O
O
O
O
O
O
OOO
O
O
O
O
OOO
O
O
OO
OOOOOOO
OO
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
OO
OO
OO
O
O
OO
O
O
O
O
OO
O
OO
O
OO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O O
O O
OO
OO
O OOO
O
O
O
O
OO
O
OOO
OO
OO
O
O
O
O
O
O
O
OO
O
O
OO
O
O
O
O
O
OO
O
O
O O
OO
O
OO
O
O
OO
O
OO
O
O
O
O
O
O
O
O
O OO
O
OO
OO
O
O
OOOO
OO
O
O
OO
O
OO
O
O
O
O
O
O
O
O
OO
O
O
O
O
O O
O
O
O
O
O
O
O O
O
OO
O
O
O
O
O
OO O
O
O
O
O
O
OO
O
O
OO
O
O
O
OO
O
O
O
OO
O
O
O
O
OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
OOO
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
O
O
O
O
O
O
O
OO
O
O
O
OO
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O O
O
O
O
OOO
O
O
O
O
OO
OOO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O O
O
OO
O
OO
O
O
O
O
O
OO
O
O
OO
OOO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
O
OO
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OOO
O
O
OO
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
OOO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O
OO
O
O
OO
O
O
O
OO
O
O
O
O
O
O
O
O
OO
O
O
OO
OO
O
O
O
O
O
O
O
O O
O
O
OO
OO
O
O
O
O
O
O O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
O
O O
OO
O
O
O
O
OO
OO
OO
O
O
O
O
OO
OO
O
O
O
O
O
O O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O OO
O
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
OOO
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
OO
O
OO
O
O
O
O
OO
O
OO
O
O
OO
O
O
O
O
O
O
O
O
OO O
O
O
O
OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O O
OO
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
OOO
OO
O
OO
O
O
O
OO
O
O
O
O
O
O
OO
O
OO
O
O
OOO
OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
OO
O
O
O
O
O
OO
O
O
O
OO
O
OO
O
O
O
O
O
OO
O
O
OO
O
O
OO
O
O
O
O
O
OO
O
O
OO
O
O
OO
OO
O
O
O
OO
O
O
OO
OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
O
O
O
O
OO
O
O O
OOO
O
O
O
O
OO
O
OO
O
O
O
O
OO
O
OOOOOOO
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O OO
O
O
O
OO
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
O O
O
O
O
O
O
O
O
O
O
O
OO
O
O
OOO
O
O
OO
O
O
O OO
O
O
OO
O
OO
O
OO
O
OOO
O
O
O
O
O
O
O
O
OO
O O
O
O
O
O
O
O
O
O
OO
O O
O
O
OO
O
O
O
O
O
O
O
O
O
O
MLEN[ k]
MLE
O
O
O
O
O
O
O
O O
O
O
OO
O
O
O O
O
OO
O
O
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
O
O
OO
OO
OO
O
O
O O
O
O
O
OOO
O
OO
O
OOO
O OO
O
O
O OOO
O
O
OOO
O
O
O
O
OO
O
O
O
OO
OO O
OO
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O
OOO O
O
OO
O
OO
O
O
O
O
OO
OO
OO
OO
O
OO O
O
OO
OO
O
O
O
O
O
O
OOO
OO
O
OO
OO
O
O
O
OO
O
OO O
OO
O
O
O
O
O
O
OO
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OOO
OO
O
O
O
OO
OOO
O O
O
O
O
OOOO
OO
O
OO
O
O
O
O
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O OO
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O O
O
O
O
O
O
O
O
O
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
OOO
O
O
O
O
O
O
O O
OOO
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O
O
O
OOO
O
OO
O
O
OOO
O
O O
O
O
O
O
O
OO
O
O
O O
O
OOOO
O
O
O O
OO
O
O
OO
O
O
O
O
OO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OOO
O
O
O
O
OO
O
O
O
O
OO
O
O
OO
O
O
O
OO
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
OOO
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OO
O
O
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O
O
O
O O
OOO
O
OO
O
O
OO
O
O
OO
O
O
O
O
OO
O
O
OO
O
O
O
O
OO
O
O
OO
O
O O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OO
O
O
O
OO
O
O
O
O
O
O
O
O
O
OO
OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
OO
OO
O
O
O
O
O
O
O
O
O
O
O
O
OOO
O
O
OO
OO
O
O
O
O
O
OO
O
O
O
OO
O
O
O
O
O
OOO
O
O
OOO
O
O
O
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O O
O
O
O
OO
O
O
OO
O
O
O
O
O
OO
O
O O
OO
O
OO
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O OO
O
OOO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O OO
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O O
O
O
O
O
OO
OO
O
O
O
O
OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OOO
O
O
OO
OO O
O
O
O
O
O
O
O
O
O
O
O
O
O
O O
O
O
O
O
O
O
O OO
O
O
O
OO
O
O
O
OO
O
O O
O
OO
O
O
O
OO
O
OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O O
O
O
O
OO
O
O
O
O
O
O OO
O
O
OOO
OO
O
OO
O
OO
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
OO O
O
O
O O
OO
O
O
O
OO
O
O
OO
O
O
O
O
O
OO
O
O
OO
O
O
O
OO
O
OO
O
O
O
O
O O
O
OO
O
O
O
O
O
O
OO
O
OOO
O
O
O
O
O
OO
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OOOO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O O
O
OO
OO
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
OOO
O
OO
O
OO
O
O
O
O
O
O
O
O
OO
O
O
OO
OO
O
O
O
O
O
O
O
OO
OO
OOO
O
O
O
OO
O
OO
O
O
OO
O
O
OOO
O
O
OO
OO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O O
O
O
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O O
OO
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O OO
O
O
O
OO
OO
OO
O
O
OO
O
O
O
O
O
O
O
OO
O
OO
O
O
OO
O
O
O
O
O
O
O
OO
OOO
O
O
O
O
O
OO
O
O
O O
OO
O
O
O
O
O
O
O
O
O O
O
O
O
OO
OO
O
O
O
O
O
O
O
OOO
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OO
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OO
O
O
O
O
O
O
O
O
OO
O O
O
O
OO
O
O
O
O
OO
O
O
OOOO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
OOOO OO
O
O
O
O
OO
OO
O
O
O
OO
O
OO
O
OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
OOO
OO
OO
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OO
OO
O
O
O
O
OO
O OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
OO
O
OO
O
O
O
O
OO
O
O
O
OO
O
O
O
O
O
OO
O
O
O
OO
O
O
OO
O
O
O
OO
O O
O
OO
O
O
O
O
OO
O
O
O
O
O
O O
O
O
O
OOO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
OOOOO
O
O
O
O
O
OO
O OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
OO
O
O
O
OO
O
O
O
O
O
O
OO
O
OO
O
O
O
O
O
O
O
O
O
O
O OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
O
OO
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
O
O O
O
O
OO
O
O O
O
O
O
O
O
OOO
OO
O
O
O
O
O
O
O
OO
O
O
O
OO
O
O
O
O
O
O
O
OO
OO
O
O
O
O
OOO
O
O
OO OO
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O
OO
O
O
O
O
O
O
O
OOO
O
O
O
O
O
OO
O
OO
O
O
OO
OO
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
OO
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
OO
O
O
OOOO
O
O
O
O
OO
O
OO
O
O
O
O
O
O
O
OO
O
O
OO
O
O
OO
OOO O
O
OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
OO O
O
OOO O
O
O
OO
O
O
O
OO O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
OOO
O
O
OO
OO
O
O
O
O
O
OOO
O
O
O
O
O
OO
O
O
O
O
O
O
O O
O
O
O
O
OO
O
O
OOO
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
OO
O
O
O
O O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
OOO
O
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O
OO
O
O
O
OO
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O O
O
O
O
O
O
O
O
O
OOO
O
O
O
O
OOO
O
O
OO
OOOOOOO
OO
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
OO
OO
OO
O
O
OO
O
O
O
O
OO
O
OO
O
OO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O O
O O
OO
OO
O OOO
O
O
O
O
OO
O
OOO
OO
OO
O
O
O
O
O
O
O
OO
O
O
OO
O
O
O
O
O
OO
O
O
O O
OO
O
OO
O
O
OO
O
OO
O
O
O
O
O
O
O
O
O OO
O
OO
OO
O
O
OOOO
OO
O
O
OO
O
OO
O
O
O
O
O
O
O
O
OO
O
O
O
O
O O
O
O
O
O
O
O
O O
O
OO
O
O
O
O
O
OO O
O
O
O
O
O
OO
O
O
OO
O
O
O
OO
O
O
O
OO
O
O
O
O
OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
OOO
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
O
O
O
O
O
O
O
OO
O
O
O
OO
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O O
O
O
O
OOO
O
O
O
O
OO
OOO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O O
O
OO
O
OO
O
O
O
O
O
OO
O
O
OO
OOO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
O
OO
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OOO
O
O
OO
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
OOO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O
OO
O
O
OO
O
O
O
OO
O
O
O
O
O
O
O
O
OO
O
O
OO
OO
O
O
O
O
O
O
O
O O
O
O
OO
OO
O
O
O
O
O
O O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
O
O O
OO
O
O
O
O
OO
OO
OO
O
O
O
O
OO
OO
O
O
O
O
O
O O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O OO
O
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
OOO
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
OO
O
OO
O
O
O
O
OO
O
OO
O
O
OO
O
O
O
O
O
O
O
O
OO O
O
O
O
OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O O
OO
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
OOO
OO
O
OO
O
O
O
OO
O
O
O
O
O
O
OO
O
OO
O
O
OOO
OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
OO
O
O
O
O
O
OO
O
O
O
OO
O
OO
O
O
O
O
O
OO
O
O
OO
O
O
OO
O
O
O
O
O
OO
O
O
OO
O
O
OO
OO
O
O
O
OO
O
O
OO
OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
O
O
O
O
OO
O
O O
OOO
O
O
O
O
OO
O
OO
O
O
O
O
OO
O
OOOOOOO
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O OO
O
O
O
OO
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
O O
O
O
O
O
O
O
O
O
O
O
OO
O
O
OOO
O
O
OO
O
O
O OO
O
O
OO
O
OO
O
OO
O
OOO
O
O
O
O
O
O
O
O
OO
O O
O
O
O
O
O
O
O
O
OO
O O
O
O
OO
O
O
O
O
O
O
O
O
O
O
p−value
N[ k]
MLE
O
O
O
O
O
O
O
O O
O
O
OO
O
O
O O
O
OO
O
O
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
O
O
OO
OO
OO
O
O
O O
O
O
O
OOO
O
OO
O
OOO
O OO
O
O
O OOO
O
O
OOO
O
O
O
O
OO
O
O
O
OO
OO O
OO
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O
OOO O
O
OO
O
OO
O
O
O
O
OO
OO
OO
OO
O
OO O
O
OO
OO
O
O
O
O
O
O
OOO
OO
O
OO
OO
O
O
O
OO
O
OO O
OO
O
O
O
O
O
O
OO
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OOO
OO
O
O
O
OO
OOO
O O
O
O
O
OOOO
OO
O
OO
O
O
O
O
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O OO
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O O
O
O
O
O
O
O
O
O
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
OOO
O
O
O
O
O
O
O O
OOO
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O
O
O
OOO
O
OO
O
O
OOO
O
O O
O
O
O
O
O
OO
O
O
O O
O
OOOO
O
O
O O
OO
O
O
OO
O
O
O
O
OO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OOO
O
O
O
O
OO
O
O
O
O
OO
O
O
OO
O
O
O
OO
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
OOO
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OO
O
O
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O
O
O
O O
OOO
O
OO
O
O
OO
O
O
OO
O
O
O
O
OO
O
O
OO
O
O
O
O
OO
O
O
OO
O
O O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OO
O
O
O
OO
O
O
O
O
O
O
O
O
O
OO
OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
OO
OO
O
O
O
O
O
O
O
O
O
O
O
O
OOO
O
O
OO
OO
O
O
O
O
O
OO
O
O
O
OO
O
O
O
O
O
OOO
O
O
OOO
O
O
O
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O O
O
O
O
OO
O
O
OO
O
O
O
O
O
OO
O
O O
OO
O
OO
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O OO
O
OOO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O OO
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O O
O
O
O
O
OO
OO
O
O
O
O
OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OOO
O
O
OO
OO O
O
O
O
O
O
O
O
O
O
O
O
O
O
O O
O
O
O
O
O
O
O OO
O
O
O
OO
O
O
O
OO
O
O O
O
OO
O
O
O
OO
O
OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O O
O
O
O
OO
O
O
O
O
O
O OO
O
O
OOO
OO
O
OO
O
OO
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
OO O
O
O
O O
OO
O
O
O
OO
O
O
OO
O
O
O
O
O
OO
O
O
OO
O
O
O
OO
O
OO
O
O
O
O
O O
O
OO
O
O
O
O
O
O
OO
O
OOO
O
O
O
O
O
OO
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OOOO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O O
O
OO
OO
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
OOO
O
OO
O
OO
O
O
O
O
O
O
O
O
OO
O
O
OO
OO
O
O
O
O
O
O
O
OO
OO
OOO
O
O
O
OO
O
OO
O
O
OO
O
O
OOO
O
O
OO
OO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O O
O
O
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O O
OO
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O OO
O
O
O
OO
OO
OO
O
O
OO
O
O
O
O
O
O
O
OO
O
OO
O
O
OO
O
O
O
O
O
O
O
OO
OOO
O
O
O
O
O
OO
O
O
O O
OO
O
O
O
O
O
O
O
O
O O
O
O
O
OO
OO
O
O
O
O
O
O
O
OOO
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OO
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OO
O
O
O
O
O
O
O
O
OO
O O
O
O
OO
O
O
O
O
OO
O
O
OOOO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
OOOO OO
O
O
O
O
OO
OO
O
O
O
OO
O
OO
O
OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
OOO
OO
OO
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OO
OO
O
O
O
O
OO
O OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
OO
O
OO
O
O
O
O
OO
O
O
O
OO
O
O
O
O
O
OO
O
O
O
OO
O
O
OO
O
O
O
OO
O O
O
OO
O
O
O
O
OO
O
O
O
O
O
O O
O
O
O
OOO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
OOOOO
O
O
O
O
O
OO
O OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
OO
O
O
O
OO
O
O
O
O
O
O
OO
O
OO
O
O
O
O
O
O
O
O
O
O
O OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
O
OO
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
O
O O
O
O
OO
O
O O
O
O
O
O
O
OOO
OO
O
O
O
O
O
O
O
OO
O
O
O
OO
O
O
O
O
O
O
O
OO
OO
O
O
O
O
OOO
O
O
OO OO
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
OO
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O
OO
O
O
O
O
O
O
O
OOO
O
O
O
O
O
OO
O
OO
O
O
OO
OO
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
OO
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
OO
O
O
OOOO
O
O
O
O
OO
O
OO
O
O
O
O
O
O
O
OO
O
O
OO
O
O
OO
OOO O
O
OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
OO O
O
OOO O
O
O
OO
O
O
O
OO O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
OOO
O
O
OO
OO
O
O
O
O
O
OOO
O
O
O
O
O
OO
O
O
O
O
O
O
O O
O
O
O
O
OO
O
O
OOO
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
OO
O
O
O
O O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
OOO
O
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O
OO
O
O
O
OO
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O O
O
O
O
O
O
O
O
O
OOO
O
O
O
O
OOO
O
O
OO
OOOOOOO
OO
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
OO
OO
OO
O
O
OO
O
O
O
O
OO
O
OO
O
OO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O O
O O
OO
OO
O OOO
O
O
O
O
OO
O
OOO
OO
OO
O
O
O
O
O
O
O
OO
O
O
OO
O
O
O
O
O
OO
O
O
O O
OO
O
OO
O
O
OO
O
OO
O
O
O
O
O
O
O
O
O OO
O
OO
OO
O
O
OOOO
OO
O
O
OO
O
OO
O
O
O
O
O
O
O
O
OO
O
O
O
O
O O
O
O
O
O
O
O
O O
O
OO
O
O
O
O
O
OO O
O
O
O
O
O
OO
O
O
OO
O
O
O
OO
O
O
O
OO
O
O
O
O
OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
OOO
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
O
O
O
O
O
O
O
OO
O
O
O
OO
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O O
O
O
O
OOO
O
O
O
O
OO
OOO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O O
O
OO
O
OO
O
O
O
O
O
OO
O
O
OO
OOO
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
O
OO
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
OOO
O
O
OO
OO
O
O
O
O
OO
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
OOO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
OO
O
OO
O
O
OO
O
O
O
OO
O
O
O
O
O
O
O
O
OO
O
O
OO
OO
O
O
O
O
O
O
O
O O
O
O
OO
OO
O
O
O
O
O
O O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
OO
O
O
O O
OO
O
O
O
O
OO
OO
OO
O
O
O
O
OO
OO
O
O
O
O
O
O O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O OO
O
O O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
OOO
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
OO
O
O
O
OO
O
OO
O
O
O
O
OO
O
OO
O
O
OO
O
O
O
O
O
O
O
O
OO O
O
O
O
OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O O
OO
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
OOO
OO
O
OO
O
O
O
OO
O
O
O
O
O
O
OO
O
OO
O
O
OOO
OO
OO
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
OO
O
O
O
O
O
OO
O
O
O
OO
O
OO
O
O
O
O
O
OO
O
O
OO
O
O
OO
O
O
O
O
O
OO
O
O
OO
O
O
OO
OO
O
O
O
OO
O
O
OO
OOO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
O
O
O
O
OO
O
O O
OOO
O
O
O
O
OO
O
OO
O
O
O
O
OO
O
OOOOOOO
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O OO
O
O
O
OO
O
O
O
OO
O
O
O
O
O
O
O
O
O
O
O
O
O
OO
O
O
O
OO
O O
O
O
O
O
O
O
O
O
O
O
OO
O
O
OOO
O
O
OO
O
O
O OO
O
O
OO
O
OO
O
OO
O
OOO
O
O
O
O
O
O
O
O
OO
O O
O
O
O
O
O
O
O
O
OO
O O
O
O
OO
O
O
O
O
O
O
O
O
O
O
posterior mean
1
1/4
1/16
0
−1/16
−1/4
−1
(X−R)/(X+R)
RNAiexample
Bayes?
Improved ranking and selection 7
Figure 5 compares r-value rankings with three other methods in the RNAi example fromFigure 2. In this example, Di = (ni, yi) holds binomial information (set size ni and numberyi of genes in set i that were identified by RNAi). The target parameters !i are treated asdraws from a Beta(a, b) distribution, with shape parameters estimated by marginal maxi-mum likelihood, and the conditional tail probability V!(Di) becomes the probability that aBeta(a+ yi, b + ni ! yi) variable exceeds !!. R-value computation (see Section 4) requiresthe sampling distribution of these tail probabilities, which we approximated using the datafrom all 5719 sets under study. The methods compared in Figure 5 agree to some extent onthe ranking of the most interesting sets, but systematic di!erences are apparent. Rankingby yi/ni over-ranks small sets; ranking by p-value over-ranks large sets; and ranking byposterior mean (yi + a)/(ni+ a+ b) also over-ranks large sets, though to a lesser degree, allcompared to the r-value ranking.
R-values may be computed in all sorts of hierarchical modeling e!orts, including semi-parametric models and cases where Markov chain Monte Carlo (MCMC) is used to approx-imate the marginal posterior distribution of each !i given available data. Figure 6 comparesthe r-value ranking with other rankings in an example from gene-expression analysis, whereevidence suggested that the expression of a large fraction of the human genome was associ-ated with the status of a certain viral infection (Pyeon, et al., 2007). A multi-level modelinvolving both null and non-null genes as well as t!distributed non-null e!ects !i exhibitedgood fit to the data, but did not admit a closed form for V!(Di). R-values, computed usingMCMC output, again reveal systematic ranking di!erences from other approaches.
Multi-level models drive statistical inference and software in a variety of genomic do-mains: for example, limma (Smyth, 2004), EBarrays (Kendziorski et al. 2003), EBSeq (Lenget al. 2013), among others. Since these models happen to specify distributional forms forparameters of interest, the associated code could be augmented to compute posterior tailprobabilities V!(Di) and thus r-values for ranking. The limma system utilizes a conjugatenormal, inverse-gamma model, and so V!(Di) involves the tail probability of a non-centralt distribution. The EBSeq system entails a conjugate beta, negative-binomial model, andso V!(Di) for di!erential expression involves tail probabilities in a certain ratio distribution(Coelho and Mexia, 2007). One expects the benefits of r-value computation to show espe-cially in cases involving many non-null units and relatively high variation among units intheir variance parameters (e.g., sequence read depth).
3. Connections
3.1. Connection to Bayes ruleThe proposed ranking procedure is a kind of Bayes rule for a population relative rank whenconsidering multiple loss functions and a distributional constraint. To see how, introduce acollection of loss functions
L!(a, !i) = 1! 1 (a " ", !i # !!)
where action a is a relative rank value in (0, 1), " $ (0, 1) indexes the collection, and again!! = F!1(1 ! ") is a quantile in the population of interest. Specifically, no "!loss occursif the inferred relative rank a and the actual relative rank 1 ! F (!i) both are less than ".The marginal (pre-posterior) Bayes risk of rule #(Di) is
risk! = 1! P {#(Di) " " , ! # !!} , (9)
Improved ranking and selection 7
Figure 5 compares r-value rankings with three other methods in the RNAi example fromFigure 2. In this example, Di = (ni, yi) holds binomial information (set size ni and numberyi of genes in set i that were identified by RNAi). The target parameters !i are treated asdraws from a Beta(a, b) distribution, with shape parameters estimated by marginal maxi-mum likelihood, and the conditional tail probability V!(Di) becomes the probability that aBeta(a+ yi, b + ni ! yi) variable exceeds !!. R-value computation (see Section 4) requiresthe sampling distribution of these tail probabilities, which we approximated using the datafrom all 5719 sets under study. The methods compared in Figure 5 agree to some extent onthe ranking of the most interesting sets, but systematic di!erences are apparent. Rankingby yi/ni over-ranks small sets; ranking by p-value over-ranks large sets; and ranking byposterior mean (yi + a)/(ni+ a+ b) also over-ranks large sets, though to a lesser degree, allcompared to the r-value ranking.
R-values may be computed in all sorts of hierarchical modeling e!orts, including semi-parametric models and cases where Markov chain Monte Carlo (MCMC) is used to approx-imate the marginal posterior distribution of each !i given available data. Figure 6 comparesthe r-value ranking with other rankings in an example from gene-expression analysis, whereevidence suggested that the expression of a large fraction of the human genome was associ-ated with the status of a certain viral infection (Pyeon, et al., 2007). A multi-level modelinvolving both null and non-null genes as well as t!distributed non-null e!ects !i exhibitedgood fit to the data, but did not admit a closed form for V!(Di). R-values, computed usingMCMC output, again reveal systematic ranking di!erences from other approaches.
Multi-level models drive statistical inference and software in a variety of genomic do-mains: for example, limma (Smyth, 2004), EBarrays (Kendziorski et al. 2003), EBSeq (Lenget al. 2013), among others. Since these models happen to specify distributional forms forparameters of interest, the associated code could be augmented to compute posterior tailprobabilities V!(Di) and thus r-values for ranking. The limma system utilizes a conjugatenormal, inverse-gamma model, and so V!(Di) involves the tail probability of a non-centralt distribution. The EBSeq system entails a conjugate beta, negative-binomial model, andso V!(Di) for di!erential expression involves tail probabilities in a certain ratio distribution(Coelho and Mexia, 2007). One expects the benefits of r-value computation to show espe-cially in cases involving many non-null units and relatively high variation among units intheir variance parameters (e.g., sequence read depth).
3. Connections
3.1. Connection to Bayes ruleThe proposed ranking procedure is a kind of Bayes rule for a population relative rank whenconsidering multiple loss functions and a distributional constraint. To see how, introduce acollection of loss functions
L!(a, !i) = 1! 1 (a " ", !i # !!)
where action a is a relative rank value in (0, 1), " $ (0, 1) indexes the collection, and again!! = F!1(1 ! ") is a quantile in the population of interest. Specifically, no "!loss occursif the inferred relative rank a and the actual relative rank 1 ! F (!i) both are less than ".The marginal (pre-posterior) Bayes risk of rule #(Di) is
risk! = 1! P {#(Di) " " , ! # !!} , (9)
8 Henderson and Newton
which is one minus the agreement (2). In the absence of other considerations, the Bayesrule for loss L! degenerates to !(Di) = 0. Degeneration is avoided if we enforce on theestimated rank the additional structure that it share with the true rank 1 ! F ("i) theproperty of being uniformly distributed over the population of units. Such a constrainedBayes rule then minimizes the modified objective function:
risk! + #!P {!(Di) " $}
where #! is chosen to enforce the size constraint P {!(Di) " $} = $.The constrained Bayes rule is computed conditionally, per observed Di, by minimizing
the (modified) posterior expected loss (PEL)
PEL! = 1! P {!(Di) " $ , "i # "!|Di}+ #!1 {!(Di) " $} (10)
=
!
1! V!(Di) + #! if !(Di) " $1 if !(Di) > $
where V!(Di) is the upper posterior probability P ("i # "!|Di) appearing in Section 2.Curiously, finding the rule to minimize PEL! is not determined at a single $, since mini-mization in (10) requires only that
!(Di) " $ $% V!(Di) # #!. (11)
However, taking all losses together does fix a procedure:
!(Di) = inf {$ : V!(Di) # #!} . (12)
The thresholds #! are determined by the uniformity constraint, and we have #! = H!1! (1!
$), where H! is the marginal distribution of V!(Di), counting all sources of variation,and so #! = %! from the previous section. In other words, the procedure obtained by thisconstrained, multi-loss Bayes calculation is equivalent to the r-value introduced in Section 2.
3.2. Beyond p’s and q’sIn testing a single hypothesis H0, the sample space may be structured as a nested sequenceof subsets, {!! : $ & (0, 1)}, say, such that rejection of a size $ test is equivalent to data Dlanding in set (i.e., rejection region) !!. Then, the p-value of the test is
p(D) = inf{$ : D & !!}.
Storey (2003) extended this idea to multiple testing and the positive false discovery rate withthe introduction of the q-value. Specifically, with another nested sequence {!! : $ & (0, 1)}indexed such that
P (H0|D & !!) = $,
the q-value is q(D) = inf{$ : D & !!}. Where p-values refer to the distribution of D onH0, and q-values the conditional probability of H0 given sample information, the proposedr-values refer to a marginal probability. The size constraint (1) corresponds to anothersequence of subsets, {!!}, say, for which the marginal constraint holds: P (D & !!) = $.Analogously, the r-value is r(D) = inf{$ : D & !!}. In principle an r-value could be definedfor any indexed ranking method, though we have reserved the definition for that methodwhich maximizes agreement (2).
multi-loss
marginally constrained
• ranking to maximize agreement• large-scale, non-sparse settings• empirical Bayes inference
• R-package at CRAN: rvalues• manuscript at arXiv: 1312.5776
summary
context
Karen Montgomery
David Schwartz
David Schwartz
Christina KendziorskiVijay Setaluri
Rich Halberg
Mark Albertini
Bill Dove
Paul Ahlquist Audrey Gasch
interdisciplinary biostatistics
Karen Montgomery
David Schwartz
David Schwartz
Christina KendziorskiVijay Setaluri
Rich Halberg
Mark Albertini
Bill Dove
Paul Ahlquist Audrey Gasch
genome technology: optical mapping
Karen Montgomery
David Schwartz
David Schwartz
Christina KendziorskiVijay Setaluri
Rich Halberg
Mark Albertini
Bill Dove
Paul Ahlquist Audrey Gasch
cancer biology: polyclonality
Karen Montgomery
David Schwartz
David Schwartz
Christina KendziorskiVijay Setaluri
Rich Halberg
Mark Albertini
Bill Dove
Paul Ahlquist Audrey Gasch
melanoma: T cell diversity
Karen Montgomery
David Schwartz
David Schwartz
Christina KendziorskiVijay Setaluri
Rich Halberg
Mark Albertini
Bill Dove
Paul Ahlquist Audrey Gasch
virology: host/virus interactions
Karen Montgomery
David Schwartz
David Schwartz
Christina KendziorskiVijay Setaluri
Rich Halberg
Mark Albertini
Bill Dove
Paul Ahlquist Audrey Gasch
genome biology: stress response
Karen Montgomery
David Schwartz
David Schwartz
Christina KendziorskiVijay Setaluri
Rich Halberg
Mark Albertini
Bill Dove
Paul Ahlquist Audrey Gasch
stem cells: iPS vs hES
Karen Montgomery
David Schwartz
David Schwartz
Christina KendziorskiVijay Setaluri
Rich Halberg
Mark Albertini
Bill Dove
Paul Ahlquist Audrey Gasch
colon cancer: APC modifiers
Karen Montgomery
David Schwartz
David Schwartz
Christina KendziorskiVijay Setaluri
Rich Halberg
Mark Albertini
Bill Dove
Paul Ahlquist Audrey Gasch
melanoma: transcript regulation
Karen Montgomery
David Schwartz
David Schwartz
Christina KendziorskiVijay Setaluri
Rich Halberg
Mark Albertini
Bill Dove
Paul Ahlquist Audrey Gasch
genomics: gene set analysis