def groupTestExclusivity(events, bg):
logP = numpy.log(bg)
logNotP = numpy.log(1 - bg)
pExactlyOne = numpy.exp(
numpy.logaddexp.reduce(
numpy.repeat(logNotP.sum(0)[numpy.newaxis], bg.shape[0], 0) +
logP - logNotP, 0))
x = numpy.sum(events.sum(0) == 1)
return 1 - poisbinom.cdf(pExactlyOne, x - 1)
def groupTestExclusivity(events, bg):
logP = numpy.log(bg)
logNotP = numpy.log(1 - bg)
pExactlyOne = numpy.exp(
numpy.logaddexp.reduce(
numpy.repeat(logNotP.sum(0)[numpy.newaxis], bg.shape[0], 0) +
logP - logNotP, 0))
x = numpy.sum(events.sum(0) == 1)
# the numpy.minimum(...) fix should probably be in the poisbinom
# module rather than here, but it solves the problem for now
return 1 - numpy.minimum(1.0, poisbinom.cdf(pExactlyOne, x - 1))
def groupTestImpurity(events, bg):
logP = numpy.log(bg)
logNotP = numpy.log(1 - bg)
pNone = numpy.exp(logNotP.sum(0)[numpy.newaxis])
pExactlyOne = numpy.exp(
numpy.logaddexp.reduce(
numpy.repeat(logNotP.sum(0)[numpy.newaxis], bg.shape[0], 0) +
logP - logNotP, 0))
pMoreThanOne = 1 - pNone - pExactlyOne
# Due to floating point precision issues the above probabilities
# might end up being slightly smaller than 0; clip them at 0
pMoreThanOne = numpy.maximum(0, pMoreThanOne)
x = numpy.sum(events.sum(0) > 1)
return poisbinom.cdf(pMoreThanOne, x)
def groupTestCoverage(events, bg):
pCovered = 1 - numpy.exp(numpy.log1p(-bg).sum(0))
return 1 - poisbinom.cdf(pCovered, events.any(0).sum() - 1)