aziztest (version 0.2.1)

calibrate_test: Calibration of p-values in a slow multi-hypothesis setting

Description

Compute the null distribution of test statistics on one gaussian variable. Useful if testing a large number of variables at once since it allows running permutations only once beforehand rather than for every variable. Used in conjunction with "get_calibrated_pvalues"

Usage

calibrate_test(
  y,
  w = NULL,
  rep = 1e+07,
  doall = TRUE,
  unidirectional = 0,
  flatten = 0.5,
  ignoremax = 0,
  normmethod = 1,
  novariance = F
)

Arguments

y

A binary vector of sample labels (cases=1, controls=0).

w

Default = NULL. Optional numerical vector of weights. 1 means all weights are equal to 1 and only the ordering is considered. If NULL (default), a standardisation of x is used to calculate the weights giving larger weights to aberrations of larger magnitude.

rep

Default=100000. Number of permutations to be used to calculate p-values.

doall

Default=TRUE. All permutations are performed

unidirectional

Default = 0. Can be 0, 1 or -1. 0 is for testing both directions of effect. 1 is for testing cases<controls and -1 is for testing cases>controls.

flatten

Default = 0.5. Numeric value recommended between 0 and 1. If weights are not given, we take the max of flatten and the absolute value of the Z-score of x as the weights (Default behavior).

ignoremax

Default=0. Optional value indicating if we should ignore the first few values when selecting the maximal enrichment score. Alternatively, it can be viewed as the minimal size considered for the aberrant interval.

normmethod

Default=1. If w=NULL the weights are generated by subtracting the mean and dividing by the standard deviation. If normmethod=2, the median and MAD are used instead, for a better treatment of outliers.

novariance

Default=FALSE. aziz.test is able to detect a difference in variance between cases and controls as an association (when variance of cases is larger than the variance of controls). novariance=True changes the behaviour and penalizes scenarios with outliers going both ways in the cases. This will remove the associations that would usually be picked by a Levene test. Consider using this when using unidirectional testing if variance changes between groups are irrelevant in your considered problem. Results in loss of power.

Value

A vector that can be used in get_calibrated_pvalues()

See Also

get_calibrated_pvalues, aziz.test

Examples

# NOT RUN {
y = c(rep(1,200),rep(0,200))
x = rnorm(400)
calibration = calibrate_test(y,rep=100)
es = aziz.test(y,x,rep=0)$es #No need for permutations, p-values computed from calibration
get_calibrated_pvalues(calibration,es)

# }