binomial.bias.test: Perform a one-tailed sign test on p values between 0.03 and...

In roblanf/pcurver: Tools to analyse distributions of P-values from the published literature

Description

This function tests the null hypothesis that there are at least as many p values in the 0.03 - 0.04 bin as in the 0.04 - 0.05 bin. Note that p values of exactly 0.04 are excluded, because they do not fall in either bin. The test uses a one-tailed sign test. Significantly more p values in the smaller bin is consistent with collections of p values with evidential value. Significantly more p values in the larger bin is consistent with p-hacking or publication bias. This is a more sensitive test of p-hacking or publication bias than the related binomial.all.test

Usage

binomial.bias.test(p)

Arguments

p	a vector of p values between 0.0 and 0.05 (inclusive)

Value

a list giving the number of p values in each bin, and the p value of the two-tailed sign test.

See Also

binomial.all.test, binomial.sns.test

Examples

# here are some p values you might get from a strong effect size
p <- rexp(1000, 200)
p <- p[p<0.05]

# let's add some you might get from p-hacking and/or publication bias
h <- -1 * rexp(100, 200) + 0.05
h <- h[h>0.00]

p <- c(p, h)

# the binomial.all.test should show significant right skew
# that's expected - it uses all the data from 0.00 to 0.05
binomial.all.test(p)

# the binomial.bias.test is more sensitive to p-hacking and/or
# publication bias - it uses just the data from 0.03 to 0.05
binomial.bias.test(p)

roblanf/pcurver documentation built on May 27, 2019, 11:58 a.m.