Description

dBUMfit is supposed to take as input a vector of p-values for deriving their distribution under beta-uniform mixture model (see Note below). The density distribution of input p-values is expressed as a mixture of two components: one for the null hypothesis (the noise component) and the other for the alternative hypothesis (the signal component). The noise component is the uniform density, while the signal component is the remainder of the mixture distribution. It returns an object of class "BUM".

Usage

dBUMfit(x, ntry = 1, hist.bum = T, contour.bum = T, verbose = T)

Arguments

x

a vector containing input p-values

ntry

an integeter specifying how many trys are used to find the optimised parameters by maximum likelihood estimation

hist.bum

logical to indicate whether the histogram graph should be drawn

contour.bum

logical to indicate whether a contour plot should be drawn to show the log likelihood as a function of two parameters (a and lambda) in the beta-uniform mixture model

verbose

logical to indicate whether the messages will be displayed in the screen. By default, it sets to true for display

Value

an object of class "BUM", a list with following elements:

• lambda: estimated mixture parameter
• a: estimated shape parameter
• NLL: Negative log-likelihood
• pvalues: the input pvalues
• call: the call that produced this result

Note

The probability density function of p-values under the Beta-Uniform Mixture model is formulated as: f(x|\lambda,a) = \lambda + (1-\lambda)*a*x^{a-1}. The model names after mixing two distributions:

• the uniform distribution with the density function as \frac{1}{b-a}|_{a=0}^{b=1}=1
• the beta distribution with the density function as \frac{\Gamma(a+b)}{\Gamma(a)+\Gamma(b)}*x^{a-1}*(1-x)^{b-1}|_{b=1}=a*x^{a-1}

Both are mixed via \lambda. The mixture parameter \lambda measures the contribution from the uniform distribution. Accordingly, 1-\lambda measures the contribution from the beta distribution. Notably, the probability density function of the beta distribution can be splitted into two parts (rather than the exclusitive signal):

• the constant part as noise: a*x^{a-1}|_{x=1}=a
• the rest part as signal: a*(x^{a-1}-1)

In other words, there is no signal at x=1 but all being noise. It is a conservative, upper bound estimation of the noise. Therefore, the probability density function in the model can be decomposed into signal-noise components:

• the signal component: (1-\lambda)*a*(x^{a-1}-1)
• the noise component: \lambda + (1-\lambda)*a

It is misleading to simply view \lambda as the noise component and (1-\lambda)*a*x^{a-1} as the signal component, just as wrongly do in the literatures (e.g. http://www.ncbi.nlm.nih.gov/pubmed/18586718)

Examples

# 1) generate an vector consisting of random values from beta distribution
x <- rbeta(1000, shape1=0.5, shape2=1)

# 2) fit a p-value distribution under beta-uniform mixture model
fit <- dBUMfit(x)

	A total of p-values: 1000
	Maximum Log-Likelihood: 311.1
	Mixture parameter (lambda): 0.000
	Shape parameter (a): 0.498
fit$lambda


[1] 1e-05

fit$a

[1] 0.4979076

Source code

Source man

See also