mandel.kh: Calculate Mandel's h and k statistics for replicate observations

Description

mandel.kh calculates Mandel's h and k statistics for replicate observations. These are traditionally used to provide a rapid graphical summary of results from an inter-laboratory exercise in which each organisation provides replicate observations of one or more measurands on one or more test items. Mandel's h is an indication of relative deviation from the mean value; Mandel's k is an indicator of precision compared to the pooled standard deviation across all groups.

Usage

mandel.kh(x, g = NULL, m = NULL, na.rm = T, rowname = NULL, 
			type = c("h", "k"), method=c("classical", "robust"), n = NA, ...)

	## S3 method for class 'default':
mandel.kh(x, g = NULL, m = NULL, na.rm = T, rowname = NULL, 
			type = c("h", "k"), method=c("classical", "robust"), n = NA, ...)

	## S3 method for class 'data.frame':
mandel.kh(x, g = NULL, m = NULL, na.rm = T, rowname = NULL, 
			type = c("h", "k"), method=c("classical", "robust"), n = NA, ...)

	## S3 method for class 'matrix':
mandel.kh(x, g = NULL, m = NULL, na.rm = T, rowname = NULL, 
			type = c("h", "k"), method=c("classical", "robust"), n = NA, ...)

	## S3 method for class 'array':
mandel.kh(x, g = NULL, m = NULL, na.rm = T, rowname = NULL, 
			type = c("h", "k"), method=c("classical", "robust"), n = NA, ...)

	## S3 method for class 'ilab':
mandel.kh(x, g = NULL, m = NULL, na.rm = T, rowname = NULL, 
			type = c("h", "k"), method=c("classical", "robust"), n = NA, ...)

Arguments

An R object (see Details below), which contains replicate observations or, if g is absent, means or standard deviations.

A primary grouping factor, usually corresponding to Laboratory in an inter-laboratory study. If not present, x is taken as a set of means or standard deviations (depending on whether type is "h" or <

A secondary grouping factor, usually corresponding to test item or measured quantity. m is ignored if x has more than one column.

na.rm

A logical value indicating whether 'NA' values should be stripped before the computation proceeds. Passed to functions such as mean and sd.

rowname

A single character label for the primary grouping factor (e.g. "Lab", "Organisation").

type

Character denoting the statistic to be calculated; may be "h" or "k".

method

Character scalar giving the calculation method. "classical" gives the traditional calculation; "robust" gives a robust variant (see Details).

scalar number of observations per group. Required only if x consists of calculated standard deviations.

...

Additional parameters passed to hubers when method="robust" and type="h".

Value

mandel.kh returns an object of class "mandel.kh", which is a data frame consisting of the required Mandel's statistics and in which each row corresponds to a level of g and each column to a level of m or (if x was a matrix or data frame) to the corresponding column in x. In addition to the class, the object has attributes: [object Object],[object Object],[object Object]

Details

mandel.kh can be called directly, but is usually intended to be called via convenience functions mandel.h or mandel.k. mandel.kh is a generic, with methods for numeric vectors, arrays, data frames, matrices and objects of class 'ilab'. Mandel's statistics are simple indicators of relative deviation or precision for grouped sets of observations. Given a set of observations $x_{ijl}$ where $i, j, l$ denotes observation $l$, $l=1, 2, ... n$ for measurand or test item $j$ and group (usually laboratory) $i$, $i=1, 2, ... p$, Mandel's $h$ and $k$ are given by: $$h=\frac{\bar{x_{ij}}-\bar{x_j}}{s_j}$$ where $s_j=\sqrt{\sum_{i=1}^p{\frac{(\bar{x_{ij}}-\bar{x_j})}{p-1}}}$ and $$k=\sqrt{\frac{s_{ij}^2}{\sum_{i=1}^p{s_{ij}^2/p}}}$$ where $s_{ij}$ is the standard deviation of values $x_{ijk}$ over $k=1, 2, ..., n$. If x is a vector, one-dimensional array or single-column matrix, values are aggregated by g and, if present, by m. If x is a data frame or matrix, each column is aggregated by g and m silently ignored if present. In all cases, if g is NULL or missing, each row (or value, if a vector) in x is taken as a pre-calculated mean (for Mandel's $h$) or standard deviation (for Mandel's $k$). If x is an object of class 'ilab', g defaults to '$org' and m to $measurand. The returned object includes a label ('grouped.by') for the primary grouping factor. For the 'ilab' method, this is "Organisation". For other methods, If rowname is non-null, rowname is used. If rowname is NULL, the default is deparse(substitute(g)); if g is also NULL or missing, "Row" is used. If method="robust", Mandel's $h$ is replaced by a robust z score calculated by replacing $\bar{x_j}$ and $s_j$ with the robust estimates of location and scale obtained using Huber's estimate with tuning constant k set to 1.5 (or as included in ...), and Mandel's $k$ is calculated by replacing the classical pooled standard deviation in the denominator with the robust pooled standard deviation calculated by algorithm S (see algS).

References

Accuracy (trueness and precision) of measurement methods and results -- Part 2: Basic method for the determination of repeatability and reproducibility of a standard measurement method. ISO, Geneva (1994).

Examples

Run this code

data(RMstudy)

	#Data frame examples: note no secondary grouping factor
	h <- with(RMstudy, mandel.kh(RMstudy[2:9], g=Lab, type="h"))
	plot(h, las=2)

	k <- with(RMstudy, mandel.kh(RMstudy[2:9], g=Lab, type="k"))
	plot(k, las=2)

	#Vector variant
	RMstk <- stack(RMstudy[,2:9])
	names(RMstk) <- c("x", "meas")
		#names replace 'values' and 'ind'
	RMstk$Lab <- rep(RMstudy$Lab, 8)
	h2 <- with(RMstk, mandel.kh(x, g=Lab, m=meas, rowname="Laboratory"))
		#Note use of rowname to override g
	plot(h2, las=2)
	
	#ilab method
	RM.ilab <- with(RMstk, construct.ilab(org=Lab, x=x, measurand=meas, 
		item=factor(rep("CRM", nrow(RMstk))) ) )

	plot(mandel.kh(RM.ilab, type="h"))
	
	#Robust variants
	hrob <- with(RMstudy, mandel.kh(RMstudy[2:9], g=Lab, type="h", method="robust"))
	plot(hrob, las=2)
	
	krob <- with(RMstudy, mandel.kh(RMstudy[2:9], g=Lab, type="k", method="robust"))
	plot(krob, las=2)

Run the code above in your browser using DataLab