imageData (version 0.1-61)

PVA: Selects a subset of variables using Principal Variable Analysis (PVA)

Description

Principal Variable Analysis (PVA) (Cummings, 2007) selects a subset from a set of the variables such that the variables in the subset are as uncorrelated as possible, in an effort to ensure that all aspects of the variation in the data are covered.

Usage

PVA(responses, data, nvarselect = NULL, p.variance = 1, include = NULL, 
    plot = TRUE, ...)

Value

A data.frame giving the results of the variable selection. It will contain the columns Variable, Selected,

h.partial, Added.Propn and Cumulative.Propn.

Arguments

responses

A character giving the names of the columns in data from which the variables are to be selected.

data

A data.frame containing the columns of variables from which the selection is to be made.

nvarselect

A numeric specifying the number of variables to be selected, which includes those listed in include. If nvarselect = 1, as many variables are selected as is need to satisfy p.variance.

p.variance

A numeric specifying the minimum proportion of the variance that the selected variables must account for,

include

A character giving the names of the columns in data for the variables whose selection is mandatory.

plot

A logical indicating whether a plot of the cumulative proportion of the variance explained is to be produced.

...

allows passing of arguments to other functions

Author

Chris Brien

Details

The variable that is most correlated with the other variables is selected first for inclusion. The partial correlation for each of the remaining variables, given the first selected variable, is calculated and the most correlated of these variables is selects for inclusion next. Then the partial correlations are adjust for the second included variables. This process is repeated until the specified criteria have been satisfied. The possibilities are:

1. the default (nvarselect = NULL and p.variance = 1), which selects all variables in increasing order of amount of information they provide;

2. to select exactly nvarselect variables;

3. to select just enough variables, up to a maximum of nvarselect variables, to explain at least p.variance*100 per cent of the total variance.

References

Cumming, J. A. and D. A. Wood (2007) Dimension reduction via principal variables. Computational Statistics and Data Analysis, 52, 550--565.

See Also

intervalPVA, rcontrib

Examples

data(exampleData)
responses <- c("Area","Area.SV","Area.TV", "Image.Biomass", "Max.Height","Centre.Mass",
               "Density", "Compactness.TV", "Compactness.SV")
results <-  PVA(responses, longi.dat, p.variance=0.9, plot = FALSE)