rankComb: Ranking Based Binary Partitions

Description Usage Arguments Value Author(s) See Also Examples

View source: R/rankComb.R

Description

Blindly fitting a model to all possible partitions is wasteful use of resources. Instead, one can rank the K levels (strata) based on expected response values to explore only K-1 binary partitions along the gradient defined by the ranks of the expected values.

Usage

1
2
oComb(x, collapse)
rankComb(Y, X, Z, dist = "gaussian", collapse, ...)

Arguments

Y

numeric, vector of observations.

X

numeric, design matrix.

Z

factor, must have at least 2 unique levels.

dist

character, distribution argument passed to underlying functions, see listed on the help page of opticut.

x

and a numeric vector.

collapse

character, what to paste between levels. Defaults to getOption("ocoptions")$collapse.

...

other arguments passed to the underlying functions, see opticut.

Value

oComb returns the 'contrast' matrix based on the rank vector as input. Ranked from lowest to highest expected value among the partitions.

The function rankComb fits the model with multiple (K > 2) factor levels to find out the ranking, and returns a binary classification matrix as returned by oComb corresponding to the ranking.

Author(s)

Peter Solymos <[email protected]>

See Also

allComb for alternative partitioning algorithm.

opticut for the user interface.

Examples

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
## simulate some data
set.seed(1234)
n <- 200
x0 <- sample(1:4, n, TRUE)
x1 <- ifelse(x0 %in% 1:2, 1, 0)
x2 <- rnorm(n, 0.5, 1)
lam <- exp(0.5 + 0.5*x1 + -0.2*x2)
Y <- rpois(n, lam)

## binary partitions
head(rc <- rankComb(Y, model.matrix(~x2), as.factor(x0), dist="poisson"))
attr(rc, "est") # expected values in factor levels
aggregate(exp(0.5 + 0.5*x1), list(x0=x0), mean) # true values

## simple example
oComb(1:4, "+")
## using estimates
oComb(attr(rc, "est"))

psolymos/opticut documentation built on April 29, 2018, 10:17 a.m.