pdi {mcca}

Calculate PDI Value

Description

compute the Polytomous Discrimination Index (PDI) value of two or three or four categories classifiers with an option to define the specific model or user-defined model.

Usage

pdi(y, d, method="multinom", ...)

Arguments

y

The multinomial response vector with two, three or four categories. It can be factor or integer-valued.

d

The set of candidate markers, including one or more columns. Can be a data frame or a matrix; if the method is "prob", then d should be the probablity matrix.

method

Specifies what method is used to construct the classifier based on the marker set in d. Available option includes the following methods:"multinom": Multinomial Logistic Regression which is the default method, requiring R package nnet;"tree": Classification Tree method, requiring R package rpart;"svm": Support Vector Machine (C-classification and radial basis as default), requiring R package e1071;"lda": Linear Discriminant Analysis, requiring R package lda;"prob": d is a risk matrix resulted from any external classification algorithm obtained by the user.

...

Additional arguments in the chosen method's function.

Details

The function returns the PDI value for predictive markers based on a user-chosen machine learning method. Currently available methods include logistic regression (default), tree, lda, svm and user-computed risk values. This function is general since we can evaluate the accuracy for marker combinations resulted from complicated classification algorithms.

Value

Returns an object of class "mcca.pdi". The PDI value of the classification using a particular learning method on a set of marker(s).

An object of class "mcca.pdi" is a list containing at least the following components:

call

the matched call.

measure

the value of measure.

table

the category-specific value of measure.

Note

Users are advised to change the operating settings of various classifiers since it is well known that machine learning methods require extensive tuning. Currently only some common and intuitive options are set as default and they are by no means the optimal parameterization for a particular data analysis. Users can put machine learning methods' parameters after tuning. A more flexible evaluation is to consider "method=prob" in which case the input d should be a matrix of membership probabilities with k columns and each row of d should sum to one.

Author(s)

Ming Gao: [e-mail]

Jialiang Li: [e-mail]

References

Li, J., Gao, M., D’Agostino, R. (2019). Evaluating Classification Accuracy for Modern Learning Approaches. Statistics in Medicine (Tutorials in Biostatistics). 38(13): 2477-2503.

Van Calster B, Vergouwe Y, Looman CWN, Van Belle V, Timmerman D and Steyerberg EW. Assessing the discriminative ability of risk models for more than two outcome categories. European Journal of Epidemiology 2012; 27: 761 C 770.

Li, J., Feng, Q., Fine, J.P., Pencina, M.J., Van Calster, B. (2018). Nonparametric estimation and inference for polytomous discrimination index. Statistical Methods in Medical Research. 27(10): 3092—3103.

See Also

hum

Examples

str(iris)
data <- iris[, 3]
label <- iris[, 5]
pdi(y = label, d = data,method = "multinom")

## Call:
## pdi(y = label, d = data, method = "multinom")

## Overall Polytomous Discrimination Index:
##  0.9845333

## Category-specific Polytomous Discrimination Index:
##   CATEGORIES VALUES
## 1          1 1.0000
## 2          2 0.9768
## 3          3 0.9768

pdi(y = label, d = data,method = "tree")
pdi(y = label, d = data,method = "tree",control = rpart::rpart.control(minsplit = 200))

data <- data.matrix(iris[, 3])
label <- as.numeric(iris[, 5])
# multinomial
require(nnet)
# model
fit <- multinom(label ~ data, maxit = 1000, MaxNWts = 2000)
predict.probs <- predict(fit, type = "probs")
pp<- data.frame(predict.probs)
# extract the probablity assessment vector
head(pp)
pdi(y = label, d = pp, method = "prob")