OutlierDC (version 0.2.1)

odc: Outlier detection using quantile regression for censored data

Description

outlier detection algorithms using quantile regression for censored data

Usage

odc(formula, data, method = c("score", "boxplot","residual"), 
      rq.model = c("Wang", "PengHuang", "Portnoy"), k = 1.5, h = .05 )

Arguments

formula

a type of Formula object with a survival object on the left-hand side of the ~ operator and covariate terms on the right-hand side. The survival object with survival time and its censoring status is constructed by the

data

a data frame with variables used in the formula. It needs at least three variables, including survival time, censoring status, and covariates.

method

the outlier detection method to be used. The options "socre", "boxplot", and "residual" conduct the scoring, boxplot, and residual-based algorithm, respectively. The default algorithm is "score".

rq.model

the type of censored quantile regression to be used for fitting. The options "Wang", "Portnoy", and "PengHuang" conduct Wang and Wang's, Portnoy's, and Peng and Huang's censored quantile regression approaches, respectively. The d

k

a constant to control the tightness of cut-offs for residual-based and boxplot algorithms with a default value of 1.5.

h

bandwidth for locally weighted censored quantile regression with a default value of 0.05.

Value

• an object of the S4 class "OutlierDC" with the following slots: call: evaluated function call formula: formula to be used raw.data: data to be used for model fitting refined.data: the data set after removing outliers coefficients: the estimated censored quantile regression coefficient matrix consisting of 10th, 25th, 50th, 75th, and 90th quantiles fitted.mat: the censored quantile regression fitted value matrix consisting of 10th, 25th, 50th, 75th, and 90th quantiles score: outlying scores (scoring algorithm) or residuals (residual-based algorithm) cutoff: estimated scale parameter for the residual-based algoritm lower: lower fence vector used for the boxplot and scoring algorithms upper: upper fence vector used for the boxplot and scoring algorithms outliers: logical vector to determine which observations are outliers n.outliers: number of outliers detected method: outlier detection method to be used rq.model: censored quantile regression to be used k: constant value to be used for the tightness of cut-offs in the residual-based and boxplot algorithms UB: upper fence used for the scoring algorithm with the update function LB: lower fence used for the scoring algorithm with the update function

Details

The odc function conducts three outlier detection algorithms on the basis of censored quantile regression. Three outlier detection algorithms were implemented: residual-based, boxplot, and scoring algorithms. The residual-based algorithm detects outlying observations using constant scale estimates; however, it does not account for the heterogeneity of variablity. When the data is extremely heterogeneous, the boxplot algorithm with censored quantile regression is more effective. The residual-based and boxplot algorithms produce cut-offs to determine whether observations are outliers. In contrast, the scoring algorithm provides the outlying magnitude or deviation of each point from the distribution of observations. Outlier detection is achieved by visualizing the scores.

References

Eo SH, Hong SM and Cho H (2013+). OutlierDC: Outlier Detection Algorithms using Quantile Regression for Censored Survival Data. submitted.

See Also

OutlierDC-package coef, plot, show, update

Examples

require(OutlierDC)
    data(ebd)
    
    #scoring algorithm
    fit <- odc(Surv(log(time), status) ~ meta, data = ebd)
    fit
    coef(fit)
    plot(fit)
    fit1 <- update(fit, UB = 4, LB = -4)
    fit1
    plot(fit1)

    fit2 <- update(fit, UB = NA, LB = -4)
    fit3 <- update(fit, UB = 4, LB = NA)
 
    # residual-based algorithm
    fit4 <- odc(Surv(log(time), status) ~ meta, data = ebd, method = "residual", k = 3)
    fit4
    plot(fit4)
    
    # boxplot algorithm
    fit5 <- odc(Surv(log(time), status) ~ meta, data = ebd, method = "boxplot", k = 1.5)
    fit5
    plot(fit5, ylab = "log survival times", xlab = "metastasis lymph nodes")