cchart: Contribution chart.

Description

Display a graph explaining how the risk prediction of a new observation is obtained from the model. All contributions to the linear predictor or latent variable are visualized and summed to obtain the linear predictor. The latter is then transformed into a risk estimate.

Usage

cchart(x, obs, filename, zerolevel = "zero", risklabel, sorted = FALSE,
  time, xmin, xmax)

Arguments

glm, coxph, mfp, multinom or ksvm object.

obs

A data.frame containing the predictor values of the observation for which the chart should be made.

filename

The name of the resulting file (default: cchart).

zerolevel

The value of the contributions that should be put to zero. If "zero", the contributions are represented as they are. If "min", for each predictor or set of predictors contributing to an interaction, the minimal observed value of the contribution in the training data is substracted from the contribution to ensure that the contribution is always positive. If "median" or "mean", the median or mean value is substracted from the contributions, respectively (default="zero"). See below for more details.

risklabel

A character string representing the label for the represented risk. For multinomial logistic regression models, a vector of risk labels should be provided. See the examples for an illustration of the approach.

sorted

logical. If TRUE the contributions are sorted in increasing order (default=FALSE).

time

The time at which the estimated survival should be calculated. As default, the estimated survival at median survival time is reported. If the median survival time can not be calculated, the estimated survival at the latest event time is reported. For objects that are not a member of the coxph class, this is redundant.

xmin

Minimal values of input variables to be represented on the visualization. These values only have an influence on continuous input variables.

xmax

Maximal values of input variables to be represented on the visualization. These values only have an influence on continuous input variables.

Details

The contribution chart is a bar plot representing the contribution of each predictor or set of predictors to the score (translated linear predictor) by means of bars. Depending on the value of zerolevel, the visualized contributions are slightly different. If zerolevel="zero", the contribution for variable \(x^p\) is \(\beta_pf_p(x^p)\), with \(\beta_p\) the model coefficient corresponding to this predictor and \(f_p(x^p)\) a (possible) transformation of \(x^p\). If zerolevel is "min", "median" or "mean", a value equal to the minimum, median and mean of the contribution \(\beta_pf_p(x^p)\) in the training data, respectively, is substracted from the contribution. See the references for more information. The sum of all the contributions is the score (i.e. translated linear predictor or latent variable) which is transformed into the risk estimate. The range of all contributions in the training set are visualized by means of black horizontal lines.

References

Van Belle V., Van Calster B., Visualizing risk prediction models, PLoS ONE, 10(7):e0132614. doi:10.1371/journal.pone.0132614 (2015).

Van Belle V., Van Calster B., Suykens J.A.K., Van Huffel S. and Lisboa P., Explaining support vector machines: a color based nomogram, Internal Report 16-27, ESAT-Stadius, KU Leuven (Leuven, Belgium), 2016

Van Belle V., Van Huffel S., Timmerman D., Froyman W., Bourne T. and Van Calster B., A color based nomogram for Multinomial Logistic Regression, Internal Report 16-28, ESAT-Stadius, KU Leuven (Leuven, Belgium), 2016

Examples

Run this code

# NOT RUN {
#### logistic regression
mydata <- read.csv("https://stats.idre.ucla.edu/stat/data/binary.csv")
mydata$rank <- factor(mydata$rank)
fit <- glm(admit ~ gre + gpa + rank, data = mydata, family = "binomial")
patient1=data.frame(gre=386,gpa=3.58,rank=3)
cchart(fit,obs=patient1,filename="cchart1")

#### cox proportional hazard regression
library(mfp)
data(GBSG)
fit<-coxph(Surv(rfst, cens) ~ age+tumsize+posnodal+prm+esm+menostat+tumgrad, data = GBSG, 
		model=TRUE)
patient1=data.frame(age=52,tumsize=45,posnodal=23,prm=1025,esm=562,menostat=2,tumgrad=3)
# Indicate the risk estimate at 500 days
cchart(fit,patient1,time=500,filename="cchart6")


#### multinomial logistic regression model
library(nnet)
library(VGAMdata)
data(xs.nz)
marital.nz <- xs.nz[,c("marital","sex","age","height","weight")]
mydata <- marital.nz[complete.cases(marital.nz),]
fit <- multinom(marital ~ sex + age + height + weight, data = mydata,model=TRUE)
# for multinimial logistic regression, a vector of risk labels needs to be made 
# and provided to the colplot function
outnames=colnames(fitted(fit))
labels=c(paste("Linear predictor for",outnames[-1]),paste
				("Predicted chance of being",outnames))
patient1=data.frame(sex="F",age=27,height=1.68,weight=58.6)
cchart(fit,obs=patient1,filename="multinom",risklabel=labels,zerolevel="zero")

#### Support Vector Machine classifier
# }
# NOT RUN {
	library(kernlab)
	data(iris)
	levels(iris$Species)[levels(iris$Species)=="setosa"] <- "other"
	levels(iris$Species)[levels(iris$Species)=="virginica"] <- "other"
	names(iris)=c("SL","SW","PL","PW","Species")
	set.seed(100)
	model <-ksvm(Species ~ ., data = iris,prob.model=TRUE,kpar=list(0.03),C=10)
	obs1=data.frame(SL=5.2,SW=3.0,PL=1.5,PW=0.3)
# The plot should be based on all training data, so the following code should be used:
	newmodel=preplotperf(model,iris,indy=5,zerolevel="median")
	cchart(newmodel,obs=obs1,filename="iris2",zerolevel="median")
# }