knitr::opts_chunk$set( collapse = TRUE, comment = "#>" ) library(knitr) library(reshape) library(trinROC) options(digits=5)
The package trinROC
helps to assess three-class Receiver Operating Characteristic (ROC) type data. It provides functions for three statistical tests as described in @noll2019 along with functions for Exploratory Data Analysis (EDA) and visualization.
We assume that the reader has some background in ROC analysis as well as some understanding of the terminology associated with Area Under the ROC Curve (AUC) and Volume Under the ROC Surface (VUS) indices. See @nakas2014 or @noll2019 for a concise overview.
This vignette consists of the following parts:
The package also contains two small data sets cancer
and krebs
mimicking results of clinical studies. The datasets are fairly small and used in this vignette (and in the examples of the help files) to illustrate the functionality of the functions.
We assume a three-class setting, where one or more classifier yields measurements $X = x$ on a continuous scale for the groups of healthy ($D^-$), intermediate ($D^0$) and diseased ($D^+$) individuals. By convention, larger values of $x$ represent a more severe status of the disease. The package trinROC
provides statistical tests to assess the discriminatory power of such classifiers. These tests are:
trinROC.test
), developed by @noll2019;trinVUS.test
), developed by @xiong2007;boot.test
), developed by @nakas2004.In this document, we refer to the tests by the corresponding R function name. All tests can assess a single classifier, as well as compare two paired or unpaired classifiers.
As their names suggest, the underlying testing approach is different and either based on VUS or on ROC. Hence, their null hypotheses differ as well, as illustrated now.
Given two classifiers, boot.test
and trinVUS.test
are based on the null hypothesis $VUS_1 = VUS_2$ with the $Z$-statistic
\begin{align} Z = \frac{\widehat{VUS}_1 - \widehat{VUS}_2}{\sqrt{\widehat{Var}(\widehat{VUS}_1) + \widehat{Var}(\widehat{VUS}_2) - 2 \widehat{Cov}(\widehat{VUS}_1,\widehat{VUS}_2)}}. \end{align}
If the data of the two classifiers is unpaired, the term $Cov(VUS_1,VUS_2)$ is zero. If a single classifier is investigated, the null hypothesis is $VUS_1 = 1/6$ with the $Z$-statistic
\begin{align} Z = \frac{\widehat{VUS}_1 - 1/6}{\sqrt{\widehat{Var}(\widehat{VUS}_1)}}, \end{align}
which is equivalent to compared the VUS of the classifier to the volume of an uninformative classifier. Details about the estimators are given in the aforementioned papers.
In the trinormal model, the ROC surface is given by
\begin{align} ROC(t_-,t_+) = \Phi \left(\frac{\Phi^{-1} (1-t_+) + d}{ c} \right) - \Phi \left(\frac{\Phi^{-1} (t_-)+b}{a} \right), \end{align}
where $\Phi$ is the cdf of the standard normal distribution and $a$, $b$, $c$ and $d$ are functions of the means and standard deviations of the three groups:
\begin{align} a = \frac{{\sigma}0}{{\sigma}-}, \qquad b = \frac{ {\mu}- - {\mu}_0}{{\sigma}-}, \qquad c = \frac{{\sigma}0}{{\sigma}+}, \qquad d = \frac{ {\mu}+ - {\mu}_0}{{\sigma}+}. \end{align}
Given two classifiers, trinROC.test
investigates the shape of the two ROC surfaces. Hence, the resulting null hypothesis is $a_1=a_2$, $b_1=b_2$, $c_1=c_2$ and $d_1=d_2$, i.e., the the surfaces have the same shape.
Under the null hypothesis, the test statistic
\begin{align}
\chi^2 =
\begin{pmatrix} \widehat{a}_1 - \widehat{a}_2 &\widehat{b}_1 -\widehat{b}_2 & \widehat{c}_1-\widehat{c}_2 & \widehat{d}_1-\widehat{d}_2 \end{pmatrix}
{ \widehat{\boldsymbol{W}}}^{-1}
\begin{pmatrix} \widehat{a}_1 - \widehat{a}_2 \
\widehat{b}_1 - \widehat{b}_2 \ \widehat{c}_1-\widehat{c}_2 \ \widehat{d}_1- \widehat{d}_2 \end{pmatrix},
\end{align}
is distributed approximately as a chi-squared random variables with four degrees of freedom, where $\widehat{\boldsymbol{W}}$ contains the corresponding estimated variances and covariances. The test rejects if $\chi^2 > \chi_{\alpha}^2$, i.e., if the test statistics exceeds the chi-squared quantile with four degrees of freedom of a pre-defined confidence level $\alpha$.
When the data is unpaired, $\widehat{a}_1$, $\widehat{b}_1$, $\widehat{c}_1$, $\widehat{d}_1$ are independent from $\widehat{a}_2$, $\widehat{b}_2$, $\widehat{c}_2$, $\widehat{d}_2$, respectively, and hence every combination of covariances between them is zero.
It is also possible to investigate a single classifier with the above method. Instead of an existing second classifier, we compare the estimates $\widehat{a}1$ , $\widehat{b}_1$, $\widehat{c}_1$ and $\widehat{d}_1$ of a single marker with those from an artificial marker. If we want to detect whether a single marker is significantly better in allocating individuals to the three classes than a random allocation function we would set the parameters of our null hypothesis $a{Ho} = 1= c_{Ho}$, as we assume equal spread in the classes and $b_{Ho} = 0 = d_{Ho}$, as we impose equal means. This yields the null hypothesis $a_1=1$, $b_1=0$, $c_1=1$, $d_1=0$, which leads to the chi-squared test
\begin{align} {\chi}^2 = & \begin{pmatrix} \widehat{a}_1 -1 & \widehat{b}_1 & \widehat{c}_1 -1 & \widehat{d}_1 \end{pmatrix} \widehat{\boldsymbol{W}}^{-1} \begin{pmatrix} \widehat{a}_1 -1 \ \widehat{b}_1 \ \widehat{c}_1 - 1 \ \widehat{d}_1 \end{pmatrix}. \end{align}
We now illustrate the use of the test functions with the artificial dataset cancer
.
data(cancer) str(cancer)
The first column is a factor indicating the (true) class membership of each individual. The three levels have to be ordered according to heaviness of disease, i.e., healthy, intermediate and diseased (nor the names nor the sorting of the elements plays a role). The other columns contain the measurements yielded by Classifier 1 to 9. Further we note that some columns where multiplied by $-1$ in order to fulfill the convention that more diseased individuals have (in general) higher measurements.
For illustration, let us assess Class2
of the data set cancer
using the function trinROC.test
.
out <- trinROC.test(dat = cancer[,c("trueClass","Class2")]) out
The function returns a list of class "htest"
, similar to other tests from the stats
package. Additionally to the standard list elements, we also obtain detailed information about data and the sample estimates (VUS, and if applicable, estimates of $a$, $b$, $c$ and $d$)
out[ c("estimate", "Summary", "CovMat")]
More specifically:
$Summary
displays a summary table of $n_\ell$, $\mu_\ell$ and $\sigma_\ell$ for $\ell = -,0,+$, the three classes $D^-$, $D^0$ and $D^+$.$CovMat
or $Sigma
displays the covariance matrix of the test.The tests trinVUS.test
and boot.test
work analogously.
ROCsin <- trinROC.test(dat = cancer[,c(1,3)]) VUSsin <- trinVUS.test(dat = cancer[,c(1,3)]) bootsin <- boot.test(dat = cancer[,c(1,3)], n.boot = 250) c( ROCsin$p.value, VUSsin$p.value, bootsin$p.value)
The test functions trinROC.test
, trinVUS.test
and boot.test
handle either data frames that have the same form as cancer
or single vectors specifying the three groups. For example
(x1 <- with(cancer, cancer[trueClass=="healthy", 3])) (y1 <- with(cancer, cancer[trueClass=="intermediate", 3])) (z1 <- with(cancer, cancer[trueClass=="diseased", 3])) ROCsin2 <- trinROC.test(x1, y1, z1) ## All numbers are equal; sole difference is name of data: all.equal(ROCsin, ROCsin2, check.attributes = FALSE)
Assume we now want to compare Class2
with Class4
. If the data is paired, we take this into account by setting paired = TRUE
.
ROCcomp <- trinROC.test(dat = cancer[,c(1,3,5)], paired = TRUE) ROCcom <- trinROC.test(dat = cancer[,c(1,3,5)]) ROCcomp$p.value ROCcom$p.value # is equal to: x2 <- with(cancer, cancer[trueClass=="healthy", 5]) y2 <- with(cancer, cancer[trueClass=="intermediate", 5]) z2 <- with(cancer, cancer[trueClass=="diseased", 5]) ROCcomp2 <- trinROC.test(x1, y1, z1, x2, y2, z2, paired = TRUE)
Beside the argument paired
it is also possible to adjust the confidence level (default is conf.level = 0.95
). For the trinVUS.test
and the boot.test
one can also specify the alternative hypothesis (alternative = c("two.sided", "less", "greater")
). This not possible for the trinROC.test
, as this is a chi-squared test.
In this section, we outline the code used to construct the empirical power curves of Figures 2, 3 etc. For simplicity, here and in the paper, we set the mean and variances if the healthy group to zero and one, respectively. The variances of the intermediate and diseased groups are pre-specified ("slight crossing"). Finally the remaining means are found based on a desired VUS.
require( ggplot2, quietly = TRUE) require( MASS, quietly = TRUE) N <- 25 reps <- 99 # Is set to 1000 in the paper rho <- 0.5 # paired setting if rho!=0 sd.y1 <- 1.25; sd.y2 <- 1.5 # this corresponds to medium crossing sd.z1 <- 1.5; sd.z2 <- 2 Vus <- c(0.2, 0.25, 0.3, 0.35, 0.4, 0.45) lVus <- length(Vus) result <- matrix(0, lVus, reps) tmp <- findmu(sdy=sd.y1, sdz=sd.z1, VUS=Vus[1]) mom1 <- tmp[,2] names(mom1) <- tmp[,1] for (m in 1:lVus){ # cycle over different VUS mom2 <- findmu(sdy=sd.y2, sdz=sd.z2, VUS=Vus[m])[,2] names(mom2) <- tmp[,1] for( i in 1:reps) { # cycle over replicates SigmaX <- matrix(c(1, rho, rho, 1), 2, 2) SigmaY <- matrix(c(sd.y1^2, sd.y1*sd.y2*rho, sd.y1*sd.y2*rho, sd.y2^2), 2, 2) SigmaZ <- matrix(c(sd.z1^2, sd.z1*sd.z2*rho, sd.z1*sd.z2*rho, sd.z2^2), 2, 2) x <- mvrnorm(N, c(0, 0), SigmaX) y <- mvrnorm(N, c(mom1["muy"], mom2["muy"]), SigmaY) z <- mvrnorm(N, c(mom1["muz"], mom2["muz"]), SigmaZ) MT <- trinROC.test(x1 = x[,1], y1 = y[,1], z1 = z[,1], x2 = x[,2], y2 = y[,2], z2 = z[,2], paired = (rho!=0)) result[m,i] <- MT$p.value } } empPow <- data.frame(x = Vus, value = rowMeans(result<0.05)) ggplot(data = empPow, aes(x = Vus, y = value)) + geom_line() + geom_point() + ylab("Empirical Power") + scale_y_continuous(breaks = c(0.05, 0.25, 0.5, 1))
For the simulation study, we additionally calculated the p-values of boot.test
and varied $VUS_1$, $VUS_2$ and $N$.
In this section we discuss functions of the package trinROC
that help in the process of exploring, preparing and visualizing the data in the context of ROC analysis.
Formal statistical testing is typically preceded by an exploratory data analysis.
The function roc.eda
serves this purpose and provides three different viewpoints on its input data:
rgl
.The last option can be turned off by setting plotVUS = FALSE
. If only the interactive three-dimensional plot is desired, one can also use the functions rocsurf.emp
for empirical ROC surfaces and rocsurf.trin
for trinormal ROC surfaces.
data( cancer) roc.eda(dat = cancer[,c(1,5)], type = "trinormal", plotVUS = FALSE, saveVUS = TRUE)
roc.eda(dat = cancer[,c(1,5)], type = "empirical", sep.dens = TRUE, scatter = TRUE, verbose = FALSE) ## last call is equal to: # x <- with(cancer, cancer[trueClass=="healthy", 5]) # y <- with(cancer, cancer[trueClass=="intermediate", 5]) # z <- with(cancer, cancer[trueClass=="diseased", 5]) # roc.eda(x, y, z, type = "trinormal")
By setting plotVUS = TRUE
an interactive rgl plot window is opened, displaying the ROC surface computed from the measurements. Depending the argument type
the empirical or trinormal ROC surface is computed. The below Figures displays the empirical and trinormal snapshot of the ROC surfaces of the example data used in this section.
include_graphics(c("Figures//empVUS.png", "Figures//trinVUS.png"))
To calculate the VUS (empirical or estimate based on the trinormal assumption) the following code can be used.
data( cancer) x <- with(cancer, cancer[trueClass=="healthy", 5]) y <- with(cancer, cancer[trueClass=="intermediate", 5]) z <- with(cancer, cancer[trueClass=="diseased", 5]) emp.vus(x, y, z) trinVUS.test(x, y, z)$estimate trinROC.test(x, y, z)$estimate[1]
The ROC surface itself is visualized using rocsurf.emp(x, y, z)
or rocsurf.trin(x, y, z)
.
boxcoxROC
The trinormal model based test, trinROC.test
or trinVUS.test
are build upon a normality assumption of the data. If this assumption is violated, the trinormal tests may yield incorrect results. A common way to test for normality is the shapiro.test
. If the hypothesis of normally distributed data is rejected, there is a possibility to apply the function boxcoxROC
to the data. This function takes three vectors x
, y
and z
and computes a Box-Cox transformation, see @box1964 and @bantis2017. Consider this short example:
set.seed(712) x <- rchisq(20, 2) y <- rchisq(20, 6) z <- rchisq(20, 10) boxcoxROC(x, y, z)
roc3.test
The function roc3.test
computes every one by one combination of classifiers it contains to a desired combination of the three statistical tests trinROC.test
, trinVUS.test
and boot.test
in one step. Furthermore, single classifier assessment tests are automatically computed as well.
The output consists of:
Two data frames that contain information about each marker on its own. In the first data frame overview
the markers are sorted by their empirical VUS, while in the second data frame they are shown in their original order.
For each statistical test that has been chosen, two strictly upper triangular matrices are returned, one containing the p-values and one the test values.
out <- roc3.test(cancer[,1:8], type = c("ROC", "VUS"), paired = TRUE) out[c(1,3)]
Several of the arguments of roc3.test
are passed to the corresponding test functions specified by type
(one or several of "ROC"
, "VUS"
, "Bootstrap"
). These are (with corresponding default values) paired = FALSE
, conf.level = 0.95
and n.boot = 1000
.
The FDR p-value adjustment to be applied set p.adjust = TRUE
.
roc3.test(cancer[,1:8], type = c("ROC", "VUS"), paired = TRUE, p.adjust = TRUE)$P.values$trinROC out$P.values$trinROC
@xiong2007 also show how to apply the trinormal VUS test to a set of more than two classifiers at once.
Any scripts or data that you put into this service are public.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.