ThresholdROC (version 2.2)

diagnostic: Diagnostic tests

Description

This function performs a diagnostic test in a test outcome - condition (presence/absence of disease) 2x2 table.

Usage

diagnostic(tab, method = c("par", "exact"),
  casecontrol = FALSE, p = NULL, conf.level = 0.95)

Arguments

tab

an object of class table or matrix in the following form: TP FP FN TN where TP=number of true positives, FP=number of false positives, FN=number of false negatives and TN=number of true negatives; that is

method

method for computing the confidence intervals for sensitivity, specificity, predictive values, accuracy and error rate. The user can choose between "par" (parametric) and "exact" (exact). The last one is recommended for small sam

casecontrol

were data collected in a case-control study? Default, FALSE.

p

disease prevalence (only when casecontrol=T; otherwise, this parameter is ignored).

conf.level

confidence level for the confidence intervals. Default, 95%.

Value

• A data.frame with ten rows and three columns containing the point estimate and confidence intervals for the following statistical measures: sensitivity, specificity, positive predictive value, negative predictive value, positive likelihood ratio, negative likelihood ratio, odds ratio, Youden index, accuracy and error rate.

Details

For details about the expressions for the statistical measures computed by this function, see References. Since sensitivity, specificity, predictive values, accuracy and error rate are proportions, their confidence intervals are computed using the functions prop.test (if method = "par") and binom.test (if method = "exact") from the stats package. Confidence intervals for the likelihood ratios are computed using the formulas proposed in Zhou et al (2002), Section 4.1.3. Furthermore, when likelihood ratios can not be computed due to division by 0, a correction is applied: 0.5 units are added to tab. Confidence intervals for the odds ratio are computed using the formulas proposed in Zhou et al (2002), Section 4.1.4. The same correction described before is applied when the odds ratio can not be computed due to division by 0. Confidence intervals for the Youden index are computed using the expression $$CI(1-\alpha) = (Y-z_{1-\alpha/2}*Var(Y), Y+z_{1-\alpha/2}*Var(Y)),$$ where $Y$ is the Youden index estimate, $z_{1-\alpha/2}$ is the $1-\alpha/2$ quantile of a $N(0, 1)$ distribution and $Var(Y)$ is the variance of the Youden index estimator, which is computed as $Var(Sensitivity)+Var(Specificity)$.

References

Zhou XH, Obuchowski NA and McClish DK. (2002). Statistical methods in diagnostic medicine. John Wiley and sons. Youden, WJ. (1950). Index for rating diagnostic tests. Cancer 3:32-35.

See Also

thres2

Examples

# example 1 (Zhou et al (2002))
japan <- matrix(c(56,23,6,78), ncol=2, byrow=TRUE)
colnames(japan) <- c("D","nD")
rownames(japan) <- c("+","-")
japan
p <- 0.196 # disease prevalence
diagnostic(japan, "par", casecontrol=TRUE, p=p)

# example 2
table <- matrix(c(22,2,3,3), ncol=2, byrow=TRUE)
diagnostic(table, "par")
diagnostic(table, "exact")

# example 3
table2 <- matrix(c(22,2,0,3), ncol=2, byrow=TRUE)
diagnostic(table2, "exact")

# example 4
# generate a random sample of diseased and non-diased subjects
n1 <- 100
n2 <- 100
set.seed(1234)
par1.1 <- 0
par1.2 <- 1
par2.1 <- 2
par2.2 <- 1
k1 <- rnorm(n1, par1.1, par1.2) # non-diseased
k2 <- rnorm(n2, par2.1, par2.2) # diseased
# threshold estimation
rho <- 0.2 # prevalence
thres <- thres2(k1, k2, rho, method="eq", ci.method="d")$T$thres
# diagnostic test using the threshold estimate
marker <- c(k1, k2) # biomarker
condition <- c(rep("nD", length(k1)), rep("D", length(k2))) # condition
test <- ifelse(marker<thres, "-", "+") # test outcome according to thres
# build the table
table3 <- table(test, condition)[2:1, ]
# diagnostic test
round(diagnostic(table3), 3)