C-Statistic (Concordance Statistic): Formula, AUROC Equivalence, and Harrell’s C-Index

The C-statistic (concordance statistic, or C-index) is a non-parametric, rank-based measure of a statistical or machine learning model’s discrimination—its ability to separate individuals who develop an outcome of interest from those who do not (Harrell et al., 1982, 1996). Across clinical epidemiology, biostatistics, and TRIPOD+AI prediction model reporting, the C-statistic serves as the primary summary metric for discriminative accuracy in both binary logistic regression and right-censored time-to-event survival analysis.

What Is the C-Statistic?

Suppose a prediction model outputs a continuous risk score or predicted probability $hat{p}_i$ for each individual $i$. If we form all permissible pairs of individuals $(i, j)$ in which subject $i$ experiences the event (or experiences it earlier) and subject $j$ does not (or experiences it later), the C-statistic is the proportion of those pairs in which the model assigns a higher predicted risk to subject $i$ ($hat{p}_i > hat{p}_j$), with tied predictions receiving half-credit ($0.5$):

$$C = frac{n_{text{concordant}} + 0.5 cdot n_{text{tied}}}{n_{text{permissible}}}$$

The C-statistic ranges from $0.0$ to $1.0$:

• $C = 0.50$: No discriminative ability; the model orders pairs no better than a fair coin toss.
• $0.70 le C < 0.80$: Moderate discrimination (typical of many multivariable clinical risk scores such as cardiovascular risk equations).
• $0.80 le C < 0.90$: Strong discrimination (typical of acute organ-failure severity scores such as the MELD score).
• $C = 1.00$: Perfect discrimination; every case in the dataset receives a strictly higher predicted risk score than every non-case.
• $C < 0.50$: Inverse rank ordering (predictions are systematically reversed relative to the outcome).

1. Binary Outcomes: Mathematical Equivalence to AUROC and Mann–Whitney $U$

When the outcome $Y in {0, 1}$ is binary and completely observed (no censoring before the fixed assessment horizon), every pair consisting of one case ($Y_i = 1$) and one non-case ($Y_j = 0$) is permissible. If there are $n_1$ events and $n_0$ non-events, there are $n_{text{permissible}} = n_1 cdot n_0$ total pairs, and the binary C-statistic is:

$$C_{text{binary}} = P(hat{p}_i > hat{p}_j mid Y_i = 1, Y_j = 0) + frac{1}{2} P(hat{p}_i = hat{p}_j mid Y_i = 1, Y_j = 0)$$

As proven by Hanley and McNeil (1982), for binary outcomes:

1. $C_{text{binary}} equiv text{AUROC}$: The binary C-statistic equals the exact area under the Receiver Operating Characteristic (ROC) curve plotting sensitivity against $1 – text{specificity}$ across all possible cutpoints (including the threshold selected by Youden’s J index).
2. $C_{text{binary}} equiv frac{U}{n_1 n_0}$: The binary C-statistic equals the two-sample Wilcoxon–Mann–Whitney $U$ statistic divided by the number of case–control pairs $n_1 n_0$, and relates linearly to Somers’ $D_{xy}$ rank correlation via $D_{xy} = 2(C – 0.5)$.

2. Survival Analysis: Harrell’s C-Index and Permissible Pairs

In time-to-event studies modeled with Cox proportional hazards regression or accelerated failure time models, follow-up times $T_i = min(T_i^*, C_i^*)$ are subject to right censoring ($delta_i = 1$ if the event occurred at $T_i$; $delta_i = 0$ if censored at $T_i$). Standard binary AUROC cannot be computed directly across all pairs because the true ordering of event times is unknown when censoring occurs first.

Harrell’s C-index (Harrell et al., 1982, 1996) solves this by restricting the denominator to permissible (comparable) pairs—those pairs $(i, j)$ in which it is known with certainty which subject experienced the event earlier:

Pair Observation Pattern (Assuming $T_i le T_j$)	Censoring Indicators $(delta_i, delta_j)$	Permissible Pair?	Reason
$T_i < T_j$	$delta_i = 1,; delta_j = 1$	Yes	Both subjects had the event, and subject $i$ had the event strictly earlier than subject $j$.
$T_i < T_j$ (or $T_i = T_j$)	$delta_i = 1,; delta_j = 0$	Yes	Subject $i$ experienced the event at $T_i$, whereas subject $j$ was known to be event-free at least through $T_j ge T_i$.
$T_i < T_j$	$delta_i = 0,; delta_j = 1$	No (Excluded)	Subject $i$ was censored at $T_i$ before subject $j$’s event at $T_j$; it is unknown whether $i$’s unobserved event occurs before or after $T_j$.
Any $T_i, T_j$	$delta_i = 0,; delta_j = 0$	No (Excluded)	Both subjects are right-censored; true event order is unknown.

For each permissible pair where subject $i$ failed earlier ($T_i < T_j$ with $delta_i = 1$), the pair is concordant if the model assigns a higher predicted risk (or higher linear predictor $hat{eta}_i > hat{eta}_j$ in a Cox model) to subject $i$.

Uno’s IPCW C-Statistic for Heavy Censoring

Because Harrell’s C-index omits pairs where the earlier observation is censored, its expected value shifts upward as censoring increases or follow-up length changes. Uno’s C-statistic (Uno et al., 2011) weights each permissible pair by the inverse squared Kaplan–Meier estimate of the censoring survival distribution $hat{G}(T_i)^{-2}$ up to a pre-specified clinical truncation time $tau$, yielding a consistent estimator that does not depend on the study censoring mechanism.

Worked Example: Calculating Harrell’s C-Index on 4 Patients

Consider $N = 4$ patients in a survival study with observed follow-up time $T$, event indicator $delta$ ($1 = text{event}$, $0 = text{censored}$), and predicted 5-year risk score $hat{r}$:

• Patient A: $T_A = 4$ months, $delta_A = 1$ (event), predicted risk $hat{r}_A = 0.82$
• Patient B: $T_B = 6$ months, $delta_B = 0$ (censored), predicted risk $hat{r}_B = 0.75$
• Patient C: $T_C = 10$ months, $delta_C = 1$ (event), predicted risk $hat{r}_C = 0.60$
• Patient D: $T_D = 15$ months, $delta_D = 0$ (censored), predicted risk $hat{r}_D = 0.65$

Among all $binom{4}{2} = 6$ possible pairs, we evaluate permissibility and concordance:

1. Pair (A, B): $T_A = 4 < T_B = 6$ and $delta_A = 1$. Permissible. Since $hat{r}_A (0.82) > hat{r}_B (0.75)$, the pair is Concordant ($1.0$).
2. Pair (A, C): $T_A = 4 < T_C = 10$ and $delta_A = 1$. Permissible. Since $hat{r}_A (0.82) > hat{r}_C (0.60)$, the pair is Concordant ($1.0$).
3. Pair (A, D): $T_A = 4 < T_D = 15$ and $delta_A = 1$. Permissible. Since $hat{r}_A (0.82) > hat{r}_D (0.65)$, the pair is Concordant ($1.0$).
4. Pair (B, C): $T_B = 6 < T_C = 10$, but $delta_B = 0$ (Patient B was censored before Patient C’s event). Not permissible (excluded).
5. Pair (B, D): Both censored ($delta_B = 0, delta_D = 0$). Not permissible (excluded).
6. Pair (C, D): $T_C = 10 < T_D = 15$ and $delta_C = 1$. Permissible. However, $hat{r}_C (0.60) < hat{r}_D (0.65)$, so the pair is Discordant ($0.0$).

With $4$ permissible pairs ($3$ concordant, $1$ discordant, $0$ tied), Harrell’s C-index is:

$$C = frac{3 + 0.5(0)}{4} = 0.75$$

Why the C-Statistic Must Be Paired With Calibration and Net Benefit

Because the C-statistic evaluates only rank order, it is completely insensitive to systematic miscalibration. If a clinical prediction model’s predicted probabilities are all divided by $3$, its C-statistic remains identical even though it severely underpredicts absolute risk at clinical decision cutoffs. Therefore, prediction model validation guidelines require reporting the C-statistic alongside calibration slope and calibration-in-the-large, calibration plots, and clinical utility via decision curve analysis.

Sources

• Harrell, F. E., Jr., Califf, R. M., Pryor, D. B., Lee, K. L., & Rosati, R. A. (1982). Evaluating the Yield of Medical Tests. JAMA, 247(18), 2543–2546. https://pubmed.ncbi.nlm.nih.gov/7069920/
• Harrell, F. E., Jr., Lee, K. L., & Mark, D. B. (1996). Multivariable Prognostic Models: Issues in Developing Models, Evaluating Assumptions and Adequacy, and Measuring and Reducing Errors. Statistics in Medicine, 15(4), 361–387. https://pubmed.ncbi.nlm.nih.gov/8668867/
• Uno, H., Cai, T., Pencina, M. J., D’Agostino, R. B., & Wei, L. J. (2011). On the C-statistics for Evaluating Overall Adequacy of Risk Prediction Procedures with Censored Survival Data. Statistics in Medicine, 30(10), 1105–1117. https://pmc.ncbi.nlm.nih.gov/articles/PMC3079915/