Compute the concordance statistic for data or a model

Description

The concordance statistic compute the agreement between an observed response and a predictor. It is closely related to Kendall's tau-a and tau-b, Goodman's gamma, and Somers' d, all of which can also be calculated from the results of this function.

Usage

concordance(object, ...)
## S3 method for class 'formula'
concordance(object, data, weights, subset, na.action,
  cluster, ymin, ymax, timewt= c("n", "S", "S/G", "n/G2", "I"),
  influence=0, ranks = FALSE, reverse=FALSE, timefix=TRUE, keepstrata=10, ...)
## S3 method for class 'lm'
concordance(object, ..., newdata, cluster, ymin, ymax,
  influence=0, ranks=FALSE, timefix=TRUE, keepstrata=10)
## S3 method for class 'coxph'
concordance(object, ..., newdata, cluster, ymin, ymax,
  timewt= c("n", "S", "S/G", "n/G2", "I"), influence=0,
  ranks=FALSE, timefix=TRUE, keepstrata=10)
## S3 method for class 'survreg'
concordance(object, ..., newdata, cluster, ymin, ymax,
  timewt= c("n", "S", "S/G", "n/G2", "I"), influence=0,
  ranks=FALSE, timefix=TRUE, keepstrata=10)

Arguments

Details

The concordance is an estimate of Pr(x_i < x_j | y_i < y_j), for a model fit replace x with \hat y, the predicted response from the model. For a survival outcome some pairs of values are not comparable, e.g., censored at time 5 and a death at time 6, as we do not know if the first observation will or will not outlive the second. In this case the total number of evaluable pairs is smaller.

Relatations to other statistics: For continuous x and y, 2C- 1 is equal to Somers' d. If the response is binary, C is equal to the area under the receiver operating curve or AUC. For a survival response and binary predictor C is the numerator of the Gehan-Wilcoxon test.

A naive compuation requires adding up over all n(n-1)/2 comparisons, which can be quite slow for large data sets. This routine uses an O(n log(n)) algorithm. At each uncensored event time y, compute the rank of x for the subject who had the event as compared to the x values for all others with a longer survival, where the rank has value between 0 and 1. The concordance is a weighted mean of these ranks, determined by the timewt option. The rank vector can be efficiently updated as subjects are added to the risk set. For further details see the vignette.

The variance is based on an infinetesimal jackknife. One advantage of this approach is that it also gives a valid covariance for the covariance based on multiple different predicted values, even if those predictions come from quite different models. See for instance the example below which has a poisson and two non-nested Cox models. This has been useful to compare a machine learning model to a Cox model fit, say. It is absolutely critical, however, that the predicted values line up exactly, with the same observation in each row; otherwise the result will be nonsense. (Be alert to the impact of missing values.)

The timewt option is only applicable to censored data. In this case the default corresponds to Harrell's C statistic, which is closely related to the Gehan-Wilcoxon test; timewt="S" corrsponds to the Peto-Wilcoxon, timewt="S/G" is suggested by Schemper, and timewt="n/G2" corresponds to Uno's C. It turns out that the Schemper and Uno weights are computationally identical, we have retained both option labels as a user convenience. The timewt= "I" option is related to the log-rank statistic.

When the number of strata is very large, such as in a conditional logistic regression for instance (clogit function), a much faster computation is available when the individual strata results are not retained; use keepstrata=FALSE or keepstrata=0 to do so. In the general case the keepstrata = 10 default simply keeps the printout managable: it retains and prints per-strata counts if the number of strata is <= 10.

Value

An object of class concordance containing the following components:

Note

A coxph model that has a numeric failure may have undefined predicted values, in which case the concordance will be NULL.

Computation for an existing coxph model along with newdata has some subtleties with respect to extra arguments in the original call. These include

• tt() terms in the model. This is not supported with newdata.

• subset. Any subset clause in the original call is ignored, i.e., not applied to the new data.

• strata() terms in the model. The new data is expected to have the strata variable(s) found in the original data set, with concordance computed within strata. The levels of the strata variable need not be the same as in the original data.

• id or cluster directives. This has not yet been sorted out.

Author(s)

Terry Therneau

References

F Harrell, R Califf, D Pryor, K Lee and R Rosati, Evaluating the yield of medical tests, J Am Medical Assoc, 1982.

R Peto and J Peto, Asymptotically efficient rank invariant test procedures (with discussion), J Royal Stat Soc A, 1972.

M Schemper, Cox analysis of survival data with non-proportional hazard functions, The Statistician, 1992.

H Uno, T Cai, M Pencina, R D'Agnostino and Lj Wei, On the C-statistics for evaluating overall adequacy of risk prediction procedures with censored survival data, Statistics in Medicine, 2011.

See Also

coxph

Examples

fit1 <- coxph(Surv(ptime, pstat) ~ age + sex + mspike, mgus2)
concordance(fit1, timewt="n/G2")  # Uno's weighting

# logistic regression 
fit2 <- glm(I(sex=='M') ~ age + log(creatinine), binomial, data= flchain)
concordance(fit2)  # equal to the AUC

# compare multiple models 
options(na.action = na.exclude)   # predict all 1384 obs, including missing
fit3 <- glm(pstat ~ age + sex + mspike + offset(log(ptime)), 
            poisson, data= mgus2)
fit4 <- coxph(Surv(ptime, pstat) ~ age + sex + mspike, mgus2)
fit5 <- coxph(Surv(ptime, pstat) ~ age + sex + hgb + creat, mgus2)

tdata <- mgus2; tdata$ptime <- 60   # prediction at 60 months
p3 <- -predict(fit3, newdata=tdata) 
p4 <- -predict(fit4) # high risk scores predict shorter survival
p5 <- -predict(fit5)
options(na.action = na.omit)      # return to the R default

cfit <- concordance(Surv(ptime, pstat) ~p3 +  p4 + p5, mgus2)
cfit
round(coef(cfit), 3)
round(cov2cor(vcov(cfit)), 3)  # high correlation

test <- c(1, -1, 0)  # contrast vector for model 1 - model 2 
round(c(difference = test %*% coef(cfit),
        sd= sqrt(test %*% vcov(cfit) %*% test)), 3)