
Net Benefit for Dynamic Predictions under Left-Truncated and Right-Censored Data
tvNB.RdComputes estimates of the net benefit for a given decision threshold (decision_threshold) at a specified horizon time (Thoriz), using longitudinal information available up to a given starting time point (Tstart). The function supports joint models, Cox proportional hazards models and predictions obtained from any method. It provides inverse probability weighting and model-based estimation methods.
Usage
tvNB(object, ...)
# Default S3 method
tvNB(object, Time, event, LT_Time = NULL, Tstart,
decision_threshold = NULL, Thoriz = NULL, Dt = NULL,
type_weights = c("model-based", "IPCW"),
preds_MB = NULL, name_model = NULL, ...)
# S3 method for class 'jm'
tvNB(object, newdata, Tstart,
decision_threshold = NULL, Thoriz = NULL, Dt = NULL,
type_weights = c("model-based", "IPCW"), name_model = NULL, ...)
# S3 method for class 'coxph'
tvNB(object, newdata, Tstart,
decision_threshold = NULL, Thoriz = NULL, Dt = NULL,
type_weights = c("model-based", "IPCW"), name_model = NULL, ...)Arguments
- object
an object inheriting from class
jmorcoxph, or a numeric vector containing the dynamic predictions to be evaluated.- newdata
a data.frame that contains the longitudinal and covariate information for the subjects for which prediction of survival probabilities is required. If a
jmobject is used, the names of the variables in this data.frame must be the same as in the data.frames that were used to fit the linear mixed effects and the event process model that were supplied as the two first argument ofjm. If acoxphis used it only applies to the event process. If the default method is used, the argument is not needed.- Tstart
numeric scalar denoting the time point up to which longitudinal information is to be used to derive predictions.
- decision_threshold
numeric scalar between 0 and 1 denoting the decision threshold used in the net benefit.
- Time, event, LT_Time
numeric vectors that must be specified when using the default method to compute the metric by directly using a vector of predictions, observed times (
Time), the event indicator (event) and (possibly) left-truncated times (LT_Time).- Thoriz
numeric scalar denoting the time point for which a prediction of the survival status is of interest;
Thorizmust be later thanTstartand eitherDtorThorizmust be specified. IfThorizisNULLis set equal toTstart + Dt.- Dt
numeric scalar denoting the length of the time interval of prediction; either
DtorThorizmust be specified.- type_weights
character string denoting the type of weights to use to account for censorting. Options are model-based (default) and inverse probability weighting (using the Kaplan-Meier estimate of the censoring distribution, and adjusting for left truncation if necessary).
- preds_MB
a numeric vector containing the model-based weights to be used for the individuals censored within (
Tstart,Thoriz]. It must be specified when using the default method and themodel-basedestimator.- name_model
character string denoting the name of the model or object used for prediction evaluation; default is
NULL, meaning that the name of the object in the environment will be used.- ...
additional arguments passed to
predict.jm().
Value
A list of class tvMetric with components:
- metric
a numeric scalar denoting the estimated net benefit.
- metric_type
the character string
'NB'.- nr
a numeric scalar denoting the number of subjects at risk at time
Tstart.- nint
a numeric scalar denoting the number of subjects with an event in (
Tstart,Thoriz].- ncens
a numeric scalar denoting the number of subjects with a censored time in (
Tstart,Thoriz].- Tstart
a copy of the
Tstartargument.- Thoriz
a copy of the
Thorizargument.- type_weights
a copy of the
type_weightsargument.- metric_info
a copy of the
decision_thresholdargument.- LT_MS
a logical that is
TRUEif left-truncated times have been used.- classObject
the class of
object.- nameObject
the name of the model; either a copy of the
name_modelargument, or the name ofobject.
References
Vickers, Andrew J and Elkin, Elena B (2006). Decision curve analysis: a novel method for evaluating prediction models. Medical Decision Making 26, 565–574.
Author
Arnau Garcia-Fernandez a.garciafernandez@erasmusmc.nl, Pedro Miranda-Afonso p.mirandaafonso@erasmusmc.nl and Dimitris Rizopoulos d.rizopoulos@erasmusmc.nl
Examples
# \donttest{
# We fit a multivariate joint model
pbc2.id$status2 <- as.numeric(pbc2.id$status != 'alive')
CoxFit <- coxph(Surv(years, status2) ~ sex, data = pbc2.id)
fm1 <- lme(log(serBilir) ~ ns(year, 3) * sex, data = pbc2,
random = ~ ns(year, 3) | id, control = lmeControl(opt = 'optim'))
fm2 <- lme(prothrombin ~ ns(year, 2) * sex, data = pbc2,
random = ~ ns(year, 2) | id, control = lmeControl(opt = 'optim'))
jointFit <- jm(CoxFit, list(fm1, fm2), time_var = "year", n_chains = 1L)
# Net benefit with decision threshold 0.3, prediction horizon 7
# and landmark time 4
tvNB(jointFit, newdata = pbc2, Tstart = 4, Dt = 3, decision_threshold = 0.3,
cores = 1L)
#>
#> Accuracy Metric for the Joint Model 'jointFit'
#>
#> Estimated Net Benefit with decision threshold 0.3: 0.0892
#> At time: 7
#> For the 225 subjects at risk at time 4
#> Number of subjects with an event in [4, 7): 34
#> Number of subjects with a censored time in [4, 7): 62
#> Accounting for censoring using model-based weights
#>
# Considering the Cox model and IPW approach for estimation:
tvNB(CoxFit, newdata = pbc2.id, Tstart = 4, Dt = 3, decision_threshold = 0.3,
type_weights = "IPCW")
#>
#> Accuracy Metric for the Cox Model 'CoxFit'
#>
#> Estimated Net Benefit with decision threshold 0.3: -0.0016
#> At time: 7
#> For the 225 subjects at risk at time 4
#> Number of subjects with an event in [4, 7): 50
#> Number of subjects with a censored time in [4, 7): 46
#> Accounting for censoring using
#> inverse probability of censoring Kaplan-Meier weights
#>
# Considering the vectors with predictions, observed times, event indicator and
# left-truncated times. We use the default method:
pbc2.id$lt <- runif(nrow(pbc2.id), 0, pbc2.id$years)
sfit0 <- summary(survfit(CoxFit, newdata = pbc2.id), times = 5)
sfit <- summary(survfit(CoxFit, newdata = pbc2.id), times = 5 + 3)
predictions <- 1.0 - (as.matrix(sfit$surv)[1L, ] / as.matrix(sfit0$surv)[1L, ])
tvNB(predictions, pbc2.id$years, pbc2.id$status2, pbc2.id$lt, Tstart = 5, Thoriz = 8,
type_weights = "IPCW", decision_threshold = 0.25)
#>
#> Accuracy Metric for the predictions 'predictions',
#> observed times 'pbc2.id$years', event indicator 'pbc2.id$status2'
#> and left-truncated times 'pbc2.id$lt'
#>
#> Estimated Net Benefit with decision threshold 0.25: 0.0261
#> At time: 8
#> For the 125 subjects at risk at time 5
#> Number of subjects with an event in [5, 8): 33
#> Number of subjects with a censored time in [5, 8): 44
#> Accounting for left truncation and right censoring using
#> nonparametric inverse probability weights
#>
# Now, we consider the default method, but using the model-based estimator:
ind_censoring <- pbc2.id$years > 5 & pbc2.id$lt < 5 & pbc2.id$years < 8 & pbc2.id$status2 == 0
sfit2_8 <- summary(survfit(CoxFit, newdata = pbc2.id[ind_censoring, ]),
times = 8)
sfit_cens <- summary(survfit(CoxFit, newdata = pbc2.id[ind_censoring, ]),
times = pbc2.id$years[ind_censoring])
preds_MB <- rep(NA, length(pbc2.id$years))
preds_MB[ind_censoring] <- 1.0 - (as.matrix(sfit2_8$surv)[1L, ] / as.matrix(sfit_cens$surv)[1L, ])
tvNB(predictions, Time = pbc2.id$years, event = pbc2.id$status2, LT_Time = pbc2.id$lt,
Tstart = 5, Dt = 3, type_weights = "model-based", preds_MB = preds_MB,
decision_threshold = 0.3)
#>
#> Accuracy Metric for the predictions 'predictions',
#> observed times 'pbc2.id$years', event indicator 'pbc2.id$status2'
#> and left-truncated times 'pbc2.id$lt'
#>
#> Estimated Net Benefit with decision threshold 0.3: 0.0124
#> At time: 8
#> For the 125 subjects at risk at time 5
#> Number of subjects with an event in [5, 8): 33
#> Number of subjects with a censored time in [5, 8): 44
#> Accounting for left truncation and right censoring using
#> model-based weights
#>
# }