DTDA.cif (version 1.0.2)

DTDAcif: Doubly Truncated Data Analysis, Cumulative Incidence Functions

Description

This function computes a nonparametric estimator of the cumulative incidences of competing risks under double truncation. The estimator generalizes the Efron-Petrosian NPMLE (Non-Parametric Maximun Likelihood Estimator) to the competing risks setting.

Usage

DTDAcif(x, u, v, comp.event, method = c("indep", "dep"), boot = F,
  B = 300, N.iter = 100, error = 1e-06)

Arguments

x

Numeric vector corresponding to the variable of ultimate interest.

u

Numeric vector corresponding to the left truncation variable.

v

Numeric vector corresponding to the right truncation variable.

comp.event

Competing risk indicator.

method

The method used to compute the nonparametric estimator. Use <U+2018>indep<U+2019> for independent truncation variables and ``dep`` for truncation variables possibly depending on the competing risk.

boot

Logical. If TRUE the bootstrap standard deviation of the cumulative incidences is calculated.

B

Number of bootstrap replicates.

N.iter

Maximum number of iterations.

error

Error criterion for convergence.

Value

A list containing:

• method: The method used to compute the estimator.

• biasf: The biasing function which reports the sampling probability for each Xi.

• cif.mas: The mass attached to each (Xi,Zi). The cumsum of cif.mas for Zi=j is the estimator of the j-th cumulative incidence function.

• data: The data corresponding to (X,Z) ordered with respect to X within each Z-value.

• sd.boot: The bootstrap standard deviation.

Acknowledgements

• Jacobo de U<U+00F1>a-<U+00C1>lvarez was supported by Grant MTM2017-89422-P (MINECO/AEI/FEDER, UE).

• Jos<U+00E9> Carlos Soage was supported by Grupos de Referencia Competitiva, Consolidaci<U+00F3>n y Estructuraci<U+00F3>n de Unidades de Investigaci<U+00F3>n Competitivas del SUG, Cons. de Cultura, Educaci<U+00F3>n e OU, Xunta de Galicia (GRC ED431C 2016/040).

Details

The nonparametric estimator is based on the Efron-Petrosian NPMLE (Efron and Petrosian, 1999). Actually, each pair (Xi,Zi) -where Xi stands for the variable of interest and Zi is the competing event indicator- is weighted by the jump of the Efron-Petrosian NPMLE at Xi (method=``indep"), or by a normalized version of the Efron-Petrosian NPMLE computed from the subset of (Xs,Zs)'s such that Zs=Zi (method=``dep''). The former is suitable when the truncating couple (U,V) is independent of (X,Z), while the latter is recommended when (U,V) and X are only conditionally independent given Z; see de U<U+00F1>a-<U+00C1>lvarez (2019) for a full description of the estimators and of their properties. When the competing event indicator is missing, the function simply computes the Efron-Petrosian NPMLE and the argument method has no role.

References

• de U<U+00F1>a-<U+00C1>lvarez, J. (2019). Nonparametric estimation of the cumulative incidences of competing risks under double truncation. Preprint.

• Efron, B. and Petrosian, V. (1999). Nonparametric methods for doubly truncated data. Journal of the American Statistical Association 94, 824-834.

Examples

# NOT RUN {
# }
# NOT RUN {
# }
# NOT RUN {
set.seed(1234)
n <- 50  # sample size

x <- runif(n, 0, 1)  # time variable of interest
z <- rbinom(n, 1, 1 / 4)   # competing event indicator

# truncation variables

u <- runif(n, -.25, .5)  # left truncation variable
v <- u + .75   # right truncation variable

# note: (u,v) is independent of (x,z) so both estimation methods are consistent

# truncating the sample:

for (i in 1:n) {
  while (u[i] > x[i] | v[i] < x[i]) {
    x[i] <- runif(1, 0, 1)
    z[i] <- rbinom(1, 1, 1 / 4)
    u[i] <- runif(1, -.25, .5)
    v[i] <- u[i] + .75
  }
}

# note: (u,v) since is independent of (x,z)
# both estimation methods are consistent:

res.i <- DTDAcif(x, u, v, z, method = "indep", boot = TRUE)
res.d <- DTDAcif(x, u, v, z, method = "dep", boot = TRUE)

oldpar <- par(mfrow=c(1,2))
plot(res.i, main = "Indep trunc", intervals = TRUE)
plot(res.d, main = "Cond indep trunc", intervals = TRUE)

summary(res.i)
summary(res.d)

plot(res.i$data$x, res.i$biasf, type = "s")  # the observational bias
# the observational bias, event 1
plot(res.d$data$x[res.d$data$z == 1], res.d$biasf$biasf_1, type = "s")
# the observational bias, event 2
lines(res.d$data$x[res.d$data$z == 2], res.d$biasf$biasf_2, type = "s", col = 2)
par(oldpar)
# }
# NOT RUN {
# }