IMR: Invariant Model Ranking

In Laksafoss/ICPSurv: Invariant Causal Prediction for Survival Data

Description

A method for ranking models based on an invariance score.

Usage

IMR(M, method = "BS", ...)

Arguments

M	A list of models. The models may be the result of a call to coxph, aareg, timecox, aalen and cox.aalen.
method	The method used for calculating invariance scores. At the moment only two ranking methods have been implemented, namely the BS and rss method. The default method is BS, which fits a b-spline and reports the curvature as an invariance score.
...	additional arguments to be passed to the internal construct_analysis functions.

Details

The IMR function takes a list of fitted models as input and computes an invariance ranking. The invariance scoring method is specified via method. It is the users responsibility to input comparable models in M and choose a suiteble ranking scheem via method; only rudimentery sanity check of the inputted models in M is conducted.

At the moment the IMR function only takes hazard models as input, and only two ranking methods have been implementd, namely the BS and the rss methods:

The BS method: For each model a basis spline is fitted to the cumulative coefficients and the integrated curvature is reported as a measure of invariance. Before the spline is fitted the cumulative coefficients are normalized. A smaller score is evidence of a more invariant model.

The rss method: For each model the corresponding invariant model is found and the residual square error is found and reported as a measure of invariance. Before the invariant model and residual square error is found the cumulative coefficients are normalized. A smaller score is evidence of a more invariant model.

Value

IMR returns an object of class "IMR" with the following components:

ranking	A data.frame summarizing the findings. For detailes on how to interpret this see the method specific details above.
results	A list with the detailed results of the analysis.
method	A list with the method, a data.frame method_expanded with a more detailed summary of the called method, the character critical giving a one line interpretation guide to the ranking data.frame. This list may also contain other method specific components.
call	The matched call.

Examples

## Generate Data
n <- 1000
E <- sample(4L, n, replace = TRUE)
X <- data.frame(X1 = rnorm(n, 2 * E, 1),
                X2 = rnorm(n, 2 * (E == 1) + 5 * (E == 4)  , 1),
                X3 = rbinom(n, 1, 0.7))
Y <- rexp(n, exp(4 - 0.6 * X$X1 + 0.4 * X$X2))
C <- rexp(n, exp(0.5))
time <- pmin(Y, C)
status <- time == Y
# Note that X1 and X2 are the true causal predictors

## Fit Models
m1 <- survival::coxph(survival::Surv(time, status) ~ X1, data = X)
m2 <- survival::coxph(survival::Surv(time, status) ~ X2, data = X)
m3 <- survival::coxph(survival::Surv(time, status) ~ X3, data = X)
m12 <- survival::coxph(survival::Surv(time, status) ~ X1 + X2, data = X)
m13 <- survival::coxph(survival::Surv(time, status) ~ X1 + X3, data = X)


# Ranking via Basis Splines
IMR(list(m1, m2, m3, m12, m13), method = "BS")

# Ranking via Residual Square Error
IMR(list(m1, m2, m3, m12, m13), method = "rss")

Laksafoss/ICPSurv documentation built on Feb. 26, 2020, 11:32 a.m.