Skip to contents

IPTW Cox Regression (Inverse Probability of Treatment Weighted)

Source: R/phreg.R
phreg_IPTW.Rd

Fits a Cox model with treatment weights $$w(A) = \sum_a I(A=a)/\pi(a|X)$$, where $$\pi(a|X) = P(A=a|X)$$.

Usage

phreg_IPTW(
  formula,
  data,
  treat.model = NULL,
  treat.var = NULL,
  weights = NULL,
  estpr = 1,
  pi0 = 0.5,
  se.cluster = NULL,
  ...
)

Arguments

formula

Formula for phreg.

data

Data frame for risk averaging.

treat.model

Propensity score model (binary or multinomial).

treat.var

A 1/0 variable indicating when treatment is given (for time-dependent weights).

weights

Optional weights to multiply with the IPTW weights.

estpr

(=1, default) to estimate propensity scores and include their uncertainty in the influence function.

pi0

Fixed simple weights (if estpr=0).

se.cluster

To compute GEE-type standard errors when additional cluster structure is present.

...

Arguments for phreg call.

Value

An object of class "phreg" with additional IPTW components:

IID

Influence functions including propensity score uncertainty.

iptw

IPTW weights used.

naive.var

Naive variance ignoring propensity score uncertainty.

Details

Standard errors are computed via influence functions that are returned as the IID argument. Propensity scores are fitted using either logistic regression (glm) or the multinomial model (mlogit) when there are more than two treatment categories.

The treatment variable must be a factor and is identified on the RHS of the treat.model. Recurrent events can be considered with a start-stop structure, requiring cluster(id). Robust standard errors are computed in all cases.

Time-dependent propensity score weights can be computed when treat.var is used. This weight be 1 at the time of first (A_0) and 2nd treatment (A_1), then uses weights $$w_0(A_0) * w_1(A_1)^{t>T_r}$$ where $$T_r$$ is time of 2nd randomization. The weights are constructed using a glm or mlogit model based on the data where treat.var=1. The propensity score can be constructed for any number of treatments in a similar manner.

Author

Thomas Scheike

Examples

data <- mets:::simLT(0.7, 100, beta = 0.3, betac = 0, ce = 1, betao = 0.3)
dfactor(data) <- Z.f ~ Z
out <- phreg_IPTW(Surv(time, status) ~ Z.f, data = data, treat.model = Z.f ~ X)
summary(out)
#> 
#>    n events
#>  100     38
#> 
#>  100 clusters
#> coefficients:
#>      Estimate     S.E.  dU^-1/2 P-value
#> Z.f1 -0.84702  0.27941  0.23943  0.0024
#> 
#> exp(coefficients):
#>      Estimate    2.5%  97.5%
#> Z.f1  0.42869 0.24792 0.7413
#>