Inverse Probability of Censoring Weights

library(trtswitch)
library(dplyr, warn.conflicts = FALSE)
library(ggplot2)

Introduction

The inverse probability of censoring weights (IPCW) method is a powerful tool for adjusting survival analysis in the presence of treatment switching. In scenarios where patients switch treatments, survival data are artificially censored, leading to biased estimates if not properly addressed. This bias arises because the decision to switch treatments may depend on prognostic factors that also influence survival outcomes.

To mitigate this bias, researchers gather data on prognostic covariates at baseline and at regular intervals until one of the following occurs: a treatment switch, death, drop-out, or the end of the follow-up period. A “switching model” is then developed to estimate weights for each patient at each time point. These weights represent the inverse probability of remaining unswitched over time. Each patient retains a weight greater than 1, which accounts for both their own data and that of switchers with similar prognostic profiles.

Finally, an “outcome” model utilizing IPCW-weighted survival times is employed to estimate the treatment effect, adjusted for switching.

Estimation of weights

The switching model can be implemented using either pooled logistic regression or proportional hazards regression with time-dependent covariates.

Pooled Logistic Regression Model

Let \(A_{i,j}\) denote the indicator of alternative therapy for subject \(i\) in treatment cycle \(j\). We assume that once a patient switched to an alternative therapy, he/she will stay on the alternative therapy. Let \(X_i\) denote the baseline covariates for subject \(i\) and \(L_{i,j}\) denote the time-dependent covariates for subject \(i\) at the start of cycle \(j\). For the pooled logistic regression model for treatment switching, the full data will be used for patients who didn’t switch treatment, while the partial data up to and including the treatment cycle when the patient switched treatment will be used for patients who switched treatment. The logistic regression model is used to estimate \[P(A_j = a_{i,j} | A_{j-1} = 0, \bar{L}_{j} = \bar{l}_{i,j}, X = x_i)\] where \(x_i\), \(l_{i,j}\), and \(a_{i,j}\) denote the observed value of baseline covariates, time-dependent covariates at the start of cycle \(j\), and alternative therapy status for cycle \(j\) of subject \(i\), respectively. A subject is on alternative therapy in cycle \(j\) if and only if treatment switching has occurred at the start of that cycle, which is equivalent to switching having occurred by the end of cycle \(j−1\). Consequently, lagged values of time-dependent covariates are often used in the logistic regression model. Natural cubic splines can be used to model visit-specific intercepts for the pooled logistic regression model. The unstabilized weights for subject \(i\) for cycle \(t\) are given by \[w_i(t) = \prod_{j=1}^{t} \frac{1}{P(A_j = a_{i,j} | \bar{A}_{j-1} = \bar{a}_{i,j-1}, \bar{L}_{j} = \bar{l}_{i,j}, X = x_i)}\] In the IPCW method, all visits that occur after treatment switching are excluded from the weighted outcome model. Accordingly, the weights are defined as the inverse probability of not having switched at the start of each cycle.

When using pooled logistic regression for the switching model, the following specifications can be made:

Proportional Hazards Model

When using a Cox model with time-dependent covariates as the switching model, adjacent observations within a subject may be combined as long as the values of time-dependent covariates do not change over each combined interval. This approach saves data storage but observations across subjects will not be aligned at regular intervals. In this case, unique event times across treatment arms should be replicated within each subject, enabling the availability of weights at each event time.

In the time-dependent Cox switching model, survival times are artificially censored at the point of treatment switching. The IPCW weight is then defined as the probability of not having switched treatment by the end of the time interval. This is to account for the irregular visit structure for the time-dependent Cox switching model.

Stabilized Weights

To address the high variability in “unstabilized” weights, researchers often opt for “stabilized” weights. The switching model for stabilized weights involves a model for the numerator of the weight in addition to the model for the denominator. Typically, the numerator model mirrors the denominator model but includes only prognostic baseline variables. Any baseline variables included in the numerator must be part of the weighted outcome model. Any baseline variables included in the weighted outcome model must also be part of the denominator.

Truncation of Weights

You can specify weight truncation using the trunc and trunc_upper_only parameters of the ipcw function. The trunc parameter is a number between 0 and 0.5 that represents the fraction of the weight distribution to truncate. The trunc_upper_only parameter is a flag that indicates whether to truncate weights only from the upper end of the weight distribution.

Estimation of Hazard Ratio

Once weights have been obtained for each event time in the data set, we fit a (potentially stratified) Cox proportional hazards model to this weighted data set to obtain an estimate of the hazard ratio. The confidence interval for the hazard ratio can be derived by bootstrapping the weight estimation and the subsequent model-fitting process.

Example for Pooled Logistic Regression Switching Model

First we prepare the data.

sim1 <- tssim(
  tdxo = 1, coxo = 1, allocation1 = 1, allocation2 = 1,
  p_X_1 = 0.3, p_X_0 = 0.3, 
  rate_T = 0.002, beta1 = -0.5, beta2 = 0.3, 
  gamma0 = 0.3, gamma1 = -0.9, gamma2 = 0.7, gamma3 = 1.1, gamma4 = -0.8,
  zeta0 = -3.5, zeta1 = 0.5, zeta2 = 0.2, zeta3 = -0.4, 
  alpha0 = 0.5, alpha1 = 0.5, alpha2 = 0.4, 
  theta1_1 = -0.4, theta1_0 = -0.4, theta2 = 0.2,
  rate_C = 0.0000855, accrualIntensity = 20/30, 
  followupTime = 600, fixedFollowup = 0, days = 30,
  n = 500, NSim = 100, seed = 314159)

Given that the observations are collected at regular intervals, we can use a pooled logistic regression switching model. In this model, the prognostic baseline variable is represented by bprog, while L serves as a time-dependent covariate. Additionally, ns1, ns2, and ns3 are the three terms for the natural cubic spline with \(\nu=3\) degrees of freedom for visit-specific intercepts.

fit1 <- ipcw(
  sim1[[1]], id = "id", tstart = "tstart", 
  tstop = "tstop", event = "Y", treat = "trtrand", 
  swtrt = "xo", swtrt_time = "xotime", 
  base_cov = "bprog", numerator = "bprog", 
  denominator = c("bprog", "L"),
  logistic_switching_model = TRUE, ns_df = 3,
  swtrt_control_only = TRUE, boot = FALSE)

The fits for the denominator and numerator switching models for the control arm are as follows, utilizing robust sandwich variance estimates to account for correlations among observations within the same subject.

# denominator switching model fit
fit1$fit_switch[[1]]$fit_den$parest[, c("param", "beta", "sebeta", "z")]
#>         param        beta    sebeta            z
#> 1 (Intercept) -3.68101349 0.3623806 -10.15786448
#> 2       bprog  0.02169405 0.2398362   0.09045364
#> 3           L  0.35036304 0.2728040   1.28430313
#> 4         ns1 -0.25774949 0.5444454  -0.47341658
#> 5         ns2  0.65961680 0.7229768   0.91236232
#> 6         ns3  0.05395416 0.6817778   0.07913746

# numerator switching model fit
fit1$fit_switch[[1]]$fit_num$parest[, c("param", "beta", "sebeta", "z")]
#>         param        beta    sebeta           z
#> 1 (Intercept) -3.44467907 0.3054745 -11.2764849
#> 2       bprog  0.07952921 0.2362668   0.3366077
#> 3         ns1 -0.25793955 0.5435290  -0.4745645
#> 4         ns2  0.70028647 0.7197425   0.9729680
#> 5         ns3  0.07212922 0.6776669   0.1064376

Since treatment switching is not allowed in the experimental arm, the weights are identical to 1 for subjects in the experimental group. The unstabilized and stablized weights for the control group are plotted below.

# unstabilized weights
ggplot(fit1$data_outcome %>% filter(trtrand == 0), 
       aes(x = unstabilized_weight)) + 
  geom_density(fill="#77bd89", color="#1f6e34", alpha=0.8) +
  scale_x_continuous("unstabilized weights")


# stabilized weights
ggplot(fit1$data_outcome %>% filter(trtrand == 0), 
       aes(x = stabilized_weight)) + 
  geom_density(fill="#77bd89", color="#1f6e34", alpha=0.8) +
  scale_x_continuous("stabilized weights")

Now we fit a weighted outcome Cox model and compare the treatment hazard ratio estimate with the reported.

fit1$fit_outcome$parest[, c("param", "beta", "sebeta", "z")]
#>     param       beta    sebeta         z
#> 1 treated -0.4198107 0.1133337 -3.704200
#> 2   bprog  0.3779653 0.1164952  3.244472
    
exp(fit1$fit_outcome$parest[1, c("beta", "lower", "upper")])
#>        beta     lower     upper
#> 1 0.6571712 0.5262701 0.8206318

Example for Time-Dependent Cox Switching Model

Now we apply the IPCW method using a Cox proportional hazards model with time-dependent covariates as the switching model.

fit2 <- ipcw(
  shilong, id = "id", tstart = "tstart", tstop = "tstop", 
  event = "event", treat = "bras.f", swtrt = "co", 
  swtrt_time = "dco", 
  base_cov = c("agerand", "sex.f", "tt_Lnum", "rmh_alea.c", 
               "pathway.f"),
  numerator = c("agerand", "sex.f", "tt_Lnum", "rmh_alea.c", 
                "pathway.f"),
  denominator = c("agerand", "sex.f", "tt_Lnum", "rmh_alea.c",
                  "pathway.f", "ps", "ttc", "tran"),
  swtrt_control_only = FALSE, boot = FALSE)

The fits for the denominator and numerator switching models for the control arm are as follows, utilizing robust sandwich variance estimates to account for correlations among observations within the same subject.

# denominator switching model for the control group
fit2$fit_switch[[1]]$fit_den$parest[, c("param", "beta", "sebeta", "z")]
#>                    param         beta     sebeta           z
#> 1                agerand  0.007642073 0.01023088  0.74696131
#> 2            sex.fFemale -0.364211003 0.28631111 -1.27208128
#> 3                tt_Lnum  0.042406215 0.05314773  0.79789331
#> 4             rmh_alea.c -0.351616244 0.27416745 -1.28248719
#> 5            pathway.fHR -0.041768569 0.43444787 -0.09614173
#> 6 pathway.fPI3K.AKT.mTOR  0.267055320 0.43736340  0.61060281
#> 7                     ps  0.103478553 0.18544824  0.55799156
#> 8                    ttc -0.480378314 0.33557237 -1.43151926
#> 9                   tran  0.295828936 0.45806296  0.64582592

# numerator switching model for the control group
fit2$fit_switch[[1]]$fit_num$parest[, c("param", "beta", "sebeta", "z")]
#>                    param        beta      sebeta           z
#> 1                agerand  0.01059477 0.009501694  1.11503993
#> 2            sex.fFemale -0.32452112 0.283829120 -1.14336794
#> 3                tt_Lnum  0.05956492 0.051120376  1.16518947
#> 4             rmh_alea.c -0.29758402 0.266335628 -1.11732711
#> 5            pathway.fHR  0.02467202 0.429102680  0.05749678
#> 6 pathway.fPI3K.AKT.mTOR  0.28245603 0.434306994  0.65036030

The fits for the denominator and numerator switching models for the experimental arm are as follows:

# denominator switching model for the experimental group
fit2$fit_switch[[2]]$fit_den$parest[, c("param", "beta", "sebeta", "z")]
#>                    param         beta     sebeta          z
#> 1                agerand -0.002639441 0.01796294 -0.1469381
#> 2            sex.fFemale  0.428028469 0.47154276  0.9077193
#> 3                tt_Lnum -0.152758042 0.10774494 -1.4177746
#> 4             rmh_alea.c  0.210365664 0.50502449  0.4165455
#> 5            pathway.fHR  1.774466817 1.04918789  1.6912765
#> 6 pathway.fPI3K.AKT.mTOR  0.859950234 1.08995673  0.7889765
#> 7                     ps  0.261142698 0.27003546  0.9670682
#> 8                    ttc -0.408175689 0.53635218 -0.7610218
#> 9                   tran  1.053347294 0.69246856  1.5211482

# numerator switching model for the experimental group
fit2$fit_switch[[2]]$fit_num$parest[, c("param", "beta", "sebeta", "z")]
#>                    param         beta     sebeta           z
#> 1                agerand  0.001625622 0.01812956  0.08966694
#> 2            sex.fFemale  0.476979559 0.45994387  1.03703863
#> 3                tt_Lnum -0.160020273 0.10615092 -1.50747886
#> 4             rmh_alea.c  0.154833260 0.48178554  0.32137382
#> 5            pathway.fHR  1.841535014 1.03960949  1.77137188
#> 6 pathway.fPI3K.AKT.mTOR  1.064637540 1.06675981  0.99801055

Below, the unstabilized and stabilized weights are plotted by treatment group: the left panel displays data for the control group, while the right panel shows data for the experimental group.

# unstabilized weights
ggplot(fit2$data_outcome, aes(x = unstabilized_weight)) + 
  geom_density(fill="#77bd89", color="#1f6e34", alpha=0.8) + 
  scale_x_continuous("unstabilized weights") + 
  facet_wrap(~treated) 


# stabilized weights
ggplot(fit2$data_outcome, aes(x = stabilized_weight)) + 
  geom_density(fill="#77bd89", color="#1f6e34", alpha=0.8) + 
  scale_x_continuous("stabilized weights") + 
  facet_wrap(~treated)

Finally, we fit the weighted outcome Cox model and compare the treatment hazard ratio estimate with the reported.

fit2$fit_outcome$parest[, c("param", "beta", "sebeta", "z")]
#>                    param         beta     sebeta          z
#> 1                treated  0.356390611 0.25526832  1.3961412
#> 2                agerand -0.006047034 0.01018596 -0.5936636
#> 3            sex.fFemale -0.487409540 0.24458876 -1.9927716
#> 4                tt_Lnum  0.011244574 0.04147245  0.2711336
#> 5             rmh_alea.c  0.941651485 0.25315061  3.7197283
#> 6            pathway.fHR -0.127273307 0.35878557 -0.3547336
#> 7 pathway.fPI3K.AKT.mTOR -0.166035907 0.34997206 -0.4744262

c(fit2$hr, fit2$hr_CI)
#> [1] 1.4281653 0.8659517 2.3553924