PosViolMSM: Positivity Violations in Marginal Structural survival Models

Code for implementing a simulation study to investigate the effect of near positivity violations on the estimates from IPTW-based marginal structural survival models in a survival context with longitudinal exposure and time-dependent confounding.

Reference

Spreafico M (2025). Impact of Near-Positivity Violations on IPTW-Estimated Marginal Structural Survival Models With Time-Dependent Confounding. Biometrical Journal, 67(6):e70093. doi: 10.1002/bimj.70093

Description

• Files:

  • AI_1 simulations.R: Simulation study using Algorithm I - Investigating various scenarios (see Section 4.2.1).
  • AI_2 results.R: Simulation study using Algorithm I - Results (see Section 4.2).
  • AI_3 figures_2_3.R: Results across all scenarios - reproducing Figures 2 and 3 (see Section 4.2.2).
  • AI_4 figures_4_5_6 table_1.R: Focused examination of weight truncation in selected scenarios - reproducing Figures 4 to 6 and Table 1 (see Section 4.2.3).
  • AII_1 simulations.R: Simulation study using Algorithm II - Investigating various scenarios (see Section 5.2.1).
  • AII_results.R: Simulation study using Algorithm II - Results (see Section 5.2).
  • AII_3 figures 7_8.R: Results across all scenarios - reproducing Figures 7 and 8 (see Section 5.2.2).
  • AII_4 figures_9_10_11 tables_2_3.R: Focused examination of weight truncation in selected scenarios - reproducing Figures 9 to 11 and Tables 2 and 3 (see Section 5.2.3).

• Sub-folder ./functions/ contains some auxiliary functions to run the main files:

  • algorithm_I.R: Code for Algorithm I (see Section 4.1.2).
  • algorithm_II.R: Code for Algorithm II (see Section 5.1.2).
  • eval_measuresI.R: Functions to evaluate the results for the scenarios simulated using Algorithm I.
  • eval_measuresII.R: Functions to evaluate the results for the scenarios simulated using Algorithm II.
  • iptw_I.R: Functions to perform IPTW in Simulation Study I.
  • iptw_II.R: Functions to perform IPTW in Simulation Study II.
  • mc_simI_functions.R: Functions to estimate the logit-MSMs using the longitudinal datasets simulated from Algorithm I.
  • mc_simII_functions.R: Functions to estimate the Aalen-MSMs using the longitudinal datasets simulated from Algorithm II.
  • plots_simI.R: Functions to plot results for Algorithm I.
  • plots_simII.R: Functions to plot results for Algorithm II.
  • specific_plots_simI.R: Functions to plot results for Algorithm I for the three specific scenarios examined in Section 4.2.3.
  • specific_plots_simII.R: Functions to plot results for Algorithm II for the three specific scenarios examined in Section 5.2.3.

• After running the main scripts, sub-folder ./results/ will contain the results from the various scenarios and relative performance.

Software

• R software, version 4.3.1
• Packages: data.table, doParallel, foreach, ggplot2, ggpubr, tidyr, timereg.

Additional info

To reproduce the Simulation Study I in Section 4.2 of the manuscript:

1. first run the script AI_1 simulations.R to simulate the $B=1000$ datasets under the different scenarios detailed in Section 4.2.1. Parallelisation details/remarks:
   • This code is parallelised over the different sample sizes considered, so 5 cores are used by default.
   • If your machine has fewer than 5 cores, you should adjust the value of n.cores when calling the function run_simulation_studyI() (line 73). In this case, n.sizes should be a vector whose length matches the number of cores.
   • Note that if the number of cores is less than the number of sample sizes, not all sample sizes will be processed in a single run. You will need to re-run the code multiple times to obtain results for all sample sizes.
   • The running time for n.sim = 5 is 6.33 minutes. Based on this, the expected running time for n.sim = 1000 is approximately 21–22 hours.
2. Then run the scripts AI_2 results.R to estimate the performance measures in Section 4.2.2 and AI_3 figures_2_3.R for reproducing Figures 2 and 3;
3. Finally, run the script AI_4 figures_4_5_6 table_1.R for reproducing Figures 4, 5, 6 and Table 1 in Section 4.2.3.

To reproduce the Simulation Study II in Section 5.2 of the manuscript:

1. first run the script AII_1 simulations.R to perform simulations under the different scenarios detailed in Section 5.2.1. Parallelisation details/remarks:
   • This code is parallelised over the different sample sizes considered, so 5 cores are used by default.
   • If your machine has fewer than 5 cores, you should adjust the value of n.cores when calling the function run_simulation_studyII() (line 88). In this case, n.sizes should be a vector whose length matches the number of cores.
   • Note that if the number of cores is less than the number of sample sizes, not all sample sizes will be processed in a single run. You will need to re-run the code multiple times to obtain results for all sample sizes.
   • The running time for n.sim = 5 is 41.36 seconds. Based on this, the expected running time for n.sim = 1000 is approximately 2.5 hours.
2. Then run the scripts AII_2 results.R to estimate the performance measures in Section 5.2.2 and AII_3 figures_4_5.R for reproducing Figures 4 and 5;
3. finally, run the script AII_4 figures_9_10_11 tables_2_3.R for reproducing Figures 9, 10, 11 and Tables 2 and 3 in Section 5.2.3.

(Last update: November 3rd, 2025)