/registry_simulations

targets-based pipeline for longitudinal simulation studies from EHR data similar to the Danish registry

Primary LanguageR

Registry data simulations for the analysis of GLP1 use and dementia in a cohort of diabetes patients

Reproducible analysis pipeline for a longitudinal simulation study of the effect of 2nd line diabetes drugs from EHR data similar to the Danish registry. Simulation results found in the arXiv paper here.

Overview

This repository contains a pipeline for longitudinal simulation studies from EHR data similar to the Danish registry. The pipeline is adapted from the applied analysis functions used to analyze the Danish registry, which use the targets package framework and reproducible pipeline. The simulations study, however, does not use a targets pipeline itself due to computation time and for increased flexibility in running parts of the simulation. It uses a series of scripts that can be run in order to reproduce the simulation study using the same functions used in the applied analysis. The pipeline is designed to be reproducible and to be run or adapted by any reader of the manuscript who has access to a high-performance computing cluster, as the whole simulation is computationally expensive and prohibitively slow without parallelization.

Table of Contents

  1. Installation
  2. Functions
  3. Usage
  4. License

Installation

  1. Install required packages by running:
   install.packages(c("tidyverse",  "fst","lava","ltmle","data.table","tidyverse","glmnet","Matrix","Publish","matrixStats","speedglm","parallel","caret","foreach","clustermq"))

Functions

The simulation study primarily runs off functions sourced from the Ltmle folder, which are the ltmle package functions for longitudinal maximum likelihood estimation, adapted to run without error on Statistics Denmark, the remote server location of the Danish Registry and to allow different estimator specifications outside of GLM and SuperLearner. Other functions in the /functions/ folder are wrapper functions to run the simulation and applied analysis without error on Statistics Denmark as well as the remote server used to run the simulation.

Usage

The scripts to reproduce the simulation are in the /src/ folder labeled in order, and the script 0_run_simulations.R runs each script in order. The file contents, in brief, are as follows:

  • 1_simulate_data.R: using the lava package and exported coefficients using the Danish registry, we simulate the datasets to be used for the simulation and save this in the /data/ subfolder.
  • 2_calculate_truth.R: use function calc_realistic_truth to intervene on the DGP to set desired causal contrasts and simulate from them with large N, approximating the truth.
  • 3_run_simulation_for_point_estimators.R: With wrapper functions for parallelization, we run the simulation for the different estimator options across 1000 iterations of simulated data.
  • 4_calc_performance.R: Calculates the performance of the estimators across multiple metrics, including bias and empirical and oracle variance and coverage.
  • 5_run_bootstrap_variance.R: Estimate the bootstrapped confidence intervals the winning point estimator
  • 6_calc_bootstrap_performance.R: Compare the coverage of bootstrapped variance compared to the influence curve and TMLE based variance.
  • 7_simulate_data_null.R: using the same process as 1_simulate_data.R but with the null hypothesis of no effect of the treatment on dementia or death.
  • 8_run_null_simulation_for_point_estimators.R: With wrapper functions for parallelization, we run the simulation for the different estimator options across 1000 iterations of simulated null data.
  • 9_calc_performance_null.R: Calculates the performance of the null simulation.
  • 10_create_latex_tables.R: creates the manuscript results tables
  • 11_example_DAG.R: creates the example DAG used in the publication.

Notes

  • The deterministic Q function (which encodes the setting of the probability of the outcome dementia deterministically to 1 once death, the competing event, has occurred) is built into the updates to the ltmle package, found in the repo. The exact code used for this can be found here: Ltmle/Augmentation/event_node_manipulator.R

  • When using undersmoothed penalized regressions, we selected the lambda value by choosing the minimum penalization (i.e., the largest lambda value) across a range of candidate values that did not cause the algorithm to crash. The exact code used for this can be found on lines 80-96 here: Ltmle/Augmentation/ltmle.glmnet.R

License

This project is licensed under the MIT License - see the LICENSE.md file for details.

For any questions or issues, please open an issue on the GitHub repository or contact the maintainer at amertens@berkeley.edu.