ImputationPackage

The goal of ImputationPackage is to simulate missingness in PatientLevelPrediction data, apply imputation techniques and evaluate downstream model performance.

Overview

This package is designed for studies that want to evaluate how different missing-data structures and imputation methods affect predicitive modeling results.

Features

• Build PLP-ready data objects
• Simulate missingness under different mechanisms and allow different missingness structures in one data object
• Run multiple imputation methods
• Fit prediction models
• Aggregate and summarise evaluation outputs

Prerequisites

• R >= 4.2.0
• Required packages:
  • Andromeda
  • CirceR
  • DatabaseConnector
  • dplyr
  • FeatureExtraction
  • margrittr
  • PatientLevelPrediction
  • readr
  • rlang
  • sqlRender
  • tibble

Installation

You can install the development version of ImputationPackage like so:

# FILL THIS IN! HOW CAN PEOPLE INSTALL YOUR DEV PACKAGE?

Workflow

1. Build a (complete-case) PLP data object
2. Simulate missingness
3. Apply imputation methods
4. Fit prediction models
5. Evaluate and summarise prediction results

Important Functions

• buildPopulationPLPData(): to build a (complete-case) PLP data object.
• runMissingnessSimulation(): to run a full simulation study.
• createPLPImputationMethods(): to generate the PLP imputation methods from the PatientLevelPrediction package.
• runPlpImputers(): to run imputation methods on train- and test data.
• runPlpModels(): to run prediction models on train- and test data and store the evaluation metrics.
• collectEvaluationTables(): to collect model evaluation outputs into tables for comparison and evaluation summaries.

Layout

• R/ : main package functions
• extras/: data exploration scripts and example scripts
• tests/: test folder

Example

An example script to run a simulation is in extras/run_workflow.R