Causal Tools Project

September 2021 Newsletter

Code Quality Assurance for Statistical Research

The paper “Best Practices in Statistical Computing” (Sanchez, Griffin, Pane, and McCaffrey, 2021) was recently published online as a featured article in Statistics in Medicine. In this paper, we describe the key steps for implementing a code quality assurance (QA) process that researchers can follow to improve their coding practices throughout a research project to assure the quality of the final data, code, analyses, and results. Following these steps can also make the code more accessible for sharing to support open science and replication.

The QA steps include: (i) adherence to principles for code writing and style that follow best practices; (ii) clear written documentation that describes code, workflow, and key analytic decisions; (iii) careful version control; (iv) good data management; and (v) regular testing and review. The steps are clearly demonstrated using a case study on treatment for substance use disorder taken from our team’s work. All the code and documentation used in the case study is available via Github.

SHINY App for Time-Varying Treatments

This is a user-friendly, menu-driven Shiny application (app) that can be used to estimate inverse probability of treatment weights (IPTW) for time-varying treatments. It also can be used to estimate three different types of treatment effects using those weights. The app includes a tutorial, which demonstrates use of the TWANG Shiny app through an illustrative example and toy dataset. It provides a nice complement to our binary SHINY App for Two Treatments and our Shiny App for Three or More Treatments.

Covariate Web App (CoBApp)

This user-friendly, menu-driven app, created by Andreas Markoulidakis of Cardiff University in partnership with RAND, estimates balancing weights for nine different propensity score and entropy balancing methods, selects the optimal set of weights, performs treatment effect estimation, and helps users carefully assess the validity of both the overlap and unobserved confounding assumptions required for valid inference.

New Packages on CRAN

Our team has released several new R libraries on CRAN that can be used to aid causal inference.

• Omitted variable sensitivity analysis tool (OVtool)
  Assesses the sensitivity of treatment effects and statistical significance to an unobserved confounder.
• Entropy balancing for binary and continuous trts (entbal)
  Estimates causal treatment effects for binary and continuous treatments using entropy balancing (EB).
• Continuous exposures/treatment tool (twangContinuous)
  Estimates generalized propensity score weights for continuous exposures/treatments using generalized boosted models (GBM).
• Causal mediation tool (wgtmed)
  Conducts causal mediation to estimate natural direct and indirect effects for one or more mediators using a weighting approach with weight estimated by GBM or logistic regression.
• Selection bias decomposition (SBdecomp)
  Quantifies the proportion of the estimated selection bias explained by each observed confounder when estimating causal effects using propensity score weights.
• Twang 2.0
  Our updated TWANG package expands the original package to include xgboost for estimation of propensity score weights with the goal of improving the performance of the TWANG package for big data. The twang 2.0 R library on CRAN contains:
  • ps command: to estimate propensity score weights for two groups
  • mnps command: to estimate propensity score weights for more than two groups
  • iptw command: to estimate inverse probability of treatment weights for time-varying treatments.

SAS Tools

• TWANG Macro Package: This package contains commands for estimating propensity score weights for two groups (using the %ps command) and for more than two groups (using the %mnps command). It also contains a command (using the %dxwts command) for assessing the covariate balance achieved by any set of weights.

Stata Tools

• Support Files for PS and MNPS: This set of files supports estimating propensity score weights for two groups (using the ps command) and more than two groups (using the ps and mnps command) and for checking the balance of groups after weighting.