Literature review on existing methods and packages

[Kuan-Liu]

@larryshamalama To summarize the pros and cons for each approach (point-treatment and binary or continuous outcome) and document published packages on their scope and flexibility. In particular, we focus on the following methods. We will discuss the lit review summary in the week of Nov 6th.

• Approach category 1 (frequentist): Confounding functions: adjust causal estimates using confounding functions that describe the degree of unmeasured confounding.

1. Propensity Score-based Sensitivity Analysis Method for Uncontrolled Confounding, https://doi.org/10.1093%2Faje%2Fkwr096 (no R package, to code)
2. Assessing the impact of unmeasured confounding for binary outcomes using confounding functions, https://doi.org/10.1093/ije/dyx023 (no R package, to code)

• Approach category 2 (more advanced): Latent confounding variable approach using bias parameters

1. full parametric probabilistic approaches: Bayesian and Monte Carlo Sensitivity analysis, A comparison of Bayesian and Monte Carlo sensitivity analysis for unmeasured confounding, https://doi.org/10.1002/sim.7298 (no R package, the paper estimate conditional treatment effect, coding is needed to estimate marginal treatment effect and ATE)
2. Semi-parametric probabilistic approach using BART, A flexible, interpretable framework for assessing sensitivity to unmeasured confounding, https://doi.org/10.1002/sim.6973 (treatSens package available on GitHub, https://github.com/vdorie/treatSens, explore it's scope for example does it work for binary or continuous outcomes)
3. Semi-parametric approach, A Semi-Parametric Approach to Model-Based Sensitivity Analysis in Observational Studies, https://doi.org/10.1111/rssa.12946 (R code available, https://github.com/bzhangupenn/Code_for_reproducing_semi_SA but it's not a package, study scope and we will embed this in our package)

• When reading these methods, take notes on the analysis and visual output of each data analysis application. This will help guide us in coding our package model output!

[Kuan-Liu]

@larryshamalama Hi Larry, I created four functions to generate testing data for the package. For some reason I can't uploaded to the git repo. I am getting errors "[Kuan-Liu-Lab/causens] Run failed: lint-project - main (f957a1b)"...

The code is here in txt format you can convert to R script,simData.txt.

• U is the unmeasured confounder, U can be binary or continuous
• Y is the outcome, it can be binary or continuous.
• The 2*2 combo is coded in the txt file.

After reading the reference, for those that have existing packages, can you implement these methods using test data (generated from the simulation code). It's best we make sure all methods run before your code into functions. What is the timeline for this for you? We should aim to complete implementation of all existing methods by the end of this month.

[larryshamalama]

@larryshamalama Hi Larry, I created four functions to generate testing data for the package. For some reason I can't uploaded to the git repo. I am getting errors "[Kuan-Liu-Lab/causens] Run failed: lint-project - main (f957a1b)"...

The code is here in txt format you can convert to R script,simData.txt.

This is because of the linter that I put. I need to add a "styler", as the R ecosystem calls it. Not to worry about, thanks for the code.

• U is the unmeasured confounder, U can be binary or continuous
• Y is the outcome, it can be binary or continuous.
• The 2*2 combo is coded in the txt file.

After reading the reference, for those that have existing packages, can you implement these methods using test data (generated from the simulation code). It's best we make sure all methods run before your code into functions. What is the timeline for this for you? We should aim to complete implementation of all existing methods by the end of this month.

This timeline should work with me.