Sensitivity Analysis for Covariates Observed in Two-Phase Samples

Provides a structured method-comparison sensitivity-analysis workflow for linear regression when important continuous adjustment covariates are observed only in a Phase-2 subsample. Implements naive analysis, complete-case analysis, fully conditional specification multiple imputation, joint-model multiple imputation, inverse probability weighting, and augmented inverse probability weighting, with standardized tables, diagnostics, and publication-style forest plots. Methodological background is provided by Little and Rubin (2019) for general missing-data methods, White and Carlin (2010) for complete-case analysis with missing covariates, White, Royston and Wood (2011) and van Buuren and Groothuis-Oudshoorn (2011) for chained-equations multiple imputation, Quartagno, Grund and Carpenter (2019) for joint-model multiple imputation, Robins, Rotnitzky and Zhao (1994) for inverse-probability weighting with missing regressors, and Bang and Robins (2005) for doubly robust estimation.

twoPhaseSens

twoPhaseSens is an R package for method-comparison sensitivity analyses when important continuous adjustment covariates are observed only in a Phase-2 subsample.

Version 0.1.0 compares six approaches for a linear-regression exposure coefficient:

1. Naive Phase-1 analysis
2. Complete-case analysis (CCA)
3. Fully conditional specification multiple imputation (FCS-MI)
4. Joint-model multiple imputation (JM-MI)
5. Inverse probability weighting (IPW)
6. Augmented inverse probability weighting (AIPW)

Release status

Version 0.1.0 is the first public release of twoPhaseSens. The six analytic methods have been generalized into the package interface and numerically checked against the validated manuscript implementation for both MIDUS outcomes. Automated tests, cross-platform R CMD checks, source-package checks, Win-builder checks, and a clean source-install smoke test are passing.

Installation

Install version 0.1.0 from GitHub with:

# install.packages("remotes")
remotes::install_github("rophenceojiambo/twoPhaseSens")

By default, remotes may ask whether to update older installed dependencies. Users who specifically want to avoid upgrading already-installed packages can instead use upgrade = "never".

After the package is available on CRAN, it can also be installed with install.packages("twoPhaseSens").

Basic workflow

library(twoPhaseSens)

dat <- twophase_example_data(
  n = 300,
  seed = 2026
)

fit <- twophase_sensitivity(
  data = dat,
  outcome = c("outcome1", "outcome2"),
  exposure = "exposure",
  covariates = c("age", "sex"),
  phase2_covariates = c("marker1", "marker2", "marker3"),
  phase2 = "phase2",
  methods = c("naive", "cca", "ipw"),
  seed = 20260928
)

tbl_sensitivity(fit)
diagnostics(fit)
plot_sensitivity(fit)

The example above uses three fast methods for a quick first run. Omit the methods argument to run all six methods in version 0.1.0.

Plot display and export

plot_sensitivity() returns a publication-style ggplot object. Annotated plots can look compressed in a small RStudio Plots pane, especially with two outcomes. Use Zoom, enlarge the pane, or set annotate = FALSE for a compact interactive preview.

For exported output, specify the graphics size explicitly. As practical starting points, use approximately 9.5 x 6 inches for one annotated outcome and 14 x 6 inches for two annotated outcomes.

p <- plot_sensitivity(
  fit,
  estimate_label = "Adjusted beta"
)

ggplot2::ggsave(
  "sensitivity_plot.png",
  p,
  width = 9.5,
  height = 6,
  dpi = 320,
  bg = "white"
)

See the Getting Started vignette for single- and multiple-outcome examples, fixed versus free x-axis scales, custom labels, and export guidance.