confoundvis
Visualization Tools for Sensitivity Analysis of Unmeasured Confounding
Visualization and reporting tools for sensitivity analysis to
unmeasured confounding in observational studies. A common
'confoundsens' object stores a sensitivity path (the treatment effect
as a function of hypothetical confounder strength) regardless of the
framework that produced it, so the same robustness curves, contour
plots, covariate benchmark ("sensitivity Love") plots, and plain-language
reports can be drawn for impact threshold analysis (Frank, 2000,
confoundvis
confoundvis draws and reports sensitivity analyses for unmeasured
confounding. A single confoundsens object stores a sensitivity path,
the treatment estimate as a function of the strength of a hypothetical
omitted confounder, whichever framework produced it. The same robustness
curves, covariate benchmark plots, contour plots, and plain-language
reports then work for:
| Framework | Reference | Strength index | Path source |
|---|---|---|---|
| Impact threshold (ITCV) | Frank (2000); Frank et al. (2013) | impact r(D,U) x r(Y,U) |
itcv_lm(), from_konfound() |
| Partial R-squared / robustness value | Cinelli & Hazlett (2020) | partial R-squared of the confounder | sens_path_lm(), from_sensemakr() |
| E-value | VanderWeele & Ding (2017) | confounder risk ratio | from_evalue() |
Installation
install.packages("confoundvis")
# development version
# pak::pak("subirhait/confoundvis")
Workflow
library(confoundvis)
fit <- lm(mpg ~ am + wt + hp + qsec, data = mtcars)
# 1. sensitivity paths computed from the fitted model
path <- sens_path_lm(fit, treatment = "am") # partial R-squared
it <- itcv_lm(fit, treatment = "am") # ITCV
# 2. plots
plot_robustness_curve(path)
plot_robustness_curve(it$path)
# 3. benchmark against observed covariates
imp <- covariate_impacts(fit, "am")
plot_sensitivity_love(imp)
plot_sensitivity_contour(attr(imp, "threshold"), benchmarks = imp)
# 4. report
sens_report(path)
Results already produced by sensemakr, konfound, or EValue
can be converted with from_sensemakr(), from_konfound(), and
from_evalue(); as_confoundsens() also accepts a data frame of
precomputed paths, including multilevel (within/between) paths.
See vignette("confoundvis-workflow") for a complete example with the
public darfur data.
Scope
confoundvis is a presentation layer. Its computations reproduce each
framework’s published formulas (tests compare them with sensemakr,
konfound, and EValue), and it inherits each framework’s assumptions. A
sensitivity display shows how strong confounding would have to be; it
cannot show whether such a confounder exists, and it cannot repair a
flawed identification strategy. plot_reversal_cone() and
plot_taylor_panels() are conceptual illustrations built on stylized
models.
References
Cinelli, C., & Hazlett, C. (2020). Making sense of sensitivity: Extending omitted variable bias. JRSS-B, 82(1), 39–67.
Frank, K. A. (2000). Impact of a confounding variable on a regression coefficient. Sociological Methods & Research, 29(2), 147–194.
Frank, K. A., Maroulis, S. J., Duong, M. Q., & Kelcey, B. M. (2013). What would it take to change an inference? Educational Evaluation and Policy Analysis, 35(4), 437–460.
VanderWeele, T. J., & Ding, P. (2017). Sensitivity analysis in observational research: Introducing the E-value. Annals of Internal Medicine, 167(4), 268–274.
Citation
citation("confoundvis")