confoundvis 0.3.0

Scope

• confoundvis is now a common interface to established sensitivity analyses only. The illustration functions fit_local_quadratic(), plot_local_taylor(), plot_taylor_panels(), plot_figure2_taylor_panels(), simulate_taylor_demo(), plot_reversal_cone(), and plot_cone_comparison() are removed. They did not analyse data and did not correspond to any established framework.
• The introductory vignette built on stylized paths is removed; the workflow vignette covers every feature with real data.
• ‘grid’ is no longer imported and ‘gridExtra’ is no longer suggested.

New features

• sens_path_pairs() computes a sensitivity path for matched-pair studies under Rosenbaum’s sensitivity model. For each value of Gamma it stores the deviate equivalent to the upper bound on the one-sided P-value, so robustness_points(), plot_robustness_curve(), sens_report(), and sens_table() apply unchanged. Two statistics are available: Wilcoxon’s signed rank statistic (method = "wilcoxon", agreeing with DOS2::senWilcox()) and the exact bound for McNemar’s statistic for binary outcomes (method = "binary", agreeing with sensitivity2x2xk::mh()). The sensitivity value is also found by root finding and stored in meta$sens_value.
• rir_lm() computes the robustness of inference to replacement (RIR; Frank et al., 2013) from a fitted linear model: the threshold for significance, the percent of the estimate that would have to be due to bias, the number of cases that would have to be replaced with null-effect cases, and the corresponding path (estimate after replacing a proportion lambda of the cases). Results agree with konfound::konfound(index = "RIR") and reproduce the worked example of Frank et al.’s (2025) practical guide. A threshold argument allows a substantive effect threshold in place of statistical significance. from_konfound() gains index = "RIR" to import konfound’s RIR output as a path.
• ITCV benchmarks following Frank et al.’s (2025) practical guide: itcv_lm() returns benchmark_corr_product (R_XZ * R_YZ, the impact of all covariates together) and itcv_ratio_to_benchmark (eq. 12), and covariate_impacts() adds each term’s zero-order correlations, their product impact_zero, and itcv_ratio, the unconditional ITCV divided by that impact (eq. 10).
• sens_surface() returns the two-parameter sensitivity surface of which a path is one slice, and plot_sensitivity_surface() draws it with shaded zones, the null and significance boundaries, and optional benchmarks (sensemakr formal bounds are added automatically). Surfaces are available for partial R2 (Cinelli and Hazlett, 2020), ITCV (exact boundaries), E-values (joint bounding factor of Ding and VanderWeele, 2016), and Rosenbaum bounds (amplification of Rosenbaum and Silber, 2009, agreeing with sensitivitymv::amplify()). Boundaries are computed in closed form or by one-dimensional root finding.
• sens_table() collects several confoundsens objects into one table: framework, meaning of the sensitivity parameter, the framework’s own summary statistic, null and significance crossings, and the rule used. Rows keep their own scales; the printed table says so.
• plot_sensitivity_contour() gains an itcv argument. Given the output of itcv_lm(), it draws, in closed form, the exact boundaries at which the magnitude of the confounder-adjusted partial correlation equals the critical value: in the original direction (solid) and in the reversed direction (dotted). It shades the region where the estimate is no longer significant, the region where it becomes significant with the opposite sign, and correlation pairs that cannot occur (non-positive-definite correlation matrix). The product curve is kept as a dashed reference. threshold may now be omitted when itcv is supplied, and a new limits argument sets the axis range.
• Benchmark labels in plot_sensitivity_contour() repel one another when ggrepel is installed (suggested), instead of overlapping.

Bug fixes

• itcv_lm(): for weighted fits, the R-squared values used for the unconditional ITCV were computed around an unweighted mean. They now equal the weighted R-squared reported by summary.lm(). Unweighted fits are unaffected.
• sens_report() no longer points to covariate_impacts() for frameworks that it does not benchmark (E-values, Rosenbaum bounds).
• robustness_points() now takes the direction of the original conclusion from the stored threshold when there is one. Paths that start on the wrong side of their threshold are therefore reported as not significant at the first grid value instead of returning NA. Results for paths produced by earlier versions are unchanged.

Other changes

• DOS2 and sensitivity2x2xk are suggested for the cross-checks in the test suite.
• The workflow vignette adds the exact contour, a matched-pair example, and sens_table().

confoundvis 0.2.0

Bug fixes

• Added the missing tests/testthat.R runner. In 0.1.0 the test files were shipped but never executed by R CMD check.
• grid is now declared in Imports (it was used via grid::).
• validate_confoundsens() now checks that lambda is nondecreasing within each level, so stacked subgroup paths no longer trigger a spurious warning.

New features: from a fitted model to a report

• sens_path_lm() computes the partial R-squared sensitivity path (Cinelli and Hazlett, 2020) directly from an lm fit, including weighted fits.
• itcv_lm() computes the conditional and unconditional impact threshold (Frank, 2000; Frank et al., 2013) and the ITCV path.
• covariate_impacts() benchmarks every observed covariate term (including multi-column factor terms for the partial R-squared metric) and returns a data frame that plot_sensitivity_love() and plot_sensitivity_contour() accept directly.
• robustness_points() locates where a path reaches the null and where significance is lost; sens_report() turns these into a short plain-language report that states what the analysis cannot establish.

Interoperability

• as_confoundsens() is now generic, with methods for data frames (as before), lm fits, and sensemakr objects.
• from_sensemakr(), from_konfound(), and from_evalue() convert results from the ‘sensemakr’, ‘konfound’, and ‘EValue’ packages. These packages are only suggested; the adapters read documented fields and do not call them.
• The test suite checks the computations against those packages when they are installed (adjusted estimates and standard errors, robustness values, conditional and unconditional ITCV, and E-values).

Other changes

• confoundsens objects gain an optional meta list (framework, labels, null value, critical value). Objects created by 0.1.0 remain valid.
• plot_robustness_curve() uses axis labels from meta and gains reference to draw the null and critical values.
• plot_sensitivity_love() takes its threshold from the "threshold" attribute of df when not supplied, and drops rows with missing impact.
• plot_figure2_taylor_panels() is renamed plot_taylor_panels(); the old name is deprecated.
• plot_reversal_cone() documentation now states that the display uses a stylized bilinear bias model and that the annotated share depends on the plotted ranges.
• New vignette, “From a fitted model to a sensitivity report”, using the public darfur data and including a limitations section.

confoundvis 0.1.0

• First release on CRAN.