Changelog • ineqx

ineqx 0.5.0

The causal decomposition now targets marginal counterfactual inequality by default, rather than residual (conditional) inequality. Within-group variance is computed by the law of total variance over the covariate distribution of the treated within each group: the average conditional residual variance plus the dispersion of predicted conditional means across covariate profiles. Control variables that adjust the counterfactual predictions therefore contribute to within-group inequality. When the model contains no within-group control variation, the marginal and residual results coincide, so analyses without controls are unaffected.

New estimand argument on ineqx() and ineqx_params(): "marginal" (default) or "residual". The "residual" variant reproduces the previous behavior (conditional scale parameter only, omitting the predicted-mean dispersion term). The chosen estimand is recorded on the params/result objects and shown by print().
Marginal delta-method standard errors use a numerical Jacobian of the full prediction + g-computation routine (the marginal group variance is nonlinear in both the mean- and scale-equation coefficients); the residual variant keeps the fast analytical Jacobian. The bootstrap path honors estimand throughout. The marginal Jacobian is evaluated with incremental finite differences (each coefficient perturbation is a rank-1 update of the cached predictions), making it O(N * p) rather than O(N * p^2) – nearly as fast as the analytical path even for models with large coefficient counts.

Merged ineq() and ineqx() into a single ineqx() function
- If no treat is specified: descriptive decomposition (was ineq())
- If treat is specified or params provided: causal decomposition
ineq() removed — use ineqx() without treat instead
y is now a separate argument (character, column name in data); formula_mu is now one-sided (like formula_sigma)
boot_config() now requires y argument and one-sided formula_mu
Default order changed from c("behavioral", "compositional", "pretreatment") to "shapley"
Argument order: y, ystat, treat, post, group, time, ref, order, formula_mu, formula_sigma, params, weights, se, data
Renamed modes: “Mode A” → “integrated estimation”, “Mode B” → “externally estimated model”

Fully parallel argument structure: descriptive and causal paths share y, group, time, ref, ystat, order, data
Public API reduced to 2 main functions: ineqx() and ineqx_params()

Streamlined to 3 user-facing functions: ineq(), ineqx(), ineqx_params()
desc_decompose() replaced by ineq() with enhanced counterfactual decomposition
extract_params() merged into ineqx_params(model = ...)
causal_decompose_cross(), causal_decompose_longit(), causal_shapley(), fit_ineqx_model() are now internal
ineqx() se_method/boot arguments replaced by unified se argument: "delta" (default), "none", or boot_config() object
ineqx() no longer returns ineqx_result wrapper; returns decomposition object directly

ineq(): descriptive W/B decomposition with counterfactual parameter-switching decomposition (mu, sigma, pi) and Shapley averaging
ineqx() integrated estimation: GAMLSS fitting via formula_mu/formula_sigma/data arguments
order = "shapley" supported for ordering-robust decomposition
ineqx_params() accepts either manual data.frame or fitted gamlss model (via model arg)

Complete package redesign with new modular architecture
Old functions removed: wibe(), calcAME(), plot.ineqx()
Replaced by: desc_decompose(), extract_params(), new S3 plot methods
The c./i. variable prefix convention has been removed; use plain column names
zeallot dependency removed; gamlss moved from Depends to Suggests

ineqx_params(): Standardized input format that decouples estimation from decomposition. Users can construct parameters from any estimation method.
causal_decompose_cross(): Cross-sectional causal decomposition (treatment effect on variance = within + between)
causal_decompose_longit(): Longitudinal causal decomposition with configurable decomposition ordering (6 possible orderings via the order parameter)
causal_shapley(): Shapley values averaged across all 6 orderings as robustness check
desc_decompose(): Clean descriptive within/between decomposition
extract_params(): Extract decomposition parameters from fitted GAMLSS models
fit_ineqx_model(): Convenience wrapper for fitting GAMLSS models
Both Var (variance) and CV2 (squared coefficient of variation) supported throughout
Print, summary, and plot methods for all result classes