Skip to contents

Bias-corrected bounds are experimental (pre-beta research), incomplete, and not recommended for general-purpose inference. See the experimental-status section below. Selects explicitly the experimental local-constant conditional inference constructions. Existing plots and methods are unchanged when this control is omitted.

Usage

np_conditional_bias_control(construction = c("analytic", "jackknife", "jackknife-mp1"))

Arguments

construction

The corrected functional. Use "analytic" for conditional CDF inference and conditional density bootstrap inference; use "jackknife" for conditional density asymptotic inference. Use "jackknife-mp1" for the separate asymmetric studentized conditional density bootstrap described below.

Value

A validated control object for the bias_control argument of plot applied to npcdist or npcdens fits or bandwidths.

Details

These are experimental opt-ins for local constant, one continuous explanatory variable and one continuous response, fixed unbounded second-order Gaussian kernels. Gradients, quantiles, shape projections, nonfixed bandwidths and mixed variables are not admitted. Original fits and selected bandwidths remain fixed.

Specify center="bias-corrected" and an explicit band: "pointwise", "simultaneous", "bonferroni" or "all". With errors="bootstrap", explicitly specify bootstrap="conditional-pilot" and np_boot_control(calibration="symmetric-sd") or np_boot_control(calibration="asymmetric-sd"). This law keeps explanatory observations fixed and draws each response from its conditional Gaussian donor mixture, with pilot bandwidths multiplied by \(\sqrt{2}\). It is not ordinary pairs resampling and does not replace inid, fixed, geom, residual or wild resampling.

The analytic correction contains both explanatory-weight and response-kernel curvature. Each bootstrap root includes the complete corrected functional, centered by its conditional-pilot expectation. A common marginal bootstrap SD, maximized within each explanatory slice, scales the simultaneous signed or absolute extremes. The asymmetric method inverts the signed extremes; it does not use two unrelated lower/upper SD estimates.

With errors="asymptotic", complete conditional-pilot covariances feed Gaussian finite-grid calibration. For CDFs the analytic correction and slice scales are used. For densities the corrected estimate is twice the estimate at the fitted bandwidths minus the estimate at bandwidths enlarged by \(\sqrt{2}\) in both coordinates. All cross-covariance terms are retained. The density asymptotic and bootstrap corrections therefore differ deliberately.

The separate density construction "jackknife-mp1" requires errors="bootstrap", bootstrap="conditional-pilot" and calibration="asymmetric-sd" in np_boot_control(). It uses the same jackknife corrected functional as density asymptotic inference. Each complete corrected bootstrap root is divided by its own MP1 standard error. MP1 averages the original and resample pilot covariance matrices and adds one quarter of the sum of per-observation mean-difference outer products. This is the covariance of independent equal pilot mixtures for each training observation, retaining every correction cross term. Its variance is at least half the original pilot variance; this is an algebraic property, not a variance floor. Nonpositive or nonfinite component variances cause an error. Signed studentized roots are inverted using the original standard errors. The selector is currently restricted to asymmetric calibration; it does not change either existing analytic SD bootstrap construction.

A nominal marginal sharp-score safeguard uses the certified range of each signed contribution and its heterogeneous variance bound. It augments the Gaussian or bootstrap intervals. The density asymptotic and jackknife-MP1 constructions use a marginal hull; the other constructions also propagate required widths within explanatory slices. Companion hulls protect interval nesting. These are approximate confidence constructions, not finite-sample honesty guarantees. No clipping, positivity projection or forced properness is applied to corrected centers or endpoints.

The family consists of the actual requested finite evaluation points within one panel or surface. Separate panels are not a joint family; a finite grid is not the entire continuous domain. Initial admission is at most 256 target points and 25 distinct response targets. Choose, for example, neval=9 for an 81-point surface; the ordinary default 50-by-50 surface exceeds this new method's scope. Inputs exceeding the numerical workspace budget of 256 MiB fail before resampling; the grid and draw count are never silently reduced. The budget is not a bound on total R-process memory. Bootstrap B and asymptotic gaussian_draws must be integers at least two; defaults are 9999. Low finite-draw tail resolution is reported. Alpha is propagated throughout in \((0,0.5)\); population coverage evidence is currently for nominal 95% only.

Data-returning plots use np_plot_data. The value tables contain original and corrected estimates and available SEs; interval tables contain absolute endpoints for every requested band. Compact inference records identify the construction, law, calibration and available studentization diagnostics. Corrected centers may be signed and need not be interval midpoints. Internal pilot samples, covariance workspaces and RNG states are not returned.

Saved Gaussian-design studies support full-grid CDF coverage at n=100 and 400, and density coverage at n=400 and 800. CDF plot subfamilies and marginal intervals can be conservative above 99%. Density bootstrap small-n studies failed lower coverage gates. These failures remain limitations; they do not establish uniform coverage, robustness to arbitrary DGPs, or adjustment for bandwidth selection uncertainty. Fresh paired Gaussian studies of asymmetric MP1 at n=400 and 800 retained 481/512 and 488/512 simultaneous coverages, respectively, on the studied 81-point random grids. These results qualify this experimental opt-in, not a general coverage or optimality guarantee. The options are intended for further experimentation.

Experimental status of bias-corrected inference

All bias-corrected confidence intervals and bands are experimental (pre-beta research), incomplete, and not recommended for general-purpose inference. This status remains in effect unless the documentation explicitly states otherwise for a particular method and scope. It applies wherever bias-corrected bounds are available: asymptotic and bootstrap constructions, including legacy methods, pointwise intervals, Bonferroni intervals and simultaneous bands, for levels and supported derivatives.

These methods form an ongoing research project. Implementation support, numerical checks and favorable simulation results in selected settings do not establish generally valid coverage. Coverage can fail outside the studied settings, and available options do not constitute a general solution to bias-corrected inference. Method-specific assumptions, restrictions and known failures must be considered when experimenting with these bounds.

Examples

if (FALSE) { # \dontrun{
set.seed(42)
x <- rnorm(100)
y <- 0.5*x + rnorm(100)
model <- npcdist(y ~ x, regtype="lc")
p <- plot(model, errors="asymptotic", center="bias-corrected",
          bias_control=np_conditional_bias_control("analytic"),
          band="all", neval=9, output="plot-data")
p <- plot(model, errors="bootstrap", bootstrap="conditional-pilot",
          center="bias-corrected", band="all", neval=9,
          bias_control=np_conditional_bias_control("analytic"),
          boot_control=np_boot_control(calibration="asymmetric-sd"),
          output="plot-data")

model.d <- npcdens(y ~ x, regtype="lc")
p.d <- plot(model.d, errors="bootstrap", bootstrap="conditional-pilot",
            center="bias-corrected", band="all", neval=9,
            bias_control=np_conditional_bias_control("jackknife-mp1"),
            boot_control=np_boot_control(calibration="asymmetric-sd"),
            output="plot-data")
} # }