Conformal Local Influence Screening for Bounded-Response Regression

Provides fast, statistically calibrated influence diagnostics for zero-or-one inflated beta (BIc) regression models with variable dispersion. The core idea is to use the conformal normal curvature of Poon and Poon (1999) as a non-conformity score within a split-conformal testing procedure, yielding per-observation conformal p-values whose Benjamini-Hochberg adjustment controls the false discovery rate at a user-specified level (Bates and others, 2023) . Unlike classical local influence diagnostics, which rely on visual inspection of index plots and do not scale beyond a few hundred observations, 'clis' provides a finite-sample error guarantee and runs in linear time per observation after a single model fit. Methods for four perturbation schemes, block decomposition of influence into the inflation-probability and conditional-mean/precision components, penalised additive (semiparametric) submodels, and a full suite of diagnostic plots are included.

clis

clis provides scalable, statistically calibrated influence diagnostics for zero-or-one inflated beta (BIc) regression models with variable dispersion. It turns the conformal normal curvature of an observation into a non-conformity score, then wraps it in a split-conformal testing procedure so that the declared influential set has its false discovery rate controlled at a level you choose.

Why clis?

Classical local influence analysis (Cook, 1986; Poon & Poon, 1999) ranks observations by curvature and asks you to eyeball an index plot. That does not scale to large data and gives no error guarantee. clis fixes both:

• Finite-sample FDR control via conformal $p$-values and Benjamini-Hochberg (Bates et al., 2023).
• Linear time per observation after a single model fit, so it scales to large $n$.
• Block decomposition that attributes each observation's influence to the inflation-probability submodel or the conditional-mean/precision submodel.

Installation

install.packages("clis_0.3.6.tar.gz", repos = NULL, type = "source")

Quick start

library(clis)
library(gamlss)

vaccination <- load_vaccination()

fit <- gamlss(
  dtp3 ~ ln_gdp + urb,
  sigma.formula = ~ ln_gdp + ln_pop,
  nu.formula    = ~ hdi,
  family  = gamlss.dist::BEOI,
  data    = vaccination,
  control = gamlss.control(trace = FALSE)
)

res <- clis_screen(fit, alpha = 0.1, seed = 1)
res
plot_clis(res)

Learn more

See vignette("clis-intro") for a full walkthrough. The scripts under data-raw/ reproduce every table and figure of the accompanying paper; data-raw/README.md is the runbook.

References

• Bates, S., Candès, E., Lei, L., Romano, Y., & Sesia, M. (2023). Testing for outliers with conformal p-values. The Annals of Statistics, 51(1), 149–178.
• Ospina, R., & Ferrari, S. L. P. (2012). A general class of zero-or-one inflated beta regression models. Computational Statistics & Data Analysis, 56(6), 1609–1623.
• Poon, W.-Y., & Poon, Y. S. (1999). Conformal normal curvature and assessment of local influence. JRSS-B, 61(1), 51–61.