qrcmNP 1.0.0

Breaking changes

• piqr() now solves by default the lasso problem of Sottile et al. (2020, eq. 3.4) (method = "pnewton"): results differ from version 0.2.x. The previous algorithm is available with method = "legacy".

• library(qrcmNP) no longer attaches ‘qrcm’ and ‘survival’ (moved to Imports); slp(), plf() and test.fit() are re-exported. Use library(qrcm), library(survival) or qrcm::iqr(), survival::Surv() where needed.

• gof.piqr(): the default df counts the nonzero coefficients ("parameters"); use df = "covariates" for the previous definition (default with method = "legacy").

• Default output is silent (verbose = FALSE); ask defaults to interactive().

• niqr(): standard errors are now computed with the analytic Hessian, so they differ from those of version 0.2.x.

• Removed (with an informative error when used):

  Removed in 1.0.0	Use instead
  testfit.niqr(obj, R)	test.fit(object, R)
  niqr(fun, fun2, x0, X, y, control)	niqr(formula, data, fun, fun2, x0)
  piqr(..., display = )	piqr(..., verbose = )
  gof.piqr(..., df.new = )	gof.piqr(..., df = )
  niqr() control names display, n_p, low_p, up_p, maxit_start, h1, meth, …	niqr.control(verbose, n.p, eps, ...); cluster is an argument of niqr()
  components x, fx, gx, Hx, p.star.y, code, iter of niqr objects	coefficients, obj.function, gradient, hessian, CDF, converged, n.it
  matrix newdata (with a column "y") in predict.niqr()	a data frame

Penalized models (piqr())

• New solver (method = "pnewton", default): proximal Newton algorithm with exact gradient and Hessian of the integrated loss, coordinate descent for the penalized quadratic subproblem, warm starts, sequential strong rules and KKT checks. It solves exactly the penalized problem of Sottile et al. (2020, eq. 3.4).
• New argument penalty = c("lasso", "group"): the group lasso selects covariates as a whole (all the coefficients of a covariate are zero or nonzero).
• New argument penalty.factor (covariate-specific penalty weights; 0 = unpenalized).
• The algorithm of version 0.2.x is still available with method = "legacy".
• New refit_piqr(): unpenalized re-estimation of the selected model with qrcm::iqr() (post-selection inference, Sottile et al., 2020, Section 3.2).
• New cv_piqr(): K-fold cross-validation of the integrated loss.
• gof.piqr(): Cn can be a number; new argument df ("parameters", the default for the new solver, or "covariates", the default for method = "legacy"); the result has a print() method and contains the criterion along the path.
• obj.function is now computed also for censored and truncated data, so that gof.piqr() works with Surv() responses.
• coef(), summary() and predict() gain the argument lambda; they no longer fail when the path contains a single value of lambda.
• plot(xvar = "beta"): the QQ-plot now uses ppoints() (it previously used sorted uniform random numbers); ask defaults to interactive().
• Messages are issued with message() (argument verbose).
• Covariance matrices are computed only by method = "legacy"; use refit_piqr().

Nonlinear models (niqr())

• New formula interface: niqr(formula, data, fun, fun2, x0, weights, cluster, control).
• The gradient is now correct also when fun2 is nonlinear in beta or in the covariates (e.g., Hill dose-response models); it was correct only for models linear in the covariates.
• Integrals over p are computed on a grid that is denser in the tails, splitting the cell that contains F(y | x), which makes loss and gradient smooth and consistent. The loss no longer requires O(n^2) operations.
• Optimization by Levenberg-Marquardt damped Newton with the analytic Gauss-Newton Hessian, instead of BFGS.
• Standard errors: sandwich formula with the analytic Hessian (previously the BFGS approximation, or the identity matrix if the algorithm stopped within 10 iterations), cluster-robust sandwich, bootstrap and cluster bootstrap (niqr.control(se = "boot")).
• New niqr.control().
• New methods coef(), vcov(), nobs(), terms(), model.matrix(); predict() gains type = "PDF" and accepts data frames as newdata; summary() reports the information criteria of Bottai and Cilluffo (2020).
• New gof.niqr() (AIC, BIC and their log versions, for one or more models).
• test.fit(): the Kolmogorov-Smirnov statistic is now two-sided.
• New test.fit() method for niqr objects; failed bootstrap replications no longer stop the procedure.

Package infrastructure

• The package no longer calls non-exported functions of ‘qrcm’ through getFromNamespace(): the required internal routines (adapted from ‘qrcm’, with attribution) are now part of ‘qrcmNP’. Jacobians are computed with a single cross-product of Khatri-Rao matrices.
• ‘qrcm’ and ‘survival’ moved from Depends to Imports; slp(), plf() and test.fit() are re-exported from ‘qrcm’. Note that library(qrcmNP) no longer attaches ‘qrcm’ and ‘survival’: scripts calling, e.g., iqr() or Surv() need library(qrcm) / library(survival) (or qrcm::iqr(), survival::Surv()).
• Documentation written with roxygen2; new vignette, tests (testthat) and NEWS.

qrcmNP 0.2.2

• Included the Authors@R field in the DESCRIPTION.
• Minor fixes to the linking targets problem.

qrcmNP 0.2.0

• Renamed internal functions.
• Improved starting points.
• New dependency ‘pch’ (>= 2.0).