get_ns_spline {SimSurvNMarker}

Faster Pointwise Function than ns

Description

Creates a function which can evaluate a natural cubic spline like ns.

The result may differ between different BLAS and LAPACK implementations as the QR decomposition is not unique. However, the column space of the returned matrix will always be the same regardless of the BLAS and LAPACK implementation.

Usage

get_ns_spline(knots, intercept = TRUE, do_log = TRUE)

Arguments

knots

sorted numeric vector with boundary and interior knots.

intercept

logical for whether to include an intercept.

do_log

logical for whether to evaluate the spline at log(x) or x.

Examples

# compare with splines
library(splines)
library(SimSurvNMarker)
xs <- seq(1, 5, length.out = 10L)
bks <- c(1, 5)
iks <- 2:4

# we get the same
if(require(Matrix)){
  r1 <- unclass(ns(xs, knots = iks, Boundary.knots = bks, intercept = TRUE))
  r2 <- get_ns_spline(knots = sort(c(iks, bks)), intercept = TRUE,
                      do_log = FALSE)(xs)

  cat("Rank is correct:      ", rankMatrix(cbind(r1, r2)) == NCOL(r1), "\n")

  r1 <- unclass(ns(log(xs), knots = log(iks), Boundary.knots = log(bks),
                   intercept = TRUE))
  r2 <- get_ns_spline(knots = log(sort(c(iks, bks))), intercept = TRUE,
                      do_log = TRUE)(xs)
  cat("Rank is correct (log):", rankMatrix(cbind(r1, r2)) == NCOL(r1), "\n")
}

# the latter is faster
system.time(
  replicate(100,
            ns(xs, knots = iks, Boundary.knots = bks, intercept = TRUE)))
system.time(
  replicate(100,
            get_ns_spline(knots = sort(c(iks, bks)), intercept = TRUE,
                          do_log = FALSE)(xs)))
func <- get_ns_spline(knots = sort(c(iks, bks)), intercept = TRUE,
                      do_log = FALSE)
system.time(replicate(100, func(xs)))