addNonlinearity {Greg}

Add a nonlinear function to the model

Description

This function takes a model and adds a non-linear function if the likelihood-ratio supports this (via the anova(..., test = "chisq") test for stats while for rms you need to use the rcs() spline that is automatically evaluated for non-linearity).

Usage

addNonlinearity(
  model,
  variable,
  spline_fn,
  flex_param = 2:7,
  min_fn = AIC,
  sig_level = 0.05,
  verbal = FALSE,
  workers,
  ...
)

## S3 method for class 'negbin'
addNonlinearity(model, ...)

Arguments

model

The model that is to be evaluated and adapted for non-linearity

variable

The name of the parameter that is to be tested for non-linearity. Note that the variable should be included plain (i.e. as a linear variable) form in the model.

spline_fn

Either a string or a function that is to be used for testing alternative non-linearity models

flex_param

A vector with values that are to be tested as the default second parameter for the non-linearity function that you want to evaluate. This defaults to 2:7, for the ns() it tests the degrees of freedom ranging between 2 and 7.

min_fn

This is the function that we want to minimized if the variable supports the non-linearity assumption. E.g. BIC() or AIC, note that the BIC() will in the majority of cases support a lower complexity than the AIC().

sig_level

The significance level for which the non-linearity is deemed as significant, defaults to 0.05.

verbal

Set this to TRUE if you want print statements with the anova test and the chosen knots.

workers

The function tries to run everything in parallel. Under some circumstances you may want to restrict the number of parallel threads to less than the default detectCores() - 1, e.g. you may run out of memory then you can provide this parameter. If you do not want to use parallel then simply set workers to FALSE. The cluster created using makeCluster() function.

...

Passed onto internal prNlChooseDf() function.

Examples

library(Greg)
data("melanoma", package = "boot", envir = environment())

library(dplyr)
melanoma <- mutate(melanoma, 
                   status = factor(status,
                                   levels = 1:3,
                                   labels = c("Died from melanoma", 
                                              "Alive", 
                                              "Died from other causes")),
                   ulcer = factor(ulcer,
                                  levels = 0:1,
                                  labels = c("Absent", "Present")),
                   time = time/365.25, # All variables should be in the same time unit
                   sex = factor(sex,
                                levels = 0:1,
                                labels = c("Female", "Male")))

library(survival)
model <- coxph(Surv(time, status == "Died from melanoma") ~ sex + age,
  data = melanoma
)

nl_model <- addNonlinearity(model, "age",
  spline_fn = "pspline",
  verbal = TRUE,
  workers = FALSE
)
# Note that there is no support for nonlinearity in this case