TreatmentSelection (version 2.1.1)

trtsel: create a trtsel object

Description

Creates an object of class "trtsel" given a data.frame containing marker, treatment, and adverse event status information. The functions "plot.trtsel", "evaluate.trtsel", and "calibrate.trtsel" can then be used to plot risk and treatment effect curves, estimate summary measures, and check model calibration. The function "compare.trtsel" can also be used to compare two treatment selection markers.

Usage

trtsel(formula, treatment.name, data, fittedrisk.t0 = NULL,
  fittedrisk.t1 = NULL, thresh = 0, prediction.time = NULL,
  study.design = c("RCT", "nested case-control",
  "stratified nested case-control"), cohort.attributes = NULL, family,
  default.trt = c("trt all", "trt none"))

Arguments

formula

a 'formula' object consisting of outcome ~ markers and marker by treatment interactions for the treatment selection model to be evaluated. The outcome can be continuous, binary or a 'Surv' object for time-to-event outcomes. Binary variable should equal 1 for cases and 0 for controls. For continuous outcomes, we assume that larger values are associated with worse outcomes.

treatment.name

Name of the treatment variable in data.frame "data". The treatment variable must be coded 1 for "treated" and 0 for "un-treated."

data

data.frame object used to fit and evaluate the model.

fittedrisk.t0

Instead of providing a marker, fitted risks for T=0 and T=1 may be provided. This should be the column name of the fitted risk for T=0 that can be found in 'data'. If fitted risks are provided, a model will not be fit, and all bootstrap confidence intervals will be conditional on the model fit provided.

fittedrisk.t1

Name of for the fitted risks given treatment = 1.

thresh

The treatment effect threshold used to define the treatment rule: Do not treat if the marker-specific treatment effect is less than "thresh". This is a numeric constant with a default value of 0.

prediction.time

a landmark prediction time used only when the outcome is a time-to-event Surv object

study.design

Character string indicating the study design used to collect the data. The three options are "RCT" (default), "nested case-control", or "stratified nested case-control". A "RCT" design is simply a randomized trial comparing T = 0 to T = 1 with the marker measured at baseline. A nested case-control or stratified nested case-control study samples cases and controls for marker measurement, perhaps stratified on treatment assignment, from a randomized trial comparing T = 0 to T = 1. See Janes et al. (2013) for a full description of these designs.

If a "nested case-control" or "stratified nested case-control" design is specified, "cohort.attributes"" must be provided (see below).

cohort.attributes

If a "nested case-control" or "stratified nested case-control" design is specified, "cohort.attributes" must be provided. Order does matter.

For the "nested case-control" design, specify the following attributes of the randomized trial "cohort" from which the case-control sample was selected: cohort.attributes = c(cohort sample size, proportion treated in cohort (Pr(trt==1)), adverse event prevalance in cohort (Pr(event==1)), fraction of cases sampled from cohort)

For the "stratitified nested case-control" design, specify the following attributes of the randomized trial "cohort" from which the stratified case-control sample was selected: cohort.attributes = c(cohort sample size, Pr(trt==0 & event==0) in cohort, Pr(trt==0 & event==1) in cohort, Pr(trt==1 & event==0) in cohort, fraction of cases with trt == 0 sampled from cohort, fraction of cases with trt == 1 sampled from cohort )

family

function passed to `glm` describing the error distribution and link function to be used in the model. Defaults to binomial(link = "logit") when the outcome is numeric with two values 0,1 and gaussian(link = "identity") otherwise. Ignored for time-to-event outcomes. For non RCT study designs, methods only exist for family = binomial(link = "logit").

default.trt

The default treatment assignment to compare with marker-based treatment. Can either be set at "trt all" (default) or "trt none". Use "trt all" if everyone is treated and the aim is to discover those who would benefit from no treatment, but use "trt none" if the common practice is to treat no-one and the goal is to discover those who would benefit from treatment.

Value

An object of class "trtsel," which is a list containing:

model.fit

A list containing "coefficients" -- a 4 x 4 matrix with columns for coefficient estimates, standard errors, t-statistics, and two-sided p-values. "cohort.attributes" -- the vector of cohort.attributes provided "study.design" -- character string of study.design provided

derived.data

A data.frame with event, trt, and marker information along with "fittedrisk.t0" (risk estimates given no treatment), "fittedrisk.t1" (risk estimates given treatment), "trt.effect" (treatment effect estimates), and "trt.rec" (indicator of trt recommendation) columns.

functions

For internal package use only

References

Janes, Holly; Brown, Marshall D; Pepe, Margaret; Huang, Ying; "An Approach to Evaluating and Comparing Biomarkers for Patient Treatment Selection" The International Journal of Biostatistics. Volume 0, Issue 0, ISSN (Online) 1557-4679, ISSN (Print) 2194-573X, DOI: 10.1515/ijb-2012-0052, April 2014

See Also

plot.trtsel for plotting risk curves and more, evaluate.trtsel for evaluating marker performance, calibrate.trtsel for assessing model calibration, and compare.trtsel to compare two trtsel object.

Examples

# NOT RUN {

data(tsdata)

###########################
## Create trtsel objects 
###########################

trtsel.Y1 <- trtsel(event ~ Y1*trt, 
                   treatment.name = "trt", 
                   data = tsdata, 
                   study.design = "RCT",
                   family = binomial("logit"), 
                   default.trt = "trt all")

trtsel.Y1

trtsel.Y2 <- trtsel(event ~ Y2*trt, 
                   treatment.name = "trt", 
                   data = tsdata, 
                   default.trt = "trt all")
trtsel.Y2


# calculate fitted risks using a logistic model 
#(one can use any model here, the point is that the fitted risks are provided )
mymod <- glm(event~trt*Y2, data= tsdata, family = binomial("logit"))

tsdata$fitted.t0 <- predict(mymod, newdata=data.frame(trt = 0, Y2 = tsdata$Y2), type = "response")
tsdata$fitted.t1 <- predict(mymod, newdata=data.frame(trt = 1, Y2 = tsdata$Y2), type = "response")

#all bootstrapping done using this object will be conditional on the model fit. 

myfitted.trtsel <- trtsel( event~trt, treatment.name = "trt", 
                           data = tsdata,
                           fittedrisk.t0 = "fitted.t0",
                           fittedrisk.t1 = "fitted.t1",
                           default.trt = "trt all")



# }