Using Custom Outcome Models in gfoRmula
In gfoRmula: Parametric G-Formula

knitr::opts_chunk$set(echo = TRUE)

By default, the \verb|gfoRmula| package uses a pooled logistic regression model for survival outcomes, logistic regression model for binary end-of-follow-up outcomes, and a linear regression model for continuous end-of-follow-up outcomes. Starting from version 1.1.0, the \verb|gfoRmula| package allows users to apply their own type of outcome models. This document describes how to specify such custom outcome models. This document assumes that readers have read the long-form package documentation of McGrath et al. (2020).

Specifying custom outcome models

ymodel_fit_custom <- function(ymodel, obs_data){
  return(randomForest::randomForest(formula = ymodel, data = obs_data))
}

The function for obtaining estimates from the model must take the parameters \verb|fit| (the fitted outcome model) and \verb|newdf| (a \verb|data.table| containing the simulated dataset at time $t$). This function must return the estimated probability of the outcome for survival and binary end-of-follow-up outcomes or the estimated mean of the outcome for continuous end-of-follow-up outcomes in \verb|newdf|. Continuing with the random forest example, the code below obtains the estimated outcome mean for a continuous end-of-follow-up outcome. This code leverages the \verb|predict.randomForest| function in the \verb|randomForest| package.

ymodel_predict_custom <- function(fit, newdf){
  return(as.numeric(predict(object = fit, newdata = newdf)))
}

Example

We perform an analysis similar to that Example 3 in McGrath et al. (2020), except we use the custom outcome model from the previous section.

library('gfoRmula')
library('data.table')

library('Hmisc')
id <- 'id'
time_name <- 't0'
covnames <- c('L1', 'L2', 'A')
outcome_name <- 'Y'
outcome_type <- 'continuous_eof'
covtypes <- c('categorical', 'normal', 'binary')
histories <- c(lagged)
histvars <- list(c('A', 'L1', 'L2'))
covparams <- list(covmodels = c(L1 ~ lag1_A + lag1_L1 + L3 + t0 +
                                  rcspline.eval(lag1_L2, knots = c(-1, 0, 1)),
                                L2 ~ lag1_A + L1 + lag1_L1 + lag1_L2 + L3 + t0,
                                A ~ lag1_A + L1 + L2 + lag1_L1 + lag1_L2 + L3 + t0))
ymodel <- Y ~ A + L1 + L2 + lag1_A + lag1_L1 + lag1_L2 + L3
intervention1.A <- list(static, rep(0, 7))
intervention2.A <- list(static, rep(1, 7))
int_descript <- c('Never treat', 'Always treat')
nsimul <- 10000

gform_cont_eof <- gformula(obs_data = continuous_eofdata,
                           id = id, time_name = time_name,
                           covnames = covnames, outcome_name = outcome_name,
                           outcome_type = outcome_type, covtypes = covtypes,
                           covparams = covparams, ymodel = ymodel,
                           ymodel_fit_custom = ymodel_fit_custom, 
                           ymodel_predict_custom = ymodel_predict_custom,
                           intervention1.A = intervention1.A,
                           intervention2.A = intervention2.A,
                           int_descript = int_descript,
                           histories = histories, histvars = histvars,
                           basecovs = c("L3"), nsimul = nsimul, seed = 1234)
gform_cont_eof

References

McGrath S, Lin V, Zhang Z, Petito LC, Logan RW, Hernán MA, Young JG. gfoRmula: an R package for estimating the effects of sustained treatment strategies via the parametric g-formula. Patterns. 2020 Jun 12;1(3).