--- title: "Estimation of Flexible ITRs with xgboost" author: "Jared Huling" date: "`r Sys.Date()`" output: rmarkdown::html_vignette: fig_width: 7 fig_height: 5 toc: true toc_depth: 3 number_sections: true self_contained: true preamble: > \usepackage{amsmath,amssymb,amsfonts,bm,bbm,amsthm} \usepackage{longtable} \usepackage{booktabs} \usepackage{bbm, bm} \def\bsx {\boldsymbol x} \def\bsX {\boldsymbol X} \def\bsT {\boldsymbol T} \def\bsbeta {\boldsymbol \beta} \def\bsgamma {\boldsymbol \gamma} \def\bsW {\boldsymbol W} \def\bsy {\boldsymbol y} \def\bsY {\boldsymbol Y} \def\bsM {\boldsymbol M} \def\bfx {\mathbf x} \def\bfX {\mathbf X} \def\bfT {\mathbf T} \def\bfW {\mathbf W} \def\bfy {\mathbf y} \def\bfY {\mathbf Y} \def\bfM {\mathbf M} \def\bfU {\mathbf U} vignette: > %\VignetteIndexEntry{Estimation of Flexible ITRs with xgboost} %\VignetteEngine{knitr::rmarkdown} --- # First simulate data with complicated conditional average treatment effect/benefit score To demonstrate how to estimate flexible individualized treatment rules using `xgboost` in the `personalized` package, we simulate data with a binary treatment and a complex relationship between covariates and the effect of the treatment. The treatment assignments are based on covariates and thus mimic an observational setting with no unmeasured confounders. ```{r loadpkg, message=FALSE, warning=FALSE, echo=FALSE} library(personalized) ``` In this simulation, the treatment assignment depends on covariates and hence we must model the propensity score $\pi(x) = Pr(T = 1 | X = x)$. In this simulation we will assume that larger values of the outcome are better. ```{r sim_data_1, message = FALSE, warning = FALSE} library(personalized) set.seed(1) n.obs <- 500 n.vars <- 10 x <- matrix(rnorm(n.obs * n.vars, sd = 1), n.obs, n.vars) # simulate non-randomized treatment xbetat <- 0.5 + 0.25 * x[,1] - 0.25 * x[,5] trt.prob <- exp(xbetat) / (1 + exp(xbetat)) trt <- rbinom(n.obs, 1, prob = trt.prob) # simulate delta (CATE) as a complex function of the covariates delta <- 2*(0.25 + x[,1] * x[,2] - x[,3] ^ {-2} * (x[,3] > 0.35) + (x[,1] < x[,3]) - (x[,1] < x[,2])) # simulate main effects g(X) xbeta <- x[,1] + x[,2] + x[,4] - 0.2 * x[,4]^2 + x[,5] + 0.2 * x[,9] ^ 2 xbeta <- xbeta + delta * (2 * trt - 1) * 0.5 # simulate continuous outcomes y <- drop(xbeta) + rnorm(n.obs) ``` # Setup To estimate any ITR, we first must construct a propensity score function. We also optionally (and highly recommended for performance) can define an augmentation function that estimates main effects of covariates. The `personalized` package has functionality for doing so using cross-fitting (see the vignette for augmentation): ```{r} # arguments can be passed to cv.glmnet via `cv.glmnet.args`. # normally we would set the number of crossfit folds and internal folds to be larger, # but have reduced it to make computation time shorter prop.func <- create.propensity.function(crossfit = TRUE, nfolds.crossfit = 4, cv.glmnet.args = list(type.measure = "auc", nfolds = 3)) ``` ```{r} aug.func <- create.augmentation.function(family = "gaussian", crossfit = TRUE, nfolds.crossfit = 4, cv.glmnet.args = list(type.measure = "mse", nfolds = 3)) ``` For the sake of comparison, first fit a linear ITR. We have set `nfolds` to 3 for computational reasons; it should generally be higher, such as 10. ```{r} subgrp.model.linear <- fit.subgroup(x = x, y = y, trt = trt, propensity.func = prop.func, augment.func = aug.func, loss = "sq_loss_lasso", nfolds = 3) # option for cv.glmnet (for ITR estimation) summary(subgrp.model.linear) ``` # Using xgboost for estimation of ITRs The `{personalized}` package allows use of `{xgboost}` routines for direct estimation of the CATE (conditional average treatment effect) based on gradient boosted decision trees. This allows for highly flexible estimates of the CATE and thus benefit scores. Several arguments used by the `xgb.train()` and `xgb.cv()` functions from `{xgboost}` must be specified; they are: - `params`: the list of parameters for the underlying `xgboost` model (see help file for `xgb.train()`: this includes `eta`, `max_depth`, `nthread`, `subsample`, `colsample_bytree`, etc). However, note that `objective` and `eval_metric` will be overwritten, as they need to be set to custom values to work within `{personalized}`. - `nfold`: number of cross validation folds to be used by `xgb.cv()` for tuning - `nrounds`: the number of boosting iterations - `early_stopping_rounds`: can optionally be set. If set to an integer `k`, training will stop if the performance doesn't improve for `k` rounds. We have set `nfolds` to 3 for computational reasons; it should generally be higher, such as 10. ```{r} ## xgboost tuning parameters to use: param <- list(max_depth = 5, eta = 0.01, nthread = 1, booster = "gbtree", subsample = 0.623, colsample_bytree = 1) subgrp.model.xgb <- fit.subgroup(x = x, y = y, trt = trt, propensity.func = prop.func, augment.func = aug.func, ## specify xgboost via the 'loss' argument loss = "sq_loss_xgboost", nfold = 3, params = param, verbose = 0, nrounds = 5000, early_stopping_rounds = 50) subgrp.model.xgb ``` # Comparing performance with augmentation We first run the training/testing procedure to assess the performance of the non-augmented estimator (ideally, with the number of replications larger than `B=100`). Note we do not run this part due to computation time. ```{r eval=FALSE} valmod.lin <- validate.subgroup(subgrp.model.linear, B = 100, method = "training_test", train.fraction = 0.75) valmod.lin ``` Then we compare with the augmented estimator. Although this is based on just 5 replications, we can see that the augmented estimator is much better at discriminating between benefitting and non-benefitting patients, as evidenced by the large treatment effect among those predicted to benefit by the augmented estimator versus the smaller conditional treatment effect above. ```{r eval=FALSE} valmod.xgb <- validate.subgroup(subgrp.model.xgb, B = 100, method = "training_test", train.fraction = 0.75) valmod.xgb ``` We also plot the estimated CATE versus the true CATE for each approach: ```{r, fig.height=10} ## RMSE (note: this is still on the in-sample data so ## out-of-sample RMSE is preferred to evaluate methods) sqrt(mean((delta - treatment.effects(subgrp.model.linear)$delta) ^ 2)) sqrt(mean((delta - treatment.effects(subgrp.model.xgb)$delta) ^ 2)) par(mfrow = c(2,1)) plot(delta ~ treatment.effects(subgrp.model.linear)$delta, ylab = "True CATE", xlab = "Estimated Linear CATE") abline(a=0,b=1,col="blue") plot(delta ~ treatment.effects(subgrp.model.xgb)$delta, ylab = "True CATE", xlab = "CATE via xgboost") abline(a=0,b=1,col="blue") ```