We can estimate ITR with various machine learning algorithms and then
compare the performance of each model. The package includes all ML
algorithms in the caret
package and 2 additional algorithms
(causal
forest and bartCause).
The package also allows estimate heterogeneous treatment effects on
the individual and group-level. On the individual-level, the summary
statistics and the AUPEC plot show whether assigning individualized
treatment rules may outperform complete random experiment. On the
group-level, we specify the number of groups through ngates
and estimating heterogeneous treatment effects across groups.
library(evalITR)
#> Loading required package: MASS
#>
#> Attaching package: 'MASS'
#> The following object is masked from 'package:dplyr':
#>
#> select
#> Loading required package: Matrix
#> Loading required package: quadprog
# specify the trainControl method
<- caret::trainControl(
fitControl method = "repeatedcv",
number = 2,
repeats = 2)
# estimate ITR
set.seed(2021)
<- estimate_itr(
fit_cv treatment = "treatment",
form = user_formula,
data = star_data,
trControl = fitControl,
algorithms = c(
"causal_forest",
# "bartc",
# "rlasso", # from rlearner
# "ulasso", # from rlearner
"lasso" # from caret package
# "rf" # from caret package
# from caret package
), budget = 0.2,
n_folds = 2)
#> Evaluate ITR with cross-validation ...
#> Loading required package: lattice
#> Loading required package: ggplot2
#> Warning: model fit failed for Fold1.Rep1: fraction=0.9 Error in elasticnet::enet(as.matrix(x), y, lambda = 0, ...) :
#> Some of the columns of x have zero variance
#> Warning: model fit failed for Fold1.Rep2: fraction=0.9 Error in elasticnet::enet(as.matrix(x), y, lambda = 0, ...) :
#> Some of the columns of x have zero variance
#> Warning in nominalTrainWorkflow(x = x, y = y, wts = weights, info = trainInfo,
#> : There were missing values in resampled performance measures.
# evaluate ITR
<- evaluate_itr(fit_cv)
est_cv #>
#> Attaching package: 'purrr'
#> The following object is masked from 'package:caret':
#>
#> lift
# summarize estimates
summary(est_cv)
#> -- PAPE ------------------------------------------------------------------------
#> estimate std.deviation algorithm statistic p.value
#> 1 1.7 1.3 causal_forest 1.4 0.18
#> 2 1.2 1.0 lasso 1.1 0.26
#>
#> -- PAPEp -----------------------------------------------------------------------
#> estimate std.deviation algorithm statistic p.value
#> 1 1.27 0.93 causal_forest 1.37 0.17
#> 2 -0.22 0.80 lasso -0.27 0.79
#>
#> -- PAPDp -----------------------------------------------------------------------
#> estimate std.deviation algorithm statistic p.value
#> 1 1.5 0.99 causal_forest x lasso 1.5 0.13
#>
#> -- AUPEC -----------------------------------------------------------------------
#> estimate std.deviation algorithm statistic p.value
#> 1 1.56 0.83 causal_forest 1.87 0.062
#> 2 0.52 1.20 lasso 0.43 0.666
#>
#> -- GATE ------------------------------------------------------------------------
#> estimate std.deviation algorithm group statistic p.value upper lower
#> 1 -46 82 causal_forest 1 -0.566 0.57 114 -207
#> 2 -59 59 causal_forest 2 -1.007 0.31 56 -175
#> 3 72 79 causal_forest 3 0.913 0.36 227 -83
#> 4 -33 65 causal_forest 4 -0.509 0.61 94 -160
#> 5 85 59 causal_forest 5 1.441 0.15 201 -31
#> 6 27 83 lasso 1 0.321 0.75 188 -135
#> 7 -60 80 lasso 2 -0.750 0.45 96 -215
#> 8 80 76 lasso 3 1.043 0.30 230 -70
#> 9 -4 82 lasso 4 -0.048 0.96 156 -164
#> 10 -24 83 lasso 5 -0.293 0.77 138 -187
We plot the estimated Area Under the Prescriptive Effect Curve for the writing score across different ML algorithms.
# plot the AUPEC with different ML algorithms
plot(est_cv)