Sample Splitting

This is an example using the star dataset (for more information about the dataset, please use ?star).

We start with a simple example with one outcome variable (writing scores) and one machine learning algorithm (causal forest). Then we move to incoporate multiple outcomes and compare model performances with several machine learning algorithms.

To begin, we load the dataset and specify the outcome variable and covariates to be used in the model. Next, we utilize a random forest algorithm to develop an Individualized Treatment Rule (ITR) for estimating the varied impacts of small class sizes on students’ writing scores. Since the treatment is often costly for most policy programs, we consider a case with 20% budget constraint (budget = 0.2). The model will identify the top 20% of units who benefit from the treatment most and assign them to with the treatment. We train the model through sample splitting, with the split_ratio between the train and test sets determined by the split_ratio argument. Specifically, we allocate 70% of the data to train the model, while the remaining 30% is used as testing data (split_ratio = 0.7).

library(dplyr)
library(evalITR)

# specifying the outcome
outcomes <- "g3tlangss"

# specifying the treatment
treatment <- "treatment"

# specifying the data (remove other outcomes)
star_data <- star %>% dplyr::select(-c(g3treadss,g3tmathss))

# specifying the formula
user_formula <- as.formula(
  "g3tlangss ~ treatment + gender + race + birthmonth + 
  birthyear + SCHLURBN + GRDRANGE + GKENRMNT + GKFRLNCH + 
  GKBUSED + GKWHITE ")


# estimate ITR 
fit <- estimate_itr(
  treatment = treatment,
  form = user_formula,
  data = star_data,
  algorithms = c("causal_forest"),
  budget = 0.2,
  split_ratio = 0.7)
#> Evaluate ITR under sample splitting ...


# evaluate ITR 
est <- evaluate_itr(fit)
#> Cannot compute PAPDp

Statistics	Description
`PAPE`	population average prescriptive effect
`PAPEp`	population average prescriptive effect with a budget constraint
`PAPDp`	population average prescriptive effect difference with a budget constraint (this quantity will be computed with more than 2 machine learning algorithms)
`AUPEC`	area under the prescriptive effect curve
`GATE`	grouped average treatment effects

# summarize estimates
summary(est)
#> -- PAPE ------------------------------------------------------------------------
#>   estimate std.deviation     algorithm statistic p.value
#> 1      1.2           1.3 causal_forest      0.91    0.36
#> 
#> -- PAPEp -----------------------------------------------------------------------
#>   estimate std.deviation     algorithm statistic p.value
#> 1      1.6           1.1 causal_forest       1.4    0.15
#> 
#> -- PAPDp -----------------------------------------------------------------------
#> Cannot compute PAPDp
#> 
#> -- AUPEC -----------------------------------------------------------------------
#>   estimate std.deviation     algorithm statistic p.value
#> 1      1.3          0.96 causal_forest       1.4    0.18
#> 
#> -- GATE ------------------------------------------------------------------------
#>   estimate std.deviation     algorithm group statistic p.value upper lower
#> 1   -210.7           107 causal_forest     1    -1.976   0.048  -386   -35
#> 2     37.4           108 causal_forest     2     0.346   0.729  -140   215
#> 3     -4.0           109 causal_forest     3    -0.037   0.971  -183   175
#> 4      7.4           108 causal_forest     4     0.068   0.945  -171   186
#> 5    195.9           108 causal_forest     5     1.813   0.070    18   374

We can extract estimates from the est object. The following code shows how to extract the GATE estimates for the writing score with the causal forest algorithm. We can also plot the estimates using the plot_estimate() function and specify the type of estimates to be plotted (GATE, PAPE, PAPEp, PAPDp).

# plot GATE estimates
library(ggplot2)
gate_est <- summary(est)$GATE

plot_estimate(gate_est, type = "GATE") +
  scale_color_manual(values = c("#0072B2", "#D55E00"))

We plot the estimated Area Under the Prescriptive Effect Curve for the writing score across a range of budget constraints for causal forest.