Compare Estimated and User Defined ITR

Estimated vs. User Defined ITR

The package allows to compare the performance of estimated ITRs with user defined ITRs. The estimate_itr function takes the following arguments:

Argument Description
fit a fitted object from the estimate_itr function
user_itr a function defined by users that returns a unit-level continuous score for treatment assignment (we assume those that have score less than 0 should not have treatment)
data a data frame
treatment a character string specifying the treatment variable in the data
outcome a character string specifying the outcome variable in the data
budget a numeric value specifying the maximum percentage of population that can be treated under the budget constraint

The function returns an object that contains the estimated GATE, ATE, and AUPEC for the user defined ITR.


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

# user's own ITR
score_function <- function(data){
  data %>% 
    mutate(score = case_when(
      school_urban == 1 ~ 0.1, # inner-city
      school_urban == 2 ~ 0.2, # suburban
      school_urban == 3 ~ 0.4, # rural
      school_urban == 4 ~ 0.3, # urban
    )) %>%
    pull(score) -> score
    
  return(score)
}

# evalutate ITR
compare_itr <- evaluate_itr(
  fit = fit,
  user_itr = score_function,
  data = star_data,
  treatment = "T",
  outcome = outcomes,
  budget = 0.2)
#> Cannot compute PAPDp

# summarize estimates
summary(compare_itr)
#> -- PAPE ------------------------------------------------------------------------
#>   estimate std.deviation     algorithm statistic p.value
#> 1    -0.37           1.4 causal_forest     -0.26     0.8
#> 2     0.00           0.0      user_itr       NaN     NaN
#> 
#> -- PAPEp -----------------------------------------------------------------------
#>   estimate std.deviation     algorithm statistic p.value
#> 1    -0.11          1.08 causal_forest      -0.1    0.92
#> 2     1.06          0.67      user_itr       1.6    0.11
#> 
#> -- PAPDp -----------------------------------------------------------------------
#> Cannot compute PAPDp
#> 
#> -- AUPEC -----------------------------------------------------------------------
#>   estimate std.deviation     algorithm statistic p.value
#> 1    -0.63          1.07 causal_forest     -0.59   0.555
#> 2    -0.91          0.42          <NA>     -2.19   0.028
#> 
#> -- GATE ------------------------------------------------------------------------
#>    estimate std.deviation     algorithm group statistic p.value  upper lower
#> 1       -34           108 causal_forest     1     -0.32   0.750 -212.2   143
#> 2      -223           107 causal_forest     2     -2.08   0.038 -399.9   -47
#> 3       181           108 causal_forest     3      1.68   0.093    3.8   359
#> 4       -28           108 causal_forest     4     -0.26   0.795 -205.4   149
#> 5       126           106 causal_forest     5      1.19   0.235  -48.7   301
#> 6       126            58      user_itr     1      2.16   0.031   30.2   222
#> 7        96            59      user_itr     2      1.62   0.105   -1.4   194
#> 8       -33            59      user_itr     3     -0.56   0.579 -129.7    64
#> 9      -139            59      user_itr     4     -2.36   0.018 -236.5   -42
#> 10      -32            59      user_itr     5     -0.54   0.589 -129.4    65

We plot the estimated Area Under the Prescriptive Effect Curve (AUPEC) for the writing score across a range of budget constraints for user defined ITR and estimated ITRs. The plot shows that the estimated ITRs have better performance than the user defined ITR.

# plot the AUPEC 
plot(compare_itr)

Existing Model vs. User-Defined Model

The package also allows to compare the performance of estimated ITRs of existing ML packages with user defined models. The following code shows an example using causal forest from the grf package with sample splitting. The estimate_itr function takes the following arguments:

Argument Description
treatment a character string specifying the treatment variable in the data
form a formula specifying the outcome and covariates
data a data frame
algorithms a character vector specifying the ML algorithms to be used
budget a numeric value specifying the maximum percentage of population that can be treated under the budget constraint
split_ratio a character string specifying the outcome variable in the data
user_model a character string specifying the user defined model

The user_model input should be a function that takes two arguments: training_data and test_data. The function will make use of the training_data to fit a model and then use the test_data to estimate CATE or other metrics of interest. It should also specify the way to get the ITR, based on the estimated effects.

In the following example, we fit a linear model with sample splitting and use the estimated CATE. We compute the ITR by assigning treatment to those with positive CATE and no treatment to those with negative CATE. The function user_model takes in the training data and test data and return a list that contains (1) an ITR; (2) a fitted model; and (3) a continuous score with the same length as the input data.

# user-defined model
user_model <- function(training_data, test_data){

  # model fit on training data
  fit <- train_model(training_data)
  
  # estimate CATE on test data
  compute_hatf <- function(fit, test_data){

    score <- fit_predict(fit, test_data)  
    itr   <- score_function(score)
    
    return(list(itr = itr, score = score))
  }

  hatf <- compute_hatf(fit, test_data)
  
  return(list(
    itr = hatf$itr, 
    fit = fit, 
    score = hatf$score))
}

Note that the user defined model can be any model that returns a unit-level continuous score for treatment assignment. It does not have to be a linear model or model that estimate CATE. We can specify custom functions in the train_model function and the fit_predict function to compute the score. If the model does not have a default predict function, we need to write up a custom function with fit_predict.

# train model
train_model <- function(data){
  fit <- lm(
    Y ~ T*(cov1 + cov1 + cov3), 
    data = data)
  return(fit)
}

# predict function
fit_predict <- function(fit, data){
  # need to change this function if 
  # the model does not have a default predict function
  score <- predict(fit, data) 
  return(score)
}

In addition, we can also choose any scoring rule that maps the score to a binary indicator of treatment assignment.

# score function
score_function <- function(score){
  itr <- (score >= 0) * 1
  return(itr)
}

If split_ratio is specified, the function will split the data into training and test data. The split_ratio should be a numeric value between 0 and 1. Alternatively, if n_folds is specified, the function will use the entire data to fit the user defined model via cross-validation.

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


# evaluate ITR 
compare_est <- evaluate_itr(compare_fit)

# summarize estimates
summary(compare_est)
#> -- PAPE ------------------------------------------------------------------------
#>   estimate std.deviation     algorithm statistic p.value
#> 1  2.9e-01           1.2 causal_forest      0.25    0.81
#> 2 -2.2e-16           0.0    user_model      -Inf    0.00
#> 
#> -- PAPEp -----------------------------------------------------------------------
#>   estimate std.deviation     algorithm statistic p.value
#> 1     2.56           1.1 causal_forest      2.31   0.021
#> 2     0.62           1.0    user_model      0.61   0.542
#> 
#> -- PAPDp -----------------------------------------------------------------------
#>   estimate std.deviation                  algorithm statistic p.value
#> 1      1.9           1.6 causal_forest x user_model       1.2    0.23
#> 
#> -- AUPEC -----------------------------------------------------------------------
#>   estimate std.deviation     algorithm statistic p.value
#> 1     0.20          0.93 causal_forest      0.22    0.83
#> 2    -0.11          0.81    user_model     -0.13    0.89
#> 
#> -- GATE ------------------------------------------------------------------------
#>    estimate std.deviation     algorithm group statistic p.value upper lower
#> 1       -72           108 causal_forest     1     -0.66 5.1e-01  -250   106
#> 2       107           107 causal_forest     2      1.00 3.2e-01   -69   284
#> 3       -94           108 causal_forest     3     -0.87 3.9e-01  -272    84
#> 4        57           108 causal_forest     4      0.53 6.0e-01  -121   235
#> 5        17           107 causal_forest     5      0.16 8.7e-01  -159   193
#> 6      -137           106    user_model     1     -1.30 1.9e-01  -311    37
#> 7      -571           104    user_model     2     -5.47 4.5e-08  -743  -400
#> 8      -529           105    user_model     3     -5.02 5.2e-07  -702  -355
#> 9       436           107    user_model     4      4.06 4.9e-05   260   613
#> 10      817           104    user_model     5      7.83 5.1e-15   645   989
plot(compare_est)