Estimating and Evaluating the Optimal Subgroup

library(polle)
library(data.table)

This vignette showcase how policy_learn() and policy_eval() can be combined to estimate and evaluate the optimal subgroup in the single-stage case. We refer to (Nordland and Holst 2023) for the syntax and methodological context.

Setup

From here on we consider the single-stage case with a binary action set \(\{0,1\}\). For a given threshold \(\eta > 0\) we can formulate the optimal subgroup function via the conditional average treatment effect (CATE/blip) as

\[\begin{align*} d^\eta_0(v)_ = I\{B_0(v) > \eta\}, \end{align*}\] where \(B_0\) is the CATE defined as \[\begin{align*} E\left[U^{(1)} - U^{(0)} \big | V = v \right]. \end{align*}\]

The average treatment effect in the optimal subgroup is now defined as \[\begin{align*} E\left[U^{(1)} - U^{(0)} \big | d^\eta_0(V) = 1 \right], \end{align*}\]

which under consistency, positivity and randomization is identified as

\[\begin{align*} E\left[Z(1,g_0,Q_0)(O) - Z(0,g_0,Q_0)(O) \big | d^\eta_0(V) = 1 \right], \end{align*}\]

where \(Z(a,g,Q)(O)\) is the doubly robust score for treatment \(a\) and

\[\begin{align*} d^\eta_0(v) &= I\{B_0(v) > \eta\}\\ B_0(v) &= E\left[Z(1,g_0,Q_0)(O) - Z(0,g_0,Q_0)(O) \big | V = v \right] \end{align*}\]

Threshold policy learning

In polle the threshold policy \(d_\eta\) can be estimated using policy_learn() via the threshold argument, and the average treatment effect in the subgroup can be estimated using policy_eval() setting target = subgroup.

Here we consider an example using simulated data:

par0 <- list(a = 1, b = 0, c = 3)
sim_d <- function(n, par=par0, potential_outcomes = FALSE) {
  W <- runif(n = n, min = -1, max = 1)
  L <- runif(n = n, min = -1, max = 1)
  A <- rbinom(n = n, size = 1, prob = 0.5)
  U1 <- W + L + (par$c*W + par$a*L + par$b) # U^1
  U0 <- W + L # U^0
  U <- A * U1 + (1 - A) * U0 + rnorm(n = n)
  out <- data.table(W = W, L = L, A = A, U = U)
  if (potential_outcomes == TRUE) {
    out$U0 <- U0
    out$U1 <- U1
  }
  return(out)
}

Note that in this simple case \(U^{(1)} - U^{(0)} = cW + aL + b\).

set.seed(1)
d <- sim_d(n = 200)
pd <- policy_data(
    d,
    action = "A",
    covariates = list("W", "L"),
    utility = "U"
)

We set a correctly specified policy learner using policy_learn() with type = "blip" and set a threshold of \(\eta = 1\):

pl1 <- policy_learn(
  type = "blip",
  control = control_blip(blip_models = q_glm(~ W + L)),
  threshold = 1
)

When then apply the policy learner based on the correctly specified nuisance models. Furthermore, we extract the corresponding policy actions, where \(d_N(Z,L) = 1\) identifies the optimal subgroup for \(\eta = 1\):

po1 <- pl1(
  policy_data = pd,
  g_models = g_glm(~ 1),
  q_models = q_glm(~ A * (W + L))
)
pf1 <- get_policy(po1)
pa <- pf1(pd)

In the following plot, the black line indicates the boundary for the true optimal subgroup. The dots represent the estimated threshold policy:

Similarly, we can also use type = "ptl" to fit a policy tree with a given threshold for not choosing the reference action (first action in action set in alphabetical order)

get_action_set(pd)[1] # reference action
## [1] "0"
pl1_ptl <- policy_learn(
    type = "ptl",
    control = control_ptl(policy_var = c("W", "L")),
    threshold = 1
)
## Loading required namespace: policytree
po1_ptl <- pl1_ptl(
  policy_data = pd,
  g_models = g_glm(~ 1),
  q_models = q_glm(~ A * (W + L))
)
po1_ptl$ptl_objects
## $stage_1
## $stage_1$threshold_1
## policy_tree object 
## Tree depth:  2 
## Actions:  1: 0 2: 1 
## Variable splits: 
## (1) split_variable: W  split_value: -0.0948583 
##   (2) split_variable: W  split_value: -0.107529 
##     (4) * action: 1 
##     (5) * action: 2 
##   (3) split_variable: W  split_value: 0.197522 
##     (6) * action: 1 
##     (7) * action: 2

Subgroup average treatment effect

The true subgroup average treatment effect is given by:

\[\begin{align*} E[cW + aL + b | cW + aL + b \geq \eta ], \end{align*}\]

which we can easily approximate:

set.seed(1)
approx <- sim_d(n = 1e7, potential_outcomes = TRUE)
(sate <- with(approx, mean((U1 - U0)[(U1 - U0 >= 1)])))
## [1] 2.082982
rm(approx)

The subgroup average treatment effect associated with the learned optimal threshold policy can be directly estimated using policy_eval() via the target argument:

(pe <- policy_eval(
  policy_data = pd,
  policy_learn = pl1,
  target = "subgroup"
 ))
##                                 Estimate Std.Err  2.5% 97.5%   P-value
## E[Z(1)-Z(0)|d=1]: d=blip(eta=1)    1.941  0.2614 1.428 2.453 1.136e-13

We can also estimate the subgroup average treatment effect for a set of thresholds at once:

pl_set <- policy_learn(
  type = "blip",
  control = control_blip(blip_models = q_glm(~ W + L)),
  threshold = c(0, 1)
)

policy_eval(
  policy_data = pd,
  g_models = g_glm(~ 1),
  q_models = q_glm(~ A * (W + L)),
  policy_learn = pl_set,
  target = "subgroup"
)
##                                 Estimate Std.Err  2.5% 97.5%   P-value
## E[Z(1)-Z(0)|d=1]: d=blip(eta=0)    1.641  0.2161 1.217 2.064 3.118e-14
## E[Z(1)-Z(0)|d=1]: d=blip(eta=1)    1.935  0.2612 1.423 2.447 1.268e-13

Asymptotics

The data adaptive target parameter

\[\begin{align*} E[U^{(1)} - U^{(0)}| d_N(V) = 1] = E[Z_0(1,g,Q)(O) - Z_0(0,g,Q)(O)| d_N(V) = 1] \end{align*}\]

is asymptotically normal with influence function

\[\begin{align*} \frac{1}{P(d'(\cdot) = 1)} I\{d'(\cdot) = 1\}\left\{Z(1,g,Q)(O) - Z(0,g,Q)(O) - E[Z(1,g,Q)(O) - Z(0,g,Q)(O) | d'(\cdot) = 1]\right\}, \end{align*}\]

where \(d'\) is the limiting policy of \(d_N\). The fitted influence curve can be extracted using IC():

IC(pe) |> head()
##           [,1]
## [1,]  0.000000
## [2,]  0.000000
## [3,]  0.000000
## [4,] -1.780405
## [5,]  0.000000
## [6,]  9.520253

References

Nordland, Andreas, and Klaus K. Holst. 2023. “Policy Learning with the Polle Package.” https://doi.org/10.48550/arXiv.2212.02335.