Compute a value score under a candidate treatment rule

Evaluates the mean outcome under a candidate treatment rule by assigning, for each subject, the outcome corresponding to the treatment actually recommended by the rule. If the predicted treatment matches the observed treatment, the observed outcome under the received treatment is used; otherwise, the alternative counterfactual or model-based outcome is used.

Usage

my_score.Surv(pred, A, Q, pQ)

Arguments

pred

A vector of predicted treatment assignments under the candidate rule.

A

A vector of observed treatment assignments.

Q

A numeric vector giving the outcome value to use when the predicted treatment matches the observed treatment, i.e., for subjects with pred == A.

pQ

A numeric vector giving the outcome value to use when the predicted treatment does not match the observed treatment, i.e., for subjects with pred != A.

Value

A single numeric value equal to the mean outcome under the candidate treatment rule.

Details

This function is useful in dynamic treatment regime or policy-learning settings where one wants to estimate the value of a learned treatment rule using observed and predicted treatment assignments together with outcome quantities under concordant and discordant treatment choices.

For each subject, the function constructs $$ Q_i^* = \begin{cases} Q_i, & \text{if } pred_i = A_i, \\ pQ_i, & \text{if } pred_i \neq A_i. \end{cases} $$ and returns the sample mean of \(Q_i^*\).

The function assumes that pred, A, Q, and pQ are all aligned and of equal length.

Examples

pred <- c(1, -1, 1, -1)
A    <- c(1,  1, 1, -1)
Q    <- c(10, 12,  8, 15)
pQ   <- c( 9, 11,  7, 14)

my_score.Surv(pred = pred, A = A, Q = Q, pQ = pQ)
#> [1] 11