R/target.R
target_Qbarbar_M1_times_M2_star_a.Rd
There are two interesting features of this
targeting problem. First, we see that the nuisance parameter Qbarbar_M1_times_M2_star_a
can be viewed in two ways: (1) the conditional mean of Qbarbar_M1_a given C with respect
to the marginal of M_2 given A = a_star, C; (2) the conditional mean of Qbarbar_M2_star_a given C
with respect to the marginal of M_1 given A = a, C. The natural inclination then
is to use a sum loss function. It seems that we're able to do that here.
However, to generate the proper score, we would need to
consider a submodel for the conditional mean of Qbarbar_M1_a/Qbarbar_M2_star_a given A and C;
since, the inverse probability of treatment weight is a function of A. We also cannot
include the IPTW as part of the loss function since we need one of the weights to be
negative. So, we resort to an iterative approach, where we define a submodel and loss
for the conditional mean of Qbarbar_M1_a and then a loss for the conditional mean of
Qbarbar_M2_star_a. We iterate until the empirical mean of this portion of the
canonical gradient is smaller than tol
.
target_Qbarbar_M1_times_M2_star_a( Qbarbar, Y, A, a, a_star, gn, tol = 1/(sqrt(length(Y)) * log(length(Y))), max_iter = 25, iterative = FALSE, ... )
Qbarbar | Iterated mean estimates |
---|---|
Y | A vector of continuous or binary outcomes. |
A | A vector of binary treatment assignment (assumed to be equal to 0 or 1). |
a | The label for the treatment. The effects estimates returned pertain
to estimation of interventional effects of |
a_star | The label for the treatment. The effects estimates returned pertain
to estimation of interventional effects of |
gn | Power users may wish to pass in their own properly formatted list of the
propensity score so that
nuisance parameters can be fitted outside of |
tol | The tolerance for stopping the iterative targeting procedure. |
max_iter | The maximum number of iterations for the TMLE |