[BUG] y_hat_test specification under adaptive coding

[MaggieWang0614]

Description

Language: R
Model: BCF
Question: Is the y_hat_test correctly specified under adaptive coding?

Bug

stochtree/R/bcf.R

Line 2534 in 4f1c3e8

y_hat_test <- mu_hat_test + treatment_term_test

Based on this line, we would get
y(1) = mu(x) + tau_raw(x) * (b1 - b0)
y(0) = mu(x)

Expected behavior

With adaptive coding, we should expect
y(1) = mu(x) + tau_raw(x) * b1
y(0) = mu(x) + tau_raw(x) * b0

[andrewherren]

Thank you for bringing this up! I just checked and it looks like stochtree is currently missing a component of the control prediction, although the treatment prediction is correct. When we fit BCF with adaptive coding, we are modeling

$$ y_i = \mu(x_i) + \tau(x_i) [b_1 z_i + b_0 (1 - z_i)] + \epsilon_i $$

so that the control and treated conditional expectations are

$$ E[y_i \mid x_i, z_i = 0] = \mu(x_i) + \tau(x_i) b_0 $$

and

$$ E[y_i \mid x_i, z_i = 1] = \mu(x_i) + \tau(x_i) (b_1 - b_0) $$

I will fix that this evening, so that it's reflected in both the cached predictions of the bcf() function (the y_hat_test you linked to) and the results of the predict.bcfmodel() function.

[andrewherren]

Closed as this has now been address in PR #232