Using the best_linear_projection to get coefficient estimates for treatment and control separately

[aczura2024]

I am fitting a causal forest with 2 treatment arms and want to use the best_linear_projection() function to estimate the regression coefficients on the treatment and control groups separately. When I make a separate matrix corresponding to the observations solely in the treatment group, I get an error Error in best_linear_projection(mycrf, A = treat_covar[, 2:ncol(treat_covar)]) :
The number of rows of A does not match the number of training examples.

Any ideas?

[erikcs]

That error message tells you that the number of rows in treat_covar is different from the number of samples used to train mycrf. BLP fits a linear model to $\tau(X)$. If you want to fit a linear model to either a) $E[Y_i(1)|X_i]$ or b) $E[Y_i(0)|X_i]$ you could construct doubly robust scores for those and run a linear model similarily to how BLP does it under the hood. policytree does the DR score calculation for a and b here