MCPanel+

This repo is forked from MCPanel. The model is described in the paper Matrix Completion Methods for Causal Panel Data Models by Athey et al.

The code in this fork allows for unit-time specific covariates in the form of a N x T matrix C in lieu of the vectors of unit-specific and time-specific covariates in the parent repo.

The code also allows for a N x T matrix W used to weight the loss function, as described in Sec. 8.3 of the Athey et al. paper.

Prerequsites

R >= 3.5.0 (tested on 3.6.1)
Rcpp, evalCpp, glmnet, latex2exp, ggplot2

To install this package in R, run the following commands:

install.packages("devtools")
install.packages("latex2exp")
library(devtools) 
install_github("jvpoulos/MCPanel")

Example usage:

library(MCPanel)

T <- 50 # No. time periods
N <- 50 # No. units

Y <- replicate(T,rnorm(N)) # simulated observed outcomes

X <- replicate(T,rnorm(N)) # simulated covariates

treat_mat <- stag_adapt(M = Y, N_t = (N/2+1), T0= T/2, treat_indices=seq(N/2, N, 1)) # 0s are treated

Y_obs <- Y * treat_mat

# Estimate weights by matrix completion

est_weights <- mcnnm_wc_cv(M = treat_mat, C=X, mask = matrix(1, nrow(treat_mat), ncol(treat_mat)), W = matrix(1, nrow(treat_mat), ncol(treat_mat)), 
	to_estimate_u = 1, to_estimate_v = 1, num_lam_L = 5, num_lam_B = 5, niter = 100, rel_tol = 1e-05, cv_ratio = 0.8, num_folds = 2, is_quiet = 0) # no missing values

W <- plogis(est_weights$L + X%*%replicate(T,as.vector(est_weights$B)) + replicate(T,est_weights$u) + t(replicate(N,est_weights$v)))

weights <- (1-treat_mat) + (treat_mat)*((1-W)/(W))  # weight adjustment (treated are 0)

# Model with covariates
est_model_MCPanel_w <- mcnnm_wc_cv(M = Y_obs, C = X, mask = treat_mat, W = weights, to_normalize = 1, to_estimate_u = 1, to_estimate_v = 1, num_lam_L = 5, num_lam_B = 5, niter = 100, rel_tol = 1e-05, cv_ratio = 0.8, num_folds = 2, is_quiet = 0)

est_model_MCPanel_w$Mhat <- est_model_MCPanel_w$L + X%*%replicate(T,as.vector(est_model_MCPanel_w$B)) + replicate(T,est_model_MCPanel_w$u) + t(replicate(N,est_model_MCPanel_w$v))
est_model_MCPanel_w$msk_err <- (Y-est_model_MCPanel_w$Mhat)*(1-treat_mat)
est_model_MCPanel_w$att<- (1/sum(1-treat_mat)) * sum(est_model_MCPanel_w$msk_err)
est_model_MCPanel_w$att