Skip to contents

Treatment-by-moderator model with LASSO-selected covariates

Source: R/fit_lasso.R
lm_moderator_lasso.Rd

Fits outcome ~ treatment * moderator + selected covariates, where the covariates are chosen by lasso_select_covariates from the focal outcome ~ treatment equation.

Usage

lm_moderator_lasso(
  formula,
  moderator,
  data,
  covariates = NULL,
  clusters = NULL,
  ci = TRUE,
  alpha = 0.05,
  lasso_args = list()
)

Arguments

formula

A two-sided formula outcome ~ treatment, the focal estimand equation passed to selection. The moderator is supplied separately.

moderator

Character scalar naming the moderator column.

data

A data frame.

covariates

Candidate covariates: a one-sided formula, a character vector of column names, or a bare expression.

clusters, ci, alpha

Passed through to lm_robust.

lasso_args

A named list of further arguments for lasso_select_covariates.

Value

An lm_robust object, carrying the attributes described in adjustment. When the interacted fit with selected covariates cannot be produced, the fallback is the same interaction model fitted on the full candidate covariate set, not an unadjusted model.

Details

The selected covariates enter additively and are deliberately not interacted with treatment or with the moderator. Adding those interactions would spend degrees of freedom on terms that are not the estimand and, with a data-driven covariate set, would make the reported interaction sensitive to which covariates happened to be selected.

The treatment-by-moderator coefficient describes how the treatment effect varies across levels of an observed variable. The moderator is not randomly assigned, so that variation is descriptive: it is not the causal effect of the moderator, and it does not decompose the treatment effect into a pathway.

See also

Other adjusted estimators: lm_lin_lasso(), lm_robust_lasso()

Examples

set.seed(1)
n <- 600
dat <- data.frame(
  Z = rep(0:1, n / 2),
  X_pid = rep(c(0, 1), each = n / 2),
  X1 = rnorm(n), X2 = rnorm(n)
)
dat$Y <- 0.2 * dat$Z + 0.6 * dat$Z * dat$X_pid + 1.5 * dat$X1 + rnorm(n)

lm_moderator_lasso(Y ~ Z, moderator = "X_pid", data = dat, covariates = ~ X1 + X2)
#>                Estimate Std. Error    t value      Pr(>|t|)    CI Lower
#> (Intercept)  0.03729435 0.08620065  0.4326459  6.654288e-01 -0.13200019
#> Z            0.22121165 0.11711298  1.8888738  5.939482e-02 -0.00879345
#> X_pid       -0.21173121 0.12763984 -1.6588176  9.767956e-02 -0.46241063
#> X1           1.53575730 0.04090927 37.5405720 4.741440e-159  1.45541318
#> Z:X_pid      0.79325275 0.17013819  4.6624026  3.861417e-06  0.45910832
#>              CI Upper  DF
#> (Intercept) 0.2065889 595
#> Z           0.4512167 595
#> X_pid       0.0389482 595
#> X1          1.6161014 595
#> Z:X_pid     1.1273972 595