Post-double selection estimator for interactions

Source: R/lasso_interactions.R

post_ds_interaction applies post-double selection to the estimation of an interaction in a linear model.

post_ds_interaction(
  data,
  treat,
  moderator,
  outcome,
  control_vars,
  panel_vars = NULL,
  moderator_marg = TRUE,
  cluster = NULL,
  method = "double selection"
)

Arguments

data

data.frame to find the relevant variables.

treat

string with the name of the treatment variable.

moderator

string with the name of the moderating variable.

outcome

string with the name of the outcome variable.

control_vars

vector of strings with the names of the control variables to include.

panel_vars

vector of strings with the names of categorical variables to include as fixed effects.

moderator_marg

logical indicating if the lower-order term of the moderator should be included ()

cluster

string with the name of the cluster variable.

method

string indicating which method to use. The default is "double selection" selects variables based on the outcome and treatment/interaction variables and "single selection" only selects on the outcome.

Value

Returns an object of the class lm with an additional clustervcv object containing the cluster-robust variance matrix estimate when cluster is provided.

Details

The post_ds_interaction implements the post-double selection estimator of Belloni et al (2014) as applied to interactions, which was proposed by Blackwell and Olson (2019). Variables passed to panel_vars are considered factors for fixed effects and whose "base effects" are removed by demeaning all variables by those factors. Interactions between the moderator and all variables (including the factors generated by panel_vars) are generated and passed to the post-double selection procedure. Base terms for the treatment, moderator, and control variables are forced to be included in the final post-double selection OLS. The cluster argument adjusts the lasso

References

Alexandre Belloni, Victor Chernozhukov, Christian Hansen, Inference on Treatment Effects after Selection among High-Dimensional Controls, The Review of Economic Studies, Volume 81, Issue 2, April 2014, Pages 608-650, doi:10.1093/restud/rdt044

Matthew Blackwell and Michael Olson.. "Reducing Model Misspectation and Bias in the Estimation of Interactions." Political Analysis, 2021.

Examples

data(remit)

controls <- c("l1gdp", "l1pop", "l1nbr5", "l12gr", "l1migr",
"elec3")

post_ds_out <- post_ds_interaction(
  data = remit, treat = "remit",
  moderator = "dict", outcome = "Protest",
  control_vars = controls,
  cluster = "caseid"
)