R/estimate_semipmetric_erf.R

Defines functions estimate_semipmetric_erf

Documented in estimate_semipmetric_erf

#' @title
#' Estimate Semi-exposure-response Function (semi-ERF).
#'
#' @description
#' Estimates the smoothed exposure-response function using a generalized
#' additive model with splines.
#'
#' @param formula a vector of outcome variable in matched set.
#' @param family a description of the error distribution (see ?gam).
#' @param data dataset that formula is build upon.
#' @param ci_appr causal inference approach (matching or weighting).
#'
#' @details
#' This approach uses Generalized Additive Model (gam) using mgcv package.
#'
#' @return
#' returns an object of class gam
#'
#' @export
#'
#' @examples
#' m_d <- generate_syn_data(sample_size = 100)
#' pseudo_pop <- generate_pseudo_pop(m_d$Y,
#'                                   m_d$treat,
#'                                   m_d[c("cf1","cf2","cf3","cf4","cf5","cf6")],
#'                                   ci_appr = "matching",
#'                                   pred_model = "sl",
#'                                   sl_lib = c("m_xgboost"),
#'                                   params = list(xgb_nrounds=c(10,20,30),
#'                                    xgb_eta=c(0.1,0.2,0.3)),
#'                                   nthread = 1,
#'                                   covar_bl_method = "absolute",
#'                                   covar_bl_trs = 0.1,
#'                                   covar_bl_trs_type = "mean",
#'                                   max_attempt = 1,
#'                                   matching_fun = "matching_l1",
#'                                   delta_n = 1,
#'                                   scale = 0.5)
#'
#' outcome_m <- estimate_semipmetric_erf (formula = Y ~ w,
#'                                        family = gaussian,
#'                                        data = pseudo_pop$pseudo_pop,
#'                                        ci_appr = "matching")
#'
#'
estimate_semipmetric_erf <- function(formula, family, data, ci_appr){

  counter_weight <- NULL

  if (ci_appr == "matching"){

    # If the approach is not optimized, the counter will be zero, which causes
    # problem in generating prediction model.
    if (sum(data$counter_weight) == 0) {
      data$counter_weight <- data$counter_weight + 1
      logger::log_debug("Giving equal weight for all samples.")
    }

    suppressWarnings(gam_model <- gam::gam(formula = formula,
                                           family = family,
                                           data = data,
                                           weights = counter_weight))
  } else if (ci_appr == "weighting"){
    suppressWarnings(gam_model <- gam::gam(formula = formula,
                                           family = family,
                                           data = data,
                                           weights = counter_weight))
  } else {
    stop(paste("ci_appr: ", ci_appr, " is not a valid causal inference."))
  }


  if (is.null(gam_model)) {
    stop("gam model is null. Did not converge.")
  }

  return(gam_model)
}
wxwx1993/GPSmatching documentation built on March 1, 2023, 9:32 p.m.