Univariate Heterogeneous Treatment Effect Modifier Estimator

unihtee() estimates treatment effect modifier variable importance parameters (TEM-VIPs). Both absolute and relative TEM-VIPs can be estimated using one-step or targeted maximum likelihood estimators.

Usage

unihtee(
  data,
  confounders,
  modifiers,
  exposure,
  outcome,
  censoring = NULL,
  time_cutoff = NULL,
  outcome_type = c("continuous", "binary", "time-to-event"),
  effect = c("absolute", "relative"),
  estimator = c("tmle", "onestep"),
  cross_fit = FALSE,
  cross_fit_folds = 5,
  cond_outcome_estimator = sl3::Lrnr_glm_fast$new(),
  prop_score_estimator = sl3::Lrnr_glm_fast$new(),
  prop_score_values = NULL,
  failure_hazard_estimator = sl3::Lrnr_xgboost$new(),
  censoring_hazard_estimator = sl3::Lrnr_xgboost$new(),
  parallel = FALSE
)

Arguments

data: A data.table containing the observed data.
confounders: A character vector of column names corresponding to baseline covariates.
modifiers: A character vector of columns names corresponding to the suspected effect modifiers. This vector must be a subset of confounders.
exposure: A character corresponding to the exposure variable.
outcome: A character corresponding to the outcome variable.
censoring: A character indicating the right censoring indicator variable. Only used with time-to-event outcomes. Defaults to NULL.
time_cutoff: A numeric representing the time point at which to evaluate the time-to-event parameter. Only used with time-to-event outcomes. Defaults to NULL.
outcome_type: A character indicating the outcome type. "continuous", "binary" and "time-to-event" are currently supported.
effect: A character indicating the type of treatment effect modifier variable importance parameter. Currently supports "absolute" and "relative".
estimator: A character set to either "tmle" or "onestep". The former results in unihtee() to use a targeted maximum likelihood estimators to estimate the desired TEM-VIP, while the latter uses a one step estimator.
cross_fit: A logical determining whether cross-fitting should be used. Defaults to FALSE.
cross_fit_folds: A numeric stating the number of folds to use in the cross-fitting procedure. Defaults to 5.
cond_outcome_estimator: A Stack, or other learner class (inheriting from Lrnr_base), containing a set of learners from sl3 to estimate the conditional outcome. Defaults to a generalized linear model with one- and two- way interactions among all confounders and exposure variables. Only used with continuous and binary outcomes.
prop_score_estimator: A Stack, or other learner class (inheriting from Lrnr_base), containing a set of learners from sl3 to estimate the propensity score. Defaults to a generalized linear model with one- and two- way interactions among all confounders variables.
prop_score_values: An optional character corresponding to the (known) propensity score values for each observation in data. Defaults to NULL.
failure_hazard_estimator: A Stack, or other learner class (inheriting from Lrnr_base), containing a set of learners from sl3 to estimate the conditional failure hazard function. Defaults to an XGBoost learner with confounders and exposure variables as covariates. Only used with time-to-event outcomes.
censoring_hazard_estimator: A Stack, or other learner class (inheriting from Lrnr_base), containing a set of learners from sl3 to estimate the conditional censoring hazard function. Defaults to an XGBoost learner with confounders and exposure variables as covariates. Only used with time-to-event outcomes.
parallel: A logical stating if origami's built-in parallelized cross-validation routines should be used when cross_fit = TRUE. The future suite is used. Defaults to FALSE.

Value

A data.table containing the effect estimates and (adjusted) p-values of the modifiers. The suspected treatment effect modifiers ordered according to ascending p-values.