OneStepEstimator: OneStepEstimator

In frbl/OnlineSuperLearner: Online SuperLearner package

Description

The one step estimator (OOS) is a technique that improves our initial estimates of the parameter of interest and targets them towards this parameter of interest. In order to use the OOS, one has to solve the Efficient Influence Curve equation, which can be done using a Monte-Carlo approximation. The process of Monte-Carlo approximating the efficient influence curve consists of two main steps: (i) compute the so-called 'h-ratios', and (ii) compute a number of conditional expectations. This procedure is implemented in this class and can be started by calling the perform method of an instance of this class.

Usage

OneStepEstimator

Format

An object of class R6ClassGenerator of length 24.

Methods

initialize(osl, relevantVariables, N, B, pre_processor, tau, intervention, variable_of_interest, discrete = TRUE, parallel= TRUE, online = FALSE, verbose = FALSE, minimal_measurements_needed = 1)

Initializes the online one step estimator. It uses an earlier fitted online super learner to sample from the conditional densities.

@param osl the online superlearner, which was fitted earlier on the data

@param relevantVariables a list of relevant variables used for fitting the OSL

@param N the number of measurements in a timeseries

@param B the number of iterations we should do while sampling from the conditional expectations

@param pre_processor the PreProcessor object used to normalize the data.

@param tau the time at which we want to measure the effect of an intervention.

@param intervention the intervention we want to perform. See InterventionParser for more details

@param variable_of_interest the variable we are interested in (e.g., the Y relevant variable)

@param discrete (default = TRUE) whether we should use the discrete (true) or continuous (false) super learner

@param parallel (default = TRUE) should the estimation run in parallel?

@param online (default = TRUE) should the estimation run online?

@param verbose (default = FALSE) the verbosity of the etimation process

@param minimal_measurements_needed (default = 1) the minimal number of measurements needed to sample a completely new block. I.e., that all lags are renewed.

perform(initial_estimate, data, truth = NULL

This method actually runs the oos. Based on an initial estimate, it calculates an update term to add to this estimate. This will make the estimator well behaved (i.e., normally distributed). The function will add this correction term to the initial estimate and return the estimated variance of the estimator.

@param initial_estimate the initial estimate of the target parameter, as calculated using OSL

@param data the data to seed the sampling procedure

@param truth double (default = NULL) the true value of the parameter of interest. This can only be specified when running a simulation study and can be useful for logging purposes.

@return a list containing two elements: oos_estimate and oos_variance. This first element (oos_estimate) contains the updated estimate of the target parameter. The second element (oos_variance), contains the variance of this estimator, which can be used to derive confidence intervals.

calculate_full_oos(initial_estimate, data, truth = NULL)

Calculates the one step estimation based on the data and the initial estimate. This is the function called by the perform method, see that function for more details.

@param initial_estimate double the initial estimate that should be improved.

@param data the data to seed the sampling procedure

@param truth double (default = NULL) the true value of the parameter of interest. This can only be specified when running a simulation study and can be useful for logging purposes.

calculate_oos_variance()

NOTE! NOT YET IMPLEMENTED Calculates the variance of the estimator. This can be used to determine the confidence bands of an estimand.

get_h_ratio_estimators(data, last_h_ratio_estimators = NULL)

Calculates a list of h-ratio estimators, one for each relevantvariable. This function calls the calculate_h_ratio_predictors and returns the result (== actual estimators) in a list.

@param data the data used to seed the sampling procedure.

@param last_h_ratio_estimators = NULL currently unused, could be used the reuse (and update) the previous set of h-ratio estimators.

@return a list of H-ratio estimators. This list is 2 dimentional. THe outer list is a list per time s. The inner list is a list with an estimator for each relevantVariable.

evaluation_of_conditional_expectations(data, h_ratio_predictors

In this function one can perform the second step of OOS, calculate the conditional expectations / difference in conditional expectations.

@param data the data to seed the sampling procedure and thereby conditional expectation evaluation.

@param h_ratio_predictors the list of predictors, in a format returned using the get_h_ratio_estimators format.

calculate_h_ratio_predictors(Osample_p, Osample_p_star)

This method can be used to perform the first step of OOS. It can be used to retrieve a list of h-ratio estimators. This method returns an estimator for each W,A,Y, and for each time s from 1 to tau. This method uses the more efficient way, as described in Van der Laan 2017.

@param Osample_p the observations sampled from the normal (not intervened) distribution.

@param Osample_p the observations sampled from the intervened distribution.

@return a list of H-ratio estimators. This list is 2 dimentional. THe outer list is a list per time s. The inner list is a list with an estimator for each relevantVariable.

calculate_h_ratio(h_ratio_predictors, s, formula, data)

Calculates a specific h-ratio for a specific formula (== the variable of interest and the needed covariates) and a specific time s. The result of this function is an estimated h-ratio

@param h_ratio_predictors list a list of h-ratio predictors (this should be the same format as the one exported by the get_h_ratio_estimators).

@param s the time at which one wants to calculate the h-ratio

@param formula the formula of the h-ratio estimator to use for calculating the estimate.

@param data the data to do the prediction with

@return the actual estimated h-ratio value.

calculate_difference_in_expectations(s, dat, formula, current_rvs)

The efficient influence curve equation consists of two equations. One is the h-ratio calculation, the other is the difference between two expectations. In this function we calculate the difference in expectations for a specified relevant variable.

@param s the time at which the difference in expectations needs to be calculated.

@param dat data.table the initial data to use for estimating the difference in expectations

@param current_rvs the current list of relevant variables

@return the difference in expectations

get_next_and_current_rv(current_rv_index)

<<description>> @param current_rv_index

get_data_cache(star = FALSE)

<<description>> @param star = FALSE

frbl/OnlineSuperLearner documentation built on Feb. 9, 2020, 9:28 p.m.