Prepare data for the sequence of emulated target trials — data

Usage

data_preparation(
  data,
  id = "id",
  period = "period",
  treatment = "treatment",
  outcome = "outcome",
  eligible = "eligible",
  model_var = NULL,
  outcome_cov = ~1,
  estimand_type = c("ITT", "PP", "As-Treated"),
  switch_n_cov = ~1,
  switch_d_cov = ~1,
  first_period = NA,
  last_period = NA,
  use_censor_weights = FALSE,
  cense = NA,
  pool_cense = c("none", "both", "numerator"),
  cense_d_cov = ~1,
  cense_n_cov = ~1,
  eligible_wts_0 = NA,
  eligible_wts_1 = NA,
  where_var = NULL,
  data_dir,
  save_weight_models = FALSE,
  glm_function = "glm",
  chunk_size = 500,
  separate_files = FALSE,
  quiet = FALSE,
  ...
)

Arguments

data: A data.frame containing all the required variables in the person-time format, i.e., the `long' format.
id: Name of the variable for identifiers of the individuals. Default is `id'.
period: Name of the variable for the visit/period. Default is `period'.
treatment: Name of the variable for the treatment indicator at that visit/period. Default is `treatment'.
outcome: Name of the variable for the indicator of the outcome event at that visit/period. Default is `outcome'.
eligible: Name of the variable for the indicator of eligibility for the target trial at that visit/period. Default is `eligible'.
model_var: Treatment variables to be included in the marginal structural model for the emulated trials. model_var = "assigned_treatment" will create a variable assigned_treatment that is the assigned treatment at the trial baseline, typically used for ITT and per-protocol analyses. model_var = "dose" will create a variable dose that is the cumulative number of treatments received since the trial baseline, typically used in as-treated analyses.
outcome_cov: A RHS formula with baseline covariates to be adjusted for in the marginal structural model for the emulated trials. Note that if a time-varying covariate is specified in outcome_cov, only its value at each of the trial baselines will be included in the expanded data.
estimand_type: Specify the estimand for the causal analyses in the sequence of emulated trials. estimand_type = "ITT" will perform intention-to-treat analyses, where treatment switching after trial baselines are ignored. estimand_type = "PP" will perform per-protocol analyses, where individuals' follow-ups are artificially censored and inverse probability of treatment weighting is applied. estimand_type = "As-Treated" will fit a standard marginal structural model for all possible treatment sequences, where individuals' follow-ups are not artificially censored but treatment switching after trial baselines are accounted for by applying inverse probability of treatment weighting.
switch_n_cov: A RHS formula to specify the logistic models for estimating the numerator terms of the inverse probability of treatment weights. A derived variable named time_on_regime containing the duration of time that the individual has been on the current treatment/non-treatment is available for use in these models.
switch_d_cov: A RHS formula to specify the logistic models for estimating the denominator terms of the inverse probability of treatment weights.
first_period: First time period to be set as trial baseline to start expanding the data.
last_period: Last time period to be set as trial baseline to start expanding the data.
use_censor_weights: Require the inverse probability of censoring weights. If use_censor_weights = TRUE, then the variable name of the censoring indicator needs to be provided in the argument cense.
cense: Variable name for the censoring indicator. Required if use_censor_weights = TRUE.
pool_cense: Fit pooled or separate censoring models for those treated and those untreated at the immediately previous visit. Pooling can be specified for the models for the numerator and denominator terms of the inverse probability of censoring weights. One of "none", "numerator", or "both" (default is "none" except when estimand_type = "ITT" then default is "numerator").
cense_d_cov: A RHS formula to specify the logistic models for estimating the denominator terms of the inverse probability of censoring weights.
cense_n_cov: A RHS formula to specify the logistic models for estimating the numerator terms of the inverse probability of censoring weights.
eligible_wts_0: See definition for eligible_wts_1
eligible_wts_1: Exclude some observations when fitting the models for the inverse probability of treatment weights. For example, if it is assumed that an individual will stay on treatment for at least 2 visits, the first 2 visits after treatment initiation by definition have a probability of staying on the treatment of 1.0 and should thus be excluded from the weight models for those who are on treatment at the immediately previous visit. Users can define a variable that indicates that these 2 observations are ineligible for the weight model for those who are on treatment at the immediately previous visit and add the variable name in the argument eligible_wts_1. Similar definitions are applied to eligible_wts_0 for excluding observations when fitting the models for the inverse probability of treatment weights for those who are not on treatment at the immediately previous visit.
where_var: Specify the variable names that will be used to define subgroup conditions when fitting the marginal structural model for a subgroup of individuals. Need to specify jointly with the argument where_case.
data_dir: Directory to save model objects when save_weight_models=TRUE and expanded data as separate CSV files names as trial_i.csvs if separate_files = TRUE. If the specified directory does not exist it will be created. If the directory already contains trial files, an error will occur, other files may be overwritten.
save_weight_models: Save model objects for estimating the weights in data_dir.
glm_function: Specify which glm function to use for the marginal structural model from the stats or parglm packages. The default function is the glm function in the stats package. Users can also specify glm_function = "parglm" such that the parglm function in the parglm package can be used for fitting generalized linear models in parallel. The default control setting for parglm is nthreads = 4 and method = "FAST", where four cores and Fisher information are used for faster computation. Users can change the default control setting by passing the arguments nthreads and method in the parglm.control function of the parglm package, or alternatively, by passing a control argument with a list produced by parglm.control(nthreads = , method = ).
chunk_size: Number of individuals whose data to be processed in one chunk when separate_files = TRUE
separate_files: Save expanded data in separate CSV files for each trial.
quiet: Suppress the printing of progress messages and summaries of the fitted models.
...: Additional arguments passed to glm_function. This may be used to specify initial values of parameters or arguments to control. See stats::glm, parglm::parglm and parglm::parglm.control() for more information.

Value

An object of class TE_data_prep, which can either be sampled from (case_control_sampling_trials) or directly used in a model (trial_msm). It contains the elements

data: the expanded dataset for all emulated trials. If separate_files = FALSE, it is a data.table; if separate_files = TRUE, it is a character vector with the file path of the expanded data as CSV files.
min_period: index for the first trial in the expanded data
max_period: index for the last trial in the expanded data
N: the total number of observations in the expanded data
data_template: a zero-row data.frame with the columns and attributes of the expanded data
switch_models: a list of summaries of the models fitted for inverse probability of treatment weights, if estimand_type is "PP" or "As-Treated"
censor_models: a list of summaries of the models fitted for inverse probability of censoring weights, if use_censor_weights=TRUE
args: a list contain the parameters used to prepare the data and fit the weight models

Details

This function expands observational data in the person-time format (i.e., the `long' format) to emulate a sequence of target trials and also estimates the inverse probability of treatment and censoring weights as required.

The arguments chunk_size and separate_files allow for processing of large datasets that would not fit in memory once expanded. When separate_files = TRUE, the input data are processed in chunks of individuals and saved into separate files for each emulated trial. These separate files can be sampled by case-control sampling to create a reduced dataset for the modelling.