Skip to contents

Doubly-robust causal inference with missing data

Source: R/drcmd.R
drcmd.Rd

Doubly-robust estimation of counterfactual means in the presence of missing data. The drcmd() function estimates counterfactual means for binary point treatments and reports average treatment effects, as well as causal risk ratios and odds ratios for binary outcomes. General missingness patterns in the data are allowed and automatically determined by the function – the only requirement is that any missingness occurs at random conditional on variables that are always available. For scenarios where non-missingness probabilities are known, as is common in two-phase sampling designs, users can provide the non-missingness probabilities through the Rprobs argument. Users can fit nuisance functions through either highly-adaptive LASSO (HAL) or SuperLearner, the latter of which the user must specify libraries.

Usage

drcmd(
  Y,
  A,
  X,
  W = NA,
  R = NA,
  default_learners = NULL,
  m_learners = NULL,
  g_learners = NULL,
  r_learners = NULL,
  po_learners = NULL,
  eem_ind = FALSE,
  tml = FALSE,
  Rprobs = NA,
  k = 1,
  cutoff = 0.025,
  quiet = TRUE,
  cv_folds = 5,
  parallel = FALSE,
  att = FALSE,
  atc = FALSE
)

Arguments

Y

Outcome variable. Can be continuous or binary

A

A binary treatment variable (1=treated, 0=control)

X

Dataframe containing baseline covariates

W

(optional) Dataframe containing variables solely predictive of missingness, but not a cause of the outcome or exposure.

R

(optional) A character string specifying the missingness indicator, where 0 indicates missing data. If not specified, the function will create the missingness indicator by identifying the missingness pattern in the data

default_learners

A character vector containing SuperLearner libraries to use for estimating all nuisance functions. User can alternatively specify libraries for each nuisance function for added flexibility

m_learners

A character vector containing learners to be used for the outcome regression. A vector of SuperLearner library names

g_learners

A character vector containing learners to be used for the propensity score. A vector of SuperLearner library names

r_learners

A character vector containing learners to be used for the missingness indicator regression. A vector of SuperLearner library names

po_learners

A character vector containing learners to be used for the pseudo-outcome regression. A vector of SuperLearner library names

eem_ind

A logical indicating whether to use empirical efficiency maximization

tml

A logical indicating whether to obtain estimates through targeted maximum likelihood estimation (TML). If TRUE, estimates obtained by TML. If FALSE (default setting), estimates are obtained through a one-step estimator

Rprobs

A vector of probabilities for the missingness indicator. Only suitable for study designs where researcher controls mechanism by which variables are missing (e.g. two-phase sample designs). Defaults to NA, in which case missingness probabilities are estimated.

k

A numeric indicating the number of folds for cross-fitting

cutoff

Cutoff for treatment and complete case propensity scores. Estimates outside of [cutoff, 1-cutoff] are set to cutoff or 1-cutoff, respectively

quiet

Logical indicating whether to suppress progress messages. Default is TRUE. Set to FALSE to see which nuisance function is being estimated.

cv_folds

Number of cross-validation folds used internally by SuperLearner for model selection. Default is 5. Lower values speed up estimation.

parallel

Logical indicating whether to run cross-fitting folds in parallel using parallel::mclapply. Only used when k > 1. Note: not supported on Windows.

att

Logical indicating whether to additionally estimate the average treatment effect on the treated (ATT). Default is FALSE.

atc

Logical indicating whether to additionally estimate the average treatment effect on the controls (ATC). Default is FALSE.

Value

An S3 object of class "drcmd" containing estimation results, information on the missing data structure, and parameters used in the estimation

params

A list containing parameters used in the estimation

Z

A character vector containing always-available variables

R

A character vector containing the complete case indicator values

U

A character vector containing partially-missing variables

results

A list of dataframes storing (i) point estimates, (ii) standard errors, and (iii) nuisance function estimates

Examples

set.seed(1)
n <- 200
X <- rnorm(n)
A <- rbinom(n, 1, plogis(X))
Y <- rnorm(n) + A + X
R_ind <- rbinom(n, 1, plogis(X))
Y[R_ind == 0] <- NA
covariates <- data.frame(X = X)

fit <- drcmd(Y, A, covariates,
             default_learners = "SL.glm",
             k = 1)
#> Loading required package: nnls
fit
#> drcmd results
#> -------------------------------------------------
#> ATE estimate:  0.9474179 
#> -------------------------------------------------
#> Variables with missingness (U):  Y 
#> -------------------------------------------------
#> Variables without missingness (Z):  X A 
#> -------------------------------------------------
#> Validity of results requires causal assumptions to hold
#> As well as the assumption that U is independent of R given Z
#> Number of cross-fitting folds (k): 1