moi: Average Treatment Effect Estimation with Missing Outcome Imputation

Description

Estimates the Average Treatment Effect (ATE) in settings where the outcome may be missing (not observed for all individuals). The treatment effect implied by a parametric imputation model is targeted directly through an efficient one-step estimator constructed from its influence function (Nordland et al., 2026).

Usage

moi(
  data,
  response.model,
  treatment.model,
  missing.model,
  imputation.model,
  imputation.subset = NULL,
  imputation.augmentation = FALSE,
  imputation.augmentation.model = NULL,
  return.all = FALSE,
  nfolds = 1,
  silent = FALSE,
  stratify = FALSE,
  mc.cores = NULL,
  second.order = TRUE
)

Value

An object of class moi.targeted (inheriting from targeted), a list with components:

call: The matched call.
estimate: A lava::estimate object containing the per-arm expected potential outcomes $E[\tilde{Y}|A=a]$ and the ATE contrast $E[\tilde{Y}|A=1] - E[\tilde{Y}|A=0]$, with influence-function-based standard errors. Row labels follow the cate convention: per-arm rows are labeled $E[\tilde{y}(1)]$ and $E[\tilde{y}(0)]$ (or E[tildeY(1)] / E[tildeY(0)] in non-UTF-8 locales), and the contrast row is labeled $E[\tilde{y}(1)]-E[\tilde{y}(0)]$.
levels: Treatment levels (character).
intermediate: (only if return.all = TRUE) Intermediate estimates: $E[\Delta Y|A=a]$, $P(\Delta=0|A=a)$, and $E[U|A=a, \Delta=0]$.

Standard methods (print, summary, coef, vcov, IC) are provided.

Arguments

data: A data.frame containing all variables required by the models. data.table and tbl_df objects are automatically coerced to data.frame.
response.model: A formula or learner object specifying the response/outcome and the associated baseline adjusted model. If a formula is provided, it is automatically wrapped in learner_glm. Used to estimate $E[\Delta Y | A = a]$.
treatment.model: A base R stats formula specifying the binary treatment variable. Only an intercept is allowed on the right-hand side, e.g., A ~ 1.
missing.model: A formula or learner object specifying the model for the probability of the outcome being observed/non-missing (i.e., $P(\Delta = 1 | A = a)$). If a formula is provided, it is wrapped in learner_glm(..., family = binomial()). Used to estimate $P(\Delta = 0 | A = a)$.
imputation.model: A formula or learner_glm object specifying the missing outcome imputation model. If a formula is provided, it is wrapped in learner_glm. Used to estimate $E[U(X, A, Z; \theta) | A = a, \Delta = 0]$.
imputation.subset: Optional character string giving an R expression that evaluates to a logical vector indicating which rows of data to use when fitting the imputation model. The expression is parsed and evaluated in the context of data; for example, imputation.subset = "!is.na(y)" restricts the fit to the observed outcomes. If NULL (default), all rows are used.
imputation.augmentation: Logical. If TRUE, an augmentation term is added to the imputation estimator for improved efficiency. Default is FALSE.
imputation.augmentation.model: A formula, learner, or NULL specifying a working model for the conditional imputation mean $E[U(X, A, Z; \theta) \mid W, A]$, used to augment the imputation estimator. Only used when imputation.augmentation = TRUE; if NULL, the imputation model $U$ itself is used. Default is NULL.
return.all: Logical. If TRUE, the returned object includes all intermediate estimates in addition to the final ATE estimate. Default is FALSE.
nfolds: number of folds (positive integer), or a pre-specified list of fold indices where each element is an integer vector of observation indices forming a partition of 1:nrow(data).
silent: suppress all messages and progressbars
stratify: if TRUE the response.model and missing.model (if provided) will be stratified by treatment
mc.cores: (optional) number of cores. parallel::mcmapply used instead of future
second.order: add seconder order term to IF to handle misspecification of outcome models

Author

Andreas Nordland

Details

The moi function implements an estimator for the Average Treatment Effect where missing outcomes are imputed using a parametric (glm) model.

The function estimate the target parameter

$$E[\tilde{Y}| A = 1] - E[\tilde{Y}| A = 0],$$

where

$$E[\tilde{Y}| A = a] = E[\Delta Y | A=a] + P(\Delta=0 | A=a) \cdot E[U(X, A, Z; \theta) | A=a, \Delta=0],$$

and $\Delta$ denotes the non-missing indicator, and $U$ denotes the imputation model possibly depending on baseline covariates $X$, the treatment $A$, and a post randomization variable $Z$.

Inference in based on the estimated influence functions (IFs) of the associated (covariate adjusted) one-step estimators.

When imputation.augmentation = TRUE, an augmentation term built from the efficient influence function is added, giving an efficient one-step estimator of the treatment effect implied by the imputation model (Nordland et al., 2026). The augmentation uses a working model for the conditional imputation mean $E[U(X, A, Z; \theta) \mid W, A]$ (see imputation.augmentation.model) together with the missingness model.

If no observations are missing in an arm $a$, the imputation contribution for that arm vanishes ($P(\Delta = 0 | A = a) = 0$) and $E[\tilde{Y} | A = a] = E[Y | A = a]$. If no observations are missing in any arm, moi reduces to a standard cate call with cate.model = ~ 1.

References

Nordland, A., Holst, K. K., Redek, D., Pipper, C. B. & Iversen, A. T. (2026) One-step Outcome Imputation: An Alternative to Multiple Imputation. arXiv: https://arxiv.org/abs/2606.07174.

Examples

Run this code

sim_moi <- function(n = 1000, ...) {
  w <- rnorm(n)
  a <- rbinom(n, 1, 0.5)
  y <- 1 + a + w + rnorm(n)
  ## outcome observed (delta = 1) with probability depending on w
  delta <- rbinom(n, 1, lava::expit(1 + w))
  y[delta == 0] <- NA
  data.frame(y, a, w)
}

d <- sim_moi(1000)
## ATE with missing outcomes imputed by a working glm model
moi(data = d,
    response.model = y ~ a + w,
    treatment.model = a ~ 1,
    missing.model = ~ a + w,
    imputation.model = y ~ a + w,
    imputation.subset = "!is.na(y)")

Run the code above in your browser using DataLab