Getting Started with remoteoutcome

Overview

remoteoutcome estimates average treatment effects (ATEs) when the outcome is observed only in an observational sample but not in the experimental sample. The key idea is to use remotely sensed variables (RSVs) — covariates observed in both samples that are informative about the outcome — to bridge the two samples.

This vignette walks through the full workflow on a simulated binary-outcome dataset:

1. Simulate data with sim_rsv_data()
2. Estimate the ATE with cv.rsv() (cross-fitting)
3. Add standard errors with add_se()
4. Inspect the fit: coefficients, relevance, and the J-test

Setup

library(remoteoutcome)

1. Data structure

The package expects a stacked dataset containing both the experimental and observational samples. Each row is a unit; the two samples are distinguished by indicator columns S_e (experimental) and S_o (observational).

sim_rsv_data() generates data from this design. The true ATE is controlled by tau (default 0.10). Overlap units (S_e = 1 & S_o = 1) are created by setting n_v > 0; they have both Y and D observed and enable direct assumption tests (see vignette("assumption-testing")).

dat <- sim_rsv_data(n_e = 300, n_o = 700, tau = 0.10, seed = 42)
head(dat)

# Outcome is NA for purely experimental units (S_e=1, S_o=0)
table(S_e = dat$S_e, S_o = dat$S_o, Y_observed = !is.na(dat$Y))

2. Fitting the RSV estimator

cv.rsv() uses K-fold cross-fitting: models are fit on held-out folds to avoid overfitting the predictions used in the moment conditions.

The models argument specifies the learner for each nuisance component. "logit" (logistic regression) is fast and works well for small-to-medium samples. For larger datasets or complex relationships, use list(model = "rf", num.trees = 500) to fit a random forest.

R <- as.matrix(dat[, paste0("R", 1:5)])

fit <- cv.rsv(
  Y   = dat$Y,
  D   = dat$D,
  S_e = dat$S_e,
  S_o = dat$S_o,
  R   = R,
  models = list(
    Y   = list(model = "logit"),
    D   = list(model = "logit"),
    S_e = list(model = "logit"),
    S_o = list(model = "logit")
  ),
  nfolds = 5,
  seed   = 42
)

print(fit)

The coefficients field contains the estimated ATE. The true value is 0.10; the RSV estimate should be close.

fit$coefficients

3. Adding standard errors

Standard errors are computed via the influence-function approach by default, which is fast and suitable for most use cases.

fit <- add_se(fit, method = "influence", B = 500, seed = 42)
summary(fit)

score_bootstrap re-draws influence scores rather than refitting the full model, giving a quick bootstrap approximation:

fit <- add_se(fit, method = "score_bootstrap", B = 1000, seed = 42)

For fully non-parametric uncertainty estimates, method = "bootstrap" refits cv.rsv() on each bootstrap sample (slow but most general):

fit <- add_se(fit, method = "bootstrap", B = 200, seed = 42)

4. Relevance

Relevance measures how strongly the RSVs predict treatment-induced variation in the outcome. Higher relevance indicates more informative RSVs and tighter estimates.

relevance(fit)

The table shows the RSV estimator alongside the naive estimator (which ignores the experimental sample structure). The RSV relevance is typically higher when the nuisance models are well-specified.

5. J-test

The J-test checks whether the efficient (RSV) and naive moment conditions agree on the same treatment effect — an indirect test of the identifying assumptions (stability and no direct effect).

jtest(fit)

A large p-value means the two moment conditions are consistent, which supports the assumptions. A small p-value suggests misspecification in the nuisance models or a violation of an assumption.

6. Clustered data

When observations are grouped (e.g., villages, schools), pass cluster IDs to cv.rsv() so the train/test split respects cluster boundaries, and to add_se() for cluster-robust standard errors:

cluster_id <- rep(1:100, each = 10)  # 100 clusters of 10

fit_cl <- cv.rsv(
  Y = dat$Y, D = dat$D, S_e = dat$S_e, S_o = dat$S_o, R = R,
  models  = list(
    Y   = list(model = "logit"),
    D   = list(model = "logit"),
    S_e = list(model = "logit"),
    S_o = list(model = "logit")
  ),
  clusters = cluster_id,
  nfolds   = 5,
  seed     = 42
)

fit_cl <- add_se(fit_cl, method = "bootstrap", B = 200,
                 clusters = cluster_id, seed = 42)

Summary

See vignette("assumption-testing") for how to test the stability and no-direct-effect assumptions when the data allow it.