Package {dsFedAte}

Add study membership dummy columns to a server-side data frame

Description

Server-side assign function called once per server before the study-effects arm-specific regression models are fitted via ds.glm(). Called automatically when study_effects = TRUE is passed to dsFedAteClient::ds.EstimateAte(). Appends K-1 explicit 0/1 integer columns named study_id2, ..., study_idK to the server-side data frame.

Usage

addStudyIdDS(base_name, study_index, K_studies)

Arguments

Details

These are dummy columns: column study_idk equals 1 on the server whose study index is k and 0 on every other server. Study 1 is the reference level (all dummies zero).

Because the columns are plain integers, ds.glm() includes them in the model formula without invoking R's factor-contrast machinery. This avoids the "contrasts can be applied only to factors with 2 or more levels" error that arises when a factor column carries only one level locally (every server holds observations from exactly one study).

The column values are entirely determined by the study index supplied by the client — they are constants, not individual-level data — so appending them carries no disclosure risk beyond the existing nfilter.subset check on the dataset size.

Value

The augmented data frame with K-1 integer dummy columns (study_id2, ..., study_idK) appended. The return value is assigned to a server-side symbol by DSI::datashield.assign() on the client.

Privacy control level

• Permitted: permissive, banana.

• Blocked: carrot, avocado, non-permissive.

References

Khellaf, R., Bellet, A., and Josse, J. (2025). Federated Causal Inference: Multi-Study ATE Estimation beyond Meta-Analysis. Proceedings of the 28th International Conference on Artificial Intelligence and Statistics (AISTATS 2025).

See Also

localAteDS, dsFedAteClient::ds.EstimateAte().

Examples

## Not run: 
# Called automatically by dsFedAteClient::ds.EstimateAte(..., study_effects = TRUE).
# Direct invocation (per server, from the client):
cally <- call("addStudyIdDS", "study_data", "2", "3")
DSI::datashield.assign(conns["site2"], "study_data_adj", cally)

## End(Not run)

Compute local statistics for the federated average treatment effect (ATE) and (optionally) its variance

Description

Server-side aggregate function called after the federated GLM has converged. Always returns the quantities needed for the ATE point estimate. When compute_variance = TRUE (the default) it also returns the statistics required for the delta-method variance. Set compute_variance = FALSE to limit disclosure to the minimum needed for the point estimate only.

Usage

localAteDS(
  base_name,
  outcome_name,
  covariate_names_str,
  treatment_name,
  alpha0_str,
  beta0_str,
  alpha1_str,
  beta1_str,
  compute_variance = TRUE,
  heterogeneous = TRUE
)

Arguments

Value

A named list. The following three elements are always present:

n_k

Total sample size on this server.

n_k_1

Number of treated observations on this server.

tau_k

Local average treatment effect estimate \hat\tau_k.

When compute_variance = TRUE, one further element is added:

s_k

Within-site sum of squared deviations of the individual treatment-effect projections from their site mean: \sum_{i \in k}(X_i^\top\hat\gamma - \bar{X}_k^\top\hat\gamma)^2, where \hat\gamma = \hat\beta_1 - \hat\beta_0 is the difference of the arm-specific slope vectors.

When compute_variance = TRUE and heterogeneous = TRUE, two additional elements:

g_raw_k

Column sums of the augmented design matrix (1, X), i.e. the vector

(n_k,\,\sum\nolimits_i X_{i1},\,\ldots,\,\sum\nolimits_i X_{id}).

Accumulated across servers by the client to obtain the global covariate mean used for both arm gradients.

Privacy control level

• Permitted: permissive, banana.

• Blocked: carrot, avocado, non-permissive.

References

Khellaf, R., Bellet, A., and Josse, J. (2025). Federated Causal Inference: Multi-Study ATE Estimation beyond Meta-Analysis. Proceedings of the 28th International Conference on Artificial Intelligence and Statistics (AISTATS 2025).

See Also

addStudyIdDS, dsFedAteClient::ds.EstimateAte().

Examples

## Not run: 
df <- data.frame(
  X1 = rnorm(30), X2 = rnorm(30),
  W  = rbinom(30, 1, 0.5)
)
df$Y <- 1 + 0.5 * df$W + 0.3 * df$X1 + rnorm(30, sd = 0.5)
options(datashield.privacyControlLevel = "banana", nfilter.subset = 3)
localAteDS("df", "Y", "X1,X2", "W", "1.0", "0.3,0.0", "1.5", "0.3,0.0")

## End(Not run)