Skip to contents

Aggregate estimands across treatment groups

Source: R/aggregate_estimands.R
aggregate_estimands.Rd

Takes the stacked output of multiple_treatment_group_analysis() and computes three aggregate estimands across treatment groups for each event time:

Usage

aggregate_estimands(
  results,
  weights = "sample",
  methods = c("DID_Female", "DID_Male", "TD", "NTD_Conv", "NTD_New"),
  include_pre = FALSE
)

Arguments

results

A data.frame as returned by multiple_treatment_group_analysis(), with at minimum the columns d, event_time, estimand, method, est, and se. If results carries influence-function data (attached automatically by multiple_treatment_group_analysis()), standard errors account for shared control groups across treatment groups.

weights

How to weight treatment groups. One of:

  • "sample" (default): use sample-proportion weights as in Leventer (2025). Within-gender weights \(w_{g,d} = n_{g,d}/\sum n_{g,\tilde d}\) are used for DID_Female and DID_Male; cross-gender weights \(w_d = n_d / \sum n_{\tilde d}\) for TD, NTD_Conv, and NTD_New. Standard errors account for estimation of the weights.

  • NULL: uniform weights (equal weight per group).

  • A named numeric vector: custom fixed weights (names = treatment groups as characters). Values are renormalised to sum to 1.

methods

Character vector of methods to aggregate. Defaults to all five main methods.

include_pre

Logical. If TRUE, also aggregate pre-treatment event times (event_time < 0). Default FALSE.

Value

A data.frame with one row per event_time by estimand by method by agg_type combination, containing:

  • event_time – event time

  • estimand – "APO", "ATE", "theta", or "Delta_rho"

  • method – method name

  • agg_type – one of "avg_of_ratios", "ratio_of_avgs", "gender_ineq"

  • est – aggregate estimate

  • se – standard error (see Details)

  • ci_l, ci_h – 95 \

  • n_groups – number of treatment groups contributing

Details

avg_of_ratios (\(\theta_{\text{Agg},1}\))

Weighted average of the group-specific normalised effects \(\theta(g, d, d+e)\) across treatment groups \(d\). This is the preferred estimand because it averages effects that are already scaled by each group's baseline.

ratio_of_avgs (\(\theta_{\text{Agg},2}\))

Ratio of the weighted-average ATE to the weighted-average APO. The implicit weight on each group is \(p_d \cdot \text{APO}_d\), giving higher-earning groups more influence.

gender_ineq (\(\Delta\rho_{\text{Agg}}\))

Weighted average of NTD_New (estimand == "Delta_rho") across treatment groups – the aggregate gender-inequality estimand.

Standard errors. When the results object carries influence-function (IF) data from multiple_treatment_group_analysis(), aggregate SEs account for dependence across treatment groups caused by shared control individuals.

With weights = "sample", the IF additionally accounts for estimation of the weights, following the formula in Leventer (2025, Appendix G): $$\psi_{A(e)} = \sum_d \left[ w_d\,\psi_{B_d} + \frac{B_d - A(e)}{M}\,\psi_{p_d} \right]$$ where \(M = \sum_d p_d\) and \(\psi_{p_d}\) is the IF of the group proportion.

With fixed weights (NULL or a named vector), the second term drops out and the IF reduces to \(\sum_d w_d\,\psi_{B_d}\).

For ratio_of_avgs, the delta method is applied to the ratio \(\bar\mu_{\text{ATE}} / \bar\mu_{\text{APO}}\) using the aggregate IFs for the numerator and denominator.

If IF data is not available (e.g., when the user supplies a manually constructed results table), SEs are computed under an independence approximation with a warning.

Handling missing cells. Not every treatment group produces an estimate for every event time (due to max_age / min_age bounds). The function operates on whichever groups are present for each cell and reports how many via n_groups. If weights is supplied as a named vector, only the entries whose names appear in the observed treatment groups are used; the remaining weights are dropped and the retained weights are renormalised.

Examples

# \donttest{
set.seed(1)
sim <- simulate_data(n_individuals = 500)
res <- multiple_treatment_group_analysis(sim, treatment_groups = 24:25,
                                         periods_post = 2, verbose = FALSE)
agg <- aggregate_estimands(res)
head(agg)
#>   event_time estimand     method      agg_type         est         se n_groups
#> 1          0    theta DID_Female avg_of_ratios -0.29648428 0.02688094        2
#> 2          1    theta DID_Female avg_of_ratios -0.27864063 0.02450991        2
#> 3          2    theta DID_Female avg_of_ratios -0.26867246 0.02001048        2
#> 4          0    theta   DID_Male avg_of_ratios -0.02789507 0.03252366        2
#> 5          1    theta   DID_Male avg_of_ratios -0.02730183 0.03285187        2
#> 6          2    theta   DID_Male avg_of_ratios -0.07414387 0.03154374        2
#>          ci_l        ci_h
#> 1 -0.34917093 -0.24379764
#> 2 -0.32668006 -0.23060119
#> 3 -0.30789301 -0.22945192
#> 4 -0.09164144  0.03585129
#> 5 -0.09169150  0.03708783
#> 6 -0.13596961 -0.01231814
# }