Timing-group summary

One row per timing group (sub-experiment) with masses, precision vs corrective weights, and cohort ATTs — plus the two aggregation identities that connect the cohort ATTs to the stacked TWFE estimates.

R stack_summary()   ↔︎   Stata stacked summary

Usage

stack_summary(
  stack_data, outcome_var,
  cluster_var = NULL, ref_period = -1L, covariates = NULL,
  model = c("event_study", "att"),
  fe = c("interacted", "saturated"),
  unit_var = NULL, time_var = NULL,
  weight_type = NULL, pop_var = NULL, weight_var = NULL,
  conf_level = 0.95, digits = 4
)

Arguments

Argument Description
stack_data, outcome_var Stack from build_stack() and outcome name.
cluster_var, ref_period, covariates As in stackreg().
model "event_study" (default; average post-period ATT per cohort) or "att".
fe "interacted" (default here — the precision identity is exact only for the canonical TWFE) or "saturated".
unit_var, time_var Read from stacked_meta when NULL; unit_var needed to count treated units.
weight_type, pop_var, weight_var Override the mass basis; default read from stacked_meta.
conf_level Confidence level for the delta-method intervals (default 0.95).
digits Print digits (default 4). 
stacked summary depvar [, cluster(varname) ref(#) model(str) fe(str) ///
    covariates(varlist) unitvar(varname) timevar(varname) ///
    weighttype(str) popvar(varname) weightvar(varname) level(#)]

Options

Option Description
cluster(), ref(), covariates(), level() As in stacked reg.
model(str) eventstudy (default) or att.
fe(str) interacted (default here) or saturated.
weighttype(), popvar(), weightvar() Override the mass basis; default read from what stacked build stamped.

R ↔︎ Stata mapping

R Stata
outcome_var depvar
cluster_var cluster()
ref_period ref()
model / fe model() / fe()
weight_type / pop_var / weight_var weighttype() / popvar() / weightvar()
conf_level level()
digits (none; Stata uses default formatting)

Value / Stored results

R returns a stack_summary object (list) with table (per-cohort rows plus a Total row), D, d_gap, precision and corrective (each the weighted-sum aggregate and its matching TWFE coefficient), and weight_basis. The print method renders the table and the identities.

Stata returns r(summary) (matrix) and scalars r(D), r(d_gap), r(theta_precision), r(theta_corrective), r(twfe_precision), r(twfe_corrective), r(att), r(att_se).

Both print, per cohort: N obs, obs share, treated units, treated-obs share, the precision weight \(w_a^S\) (what an unweighted stacked TWFE implicitly places on the cohort), the corrective Q-weight \(w_a^T\) (the treated-mass share the design targets), their gap, and the cohort ATT/SE/p-value. Beneath: \(D = \sum_a |w_a^S - w_a^T|\), and the two identities — precision-weighted sum of ATTs = unweighted stacked TWFE; corrective-weighted sum = Q-weighted stacked TWFE (the target ATT). This combines what stacked kappa, screen and stacked reg, bygroup report separately.

Example

library(stacked)
data(medicaid)
stack <- build_stack(medicaid, "year", "state", "adopt_year",
                     kappa_pre = 3, kappa_post = 2)

stack_summary(stack, "uninsured", cluster_var = "state")
Stacked DiD summary by timing group (event study, fe = interacted)
Outcome: uninsured   |   weight basis: unit counts

 group N obs obs shr N trt trt shr precis wt correc wt wS - wT     ATT     SE
  2014   276  0.4600    28  0.8000    0.6437    0.8000 -0.1563 -0.0216 0.0065
  2015   126  0.2100     3  0.0857    0.1511    0.0857  0.0654 -0.0159 0.0052
  2016   120  0.2000     2  0.0571    0.1058    0.0571  0.0486 -0.0421 0.0041
  2019    78  0.1300     2  0.0571    0.0994    0.0571  0.0423 -0.0152 0.0039
 Total   600  1.0000    35  1.0000    1.0000    1.0000 -0.0000 -0.0219 0.0054
  p-val
 0.0009
 0.0022
 0.0000
 0.0001
 0.0001

Design divergence (from adoption dates alone; N = unit counts mass):
  S_a = (N_a^D * N_a^C) / (N_a^D + N_a^C)      [precision mass of sub-experiment a]
  precision  weight  w_a^S = S_a / sum_b S_b
  corrective weight  w_a^T = N_a^D / sum_b N_b^D
  D = sum_a |w_a^S - w_a^T| = 0.3125
  D-gap (largest single-cohort gap): cohort 2014, |w^S - w^T| = 0.1563

Precision-weighted aggregate (what the UNWEIGHTED stacked TWFE reports):
  theta_S = sum_a w_a^S * ATT_a           = -0.0222
  unweighted stacked TWFE coefficient      = -0.0222   [match]

Corrective-weighted aggregate (the target ATT):
  theta_Q = sum_a w_a^T * ATT_a           = -0.0219
  Q-weighted stacked TWFE coefficient      = -0.0219   [match]
stacked use medicaid, clear
stacked build, time(year) unit(state) adopt(adopt_year) kpre(3) kpost(2)
stacked summary uninsured, cluster(state)

See also