Estimate DiD Effects with Compositional Changes

comp_did() is the user-facing entry point for the package. It accepts a two-sided formula such as y ~ x1 + x2, or separate column-name arguments, builds the validated internal data object, fits nuisance functions, and returns the non-stationary doubly robust ATT estimator for repeated cross-section data with compositional changes.

Usage

comp_did(
  yname,
  tname,
  dname,
  xformla = NULL,
  data,
  xformula = NULL,
  continuous = NULL,
  unordered = NULL,
  ordered = NULL,
  nuisance_method = c("dml", "loo", "parametric"),
  stationary = FALSE,
  stationarity_test = FALSE,
  stabilized = TRUE,
  i.weights = NULL,
  boot = FALSE,
  nboot = NULL,
  boot_type = c("mammen", "normal", "bayes", "wild"),
  ps_min_treat = 0.005,
  inffunc = TRUE,
  ...
)

Arguments

yname: Name of the outcome column, or a two-sided formula such as y ~ x1 + x2. When a formula is supplied, tname, dname, and data must be named explicitly.
tname: Name of the post-period indicator column. Values must be coded as 0 for the pre period and 1 for the post period.
dname: Name of the treatment-group indicator column. Values must be coded as 0 for controls and 1 for the treated group.
xformla, xformula: Optional one-sided formula for covariates. xformula is the preferred spelling; xformla is accepted for users familiar with DID-package style arguments. When supplied, the resulting model matrix columns are treated as continuous covariates. For nuisance_method = "loo", formula covariates must be numeric; use continuous, unordered, and ordered instead when the local-polynomial mixed-data kernels should distinguish covariate types.
data: Data frame containing all analysis variables.
continuous, unordered, ordered: Optional covariate column names or positions identifying continuous, unordered discrete, and ordered discrete covariates. These are passed to make_did_dp().
nuisance_method: Either "dml" for cross-fitted nuisance estimation, "loo" for local-polynomial nuisance estimation, or "parametric" for fixed-dimensional parametric first steps.
stationary: Logical. If TRUE, also compute the stationary repeated cross-section estimator.
stationarity_test: Logical. If TRUE, compute the test comparing stationary and non-stationary estimates. The nonparametric and DML paths report the existing Hausman-type comparison; the parametric path reports the corresponding Wald equality test. This requires stationary = TRUE and influence functions.
stabilized: Logical. If TRUE, use stabilized weights in the DR second-stage estimators.
i.weights: Optional non-negative sampling weights.
boot: Logical. If TRUE, use multiplier-bootstrap inference.
nboot: Number of bootstrap draws when boot = TRUE.
boot_type: Multiplier-bootstrap type.
ps_min_treat: Minimum treated-group propensity used in stationary weights when stationary = TRUE or stationarity_test = TRUE. The default (0.005) clips fitted treated propensities to [ps_min_treat, 1 - ps_min_treat]; set to 0 to disable clipping.
inffunc: Logical. If TRUE, return influence functions.
...: Additional arguments passed to att_estimate().

Value

A "compdid" object containing the ATT fit, nuisance metadata, the processed data object, and optionally stationary-fit and test results.

Examples

set.seed(123)
d <- simulate_comp_did(n = 200)
out <- comp_did(
  y ~ x1 + x2 + x3 + x4 + x5 + x6,
  tname = "post",
  dname = "d",
  data = d,
  K = 2
)
out
#> Compositional-change DiD
#> Nuisance method: dml 
#> Nuisance backend: custom 
#> ATT: 9.566 
#> SE : 7.618 
#> 95% CI: [-5.366, 24.5]