didImputation: DiD regression with imputation method
In CdfInnovLab/didImputation: Staggered difference-in-differences estimation

Description Usage Arguments Details Value Estimated coefficients Effective sample a.k.a. Herfindahl condition Cohort Standard-errors Author(s) References Examples

View source: R/didImputation.R

DiD regression with imputation method

didImputation(
  y0,
  data,
  cohort,
  nevertreated.value = Inf,
  unit = NULL,
  time = NULL,
  het = NULL,
  w = NULL,
  coef = -Inf:Inf,
  OATT = FALSE,
  with.se = TRUE,
  tol = 1e-06,
  maxit = 100,
  mutatedata = FALSE,
  verbose = 0,
  effective.sample = 30
)

`y0`	Formula. model for Y(0). This is the full model without leads and lags used to predict the counterfactual outcome.
`data`	A data frame or data.table
`cohort`	Character. Time of treatment identifier. By default it expect `Inf` for never treated units. You can override this with `nevertreated.value`
`nevertreated.value`	Any, default is `Inf`. Value encoding the cohort of never treated units.
`unit`	Character, default is `NULL`. Unit identifier. If `NULL`, it default to the first fixed effect.
`time`	Character, default is `NULL`. Time period identifier. If `NULL`, it default to the second fixed effect.
`het`	Character, default is `NULL`. Heterogeneity estimation. Column name of the additional group(s). Will produce an estimate for each group.
`w`	Numerical vector or Variable name. Default is `NULL`. Sample weights.
`coef`	R Expression of leads and lags (eg `coef = -5:8`). By default estimates whole dynamic effect. See Details.
`OATT`	Logical, default is `FALSE`. If `TRUE` then the overall average treatment effect is computed instead of ATT at each horizon.
`with.se`	Logical, default is `TRUE`. Should standard errors be reported
`tol`	Numeric, default is `1e-6`. Tolerance level for weights convergence.
`maxit`	Numeric, default is `100`. Maximum number of iterations for computing weights
`mutatedata`	Logical, default is `FALSE`. USE AT YOUR OWN RISK. This option will modify your data in place instead of making a copy of it which could be useful with large dataset if your RAM is limited. It may result in lost observations and bugs.
`verbose`	Numeric, default is `0`. Level of verbosity.
`effective.sample`	Numeric, default is `30`. Effective sample size under which the function will throw a warning. See details.

See below for additional details on some arguments.

A didImputation object with the results of the imputation estimation.

`data`	Data used for estimation.
`convergence_time`	User and CPU time for weights convergence.
`pvalue`	p-value for the positive horizon estimates.
`coeftable`	Table of regression coefficients.
`wald`	Wald statistic of the pre-trend regression.
`coefs`	Average treatment effects from the imputation procedure.
`niterations`	Number of iterations it took to compute the weights.
`pre_trends`	`fixest` regression object for the pre-trends estimation.

By default, the function will estimate all coefficients available. You can customize this behavior with the coef option. The option takes an R expression of the form -{leads}:{lags}. The default behavior is to estimate all available coefficients such that coef = -Inf:Inf.

For pre-trend coefficients, by default the function will set the greatest leads as the reference group if -Inf is set.

Borusyak, K., Jaravel, X., & Spiess, J. give a condition on weights such that consistency of estimators holds. This condition states that the Herfindahl index of weights must converge to zero. Another interpretation is that the inverse of the index is a measure of 'effective sample size'. Authors recommand an effective sample size of at least 30. If the effective sample size is lower, a warning will be thrown.

The cohort argument is the date of treatment of the observation. By default it expect Inf for never treated individuals. You can override this behavior by setting another value to nevertreated.value.

The standard errors are computed using an alternating projection method. You can tweek its meta-parameter by changing tol and maxit.

Antoine Mayerowitz

Maxime Gravoueille

Borusyak, K., Jaravel, X., & Spiess, J. (2021). Revisiting event study designs: Robust and efficient estimation. Working paper.

Stata implementation from the authors:

did_imputation (https://github.com/borusyak/did_imputation)

Another R implementation using sparse matrix inversion: didimputation (https://github.com/kylebutts/didimputation)

#Load example data
data(did_simulated)

# Estimate the overall average treatment effect on treated and all available pre-trends
didImputation(y ~ 0 | i + t,
              cohort = 'g',
              OATT = TRUE,
              data = did_simulated)

# Estimate the full dynamic model
didImputation(y ~ 0 | i + t,
              cohort = 'g',
              data = did_simulated)

# Estimate positive (lags) coefficients
didImputation(y ~ 0 | i + t,
              cohort = 'g',
              coef = 0:Inf,
              data = did_simulated)

# Estimate first 3 post treatment coefficients
didImputation(y ~ 0 | i + t,
              cohort = 'g',
              coef = 0:2,
              data = did_simulated)

# Return only point estimates
didImputation(y ~ 0 | i + t,
              cohort = 'g',
              with.se = FALSE,
              data = did_simulated)