WeightedStackedDiD.jl

Implements Weighted Stacked Regressions (Wing, Freedman, and Hollingsworth, 2024) - NBER Paper

Following this code example: Link

ToDo

• Different controls - currently only supports not treated during window
  • Never Treated

Function Definition

    stacked_did_reg(data::DataFrame;yvar::Symbol,
    timeID::Symbol,
    unitID::Symbol,
    cohort::Symbol,
    kappa_pre::Int,
    kappa_post::Int,
    x_vars::Vector{Symbol}=Symbol[],
    fes::Vector{Symbol}=Symbol[],
    cluster::Union{Nothing,Symbol}=nothing,
    contrasts::Dict{Symbol, DummyCoding}=Dict{Symbol, DummyCoding}(),
    multi_thread::Bool=true,
    inc_sub_exp_fes=false)

Arguments

• data: The DataFrame
• yvar: Y variable
• timeID: The calander time variable (assumes an integer not a date variable)
• unitID: The unit/individual ID
• cohort: Time period of first treatment (assumes an integer not a date variable)
• kappa_pre: Number of pre periods
• kappa_post: Number of post periods
• x_vars: Vector of control variables
• fes: Vector of additional fixed effects
• cluster: The cluster variable (currently only supports one). If not supplied it is set at the level of the unitID
• contrasts: contrasts for DummyVariables. Required if you included a Dummy Variable in x_vars
• multi_thread: Whether to use multithreading or not in building the stacks
• inc_sub_exp_fes :If yes includes sub-experiment-unit-id and sub-experiment-event-time fixed effects

Example

julia> using CSV, DataFrames, FixedEffectModels

julia> data = CSV.File(raw"data\acs1860_unins_2008_2021.csv") |> DataFrame;

julia> data.adopt_year[ismissing.(data.adopt_year)] .=0; # Never Treated Adoption time must be set to zero

julia> wsd_res = stacked_did_reg(data,
  yvar=:unins,
  timeID=:year,
  unitID=:statefip,
  cohort=:adopt_year,
  kappa_pre=3,
  kappa_post=2,
  cluster=:statefip)
                                FixedEffectModel
================================================================================
Number of obs:                      336  Converged:                         true
dof (model):                          5  dof (residuals):                     39
R²:                               0.391  R² adjusted:                      0.369
F-statistic:                    5.57619  P-value:                          0.001
R² within:                        0.018  Iterations:                           2
================================================================================
             Estimate  Std. Error     t-stat  Pr(>|t|)    Lower 95%    Upper 95%
────────────────────────────────────────────────────────────────────────────────
ttt::-3   0.000638982  0.00451112   0.141646    0.8881  -0.00848562   0.00976358
ttt::-2  -0.00374609   0.00317994  -1.17804     0.2459  -0.0101781    0.00268595
ttt::0   -0.0181662    0.00417973  -4.34627     <1e-04  -0.0266205   -0.00971193
ttt::1   -0.0260809    0.00648439  -4.0221      0.0003  -0.0391968   -0.012965
ttt::2   -0.0303825    0.00808953  -3.75578     0.0006  -0.0467451   -0.0140199
================================================================================

Calculate the Trimmed Aggregate ATT (average post-treatment effect). Also shows the average pre-treatment effect.

julia> agg_to_ATT(wsd_res)

             Trimmed Aggregate ATT
           (average post-treatment effect)
=====================================================
Dof   : 51.0
T-ATT : -0.0219                  Std. Error: 0.0056
FVal  : 15.0847                  P-Value   : 0.0003
-----------------------------------------------------
Average pre-treatment effect:
Dof   : 51.0
Effect: -0.002                   Std. Error: 0.0032
FVal  : 0.4137                   P-Value   : 0.523
=====================================================