TreatmentPanels

Julia package that defines structs to hold panel data for use in causal inference estimation. This package is used in the package SynthControl to provide input data types for different synthetic control algorithms.

The package is registered in the General Registry and can be added via

julia> Pkg.add("TreatmentPanels")

or in the Pkg REPL

(@v1.10) add TreatmentPanels

The basic idea of the package is to combine a DataFrame with a specification of treatment assignment to construct a an outcome matrix Y, and an accompanying treatment matrix W. Both Y and W follow the convention that a row is an observational unit and a column is a time period - i.e. it can be seen as a "wide" panel data set.

julia> using DataFrames, TreatmentPanels

julia> data = DataFrame(region = repeat(["A", "B"], inner = 5), year = repeat(Date(2000):Year(1):Date(2004), 2), outcome = rand(10))
10×3 DataFrame
 Row │ region  year        outcome   
     │ String  Date        Float64   
─────┼───────────────────────────────
   1 │ A       2000-01-01  0.0605538
   2 │ A       2001-01-01  0.820218
   3 │ A       2002-01-01  0.533732
   4 │ A       2003-01-01  0.144979
   5 │ A       2004-01-01  0.353885
   6 │ B       2000-01-01  0.65294
   7 │ B       2001-01-01  0.353973
   8 │ B       2002-01-01  0.683144
   9 │ B       2003-01-01  0.477427
  10 │ B       2004-01-01  0.702888

julia> bp = BalancedPanel(data, "A" => Date(2003); id_var = "region", t_var = "year", outcome_var = "outcome")
Balanced Panel - single unit, single continuous treatment
    Treated unit: A
    Number of untreated units: 1
    First treatment period: 2003-01-01
    Number of pretreatment periods: 3
    Number of treatment periods: 2

julia> bp.Y
2×5 Matrix{Float64}:
 0.0605538  0.820218  0.533732  0.144979  0.353885
 0.65294    0.353973  0.683144  0.477427  0.702888

julia> bp.W
2×5 Matrix{Bool}:
 0  0  0  1  1
 0  0  0  0  0

The package provides simple plotting functionality to visualise treatment assignment and outcomes:

julia> using Plots

julia> plot(bp, markersize = 10)

julia> plot(bp; kind = "outcome")

There are two basic types:

• BalancedPanel{T1, T2} <: TreatmentPanel holds panel data in which each unit of observation $i = 1, ..., N$ is observed for $t = 1, ..., T$ periods
• UnbalancedPanel{T1, T2} holds panel data in which there might be a different number of periods for each unit of observation

Basic usage of the BalancedPanel constructor is

BalancedPanel(data::DataFrame, treatment_assignment; id_var, t_var, outcome_var)::BalancedPanel{T1, T2}

where

• data is a DataFrame holding the panel data, including columns holding id_var, t_var, and outcome_var
• treatment_assignment specifies the units treated and timing of treatment.
• id_var is the name of the column in data holding the unit identifier
• t_var is the name of the column in data holding the time identifier
• outcome_var is the name of the column in data holding data on the outcome variable of interest

T1 <: UnitTreatmentType and T2 <: TreatmentDurationType are types indicating the nature the treatment observed. UnitTreatmentType currently has three concrete types:

• SingleUnitTreatment - only one unit of observation is subject to treatment
• MultiUnitSimultaneousTreatment - multiple units are being treated in the same time period
• MultiUnitStaggeredTreatment - multiple units are being treated in different periods

TreatmentDurationType currently has two concrete types:

• ContinuousTreatment - once a unit is treated, it remains treated for the remainder of the observation period
• StartEndTreatment - the treatment has an end point in time, i.e. treated units revert to being untreated at some point

The types of T1 and T2 are automatically chosen based on the treatment_assignment passed.

As an example, calling BalancedPanel(data, "unit1" => Date(2000)) will return an object of type BalancedPanel{SingleUnitTreatment, ContinuousTreatment}, as only one unit is treated and only a start point for treatment is provided, so it is assumed that all periods from the year 2000 onwards are treatment periods. BalancedPanel(data, "unit1" => (Date(2000) => Date(2010)) in contrast will return a BalancedPanel{SingleUnitTreatment, StartEndTreatment}, and any observations in the year 2011 or after will be considered untreated.

While many of the convenience methods in this package are specialised on the concrete types of T1 and T2, their main role is in supporting dispatch for synthetic control models. As an example, a standard synthetic control model is designed to estimate the impact of a continuous treatment for a single unit, and hence should accept a BalancedPanel{SingleUnitTreatment, ContinuousTreatment} as input, while a staggered synthetic control model will accept a BalancedPanel{MultiUnitStaggeredTreatment, T2}. In this way, the type of the BalancedPanel passed to an estimation routine can be used to enforce the data structure assumed in the routine.