Authors: In Song Kim (insong@mit.edu), Adam Rauh (amrauh@umich.edu), Erik Wang (haixiaow@Princeton.edu), Kosuke Imai (imai@harvard.edu)
PanelMatch is an R package implementing a set of methodological tools proposed by Imai, Kim, and Wang (2021) that enables researchers to apply matching methods for causal inference on time-series cross-sectional data with binary treatments. The package includes implementations of matching methods based on propensity scores and Mahalanobis distance, as well as weighting methods. PanelMatch enables users to easily calculate a variety of possible quantities of interest, along with standard errors. The software is flexible, allowing users to tune the matching, refinement, and estimation procedures with a large number of parameters. The package also offers a variety of visualization and diagnostic tools for researchers to better understand their data and assess their results.
PanelMatch is available on CRAN and can be installed using:
install.packages("PanelMatch")You can install the most recent development version of PanelMatch using the devtools package. First you have to install devtools using the following code. Note that you only have to do this once:
if(!require(devtools)) install.packages("devtools")Then, load devtools and use the function install_github() to install PanelMatch:
library(devtools)
install_github("insongkim/PanelMatch", dependencies=TRUE, ref = "se_comparison")If you encounter problems during installation, please consult the wiki page that has some ideas for handling common issues.
Users can visualize the variation of treatment across space and time. This will help users build an intuition about how comparison of treated and control observations can be made.
library(PanelMatch)
DisplayTreatment(unit.id = "wbcode2",
time.id = "year", legend.position = "none",
xlab = "year", ylab = "Country Code",
treatment = "dem", data = dem)
PanelMatch identifies a matched set for each treated
observation. Specifically, for a given treated unit, the matched set
consists of control observations that have an identical treatment
history up to a chosen number (lag) of years. This number corresponds with the lag parameter, which must
be chosen by the user. Users must also consider various parameters regarding the refinement of created matched sets. Please consult the function documentation for a full set of descriptions, but some important arguments are described below:
refinement.method-- Users may choose between standard propensity score weighting or matching (ps.weight,ps.match), covariate balanced propensity score weighting or matching (CBPS.weight,CBPS.match), and mahalanobis distance matching (mahalanobis). Users may also opt to apply the idea of marginal structural models with theCBPS.msm.weightandps.msm.weightmethods. Alternatively users can do no refinement by setting this parameter tonone.size.match-- This sets the maximum number of control units that can be included in a matched set.covs.formula-- This parameter defines which variables are considered in measuring the similarities/distances between units. These will then affect which control units are included/excluded during refinement. This can be set to include lagged versions of any variable as well. See thePanelMatchdocumentation for more information about this parameter.match.missing-- Should matches between treatment and control units with identical patterns of missingness in the treatment variable be considered? If set to FALSE, missing data is not permitted in the lag window of the treatment variable in either treated or control units.
PM.results <- PanelMatch(lag = 4, time.id = "year", unit.id = "wbcode2",
treatment = "dem", refinement.method = "mahalanobis",
data = dem, match.missing = TRUE,
covs.formula = ~ I(lag(tradewb, 1:4)) + I(lag(y, 1:4)),
size.match = 5, qoi = "att" ,outcome.var = "y",
lead = 0:4, forbid.treatment.reversal = FALSE)
The PanelMatch function will return an object of class "PanelMatch". This is a list that contains a few specific elements: First, a matched.set object(s) that has the same name as the provided qoi -- if the qoi is "att", "atc". If qoi = "ate" then two matched.set objects will be attached, named "att" and "atc." Users can extract information about individual matched sets as well as statistics about all created matched sets from this object. Consult the Wiki page on Matched Set Objects for a more detailed walk through and description of these objects. Put simply, matched.set objects are merely lists with some assumed structure and special attributes.
The PanelMatch object also has some additional attributes: "qoi", "lead", "forbid.treatment.reversal" (a logical value that is the same as what was specified in the function call), and "outcome.var" (character value that is the same as what was specified in the function call)
You can check covariate balance using the get_covariate_balance function:
get_covariate_balance(PM.results$att, dem, covariates = c("tradewb"), plot = FALSE, ylim = c(-2,2))
tradewb
t_4 0.05459705
t_3 -0.03101839
t_2 -0.01828529
t_1 0.07784846
See the documentation for more information about this function.
Once proper matched sets are attained by PanelMatch, users can
estimate the causal quantity of interest such as the average
treatment effect using PanelEstimate. Either bootstrap or weighted
fixed effects methods can be used for standard error
calculation. Users can estimate the contemporaneous effect as well as
long-term effects. In this example, we illustrate the use of
PanelEstimate to estimate the average treatment effect on treated units (att) at time t on the outcomes from time t+0 to t+4.
PE.results <- PanelEstimate(sets = PM.results, data = dem)The PanelEstimate function returns a PanelEstimate object, which is a named list. This object will contain the point estimates, standard errors and other information about the calculations. See the wiki page about PanelEstimate objects for more information.
Users can easily obtain and visualize important information about esimtates and standard errors using the summary and plot methods for PanelEstimate objects
summary(PE.results)
Weighted Difference-in-Differences with Mahalanobis Distance
Matches created with 4 lags
Standard errors computed with 1000 Weighted bootstrap samples
Estimate of Average Treatment Effect on the Treated (ATT) by Period:
$summary
estimate std.error 2.5% 97.5%
t+0 -0.5349572 0.9231828 -2.434598 1.314614
t+1 -0.2396204 1.4565119 -2.970407 2.506002
t+2 0.5532550 1.8562746 -2.949847 4.059749
t+3 1.8425824 2.1679704 -2.421909 5.953401
t+4 1.9920680 2.3756352 -2.648517 6.569985
$lag
[1] 4
$iterations
[1] 1000
$qoi
[1] "att"plot(PE.results)