| Title: | A Big Data Implementation of Difference-in-Differences Estimation with Staggered Treatment |
| Version: | 1.0 |
| Description: | Provides a big-data-friendly and memory-efficient difference-in-differences estimator for staggered (and non-staggered) treatment contexts. |
| License: | MIT + file LICENSE |
| Encoding: | UTF-8 |
| RoxygenNote: | 7.2.3 |
| Depends: | data.table, sandwich |
| Suggests: | ggplot2, knitr, rmarkdown, scales, parallel, fixest, progress |
| VignetteBuilder: | knitr |
| URL: | https://setzler.github.io/DiDforBigData/ |
| BugReports: | https://github.com/setzler/DiDforBigData/issues |
| NeedsCompilation: | no |
| Packaged: | 2023-04-02 23:58:12 UTC; bradleysetzler |
| Author: | Bradley Setzler [aut, cre, cph] |
| Maintainer: | Bradley Setzler <bradley.setzler@gmail.com> |
| Repository: | CRAN |
| Date/Publication: | 2023-04-03 15:50:02 UTC |
Combine DiD estimates across cohorts and event times.
Description
Estimate DiD for all possible cohorts and event time pairs (g,e), as well as the average across cohorts for each event time (e).
Usage
DiD(
inputdata,
varnames,
control_group = "all",
base_event = -1,
min_event = NULL,
max_event = NULL,
Esets = NULL,
return_ATTs_only = TRUE,
parallel_cores = 1
)
Arguments
inputdata |
A data.table. |
varnames |
A list of the form varnames = list(id_name, time_name, outcome_name, cohort_name), where all four arguments of the list must be a character that corresponds to a variable name in inputdata. |
control_group |
There are three possibilities: control_group="never-treated" uses the never-treated control group only; control_group="future-treated" uses those units that will receive treatment in the future as the control group; and control_group="all" uses both the never-treated and the future-treated in the control group. Default is control_group="all". |
base_event |
This is the base pre-period that is normalized to zero in the DiD estimation. Default is base_event=-1. |
min_event |
This is the minimum event time (e) to estimate. Default is NULL, in which case, no minimum is imposed. |
max_event |
This is the maximum event time (e) to estimate. Default is NULL, in which case, no maximum is imposed. |
Esets |
If a list of sets of event times is provided, it will loop over those sets, computing the average ATT_e across event times e. Default is NULL. |
return_ATTs_only |
Return only the ATT estimates and sample sizes. Default is TRUE. |
parallel_cores |
Number of cores to use in parallel processing. If greater than 1, it will try to run library(parallel), so the "parallel" package must be installed. Default is 1. |
Value
A list with two components: results_cohort is a data.table with the DiDge estimates (by event e and cohort g), and results_average is a data.table with the DiDe estimates (by event e, average across cohorts g). If the Esets argument is specified, a third component called results_Esets will be included in the list of output.
Examples
# simulate some data
simdata = SimDiD(sample_size=200, ATTcohortdiff = 2)$simdata
# define the variable names as a list()
varnames = list()
varnames$time_name = "year"
varnames$outcome_name = "Y"
varnames$cohort_name = "cohort"
varnames$id_name = "id"
# estimate the ATT for all cohorts at event time 1 only
DiD(simdata, varnames, min_event=1, max_event=1)
Estimate DiD for a single cohort (g) and a single event time (e).
Description
Estimate DiD for a single cohort (g) and a single event time (e).
Usage
DiDge(
inputdata,
varnames,
cohort_time,
event_postperiod,
base_event = -1,
control_group = "all",
return_data = FALSE,
return_ATTs_only = TRUE
)
Arguments
inputdata |
A data.table. |
varnames |
A list of the form varnames = list(id_name, time_name, outcome_name, cohort_name), where all four arguments of the list must be a character that corresponds to a variable name in inputdata. |
cohort_time |
The treatment cohort of reference. |
event_postperiod |
Number of time periods after the cohort time at which to estimate the DiD. |
base_event |
This is the base pre-period that is normalized to zero in the DiD estimation. Default is base_event=-1. |
control_group |
There are three possibilities: control_group="never-treated" uses the never-treated control group only; control_group="future-treated" uses those units that will receive treatment in the future as the control group; and control_group="all" uses both the never-treated and the future-treated in the control group. Default is control_group="all". |
return_data |
If true, this returns the treated and control differenced data. Default is FALSE. |
return_ATTs_only |
Return only the ATT estimates and sample sizes. Default is TRUE. |
Value
A single-row data.table() containing the estimates and various statistics such as sample size. If return_data=TRUE, it instead returns a list in which the data_prepost entry is the previously-mentioned single-row data.table(), and the other argument data_prepost contains the constructed data that should be provided to OLS.
Examples
# simulate some data
simdata = SimDiD(sample_size=200)$simdata
# define the variable names as a list()
varnames = list()
varnames$time_name = "year"
varnames$outcome_name = "Y"
varnames$cohort_name = "cohort"
varnames$id_name = "id"
# estimate the ATT for cohort 2007 at event time 1
DiDge(simdata, varnames, cohort_time=2007, event_postperiod=1)
# change the base period to -3
DiDge(simdata, varnames, base_event=-3, cohort_time=2007, event_postperiod=1)
# use only the never-treated control group
DiDge(simdata, varnames, control_group = "never-treated", cohort_time=2007, event_postperiod=1)
DiD data simulator with staggered treatment.
Description
Simulate data from the model Y_it = alpha_i + mu_t + ATT*(t >= G_i) + epsilon_it, where i is individual, t is year, and G_i is the cohort. The ATT formula is ATTat0 + EventTime*ATTgrowth + \*cohort_counter\*ATTcohortdiff, where cohort_counter is the order of treated cohort (first, second, etc.).
Usage
SimDiD(
seed = 1,
sample_size = 100,
cohorts = c(2007, 2010, 2012),
ATTat0 = 1,
ATTgrowth = 1,
ATTcohortdiff = 0.5,
anticipation = 0,
minyear = 2003,
maxyear = 2013,
idvar = 1,
yearvar = 1,
shockvar = 1,
indivAR1 = FALSE,
time_covars = FALSE,
clusters = FALSE,
markets = FALSE,
randomNA = FALSE,
missingCohorts = NULL
)
Arguments
seed |
Set the random seed. Default is seed=1. |
sample_size |
Number of individuals. Default is sample_size=100. |
cohorts |
Vector of years at which treatment onset occurs. Default is cohorts=c(2007,2010,2012). |
ATTat0 |
Treatment effect at event time 0. Default is 1. |
ATTgrowth |
Increment in the ATT for each event time after 0. Default is 1. |
ATTcohortdiff |
Incrememnt in the ATT for each cohort. Default is 0.5. |
anticipation |
Number of years prior to cohort to allow 50% treatment effects. Default is anticipation=0. |
minyear |
Minimum calendar year to include in the data. Default is minyear=2003. |
maxyear |
Maximum calendar year to include in the data. Default is maxyear=2013. |
idvar |
Variance of individual fixed effects (alpha_i). Default is idvar=1. |
yearvar |
Variance of year effects (mu_i). Default is yearvar=1. |
shockvar |
Variance of idiosyncratic shocks (epsilon_it). Default is shockvar=1. |
indivAR1 |
Each individual's shocks follow an AR(1) process. Default is FALSE. |
time_covars |
Add 2 time-varying covariates, called "X1" and "X2". Default is FALSE. |
clusters |
Add 10 randomly assigned clusters, with cluster-specific AR(1) shocks. Default is FALSE. |
markets |
Add 10 randomly assigned markets, with market-specific shocks that are systematically greater for markets that are treated earlier. Default is FALSE. |
randomNA |
If TRUE, randomly assign the outcome variable with missing values (NA) in some cases. Default is FALSE. |
missingCohorts |
If set to a particular cohort (or vector of cohorts), all of the outcomes for that cohort at event time -1 will be set to missing. Default is NULL. |
Value
A list with two data.tables. The first data.table is simulated data with variables (id, year, cohort, Y), where Y is the outcome variable. The second data.table contains the true ATT values, both at the (event,cohort) level and by event averaging across cohorts.
Examples
# simulate data with default options
SimDiD()