This vignette shows a first analysis with fasttrop. You make a practice panel, you calculate the TROP estimate, and you read the result. The article “TROP in plain words” on the package website explains the ideas of the method.
The data
fasttrop uses panel data in long format. Each row is one unit in one period. The data must have four columns:
- the outcome
- the treatment indicator, with the values 0 and 1
- the unit identifier
- the period identifier
The function trop_simulate() makes a practice panel with
a known effect. In this panel, 5 of 20 units get the treatment in the
last 3 of 12 periods. The true effect is 1 for each treated cell.
library(fasttrop)
panel <- trop_simulate(n_units = 20, n_periods = 12, n_treated = 5, n_post = 3,
effect = 1, seed = 1)
head(panel)
#> unit time y d y0
#> 1 1 1 -0.5496887 0 -0.5496887
#> 2 2 1 -0.5309372 0 -0.5309372
#> 3 3 1 -1.1621093 0 -1.1621093
#> 4 4 1 2.5570765 0 2.5570765
#> 5 5 1 -0.1638954 0 -0.1638954
#> 6 6 1 -1.1801013 0 -1.1801013The column y is the outcome, and the column
d is the treatment indicator. The column y0 is
the outcome without the treatment. Real data does not have this column.
The practice panel has it, so you can compare the estimate with the
truth.
A first estimate
To calculate the estimate, give trop() the data and the
names of the four columns. In this first fit, you set the tuning
parameters. This makes the fit fast.
fit <- trop(panel, outcome = "y", treatment = "d", unit = "unit", time = "time",
lambda = c(time = 0.2, unit = 0.5, nn = 0.5), n_boot = 100, seed = 1)
fit
#> Triply Robust Panel (TROP) estimate
#> Method: local | Design: block | Units: 20 | Periods: 12 | Treated cells: 15
#> ATT: 1.119 (SE 0.2199, 95% CI 0.6879 to 1.55)
#> lambda: time = 0.2, unit = 0.5, nn = 0.5After the title, the output has three lines:
- The method, the design of the panel, the size of the panel, and the number of treated cells.
- The average treatment effect on the treated (ATT), with the bootstrap standard error (SE) and the 95% confidence interval.
- The tuning parameters.
The true ATT of this panel is 1. The interval contains this value.
Parts of the result
Use these functions to get parts of the result:
coef(fit)
#> ATT
#> 1.118779
confint(fit)
#> 2.5 % 97.5 %
#> ATT 0.6878718 1.549686
head(as.data.frame(fit))
#> unit time tau
#> 1 7 10 2.0694817
#> 2 7 11 3.3467710
#> 3 7 12 0.1396825
#> 4 11 10 0.7607983
#> 5 11 11 1.9304028
#> 6 11 12 -0.3167042The function as.data.frame() gives the effect for each
treated cell. The ATT is the mean of these effects.
Let the data select the tuning parameters
TROP has three tuning parameters:
-
timesets how fast the time weights decrease with the distance from the treated period. With 0, all periods get the same weight. -
unitsets how fast the unit weights decrease when the outcome history of a unit is different from the treated unit. With 0, all units get the same weight. -
nnis the penalty on the low-rank term. A larger value gives a simpler low-rank term. WithInf, the model has no low-rank term.
If you do not set lambda, cross-validation selects the
tuning parameters from the grid of trop_grid(). To make
this example fast, the cross-validation uses a sample of 60 untreated
cells.
fit_cv <- trop(panel, outcome = "y", treatment = "d", unit = "unit", time = "time",
cv_cells = 60, n_boot = 100, seed = 1)
summary(fit_cv)
#> Triply Robust Panel (TROP) estimate
#> Method: local | Design: block | Units: 20 | Periods: 12 | Treated cells: 15
#> ATT: 1.142 (SE 0.2267, 95% CI 0.6978 to 1.587)
#> lambda: time = 0.1, unit = 0.5, nn = 1
#> Tuning: loocv, score 18.79, 34 grid evaluations
#> Bootstrap: 100 draws (stratified), 0 failed, normal interval
#> Treated units: 5 | Skipped cells: 0 | Non-converged fits: 0
#> Time: 1.454 s in total, 1.449 s in the C code (4 threads)The function summary() also gives the result of the
tuning, the bootstrap, and the time of the calculation. With
cv_cells = NULL, the default, the cross-validation uses all
untreated cells, as in the paper. This takes more time, but it uses all
the data.
The local method and the global method
The default method is the local method of the paper. It calculates a separate fit for each treated cell. The global method is the method of the Python package from the authors. It calculates one fit for all treated cells. You can use the global method only for a block design. In a block design, all treated units start the treatment in the same period and stay treated.
Standard errors and intervals
A unit bootstrap gives the standard error. Each draw takes a sample
of units with replacement and calculates the estimate again. fasttrop
keeps the tuning parameters fixed in each draw. The argument
n_boot sets the number of draws. The default is 200
draws.
The default interval is a normal interval. To get a percentile
interval, set ci = "percentile". The bootstrap is correct
only with many treated units. For a panel with fewer than five treated
units, fasttrop gives a warning message.
Threads and seeds
fasttrop runs its model fits in C, and it can use more than one
thread. The function trop_threads() gives the default
number of threads. To set the number of threads, use
trop_control():
fit <- trop(panel, outcome = "y", treatment = "d", unit = "unit", time = "time",
control = trop_control(threads = 4))The number of threads does not change the results. With the same
seed, a fit gives the same result each time. The
seed argument does not change the random number generator
of your R session.
Next steps
- The article “How to read your results” explains each part of the output and each warning message.
- The article “What TROP can and cannot do” gives the limits of the method.
- The help page
?tropgives all arguments.