Plotting results
The plot functions take the result
of sdid() and return ggplot objects, which draw when
printed and can be extended with +. They require ggplot2,
and sdid_units_plot() uses ggrepel, when it is installed,
to keep its labels apart.
library(sdid)
data(california_prop99)
fit = sdid(california_prop99, "PacksPerCapita", "State", "Year", "treated",
vce = "placebo", seed = 1213, graph = FALSE, quietly = TRUE)Trends
sdid_plot() draws the treated units’ average and the
synthetic control. The red dashed line marks treatment. The gray band at
the bottom shows the time weights with their scale on the right-hand
axis.
style = "effects" adds a diagram of the comparison
behind the estimate, with an arrow marking the effect.
Individual units
spaghetti.units adds the trajectories of chosen units,
here the three controls with the largest weights.
w = weights(fit)$omega
top = w$unit[order(-w$`1989`)][1:3]
sdid_plot(fit, spaghetti.units = top, spaghetti.label.size = 3)
spaghetti.matrices adds any other trajectories you want,
as a list with one matrix per estimate. Each row is a trajectory with
one column per period, and its row name becomes the label. Here we add
the plain average of all controls, which is what
difference-in-differences compares against.
Y = fit$setup$Y
average = rbind("Average of controls" = colMeans(Y[w$unit, , drop = FALSE]))
sdid_plot(fit, spaghetti.matrices = list(average), spaghetti.label.size = 3)
Comparing estimators
A named list of results draws each in its own panel, stacked and labeled by its name.
estimates = lapply(c(SDID = "sdid", SC = "sc", DID = "did"), function(m) {
sdid(california_prop99, "PacksPerCapita", "State", "Year", "treated",
method = m, vce = "noinference", graph = FALSE, quietly = TRUE)
})
sdid_plot(estimates)
Synthetic control puts no weight on the periods before treatment, so
its panel has no band, and difference-in-differences weighs them
equally. Each band is drawn on its own scale.
lambda.comparable = TRUE draws them on a common scale
instead, where a weight of 1 fills the band, so that their heights can
be compared across panels.
Ticks
By default each axis gets about five ticks, and a year axis only gets
whole years. To choose the ticks yourself, use x.breaks and
y.breaks. Adding a ggplot scale to the plot instead would
replace the package’s own.
Unit weights
sdid_units_plot() shows what each control contributes.
For every control, it computes the difference-in-differences estimate
using that control alone; the overall estimate is the weighted average
of these. The default scatter plots each control’s weight against its
own estimate, and controls with no weight are drawn as open circles.
Heavily weighted controls sit on the right. If they cluster around
the red line, they agree with each other. type = "bars"
shows only the weighted controls, largest weight first, with each bar
running from the overall estimate to that control’s own estimate.
label.units sets which controls get labels: all of them,
only the weighted ones, or the n with the largest weights.
A staggered result gets one panel per cohort, with ten labels each,
unless cohort picks a single cohort.
Placebo tests
sdid_placebo_plot() draws the estimate above an in-time
placebo, which pretends treatment began earlier, at
placebo.time, using only the periods before the real
treatment. If there was no effect before the real treatment, the
placebo’s two lines should keep tracking each other after the red line
that marks the placebo treatment.
sdid_placebo() returns the placebo estimate itself.