library("microsynth") library("knitr") knitr::opts_chunk$set(message = FALSE, warning = FALSE, fig.height=3.1, fig.width=7.2, fig.show = "hold") # For fast vignette-compiling (needed for CRAN), declare function to save reduced microsynth object. # (Microsynth chunks with long run-times are set to eval=TRUE to create objects for the first time, # but these chunks are set to eval=FALSE in the package for fast vignette compiling.) # This reduces microsynth files to <1MB instead of >10 MB saveReducedMicrosynth <- function(msobject, filename) { # Proceed only if object exists (so it is always safe to run the code) if(exists(deparse(substitute(msobject)))) { # Strip data-heavy objects and save msobject$w$Weights <- NULL msobject$w$Intervention <- NULL saveRDS(msobject, filename) } }
Synthetic controls are a generalization of the difference-in-difference approach (Abadie and Gardeazabal, 2003; Abadie et al, 2010). Difference-in-difference methods often require the researcher to manually identify a control case, against which the treatment will be compared, on the basis of apparent similarity before the intervention and the plausibility that identical secular trends affect both the treatment and control equally after the intervention. Instead, the synthetic control method offers a formalized and more rigorous method for identifying comparison cases, by constructing a "synthetic" control unit that represents a weighted combination of many untreated cases. Weights are calculated in order to maximize the similarity between the synthetic control and the treatment unit in terms of specified "matching" variables. By matching on the observable characteristics between treatment and control, the method may also do a better job of matching on the unobservable characteristics (though by nature this cannot be verified).
The advantages over the general difference-in-difference approach are several: a) the observable similarity of control and treatment cases is maximized, and perhaps also similarity of unobservables, strengthening the assumptions (e.g., equal secular trends) inherent to the difference-in-difference approach; b) the method is feasible even when there exists no single untreated case adequately similar to the treatment case; and c) researchers can point to a formal and objective approach to the selection of controls, rather than having to justify ad hoc decisions which could potentially create the appearance of the researcher having his thumb on the scale.
Generally, synthetic controls have been applied in the context of a single
treatment case with a limited number (e.g., several dozens) of untreated cases
for comparison. The Synth
package has been developed for R and designed for
this type of application. But the relative dearth of treatment and comparison
data in such settings complicates efforts to a) develop a synthetic control that
matches the treatment case, b) precisely estimate the effect of treatment, c)
gauge the significance of that effect, and d) jointly incorporate multiple
outcome variables.
This package is developed to address those limitations, by incorporating high-dimensional, micro-level data into the synthetic controls framework. Therefore, in addition to what Synth provides, microsynth offers several advantages and new tools:
With the advantage of a large number of smaller-scale observations, microsynth
is often better able to calculate weights that provide exact matches between
treatment and synthetic control units (on all variables passed to match.out
[for time-variant variables] and match.covar
[for time-variant variables]).
This bolsters the conceptual framework behind the synthetic control method.
To generate an additional measure for significance, microsynth can generate
hundreds or thousands of placebo treatment units using random permutations of
the control units (e.g., with perm = 250
and jack = TRUE
). This allows
estimated effects from the actual treatment unit to be compared to effects for
the placebo treatment units, after standardization (if use.survey = TRUE
),
generating a new variance estimator and p-value. The sampling distribution of
the effects from placebo treatment units is plotted visually, along with
Synth-style plots comparing observed outcomes in the treatment and synthetic
control units over time.
An omnibus statistic is calculated to assess the statistical significance
across multiple variables (i.e., those set to omnibus.var
), as may be desired
in scenarios with limited power where several outcome variables.
Results may be estimated across multiple follow-up periods (by passing a
vector to end.post
).
Matching variables may be specified flexibly. Time-variant variables may be
aggregated across multiple time periods before matching (by passing a list to
match.out
or passing a value to period
), helping to reduce variable
sparseness and improve the likelihood of a satisfactory match.
microsynth provides parameters to assist users in finding feasible models when
a plethora of matching variables and a scarcity of data make the calculation of
satisfactory weights difficult. Users may call check.feas
or use.backup
to
call on more computationally-intensive methods to calculate weights.
Alternately, difficult-to-match variables may be passed to
match.out.min
/match.covar.min
as to seek weights that deliver the
best-possible but not necessarily exact match on those variables.
microsynth is also backwards compatible, i.e., it can be deployed on the Synth-like case of a single treatment with a limited number of untreated cases, although the relative dearth of data should be expected to decrease matching performance and limit the usefulness of the features discussed above (see Example 8).
For this example we will use \code{\link{seattledmi}} to evaluate a Drug Market Intervention using the "seattledmi" dataset provided with the microsynth package. The intervention was applied to 39 blocks, which represent the treatment; the remaining 9,603 Seattle blocks are potential comparison units from which the synthetic control may be constructed. Data are available for block-level Census demographics and incidences of crime reported by the Seattle Police Department.
colnames(seattledmi) set.seed(99199)
We would like to detect whether the program was effective at reducing the incidence of crime in those neighborhoods where the intervention was applied. Before beginning examples, we will specify the mandatory minimum parameters pursuant to the dataset and our basic research design.
The bedrock of the synthetic controls research design (like any
difference-in-difference method) involves comparing observations between
treatment (i.e., "intervention") areas versus control areas, with observations
for each unit over a certain period of time. Therefore microsynth requires we
identify the idvar
, timevar
, and intvar
columns.
In this case, we are provided with Census block-level observation units (idvar
= "ID"
) and quarterly observations (timevar = "time"
), along with a binary
variable with 0 for all untreated groups and the treated groups during the
pre-intervention period and a 1 for treated groups at the time of intervention
and later (intvar = "Intervention"
).
Next, the user can specify parameters relating to the beginning of the
pre-intervention data (start.pre
), the last time period of the pre-intervention
period (end.pre
),
and the time(s) through which post-intervention effects ought to be estimated
(end.post
). For all observations up to and including end.pre
, outcome
variables and covariates will be used to match treatment and control. (If the
data is formatted such that 0s are assigned to all end.pre
observations for
the control units and treatment units pre-intervention, and 1s assigned only
to treatment units post-intervention, then end.pre
will by default be
automatically set appropriately, such that end.pre
will equal the last period
of pre-intervention data.)
In this case, our study period begins at the first quarter of data available in
the dataset; the intervention occurs after 12 quarters of pre-intervention data
(end.pre = 12
); and our study period continues for four quarters of
post-intervention data (end.post = 16
). With this dataset, end.post
could
also be left unassigned and would be automatically set to the latest observation
in the data; likewise, we can set end.pre = NULL
, as we
expect the program's effects not to occur instantaneously, the intvar
column
is adequately formatted to allow microsynth to detect the intervention time
automatically. Note that start.pre
will default to the earliest time in the
dataset.
The last group of microsynth
's mandatory parameters relate to declaring
outcome variables and covariates. Both outcomes and covariates will be used to
match treatment units to synthetic controls during the pre-intervention period.
The key difference between outcomes and covariates is that outcome variables are
required to be time-variant, and covariates to be time-invariant (constant
overtime).
For this study, we would like to estimate the effect of the DMI on rates of
crime. Specifically we are interested in the effects on four types of incidences
of crime: felony arrests, misdemeanor arrests, drug arrests, and any criminal
arrest. Passing these variables to match.out
instructs microsynth to calculate
weights that provide exact matches on these variables; assigning result.var =
match.out
identifies them as outcome variables for which we would like effects
estimated; omnibus.var
will include them in the omnibus statistic. After the
microsynth
object is created, we can plot results with \code{plot.microsynth}
with the argument set to plot.var = match.out
to indicate variables to appear
on plots.
Treatment and synthetic control will also be matched on block-level Census
demographic data: each block's population, black residents, hispanic residents,
males aged 15-21, the number of households, the number of families per
household, the number of female-led households, the number of households that
are renters, and the number of vacant houses. As these variables are all
time-invariant, they will be set to "match.covar"
.
cov.var <- c("TotalPop", "BLACK", "HISPANIC", "Males_1521", "HOUSEHOLDS", "FAMILYHOUS", "FEMALE_HOU", "RENTER_HOU", "VACANT_HOU") match.out <- c("i_felony", "i_misdemea", "i_drugs", "any_crime")
Exact matches are not always possible, especially for variables that are sparse
(i.e., few non-zero values), containing little variation, or for which the
treatment units have values outside of the range of observations from the
un-treated units. In these cases, variables may be moved from
match.out
/match.covar
to match.out.min
/match.out.covar
as to minimize
the distance between treatment and synthetic control on those variables rather
than find exact matches. Alternately, a value may be set to period
to
aggregate all variable names in match.out
/match.covar
under the same regular
time duration; or, to set aggregation instructions with more detail,
match.out
/match.covar
may receive a list with detailed parameters.
microsynth() provides several different ways to address this problem. A variable
can be treated such that the distance between treatment and synthetic control is
minimized, even if a distance of zero is infeasible, by listing it under
match.out.min
(for time-variant outcome variables) or match.covar.min
(for
time-invariant variables). In this case, match.out
, match.out.min
,
match.covar
, and match.covar.min
may each be vectors of variable names.
There ought not be any overlap: each variable should appear in only one
argument.
Another potential response is to aggregate the variable across multiple time
periods. match.out
, match.out.min
, match.covar
, and match.covar.min
all
behave similarly in this manner. Rather than being passed a vector of variable
names, each may receive a list; each element of the list is a vector
corresponding to the time units across which each variable should be aggregated
before matching, with each element named equal to the variable name. In this
case, the element vectors represent the duration during which the variable
should be aggregated, counting backwards from the intervention time.
Combining these approaches, if match.covar.min = list("Y1" = c(1, 3, 3))
, then
the variable "Y1" will be used to match treatment to synthetic control at the
time of the intervention (t), the sum of values of "Y1" across t-1 to t-3,
and the sum across t-4 to t-6.
If the dataset contains both time-variant outcome variables and time-variant
predictor variables (i.e., belonging on the RHS of a regression rather than the
LHS), then both 1) match.out
or match.out.min
and 2) result.var
must be
specified. match.out
or match.out.min
should include all time-variant
variables used for matching, whether they are true outcomes or predictors;
result.var
should specify only the subset of those that are outcomes (for
which estimated effects will be calculated).
Note: in some cases, the term "outcome variable" may be a misnomer. Though by
default all time-variant variables assigned to match.out
and match.out.min
will be used to estimate the program effect (result.var = T
), this doesn't
have to be the case. result.var
may be set to a vector of variable names
representing a subset of the outcome variables entered into match.out
and
match.out.min
; this is useful if the dataset includes time-variant variables
that we'd like to use to match treatment and synthetic control but which we do
not want to use for the purposes of evaluating the program effect.
microsynth allows for extensive configuration, for instance, relating to the mechanics of calculating weights, plotting options, and the calculation of variance estimators through permutation tests and jackknife replication groups. These aspects will be discussed in the later examples below.
In this minimal example, we will calculate and display results in the simplest way possible. This includes:
test=lower
) result.var = match.out
) plot_microsynth()
); to save to file,
specify a .csv or .xlsx as the file
argument in plot_microsynth()
. result.file=NULL
); instead,
results can be viewed by inspecting the microsynth
object.As microsynth runs, it will display output relating to the calculation of weights, the matching of treatment to synthetic control, and the calculation of survey statistics (e.g., the variance estimator). The first table to display summarizes the matching properties after applying the main weights. It shows three columns: 1) characteristics of the treated areas for the time-variant and time-invariant variables, 2) characteristics of the synthetic control, and 3) characteristics of the entire population. Because this example is successful in creating a matching synthetic control, the first column and the second column will be nearly equal.
Note that match.out = match.out
, result.var = match.out
, and
omnibus.var =match.out
. This means that the outcome
variables that we declared as match.out
will all be matched on
exactly, will be used to report results, and will feature in the omnibus p-value.
match.covar
indicates that the specified covariates will also be matched on
exactly. (By setting result.var = match.out
, there is provided one chart per
time-variant outcome variable for which we calculate results.)
sea1 <- microsynth(seattledmi, idvar="ID", timevar="time", intvar="Intervention", start.pre=1, end.pre=12, end.post=16, match.out=match.out, match.covar=cov.var, result.var=match.out, omnibus.var=match.out, test="lower", n.cores = min(parallel::detectCores(), 2)) sea1 summary(sea1)
After the call to microsynth has been made, the \code{print} function displays
a brief description of the parameters used in the \code{microsynth} call along
with the results (if available). Also, the \code{summary} function
can be used to display a summary of the matching between treatment, synthetic
control, and the population, and the results table. Below we reproduce the
results that were saved to file in the previous example, with one row for each
of the variables entered to result.var
, which have each been used to calculate
an omnibus statistic (omnibus.var = TRUE
), and two columns corresponding to
the confidence interval (confidence
) resulting from the variance estimator
generated by linearization. The first row of the output (16
) refers to the
maximum post-intervention time used to compile results (end.post
).
Note that the p-value of the omnibus statistic is smaller than any of the individual outcome variables.
plot_microsynth(sea1)
Above are produced plots under default settings. By default, if no other
arguments are declared in the call to plot_microsynth()
, the plots will include one row
for each variable passed to result.var
in the original \code{microsynth} call.
Likewise, values for the duration of the pre- and post-intervention periods
(i.e. start.pre
, end.pre
, end.post
) can also be automatically detected from
the original \code{microsynth} object if not specified manually.
The first plot column compares
the observed outcomes among the treatment, synthetic control, and population
during the pre-intervention and post-intervention periods. Outcomes are scaled
by default (scale.var = "Intercept"
) to the number of treatment units, to
facilitate comparison. The dotted red line indicates the last time period of the
pre-intervention period (end.pre
). Because matching was successful, the treatment
and synthetic control lines track closely during the pre-intervention period; their
divergence during the post-intervention period represents an estimate of the
causal effect of the program (i.e., the red synthetic control line is treated as
the counterfactual to the black treatment line). This difference is charted on
the right plot column.
In addition to using linearization to calculate a variance estimate, microsynth
can approximate the estimator's sampling distribution by generating permuted
placebo groups. When dealing with a large number of treatment and control units,
there is a near infinite number of potential permutations. A default (perm =
250
) is set as permutations are somewhat computationally intensive.
For each placebo, weights are calculated to match the placebo treatment to a new
synthetic control, and an effect is estimated, generating a sampling
distribution and an corresponding p-value. Because the actual treatment area is
a non-random group of treatment units, while the placebo treatments are random
groups, by default microsynth will standardized the placebo treatment effects to
filter out potential design effects (use.survey = TRUE
).
We will also generate jackknife replication groups, using as many groups as the
lesser of the number of cases in the treatment group and the number of cases in
the control group (jack = TRUE
).
The output from this call to microsynth will be largely identical to the previous call, except for the appearance of the right column of plots. Now that permutation groups have been generated, the estimated effect under each of the placebo treatments (gray lines) will be shown along with the estimated effect of the real treatment. This displays the estimated treatment effect in the context of the estimator's sampling distribution.
sea2 <- microsynth(seattledmi, idvar="ID", timevar="time", intvar="Intervention", start.pre=1, end.pre=12, end.post=c(14, 16), match.out=match.out, match.covar=cov.var, result.var=match.out, omnibus.var=match.out, test="lower", perm=250, jack=TRUE, n.cores = min(parallel::detectCores(), 2))
Calling \code{summary} or \code{print} identifies other new changes to the
results. Columns are added to display the confidence intervals
(confidence = 0.9
) and p-values (test = "lower"
) from the jackknife and
permutation tests. Note that end.post=c(14,16)
in the code above, instructing
results to be calculated for two different follow-up periods, ending at t=14
and t=16 respectively. One results table will be calculated for each.
saveReducedMicrosynth(sea2, "../inst/extdata/sea2.rds") sea2 <- readRDS("../inst/extdata/sea2.rds") sea2
plot_microsynth(sea2)
Now, we will add additional outcome variables and also use them to match the treatment area to the synthetic control units. We do this at the risk of model feasibility, as each variable introduces another constraint.
match.out <- c("i_robbery", "i_aggassau", "i_burglary", "i_larceny", "i_felony", "i_misdemea", "i_drugsale", "i_drugposs", "any_crime")
In the example below, without overriding the default weight parameters,
microsynth will fail to find a feasible model. Weights would not be calculated,
and no results or plots will be generated. But we may still attempt to estimate
the model by setting check.feas = TRUE
and use.backup = TRUE
. This will
check for feasibility, and if needed, invoke the computationally intensive
LowRankQP
package to calculate the weights.
Note that the additional matching variables introduce further constraints to the calculation of weights, lengthening the output. Moreover, the introduction of additional time-variant matching variables results in a poorer match on each, shown in the left column of plots, where red and dashed-black lines no longer track perfectly in the pre-intervention period.
Also note that we need not specify values for start.pre
, end.pre
, and
end.post
, as the default settings align with our intentions. Likewise, we can
trust the default values for specifying the variables for the omnibus statistic
(omnibus.var=result.var
) by default. This way we specify the minimum
number of non-default arguments.
sea3 <- microsynth(seattledmi, idvar="ID", timevar="time", intvar="Intervention", end.pre=12, match.out=match.out, match.covar=cov.var, result.var=match.out, perm=250, jack=0, test="lower", check.feas=TRUE, use.backup = TRUE, n.cores = min(parallel::detectCores(), 2))
saveReducedMicrosynth(sea3, "../inst/extdata/sea3.rds") sea3 <- readRDS("../inst/extdata/sea3.rds") summary(sea3)
The results file now shows additional rows for the new outcome variables, and these are also displayed in plots.
plot_microsynth(sea3)
Another potential response is to aggregate sparse variables across multiple time
periods before using them to match to synthetic control. Rather than passing a
vector of variable names to match.out
and/or match.out.min
, the user may
pass a list; each element of the list is a vector corresponding to the time
units across which each variable should be aggregated before matching, with each
element named equal to the variable name. In this case, the element vectors
represent the duration during which the variable should be aggregated, counting
backwards from the intervention time.
In our dataset, incidences of drug sale are relatively scarce, and so are
aggregated every four months before matching ('i_drugsale'=rep(4,3)
);
meanwhile, larceny is relatively common and so is matched un-aggregated
('i_larceny'=rep(1, 12)'
).
Each vector indicates the time-durations for aggregation, starting from the
period directly prior to intervention and finishing with the earliest
observations in the dataset. Because our end.pre = 12
, to use the full
dataset, each vector-element in the list should add to 12. Sums less than 12
would ignore portions of the pre-intervention data; sums more than 12 will throw
an error, calling on more pre-intervention data than are available.
The aggregated variables now appear in the main weights summary table, e.g., "i_robbery.11.12", representing the sum of reported robberies in time periods 11 and 12.
match.out <- list( 'i_robbery'=rep(2, 6), 'i_aggassau'=rep(2, 6), 'i_burglary'=rep(1, 12), 'i_larceny'=rep(1, 12), 'i_felony'=rep(2, 6), 'i_misdemea'=rep(2, 6), 'i_drugsale'=rep(4, 3), 'i_drugposs'=rep(4, 3), 'any_crime'=rep(1, 12)) sea4 <- microsynth(seattledmi, idvar="ID", timevar="time",intvar="Intervention", match.out=match.out, match.covar=cov.var, result.var=names(match.out), omnibus.var=names(match.out), end.pre=12, perm=250, jack = TRUE, test="lower", n.cores = min(parallel::detectCores(), 2))
saveReducedMicrosynth(sea4, "../inst/extdata/sea4.rds") sea4 <- readRDS("../inst/extdata/sea4.rds") summary(sea4)
The aggregation of the outcome variables over time is not directly reflected in the results table, though the estimates have changed as a consequence of the aggregation.
plot_microsynth(sea4)
The following examples will demonstrate that microsynth can be used to calculate weights and variance estimators, produce results, and display charts separately, one at a time. This can be useful given the time-intensive nature of calculating weights and generating permutation groups. It allows for weights to be saved once calculated and for plots and results to be reproduced iteratively without repeating the matching process.
This setting represses reporting of results by setting
result.var
= FALSE. Only weights will be calculated.
Note that settings for permutation groups (perm
) and jackknife replication
groups (jack
) are considered when calculating weights, and then will not be
referred to again in calls that only produce plots or display results.
match.out <- c("i_felony", "i_misdemea", "i_drugs", "any_crime") sea5 <- microsynth(seattledmi, idvar="ID", timevar="time",intvar="Intervention", end.pre=12, match.out=match.out, match.covar=cov.var, result.var=FALSE, perm=0, jack=FALSE, n.cores = min(parallel::detectCores(), 2)) summary(sea5)
Appropriately, the table summarizing the main weights may be viewed, but results are unavailable.
If weights have already been calculated, then microsynth() can also be
configured to only reproduce results. Results are displayed for all outcome
variables used for exact matches (result.var = match.out
). Further, results
can now be calculated for any single or group of follow-up periods
(end.post=c(14,16)
) without having to re-calculate weights.
sea6 <- microsynth(seattledmi, idvar="ID", timevar="time", intvar="Intervention", end.pre=12, end.post=c(14, 16), result.var=match.out, test="lower", w=sea5$w, n.cores = min(parallel::detectCores(), 2)) sea6
For each follow-up period, a separate results table is provided. If saving to file, this requires the file be saved as an XLSX rather than a CSV; each table will be saved to a different XLSX tab.
saveReducedMicrosynth(sea6, "../inst/extdata/sea6.rds") sea6 <- readRDS("../inst/extdata/sea6.rds") sea6
If weights have already been calculated, then plot_microsynth() can be used to
display plots from the original microsynth object. In this case, we limit plots
to a subset of time-variant variables (plot.var=match.out[1:2]
).
plot_microsynth(sea6, plot.var=match.out[1:2])
One of the major differences between Synth and microsynth is that Synth requires that the treatment is confined to a single unit of observation, and to estimating the effect on a single outcome variable; in contrast, microsynth anticipates that treatment has been applied to multiple areas and can estimate effects with respect to multiple outcomes. But microsynth can also be applied to this simpler case.
To demonstrate, first we will create a reduced dataset with 1 treatment block and 100 control blocks.
set.seed(86872) ids.t <- names(table(seattledmi$ID[seattledmi$Intervention==1])) ids.c <- names(table(seattledmi$ID[seattledmi$Intervention==0])) ids.synth <- c(sample(ids.t, 1), sample(ids.c, 100)) seattledmi.one <- seattledmi[is.element(seattledmi$ID, as.numeric(ids.synth)), ]
Then microsynth can be run on the dataset with just a single variable passed out
match.out
, so that effect is estimated for only one variable, as with Synth.
Due to the small size of the reduced dataset, model feasibility may be an issue
(so we set use.backup = TRUE
and check.feas = TRUE
) and variance estimators
will be less reliable.
sea8 <- microsynth(seattledmi.one, idvar="ID", timevar="time", intvar="Intervention", match.out=match.out[4], match.covar=cov.var, result.var=match.out[4], test="lower", perm=250, jack=FALSE, check.feas=TRUE, use.backup=TRUE, n.cores = min(parallel::detectCores(), 2))
saveReducedMicrosynth(sea8, "../inst/extdata/sea8.rds") sea8 <- readRDS("../inst/extdata/sea8.rds") plot_microsynth(sea8) summary(sea8)
microsynth() may also be used to calculate propensity score-type weights. We will demonstrate this by transforming our panel data into a cross-sectional dataset with data corresponding to our final observed period.
seattledmi.cross <- seattledmi[seattledmi$time==16, colnames(seattledmi)!="time"]
By setting match.out = FALSE
, no outcome variables will be used to calculate
weights, only (time-invariant) covariates (match.covar
). No outcome-reporting
variables (result.var = NULL
) need be
reported. Plots are therefore inappropriate, but results (i.e., a summary of
weights only) can be saved to file or viewed using summary
.
sea9 <- microsynth(seattledmi.cross, idvar="ID", intvar="Intervention", match.out=FALSE, match.covar=cov.var, result.var=NULL, test="lower", perm=250, jack=TRUE, n.cores = min(parallel::detectCores(), 2))
saveReducedMicrosynth(sea9, "../inst/extdata/sea9.rds") sea9 <- readRDS("../inst/extdata/sea9.rds") sea9
Abadie A, Gardeazabal J (2003). “The economic costs of conflict: A case study of the Basque Country.” \emph{American Economic Review}, pp. 113-132.
Abadie A, Diamond A, Hainmueller J (2010). “Synthetic control methods for comparative case studies: Estimating the effect of California’s tobacco control program.” \emph{Journal of the American Statistical Association}, 105(490), 493-505.
Any scripts or data that you put into this service are public.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.