Estimates the effect of a treatment on an outcome, in sharp DID designs with multiple groups and periods.
did_multiplegt_old(
df,
Y,
G,
T,
D,
controls = c(),
placebo = 0,
dynamic = 0,
threshold_stable_treatment = 0,
recat_treatment = NULL,
trends_nonparam = NULL,
trends_lin = NULL,
brep = 0,
cluster = NULL,
covariance = FALSE,
average_effect = NULL,
parallel = FALSE
)did_multiplegt_old returns an object class that has the following values
effect, effect of the treatment
se_effect, standard error of the treatment when bootstraping
N_effect, number of samples used
placebo_i, estimated placebo effect i periods before switchers switch treatment, for all i in 0, 1, ..., k
se_placebo_i, estimated standard error of placebo_i, if the option brep has been specified
N_placebo_i, number of observations used in the estimation of placebo_i
placebo_i, estimated dynamic effect i periods, for all i in 0, 1, ..., k
se_placebo_i, estimated standard error of dynamic_i, if the option brep has been specified
N_placebo_i, number of observations used in the estimation of dynamic_i
the data frame for input
the name of Y variable
the name of group variable
the name of time variable
the name of treatment variable
the list of names of control variables, empty if not specified
the number of placebo estimators to be estimated. Placebo estimators compare switchers' and non-switchers' outcome evolution before switchers' treatment changes. Under the parallel trends assumption underlying the \(DID_M\) estimator, the placebo estimators should not significantly differ from 0. The number of placebos requested can be at most equal to the number of time periods in the data minus 2.
the number of dynamic treatment effects to be estimated. This option should only be used in staggered adoption designs, where each group's treatment is weakly increasing over time, and when treatment is binary. The estimators of dynamic effects are similar to the \(DID_M\) estimator, except that they make use of long differences of the outcome (e.g. from \(t-1\) to \(t+1\)) rather than first differences. The number of dynamic effects requested can be at most equal to the number of time periods in the data minus 2.
this option may be useful when the treatment is continuous, or takes a large number of values. The DIDM estimator uses as controls groups whose treatment does not change between consecutive time periods. With a continuous treatment, there may not be any pair of consecutive time periods between which the treatment of at least one group remains perfectly stable. For instance, if the treatment is rainfall and one uses a county \(\times\) year data set, there is probably not a single county*year whose rainfall is exactly the same as in the same county in the previous year. Then, one needs to specify the threshold_stable_treatment = # option, with # a positive real number. For each pair of consecutive time periods, the command will use counties whose rainfall changed in absolute value by less than # as controls. # should be large enough so that there are counties whose rainfall levels change by less than # between two consecutive years, but it should be small enough so that a change in rainfall of # would be unlikely to affect the outcome.
pools some values of the treatment together when determining the groups whose outcome evolution are compared. This option may be useful when the treatment takes a large number of values, and some very rare in the sample. For instance, assume that treatment D takes the values 0, 1, 2, 3, and 4, but few observations have a treatment equal to 2. Then, there may be a pair of consecutive time periods where one group goes from 2 to 3 units of treatment, but no group has a treatment equal to 2 at both dates. To avoid loosing that observation, one can create a variable D_recat that takes the same value when D=1 or 2 (e.g.: D_recat=(D>=1)+(D>=3)+(D>=4)), and then specify the recat_treatment = "D_recat" option. Then, the command can also use groups with a treatment equal to 1 at two consecutive dates as controls for groups going from 2 to 3 units of treatment, thus making it more likely that all switchers have a non-empty set of controls.
when this option is specified, time fixed effects interacted with varlist are included in the estimation. varlist can only include one categorical variable. For instance, if one works with county \(\times\) year data set and one wants to allow for state-specific trends, then one should write trends_nonparam = "state", where state is the state identifier.
when this option is specified, linear time trends interacted with varlist are included in the estimation. varlist can only include one categorical variable. For instance, if one works with a year data set and one wants to allow for village-specific linear trends, one should write trends_lin = "village", where village is the village identifier. The trends_nonparam= varlist and trends_lin = varlist cannot be specified at the same time.
The number of bootstrap replications to be used in the computation of estimators' standard errors. If the option is specified, did_multiplegt_old returns a graph with all the estimated treatment effects and placebos, and their 95 % confidence intervals constructed using a normal approximation. Otherwise, the command does not compute estimators' standard errors. If the option is specified, it plots a graph with all the estimated treatment effects and placebos, and their 95 % confidence intervals constructed using a normal approximation.
the standard errors of the estimators using a block bootstrap at the varname level. Only one clustering variable is allowed.
if this option and the brep = # option are specified, the command computes the covariances between all the pairs of instantaneous and dynamic effects requested, and between all the pairs of placebos requested. This option can be useful to assess whether some average of the instantaneous and dynamic effects is statistically significant. For instance, assume that you estimate the instantaneous effect, effect_0, and one dynamic effect, effect_1. You would assess whether \(2/3\) effect_0+ \(1/3\) effect_1, a weighted average of those two effects, is statistically significant. You can specify the covariances option, use the fact that Var(\(2/3\) effect_0+\(1/3\) effect_1)=4/9V(effect_0)+1/9V(effect_1)+4/9cov(effect_0,effect_1) to compute the standard error of \(2/3\) effect_0+\(1/3\) effect_1, and finally assess if this average effect is significant. This option can also be useful to run an F-test of whether the placebos are all equal to 0, when several placebos are requested.
if that option is specified, the command will compute an average of the instantaneous and dynamic effects requested. If average_effect = "simple" is specified, the command will compute ple average of the effects and its standard error. If average_effect = "prop_number_switchers" is specified, the command will compute an average where each effect receives a weight proportional to the number of switchers the effect to. When average_effect is specified, the covariances option also has to be specified, and the number of dynamic effects requested should be greater than or equal to 1.
perform bootstrap on multicore if TRUE.
did_multiplegt_old estimates the effect of a treatment on an outcome, using group- (e.g. county- or state-) level panel data with multiple groups and periods. Like other recently proposed DID estimation commands (did, didimputation...), did_multiplegt can be used with a binary and staggered (absorbing) treatment. But unlike those other commands, did_multiplegt_old can also be used with a non-binary treatment (discrete or continuous) that can increase or decrease multiple times. The panel of groups may be unbalanced: not all groups have to be observed at every period (see FAQ section for more info on that). The data may also be at a more disaggregated level than the group level (e.g. individual-level wage data to measure the effect of a regional-level minimum-wage on individuals' wages).
It computes the \(DID_M\) estimator introduced in Section 4 of Chaisemartin and D'Haultfoeuille (2019), which generalizes the standard DID estimator with two groups, two periods and a binary treatment to situations with many groups,many periods and a potentially non-binary treatment. For each pair of consecutive time periods \(t-1\) and \(t\) and for each value of the treatment \(d\), the package computes a \(DID\) estimator comparing the outcome evolution among the switchers, the groups whose treatment changes from \(d\) to some other value between \(t-1\) and \(t\), to the same evolution among control groups whose treatment is equal to \(d\) both in \(t-1\) and \(t\). Then the \(DID_M\) estimator is equal to the average of those \(DID\)s across all pairs of consecutive time periods and across all values of the treatment. Under a parallel trends assumption, \(DID_M\) is an unbiased and consistent estimator of the average treatment effect among switchers, at the time period when they switch.
The package can also compute placebo estimators that can be used to test the parallel trends assumption.
Finally, in staggered adoption designs where each group's treatment is weakly increasing over time, it can compute estimators of switchers' dynamic treatment effects, one time period or more after they have started receiving the treatment.
WARNING: To estimate event-study/dynamic effects, we strongly recommend using the much faster did_multiplegt_dyn command, available from the CRAN repository. In addition to that, did_multiplegt_dyn offers more options than did_multiplegt_old.
# estimating the effect of union membership on wages
# using the same panel of workers as in Vella and Verbeek (1998)
data("wagepan_mgt")
Y = "lwage"
G = "nr"
T = "year"
D = "union"
controls = c("hours")
did_multiplegt_old(wagepan_mgt, Y, G, T, D, controls)
Run the code above in your browser using DataLab