| Title: | Quantile Treatment Effects |
|---|---|
| Description: | Provides several methods for computing the Quantile Treatment Effect (QTE) and Quantile Treatment Effect on the Treated (QTT). The main cases covered are (i) treatment is randomly assigned, (ii) treatment is as good as randomly assigned after conditioning on covariates (selection on observables) using the methods of Firpo (2007) <doi:10.1111/j.1468-0262.2007.00738.x>, and (iii) identification is based on a Difference in Differences assumption, with support for several varieties including Athey and Imbens (2006) <doi:10.1111/j.1468-0262.2006.00668.x>, Callaway and Li (2019) <doi:10.3982/QE935>, and Callaway, Li, and Oka (2018) <doi:10.1016/j.jeconom.2018.06.008>. Version 2.0 adds a unified staggered treatment adoption API (built on 'ptetools') for all DiD-based estimators, as well as a new lagged-outcome unconfoundedness estimator ('lou_qtt'). |
| Authors: | Brantly Callaway [aut, cre] |
| Maintainer: | Brantly Callaway <[email protected]> |
| License: | GPL-3 |
| Version: | 2.1.0 |
| Built: | 2026-11-20 16:40:16 UTC |
| Source: | https://github.com/bcallaway11/qte |
Provides several methods for computing the Quantile Treatment Effect (QTE) and Quantile Treatment Effect on the Treated (QTT). The main cases covered are (i) treatment is randomly assigned, (ii) treatment is as good as randomly assigned after conditioning on covariates (selection on observables) using the methods of Firpo (2007) doi:10.1111/j.1468-0262.2007.00738.x, and (iii) identification is based on a Difference in Differences assumption, with support for several varieties including Athey and Imbens (2006) doi:10.1111/j.1468-0262.2006.00668.x, Callaway and Li (2019) doi:10.3982/QE935, and Callaway, Li, and Oka (2018) doi:10.1016/j.jeconom.2018.06.008. Version 2.0 adds a unified staggered treatment adoption API (built on 'ptetools') for all DiD-based estimators, as well as a new lagged-outcome unconfoundedness estimator ('lou_qtt').
Maintainer: Brantly Callaway [email protected]
Authors:
Brantly Callaway [email protected]
Useful links:
Report bugs at https://github.com/bcallaway11/qte/issues
Plot a QTE object as a quantile treatment effect curve
with optional confidence bands.
## S3 method for class 'QTE' autoplot(object, cband = TRUE, ylab = "QTE", ...)## S3 method for class 'QTE' autoplot(object, cband = TRUE, ylab = "QTE", ...)
object |
a |
cband |
logical; if |
ylab |
label for the y-axis. Default |
... |
unused. |
a ggplot object.
Computes the Quantile Treatment Effect on the Treated (QTT) and the Average Treatment Effect on the Treated (ATT) under parallel trends in the distribution functions of untreated potential outcomes, with staggered treatment adoption, for panel data or repeated cross sections.
Identification. For each group and period ,
with the comparison group: parallel trends for the distribution
function rather than the mean (DiD) or the quantile function (QDiD).
Roth and Sant'Anna (2023) show this is the condition under which
parallel trends holds for every transformation of the outcome. It
point-identifies the counterfactual distribution, so the QTT, for
discrete outcomes too, and the implied ATT is the usual DiD ATT.
Covariates make the assumption conditional on .
See cdf_did_gt for how the counterfactual distribution is
estimated (including the rearrangement into a valid distribution
function and the three covariate methods).
cdf_did( yname, gname, tname, idname = NULL, data, panel = TRUE, xformula = ~1, est_method = c("aipw", "ipw", "reg"), weightsname = NULL, clustervars = NULL, control_group = "notyettreated", anticipation = 0, alp = 0.05, cband = TRUE, biters = 100, cl = 1, gt_type = c("att", "qtt"), probs = NULL )cdf_did( yname, gname, tname, idname = NULL, data, panel = TRUE, xformula = ~1, est_method = c("aipw", "ipw", "reg"), weightsname = NULL, clustervars = NULL, control_group = "notyettreated", anticipation = 0, alp = 0.05, cband = TRUE, biters = 100, cl = 1, gt_type = c("att", "qtt"), probs = NULL )
yname |
Name of the outcome variable in |
gname |
Name of the treatment group variable (first treatment period; 0 for never-treated units). |
tname |
Name of the time period variable. |
idname |
Name of the unit id variable. Required when
|
data |
A data frame. |
panel |
Logical. |
xformula |
One-sided formula of covariates. Default |
est_method |
With covariates: |
weightsname |
Name of the column in |
clustervars |
Name of a column in |
control_group |
|
anticipation |
Number of time periods before treatment during which units may anticipate it. Default 0. |
alp |
Significance level. Default 0.05. |
cband |
Logical; uniform confidence bands. Default |
biters |
Number of bootstrap iterations. Default 100. |
cl |
Number of cores for the bootstrap. Default 1. |
gt_type |
|
probs |
For |
For gt_type = "att", a pte_emp_boot object; for
gt_type = "qtt", a pte_qtt object. Both include
outcome_discreteness.
The counterfactual distribution is identified for discrete outcomes as
well, unlike for rank- or quantile-based estimators such as
cic. Quantile effects still jump at mass points and their
bootstrap inference can be unreliable, so gt_type = "qtt" warns
(once per call) when the outcome looks discrete; see
cic.
cic(), qdid(), mdid(), cdf_did() and
lou_qtt(lagged_outcome_cov = FALSE) also work with repeated cross
sections (panel = FALSE; idname not needed). Without
covariates they use only each period's outcome distributions, so on
panel data the estimates are the same either way; the standard errors
differ, because the bootstrap then resamples observations rather than
units. Some things need the same units in several periods and so
require panel data: ddid() (each comparison unit's change over
time), panel_qtt() (the copula between periods), the lagged
outcome in lou_qtt(), and gt_type = "qott" (the
distribution of individual treatment effects).
Roth, Jonathan, and Pedro H. C. Sant'Anna. “When Is Parallel Trends Sensitive to Functional Form?” Econometrica 91(2), pp. 737-747, 2023.
Sant'Anna, Pedro H. C., and Jun Zhao. “Doubly Robust Difference-in-Differences Estimators.” Journal of Econometrics 219(1), pp. 101-122, 2020.
data(mpdta, package = "did") res <- cdf_did(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "qtt", probs = seq(0.1, 0.9, 0.1), biters = 20) summary(res)data(mpdta, package = "did") res <- cdf_did(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "qtt", probs = seq(0.1, 0.9, 0.1), biters = 20) summary(res)
Computes the ATT and the distributions needed for the QTT for
one group-time cell under parallel trends in the distribution functions
of untreated potential outcomes. Used as the attgt_fun in
pte by cdf_did.
cdf_did_gt(gt_data, xformula = ~1, est_method = c("aipw", "ipw", "reg"), ...)cdf_did_gt(gt_data, xformula = ~1, est_method = c("aipw", "ipw", "reg"), ...)
gt_data |
A |
xformula |
One-sided formula of covariates. Default |
est_method |
With covariates: |
... |
Additional arguments passed by |
The counterfactual distribution of untreated outcomes for the treated group in the post period is
evaluated exactly at every point of the pooled support of the three
distributions on the right. The raw need not be a valid
distribution function (it can decrease, or leave 0, 1); it is made one
by monotone rearrangement (sorting its values; Chernozhukov,
Fernandez-Val and Galichon 2010) and clipping to 0, 1. The largest
change this makes is returned as cdf_violation; large values are
evidence against the identifying assumption.
The ATT is the mean of the treated group's observed post-period outcomes
minus the mean of the raw , which is what the assumption
identifies. Without covariates this equals the usual DiD ATT.
With covariates, the comparison group's change in CDFs is averaged over
the treated group's covariate distribution. For each this is a
DiD for the outcome , estimated as in Sant'Anna and
Zhao (2020): by outcome regression (distribution regression of
on the covariates, in each period), by inverse
probability weighting, or doubly robust (AIPW, the default). IPW is
exact on the data's support points; outcome regression and AIPW use the
distribution regression's threshold grid (see cic_gt's
cond_dist = "dr"). The propensity score is a weighted logit of
treatment on the pre-period covariates.
The distribution of individual treatment effects is not identified, so
Fte is NULL.
A ptetools::attgt_noif object with the ATT and, in
extra_gt_returns, F0 (rearranged counterfactual
distribution), F1 (treated post-period distribution),
Fte = NULL, and cdf_violation.
Roth, Jonathan, and Pedro H. C. Sant'Anna. “When Is Parallel Trends Sensitive to Functional Form?” Econometrica 91(2), pp. 737-747, 2023.
Sant'Anna, Pedro H. C., and Jun Zhao. “Doubly Robust Difference-in-Differences Estimators.” Journal of Econometrics 219(1), pp. 101-122, 2020.
Chernozhukov, Victor, Ivan Fernandez-Val, and Alfred Galichon. “Quantile and Probability Curves Without Crossing.” Econometrica 78(3), pp. 1093-1125, 2010.
Computes Quantile Treatment effects on the Treated (QTT) and the Average Treatment Effect on the Treated (ATT) using the Change in Changes identification strategy of Athey and Imbens (2006). Handles two-period data and staggered treatment adoption uniformly: a two-period, two-group dataset is the degenerate single-(g,t) case. Supports both panel and repeated cross sections data.
cic( yname, gname, tname, idname = NULL, data, panel = TRUE, xformula = ~1, cond_dist = c("qr", "dr"), weightsname = NULL, clustervars = NULL, control_group = "notyettreated", anticipation = 0, alp = 0.05, cband = TRUE, biters = 100, cl = 1, ret_quantile = NULL, gt_type = "att", probs = NULL )cic( yname, gname, tname, idname = NULL, data, panel = TRUE, xformula = ~1, cond_dist = c("qr", "dr"), weightsname = NULL, clustervars = NULL, control_group = "notyettreated", anticipation = 0, alp = 0.05, cband = TRUE, biters = 100, cl = 1, ret_quantile = NULL, gt_type = "att", probs = NULL )
yname |
Name of the outcome variable in |
gname |
Name of the treatment group variable (first treatment period; 0 for never-treated units). |
tname |
Name of the time period variable. |
idname |
Name of the unit id variable. Required when
|
data |
A data frame. |
panel |
Logical; |
xformula |
One-sided formula for covariates used in the covariate
adjustment. Default |
cond_dist |
How the conditional distributions are estimated when
|
weightsname |
Name of the column in |
clustervars |
Name of a column in |
control_group |
Which units to use as the comparison group:
|
anticipation |
Number of periods of anticipation. Default |
alp |
Significance level for confidence bands. Default |
cband |
Logical; if |
biters |
Number of bootstrap iterations. Default |
cl |
Number of clusters for parallel computation. Default |
ret_quantile |
Passed through to |
gt_type |
Type of group-time effect to compute. |
probs |
For |
For gt_type = "att", a pte_results object from
ptetools. For gt_type = "qtt", a pte_qtt object
with overall, group-specific, and dynamic QTT curves and bootstrap SEs.
Quantile treatment effects usually require a continuous outcome. With
mass points (e.g. many zeros) or few distinct values, quantiles jump and
effects at quantiles inside a mass point are 0 by construction;
identification based on ranks or quantile functions (change-in-changes,
quantile DiD, the panel and distributional DiD copula arguments) only
bounds the effects (Athey and Imbens 2006, Section 4); and bootstrap
inference for quantiles can be unreliable. The estimators warn (once per
call, class qte_discrete_outcome) when more than 10\
outcome's observations share one value or it takes at most 20 distinct
values, across all groups and periods; the thresholds are the options
qte.mass_point_tol and qte.discrete_n_unique. The
diagnostics are returned as outcome_discreteness.
cdf_did is identified for discrete outcomes.
cic(), qdid(), mdid(), cdf_did() and
lou_qtt(lagged_outcome_cov = FALSE) also work with repeated cross
sections (panel = FALSE; idname not needed). Without
covariates they use only each period's outcome distributions, so on
panel data the estimates are the same either way; the standard errors
differ, because the bootstrap then resamples observations rather than
units. Some things need the same units in several periods and so
require panel data: ddid() (each comparison unit's change over
time), panel_qtt() (the copula between periods), the lagged
outcome in lou_qtt(), and gt_type = "qott" (the
distribution of individual treatment effects).
Athey, Susan and Guido Imbens. “Identification and Inference in Nonlinear Difference-in-Differences Models.” Econometrica 74(2), pp. 431-497, 2006.
data(mpdta, package = "did") ## ATT aggregated across all groups and periods res_att <- cic(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "att", biters = 20) summary(res_att) ## Full QTT curve at selected quantiles res_qtt <- cic(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "qtt", probs = seq(0.1, 0.9, 0.1), biters = 20) summary(res_qtt)data(mpdta, package = "did") ## ATT aggregated across all groups and periods res_att <- cic(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "att", biters = 20) summary(res_att) ## Full QTT curve at selected quantiles res_qtt <- cic(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "qtt", probs = seq(0.1, 0.9, 0.1), biters = 20) summary(res_qtt)
Computes the Change in Changes (CiC) ATT and counterfactual
outcome distribution for a single 2x2 (pre/post x treated/control) data
subset. Serves directly as the attgt_fun argument to
ptetools::pte. Panel vs. repeated cross sections is detected
automatically from whether the same unit ids appear in both periods.
cic_gt(gt_data, xformula = ~1, cond_dist = c("qr", "dr"), ...)cic_gt(gt_data, xformula = ~1, cond_dist = c("qr", "dr"), ...)
gt_data |
A data frame (typically a |
xformula |
One-sided formula for covariates. Default |
cond_dist |
How the conditional distributions are estimated when
|
... |
Additional arguments passed through by |
A ptetools::attgt_noif object with the ATT estimate and, in
extra_gt_returns, three objects: F1 (ECDF of observed
treated outcomes in the post period), F0 (ECDF of counterfactual
untreated outcomes for the treated group), and Fte (ECDF of
individual treatment effects under rank invariance; NULL for
repeated cross sections).
Athey, Susan and Guido Imbens. “Identification and Inference in Nonlinear Difference-in-Differences Models.” Econometrica 74(2), pp. 431-497, 2006.
Chernozhukov, Victor, Ivan Fernandez-Val and Blaise Melly. “Inference on Counterfactual Distributions.” Econometrica 81(6), pp. 2205-2268, 2013.
Computes Quantile Treatment effects on the Treated (QTT) and
the Average Treatment Effect on the Treated (ATT) using the distributional
DiD identification strategy of Callaway, Li, and Oka (2018). Handles
two-period data and staggered treatment adoption uniformly via
ptetools. Requires panel data.
Identification. Under distributional parallel trends and a copula
restriction (the rank correlation of untreated potential outcomes between
the treated and control groups in the pre-period is preserved), the
counterfactual distribution is recovered
by adding each control unit's actual change
to the treated pre-period quantile at that control unit's rank. This
point-identifies the QTT with only two time periods, in contrast to
methods (e.g. panel_qtt) that require three periods.
Panel data required. Unlike cic, qdid, and
mdid, this estimator cannot be applied to repeated cross sections
because it requires observing the within-unit outcome change for each
control unit.
ddid( yname, gname, tname, idname, data, xformula = ~1, cond_dist = c("qr", "dr"), weightsname = NULL, clustervars = NULL, control_group = "notyettreated", anticipation = 0, alp = 0.05, cband = TRUE, biters = 100, cl = 1, gt_type = "att", probs = NULL )ddid( yname, gname, tname, idname, data, xformula = ~1, cond_dist = c("qr", "dr"), weightsname = NULL, clustervars = NULL, control_group = "notyettreated", anticipation = 0, alp = 0.05, cband = TRUE, biters = 100, cl = 1, gt_type = "att", probs = NULL )
yname |
Name of the outcome variable in |
gname |
Name of the treatment group variable (first treatment period; 0 for never-treated units). |
tname |
Name of the time period variable. |
idname |
Name of the unit id variable (required). |
data |
A data frame. |
xformula |
One-sided formula for covariates. Default |
cond_dist |
How conditional distributions are estimated when
|
weightsname |
Name of the column in |
clustervars |
Name of a column in |
control_group |
Which units to use as the comparison group:
|
anticipation |
Number of periods of anticipation. Default |
alp |
Significance level for confidence bands. Default |
cband |
Logical; if |
biters |
Number of bootstrap iterations. Default |
cl |
Number of clusters for parallel computation. Default |
gt_type |
Type of group-time effect to compute. |
probs |
For |
For gt_type = "att", a pte_results object from
ptetools. For gt_type = "qtt", a pte_qtt object.
Quantile treatment effects usually require a continuous outcome. With
mass points (e.g. many zeros) or few distinct values, quantiles jump and
effects at quantiles inside a mass point are 0 by construction;
identification based on ranks or quantile functions (change-in-changes,
quantile DiD, the panel and distributional DiD copula arguments) only
bounds the effects (Athey and Imbens 2006, Section 4); and bootstrap
inference for quantiles can be unreliable. The estimators warn (once per
call, class qte_discrete_outcome) when more than 10\
outcome's observations share one value or it takes at most 20 distinct
values, across all groups and periods; the thresholds are the options
qte.mass_point_tol and qte.discrete_n_unique. The
diagnostics are returned as outcome_discreteness.
cdf_did is identified for discrete outcomes.
cic(), qdid(), mdid(), cdf_did() and
lou_qtt(lagged_outcome_cov = FALSE) also work with repeated cross
sections (panel = FALSE; idname not needed). Without
covariates they use only each period's outcome distributions, so on
panel data the estimates are the same either way; the standard errors
differ, because the bootstrap then resamples observations rather than
units. Some things need the same units in several periods and so
require panel data: ddid() (each comparison unit's change over
time), panel_qtt() (the copula between periods), the lagged
outcome in lou_qtt(), and gt_type = "qott" (the
distribution of individual treatment effects).
Callaway, Brantly, Tong Li, and Tatsushi Oka. “Quantile Treatment Effects in Difference in Differences Models under Dependence Restrictions and with Only Two Time Periods.” Journal of Econometrics 206(2), pp. 395-413, 2018.
data(mpdta, package = "did") ## ATT aggregated across all groups and periods res_att <- ddid(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "att", biters = 20) summary(res_att) ## Full QTT curve at selected quantiles res_qtt <- ddid(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "qtt", probs = seq(0.1, 0.9, 0.1), biters = 20) summary(res_qtt)data(mpdta, package = "did") ## ATT aggregated across all groups and periods res_att <- ddid(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "att", biters = 20) summary(res_att) ## Full QTT curve at selected quantiles res_qtt <- ddid(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "qtt", probs = seq(0.1, 0.9, 0.1), biters = 20) summary(res_qtt)
Computes the distributional DiD ATT and counterfactual outcome
distribution for a single 2x2 (pre/post x treated/control) data subset.
Serves directly as the attgt_fun argument to ptetools::pte.
Identification. Under distributional parallel trends and a copula
restriction (Callaway, Li, and Oka 2018), the counterfactual outcome for
each control unit is
where
is the observed change for control unit , is that unit's rank in the control
pre-period distribution, and is the quantile
function of the treated pre-period distribution. The unconditional
counterfactual distribution is then the
(weighted) empirical CDF of .
Unlike CiC, QDiD, and MDiD, the counterfactual is indexed over
control units, not treated units. Consequently F0 and the
ATT counterfactual term are weighted by w_pre_ctrl, and no
individual treatment effect distribution (Fte) is returned.
Panel data required. The estimator needs the actual change
for each control unit, which requires
observing the same units in both periods.
ddid_gt(gt_data, xformula = ~1, cond_dist = c("qr", "dr"), ...)ddid_gt(gt_data, xformula = ~1, cond_dist = c("qr", "dr"), ...)
gt_data |
A data frame (typically a |
xformula |
One-sided formula for covariates. Default |
cond_dist |
How the conditional distributions are estimated when
|
... |
Additional arguments passed through by |
A ptetools::attgt_noif object with the ATT estimate and, in
extra_gt_returns, F0 (weighted ECDF of counterfactual
outcomes indexed over control units), F1 (weighted ECDF of observed
treated post-period outcomes), and Fte = NULL (individual treatment
effect distribution is not identified for this estimator).
Callaway, Brantly, Tong Li, and Tatsushi Oka. “Quantile Treatment Effects in Difference in Differences Models under Dependence Restrictions and with Only Two Time Periods.” Journal of Econometrics 206(2), pp. 395-413, 2018.
lalonde contains data from the National Supported Work
Demonstration. This program randomly assigned applicants to the job
training program (or out of the job training program). The dataset is
discussed in Lalonde (1986). The experimental part of the dataset is
combined with an observational dataset from the Panel Study of Income
Dynamics (PSID). Lalonde (1986) and many subsequent papers (e.g.
Heckman and Hotz (1989), Dehejia and Wahba (1999), Smith and Todd (2005),
and Firpo (2007) have used this combination to study the effectiveness
of various ‘observational’ methods (e.g. regression, Heckman selection,
Difference in Differences, and propensity score matching) of estimating
the Average Treatment Effect (ATE) of participating in the job training
program. The idea is that the results from the observational method
can be compared to results that can be easily obtained from the
experimental portion of the dataset.
To be clear, the observational data combines the observations that are treated from the experimental portion of the data with untreated observations from the PSID.
data(lalonde)data(lalonde)
Four data.frames: (i) lalonde.exp contains a cross sectional version of the experimental data, (ii) lalonde.psid contains a cross sectional version of the observational data, (iii) lalonde.exp.panel contains a panel version of the experimental data, and (iv) lalonde.psid.panel contains a panel version of the observational data. Note: the cross sectional and panel versions of each dataset are identical up to their shape; in demonstrating each of the methods, it is sometimes convenient to have one form of the data or the other.
LaLonde, Robert. “Evaluating the Econometric Evaluations of
Training Programs with Experimental Data.” The American Economics Review,
pp. 604-620, 1986.
@source The dataset comes from Lalonde (1986) and has been studied in much
subsequent work. The qte package uses a version from the
causalsens package
(https://CRAN.R-project.org/package=causalsens)
The cross sectional verion of the experimental part of the
lalonde dataset. It
is loaded with all the datasets with the command data(lalonde)
The panel verion of the experimental part of the
lalonde dataset. It
is loaded with all the datasets with the command data(lalonde)
The cross sectional verion of the observational part of the
lalonde dataset. It
is loaded with all the datasets with the command data(lalonde)
The panel verion of the observational part of the
lalonde dataset. It
is loaded with all the datasets with the command data(lalonde)
Estimates the Quantile Treatment Effect on the Treated (QTT)
and Average Treatment Effect on the Treated (ATT) under a lagged-outcome
unconfoundedness assumption with staggered treatment adoption. The key
identifying assumption is
, i.e., conditional on
observed covariates and the pre-treatment outcome, treatment is as good
as randomly assigned within each cohort-period cell.
Estimation operates at the (g,t) level: for each cohort and
post-treatment period , a cross-sectional comparison is made
between the treated group and a not-yet-treated (or never-treated)
comparison group, adjusting for covariates and (optionally) the
pre-treatment outcome. Group-time estimates are then aggregated to
overall, dynamic, and group-specific summaries.
Three estimation methods are available:
"ipw"Propensity-score reweighting. Control units are
reweighted by
to approximate the covariate distribution of the treated group.
"reg"Outcome regression. A quantile regression model is
fit on control units' post-period outcomes as a function of
, then predicted at treated units to
construct the counterfactual distribution (Melly 2006;
Chernozhukov, Fernandez-Val, and Melly 2013).
"aipw"Doubly-robust augmented IPW. Combines the propensity score and outcome models. Consistent if either model is correctly specified.
When lagged_outcome_cov = FALSE and xformula = ~1, all
three methods reduce to a simple distribution comparison within each
(g,t) cell, which is consistent under unconditional unconfoundedness.
The ATT analogue (without the QTT) for staggered adoption under lagged-outcome unconfoundedness is developed in Callaway (2023).
lou_qtt( yname, gname, tname, idname = NULL, data, xformula = ~1, lagged_outcome_cov = TRUE, est_method = c("ipw", "reg", "aipw"), cond_dist = c("qr", "dr"), panel = TRUE, weightsname = NULL, clustervars = NULL, control_group = "notyettreated", anticipation = 0, alp = 0.05, cband = TRUE, biters = 100, cl = 1, gt_type = c("att", "qtt"), probs = NULL )lou_qtt( yname, gname, tname, idname = NULL, data, xformula = ~1, lagged_outcome_cov = TRUE, est_method = c("ipw", "reg", "aipw"), cond_dist = c("qr", "dr"), panel = TRUE, weightsname = NULL, clustervars = NULL, control_group = "notyettreated", anticipation = 0, alp = 0.05, cband = TRUE, biters = 100, cl = 1, gt_type = c("att", "qtt"), probs = NULL )
yname |
Name of the outcome variable in |
gname |
Name of the treatment group variable (first treatment period; 0 for never-treated units). |
tname |
Name of the time period variable. |
idname |
Name of the unit id variable. Required when
|
data |
A data frame. |
xformula |
One-sided formula for additional covariates used in the
propensity score and/or outcome model. Default |
lagged_outcome_cov |
Logical; if |
est_method |
Estimation method: |
cond_dist |
For |
panel |
Logical; |
weightsname |
Name of the column in |
clustervars |
Name of a column in |
control_group |
Which units to use as the comparison group:
|
anticipation |
Number of periods of anticipation. Default |
alp |
Significance level for confidence intervals. Default |
cband |
Logical; if |
biters |
Number of bootstrap iterations. Default |
cl |
Number of clusters for parallel bootstrap. Default |
gt_type |
Type of group-time effect to compute. |
probs |
For |
For gt_type = "att", a pte_emp_boot object. For
gt_type = "qtt", a pte_qtt object with overall,
group-specific, and dynamic QTT curves, bootstrap standard errors,
and pointwise and uniform confidence bands.
Quantile treatment effects usually require a continuous outcome. With
mass points (e.g. many zeros) or few distinct values, quantiles jump and
effects at quantiles inside a mass point are 0 by construction;
identification based on ranks or quantile functions (change-in-changes,
quantile DiD, the panel and distributional DiD copula arguments) only
bounds the effects (Athey and Imbens 2006, Section 4); and bootstrap
inference for quantiles can be unreliable. The estimators warn (once per
call, class qte_discrete_outcome) when more than 10\
outcome's observations share one value or it takes at most 20 distinct
values, across all groups and periods; the thresholds are the options
qte.mass_point_tol and qte.discrete_n_unique. The
diagnostics are returned as outcome_discreteness.
cdf_did is identified for discrete outcomes.
cic(), qdid(), mdid(), cdf_did() and
lou_qtt(lagged_outcome_cov = FALSE) also work with repeated cross
sections (panel = FALSE; idname not needed). Without
covariates they use only each period's outcome distributions, so on
panel data the estimates are the same either way; the standard errors
differ, because the bootstrap then resamples observations rather than
units. Some things need the same units in several periods and so
require panel data: ddid() (each comparison unit's change over
time), panel_qtt() (the copula between periods), the lagged
outcome in lou_qtt(), and gt_type = "qott" (the
distribution of individual treatment effects).
Callaway, Brantly. “Policy Evaluation during a Pandemic.” Journal of Econometrics 236(2), 2023.
Melly, Blaise. “Estimation of Counterfactual Distributions Using Quantile Regression.” Working paper, University of St. Gallen, 2006.
Chernozhukov, Victor, Ivan Fernandez-Val, and Blaise Melly. “Inference on Counterfactual Distributions.” Econometrica 81(6), pp. 2205–2268, 2013.
unc_qte for the cross-sectional (non-staggered)
version. cic, qdid, mdid for
alternative identification strategies with staggered adoption.
data(mpdta, package = "did") ## ATT under lagged-outcome unconfoundedness (IPW with pre-period outcome) res_att <- lou_qtt( yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, lagged_outcome_cov = TRUE, est_method = "ipw", gt_type = "att", biters = 20 ) summary(res_att) ## QTT with doubly-robust estimation res_qtt <- lou_qtt( yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, lagged_outcome_cov = TRUE, est_method = "aipw", gt_type = "qtt", probs = seq(0.1, 0.9, 0.1), biters = 20 ) summary(res_qtt)data(mpdta, package = "did") ## ATT under lagged-outcome unconfoundedness (IPW with pre-period outcome) res_att <- lou_qtt( yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, lagged_outcome_cov = TRUE, est_method = "ipw", gt_type = "att", biters = 20 ) summary(res_att) ## QTT with doubly-robust estimation res_qtt <- lou_qtt( yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, lagged_outcome_cov = TRUE, est_method = "aipw", gt_type = "qtt", probs = seq(0.1, 0.9, 0.1), biters = 20 ) summary(res_qtt)
Computes Quantile Treatment effects on the Treated (QTT) and the Average Treatment Effect on the Treated (ATT) using the Mean Difference-in-Differences identification strategy. Handles two-period data and staggered treatment adoption uniformly. Supports both panel and repeated cross sections data.
Identification. MDiD assumes the counterfactual distribution of
untreated potential outcomes is a location shift of the treated group's
pre-treatment distribution. The size of the shift is the mean DiD
, so
.
This is stronger than parallel trends in means alone: it additionally
requires that the shape of the treated group's outcome distribution is
unchanged in the counterfactual. MDiD is a special case of QDiD that
applies a single mean shift rather than a rank-specific distributional
shift.
Covariate adjustment. When xformula is specified, the
scalar shift is replaced by a unit-specific conditional mean shift
,
estimated by weighted OLS. The counterfactual for treated unit is
, and the unconditional counterfactual
distribution is the empirical CDF of these values. By the law of iterated
expectations, this consistently estimates .
mdid( yname, gname, tname, idname = NULL, data, panel = TRUE, xformula = ~1, weightsname = NULL, clustervars = NULL, control_group = "notyettreated", anticipation = 0, alp = 0.05, cband = TRUE, biters = 100, cl = 1, ret_quantile = NULL, gt_type = "att", probs = NULL )mdid( yname, gname, tname, idname = NULL, data, panel = TRUE, xformula = ~1, weightsname = NULL, clustervars = NULL, control_group = "notyettreated", anticipation = 0, alp = 0.05, cband = TRUE, biters = 100, cl = 1, ret_quantile = NULL, gt_type = "att", probs = NULL )
yname |
Name of the outcome variable in |
gname |
Name of the treatment group variable (first treatment period; 0 for never-treated units). |
tname |
Name of the time period variable. |
idname |
Name of the unit id variable. Required when
|
data |
A data frame. |
panel |
Logical; |
xformula |
One-sided formula for covariates. Default |
weightsname |
Name of the column in |
clustervars |
Name of a column in |
control_group |
Which units to use as the comparison group:
|
anticipation |
Number of periods of anticipation. Default |
alp |
Significance level for confidence bands. Default |
cband |
Logical; if |
biters |
Number of bootstrap iterations. Default |
cl |
Number of clusters for parallel computation. Default |
ret_quantile |
Passed through to |
gt_type |
Type of group-time effect to compute. |
probs |
For |
For gt_type = "att", a pte_results object from
ptetools. For gt_type = "qtt", a pte_qtt object
with overall, group-specific, and dynamic QTT curves and bootstrap SEs.
Quantile treatment effects usually require a continuous outcome. With
mass points (e.g. many zeros) or few distinct values, quantiles jump and
effects at quantiles inside a mass point are 0 by construction;
identification based on ranks or quantile functions (change-in-changes,
quantile DiD, the panel and distributional DiD copula arguments) only
bounds the effects (Athey and Imbens 2006, Section 4); and bootstrap
inference for quantiles can be unreliable. The estimators warn (once per
call, class qte_discrete_outcome) when more than 10\
outcome's observations share one value or it takes at most 20 distinct
values, across all groups and periods; the thresholds are the options
qte.mass_point_tol and qte.discrete_n_unique. The
diagnostics are returned as outcome_discreteness.
cdf_did is identified for discrete outcomes.
cic(), qdid(), mdid(), cdf_did() and
lou_qtt(lagged_outcome_cov = FALSE) also work with repeated cross
sections (panel = FALSE; idname not needed). Without
covariates they use only each period's outcome distributions, so on
panel data the estimates are the same either way; the standard errors
differ, because the bootstrap then resamples observations rather than
units. Some things need the same units in several periods and so
require panel data: ddid() (each comparison unit's change over
time), panel_qtt() (the copula between periods), the lagged
outcome in lou_qtt(), and gt_type = "qott" (the
distribution of individual treatment effects).
Athey, Susan and Guido Imbens. “Identification and Inference in Nonlinear Difference-in-Differences Models.” Econometrica 74(2), pp. 431-497, 2006.
Thuysbaert, Bram. “Distributional Comparisons in Difference in Differences Models.” Working Paper, 2007.
data(mpdta, package = "did") ## ATT aggregated across all groups and periods res_att <- mdid(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "att", biters = 20) summary(res_att) ## Full QTT curve at selected quantiles res_qtt <- mdid(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "qtt", probs = seq(0.1, 0.9, 0.1), biters = 20) summary(res_qtt)data(mpdta, package = "did") ## ATT aggregated across all groups and periods res_att <- mdid(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "att", biters = 20) summary(res_att) ## Full QTT curve at selected quantiles res_qtt <- mdid(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "qtt", probs = seq(0.1, 0.9, 0.1), biters = 20) summary(res_qtt)
Computes the MDiD ATT and counterfactual outcome distribution
for a single 2x2 (pre/post x treated/control) data subset. Serves
directly as the attgt_fun argument to ptetools::pte. Panel
vs. repeated cross sections is detected automatically from whether the
same unit ids appear in both periods.
Identification. MDiD assumes that the counterfactual distribution of untreated potential outcomes for the treated group in the post period is a location shift of the treated group's pre-period distribution:
where is
the mean DiD (the change in mean outcomes for the untreated group). The
counterfactual quantile function is therefore
.
Covariate adjustment. With covariates, the scalar shift
is replaced by a unit-specific conditional mean shift
,
estimated by weighted OLS on the control group in each period. The
counterfactual for treated pre-period unit is
. The unconditional counterfactual
distribution is the empirical CDF of these shifted values, which by the
law of iterated expectations consistently estimates
.
mdid_gt(gt_data, xformula = ~1, ...)mdid_gt(gt_data, xformula = ~1, ...)
gt_data |
A data frame (typically a |
xformula |
One-sided formula for covariates. Default |
... |
Additional arguments passed through by |
A ptetools::attgt_noif object with the ATT estimate and, in
extra_gt_returns, three objects: F1 (weighted ECDF of
observed treated outcomes in the post period), F0 (weighted ECDF
of counterfactual untreated outcomes for the treated group), and Fte
(weighted ECDF of individual treatment effects; NULL for repeated
cross sections).
Athey, Susan and Guido Imbens. “Identification and Inference in Nonlinear Difference-in-Differences Models.” Econometrica 74(2), pp. 431-497, 2006.
Thuysbaert, Bram. “Distributional Comparisons in Difference in Differences Models.” Working Paper, 2007.
Computes Quantile Treatment Effects on the Treated (QTT) and
the Average Treatment Effect on the Treated (ATT) using the three-period
panel identification strategy of Callaway and Li (2019). Handles
two-period (one post + two pre-treatment periods) and staggered treatment
adoption via ptetools. Requires panel data for both groups.
Identification. Under a copula stability assumption and
distributional parallel trends on changes, the counterfactual distribution
is recovered from three periods of panel
data without requiring rank invariance.
pre_copula. Controls which pre-treatment periods are used as the copula-transfer base for post-treatment cells (g, g+e):
"long" (default)pre2 is as many periods before pre1 as
the post period is after it, matching the copula window to the event
horizon. pre1 is the last period at least anticipation before
g (in units of the time variable). Cells where pre2 does not exist in
the data are excluded from aggregation with two-step weight
renormalization.
"short"pre2 is the period just before pre1, always anchored at treatment onset.
Pre-test placebo cells always use the sliding base regardless of
pre_copula.
panel_qtt( yname, gname, tname, idname, data, xformula = ~1, cond_dist = c("qr", "dr"), weightsname = NULL, clustervars = NULL, control_group = "notyettreated", anticipation = 0, alp = 0.05, cband = TRUE, biters = 100, cl = 1, gt_type = "att", probs = NULL, pre_copula = "long" )panel_qtt( yname, gname, tname, idname, data, xformula = ~1, cond_dist = c("qr", "dr"), weightsname = NULL, clustervars = NULL, control_group = "notyettreated", anticipation = 0, alp = 0.05, cband = TRUE, biters = 100, cl = 1, gt_type = "att", probs = NULL, pre_copula = "long" )
yname |
Name of the outcome variable. |
gname |
Name of the treatment group variable (first treatment period; 0 for never-treated units). |
tname |
Name of the time period variable. |
idname |
Name of the unit id variable (required). |
data |
A data frame. |
xformula |
One-sided formula for covariates. Default |
cond_dist |
How conditional distributions are estimated when
|
weightsname |
Name of the sampling weights column. Default |
clustervars |
Name of a column in |
control_group |
|
anticipation |
Anticipation, in units of the time variable. Default
|
alp |
Significance level. Default |
cband |
Logical; uniform confidence band if |
biters |
Number of bootstrap iterations. Default |
cl |
Number of parallel clusters. Default |
gt_type |
|
probs |
Quantile grid for |
pre_copula |
|
For gt_type = "att", a pte_emp_boot object. For
gt_type = "qtt", a pte_qtt object.
Quantile treatment effects usually require a continuous outcome. With
mass points (e.g. many zeros) or few distinct values, quantiles jump and
effects at quantiles inside a mass point are 0 by construction;
identification based on ranks or quantile functions (change-in-changes,
quantile DiD, the panel and distributional DiD copula arguments) only
bounds the effects (Athey and Imbens 2006, Section 4); and bootstrap
inference for quantiles can be unreliable. The estimators warn (once per
call, class qte_discrete_outcome) when more than 10\
outcome's observations share one value or it takes at most 20 distinct
values, across all groups and periods; the thresholds are the options
qte.mass_point_tol and qte.discrete_n_unique. The
diagnostics are returned as outcome_discreteness.
cdf_did is identified for discrete outcomes.
cic(), qdid(), mdid(), cdf_did() and
lou_qtt(lagged_outcome_cov = FALSE) also work with repeated cross
sections (panel = FALSE; idname not needed). Without
covariates they use only each period's outcome distributions, so on
panel data the estimates are the same either way; the standard errors
differ, because the bootstrap then resamples observations rather than
units. Some things need the same units in several periods and so
require panel data: ddid() (each comparison unit's change over
time), panel_qtt() (the copula between periods), the lagged
outcome in lou_qtt(), and gt_type = "qott" (the
distribution of individual treatment effects).
Callaway, Brantly and Tong Li. “Quantile Treatment Effects in Difference-in-Differences Models with Panel Data.” Quantitative Economics 10(4), pp. 1579-1618, 2019.
data(mpdta, package = "did") ## Panel QTT with rolling pre-period copula (default pre_copula = "long") res <- panel_qtt(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "qtt", probs = seq(0.1, 0.9, 0.1), biters = 20) summary(res)data(mpdta, package = "did") ## Panel QTT with rolling pre-period copula (default pre_copula = "long") res <- panel_qtt(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "qtt", probs = seq(0.1, 0.9, 0.1), biters = 20) summary(res)
Computes the panel QTT and counterfactual outcome distribution
for a single (g,t) cell using the three-period copula-stability estimator
of Callaway and Li (2019). Serves directly as the attgt_fun
argument to ptetools::pte.
Identification. Under copula stability
and distributional parallel trends on changes
the counterfactual outcome for each treated unit is
where is the treated pre1
quantile at rank
(the Rosenblatt transform), and
is the control change at rank
.
The goal is distributional: is a sample from
, not individual counterfactuals.
Panel data required for both groups. Three periods for treated (pre2, pre1, post) and two for control (pre1, post).
panel_qtt_gt(gt_data, xformula = ~1, cond_dist = c("qr", "dr"), ...)panel_qtt_gt(gt_data, xformula = ~1, cond_dist = c("qr", "dr"), ...)
gt_data |
A |
xformula |
One-sided formula for covariates. Default |
cond_dist |
How the conditional distributions are estimated when
|
... |
Additional arguments passed through by |
A ptetools::attgt_noif object with att and, in
extra_gt_returns, F0 (counterfactual ECDF), F1
(observed post ECDF), and Fte (individual-effect ECDF).
Callaway, Brantly and Tong Li. “Quantile Treatment Effects in Difference-in-Differences Models with Panel Data.” Quantitative Economics 10(4), pp. 1579-1618, 2019.
Plots a QTE object using autoplot.QTE.
## S3 method for class 'QTE' plot(x, cband = TRUE, ylab = "QTE", ...)## S3 method for class 'QTE' plot(x, cband = TRUE, ylab = "QTE", ...)
x |
a |
cband |
logical; if |
ylab |
label for the y-axis. Default |
... |
passed to |
invisibly returns the ggplot object.
Prints a summary.QTE object.
## S3 method for class 'summary.QTE' print(x, ...)## S3 method for class 'summary.QTE' print(x, ...)
x |
A |
... |
unused. |
None. Called for its side effect of printing.
Computes Quantile Treatment effects on the Treated (QTT) and the Average Treatment Effect on the Treated (ATT) using the Quantile Difference-in-Differences identification strategy of Athey and Imbens (2006). Handles two-period data and staggered treatment adoption uniformly. Supports both panel and repeated cross sections data.
qdid( yname, gname, tname, idname = NULL, data, panel = TRUE, xformula = ~1, cond_dist = c("qr", "dr"), weightsname = NULL, clustervars = NULL, control_group = "notyettreated", anticipation = 0, alp = 0.05, cband = TRUE, biters = 100, cl = 1, ret_quantile = NULL, gt_type = "att", probs = NULL )qdid( yname, gname, tname, idname = NULL, data, panel = TRUE, xformula = ~1, cond_dist = c("qr", "dr"), weightsname = NULL, clustervars = NULL, control_group = "notyettreated", anticipation = 0, alp = 0.05, cband = TRUE, biters = 100, cl = 1, ret_quantile = NULL, gt_type = "att", probs = NULL )
yname |
Name of the outcome variable in |
gname |
Name of the treatment group variable (first treatment period; 0 for never-treated units). |
tname |
Name of the time period variable. |
idname |
Name of the unit id variable. Required when
|
data |
A data frame. |
panel |
Logical; |
xformula |
One-sided formula for covariates used in the covariate
adjustment. Default |
cond_dist |
How conditional distributions are estimated when
|
weightsname |
Name of the column in |
clustervars |
Name of a column in |
control_group |
Which units to use as the comparison group:
|
anticipation |
Number of periods of anticipation. Default |
alp |
Significance level for confidence bands. Default |
cband |
Logical; if |
biters |
Number of bootstrap iterations. Default |
cl |
Number of clusters for parallel computation. Default |
ret_quantile |
Passed through to |
gt_type |
Type of group-time effect to compute. |
probs |
For |
For gt_type = "att", a pte_results object from
ptetools. For gt_type = "qtt", a pte_qtt object
with overall, group-specific, and dynamic QTT curves and bootstrap SEs.
Quantile treatment effects usually require a continuous outcome. With
mass points (e.g. many zeros) or few distinct values, quantiles jump and
effects at quantiles inside a mass point are 0 by construction;
identification based on ranks or quantile functions (change-in-changes,
quantile DiD, the panel and distributional DiD copula arguments) only
bounds the effects (Athey and Imbens 2006, Section 4); and bootstrap
inference for quantiles can be unreliable. The estimators warn (once per
call, class qte_discrete_outcome) when more than 10\
outcome's observations share one value or it takes at most 20 distinct
values, across all groups and periods; the thresholds are the options
qte.mass_point_tol and qte.discrete_n_unique. The
diagnostics are returned as outcome_discreteness.
cdf_did is identified for discrete outcomes.
cic(), qdid(), mdid(), cdf_did() and
lou_qtt(lagged_outcome_cov = FALSE) also work with repeated cross
sections (panel = FALSE; idname not needed). Without
covariates they use only each period's outcome distributions, so on
panel data the estimates are the same either way; the standard errors
differ, because the bootstrap then resamples observations rather than
units. Some things need the same units in several periods and so
require panel data: ddid() (each comparison unit's change over
time), panel_qtt() (the copula between periods), the lagged
outcome in lou_qtt(), and gt_type = "qott" (the
distribution of individual treatment effects).
Athey, Susan and Guido Imbens. “Identification and Inference in Nonlinear Difference-in-Differences Models.” Econometrica 74(2), pp. 431-497, 2006.
data(mpdta, package = "did") ## ATT aggregated across all groups and periods res_att <- qdid(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "att", biters = 20) summary(res_att) ## Full QTT curve at selected quantiles res_qtt <- qdid(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "qtt", probs = seq(0.1, 0.9, 0.1), biters = 20) summary(res_qtt)data(mpdta, package = "did") ## ATT aggregated across all groups and periods res_att <- qdid(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "att", biters = 20) summary(res_att) ## Full QTT curve at selected quantiles res_qtt <- qdid(yname = "lemp", gname = "first.treat", tname = "year", idname = "countyreal", data = mpdta, gt_type = "qtt", probs = seq(0.1, 0.9, 0.1), biters = 20) summary(res_qtt)
Computes the QDiD ATT and counterfactual outcome distribution
for a single 2x2 (pre/post x treated/control) data subset. Serves
directly as the attgt_fun argument to ptetools::pte. Panel
vs. repeated cross sections is detected automatically from whether the
same unit ids appear in both periods.
qdid_gt(gt_data, xformula = ~1, cond_dist = c("qr", "dr"), ...)qdid_gt(gt_data, xformula = ~1, cond_dist = c("qr", "dr"), ...)
gt_data |
A data frame (typically a |
xformula |
One-sided formula for covariates. Default |
cond_dist |
How the conditional distributions are estimated when
|
... |
Additional arguments passed through by |
A ptetools::attgt_noif object with the ATT estimate and, in
extra_gt_returns, three objects: F1 (weighted ECDF of
observed treated outcomes in the post period), F0 (weighted ECDF
of counterfactual untreated outcomes for the treated group), and Fte
(weighted ECDF of individual treatment effects; NULL for repeated
cross sections).
Athey, Susan and Guido Imbens. “Identification and Inference in Nonlinear Difference-in-Differences Models.” Econometrica 74(2), pp. 431-497, 2006.
summary.QTE summarizes a QTE object, returning
formatted data frames for the overall ATE and the QTE table suitable for
printing.
## S3 method for class 'QTE' summary(object, ...)## S3 method for class 'QTE' summary(object, ...)
object |
A |
... |
unused. |
A summary.QTE object (a list with overall_ate and
qte_table data frames).
Estimates the Quantile Treatment Effect (QTE) or Quantile
Treatment Effect on the Treated (QTT) under unconfoundedness, also known
as selection on observables, with cross-sectional data. The key
identifying assumption is , i.e.,
potential outcomes are independent of treatment conditional on
covariates . For staggered treatment adoption with panel or
repeated cross-section data, see lou_qtt.
Three estimation methods are available via est_method:
"ipw"Propensity-score reweighting (Firpo 2007). The
propensity score is estimated by logit or
probit and used to reweight the sample so that the covariate
distributions of the treated and untreated groups match. For the QTE,
both groups are reweighted toward the population distribution; for the
QTT, only the untreated group is reweighted toward the treated
covariate distribution.
"reg"Outcome regression via quantile regression inversion.
rq is fit on a dense internal grid of
values, yielding a conditional quantile function
for each arm . The marginal
distribution is recovered by averaging conditional quantile functions
over the empirical covariate distribution (Melly 2006; Chernozhukov,
Fernandez-Val, and Melly 2013).
"aipw"Doubly-robust augmented IPW. Combines the propensity score model with the conditional quantile outcome model. The CDF estimator is consistent if either the propensity score or the outcome model is correctly specified (semiparametric efficiency when both are correct).
When xformula = ~1 (the default), all three methods reduce to
simple quantile differences between the treated and untreated groups, which
is consistent under unconditional unconfoundedness. Covariates are required
for covariate-adjusted estimation.
Rows with a missing value in the outcome, treatment, weights or any
covariate in xformula are dropped, with a warning.
Standard errors and confidence intervals are computed via the empirical
bootstrap, resampling rows of data, or whole clusters when
clustervars is given.
unc_qte( yname, dname, data, xformula = ~1, weightsname = NULL, clustervars = NULL, probs = seq(0.05, 0.95, 0.05), alp = 0.05, biters = 100, cband = TRUE, boot_type = "empirical", method = c("logit", "probit"), est_method = c("ipw", "reg", "aipw"), target = c("qte", "qtt"), cond_dist = c("qr", "dr"), cl = 1, xformla = NULL )unc_qte( yname, dname, data, xformula = ~1, weightsname = NULL, clustervars = NULL, probs = seq(0.05, 0.95, 0.05), alp = 0.05, biters = 100, cband = TRUE, boot_type = "empirical", method = c("logit", "probit"), est_method = c("ipw", "reg", "aipw"), target = c("qte", "qtt"), cond_dist = c("qr", "dr"), cl = 1, xformla = NULL )
yname |
character; name of the outcome variable in |
dname |
character; name of the binary treatment indicator in
|
data |
data.frame containing the analysis data. |
xformula |
one-sided formula for covariates used in the propensity
score and/or outcome model, e.g. |
weightsname |
character; name of a column in |
clustervars |
character; name of a column in |
probs |
numeric vector of quantile levels at which to evaluate the
QTE or QTT. Default |
alp |
significance level for confidence intervals. Default 0.05. |
biters |
number of bootstrap iterations for standard errors and confidence bands. Default 100. |
cband |
logical; if |
boot_type |
bootstrap variant. Currently only |
method |
propensity score link function: |
est_method |
estimation method; one of:
|
target |
target parameter; one of:
|
cond_dist |
For |
cl |
number of cores for parallel bootstrap. Default 1 (sequential). |
xformla |
Deprecated; use |
An object of class QTE containing:
qtenumeric vector of estimated QTE (or QTT) at each
element of probs.
ateestimated ATE (or ATT when target = "qtt").
qte.sebootstrap standard errors for qte (the
standard deviation of the bootstrap draws).
qte.lower, qte.upper
confidence bounds for
qte: a uniform band if cband = TRUE, pointwise
intervals otherwise (see cband).
cbandwhether qte.lower/qte.upper are a
uniform band.
ate.se, ate.lower, ate.upper
SE and CI
for ate.
cbootstrap critical value for a uniform band over
probs.
probsthe probs vector passed by the user.
pscore.regfitted propensity score glm object
(without copies of the data: coef(), vcov(),
summary(), fitted() and predict(newdata = )
work; anova() and update() do not), or
NULL when est_method = "reg".
F.treated.t, F.treated.t.cf
ECDFs of the
estimated quantiles of and at probs.
Use summary() to print a formatted table and plot() to
display the QTE curve with confidence bands.
Quantile treatment effects usually require a continuous outcome. With
mass points (e.g. many zeros) or few distinct values, quantiles jump and
effects at quantiles inside a mass point are 0 by construction;
identification based on ranks or quantile functions (change-in-changes,
quantile DiD, the panel and distributional DiD copula arguments) only
bounds the effects (Athey and Imbens 2006, Section 4); and bootstrap
inference for quantiles can be unreliable. The estimators warn (once per
call, class qte_discrete_outcome) when more than 10\
outcome's observations share one value or it takes at most 20 distinct
values, across all groups and periods; the thresholds are the options
qte.mass_point_tol and qte.discrete_n_unique. The
diagnostics are returned as outcome_discreteness.
cdf_did is identified for discrete outcomes.
Firpo, Sergio. “Efficient Semiparametric Estimation of Quantile Treatment Effects.” Econometrica 75(1), pp. 259–276, 2007.
Melly, Blaise. “Estimation of Counterfactual Distributions Using Quantile Regression.” Working paper, University of St. Gallen, 2006.
Chernozhukov, Victor, Ivan Fernandez-Val, and Blaise Melly. “Inference on Counterfactual Distributions.” Econometrica 81(6), pp. 2205–2268, 2013.
data(lalonde) ## IPW, no covariates (simple quantile differences) q1 <- unc_qte(yname = "re78", dname = "treat", data = lalonde.psid, biters = 20, probs = seq(0.05, 0.95, 0.05)) summary(q1) plot(q1) ## OR with covariates, QTE xf <- ~ age + I(age^2) + education + black + hispanic + married + nodegree q2 <- unc_qte(yname = "re78", dname = "treat", data = lalonde.psid, xformula = xf, est_method = "reg", biters = 20, probs = seq(0.05, 0.95, 0.05)) summary(q2) ## AIPW with covariates, QTT q3 <- unc_qte(yname = "re78", dname = "treat", data = lalonde.psid, xformula = xf, est_method = "aipw", target = "qtt", biters = 20, probs = seq(0.05, 0.95, 0.05)) summary(q3)data(lalonde) ## IPW, no covariates (simple quantile differences) q1 <- unc_qte(yname = "re78", dname = "treat", data = lalonde.psid, biters = 20, probs = seq(0.05, 0.95, 0.05)) summary(q1) plot(q1) ## OR with covariates, QTE xf <- ~ age + I(age^2) + education + black + hispanic + married + nodegree q2 <- unc_qte(yname = "re78", dname = "treat", data = lalonde.psid, xformula = xf, est_method = "reg", biters = 20, probs = seq(0.05, 0.95, 0.05)) summary(q2) ## AIPW with covariates, QTT q3 <- unc_qte(yname = "re78", dname = "treat", data = lalonde.psid, xformula = xf, est_method = "aipw", target = "qtt", biters = 20, probs = seq(0.05, 0.95, 0.05)) summary(q3)