csa-py is an implementation of the Callaway and Sant’Anna
(2021)
(CSA) estimator in Python.
Features include:
-
Inspection of group
$\times$ time effects - Optional
tqdmprogress bar (e.g. when computing many effects on large panels) - Custom group effect aggregation
- Faster than the original
didpackage in R and comparable to thefastdidpackage in R (see benchmarking below). This is made possible by:
uv pip install git+https://github.com/jsr-p/csa-pyWith uv run the following:
uv run --with "git+https://github.com/jsr-p/csa-py" python -i -c 'import csa
import polars as pl
mpdta = pl.read_csv("https://gist.githubusercontent.com/jsr-p/c14075f8ba00bcb56c5cb7a279c82d06/raw/9e1161fbba8ee84b95407ce9a56cecee6be71936/mpdta.csv")
res = csa.estimate(
data=mpdta,
outcome="lemp",
unit="countyreal",
group="first.treat",
time="year",
covariates=["lpop"],
balanced=True,
control="never",
method="reg",
verbose=False,
)
print(res)'and inspect the results in the interactive shell; see balanced panel for more.
Estimating the ATTs:
import polars as pl
import csa
res = csa.estimate(
# data from https://bcallaway11.github.io/did/
data=(mpdta := pl.read_csv("data/mpdta.csv")),
outcome="lemp",
unit="countyreal",
group="first.treat",
time="year",
covariates=["lpop"],
balanced=True,
control="never",
method="reg",
verbose=False,
)
print(res)CsaDidResult(
estimates=pl.DataFrame(shape=(12, 6)),
IF=np.ndarray(shape=(500, 12), dtype=float64),
n=500,
atts=dict(len=12)
)
with pl.Config(tbl_rows=12):
# All ATT(g, t) estimates
print(res.estimates)shape: (12, 6)
┌─────────────┬──────┬───────────┬──────────┬───────────┬───────────┐
│ first.treat ┆ year ┆ att ┆ se ┆ lower ┆ upper │
│ --- ┆ --- ┆ --- ┆ --- ┆ --- ┆ --- │
│ i64 ┆ i64 ┆ f64 ┆ f64 ┆ f64 ┆ f64 │
╞═════════════╪══════╪═══════════╪══════════╪═══════════╪═══════════╡
│ 2004 ┆ 2004 ┆ -0.014911 ┆ 0.022056 ┆ -0.05814 ┆ 0.028317 │
│ 2004 ┆ 2005 ┆ -0.076996 ┆ 0.02836 ┆ -0.13258 ┆ -0.021412 │
│ 2004 ┆ 2006 ┆ -0.14108 ┆ 0.034836 ┆ -0.209358 ┆ -0.072802 │
│ 2004 ┆ 2007 ┆ -0.107544 ┆ 0.032738 ┆ -0.171709 ┆ -0.04338 │
│ 2006 ┆ 2004 ┆ -0.002066 ┆ 0.022122 ┆ -0.045425 ┆ 0.041293 │
│ 2006 ┆ 2005 ┆ -0.006968 ┆ 0.018346 ┆ -0.042925 ┆ 0.028989 │
│ 2006 ┆ 2006 ┆ 0.000766 ┆ 0.019196 ┆ -0.036858 ┆ 0.038389 │
│ 2006 ┆ 2007 ┆ -0.041536 ┆ 0.019717 ┆ -0.08018 ┆ -0.002891 │
│ 2007 ┆ 2004 ┆ 0.026366 ┆ 0.014019 ┆ -0.001111 ┆ 0.053842 │
│ 2007 ┆ 2005 ┆ -0.00476 ┆ 0.01567 ┆ -0.035472 ┆ 0.025953 │
│ 2007 ┆ 2006 ┆ -0.028502 ┆ 0.018132 ┆ -0.06404 ┆ 0.007036 │
│ 2007 ┆ 2007 ┆ -0.028789 ┆ 0.016168 ┆ -0.060478 ┆ 0.002899 │
└─────────────┴──────┴───────────┴──────────┴───────────┴───────────┘
A summary ala. the original R did package is also available as:
res.summary()Group-Time Average Treatment Effects:
Group Time ATT(g, t) Std. Error [95% Pointwise. Conf. Band]
2004 2004 -0.0149 0.0221 -0.0581 0.0283
2004 2005 -0.0770 0.0284 -0.1326 -0.0214 *
2004 2006 -0.1411 0.0348 -0.2094 -0.0728 *
2004 2007 -0.1075 0.0327 -0.1717 -0.0434 *
2006 2004 -0.0021 0.0221 -0.0454 0.0413
2006 2005 -0.0070 0.0183 -0.0429 0.0290
2006 2006 0.0008 0.0192 -0.0369 0.0384
2006 2007 -0.0415 0.0197 -0.0802 -0.0029 *
2007 2004 0.0264 0.0140 -0.0011 0.0538
2007 2005 -0.0048 0.0157 -0.0355 0.0260
2007 2006 -0.0285 0.0181 -0.0640 0.0070
2007 2007 -0.0288 0.0162 -0.0605 0.0029
---
Signif. codes: `*' confidence band does not cover 0
Control group: Never treated
Estimation Method: Outcome Regression
The default estimation method is the Outcome Regression
(method="reg"); to use the double robust estimator use method="dr":
csa.estimate(
data=mpdta,
outcome="lemp",
unit="countyreal",
group="first.treat",
time="year",
covariates=["lpop"],
balanced=True,
control="never",
method="dr",
verbose=False,
).summary()Group-Time Average Treatment Effects:
Group Time ATT(g, t) Std. Error [95% Pointwise. Conf. Band]
2004 2004 -0.0145 0.0221 -0.0579 0.0288
2004 2005 -0.0764 0.0287 -0.1326 -0.0202 *
2004 2006 -0.1404 0.0354 -0.2098 -0.0711 *
2004 2007 -0.1069 0.0329 -0.1714 -0.0424 *
2006 2004 -0.0005 0.0222 -0.0440 0.0431
2006 2005 -0.0062 0.0185 -0.0425 0.0300
2006 2006 0.0010 0.0194 -0.0371 0.0390
2006 2007 -0.0413 0.0197 -0.0799 -0.0026 *
2007 2004 0.0267 0.0141 -0.0008 0.0543
2007 2005 -0.0046 0.0157 -0.0354 0.0262
2007 2006 -0.0284 0.0182 -0.0641 0.0072
2007 2007 -0.0288 0.0162 -0.0606 0.0030
---
Signif. codes: `*' confidence band does not cover 0
Control group: Never treated
Estimation Method: Doubly Robust
Aggregating dynamic effects (with relative horizon
dynamic_te = csa.agg_te(res, method="dynamic")
print(dynamic_te)AggTeResult(
estimates=pl.DataFrame(shape=(7, 5)),
overall_att=-0.08078174533376158,
overall_se=0.018745854711475496,
IF=np.ndarray(shape=(500, 7), dtype=float64),
IF_overall=np.ndarray(shape=(500,), dtype=float64),
method='dynamic',
IFs_weight=dict(len=7)
)
The estimates are shown as:
print(dynamic_te.estimates)shape: (7, 5)
┌─────┬───────────┬──────────┬───────────┬───────────┐
│ k ┆ att ┆ se ┆ lower ┆ upper │
│ --- ┆ --- ┆ --- ┆ --- ┆ --- │
│ i64 ┆ f64 ┆ f64 ┆ f64 ┆ f64 │
╞═════╪═══════════╪══════════╪═══════════╪═══════════╡
│ -3 ┆ 0.026366 ┆ 0.014019 ┆ -0.001111 ┆ 0.053842 │
│ -2 ┆ -0.00413 ┆ 0.012917 ┆ -0.029447 ┆ 0.021188 │
│ -1 ┆ -0.023465 ┆ 0.014442 ┆ -0.05177 ┆ 0.00484 │
│ 0 ┆ -0.021147 ┆ 0.011481 ┆ -0.043649 ┆ 0.001356 │
│ 1 ┆ -0.053356 ┆ 0.016293 ┆ -0.08529 ┆ -0.021422 │
│ 2 ┆ -0.14108 ┆ 0.034836 ┆ -0.209358 ┆ -0.072802 │
│ 3 ┆ -0.107544 ┆ 0.032738 ┆ -0.171709 ┆ -0.04338 │
└─────┴───────────┴──────────┴───────────┴───────────┘
The overall ATT (aggregated across all groups and
print(
f"Overall ATT = {dynamic_te.overall_att:.4f} "
f"(SE = {dynamic_te.overall_se:.4f})",
)Overall ATT = -0.0808 (SE = 0.0187)
A summary ala. the original R did package is also available as:
dynamic_te.summary()Overall summary of ATT's based on event-study/dynamic aggregation:
ATT Std. Error [95% Conf. Band]
-0.0808 0.0187 -0.1175 -0.0440 *
Dynamic effects:
Event time Estimate Std. Error [95% Pointwise Conf. Band]
-3 0.0264 0.0140 -0.0011 0.0538
-2 -0.0041 0.0129 -0.0294 0.0212
-1 -0.0235 0.0144 -0.0518 0.0048
0 -0.0211 0.0115 -0.0436 0.0014
1 -0.0534 0.0163 -0.0853 -0.0214 *
2 -0.1411 0.0348 -0.2094 -0.0728 *
3 -0.1075 0.0327 -0.1717 -0.0434 *
---
Signif. codes: `*' confidence band does not cover 0
Control group: Never treated
Estimation Method: Outcome Regression
We can also bootstrap the dynamic effects to get simultaneous confidence intervals using the multiplier bootstrap (see the algorithm defined on p. 215 in CSA paper and multiplier.py):
import numpy as np
np.random.seed(42)
dynamic_te = csa.agg_te(res, method="dynamic", boot=True)
print(dynamic_te.boot.estimates)shape: (7, 5)
┌─────┬───────────┬──────────┬───────────┬───────────┐
│ k ┆ att ┆ se ┆ lower ┆ upper │
│ --- ┆ --- ┆ --- ┆ --- ┆ --- │
│ i64 ┆ f64 ┆ f64 ┆ f64 ┆ f64 │
╞═════╪═══════════╪══════════╪═══════════╪═══════════╡
│ -3 ┆ 0.026366 ┆ 0.013005 ┆ -0.007303 ┆ 0.060035 │
│ -2 ┆ -0.00413 ┆ 0.01298 ┆ -0.037734 ┆ 0.029475 │
│ -1 ┆ -0.023465 ┆ 0.014593 ┆ -0.061245 ┆ 0.014315 │
│ 0 ┆ -0.021147 ┆ 0.011862 ┆ -0.051856 ┆ 0.009562 │
│ 1 ┆ -0.053356 ┆ 0.016254 ┆ -0.095434 ┆ -0.011278 │
│ 2 ┆ -0.14108 ┆ 0.035916 ┆ -0.234061 ┆ -0.048099 │
│ 3 ┆ -0.107544 ┆ 0.031933 ┆ -0.190213 ┆ -0.024875 │
└─────┴───────────┴──────────┴───────────┴───────────┘
with overall effect:
print(
f"Overall ATT = {dynamic_te.overall_att:.4f} "
f"(SE = {dynamic_te.boot.overall.se.item():.4f})",
)Overall ATT = -0.0808 (SE = 0.0187)
or more succintly:
dynamic_te.summary()Overall summary of ATT's based on event-study/dynamic aggregation:
ATT Std. Error [95% Conf. Band]
-0.0808 0.0187 -0.1163 -0.0453 *
Dynamic effects:
Event time Estimate Std. Error [95% Simult. Conf. Band]
-3 0.0264 0.0140 -0.0011 0.0538
-2 -0.0041 0.0129 -0.0294 0.0212
-1 -0.0235 0.0144 -0.0518 0.0048
0 -0.0211 0.0115 -0.0436 0.0014
1 -0.0534 0.0163 -0.0853 -0.0214 *
2 -0.1411 0.0348 -0.2094 -0.0728 *
3 -0.1075 0.0327 -0.1717 -0.0434 *
---
Signif. codes: `*' confidence band does not cover 0
Control group: Never treated
Estimation Method: Outcome Regression
Note the
95% Simult.vs.95% Pointwisefrom before; the confidence intervals are now simultaneous across all periods at once.
Aggregating group effects:
group_te = csa.agg_te(res, method="group")
print(group_te.estimates)shape: (3, 5)
┌─────────────┬───────────┬──────────┬───────────┬───────────┐
│ first.treat ┆ att ┆ se ┆ lower ┆ upper │
│ --- ┆ --- ┆ --- ┆ --- ┆ --- │
│ i64 ┆ f64 ┆ f64 ┆ f64 ┆ f64 │
╞═════════════╪═══════════╪══════════╪═══════════╪═══════════╡
│ 2004 ┆ -0.085133 ┆ 0.024251 ┆ -0.132664 ┆ -0.037601 │
│ 2006 ┆ -0.020385 ┆ 0.017403 ┆ -0.054493 ┆ 0.013723 │
│ 2007 ┆ -0.028789 ┆ 0.016168 ┆ -0.060478 ┆ 0.002899 │
└─────────────┴───────────┴──────────┴───────────┴───────────┘
group_te.summary()Overall summary of ATT's based on group/cohort aggregation:
ATT Std. Error [95% Conf. Band]
-0.0329 0.0119 -0.0562 -0.0097 *
Group effects:
Group Estimate Std. Error [95% Pointwise Conf. Band]
2004 -0.0851 0.0243 -0.1327 -0.0376 *
2006 -0.0204 0.0174 -0.0545 0.0137
2007 -0.0288 0.0162 -0.0605 0.0029
---
Signif. codes: `*' confidence band does not cover 0
Control group: Never treated
Estimation Method: Outcome Regression
Aggregating calendar effects:
calendar_te = csa.agg_te(res, method="calendar")
print(calendar_te.estimates)shape: (4, 5)
┌──────┬───────────┬──────────┬───────────┬───────────┐
│ year ┆ att ┆ se ┆ lower ┆ upper │
│ --- ┆ --- ┆ --- ┆ --- ┆ --- │
│ i64 ┆ f64 ┆ f64 ┆ f64 ┆ f64 │
╞══════╪═══════════╪══════════╪═══════════╪═══════════╡
│ 2004 ┆ -0.014911 ┆ 0.022056 ┆ -0.05814 ┆ 0.028317 │
│ 2005 ┆ -0.076996 ┆ 0.02836 ┆ -0.13258 ┆ -0.021412 │
│ 2006 ┆ -0.046516 ┆ 0.020998 ┆ -0.087672 ┆ -0.00536 │
│ 2007 ┆ -0.039705 ┆ 0.012866 ┆ -0.064922 ┆ -0.014489 │
└──────┴───────────┴──────────┴───────────┴───────────┘
calendar_te.summary()Overall summary of ATT's based on calendar time aggregation:
ATT Std. Error [95% Conf. Band]
-0.0445 0.0149 -0.0737 -0.0154 *
Time effects:
Time Estimate Std. Error [95% Pointwise Conf. Band]
2004 -0.0149 0.0221 -0.0581 0.0283
2005 -0.0770 0.0284 -0.1326 -0.0214 *
2006 -0.0465 0.0210 -0.0877 -0.0054 *
2007 -0.0397 0.0129 -0.0649 -0.0145 *
---
Signif. codes: `*' confidence band does not cover 0
Control group: Never treated
Estimation Method: Outcome Regression
Aggregating simple effect (i.e. average effect in post-horizon periods,
simple_te = csa.agg_te(res, method="simple")
print(simple_te.estimates)shape: (1, 5)
┌─────────┬───────────┬──────────┬─────────┬───────────┐
│ effect ┆ att ┆ se ┆ lower ┆ upper │
│ --- ┆ --- ┆ --- ┆ --- ┆ --- │
│ str ┆ f64 ┆ f64 ┆ f64 ┆ f64 │
╞═════════╪═══════════╪══════════╪═════════╪═══════════╡
│ overall ┆ -0.041969 ┆ 0.011445 ┆ -0.0644 ┆ -0.019537 │
└─────────┴───────────┴──────────┴─────────┴───────────┘
simple_te.summary() ATT Std. Error [95% Conf. Band]
-0.0420 0.0114 -0.0644 -0.0195 *
Estimating the ATTs on an unbalanced panel with evolution of outcomes as:
The data is simulated with python scripts/sim_test.py simub; see
data_test in Justfile.
data = pl.read_csv("data/testing/sim_ub.csv")
print(data.head())shape: (5, 8)
┌─────┬─────┬─────┬─────┬─────┬───────────┬──────────┬──────────┐
│ id ┆ pt ┆ g ┆ t ┆ K ┆ X ┆ y0 ┆ y │
│ --- ┆ --- ┆ --- ┆ --- ┆ --- ┆ --- ┆ --- ┆ --- │
│ i64 ┆ i64 ┆ i64 ┆ i64 ┆ i64 ┆ f64 ┆ f64 ┆ f64 │
╞═════╪═════╪═════╪═════╪═════╪═══════════╪══════════╪══════════╡
│ 0 ┆ 0 ┆ 6 ┆ 3 ┆ -3 ┆ -0.406317 ┆ 5.833266 ┆ 5.833266 │
│ 0 ┆ 0 ┆ 6 ┆ 4 ┆ -2 ┆ -0.406317 ┆ 3.902666 ┆ 3.902666 │
│ 0 ┆ 0 ┆ 6 ┆ 5 ┆ -1 ┆ -0.406317 ┆ 2.942913 ┆ 2.942913 │
│ 0 ┆ 1 ┆ 6 ┆ 6 ┆ 0 ┆ -0.406317 ┆ 2.697752 ┆ 5.063043 │
│ 0 ┆ 1 ┆ 6 ┆ 7 ┆ 1 ┆ -0.406317 ┆ 2.846276 ┆ 3.785909 │
└─────┴─────┴─────┴─────┴─────┴───────────┴──────────┴──────────┘
res_ub = csa.estimate(
data=data,
group="g",
time="t",
outcome="y",
unit="id",
covariates=["X"],
balanced=False,
control="notyet",
verbose=False,
method="dr",
)
# Estimates are NaN when no valid control group exist for the effect
print(res_ub.estimates)shape: (28, 6)
┌─────┬─────┬───────────┬──────────┬───────────┬───────────┐
│ g ┆ t ┆ att ┆ se ┆ lower ┆ upper │
│ --- ┆ --- ┆ --- ┆ --- ┆ --- ┆ --- │
│ i64 ┆ i64 ┆ f64 ┆ f64 ┆ f64 ┆ f64 │
╞═════╪═════╪═══════════╪══════════╪═══════════╪═══════════╡
│ 4 ┆ 2 ┆ NaN ┆ NaN ┆ NaN ┆ NaN │
│ 4 ┆ 3 ┆ -0.67955 ┆ 0.146673 ┆ -0.967025 ┆ -0.392075 │
│ 4 ┆ 4 ┆ 0.45157 ┆ 0.134338 ┆ 0.188272 ┆ 0.714867 │
│ 4 ┆ 5 ┆ 1.074521 ┆ 0.163479 ┆ 0.754109 ┆ 1.394934 │
│ 4 ┆ 6 ┆ NaN ┆ NaN ┆ NaN ┆ NaN │
│ … ┆ … ┆ … ┆ … ┆ … ┆ … │
│ 7 ┆ 8 ┆ 2.109208 ┆ 0.151619 ┆ 1.81204 ┆ 2.406375 │
│ 8 ┆ 5 ┆ NaN ┆ NaN ┆ NaN ┆ NaN │
│ 8 ┆ 6 ┆ -0.551015 ┆ 0.163906 ┆ -0.872264 ┆ -0.229766 │
│ 8 ┆ 7 ┆ -0.147197 ┆ 0.158528 ┆ -0.457907 ┆ 0.163513 │
│ 8 ┆ 8 ┆ 0.97703 ┆ 0.158914 ┆ 0.665565 ┆ 1.288495 │
└─────┴─────┴───────────┴──────────┴───────────┴───────────┘
# all summaries :=)
csa.agg_te(res_ub, method="dynamic").summary(); print("\n***")
csa.agg_te(res_ub, method="group").summary(); print("\n***")
csa.agg_te(res_ub, method="calendar").summary()Overall summary of ATT's based on event-study/dynamic aggregation:
ATT Std. Error [95% Conf. Band]
1.1244 0.0639 0.9991 1.2497 *
Dynamic effects:
Event time Estimate Std. Error [95% Pointwise Conf. Band]
-2 -0.7474 0.0994 -0.9421 -0.5526 *
-1 -0.5886 0.0686 -0.7232 -0.4541 *
0 0.7776 0.0634 0.6533 0.9020 *
1 1.4711 0.0808 1.3127 1.6296 *
---
Signif. codes: `*' confidence band does not cover 0
Control group: Not yet treated
Estimation Method: Doubly Robust
***
Overall summary of ATT's based on group/cohort aggregation:
ATT Std. Error [95% Conf. Band]
1.0798 0.0636 0.9550 1.2045 *
Group effects:
Group Estimate Std. Error [95% Pointwise Conf. Band]
4 0.7630 0.1264 0.5154 1.0107 *
5 1.1082 0.1271 0.8591 1.3574 *
6 0.9423 0.1430 0.6621 1.2226 *
7 1.5841 0.1178 1.3532 1.8150 *
8 0.9770 0.1589 0.6656 1.2885 *
---
Signif. codes: `*' confidence band does not cover 0
Control group: Not yet treated
Estimation Method: Doubly Robust
***
Overall summary of ATT's based on calendar time aggregation:
ATT Std. Error [95% Conf. Band]
1.0337 0.0620 0.9121 1.1552 *
Time effects:
Time Estimate Std. Error [95% Pointwise Conf. Band]
4 0.4516 0.1343 0.1883 0.7149 *
5 0.9637 0.1173 0.7337 1.1936 *
6 0.9900 0.1191 0.7566 1.2234 *
7 1.1673 0.1148 0.9422 1.3924 *
8 1.5958 0.1261 1.3486 1.8430 *
---
Signif. codes: `*' confidence band does not cover 0
Control group: Not yet treated
Estimation Method: Doubly Robust
csa-py supports inspection of a given
All computed effects are stored in the dictionary atts of the results
object.
As an example, we can inspect the atts map for the estimated effect
for group
g = 2004
t = 2007
print(att := res.atts[(g, t)]) # Inspect ATT(g, t)ATT(
att=np.float64(-0.10754427467304643),
IF=np.ndarray(shape=(329, 1), dtype=float64),
g=2004,
t=2007,
ids=pl.DataFrame(shape=(329, 2))
)
From the ATT object we can see the controls used for the effect:
if isinstance(att, csa.ATT):
# G = 1 treated; G = 0 controls.
print(att.ids.head())
print(att.ids["G"].value_counts())shape: (5, 2)
┌────────────┬─────┐
│ countyreal ┆ G │
│ --- ┆ --- │
│ i64 ┆ i8 │
╞════════════╪═════╡
│ 23 ┆ 0 │
│ 24 ┆ 0 │
│ 25 ┆ 0 │
│ 26 ┆ 0 │
│ 27 ┆ 0 │
└────────────┴─────┘
shape: (2, 2)
┌─────┬───────┐
│ G ┆ count │
│ --- ┆ --- │
│ i8 ┆ u32 │
╞═════╪═══════╡
│ 0 ┆ 309 │
│ 1 ┆ 20 │
└─────┴───────┘
A helper function csa.get_controls extracts the controls for a given
ATT object and CsaDidResult results object:
import typing
# lil' typing hack
att = typing.cast(csa.ATT, att)
controls = csa.get_controls(att, res)
# unit id and group for each control
print(controls.head())shape: (5, 2)
┌────────────┬─────────────┐
│ countyreal ┆ first.treat │
│ --- ┆ --- │
│ i64 ┆ i64 │
╞════════════╪═════════════╡
│ 16017 ┆ 0 │
│ 18109 ┆ 0 │
│ 18035 ┆ 0 │
│ 13093 ┆ 0 │
│ 51515 ┆ 0 │
└────────────┴─────────────┘
print(f"#Controls = {controls.shape[0]}")
print(controls["first.treat"].value_counts())#Controls = 309
shape: (1, 2)
┌─────────────┬───────┐
│ first.treat ┆ count │
│ --- ┆ --- │
│ i64 ┆ u32 │
╞═════════════╪═══════╡
│ 0 ┆ 309 │
└─────────────┴───────┘
From the output we see that only the never-treated (first.treat = 0)
are controls for
Note that implementation wise I set
$G = 0$ for the never-treated although the notation in the paper (and in the literature) is$G = \infty$ .
In the unbalanced panel example it becomes a bit
more interesting. Consider
att = res_ub.atts[(5, 6)]
if isinstance(att, csa.ATT):
controls = csa.get_controls(att, res_ub)
print(controls["g"].value_counts().sort("g"))shape: (3, 2)
┌─────┬───────┐
│ g ┆ count │
│ --- ┆ --- │
│ i64 ┆ u32 │
╞═════╪═══════╡
│ 7 ┆ 182 │
│ 8 ┆ 151 │
│ 9 ┆ 164 │
└─────┴───────┘
I.e. the controls for those with
csa-py allows for aggregating effects into custom made groups (based
on the treatment dates). E.g. can define $$
\mathcal{G}{1} = {2004, 2005}, \ \ \mathcal{G}{2} = {2006, 2007}
$$ and aggregate the effects for those two collection of treatment
groups (the 2005 group doesn’t exist in the mpdta but the point is
still valid).
This can be done as follows:
agg_res = csa.agg_te_custom_group(
res,
# Mapping from unit id to custom group `str` val
custom_group_map={
0: "0",
2004: "2004-2005",
2005: "2004-2005",
2006: "2006-2007",
2007: "2006-2007",
},
verbose=False,
)
print(agg_res.estimates)shape: (2, 5)
┌─────────────┬───────────┬──────────┬───────────┬───────────┐
│ first.treat ┆ att ┆ se ┆ lower ┆ upper │
│ --- ┆ --- ┆ --- ┆ --- ┆ --- │
│ str ┆ f64 ┆ f64 ┆ f64 ┆ f64 │
╞═════════════╪═══════════╪══════════╪═══════════╪═══════════╡
│ 2004-2005 ┆ -0.085133 ┆ 0.024251 ┆ -0.132664 ┆ -0.037601 │
│ 2006-2007 ┆ -0.023187 ┆ 0.012482 ┆ -0.047651 ┆ 0.001278 │
└─────────────┴───────────┴──────────┴───────────┴───────────┘
Thanks tqdm!
- did is the original R package for the CSA estimator
- fastdid is a fast
alternative to the
didpackage in R. Unfortunatelyfastdiddoes not support unbalanced panels when using the double robust estimator. This is a frequent use case in practice when working with microdata and datasets that are “trimmed” i.e. each unit is observed for a fixed window around the event time (here the panel is unbalanced by construction). Thecsa-pypackage supports this use case.
Note: This package was mainly developed to sharpen my own understanding.
- This experiment is akin to the benchmarking experiment conducted in
the
fastdidpackage, see here - To reproduce the figure see
bench-fastdidin the Justfile - Note: in the figure, for
$N \in {10^{2}, 10^{3}, 10^{4}, 10^{5}}$ the number of benchmark iterations equal$25$ ; for$N = 10^{6}$ the number of benchmark iterations equal$10$ . - The fastdid
experiment
also plots the runtime for the original
didpackage; hence this package is also (much) faster than the originaldidpackage in R (also in the unbalanced panel setting)
git clone [email protected]:jsr-p/csa-py.git
cd csa-py
uv venv
uv sync --all-extras