ATTgt#

[1]:

import numpy as np
import pandas as pd

pd.set_option('display.max_rows', 6)

from differences import simulate_data, ATTgt

see plotto to configure the plots. the .plot() method takes the same arguments as the function mark_plot() in plotto

Binary Treatment#

An entity can be either treated or not treated, and once treated it remains treated, the groups are identified by the date of the treatement (cohort)

[2]:

panel_data = simulate_data()  # generate data

panel_data

[2]:
y x0 w cat.0 cat.1 effect cohort intensity
entity time
e0 1900 4.295747 0.160245 3.729827 0 0 0.00 1904.0 5.25
1901 -4.799873 -1.933404 1.691201 0 0 0.00 1904.0 5.25
1902 11.082130 -0.056813 1.786326 0 0 0.00 1904.0 5.25
... ... ... ... ... ... ... ... ... ...
e999 1905 -7.629120 -0.258604 1.801049 0 0 5.25 1904.0 5.25
1906 39.241097 1.504564 0.616657 0 0 10.50 1904.0 5.25
1907 23.619567 0.472740 0.674045 0 0 15.75 1904.0 5.25

8000 rows × 8 columns

[3]:

att_gt = ATTgt(data=panel_data, cohort_name='cohort')

[4]:

att_gt.fit(
    formula='y ~ x0',
)

Computing ATTgt [workers=1]   100%|████████████████████| 14/14 [00:00<00:00, 59.11it/s]

[4]:
ATTgtResult
analytic pointwise conf. band
ATT std_error lower upper zero_not_in_cband
cohort base_period time
1903 1900 1901 0.447031 1.407877 -2.312357 3.206418
1901 1902 0.196806 1.293725 -2.338849 2.732461
1902 1903 2.613823 1.295713 0.074272 5.153373 *
... ... ... ... ... ... ... ...
1904 1903 1905 3.701955 1.352406 1.051288 6.352621 *
1906 10.637781 1.302243 8.085431 13.190131 *
1907 13.856575 1.353726 11.203320 16.509830 *

14 rows × 5 columns

[5]:

att_gt.plot(configure_axisX={'format': 'c'})

[5]:

Aggregate#

[6]:

att_gt.aggregate('time')
att_gt.plot('time', configure_axisX={'format': 'c'})

[6]:

[7]:

att_gt.aggregate('event')
att_gt.plot('event')

[7]:

[8]:

att_gt.aggregate('cohort')
att_gt.plot('cohort', lines=False, configure_axisX={'format': 'c'})

[8]:

[9]:

att_gt.aggregate('simple')

[9]:
SimpleAggregation
analytic pointwise conf. band
ATT std_error lower upper zero_not_in_cband
0 4.008994 0.727817 2.582498 5.43549 * 
[10]:

att_gt.aggregate('simple')

[10]:
SimpleAggregation
analytic pointwise conf. band
ATT std_error lower upper zero_not_in_cband
0 4.008994 0.727817 2.582498 5.43549 * 
[11]:

att_gt.aggregate('event', overall=True)

[11]:
EventAggregationOverall
analytic pointwise conf. band
ATT std_error lower upper zero_not_in_cband
0 3.775662 0.742154 2.321068 5.230257 *

Heterogeneity & Triple Difference#

[12]:

panel_data = simulate_data(samples=3)

[13]:

att_gt = ATTgt(data=panel_data, cohort_name='cohort')

# heterogeneity

att_gt.fit(
    formula='y',
    split_sample_by='samples'
)

Computing ATTgt for samples = 0 [workers=1]100%|████████████████████| 21/21 [00:00<00:00, 119.56it/s]
Computing ATTgt for samples = 1 [workers=1]100%|████████████████████| 21/21 [00:00<00:00, 116.07it/s]
Computing ATTgt for samples = 2 [workers=1]100%|████████████████████| 21/21 [00:00<00:00, 115.64it/s]

[13]:
ATTgtResult
analytic pointwise conf. band
ATT std_error lower upper zero_not_in_cband
sample_name cohort base_period time
samples = 0 1902 1900 1901 -0.413163 2.285396 -4.892458 4.066131
1901 1902 0.016927 2.219508 -4.333228 4.367082
1903 8.410651 2.174143 4.149408 12.671893 *
... ... ... ... ... ... ... ... ...
samples = 2 1905 1904 1905 -2.266016 2.272358 -6.719756 2.187723
1906 4.853377 2.291364 0.362385 9.344369 *
1907 12.146894 2.203472 7.828168 16.465620 *

63 rows × 5 columns

[14]:

att_gt.aggregate('event')

att_gt.plot('event',  vertical_line=0, horizontal_line=0)

[14]:

[15]:

att_gt.aggregate('simple')

[15]:
SimpleAggregation
analytic pointwise conf. band
ATT std_error lower upper zero_not_in_cband
sample_name
samples = 0 11.659525 1.145702 9.413990 13.905061 *
samples = 1 13.036015 1.231696 10.621935 15.450095 *
samples = 2 14.491206 1.284529 11.973576 17.008836 * 
[16]:

att_gt.plot(
    'simple',
    y_title=['Sample', 'Names'],
    vertical_line=0,
    configure_axisY={'domain': True, 'ticks': True},
)

[16]:

Triple Difference#

[17]:

att_gt.sample_names

[17]:

['samples = 0', 'samples = 1', 'samples = 2']

[18]:

att_gt.aggregate('time',  difference=['samples = 1', 'samples = 2'])

[18]:
DifferenceTimeAggregation
analytic pointwise conf. band
ATT std_error lower upper zero_not_in_cband
difference_between time
samples = 1 - samples = 2 1902 0.478123 1.541316 -2.542801 3.499047
1903 5.549908 1.474024 2.660874 8.438943 *
1904 -5.869359 1.375390 -8.565075 -3.173643 *
1905 1.940128 1.286668 -0.581695 4.461951
1906 -1.888593 1.396793 -4.626258 0.849071
1907 -5.348880 1.394556 -8.082160 -2.615600 * 
[19]:

att_gt.plot('time',  difference=True)

[19]:

Multi-valued Treatment#

An entity can be treated with different intensities or not treated, the groups are identified by the date of the treatement (cohort) & the intensity which represents a second dimension.

[20]:

panel_data = simulate_data(intensity_by=2)  # generate data

panel_data

[20]:
y x0 w cat.0 cat.1 effect cohort strata intensity
entity time
e0 1900 14.320181 0.160245 3.729827 0 0 0.0 1904.0 0.0 10.0
1901 1.118288 -1.933404 1.691201 0 0 0.0 1904.0 0.0 10.0
1902 -0.252880 -0.056813 1.786326 0 0 0.0 1904.0 0.0 10.0
... ... ... ... ... ... ... ... ... ... ...
e999 1905 16.700577 -0.258604 1.801049 0 0 0.5 1904.0 1.0 0.5
1906 15.151684 1.504564 0.616657 0 0 1.0 1904.0 1.0 0.5
1907 0.163186 0.472740 0.674045 0 0 1.5 1904.0 1.0 0.5

8000 rows × 9 columns

[21]:

att_gt = ATTgt(data=panel_data, cohort_name='cohort', strata_name='strata')

[22]:

att_gt.fit(formula='y', n_jobs=1)

Computing ATTgt [workers=1]   100%|████████████████████| 42/42 [00:00<00:00, 44.39it/s]

[22]:
ATTgtResult
analytic pointwise conf. band
ATT std_error lower upper zero_not_in_cband
stratum cohort base_period time
0 1901 1900 1901 -1.166568 1.488863 -4.084686 1.751550
1902 9.743155 1.628144 6.552052 12.934258 *
1903 19.473230 1.626157 16.286020 22.660440 *
... ... ... ... ... ... ... ... ...
1 1905 1904 1905 1.454368 1.619994 -1.720762 4.629498
1906 0.030797 1.537388 -2.982428 3.044021
1907 1.231228 1.572191 -1.850210 4.312665

42 rows × 5 columns

Aggregate#

[23]:

att_gt.aggregate('event')

[23]:
EventAggregation
analytic pointwise conf. band
ATT std_error lower upper zero_not_in_cband
stratum relative_period
0 -4 -0.528218 1.521623 -3.510545 2.454109
-3 1.539960 1.009962 -0.439529 3.519448
-2 -0.138531 0.989853 -2.078607 1.801545
... ... ... ... ... ... ...
1 4 3.188915 1.576782 0.098478 6.279352 *
5 2.272094 1.596461 -0.856913 5.401100
6 3.896055 1.626878 0.707432 7.084679 *

22 rows × 5 columns

[24]:

att_gt.plot('event')

[24]:

[25]:

att_gt.aggregate('event', difference=[0, 1], boot_iterations=5000)

Bootstrap                     100%|████████████████████| [00:00<00:00, 10926.23it/s]

[25]:
DifferenceEventAggregation
bootstrap simult. conf. band
ATT std_error lower upper zero_not_in_cband
difference_between relative_period
0 - 1 -4 -1.121424 1.702570 -5.902428 3.659580
-3 1.860588 1.272509 -1.712757 5.433934
-2 -0.597033 1.240433 -4.080306 2.886239
... ... ... ... ... ...
4 37.084082 1.716253 32.264655 41.903510 *
5 47.373146 1.620863 42.821584 51.924708 *
6 55.768956 1.691237 51.019777 60.518135 *

11 rows × 5 columns

[26]:

att_gt.plot('event', difference=True)

[26]:

[27]:

att_gt.aggregate('simple')
att_gt.plot('simple')

[27]:

Heterogeneity & Triple Difference#

[28]:

panel_data = simulate_data(intensity_by=2, samples=3)  # generate data

panel_data

[28]:
y x0 w cat.0 cat.1 effect cohort samples strata intensity
entity time
e0 1900 2.347452 0.160245 3.729827 0 0 0.0 1906.0 2 1.0 4.3
1901 -2.880331 -1.933404 1.691201 0 0 0.0 1906.0 2 1.0 4.3
1902 -6.399243 -0.056813 1.786326 0 0 0.0 1906.0 2 1.0 4.3
... ... ... ... ... ... ... ... ... ... ... ...
e999 1905 -8.514995 -0.258604 1.801049 0 0 0.0 1906.0 2 1.0 4.3
1906 3.713039 1.504564 0.616657 0 0 0.0 1906.0 2 1.0 4.3
1907 26.942832 0.472740 0.674045 0 0 4.3 1906.0 2 1.0 4.3

8000 rows × 10 columns

[29]:

att_gt = ATTgt(data=panel_data, cohort_name='cohort', strata_name='strata')

[30]:

att_gt.fit(formula='y', split_sample_by='samples')

Computing ATTgt for samples = 2 [workers=1]100%|████████████████████| 42/42 [00:00<00:00, 92.14it/s]
Computing ATTgt for samples = 0 [workers=1]100%|████████████████████| 42/42 [00:00<00:00, 95.82it/s]
Computing ATTgt for samples = 1 [workers=1]100%|████████████████████| 42/42 [00:00<00:00, 95.95it/s]

[30]:
ATTgtResult
analytic pointwise conf. band
ATT std_error lower upper zero_not_in_cband
sample_name stratum cohort base_period time
samples = 2 0 1901 1900 1901 -2.313173 2.816396 -7.833208 3.206863
1902 3.324637 2.912590 -2.383933 9.033208
1903 11.378922 2.777983 5.934176 16.823669 *
... ... ... ... ... ... ... ... ... ...
samples = 1 1 1906 1904 1905 2.983248 2.933005 -2.765336 8.731831
1905 1906 -0.334297 2.554715 -5.341447 4.672852
1907 0.343971 3.023992 -5.582944 6.270887

126 rows × 5 columns

[31]:

att_gt.aggregate('time')

att_gt.plot(
    'time',
    configure_header={
        'titleFontSize': 18,
        'labelFontSize': 15
    },
    estimation_details=False,
    configure_axisX={'format':'c'},
)

[31]:

[32]:

att_gt.aggregate('time', difference=[0, 1])

[32]:
DifferenceTimeAggregation
analytic pointwise conf. band
ATT std_error lower upper zero_not_in_cband
sample_name difference_between time
samples = 2 0 - 1 1901 -0.366749 2.745978 -5.748767 5.015269
1902 0.689242 3.101619 -5.389819 6.768304
1903 1.007773 2.002671 -2.917389 4.932936
... ... ... ... ... ... ... ...
samples = 1 0 - 1 1905 5.320888 2.212145 0.985162 9.656613 *
1906 4.642867 1.751682 1.209633 8.076101 *
1907 6.223082 1.908472 2.482545 9.963619 *

21 rows × 5 columns

[33]:

att_gt.sample_names

[33]:

['samples = 2', 'samples = 0', 'samples = 1']

[34]:

att_gt.aggregate('time', difference=['samples = 2', 'samples = 0'])

[34]:
DifferenceTimeAggregation
analytic pointwise conf. band
ATT std_error lower upper zero_not_in_cband
stratum difference_between time
0 samples = 2 - samples = 0 1901 1.470721 2.102341 -2.649793 5.591234
1902 -10.006937 2.011322 -13.949056 -6.064819 *
1903 -6.449823 1.612146 -9.609571 -3.290075 *
... ... ... ... ... ... ... ...
1 samples = 2 - samples = 0 1905 -8.045895 1.515554 -11.016327 -5.075464 *
1906 -4.207633 1.397745 -6.947163 -1.468103 *
1907 -4.728102 1.423883 -7.518862 -1.937342 *

14 rows × 5 columns

Custom Estimation & DoubleML#

[35]:

from sklearn.ensemble import RandomForestClassifier, RandomForestRegressor

# outcome model
outcome_model = RandomForestRegressor(
    n_estimators=100,
    max_depth=2)

# propensity score model
pscore_model = RandomForestClassifier(
    n_estimators=100,
    max_depth=2)

[36]:

from functools import partial
from differences.did.double_ml import aiptw_double_ml_did_panel

# use partial to set those models for cross-fitting

aiptw = partial(
    aiptw_double_ml_did_panel,

    pscore_model=pscore_model,
    outcome_model=outcome_model,
    outcome_type='continuous'
)

[37]:

# now, instead of selecting from the menu of options available for the est_method,
# we plug in the function we composed above

att_gt.fit(
    formula='y ~ x0',
    n_jobs=-1,

    est_method=aiptw,  # <--- plug in here
)

Computing ATTgt [workers=10]  100%|████████████████████| 42/42 [00:27<00:00,  1.53it/s]

[37]:
ATTgtResult
analytic pointwise conf. band
ATT std_error lower upper zero_not_in_cband
stratum cohort base_period time
0 1901 1900 1901 -2.809343 1.806964 -6.350927 0.732242
1902 5.300236 1.724092 1.921077 8.679394 *
1903 10.981719 1.759577 7.533012 14.430427 *
... ... ... ... ... ... ... ... ...
1 1906 1904 1905 -2.160258 1.731272 -5.553488 1.232972
1905 1906 0.015155 1.767796 -3.449661 3.479971
1907 4.812429 1.722523 1.436346 8.188512 *

42 rows × 5 columns