ATTgt

class differences.attgt.attgt.ATTgt(data: DataFrame, cohort_name: str, strata_name: str | None = None, base_period: str = 'varying', anticipation: int = 0, freq: str | None = None)

Difference in differences with

• balanced panels, unbalanced panels or repeated cross-section

• two or multiple periods

• fixed or staggered treatment timing

• binary or multi-valued treatment

• heterogeneous treatment effects

based on the work by [CS2021], [CGS2022], [SZ2020]

Parameters:

• data (DataFrame) –

  pandas DataFrame

  df = df.set_index(['entity', 'time'])
  

  where df is the dataframe to use, ‘entity’ should be replaced with the name of the entity column and ‘time’ should be replaced with the name of the time column.

• cohort_name (str) – cohort name

• base_period (str, default: "varying") –

  • "universal"

  • "varying"

• anticipation (int, default: 0) – The number of time periods before participating in the treatment where units can anticipate participating in the treatment, and therefore it can affect their untreated potential outcomes

• strata_name (str, default: None) –

  The name of the column to be used in case of multi-valued treatment, used to calculate cohort-time-stratum ATT.

  If stratum name is None, fit() will return cohort-time ATT.

• freq (str | None, default: None) – the date frequency of the panel data. Required if the time index is datetime. For example, if the time column is a monthly datetime then freq=’M’. Check offset aliases, for a list of available frequencies.

group_time(feasible: bool = False) → list[dict]

Returns:

• a list of dictionaries where each dictionary keys are

• cohort, base_period, time, (stratum)

fit(formula: str, weights_name: str = None, control_group: str = 'never_treated', base_delta: str | list | dict = 'base', est_method: str | Callable = 'dr', as_repeated_cross_section: bool = None, boot_iterations: int = 0, random_state: int = None, alpha: float = 0.05, cluster_var: list | str = None, split_sample_by: Callable | str | dict = None, n_jobs: int = 1, backend: str = 'loky', progress_bar: bool = True) → DataFrame

Computes the cohort-time-(stratum) average treatment effects:

effects for each cohort, in each time, (for each stratum).

Parameters:

• formula (str) –

  Wilkinson formula for the outcome variable and covariates

  If no covariates the formula must contain only the name of the outcome variable

  # example with covariates
  formula = 'y ~ a + b + a:b'
  
  # example without covariates
  formula = 'y'
  

  Formulas are implemented using formulaic, refer to its documentation for additional details.

• weights_name (str | None, default: None) – The name of the column containing the sampling weights. If None, all observations have same weights.

• control_group (str, default: "never_treated") –

  • "never_treated"

  • "not_yet_treated"

• base_delta (str | list | dict, default: "base") –

  Use base period values for covariates and/or delta values, i.e. the change in value, between the value of covariates at time and the value at base period.

  Available options are:

  • "base"

    the value of each covariate is set to its base period value

  • "delta"

    the value of each time-varying covariate is set to the delta. Time-constant covariates included through x_formula are dropped, and a warning issued.

  • ["base", "delta"] or "base_delta"

    the value of each covariate is set to its base period value, and the value of each time-varying covariate is set to the delta.

  • {'base': ['a', 'b', ..]}

    the value of the specified covariates is set to its base period value, and the value of each time-varying covariate is set to the delta. A warning is issued if x_formula included time-constant covariates that are not included in base_delta.

  • {'delta': ['c', 'd', ..]}

    the value of each covariate is set to its base period value, and the value of the specified time-varying covariates is set to the delta. If the covariates included in ‘delta’ are not time-varying they will be removed from the list.

  • {'base': ['a', 'b', ..], 'delta': ['c', 'd', ..]}

    the value of the specified covariates is set to its base period value, and the value of the specified time-varying covariates is set to the delta. A warning is issued if x_formula included time-constant covariates that are not included in ‘delta’. If the covariates included in ‘delta’ are not time-varying they will be removed from the list.

• est_method (str, default: "dr-mle") –

  • "dr-mle" or "dr"

    for locally efficient doubly robust DiD estimator, with logistic propensity score model for the probability of being treated

  • "dr-ipt"

    for locally efficient doubly robust DiD estimator, with propensity score estimated using the inverse probability tilting

  • "reg"

    for outcome regression DiD estimator

  • "std_ipw-mle" or "std_ipw"

    for standardized inverse probability weighted DiD estimator, with logistic propensity score model for the probability of being treated

• as_repeated_cross_section (bool | None, default: None) –

• boot_iterations (int, default: 0) –

• random_state (int | None, default: None) –

• alpha (float, default: 0.05) – The significance level.

• cluster_var (str | list | None, default: None) –

• split_sample_by (str | Callable | None, default: None) –

  The name of the column along which to split the data, or a function which takes the data and returns a sample mask for a binary split, for example:

  lambda: x = x['column name'] >= x['column name'].median()
  

  The estimation of the ATT will be run separately for each specified sample; used for heterogeneity analysis.

• n_jobs (int, default: 1) –

  The maximum number of concurrently running jobs. If -1 all CPUs are used.

  If ≠ 1, concurrent jobs will be run for two separate tasks:

  • computing the cohort-time ATT; each cohort-time is assigned to a job

  • computing the bootstrap; the influence function is split into n_jobs parts and the boostrap is computed concurrently for each part

  Parallelization is implemented using joblib, refer to its documentation for additional details on n_jobs.

• backend (int, default: "loky") –

  Parallelization backend implementation.

  Parallelization is implemented using joblib, refer to its documentation for additional details on backend.

• progress_bar (bool, default: True) – If True, a progress bar will display the progress over the cohort-times iterations and/or the iterations over the number of boostrap concurrent splits (not the bootstrap iterations).

Return type:

A DataFrame with the group time ATTs

aggregate(type_of_aggregation: str | None = 'simple', overall: bool = False, difference: bool | list | dict[str, list] = False, alpha: float = 0.05, cluster_var: list | str = None, boot_iterations: int = 0, random_state: int = None, n_jobs: int = 1, backend: str = 'loky') → DataFrame

Aggregate the ATTgt

Parameters:

• type_of_aggregation (str | None, default: None) –

  • "simple"

    to calculate the weighted average of all cohort-time average treatment effects, with weights proportional to the cohort size.

  • "event" or "event"

    to calculate the average effects in each relative period: periods relative to the treatment; as in an event study.

  • "cohort"

    to calculate the average treatment effect in each cohort.

  • "time" or "time"

    to calculate the average treatment effect in each time time.

• overall (bool, default: False) –

  calculates the average effect within each type_of_aggregation.

  • if type_of_aggregation is set to "event" or "event"

    to calculate the average effect of the treatment across positive relative periods

  • if type_of_aggregation is set to "cohort"

    to calculate the average effect of the treatment across cohorts

  • if type_of_aggregation is set to "time" or "time"

    to calculate the average effect of the treatment across time times

• difference (bool | list | dict, default: False) –

  take the difference of the estimates

  Available options are:

  • True

    to calculate the difference between 2 samples or 2 strata of treatments

  Note

  • Samples difference: if the estimation is run on 2 samples and more than 2 strata, the estimates for the two samples will be subtracted, as long as there are no strata that have the same names as the samples, in that case use a dictionary as indicated below

  • strata difference: if the estimation is run on 2 strata and more than 2 samples, the estimates for the two strata will be subtracted, as long as there are no samples that have the same names as the strata, in that case use a dictionary as indicated below

  • [sample-0, sample-1] or [stratum-A, stratum-B]

    to calculate the difference between 2 samples listed in the argument or the 2 strata of treatments listed in the argument

  Note

  • Samples difference: if there are strata with the same name as the two samples listed, use a dictionary as indicated below

  • strata difference: if there are samples with the same name as the two strata listed, use a dictionary as indicated below

  • {'strata': [stratum-A, stratum-B]} or {'sample_names': [sample-0, sample-1]}

• alpha (float, default: 0.05) – The significance level.

• cluster_var (str | list | None, default: None) – cluster variables

• boot_iterations (int, default: 0) – bootstrap iterations

• random_state (int | None, default: None) – seed for bootstrap

• n_jobs (int, default: 1) –

  The maximum number of concurrently running jobs. If -1 all CPUs are used.

  If ≠ 1, concurrent jobs will be run for:

  • computing the bootstrap; the influence function is split into n_jobs parts and the boostrap is computed concurrently for each part

  Parallelization is implemented using joblib, refer to its documentation for additional details on n_jobs.

• backend (int, default: "loky") –

  Parallelization backend implementation.

  Parallelization is implemented using joblib, refer to its documentation for additional details on backend.

Return type:

A DataFrame with the requested aggregation

property sample_names

property wald_pre_test

results(type_of_aggregation: str | None = None, overall: bool = False, difference: bool = False, to_dataframe: bool = True, add_info: bool = False)

provides easy access to cached results. this method must be called after fit and/or aggregate depending on the parameters requested

Parameters:

• type_of_aggregation (str | None, default: None) –

  • "simple"

    to return the weighted average of all cohort-time average treatment effects, with weights proportional to the cohort size.

  • "event" or "event"

    to return the average effects in each relative period: periods relative to the treatment; as in an event study.

  • "cohort"

    to return the average treatment effect in each cohort.

  • "time" or "time"

    to return the average treatment effect in each time time.

• overall (bool, default: False) –

  calculates the average effect within each type_of_aggregation.

  • if type_of_aggregation is set to "event" or "event"

    to return the average effect of the treatment across positive relative periods

  • if type_of_aggregation is set to "cohort"

    to return the average effect of the treatment across cohorts

  • if type_of_aggregation is set to "time" or "time"

    to return the average effect of the treatment across time times

• difference (bool, default: False) – to return the most recent estimated difference

• to_dataframe – whether to return the result in a DataFrame or a list of namedtuples

Return type:

Either a pandas dataframe or a list of namedtuples

plot(type_of_aggregation: str | None = None, overall: bool = False, difference: bool = False, estimation_details: bool = True, estimate_in_x_axis: bool = False, **plotting_parameters)

Parameters:

• type_of_aggregation (str | None, default: None) –

  • "simple"

    to plot the weighted average of all cohort-time average treatment effects, with weights proportional to the cohort size.

  • "event" or "event"

    to plot the average effects in each relative period: periods relative to the treatment; as in an event study.

  • "cohort"

    to plot the average treatment effect in each cohort.

  • "time" or "time"

    to plot the average treatment effect in each time time.

• overall (bool, default: False) –

  to plot the average effect within each type_of_aggregation.

  • if type_of_aggregation is set to "event" or "event"

    to plot the average effect of the treatment across positive relative periods

  • if type_of_aggregation is set to "cohort"

    to plot the average effect of the treatment across cohorts

  • if type_of_aggregation is set to "time" or "time"

    to plot the average effect of the treatment across time times

• difference (bool, default: False) –

  take the difference of the estimates

  Available options are:

  • True

    to plot the difference between 2 samples or 2 strata of treatments

• estimation_details (bool | list | str, default: True) – include the estimation details in the plot. One can modify the format through plotting_parameters

• estimate_in_x_axis (bool, default: False) – whether to display the ATT estimates in the x-axis

• plotting_parameters – a set of parameters to customize the plot. Please refer to the separate documentation for the plotting functionalities built in the library

Return type:

An interactive plot for the requested estimates

Citation

[CS2021]

Callaway, Brantly, and Pedro HC Sant’Anna. “Difference-in-differences with multiple time periods.” Journal of Econometrics 225, no. 2 (2021): 200-230.

[CGS2022]

Callaway, Brantly, Andrew Goodman-Bacon, and Pedro HC Sant’Anna. “Difference-in-differences with a continuous treatment.” arXiv preprint arXiv:2107.02637 (2021).

[SZ2020]

Sant’Anna, Pedro HC, and Jun Zhao. “Doubly robust difference-in-differences estimators.” Journal of Econometrics 219, no. 1 (2020): 101-122.