Reference

This page documents every public function and method exposed by the csdid package along with the meaning of every argument. Use it to look up where an option goes (constructor, fit, aggte, etc.) and what the default behavior is.

The package is organized around the ATTgt class. A typical workflow is:

  1. Build an estimator object with csdid.att_gt.ATTgt — pass in the data and identifier columns.

  2. Call csdid.att_gt.ATTgt.fit() — choose the 2x2 DiD estimator and the base period.

  3. Inspect the disaggregated group-time effects with csdid.att_gt.ATTgt.summ_attgt() or plot them with csdid.att_gt.ATTgt.plot_attgt().

  4. Aggregate to a smaller number of parameters with csdid.att_gt.ATTgt.aggte() and plot with csdid.att_gt.ATTgt.plot_aggte().

csdid.att_gt.ATTgt

class ATTgt(yname, tname, idname, gname, data, control_group=['nevertreated', 'notyettreated'], xformla=None, panel=True, allow_unbalanced_panel=True, clustervar=None, weights_name=None, anticipation=0, cband=False, biters=1000, alp=0.05)

Group-time average treatment effects estimator. Sets up the design from the long-format panel (or repeated cross-section), runs the pre-processing step, and stores everything needed for fit.

Parameters:

  • yname (str) – Name of the outcome column in data.

  • tname (str) – Name of the time column in data.

  • idname (str) – Name of the unit identifier column. Required for panel data; ignored when panel=False.

  • gname (str) – Name of the group column. This is the period in which a unit is first treated. Never-treated units should have gname = 0.

  • data (pandas.DataFrame) – The long-format dataset.

  • control_group (str or list of str) – Which units to use as controls. Either "nevertreated" (default) or "notyettreated". The not-yet-treated group is at least as large as the never-treated group and changes across time. A list ['nevertreated', 'notyettreated'] resolves to the first entry.

  • xformla (str) – Patsy-style formula for the covariates, e.g. "lemp ~ lpop". The left-hand side is ignored; only the right-hand side is used to build the design matrix. None (or "y~1") requests unconditional parallel trends.

  • panel (bool) – True (default) treats the data as a panel. False requests repeated cross-section mode (idname is then ignored).

  • allow_unbalanced_panel (bool) – When True (default), the package does not drop units with missing observations in some periods. Set to False to force a balanced panel.

  • clustervar (str) – Name of an extra clustering variable. Up to two clustering variables are supported and one of them must be idname.

  • weights_name (str) – Name of a column with sampling weights. If None, equal weights are used.

  • anticipation (int) – Number of pre-treatment periods in which units are allowed to anticipate the treatment. Default 0.

  • cband (bool) – When True, store enough information for uniform confidence bands. Default False.

  • biters (int) – Number of multiplier-bootstrap iterations used for standard errors. Default 1000.

  • alp (float) – Significance level for confidence bands. Default 0.05.

fit(est_method='dr', base_period='varying', bstrap=True)

Estimate group-time average treatment effects.

Parameters:

  • est_method (str) –

    The 2x2 DiD estimator used for each (g, t) cell. One of:

    • "dr" (default) — Sant’Anna and Zhao doubly-robust estimator.

    • "ipw" — inverse-probability-weighted estimator.

    • "reg" — outcome-regression estimator.

  • base_period (str) –

    Reference period used to construct each \(ATT(g,t)\).

    • "varying" (default) — the reference period changes with t. For pre-treatment periods it is t - 1; for post-treatment periods it is the last period before treatment for group g.

    • "universal" — a single, fixed reference period is used for each group (the last pre-treatment period, after anticipation). The entry at the base period itself is normalized to 0. With this option the result table contains one extra row per group.

  • bstrap (bool) – When True (default), standard errors and critical values come from the multiplier bootstrap. When False, asymptotic standard errors and a normal critical value (1.96) are used.

Returns:

self (the call is chainable). After fitting, self.results, self.MP, and self.did_object are populated.

summ_attgt(n=4)

Build a tidy DataFrame of the ATT(g,t) table.

Parameters:

n (int) – Number of decimals to round to. Default 4.

Returns:

self with the table available as self.summary2.

aggte(typec='group', balance_e=None, min_e=-inf, max_e=inf, na_rm=False, bstrap=None, biters=None, cband=None, alp=None, clustervars=None)

Aggregate the group-time effects into a smaller number of parameters. Returns self; the aggregated object is stored on self.atte.

Parameters:

  • typec (str) –

    Type of aggregation:

    • "simple" — weighted average of all post-treatment ATT(g,t) (weights proportional to group size).

    • "dynamic" — average effects by length of exposure (event-study).

    • "group" (default) — average effect per treatment cohort.

    • "calendar" — average effect per calendar period.

  • balance_e (int) – Only used when typec='dynamic'. If set, drops cohorts that are not observed for at least balance_e + 1 post-treatment periods, so the composition is constant in event time. Default None (no balancing).

  • min_e (float) – Smallest event time to include in the dynamic aggregation. Default -inf.

  • max_e (float) – Largest event time to include. Default inf.

  • na_rm (bool) – When True, drop missing aggregated estimates. Default False.

  • bstrap (bool) – Override the bstrap setting stored on the fitted object. None (default) inherits from fit.

  • biters (int) – Override the number of bootstrap iterations.

  • cband (bool) – Compute uniform confidence bands (requires bstrap=True). None inherits.

  • alp (float) – Override the significance level.

  • clustervars (list) – Override the clustering variables.

plot_attgt(ylim=None, xlab=None, ylab=None, title='Group', xgap=1, ncol=1, legend=True, group=None, ref_line=0, theming=True, grtitle='Group')

Faceted plot of the group-time treatment effects, one subplot per cohort.

Parameters:

  • ylim (tuple) – (ymin, ymax) for the y-axis.

  • xlab (str) – Label for the x-axis.

  • ylab (str) – Label for the y-axis.

  • title (str) – Subplot title prefix. Default "Group".

  • xgap (int) – Spacing between x-axis ticks. Default 1.

  • ncol (int) – Number of columns in the facet grid. Default 1.

  • legend (bool) – Whether to draw the legend. Default True.

  • group (list) – Subset of groups to plot. None plots all.

  • ref_line (float) – y-value of the reference line. Default 0.

  • theming (bool) – Apply the package’s default styling. Default True.

  • grtitle (str) – Per-subplot title prefix (used together with the group value). Default "Group".

plot_aggte(ylim=None, xlab=None, ylab=None, title='', xgap=1, legend=True, ref_line=0, theming=True, **kwargs)

Plot the aggregated object stored on self.atte (produced by a prior call to aggte()). For typec='group' it draws a splot; for the other types it draws a gplot.

Parameters:

  • ylim (tuple) – y-axis limits.

  • xlab (str) – x-axis label.

  • ylab (str) – y-axis label.

  • title (str) – Plot title. Empty defaults to "Average Effect by Group" (group aggregation) or "Average Effect by Length of Exposure" (dynamic / calendar).

  • xgap (int) – Tick spacing on the x-axis.

  • legend (bool) – Show legend.

  • ref_line (float) – y-value of the reference line.

  • theming (bool) – Apply default styling.

Notes on argument inheritance

Some arguments can be specified in more than one place (for example bstrap, biters, cband, alp). The resolution order is:

  1. Value passed to ATTgt.aggte() overrides the value set on ATTgt.fit(), which overrides the value set on ATTgt.

  2. Passing None to ATTgt.aggte() means “inherit from the MP object built during fit”.

This matches the R did package convention.