Causal inference diagnostic plots

Plotting utilities help you check distributions of propensity scores and predicted outcome probabilities across treatment and control groups — the first sanity check before trusting any estimate.

Requires the plotting extra:

pip install "CausalEstimate[plotting]"

import matplotlib.pyplot as plt
from CausalEstimate.vis.plotting import plot_propensity_score_dist, plot_outcome_proba_dist

# df has columns "ps", "treatment", and "predicted_outcome"
fig, ax = plot_propensity_score_dist(df, ps_col="ps", treatment_col="treatment")
plt.show()

fig, ax = plot_outcome_proba_dist(df, outcome_proba_col="predicted_outcome", treatment_col="treatment")
plt.show()

Good overlap between the treated and control propensity distributions supports the positivity assumption; clear separation is a warning sign — consider common-support filtering or matching.

Two diagnostics-oriented plots complement the Diagnostics module: a Love plot of covariate balance and the IPW weight distribution.

from CausalEstimate.diagnostics import compute_balance_table
from CausalEstimate.vis.plotting import plot_love, plot_ps_boxplot, plot_weight_dist

table = compute_balance_table(df, covariate_cols=["age", "bmi"], ps_col="ps", treatment_col="treatment")
fig, ax = plot_love(table)
fig, ax = plot_weight_dist(df, ps_col="ps", treatment_col="treatment", weight_type="ATE")
fig, ax = plot_ps_boxplot(df, ps_col="ps", treatment_col="treatment", weight_type="ATE")

plot_ps_boxplot draws propensity-score boxplots per arm before and after IPW weighting (weighted quantiles); after weighting, the treated and control boxes should nearly coincide.

For simulation studies, plot_zipper shows confidence-interval coverage across replicates: each interval is a horizontal segment, colored by whether it covers the truth, with the empirical coverage in the legend.

from CausalEstimate.vis.plotting import plot_zipper

# truth (scalar or one value per replicate), lower and upper bounds per replicate
fig, ax = plot_zipper(true_ate, ci_lower, ci_upper)

See this notebook and the diagnostics notebook for rendered examples, and the API Reference for full signatures.