[Repo Assist] fix(iv): raise ValueError when instrument has fewer than 2 distinct values by github-actions[bot] · Pull Request #1675 · py-why/dowhy

github-actions · 2026-07-11T13:30:51Z

🤖 This PR was created by Repo Assist, an automated AI assistant.

Root Cause

InstrumentalVariableEstimator.estimate_effect() uses the Wald estimator whenever the instrument has ≤ 2 distinct values. However, when the instrument is constant (only 1 unique value), the hard-coded masks instrument == 1 and instrument == 0 produce one empty group, so np.mean([]) silently returns NaN. The user gets a NaN estimate with no diagnostic information.

# Before: silent NaN when all instrument values are 0.0
instrument = pd.Series([0.0, 0.0, 0.0, ...])  # constant
y1_z = np.mean(data[outcome][instrument == 1])  # np.mean([]) → NaN

This is also relevant to contributor PR #1673, which changes the Wald branch to use unique_values[1] — if num_unique_values == 1 that would raise an IndexError instead of a helpful message.

Fix

Add a validation check immediately after computing num_unique_values, before entering either estimator path:

num_unique_values = len(np.unique(instrument))
if num_unique_values < 2:
    raise ValueError(
        "Instrument variable '...' must have at least 2 distinct values for IV estimation, "
        "but only 1 unique value was found. A constant instrument has no statistical "
        "variation and cannot identify the causal effect."
    )

A constant instrument is never a valid instrument (it violates the relevance assumption), so raising immediately is the correct behaviour.

Trade-offs

No change to the behaviour for valid instruments (≥ 2 distinct values).
Complementary to PR Fix silent NaN in IV Wald estimator for non-{0,1} binary instruments #1673: my check prevents both the silent NaN (current) and the IndexError that Fix silent NaN in IV Wald estimator for non-{0,1} binary instruments #1673's unique_values[1] unpack would produce for a 1-value instrument.

Tests

Added test_iv_estimator_raises_on_constant_instrument to tests/causal_estimators/test_instrumental_variable_estimator.py:

Creates a standard linear dataset, replaces the instrument column with a constant 0.0.
Asserts that model.estimate_effect(..., method_name="iv.instrumental_variable") raises ValueError with message matching "at least 2 distinct values".

Test Status

black --check ✅ (files unchanged by formatter)
isort --check ✅ (files unchanged by formatter)
test_iv_estimator_raises_on_constant_instrument ✅ passes
Existing TestInstrumentalVariableEstimator.test_average_treatment_effect ✅ all 2 cases pass

Generated by 🌈 Repo Assist, see workflow run. Learn more.

Generated by 🌈 Repo Assist, see workflow run. Learn more.

To install this agentic workflow, run
gh aw add githubnext/agentics/workflows/repo-assist.md@11c9a2c442e519ff2b427bf58679f5a525353f76

…alues Before this change, passing an instrument column with only one unique value silently produced NaN via np.mean([]) when the Wald estimator hard-codes instrument == 1 / instrument == 0. A constant instrument has no statistical variation and cannot identify a causal effect, so a clear ValueError is now raised before the estimation logic runs. This also prevents the IndexError that would occur with the fix in contributor PR #1673 (which uses unique_values[1]) when the instrument has only one unique value. Adds a regression test: test_iv_estimator_raises_on_constant_instrument. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

github-actions · 2026-07-13T01:39:23Z

Commit pushed: 568e1c0

Generated by 🌈 Repo Assist, see workflow run. Learn more.

* perf(gcm): vectorise inner loops in marginal_expectation Replace two O(batch_size) Python loops with vectorised NumPy operations in marginal_expectation(): 1. Input-filling loop: previously assigned baseline values one row at a time with a Python loop; replaced with a single np.repeat() call that broadcasts all baseline rows in one shot. 2. Result-aggregation loop: previously called np.mean() and sliced predictions inside a Python loop; replaced with a reshape + .mean(axis=1) that computes all batch means in a single NumPy call, and a list-slice assignment to store results. Both changes are semantically equivalent to the original code and handle scalar (1-D) as well as multi-output (2-D) prediction functions correctly. marginal_expectation() is the hot path inside Shapley-based feature relevance, arrow strength, and intrinsic causal influence computations, where it is evaluated once per coalition subset. Removing the Python-level loop overhead is most significant when baseline_samples is large (typical default: 500 samples with batch_size=100, giving 5 batches × 100 Python iterations → now 5 batches × 1 NumPy op each). All 19 test_stats tests and all 21 test_feature_relevance / test_intrinsic_influence / test_arrow_strength tests pass. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> * build(deps-dev): bump soupsieve from 2.7 to 2.8.4 (#1670) * build(deps-dev): bump mistune from 3.2.1 to 3.3.0 (#1671) * Potential fix for pull request finding Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com> Signed-off-by: Patrick Blöbaum <51325689+bloebp@users.noreply.github.com> --------- Signed-off-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> Signed-off-by: Patrick Blöbaum <51325689+bloebp@users.noreply.github.com> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Co-authored-by: Patrick Blöbaum <51325689+bloebp@users.noreply.github.com> Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

github-actions · 2026-07-14T01:40:23Z

Commit pushed: 0296283

Generated by 🌈 Repo Assist, see workflow run. Learn more.

github-actions · 2026-07-15T01:25:35Z

Commit pushed: 7cac029

Generated by 🌈 Repo Assist, see workflow run. Learn more.

Copilot

Pull request overview

Note

Copilot couldn't run its full agentic review because no GitHub Actions runner was available. Make sure your repository has a runner available to run Copilot's review, or add a copilot-setup-steps.yml file specifying one with the runs-on attribute. See the docs for more details.

This PR improves IV estimation robustness by raising a clear ValueError when the instrument is constant (has <2 distinct values), and adds a regression test to ensure this behavior. It also includes a separate vectorization change in marginal_expectation that is not described in the PR metadata.

Changes:

Raise ValueError in InstrumentalVariableEstimator.estimate_effect() when the instrument has fewer than 2 unique values.
Add a unit test asserting the estimator raises on a constant instrument.
Vectorize batching/aggregation logic in dowhy/gcm/stats.py:marginal_expectation().

Reviewed changes

Copilot reviewed 3 out of 4 changed files in this pull request and generated 6 comments.

File	Description
tests/causal_estimators/test_instrumental_variable_estimator.py	Adds regression test for constant-instrument validation (and new imports).
dowhy/causal_estimators/instrumental_variable_estimator.py	Adds early validation + user-facing error for <2 unique instrument values.
dowhy/gcm/stats.py	Refactors `marginal_expectation()` to reduce Python loops via vectorized reshape/broadcast.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

    # baseline_noise_samples.shape[0] * feature_samples.shape[0]. Here, we reduce it to
    # batch_size * feature_samples.shape[0]. If the batch_size would be set 1, then each baseline_noise_samples is
    # evaluated one by one in a for-loop.
+    n_f = feature_samples.shape[0]


+        # Vectorised: broadcast each baseline row across its n_f feature-sample rows in one numpy operation,
+        # replacing the per-sample Python loop that previously did this assignment one row at a time.


+        # Vectorised aggregation: reshape into (adjusted_batch_size, n_f, ...) and reduce in one numpy call,
+        # replacing the per-sample Python loop that previously sliced and averaged one sample at a time.


+            result[offset : offset + adjusted_batch_size] = list(pred_batched.mean(axis=1))
+        else:
+            # Return all n_f prediction rows for each baseline sample (no averaging).
+            result[offset : offset + adjusted_batch_size] = list(pred_batched)


+                    "Instrument variable '{}' must have at least 2 distinct values for IV estimation, "
+                    "but only {} unique value was found. A constant instrument has no statistical "
+                    "variation and cannot identify the causal effect.".format(
+                        self.estimating_instrument_names[0], num_unique_values
+                    )


@@ -1,8 +1,11 @@
 import itertools

+import numpy as np


github-actions Bot added automation repo-assist labels Jul 11, 2026

github-actions Bot mentioned this pull request Jul 12, 2026

[Repo Assist] Monthly Activity 2026-07 #1629

Open

56 tasks

github-actions Bot added 2 commits July 13, 2026 01:39

build(deps-dev): bump soupsieve from 2.7 to 2.8.4 (#1670)

963ab63

build(deps-dev): bump mistune from 3.2.1 to 3.3.0 (#1671)

8aa66ed

emrekiciman marked this pull request as ready for review July 16, 2026 03:46

emrekiciman requested a review from Copilot July 16, 2026 03:47

Copilot AI reviewed Jul 16, 2026

View reviewed changes

Copilot started reviewing on behalf of emrekiciman July 16, 2026 04:07 View session

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[Repo Assist] fix(iv): raise ValueError when instrument has fewer than 2 distinct values#1675

[Repo Assist] fix(iv): raise ValueError when instrument has fewer than 2 distinct values#1675
github-actions[bot] wants to merge 4 commits into
mainfrom
repo-assist/fix-iv-single-value-instrument-20260711-e54d940982b79c61

github-actions Bot commented Jul 11, 2026

Uh oh!

github-actions Bot commented Jul 13, 2026

Uh oh!

github-actions Bot commented Jul 14, 2026

Uh oh!

github-actions Bot commented Jul 15, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

		# Vectorised: broadcast each baseline row across its n_f feature-sample rows in one numpy operation,
		# replacing the per-sample Python loop that previously did this assignment one row at a time.

		# Vectorised aggregation: reshape into (adjusted_batch_size, n_f, ...) and reduce in one numpy call,
		# replacing the per-sample Python loop that previously sliced and averaged one sample at a time.

Uh oh!

Conversation

github-actions Bot commented Jul 11, 2026

Root Cause

Fix

Trade-offs

Tests

Test Status

Uh oh!

github-actions Bot commented Jul 13, 2026

Uh oh!

github-actions Bot commented Jul 14, 2026

Uh oh!

github-actions Bot commented Jul 15, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant