Causal Inference

When to use this skill

Use when the question is "what is the causal effect of X on Y" but a clean A/B test isn't possible. Triggers:

• "Estimate the impact of <feature/policy/launch> without an A/B test"
• "We can't randomize. How do we measure effect?"
• "DiD analysis"
• "Propensity score matching"
• "Instrumental variable"
• "Did the launch of X cause Y to change?"

If a clean A/B test IS possible, use ab-test-design + ab-test-analysis — they always beat quasi-experimental methods.

If the question is "predict Y" not "what causes Y," use linear-regression or logistic-regression.

Decision framework: which method?

Answer these questions in order:

1. Is there a clear before/after time and a comparison group not affected by the change?
   YES → Difference-in-Differences (DiD)
   NO  → continue

2. Did some users opt into a feature/treatment based on observable characteristics?
   YES → Propensity Score Matching (PSM) or weighting
   NO  → continue

3. Is there a "lever" that affects treatment but doesn't directly affect outcome?
   YES → Instrumental Variable (IV)
   NO  → none of these will work cleanly; recommend RCT or accept correlational evidence with explicit caveats.

Required inputs

Method 1: Difference-in-Differences (DiD)

Setup

• A change happened at a specific time (e.g., new feature rolled out 2026-03-01).
• One group experiences the change; another doesn't (or experiences it later).
• You have pre-period AND post-period data for both groups.

Assumption (critical): Parallel trends

The treated and control groups would have followed parallel trends in the outcome absent the treatment. Show this by plotting both groups' outcome over time in the pre-period.

Estimation

import statsmodels.formula.api as smf

# panel data: one row per (unit, period)
model = smf.ols(
    formula='outcome ~ treated * post',
    data=df
).fit(cov_type='cluster', cov_kwds={'groups': df['unit_id']})

# The coefficient on `treated:post` is the DiD estimate

When DiD fails

• Parallel trends violated (treated group was already trending up before treatment)
• Spillover: control group is affected indirectly by treatment
• Confounding shock: another event hit one group at the same time
• Anticipation: treated units changed behavior before the treatment

Method 2: Propensity Score Matching (PSM)

Setup

• Some users opted into / received a feature; others didn't.
• The decision was driven by observable factors (not random).
• You want to compare "similar" treated and control users.

Steps

1. Fit a logistic regression: P(treated = 1 | confounders).
2. Get propensity scores for all users.
3. Match treated users to control users with similar scores.
4. Compare outcomes between matched pairs.

from sklearn.linear_model import LogisticRegression
from sklearn.neighbors import NearestNeighbors

# Step 1: fit propensity
p_model = LogisticRegression().fit(df[confounders], df['treated'])
df['propensity'] = p_model.predict_proba(df[confounders])[:, 1]

# Step 2: 1:1 match (treated → nearest control)
treated = df[df['treated'] == 1]
control = df[df['treated'] == 0]
nn = NearestNeighbors(n_neighbors=1).fit(control[['propensity']])
distances, indices = nn.kneighbors(treated[['propensity']])
matched_control = control.iloc[indices.flatten()].reset_index(drop=True)

# Step 3: compute ATT (average treatment effect on the treated)
att = (treated['outcome'].values - matched_control['outcome'].values).mean()

Assumption (critical): No unmeasured confounders

Anything affecting both treatment selection AND outcome must be in the confounders list. This is unfalsifiable — you must defend it with domain knowledge.

Diagnostics

• Common support: propensity score distributions overlap between treated and control. If they don't, can't match.
• Balance check: after matching, treated and control should have similar means on all confounders.

Method 3: Instrumental Variable (IV)

Setup

• Treatment is endogenous (correlated with unobserved confounders).
• There exists an "instrument" Z that:

1. Affects treatment (relevance)
2. Doesn't directly affect outcome except through treatment (exclusion)
3. Isn't correlated with unobserved confounders

Examples:

• Random assignment to receiving an offer (instrument) → actually using it (treatment) → outcome
• Distance from store (instrument) → store usage (treatment) → spend (outcome)
• Encouragement design experiments

Two-stage least squares

import statsmodels.formula.api as smf
from linearmodels.iv import IV2SLS

# Treatment = f(instrument, controls)
# Outcome = f(predicted_treatment, controls)
model = IV2SLS.from_formula(
    'outcome ~ 1 + control_x + [treatment ~ instrument]',
    data=df
).fit()

Why IV is hard

• Finding a valid instrument is rare — exclusion restriction is unfalsifiable.
• Weak instruments (low correlation with treatment) inflate SEs and bias estimates.
• Estimates apply only to "compliers" — users who change treatment status because of the instrument.

Workflow (any method)

1. Force the causal question. "What is the effect of X on Y?" — not "is X associated with Y."

1. Argue for a method using the decision framework above. State which method, why, and what assumption is being made.

1. Show the data structure. Pre/post + control? Selection on observables? Instrument?

1. Run the method.

1. Run sensitivity checks:

• DiD: plot parallel trends pre-treatment
• PSM: balance table + common support
• IV: first-stage F-statistic > 10

1. Compute the effect with CI. Cluster errors at the appropriate level.

1. State counterfactual clearly.

• DiD: "Had the policy not been implemented, treated group would have moved X% instead of X+δ%"
• PSM: "If untreated users had been treated, their outcome would have improved by δ"
• IV: "For compliers (those induced by the instrument), the effect is δ"

1. Write the readout with explicit caveats about what this evidence can and cannot support.

Output format

# Causal Inference: <treatment> → <outcome>

## Question
What is the causal effect of <treatment> on <outcome>?

## Why not a randomized experiment?
<e.g., feature is already launched, can't randomize at this point>
<e.g., regulatory constraint>
<e.g., ethical reason>

## Method: <DiD | PSM | IV>
- **Rationale:** <why this method fits this question>
- **Critical assumption:** <parallel trends | no unmeasured confounders | exclusion restriction>
- **Defense of assumption:** <evidence and domain reasoning>

## Setup
- Treatment: <definition + when>
- Outcome: <definition + grain>
- Comparison group: <who, why valid>
- Sample: N treated, N control
- Confounders adjusted for: <list>

## Estimated effect
- **Point estimate:** <δ> (units: <e.g., $ per user>)
- **95% CI:** [<low>, <high>]
- **Interpretation:** <plain English what this means>

## Sensitivity checks
- <parallel trends plot — pre-trends look parallel>
- <propensity balance table — covariates balanced after match>
- <first-stage F-stat > 10 for IV>
- <robust to alternative specifications?>

## Caveats
- <causal claim limited to ...>
- <effect estimate applies to ...>
- <what could overturn this finding>

## Decision implications
- <recommendation based on effect size and uncertainty>

Validation checks

• [ ] Method choice is justified using the decision framework
• [ ] Critical assumption is named and defended
• [ ] Sensitivity check ran and is reported
• [ ] CI reported, not just point estimate
• [ ] Counterfactual stated in plain English
• [ ] Caveats include what the evidence can NOT support

Edge cases & failure modes

• DiD with non-parallel pre-trends: don't use DiD. Try synthetic control method or accept evidence is weak.
• PSM with poor common support: large treated regions have no comparable controls. Don't match those; report effect for the overlapping region only.
• IV with weak instrument: F-stat < 10 → IV estimates are biased toward OLS, with inflated SEs. Don't use the instrument.
• Confounders that change WITH treatment: don't include as covariates (they're mediators). Including them removes the effect you're trying to measure.
• Treatment effect heterogeneity: average effects may hide important variation. Report by subgroup if pre-specified.
• External validity: effect estimated on observed population may not generalize. Note explicitly.

Scripts

• scripts/did.py — Difference-in-differences with cluster-robust SEs.
• scripts/psm.py — Propensity score matching with balance diagnostics.

Related skills

• ab-test-design / ab-test-analysis — gold standard; use whenever feasible
• linear-regression — non-causal modeling
• logistic-regression — propensity scoring underlies PSM
• stakeholder-readout — package the result with clear caveats