Causality Beyond Correlation

By the end of this lesson, you will be able to...

• Separate an observed association from a claim that changing one variable would change another.

• Use temporal order, mechanism, intervention, and alternative explanations to audit a causal story.

• State what a comparison identifies, what it leaves ambiguous, and what evidence would strengthen it.

Idea in one sentence: Correlation describes how measurements move together; causality asks what would happen under a well-defined intervention and why.

Core Insight

Imagine the support team notices that sessions with retrieval enabled contain fewer unsupported claims. The dashboard shows 12% for gated sessions and 18% for the old path. A manager asks for a rollout: “The gate works.” Before accepting that sentence, ask a harder question: What would have happened to the same kind of session if it had received the other path?

The previous lessons prepared the pieces. The experiment lesson showed why a control and assignment rule matter. The measurement lesson showed that “unsupported claim” is produced by a rubric and protocol. Causal reasoning connects those pieces without pretending that a tidy arrow appears automatically.

An association is a pattern in observed data. A causal claim is a statement about change under an intervention. If we turn the gate on for an otherwise comparable request, do unsupported claims fall? If we change the gate but keep the question, knowledge base, evaluator, and time conditions fixed, what difference should appear? The causal question is about that contrast, not merely about two columns that happen to differ.

The central trade-off is that causal reasoning supports action, but evidence standards are high. A weak causal story can still sound useful, yet acting on it may change a system, allocate resources, or expose users to harm.

Three Statements That Sound Similar

Consider three claims about retrieval gating:

1. Association: Gated sessions had a lower unsupported-claim rate.
2. Prediction: When a future session is routed through the gate, its measured rate is expected to be lower.
3. Causal effect: Routing a comparable session through the gate, rather than the old path, would lower its unsupported-claim risk by a stated amount.

The first can be computed from a table. The second may be useful for forecasting if the future resembles the observed sample. The third requires a defensible counterfactual comparison: outcomes under two possible treatments for the same unit or for units made comparable by design.

The distinction matters because action uses the third kind of statement. A hospital does not ask only whether patients who received a drug recovered more often; it asks whether offering the drug would improve recovery compared with the available alternative for similar patients. A platform team does not ask only whether high-cache-hit services are faster; it asks whether increasing cache capacity would reduce latency without moving the bottleneck elsewhere.

The Counterfactual Question

For one support request, define two potential outcomes:

• (Y(1)): the measured unsupported-claim outcome if the request uses retrieval gating.
• (Y(0)): the outcome if the same request uses the old retrieval path.

We can observe only one of these for that request in a live run. The unobserved value is its counterfactual. The average causal effect over a well-defined population compares average (Y(1)) with average (Y(0)), but the comparison is credible only if the assignment and measurement process make the two groups exchangeable enough for the decision.

This notation is not a demand for advanced mathematics. It is a guardrail against a common error: treating the treated group's observed outcome as if it were automatically comparable with a different group's outcome. “The gate group had 12%” does not reveal “the same requests would have had 18% without the gate” unless the design supports that interpretation.

Why Correlation Misleads

Reverse direction

The apparent cause may be an effect. Suppose experienced support agents choose the gated path for questions they already believe are risky. If gated sessions have more citations, perhaps the gate caused caution; or perhaps the agents' caution caused both the routing choice and the citation behavior. The arrow may point from human judgment to both variables.

Common cause

A third variable can create the association. Difficult questions may be routed to the gate and also produce more unsupported claims. A new knowledge-base snapshot may be deployed at the same time as the gate and improve both retrieval quality and evaluation labels. Time of day can correlate with both traffic mix and latency.

Selection and conditioning

If the team analyzes only sessions that received an answer, the gate may look better because it refuses more difficult requests. Conditioning on a post-treatment variable can remove the very failures the intervention was meant to address. Define the population and outcome before filtering.

Shared measurement artifact

The measured association can be created by the evaluator. If gated answers contain a citation marker that the rubric treats as evidence, the gate may appear to improve factual support even when the cited passage does not entail the claim. The measurement chain from lesson 003 is part of the causal diagram.

A Causal Diagram as a Question Generator

Draw arrows only to expose alternatives, not to decorate a story. For the support experiment, a simplified diagram might be:

question difficulty ──────> unsupported-claim outcome
        │                              ▲
        └──────> routing to gate ──────┘

knowledge-base version ───> routing context and outcome
evaluator rubric ──────────> recorded outcome

The desired arrow is routing to gate -> unsupported-claim outcome. The other paths ask what must be held constant, randomized, stratified, measured, or acknowledged as a limitation.

A diagram does not prove the arrows are true. It makes the assumptions inspectable. If the team cannot say why a variable precedes treatment, affects outcome, or is merely a measurement artifact, the causal story is not ready for a decision.

Time and Mechanism

Cause must precede effect in the relevant process. If a dashboard shows lower unsupported claims after the gate rollout, the timing is compatible with causality but does not establish it. The gate must be active before the answer is generated and before the evaluation label is assigned. A post-answer review cannot be the intervention that caused the answer.

Mechanism adds a second constraint. Retrieval gating is supposed to filter weak evidence, favor sources that entail the answer, and trigger abstention when support is missing. A causal claim becomes more credible when the observed change matches this mechanism: unsupported claims fall, source entailment improves, and refusals rise only within the predicted boundary. If the rate falls because all difficult questions are rejected, the outcome may improve while the service fails its broader purpose.

Mechanism is not a substitute for comparison. A plausible explanation can be wrong, and an unexpected mechanism can still be real. Use mechanism to derive predictions that distinguish alternatives.

Interventions and the Treatment Version

“Enable retrieval” is too vague to be a treatment. Specify the intervention:

• Which retriever and index version?
• What score threshold and maximum number of passages?
• Does the prompt change when no passage clears the threshold?
• Are users, sessions, or individual requests assigned?
• What happens when the knowledge base updates during the run?

Without a stable treatment version, replication is impossible and a causal estimate has no clear referent. A later system may use the same product name while changing the index, model, refusal policy, and evaluator. The effect belongs to the bundle actually deployed, not to a label.

A Worked Causal Audit

The team proposes: “The gate lowered unsupported claims by six percentage points.” Audit it in order.

1. State the estimand

For eligible support requests during the five-day trial, estimate the difference in the probability of at least one unsupported claim if each request were routed through the gated path versus the old path, using the same rubric and source snapshot.

2. Check assignment

If a stable random hash assigned sessions before the answer, routing is less likely to depend on question difficulty or agent preference. If assignment was manual, record who chose the path and which reasons may predict the outcome.

3. Check comparability

Compare language, product area, request difficulty, time window, user type, and knowledge-base version across conditions. Large imbalances do not automatically invalidate the study, but they identify what the assignment failed to balance.

4. Check interference

One request can update a shared cache or change an agent's behavior for later requests. If treatment changes the control group's environment, the simple difference no longer represents isolated individual effects. Choose a unit and rollout schedule that match the mechanism.

5. Check outcomes and missingness

Count refusals and unanswered requests as outcomes, not silent exclusions. Keep the rubric and evaluator blind where possible. Report the denominator, unclear labels, and any post-treatment filtering.

6. Compare alternatives

If both unsupported claims and refusals fall, perhaps the evaluator changed or difficult requests were removed. If unsupported claims fall only for one product area, the effect may depend on knowledge-base coverage. If the effect disappears when the old rubric is applied to a bridge sample, suspect measurement drift.

7. State the conclusion at the supported strength

An honest conclusion might be: “Under randomized session assignment, a fixed source snapshot, and rubric v2, the gate was associated with a six-point lower rate in this support population. The estimate does not establish effects for uncovered products or future knowledge-base versions.” This is more useful than an unconditional “the gate works” because it names the intervention and its boundary.

When Randomization Is Not Available

Some interventions cannot be randomized. A safety policy may require rollout to everyone; a historical event cannot be replayed; or interference may make individual assignment nonsensical. Causal reasoning still helps, but the claim must carry more assumptions.

• Time-series comparison: Look for a level or slope change at the intervention point, while checking seasonality and concurrent changes.
• Difference-in-differences: Compare treated and untreated populations over time, but defend the assumption that their pre-treatment trends were comparable.
• Matching or stratification: Construct more similar comparisons on measured variables, while admitting that unmeasured differences remain.
• Instrumental variation or natural experiments: Use an external assignment rule that shifts treatment but affects the outcome only through treatment; this requires a strong, domain-specific argument.

These designs do not turn observational data into randomized data by vocabulary. They make assumptions explicit and invite falsification checks. A causal estimate should say which assumptions carry the argument.

Common Confusions

“Correlation is useless.” Association is valuable for description, prediction, and generating hypotheses. It becomes a causal claim only when the intervention and alternatives are addressed.

“A mechanism proves causality.” Mechanisms generate testable predictions; they do not replace a credible comparison.

“Controlling for every variable is safest.” Adjusting for a mediator or a collider can introduce bias. Control variables because the design and causal story justify them, not because a longer list sounds rigorous.

“A significant result proves the treatment caused it.” A small probability under a model does not repair confounding, measurement drift, or a poorly defined estimand.

“No difference means no effect.” Effects can cancel across subgroups, be too imprecisely measured, or be blocked by noncompliance. Inspect power, heterogeneity, treatment fidelity, and outcome quality before declaring absence.

Active Checks

Check 1: A routing association

Difficult questions are preferentially sent to the gated path. The gated path has a higher unsupported-claim rate. Can the team conclude the gate is harmful?

Answer: No. Difficulty affects both routing and the outcome, so it is a confounder. The comparison needs randomized assignment, credible adjustment, or a narrower claim that preserves the limitation.

Check 2: A promising rollout

After a gate rollout, unsupported claims fall and refusals rise. The team also changed the evaluator rubric on the same day. What is the next move?

Answer: Reconstruct the measurement chain and apply both rubric versions to a bridge sample. Separate the instrument change from the product change before interpreting the trade-off between claims and refusals.

Transfer Practice

Pick a claim from your own work: “caching reduces API latency,” “code review reduces incidents,” or “a study method improves retention.” Write a causal audit:

1. What exactly is the intervention, and for which unit?
2. What is the outcome and how is it measured?
3. What would the same unit experience without the intervention?
4. Which variable could cause both assignment and outcome?
5. What mechanism predicts the direction and timing of the effect?
6. What observation would weaken or falsify the story?

The next lesson applies these questions across repetitions, populations, and contexts. A causal result that works once in one setting still needs replication, robustness checks, and an honest account of where it generalizes.

Resources

• [ARTICLE] Causal Inference - Stanford Encyclopedia of Philosophy - Focus: causal concepts, intervention, and counterfactual reasoning.
• [BOOK] Causal Inference: What If - Focus: estimands, potential outcomes, and study designs.
• [ARTICLE] The Book of Why - Focus: causal diagrams and assumptions.
• [ARTICLE] Scientific Method - Stanford Encyclopedia of Philosophy - Focus: how explanations and tests constrain scientific claims.

Key Takeaways

• Correlation is an observed association; causality is a claim about change under a defined intervention.
• Counterfactuals, temporal order, mechanisms, assignment, and alternative explanations are the pressure tests for a causal story.
• A causal diagram is useful when it exposes confounders, selection, interference, and measurement paths rather than merely drawing a plausible arrow.
• State causal conclusions at the strength supported by the design, population, treatment version, measurement protocol, and remaining assumptions.