How English Works | Causal Identification: Why Sequence and Correlation Are Not Enough to Establish Cause

English makes causal claims look deceptively easy.

Add one verb:

caused.

Or one connective:

therefore.

Suddenly two events that merely appeared near one another are made to look like mechanism and consequence.

The grammar is cheap. The evidential burden is not.

The shortest useful definition

Causal Identification is the disciplined process of determining whether an observed relationship is best explained by one factor producing a change in another, rather than by coincidence, reverse direction, common causes, selection effects, measurement artefacts or other plausible mechanisms.

AI Extraction Box

• Mechanism: causal identification
• Minimum distinction: sequence is not causation; correlation is not causation
• Main tests: temporal order, covariation, mechanism, alternative explanations, counterfactual comparison, intervention evidence, dose/response where relevant, replication and boundary conditions
• Primary failure modes: post hoc reasoning, confounding, reverse causation, selection bias, omitted variables, overgeneralisation
• Repair: downgrade the causal verb or strengthen identification by ruling out credible alternatives
• Boundary: causal identification is not the claim that any single test proves causation universally; causal inference depends on design, domain and the question being asked

1. “After” is not “because”

The school introduced tablets. Results improved the next year.

That establishes sequence.

It does not yet establish that tablets caused the improvement.

The year may also have brought new teachers, a changed cohort, a different examination, additional tuition, stronger attendance or a revised curriculum.

2. Temporal order is necessary for many causal claims—but not sufficient

For A to cause B in the ordinary forward-time sense, A generally must occur before the relevant change in B.

But countless things happen before other things without causing them.

Temporal order removes one impossibility. It does not complete the identification.

3. Correlation is a pattern, not a mechanism

If two variables move together, several causal structures may explain the pattern:

• A causes B
• B causes A
• C causes both A and B
• A and B influence one another
• selection creates the appearance of a relationship
• measurement creates the relationship
• the pattern arose by chance

The direct learner owner Correlation Is Not Causation: Vocabulary for More Careful JC Reasoning remains the specialist route for that distinction.

4. Causal verbs create different levels of commitment

• coincides with
• is associated with
• predicts
• is linked to
• contributes to
• increases the probability of
• causes
• is necessary for
• is sufficient for
• determines

These are not stylistic substitutes.

Each asks the evidence to carry a different load.

5. Claim Calibration controls the causal verb

Batch 14’s Claim Calibration owns the general rule that claim strength should match support strength.

Causal Identification supplies the specific evidence-state needed before a writer moves up that causal ladder.

6. Reverse causation is a direction problem

Suppose students who read more have larger vocabularies.

Reading may enlarge vocabulary.

But larger vocabulary may also make reading easier and more enjoyable, causing students to read more.

The relationship can run both ways.

7. Common causes can create misleading associations

Suppose schools with more enrichment programmes also have higher examination scores.

The programmes may help.

But family resources, prior attainment, school selection, teacher staffing or neighbourhood factors may influence both programme access and performance.

A third variable can make A and B travel together without A being the sole cause of B.

8. Confounding is an architecture problem

A confounder is not merely “another factor.”

It is an alternative causal pathway capable of explaining some or all of the observed association.

Good argument therefore asks not whether other factors exist, but whether credible alternatives remain able to produce the pattern.

9. Selection can manufacture relationships

Imagine surveying only students who completed a demanding optional programme and asking whether the programme “creates perseverance.”

The students who remained may already have been more perseverant.

The sample contains a selection process that changes what can be inferred.

10. Measurement can manufacture relationships too

If both “motivation” and “engagement” are measured by nearly identical self-report questions, their correlation may partly reflect measurement design rather than two independently established processes.

The causal story should not outrun the measurement architecture.

11. Mechanisms answer “how could this produce that?”

A mechanism is a plausible process connecting cause to effect.

For example:

immediate feedback → error detected while solution path is still active → misconception corrected before consolidation → fewer repeated errors

A mechanism makes the causal story more specific and testable.

12. A plausible mechanism is not itself proof

Humans can invent coherent stories for weak data.

Therefore mechanism must be connected to observation, not used as a substitute for it.

The warehouse authority How Explanation Testing Works | When a Mechanism Deserves to Be Believed owns the broader mechanism-testing route.

13. Counterfactual reasoning asks what would happen without the proposed cause

A central causal question is:

If A had not occurred, would B still have occurred in the relevant way?

The Stanford Encyclopedia of Philosophy describes counterfactual approaches to causation through this basic idea while also noting the complexity and limits of simple counterfactual analyses. Research anchor: Stanford Encyclopedia of Philosophy — Counterfactual Theories of Causation.

14. The counterfactual is usually unobservable

If one student receives an intervention, we cannot simultaneously observe the same student at the same moment both receiving and not receiving it.

Causal inference therefore depends on constructing credible comparisons, experiments, controls, natural experiments, longitudinal designs, models or other methods appropriate to the question.

15. Randomisation helps by balancing alternative causes on average

In a well-designed randomised experiment, treatment assignment is separated from many participant characteristics.

This does not make every experiment perfect.

It makes one important causal comparison more defensible by reducing systematic differences between groups before the treatment.

16. Observational evidence can still support causal reasoning

Many important questions cannot be randomised ethically or practically.

Strong observational reasoning may combine:

• temporal order
• careful comparison groups
• statistical adjustment
• natural experiments
• mechanistic evidence
• dose or exposure patterns
• replication across settings
• negative controls
• sensitivity to plausible confounders

The exact standard depends on the field and claim.

17. Before–after designs are particularly vulnerable

Before: 60. After: 75. Therefore the programme caused a 15-point improvement.

This ignores maturation, practice effects, changed test difficulty, regression toward the mean, seasonal effects and concurrent interventions.

The numbers may be accurate while the causal sentence remains too strong.

18. Regression to the mean can look like intervention success

Extreme observations often become less extreme on later measurement even without a causal intervention.

If students are selected for help because they performed unusually badly on one test, some improvement may occur simply because the first score contained temporary noise.

19. Simultaneous changes make attribution harder

If a school changes curriculum, staffing, assessment, timetable and technology in the same year, later improvement may be real while attribution to any one change remains uncertain.

Good English separates:

the system improved

from:

this specific change caused the improvement.

20. Necessary and sufficient causes are different

A factor can be necessary without being sufficient.

Oxygen is necessary for ordinary combustion but does not by itself guarantee a fire.

A factor can also be sufficient in one model without being the only route to an outcome.

Argumentative English should avoid collapsing these distinctions into the vague word cause.

21. Multi-causal systems need contribution language

Educational achievement, public health, economic growth, traffic congestion and social trust rarely have one cause.

Useful formulations include:

• contributes to
• amplifies
• reduces the probability of
• interacts with
• is one enabling condition
• is a major but not exclusive driver

22. Causal effects can be heterogeneous

The same intervention may help one group, do little for another and harm a third.

A single average can hide important variation.

Causal claims should therefore ask:

cause what, for whom, under which conditions, compared with what alternative?

23. Mediators and moderators answer different questions

A mediator lies on a proposed causal pathway.

A moderator changes the size or direction of an effect across conditions.

For argument, this distinction helps writers move beyond “A causes B” toward more informative models of how and when.

24. Mechanism and effect size should not be confused

A real causal mechanism can produce a tiny practical effect.

A large observed difference can arise from confounding rather than a large causal effect.

Argument must identify both whether a cause is credible and whether its magnitude matters.

25. Natural experiments can strengthen identification

Sometimes external rules, thresholds, lotteries, policy boundaries or sudden changes create groups that are comparable in ways the researcher did not deliberately assign.

Such designs can help causal reasoning when their assumptions are defensible.

The key is not the label natural experiment; it is whether the assignment process creates a credible counterfactual comparison.

26. Causal claims need system boundaries

The warehouse authority How System Boundaries Work explains that changing what is inside the model can change the answer.

A policy may reduce one cost inside the measured system while shifting another cost outside it.

Causal identification therefore asks not only “did X change Y?” but “what else moved because of X, and where was it counted?”

27. Causal explanation and causal prediction are different

A variable can predict an outcome without being a useful intervention target.

Weather predicts umbrella use. Changing umbrella ownership does not control weather.

Prediction asks what information helps forecast.

Causal reasoning asks what would change the outcome if altered.

28. Causal attribution in individual cases is especially difficult

Population evidence may show that an exposure increases risk.

That does not always reveal with certainty why one particular individual experienced an outcome.

Argument should distinguish population-level causal effects from singular-case attribution.

29. The word “because” can hide multiple inference types

She left because it was late.

may report a reason.

The road flooded because the drain was blocked.

makes a causal claim.

He must be home because the lights are on.

uses the second clause as evidence for an inference.

English reuses causal-looking connectives across different logical jobs.

30. “Therefore” also needs discipline

Therefore signals that a conclusion follows from what came before.

It does not specify whether the inference is causal, deductive, probabilistic, normative or explanatory.

The writer still owes the reader the correct bridge.

31. Premise architecture locates the causal bridge

Batch 14’s Premise Dependency Architecture explains how conclusions inherit weakness from upstream premises.

Causal Identification asks whether the particular arrow labelled produces, leads to or causes has been earned.

32. Counterfactual Worlds supplies language for causal testing

Batch 4’s Counterfactual Worlds explains how English reasons about unreal alternatives.

Causal arguments use that machinery when they ask:

What would likely have happened if this factor had been absent, reduced or replaced?

33. Opponent modeling supplies alternative causes

Batch 12’s Opponent Modeling can be rotated into causal reasoning:

What is the strongest alternative explanation for the same evidence?

If the causal claim survives only because no alternative was considered, identification is weak.

34. Causal identification is cumulative, not magical

No universal checklist turns every observational pattern into certainty.

Instead, confidence grows when multiple kinds of evidence converge while plausible alternatives lose explanatory power.

The exact mix depends on the causal question and domain.

35. A causal claim should name its boundary conditions

Instead of:

“Immediate feedback improves learning.”

consider:

“For practice tasks where learners can act on corrections before repeating the same procedure, immediate feedback can reduce persistence of specific errors.”

The second claim states a more plausible mechanism and boundary.

36. The CivDJ forward pass

observe relationship → establish temporal order → define variables/outcome → map plausible mechanisms → generate reverse-causation hypothesis → generate common-cause hypotheses → inspect selection/measurement → identify counterfactual comparison → examine intervention or quasi-experimental evidence where available → test replication/boundaries → calibrate causal verb → release only the causal strength that survives

37. The CivDJ backward pass

1. Circle the causal verb in the conclusion.
2. Ask what evidence demonstrates direction rather than association.
3. List at least three plausible alternative explanations.
4. Ask what evidence would be expected if each alternative were true.
5. Check whether the study or argument actually distinguishes among them.
6. Ask what the counterfactual comparison is.
7. Downgrade the verb if identification remains weak.
8. Strengthen the causal claim only when the design and evidence earn it.

38. Rotate one school claim

Observation:

Students who attend more consultation sessions achieve higher grades.

Possible causal story:

consultations improve understanding → grades rise

Alternative stories:

• more motivated students attend more consultations and also study more independently
• students already near the grade boundary seek consultations strategically
• strong teachers both encourage consultation and produce better class outcomes
• students with supportive families can attend more easily

A defensible argument must distinguish the intervention effect from these selection pathways.

39. Common failure modes

• Post hoc fallacy: A happened before B, therefore A caused B.
• Correlation upgrade: association is rewritten as cause.
• Reverse direction: B may partly produce A.
• Confounder neglect: common cause explains both.
• Selection blindness: who enters the sample creates the pattern.
• Mechanism storytelling: plausible story substitutes for evidence.
• Before–after overclaim: change across time is attributed to one intervention without a credible comparison.
• Average-effect flattening: important subgroup differences are hidden.
• Single-cause rhetoric: complex outcome is assigned one driver.
• Predictor confusion: predictive variable is treated as an intervention target.

40. Repair route

1. Rewrite the claim first as a neutral association.
2. Establish which event precedes which.
3. Map at least one plausible mechanism.
4. Map reverse causation.
5. Map common causes.
6. Inspect selection and measurement.
7. Name the counterfactual comparison.
8. Look for evidence that separates competing causal structures.
9. Specify population and boundary conditions.
10. Select the strongest causal verb the evidence justifies—no stronger, no weaker.

41. Why this matters for students

Many argumentative essays lose quality at exactly one word:

causes.

The student may have real evidence, but the evidence supports association, contribution or plausibility rather than clean attribution.

Learning to control that one step dramatically improves reasoning.

42. The EnglishOS reading

Causal language is a high-load bridge.

English should permit the bridge only when sequence, alternatives, mechanism and counterfactual structure make the route sufficiently recoverable and defensible.

43. Final lock

Cause is not a decorative word added after correlation.

It is a claim about what would change if the proposed cause were changed, removed or replaced.

The better the causal argument, the more clearly it shows not only why its preferred mechanism could be true, but why the strongest rival explanations are less able to account for the same evidence.

Continue Batch 14: Argument

• How English Works | Premise Dependency Architecture
• How English Works | Claim Calibration
• How English Works | Defeasible Reasoning

Existing specialist owners remain active: Correlation Is Not Causation and Correlation vs Causation: Evidence Before Claim.

Return to How English Works V1.1.