HOW INTELLIGENCE WORKS · CAUSAL GENERALISATION · eduKateSG
How Intelligence Decides Whether a Cause Learned Here Should Still Work Somewhere Else
Causal generalisation is the intelligence process that decides whether a causal relationship learned in one setting should remain valid in another population, environment, time period or operating regime. It asks which parts of the mechanism are stable enough to transport and which parts depend on local conditions.
Source setting → causal mechanism → effect modifiers → target setting → compare structure → preserve stable relations → adjust changed relations → transport or refuse.
This article belongs to the How Intelligence Works series. Transfer and Recomposition owns broad transfer of knowledge into changed problems. Invariance Detection owns discovering what survives surface change. Causal generalisation owns a narrower question: does the causal relation itself survive when the population or environment changes?
A Cause Can Be Real Here and Weak Somewhere Else
A treatment, teaching method or policy may genuinely work in the setting where it was studied and still produce a different effect elsewhere because baseline risk, implementation, resources, context or effect modifiers changed.
Internal causal credibility and transport to a target population are different problems.
A causal effect is never just a number. It is a relation generated under conditions.
1. Generalisation Begins by Naming the Target
“Does it generalise?” is incomplete. To whom, where, when and under what implementation conditions?
The target population must be explicit before differences between source and target can be inspected.
2. Stable Mechanisms and Effect Modifiers Must Be Separated
| Element | Question |
|---|---|
| Core mechanism | What causal pathway is expected to operate in both settings? |
| Baseline state | Does the target begin from the same risk or capability level? |
| Effect modifier | Which variable changes the strength or direction of the effect? |
| Implementation | Is the intervention delivered with comparable fidelity and intensity? |
| Context | Do institutions, incentives or constraints alter the mechanism? |
| Outcome definition | Is the same outcome actually being measured? |
3. Surface Similarity Is Not Transportability
Two schools can look similar demographically and differ in timetable, teacher expertise, curriculum sequence or assessment incentives. Two patients can share a diagnosis and differ on a factor that modifies treatment response.
Generalisation depends on causal similarity, not visual resemblance alone.
4. Invariance Gives a Candidate Bridge
If a causal relation remains stable across several varied source settings, that invariance provides evidence that the mechanism may travel.
But invariance should be tested across meaningful changes, not only repeated samples from essentially the same regime.
Generalisation becomes credible when the relationship survives the kinds of change the target actually introduces.
5. Causal Generalisation in Mathematics and Modelling
Models often assume parameters or structural relationships remain stable beyond the sample used to estimate them. That assumption is a transport claim.
When the data-generating process changes, yesterday’s coefficient can become tomorrow’s error even if it was estimated perfectly.
6. Causal Generalisation in Education
An intervention that improves reading in one school may depend on teacher training, lesson duration, language background or existing curriculum. Copying the visible routine without those causal supports can produce a weaker result.
Educational transfer therefore asks what function must be preserved and which surface features may adapt locally.
This is how evidence can travel without becoming recipe worship.
7. Population Differences Matter Only When They Touch the Mechanism
Not every difference between source and target threatens transport. Hair colour may differ and be irrelevant. Baseline proficiency, age, prior treatment or language may change the mechanism materially.
The task is to identify differences that lie on, modify or block the causal pathway.
8. Partial Transport Is Often More Honest Than Yes or No
A direction of effect may transport while the magnitude changes. A mechanism may transport for one subgroup and not another. A policy may transport only after adaptation.
Generalisation should therefore permit bounded conclusions rather than force universal claims.
9. Causal-Generalisation Failure Atlas
| Failure | What happens | Repair |
|---|---|---|
| Source worship | One successful study is treated as universal | Name the target and compare conditions |
| Surface matching | Visible similarity substitutes for causal similarity | Map mechanisms and modifiers |
| Effect-size copying | The exact magnitude is transported without adjustment | Separate direction from size |
| Implementation blindness | Delivery differences are ignored | Measure fidelity and dose |
| Baseline neglect | Different starting states are assumed irrelevant | Model baseline-dependent effects |
| Target vagueness | “Everyone” replaces a defined population | Specify who, where and when |
| Adaptation taboo | Any local change is treated as loss of fidelity | Preserve function, adapt form carefully |
10. Causal Generalisation and Transfer Are Different
Transfer asks whether knowledge or a method can be reused in a new problem. Causal generalisation asks whether a cause–effect relation itself remains valid under changed conditions.
A method may transfer even when the causal magnitude changes; a causal relation may transport even when the operational method used to exploit it must be redesigned.
11. Teams Should Write Down the Transport Assumption
When evidence from elsewhere is imported into a decision, the team should name which mechanism it believes is stable and which contextual differences could alter the effect.
Every imported causal claim carries an invisible sentence: “we believe these differences do not change the mechanism enough to matter.” Make that sentence visible.
12. Institutions Need Local Evidence Loops After Importing External Evidence
Even strong external evidence should be followed by local monitoring when implementation begins.
The world return can reveal that a modifier, bottleneck or context feature was missed during transport reasoning.
13. Artificial Intelligence and Out-of-Distribution Causal Generalisation
AI systems often perform well when training and deployment distributions match and degrade when the environment changes.
Causal representations can help by focusing on mechanisms expected to remain stable across domains, but transport still requires assumptions about which mechanisms are invariant and which conditions shifted.
Generalisation should therefore be tested under meaningful distribution shift, not inferred from in-distribution accuracy alone.
14. The Causal Generalisation Audit
- Source: Where was the causal relation learned?
- Target: Where do we want to apply it?
- Mechanism: What pathway produced the effect?
- Differences: Which source–target variables changed?
- Modifiers: Which differences can alter the effect?
- Implementation: Is delivery comparable?
- Outcome: Is the same construct measured?
- Invariance: Which relation appears stable across settings?
- Adaptation: What must change locally while preserving function?
- Monitoring: What local return would reveal failed transport?
15. CivDJ Reading: Do Not Carry a Master Into a New Room Without Checking the Room
In the CivDJ frame, a Master that produced excellent World Return under one Receiver State may fail when the environment, constraints or receiver distribution changes.
Causal generalisation asks which wiring is fundamental and which calibration was local.
Transport the mechanism, not the decoration—and verify that the new room still supports the mechanism.
16. Return to the Cause That Must Travel
Causal generalisation prevents evidence-based practice from becoming blind copying.
It asks whether the mechanism, modifiers and implementation conditions that produced the source effect still exist in the target.
The mature mind does not ask merely whether something worked somewhere. It asks why it worked there—and whether that reason still exists here.