Correlation vs Causality¶

Decision-grade thinking

Correlation predicts. Causality intervenes.

Correlation answers “what moves together?” Causality answers “what changes what — under intervention?”. If your system will change the world, you need causal structure.

CausalGraphRAG Core primitives

The operational difference¶

flowchart TB
%% Styles (brModel Standard)
classDef i fill:#D3D3D3,stroke-width:0px,color:#000;
classDef p fill:#B3D9FF,stroke-width:0px,color:#000;
classDef r fill:#FFFFB3,stroke-width:0px,color:#000;
classDef o fill:#C1F0C1,stroke-width:0px,color:#000;
classDef s fill:#FFB3B3,stroke-width:0px,color:#000;

I_Data(["📊 Past data"]):::i
P_Fit("🎯 Fit model"):::p
R_Pred(["📈 Predict $P(Y|X)$"]):::r
O_Useful(["✅ Useful when stable (and you don't change the system)"]):::o

I_Act(["🧑‍⚖️ Proposed intervention (do X)"]):::i
R_Mech(["⚙️ Mechanism model (why X changes Y)"]):::r
P_Sim("🧪 Counterfactual reasoning"):::p
R_Effects(["📎 Side-effects + constraints (what could go wrong)"]):::r
O_Decision(["✅ Decision justified under intervention"]):::o

I_Data --> P_Fit --> R_Pred --> O_Useful
I_Act --> R_Mech --> P_Sim --> R_Effects --> O_Decision

R_Pred -. "breaks under" .-> R_Effects

%% Clickable nodes
click P_Sim "/methodology/causalgraphrag/" "CausalGraphRAG"
click R_Effects "/methodology/constraints/" "Constraints & SHACL"

🧭 This diagram separates prediction from intervention: correlation learns $P(Y|X)$, while decision-making needs mechanisms, counterfactual reasoning, and side-effect constraints.

Correlation

Often enough for prediction when environments are stable and you don’t change the system. It captures what co-moves in past data, but it can’t justify interventions or anticipate side effects.

Causality

Required for decisions when you will change the system (pricing, policy, treatment, automation). It models how actions propagate to outcomes, separating mechanism from coincidence under intervention.

Reasoning

Causality is the substrate. Reasoning adds goals, constraints, and counterfactual search over actions — so you can justify why an intervention is chosen and what could go wrong.

Two counterfactual statements¶

If we do $X=x_1$ instead of $X=x_0$, the outcome $Y$ would change.
If we remove a confounder $C$, the relationship between $X$ and $Y$ may disappear.

Diagram: confounding vs causal effect¶

flowchart TB
%% Styles (brModel Standard)
classDef i fill:#D3D3D3,stroke-width:0px,color:#000;
classDef p fill:#B3D9FF,stroke-width:0px,color:#000;
classDef r fill:#FFFFB3,stroke-width:0px,color:#000;
classDef o fill:#C1F0C1,stroke-width:0px,color:#000;
classDef s fill:#FFB3B3,stroke-width:0px,color:#000;

I_C(["🌀 Confounder C"]):::i
I_X(["X (action / exposure)"]):::i
I_Y(["Y (outcome)"]):::i

I_C --> I_X
I_C --> I_Y
I_X --> I_Y

R_Naive(["⚠️ Naive correlation mixes C → Y with X → Y"]):::r
I_C -. "unobserved" .-> R_Naive
I_X -. "observed" .-> R_Naive
I_Y -. "observed" .-> R_Naive

🌀 This diagram shows why correlation can lie: a confounder $C$ drives both $X$ and $Y$, making the observed $X$–$Y$ relationship partly (or entirely) spurious.

Diagram: interventions change the object¶

flowchart TB
%% Styles (brModel Standard)
classDef i fill:#D3D3D3,stroke-width:0px,color:#000;
classDef p fill:#B3D9FF,stroke-width:0px,color:#000;
classDef r fill:#FFFFB3,stroke-width:0px,color:#000;
classDef o fill:#C1F0C1,stroke-width:0px,color:#000;
classDef s fill:#FFB3B3,stroke-width:0px,color:#000;

I_Obs(["📊 Observational world (nobody changes X)"]):::i
R_Pyx(["📈 Learn P(Y|X)"]):::r
O_Pred(["✅ Prediction can work (if environment stays similar)"]):::o

I_Do(["🧑‍⚖️ We change X (policy / pricing / treatment)"]):::i
R_Pydo(["📐 Need P(Y|do(X))"]):::r
O_Dec(["✅ Decisions require mechanism + side-effects"]):::o

I_Obs --> R_Pyx --> O_Pred
I_Do --> R_Pydo --> O_Dec

R_Pyx -. "not equal" .-> R_Pydo

%% Clickable nodes
click R_Pydo "/methodology/causalgraphrag/" "CausalGraphRAG"

🧪 This diagram captures the core counterfactual shift: observational prediction targets $P(Y|X)$, but interventions require $P(Y|do(X))$ — a different object with different assumptions.

Common failure mode¶

A predictive model learns $P(Y|X)$. When you intervene on $X$, you need $P(Y|do(X))$. Those are not the same object.

Common traps (and what to do instead)¶

flowchart LR
%% Styles (brModel Standard)
classDef i fill:#D3D3D3,stroke-width:0px,color:#000;
classDef p fill:#B3D9FF,stroke-width:0px,color:#000;
classDef r fill:#FFFFB3,stroke-width:0px,color:#000;
classDef o fill:#C1F0C1,stroke-width:0px,color:#000;
classDef s fill:#FFB3B3,stroke-width:0px,color:#000;

I_Obs(["📊 Observational data + correlations"]):::i

R_T1(["🌀 Confounding"]):::r
R_T2(["🎯 Selection bias"]):::r
R_T3(["🌦️ Distribution shift"]):::r
R_T4(["↩️ Reverse causality"]):::r
R_T5(["🔁 Policy feedback"]):::r
R_T6(["🎯 Goodhart / proxy gaming"]):::r

R_Bad(["❌ Bad path: treat P(Y|X) as decision-grade"]):::s
P_Wrong("⚠️ Wrong intervention (do X) + missing side-effects"):::s
S_Harm(["💥 Bad outcome (unintended harm / wasted spend)"]):::s

P_F1("Model confounders / identification"):::p
P_F2("Track selection mechanism + robustness tests"):::p
P_F3("Monitor drift + revalidate assumptions"):::p
P_F4("Time ordering / instruments + structural tests"):::p
P_F5("Model feedback loops + second-order effects"):::p
P_F6("Guardrails + mechanism + anticipate adaptation"):::p

P_Good("✅ Guarded path: identify P(Y|do(X)) + governance + measurement"):::o
P_Do("🧪 Run intervention with constraints"):::p
O_Value(["✅ Better outcome + audit-ready learning"]):::o

I_Obs --> R_T1
I_Obs --> R_T2
I_Obs --> R_T3
I_Obs --> R_T4
I_Obs --> R_T5
I_Obs --> R_T6

R_T1 --> R_Bad
R_T2 --> R_Bad
R_T3 --> R_Bad
R_T4 --> R_Bad
R_T5 --> R_Bad
R_T6 --> R_Bad

R_Bad --> P_Wrong --> S_Harm

R_T1 --> P_F1 --> P_Good
R_T2 --> P_F2 --> P_Good
R_T3 --> P_F3 --> P_Good
R_T4 --> P_F4 --> P_Good
R_T5 --> P_F5 --> P_Good
R_T6 --> P_F6 --> P_Good

P_Good --> P_Do --> O_Value

%% Clickable nodes
click P_Good "/methodology/causalgraphrag/" "CausalGraphRAG"
click P_F1 "/methodology/core-primitives/" "Core primitives"
click P_F5 "/philosophy/ai-agent-vs-agentic-ai/" "Agent vs agentic"
click P_F6 "/reasoners/governance/" "Governance"
click P_Do "/methodology/constraints/" "Constraints & SHACL"

🛑 This diagram enumerates the main traps (confounding, selection, drift, reverse causality, feedback, Goodhart) and routes each into a guarded causal workflow with governance and measurement.

Confounding

A third variable drives both X and Y. Fix: model confounders explicitly or design an identification strategy.

Selection bias

Your data is a filtered subset of reality. Fix: track selection mechanisms and test robustness.

Distribution shift

The world changes after deployment. Fix: monitor drift and revalidate assumptions continuously.

Reverse causality

Y drives X (or both co-evolve), so the arrow points the other way. Fix: use time ordering, instruments, or explicit structural assumptions — then test implications.

Policy feedback

Interventions change incentives and behavior. Fix: explicitly model feedback loops and second-order effects.

Goodhart / proxy gaming

Optimizing a proxy breaks the link to the real goal. Fix: model the mechanism, include guardrail outcomes, and anticipate strategic adaptation.

Where this connects in our stack¶

Philosophy: don’t confuse predictive accuracy with reliable intervention.
Methodology: encode causal structure in memory (graphs), then constrain what paths are allowed.

Next: CausalGraphRAG.