Causal Inference Primer — Beyond Correlation

This lesson on Causal Inference Primer — Beyond Correlation is hands-on and example-driven. You will learn to distinguish between correlation, intervention, and counterfactual reasoning using Judea Pearl's Ladder of Causation. You will reorient your analytical focus from manipulating observed data to explicitly specifying the underlying causal mechanism and necessary assumptions.

What You'll Be Able To Do

• Define the three rungs of the Ladder of Causation.
• Differentiate between questions answerable by association versus intervention.
• Specify the type of assumptions required for Rung 3 analysis.
• Contrast causal modeling with standard statistical data preparation techniques.
• Identify the necessary conditions for a question to be counterfactual.

Detailed Concept Walkthrough

1. Judea Pearl's Causal Hierarchy

The Ladder of Causation organizes causal inquiry into three levels: Association (seeing), Intervention (doing), and Counterfactuals (imagining). Higher rungs require stronger, more explicit assumptions about the data generating process.

• Mechanism: The ladder represents increasing complexity and rigor in causal claims; each higher rung subsumes the capabilities of the lower ones, moving from passive observation to active imagination.
• Best Practice: Always identify the highest rung necessary to answer your research question, as this dictates the required data, modeling assumptions, and the necessary level of formalization (e.g., moving from simple regression to Structural Causal Models).
• Under the Hood: Moving up the ladder requires moving beyond standard statistical models (Rung 1) that only describe data distributions, to models that explicitly encode mechanisms and allow for hypothetical manipulation.
• Nuance: The core intellectual challenge is that causal modeling often requires pre-existing knowledge or strong theoretical assumptions about the process before the data analysis even begins.

Key Takeaway: The rung of the ladder determines the type of question you can ask and the explicit assumptions you must state.

2. Association vs. Intervention

Association (Rung 1) describes observed relationships (correlation), while Intervention (Rung 2) simulates an experiment by forcing a change in a variable. Rung 2 answers 'what if we do X?'

• Mechanism: Association uses conditional probability, $P(Y|X)$, describing what we expect to see given an observation. This is 'plain vanilla statistics' and is susceptible to confounding.
• Execution Flow: Intervention uses the $do()$ operator, $P(Y|do(X=x))$, which mathematically simulates an external manipulation by severing all incoming causal arrows to $X$ in the causal graph.
• Best Practice: Intervention analysis requires a Directed Acyclic Graph (DAG) to identify and block confounding paths, ensuring the observed effect is truly due to the manipulation and not a spurious correlation.
• Example: A Rung 1 question is: 'What is the probability of sickness given a person was observed taking the vaccine?' A Rung 2 question is: 'What is the probability of sickness if we forced everyone to take the vaccine?'

# Conceptual representation of Intervention (Rung 2)
# P(Outcome | do(Treatment=1)) calculates the effect of forcing treatment,
# ignoring factors that might normally cause someone to choose treatment.

# Standard Association (Rung 1):
# P(Sickness | Observed(Vaccine=1))

# Causal Intervention (Rung 2):
# P(Sickness | do(Vaccine=1))

Key Takeaway: Association describes what is; Intervention describes what would happen if we forced a change, simulating an experiment.

3. Counterfactual Reasoning (Rung 3)

Counterfactuals ask 'what would have happened to a specific individual if they had acted differently?' This requires modeling a hypothetical scenario that explicitly conflicts with observed reality.

• Mechanism: Rung 3 requires conditioning on two facts about a specific unit (e.g., Jane): the actual action taken (e.g., Jane did not vaccinate) and the actual outcome observed (e.g., Jane got sick).
• Advanced Concept: The inherent difficulty lies in the conflict: you must model an alternative outcome for an individual whose observed data already defines their reality, making the modeling inherently more challenging than population-level intervention.
• Conditioning: Unlike Rung 2, which conditions only on the intervention, Rung 3 conditions on the full observed history of the individual, requiring the model to be fully specified to handle the hypothetical change.
• Tooling: Counterfactual analysis relies on the full specification of the Structural Causal Model (SCM), including the functional forms of the relationships, because the DAG structure alone is insufficient to determine individual outcomes.

Key Takeaway: Counterfactuals are unique because they condition on both observed action and observed outcome before modeling the alternative.

4. Formalizing Causal Assumptions

Causal modeling requires explicitly stating assumptions about the data generating process, shifting focus from manipulating data inputs to specifying the underlying mechanism. The complexity of assumptions scales with the rung.

• Workflow: The analytical focus shifts fundamentally from standard statistical data preparation (e.g., applying log transforms, discretization, or outlier removal) to specifying the causal mechanism itself.
• Best Practice: Start by drawing a Directed Acyclic Graph (DAG) to visualize and formalize assumptions about the relationships between variables, particularly identifying potential confounders and mediators.
• Advanced Concept: For Rung 3 questions, DAGs are insufficient; you must also state non-graphical, mechanistic assumptions, such as monotonicity (assuming the treatment never negatively affects the outcome for any individual).
• Rigor: The required rigor and specificity of mechanistic assumptions increase dramatically as you move up the hierarchy, demanding that the analyst be transparent about all theoretical inputs into the model.

# Example of a non-graphical assumption required for Rung 3:
# Assumption: Monotonicity
# The treatment (T) has a non-negative effect on the outcome (Y) for all individuals.
# Y_t=1(u) >= Y_t=0(u) for all units u.
# This assumption cannot be represented solely by the DAG structure.

Key Takeaway: Causal analysis is defined by the explicit statement of assumptions about the underlying mechanism, not just data manipulation.

Topics Covered in Causal Inference Primer — Beyond Correlation

1. Causal Hierarchy Introduction (00:00 - 00:39) — The lesson introduces Judea Pearl's three-rung ladder: Association, Intervention, and Counterfactual.
2. Modeling vs. Data Prep (01:26 - 02:22) — Causality is defined as thinking about the underlying data generating process, not just manipulating observed data.
3. Rung 1 and Rung 2 Defined (02:22 - 03:07) — Association is standard statistics, while Intervention addresses 'what if' questions that emulate controlled experiments.
4. Counterfactual Reasoning (03:07 - 04:36) — Counterfactuals model hypothetical scenarios that conflict with observed data, requiring conditioning on both observed action and outcome.
5. Formalizing Assumptions (04:36 - 05:54) — Higher rungs require explicit causal assumptions, often needing mechanistic assumptions beyond the simple DAG structure.

Statistics for Data Science Cheat Sheet

• Association (Rung 1) — Observed correlation or conditional probability

  P(Y | X)
  

• Intervention (Rung 2) — Simulates an experiment by forcing a variable value

  P(Y | do(X=x))
  

• Counterfactual (Rung 3) — Hypothetical outcome for a specific unit

  Y_x(u)
  

• DAG — Graphical model of assumed causal structure

• Monotonicity — Non-graphical assumption about effect direction

  Y_1(u) >= Y_0(u)
  

Comparison Table

Common Pitfalls

• Mistake: Treating correlation (Rung 1) as causation. Avoid: Use the do() operator to test intervention claims (Rung 2).
• Mistake: Assuming a DAG is sufficient for Rung 3. Avoid: Explicitly state non-graphical mechanistic assumptions like monotonicity.
• Mistake: Focusing on data transformation instead of mechanism. Avoid: Specify the underlying data generating process first.
• Mistake: Confusing Rung 2 and Rung 3 questions. Avoid: Rung 3 must condition on observed action AND outcome.

FAQs

• What is the core difference between Rung 2 and Rung 3? Rung 2 asks about populations (policy effect); Rung 3 asks about specific individuals (hypothetical history) and requires conditioning on observed outcomes.
• Why is data manipulation insufficient for causal modeling? Causal modeling requires specifying the data generating process and its mechanisms, not just transforming the observed data to fit a statistical model.
• What is the purpose of the $do()$ operator? It mathematically simulates an external intervention, cutting off all incoming causal arrows to the manipulated variable to isolate the causal effect.