Causal Inference with LLMs

Leverage reasoning LLMs to discover causal relationships and generate directed acyclic graphs (DAGs). This skill enables Pearl's Ladder of Causation "Why" level - inferring underlying data generation processes by combining observational patterns with domain expertise.

Core Capabilities

1. Pairwise Causal Relationships (A → B)

Assess potential causal links between treatment and outcome variables using LLM reasoning and domain knowledge.

from causal import ModelSuggester

modeler = ModelSuggester('gpt-5.1-mini')
experts = modeler.suggest_domain_expertises(event_labels, n_experts=5)

modeler.suggest_pairwise_relationship(
    experts,
    'Treatment event',
    'Outcome event'
)

2. Latent Confounder Identification (A ← C → B)

Identify unobserved variables that influence both treatment and outcome.

_, latent_confounders = modeler.suggest_latent_confounders(
    treatment_variable,
    outcome_variable,
    experts
)

3. Complete Causal Graph Construction

Build full DAGs with directional relationships, reasoning, and confidence estimates.

all_factors = ["factor1", "factor2", "factor3"]
pairwise_edges = await modeler._suggest_pairwise_relationships(
    expert=str(domain_expertises),
    all_factors=all_factors,
    analysis_context="causal model"
)

4. Interactive Graph Visualization

Generate interactive HTML visualizations of causal graphs.

from causal import plot_interactive_graph
import networkx as nx

G = nx.MultiDiGraph()
for idx, label in enumerate(all_factors):
    G.add_node(idx, desc=label)

for edge, (confidence, thinking) in pairwise_edges.items():
    G.add_edge(edge[0], edge[1], relationship=confidence, thinking=thinking)

plot_interactive_graph(G, filename="output/causal_graph.html")

Workflow

Basic Analysis Flow

1. Define Variables: Specify treatment, outcome, and known factors
2. Identify Domain Expertise: Generate relevant expert perspectives
3. Discover Confounders: Find latent variables affecting both treatment and outcome
4. Build Pairwise Relationships: Assess all potential causal links
5. Construct DAG: Assemble complete causal graph
6. Visualize: Generate interactive graph representation

Example: Medical Causality

import asyncio
from causal import ModelSuggester, plot_interactive_graph
import networkx as nx

async def analyze_medical_causality():
    modeler = ModelSuggester('o3-mini')

    all_factors = ["smoking", "lung cancer", "air pollution", "genetics"]
    treatment = "smoking"
    outcome = "lung cancer"

    # Get domain expertise
    experts = modeler.suggest_domain_expertises(all_factors, n_experts=3)

    # Find confounders
    _, latent_confounders = modeler.suggest_latent_confounders(
        treatment, outcome, experts
    )

    # Build complete factor list
    all_factors += latent_confounders
    all_factors = list(set([f.lower() for f in all_factors]))

    # Discover pairwise relationships
    pairwise_edges = await modeler._suggest_pairwise_relationships(
        expert=str(experts),
        all_factors=all_factors,
        analysis_context="medical causal model"
    )

    # Build and visualize graph
    G = nx.MultiDiGraph()
    for idx, label in enumerate(all_factors):
        G.add_node(idx, desc=label)

    for edge, (confidence, thinking) in pairwise_edges.items():
        G.add_edge(edge[0], edge[1],
                   relationship=confidence, thinking=thinking)

    plot_interactive_graph(G, filename="medical_causal_graph.html")

asyncio.run(analyze_medical_causality())

Configuration

The package uses config.yaml for LLM settings. Key parameters:

• model: LLM model name (e.g., 'gpt-5.1-mini', 'o3-mini')
• temperature: Reasoning randomness
• max_tokens: Response length limit
• domain_expert_count: Number of expert perspectives (default: 3-5)

Critical Limitations

AI reasoning is based on learned patterns and may contain biases or incorrect associations. Always:

• Validate proposed relationships using additional causal inference techniques
• Cross-check with domain experts and empirical evidence
• Test causal hypotheses through experimentation when possible
• Document confidence levels and reasoning chains
• Never treat LLM suggestions as ground truth

Advanced Features

Relationship Strategies

Multiple strategies available via RelationshipStrategy enum:

• parent: Parent-child relationships
• child: Child-parent relationships
• confounder: Confounding variables
• iv: Instrumental variables
• mediator: Mediating variables
• pairwise: All pairwise relationships

Async Processing

Use async methods for efficient batch processing of multiple causal pairs.

Graph Export

Export graphs in multiple formats:

• Interactive HTML (pyvis)
• NetworkX graph objects
• JSON/GraphML for external tools

References

For theoretical background, see references/causal-theory.md.

For complete API documentation, see references/api-reference.md.