00. Getting a Local LLM Running
Sets up local model execution so later AI-assisted lectures can compare model behavior without depending on remote APIs.
AI for Causal Inference
AI
LLMs
Causal Inference
Lecture Notes
This course treats AI as an assistant to the causal analyst. LLMs, RAG systems, agents, and automation can help with ideation, documentation, critique, code scaffolding, literature synthesis, diagnostics, and reporting. Identification judgment remains the analyst’s responsibility.
By the end, a reader should be able to run and compare local models, use structured outputs, build RAG support for assumptions, automate diagnostic work, design causal-analysis agents, and evaluate where AI outputs are brittle, incomplete, or misleading.
The live parts of the course are designed for a local GPU workstation rather than a hosted API. The notebooks use Hugging Face Transformers for direct model loading, with Qwen/Qwen2.5-0.5B-Instruct as a smoke test, Qwen/Qwen2.5-7B-Instruct as the working default, Qwen/Qwen2.5-14B-Instruct and Qwen/Qwen2.5-32B-Instruct for scale comparisons, and Phi, Mistral, Gemma, and Llama models as family comparisons. A high-memory GPU, roughly 48 GB to 96 GB of VRAM, makes the larger comparisons practical. Smaller machines can still run the smoke-test and 7B workflows, or use quantized models. The technical design separates causal prompts, structured-output parsing, retrieval indices, evaluation rubrics, and model execution, so the same exercises can run through direct Transformers loading or through a local serving layer such as Ollama, llama.cpp, or vLLM when a persistent model server is preferred.
Lecture Sequence
01. AI-Assisted Causal Workflow
This lecture develops AI-Assisted Causal Workflow with examples that make assumptions, diagnostics, and interpretation visible.
02. LLM Basics for Causal Analysts
This lecture uses LLM Basics for Causal Analysts to clarify the analyst’s question, evidence, assumptions, and decision implications.
03. Turning Business Questions into Causal Questions
This lecture turns business questions into causal questions by clarifying actions, outcomes, populations, and estimands.
04. Estimand Cards and Causal Design Documents
This lecture frames Estimand Cards and Causal Design Documents as a decision problem and asks what evidence can be trusted, challenged, and communicated.
05. AI-Assisted DAG Brainstorming
This lecture builds intuition for AI-Assisted DAG Brainstorming and ties the result to model choice, uncertainty, and action.
06. DAG Critique, Variable Roles, and Backdoor Paths
This lecture applies DAG Critique, Variable Roles, and Backdoor Paths with emphasis on diagnostics, tradeoffs, and evidence limits.
07. RAG for Causal Domain Knowledge
This lecture develops RAG for Causal Domain Knowledge as a practical pattern for analysis, diagnostics, and decision support.
08. Literature Synthesis for Causal Assumptions
This lecture connects Literature Synthesis for Causal Assumptions to retrieval, structured outputs, evaluation, and AI brittleness.
This lecture develops Dataset Profiling with AI with examples that make assumptions, diagnostics, and interpretation visible.
10. Detecting Bad Controls, Post-Treatment Variables, and Leakage
This lecture uses Detecting Bad Controls, Post-Treatment Variables, and Leakage to clarify the analyst’s question, evidence, assumptions, and decision implications.
11. Synthetic Data Generation for Causal Teaching
This lecture uses synthetic data generation to make causal assumptions, failure modes, and teaching examples explicit.
12. Simulation Labs for Assumption Stress Testing
This lecture frames Simulation Labs for Assumption Stress Testing as a decision problem and asks what evidence can be trusted, challenged, and communicated.
13. AI-Assisted Method Selection
This lecture builds intuition for AI-Assisted Method Selection and ties the result to model choice, uncertainty, and action.
14. AI-Assisted Causal Code Generation
This lecture applies AI-Assisted Causal Code Generation with emphasis on diagnostics, tradeoffs, and evidence limits.
15. Automating Balance, Overlap, and Diagnostic Reports
This lecture develops Automating Balance, Overlap, and Diagnostic Reports as a practical pattern for analysis, diagnostics, and decision support.
16. AI for Sensitivity Analysis
This lecture connects AI for Sensitivity Analysis to structured prompting, evaluation, and clearer causal judgment under uncertainty.
17. AI for Experiment Design and Power Planning
This lecture develops AI for Experiment Design and Power Planning with examples that make assumptions, diagnostics, and interpretation visible.
18. AI for Quasi-Experiment Design
This lecture uses AI for Quasi-Experiment Design to clarify the analyst’s question, evidence, assumptions, and decision implications.
19. Causal Report Generation with LLMs
This lecture studies causal report generation with LLMs through structure, evidence traceability, and review quality.
This lecture frames Causal Analysis Agent as a decision problem and asks what evidence can be trusted, challenged, and communicated.
21. Multi-Agent Causal Review Workflow
This lecture builds intuition for Multi-Agent Causal Review Workflow and ties the result to model choice, uncertainty, and action.
22. Evaluating AI Outputs in Causal Workflows
This lecture applies Evaluating AI Outputs in Causal Workflows with emphasis on diagnostics, tradeoffs, and evidence limits.
23. Hallucination and Failure Modes in AI Causal Analysis
This lecture develops Hallucination and Failure Modes in AI Causal Analysis as a practical pattern for analysis, diagnostics, and decision support.
24. Capstone AI-Assisted Causal Project
Brings the AI-assisted causal workflow together in an end-to-end project, from question framing to evaluated causal output.