Understanding Causality: Fundamentals of Causal Inference

Part One of a Three-Part Series

If you’ve ever wondered, “Is it really the cause, or just a coincidence?” then this series on Causal Inference is for you. Machine learning models can unravel striking correlations, but understanding why something happens—its actual cause and effect—requires a whole new lens. That’s where Causal Inference steps in.

Yesterday, I kicked off the first session of my three-part series on Causality in Machine Learning at Professor Qiyun’s Lab in the Biodesign Center at Arizona State University. Here’s a brief recap of what we covered, why it matters, and what’s on the horizon.

Why Causal Inference?

“Data alone is not enough. To interpret data, you need a model of the process that generates the data.” – Judea Pearl.

Bridging Correlation and Causation

While conventional machine learning shines at predicting outcomes, it often stumbles on why those outcomes occur. Causal Inference provides the methodology to bridge this gap, offering insights into cause-and-effect relationships that can inform real-world decisions—from biomedical research to economic policies.

Key Highlights from Session One

1. The Essence of Causality

• Correlation vs. Causation: Why knowing that “X is correlated with Y” doesn’t guarantee “X causes Y.”
• Practical Examples: How correlation can mislead in scenarios like dietary habits and disease risk.

2. Randomized Controlled Trials (RCTs)

• The Gold Standard: RCTs are the benchmark for causal claims—think clinical drug trials.
• Limitations: Ethical and logistical constraints make RCTs impossible in many scenarios (e.g., testing harmful exposures).

3. Challenges in Causal Inference

• Confounders & Bias: These hidden variables and selection biases can skew results.
• Counterfactual Reasoning: Imagining alternate universes—what if a patient hadn’t received a specific treatment?

4. Causal Graphs & Directed Acyclic Graphs (DAGs)

• Visualizing Cause-and-Effect: DAGs are your roadmap for identifying where interventions might matter.
• Real-World Modeling: Examples of DAGs in epidemiology, public policy, and microbiome research.

5. Foundational Assumptions

• Causal Markov Condition: Connecting DAG structure to independence assumptions.
• SUTVA & Ignorability: Key pillars for ensuring accurate estimation of causal effects.

6. Estimating Average Treatment Effects (ATE & CATE)

• Understanding Treatment Impact: How interventions affect the overall population and subgroups.
• Case Studies: Using example datasets to illustrate how different treatments can yield varied effects across different segments.

Beyond RCTs: Alternative Approaches

When RCTs aren’t feasible, researchers turn to these powerful tools:

• Instrumental Variables (IVs): Using external factors to measure causality in non-random scenarios.
• Difference-in-Differences (DiD): Comparing changes between treated and untreated groups over time.
• Propensity Score Matching (PSM): Matching similar individuals in treatment and control groups to reduce selection bias.
• Synthetic Control Methods: Crafting a “synthetic” comparison group by weighing multiple untreated units.

Coming Soon: Modern Causal ML Techniques

The next session will dive deeper into cutting-edge methods that blend the best of machine learning with causal inference:

1. Structural Causal Models (SCMs): Encoding causal relationships through structural equations for more precise cause-and-effect insights.
2. Causal Discovery Algorithms
3. Double Machine Learning (DML): Harnessing ML models to manage high-dimensional confounders and refine treatment effect estimates.
4. Invariant Causal Prediction (ICP): Robust, valid causal discovery across different data distributions.
5. Counterfactual and Interventional ML

The Practical Finale: Hands-On with Microbiome Datasets

Our third session will bring everything full circle with Jupyter Notebook demos. We’ll explore:

• IBD200: Investigating the causal impact of microbiome composition on inflammatory bowel disease progression.
• HMP1 & HMP2: Identifying key microbial interactions that drive health outcomes.

We’ll implement DoWhy, EconML, and CausalML to estimate causal effects, visualize causal graphs, and validate our assumptions on real-world data.

Final Thoughts

Causal Inference is more than an academic exercise—it’s a transformative approach that empowers researchers to make data-driven decisions grounded in why something happens, not just what happens. By merging robust causal methodologies with practical machine-learning techniques, we can push the frontiers of biomedical research, economics, and countless other fields.

Stay tuned for the next installment, where we’ll explore modern causal machine learning and showcase how these techniques can revolutionize your data-driven discoveries.