Mathematics
Causal inference
Causal Inference & Experiments
Also known as correlation vs causation / A/B testing
Causal inference is how we figure out what actually causes what, instead of getting fooled by two things that just happen to move together. The core trick is imagining a parallel world: would the outcome still have happened if you'd changed one thing? That exact question shows up all over: courts use a 'but-for' test for blame, asking whether the harm would have happened otherwise, which is the same parallel-world logic statisticians use; in design, a prototype is really a cheap experiment, a guess you build fast just to try to prove it wrong; and in medicine, trials tell you what works on average, while precision medicine fires back that the 'average patient' doesn't actually exist. Learning to spot true causes is one of the most powerful mental tools there is, because our brains are wired to see cause everywhere, even where there's only coincidence.
Read
- Counterfactuals and Causal InferenceStephen L. Morgan · 2007Book
- CausalityJudea Pearl · 2000Book
- Causal Inference in StatisticsJudea Pearl · 2016Book
- Artificial Intelligence and Causal InferenceMomiao Xiong · 2022Book
Watch
- Causal Inference - EXPLAINED!CodeEmporiumVideo
- Causal Inference with Machine Learning - EXPLAINED!CodeEmporiumVideo
- Causal Inference: S/T/X Learner ExplainedXuanVideo
Listen
- Casual InferenceLucy D'Agostino McGowan and Ellie MurrayPodcast
- Linear DigressionsKatie MalonePodcast
Voices to follow
- Goran S. Milovanović@GSMilovanovic · XSerbian cognitive and data scientist
- Luigi Gresele@luigigres · XResearcher, University of Copenhagen
Debates
- Can causal inference be fully automated by AI?One view: Advanced algorithms and large datasets can potentially discover complex causal relationships without human bias. · Another: Human domain expertise is essential for defining research questions, interpreting results, and addressing ethical implications.Open question
- Should causal inference prioritize explainability or predictive accuracy?One view: Explainability is crucial for understanding mechanisms and building trust, especially in high-stakes applications. · Another: High predictive accuracy is often more valuable for practical decision-making, even if the underlying causal mechanisms are less clear.Open question
- Is it always necessary to identify the 'true' causal mechanism?One view: Understanding the precise mechanism is vital for developing effective interventions and generalizable knowledge. · Another: Sometimes, knowing *that* an intervention works is sufficient for practical purposes, even without full mechanistic understanding.Open question
Glossary
- CausalityA relationship where one event directly influences another, causing it to happen.
- CorrelationA statistical relationship between two variables that move together, but one does not necessarily cause the other.
- ConfoundingWhen a third, unobserved variable influences both the presumed cause and effect, creating a misleading association.
- Treatment EffectThe causal impact of an intervention or exposure on an outcome, compared to not receiving it.
- CounterfactualWhat would have happened to an individual or system if they had experienced an alternative treatment or exposure.
- Directed Acyclic Graph (DAG)A visual model using nodes and arrows to represent assumed causal relationships between variables.
Careers
Roles this can lead toward
Threads 9
Where this connects to other fields — and why it's worth knowing.
- Private Law: Contract, Tort & Property Law
When a court asks 'did the driver's texting cause the crash?', it uses the 'but-for' test: would the crash still have happened if he hadn't been texting? That means imagining a parallel world where he kept his eyes on the road. Statisticians do the exact same thing to prove cause and effect, comparing our world to the one that didn't happen.
- Design Thinking Arts
When a designer builds a rough prototype, they're really running a science experiment. The prototype is a guess, 'people will find this easier,' built cheap so you can quickly prove it wrong. That's the same logic as A/B testing, where you try two versions and let real results decide. Good design is testing hypotheses first, and looking pretty second.
- Psychotherapy Psychology
Compare wildly different kinds of talk therapy and a strange thing shows up: they mostly work about equally well. It's nicknamed the 'dodo bird verdict,' after the dodo who declared 'everybody has won.' That hints the fancy branded technique matters way less than we assume, which is a nightmare for researchers trying to prove their method is the best one.
- Quantum Computing Technology
Two 'entangled' particles can be split across the galaxy yet always match up perfectly when measured, spookily in sync. But here's the twist: neither one sends any signal to the other, so one can't be causing the other. It's nature's cleanest proof that two things moving perfectly together does not mean one is pulling the other's strings.
- Precision Medicine Health
A big drug trial tells you a pill helps people on average. But precision medicine points out there's no such thing as the 'average patient' sitting in the room, only this specific person with their own genes. So doctors get caught in a tug-of-war between what worked for the crowd and what will actually work for you.
- Free Will and Responsibility Philosophy
Saying 'X caused Y' secretly claims that in some other world where X didn't happen, Y wouldn't have either. That imaginary 'other world' is also how we assign blame and praise: he could have chosen differently. So a deep question about free will is quietly baked into every statistics class about cause and effect.
- AI in Medicine Health
A medical AI that just spots patterns can be confidently wrong the moment doctors change how they treat people, because the patterns it memorized shift too. To be safe it has to learn what actually causes what, not just what usually appears next to what.
- Carbon Markets & Net Zero Environment
When you pay to protect a forest as a 'carbon offset,' it only counts if those trees really would have been chopped down otherwise. So you're buying a what-if that never happened. And proving a what-if is the hardest thing in all of statistics.
- Environmental and Climate Law Law
You couldn't sue anyone over a heatwave, because who's to say it wasn't just bad luck? Now "attribution science" can measure how much more likely human pollution made this exact storm. The lawsuit was waiting on the math to catch up.