Intro Introduction: from field to plate
What the series covers and who it’s for, why telling a cause from a pattern matters at every link from farm to plate, and the journey ahead, module by module.
Why a pattern that predicts is not a cause, and what it takes to prove one.
A series of short films on cause and effect in farming and food: where the idea came from, how causal maps work, the study designs and methods that give fair answers, and what it takes to show that a farming practice changes the nutrition in food.
30 short films · about 2 hours · free
Every film, summary and transcript is free to watch and read: no sign-up needed. Get updates by email if you would like to hear what comes next.
Why a model can predict well and still give the wrong advice, and what machine learning and causal design each bring.
What the series covers and who it’s for, why telling a cause from a pattern matters at every link from farm to plate, and the journey ahead, module by module.
An AI that read the hospital instead of the lungs, and cattle that seemed to die from their antibiotics: why a pattern that predicts is not a cause, and what a confounder is.
What causal thinking gives AI, and what AI gives causal inference, through Greek cotton sowing advice and maize fertiliser in Ghana. Machine learning cleans up what you measured; design deals with the rest.
Two centuries of fair tests and arguments, from Lind’s sailors to the smoking debate, and the ideas they left us.
In 1747 James Lind gave sailors with scurvy six different remedies, and citrus worked. How a fair, controlled comparison can show that something works long before anyone knows why.
David Hume put a cause in terms of what would have happened without it. Why one of those two outcomes always stays hidden, and how correlation nearly pushed causation out of statistics.
Eijkman’s chickens and polished rice, Goldberger and pellagra: accidental changes in diet that showed how the way food is grown and processed can decide whether it nourishes.
Fisher and Neyman at Rothamsted, and a seaweed feed trial with Australian cattle: why letting chance allocate treatments, in blocks and with replicates, defeats confounders nobody has thought of.
Sewall Wright drew causes as arrows to study guinea pigs; his father Philip used an outside push to measure demand. How arrows make assumptions checkable, and what an instrument is.
Doll and Hill linked smoking to lung cancer; Fisher objected that a hidden cause might explain both. How Cornfield answered by asking how strong that hidden cause would need to be.
How to draw a causal map and read the roles variables play in it, from chains and forks to confounders and colliders.
What a causal map is, and why it is drawn before any data are analysed: boxes, arrows and the assumptions they put where others can check them, with Barn Door Farm’s published zinc map.
Chains and forks pass association along; colliders block it until you hold them fixed. The three shapes every causal map is built from, shown with zinc, bran and fibre in wheat.
How to find the back-door routes that would fool a comparison and close them without opening new ones, with Italian organic olives and Greek cotton. A covariate is a label, not a role.
A mediator sits on the path from cause to effect. Why adjusting for one removes part of the effect you want, explained with a farm-and-nutrition trial in Kenya.
Choosing which data to look at can create links that aren’t there. Collider and selection bias in laboratory samples, animals that reach slaughter and the studies that get published.
The roles that can’t simply be adjusted away: effect modifiers, instruments and proxies, with Ghana’s maize zones and cows’ own genes, and what to do about a cause nobody measured.
How a causal map turns a striking correlation on Barn Door Farm into rival explanations, an honest “not yet” and a plan for what to measure next.
The study designs that turn a question into an answer: trials, matching, before-and-after comparisons and nature’s own experiments.
Randomised trials on real farms: strips, plots, herds and whole communities. Why the unit you randomise matters, and why the outcome must be the one you actually care about.
Matching and weighting build a fair comparison without randomising, as with Italian organic olives and seed access in India. They balance what you measured, never what you didn’t.
Before and after needs a comparison. How difference-in-differences and interrupted time series estimate what would have happened anyway, with UK checkout rules, carbon labels and Mexico’s warning labels.
Sometimes the world allocates the treatment for you: John Snow’s cholera study, a farm-size threshold, California’s tobacco programme and the genes cows are born with.
When one farm’s treatment reaches its neighbour’s outcome, simple comparisons go wrong. Spillovers from Bt maize and New Zealand farmer groups, and how to design a study around them.
Observational studies said beta-carotene and hormone therapy protected health; trials said otherwise. Why observation misled, and how designing the analysis as the trial you wish you’d run helps.
What machine learning adds to causal inference, how to stress-test an answer, and what AI can and can’t do with causes.
Double machine learning, step by step: two models, their leftovers and cross-fitting. Why small errors barely move the answer, and the three assumptions no machinery removes.
Averages can hide big differences. How causal forests find where an effect is bigger or smaller, as with maize in Ghana, how honesty keeps them truthful, and what they can’t tell you.
Four ways to try to break a causal estimate before anyone relies on it, from a fake treatment and a negative control to asking how strong a hidden confounder would need to be.
What causal discovery and language models can and can’t do. They can draft a causal map; experts, tests and good design still decide whether it is right.
Following the chain from soil to health, and choosing the smallest next step that can answer a question.
Each link from soil to health needs its own evidence. Why crop minerals appear to have fallen, what the Broadbalk archive shows about breeding for yield, and how growing can change eating.
Finland’s selenium fertiliser, orange-fleshed sweet potato and a farm-and-nutrition trial in Kenya: how farming has deliberately changed what people eat, and why those cases convince.
A strong correlation starts an investigation; it doesn’t end one. How to choose the smallest next step that can answer the question, and design a fair test when one is needed.
The series ends where it began, with an AI that was right for the wrong reason: the seven things it takes to be right for the right reason, and why each matters.
Causal evidence is changing how farming is measured and rewarded. Leave your email and we will tell you when new films, courses and findings are published.
Prefer to read online? Vitagri also writes on Substack. See our privacy policy for how we use your email.
The Data Agronomy course teaches the predictive and forecasting models behind nutrient-dense food, with certificates at two levels. The causal inference glossary defines every term the films use.