How does a causal map change what you do with a strong correlation?
Before any model runs, a map lists the rival explanations, finds predictions the data could contradict, and asks whether the data can answer the question at all. On Barn Door Farm, a zinc–fibre correlation of 0.98 across five wheat samples became an honest “not yet” and a plan to measure soil, roots and grain at thirty spots.
Last time's quick check: a good instrument moves the treatment, and has no other route to the outcome. Today we put the whole module to work on one real case. Back to the wheat farm from the start of the module. Five grain samples from one harvest. Grain zinc tracks fibre with a correlation of 0.98.
Without a map the temptation is obvious: whatever raised the fibre raised the zinc, so let's do more of it. A map does three things before any model runs. First, it lists the rival explanations. The published case asks several questions of the same pattern. Could a foliar zinc spray explain it? The farming programme? Nitrogen? The structure of the grain itself? Which soil pathways might be involved? Each is a different arrangement of arrows.
Second, it finds what can be tested. If bran is a common cause of zinc and fibre, then among samples with the same bran fraction, zinc and fibre should stop moving together. That's a testable implication: a prediction the map makes that the data can contradict.
Third, the map asks the hardest question first: can these data answer the question at all? That's called identification. Here, the answer is not yet. Five samples from one season. The soil values on the published page are illustrative, not measured on the farm. Two samples have provisional field links. And for the foliar zinc question, baseline soil sits on a back door and hasn't been measured.
No algorithm fixes that. A cleverer model run on five samples would just give a confident wrong answer. So the map becomes a plan. The published case lays out six possible next steps: start from the published evidence; complete or correct the records; measure what's missing; learn from changes already happening; run a controlled trial; or hold off on changing practice.
Here, the records come first: confirm the two uncertain field links and fill in the missing sowing date. Then measure what's missing: soil, roots and grain at the same thirty spots, six in each field, followed through a season, with yield and bran fraction recorded too.
No treatment is applied and nothing about the farming changes. The case is clear about its limits. It could show whether soil biology and grain minerals move together, and whether bran fraction explains the zinc–fibre pattern. It couldn't prove that a practice causes the difference. And five samples from one season don't justify changing practice yet.
Look at what the map did. Before it, the obvious move was to act on a correlation of 0.98. After it: a list of rival explanations, a test that could rule one out, an honest "not yet", and a plan to collect data that could actually answer the question. That's what causal inference changes. Not cleverer arithmetic first, but better questions first.
Asking whether the data can answer the causal question at all, before any model runs. In the wheat case the answer was “not yet”: five samples from one season, soil values on the published page that were illustrative rather than measured, two provisional field links, and baseline soil on a back door that hadn’t been measured.
Start from the published evidence; complete or correct the records; measure what’s missing; learn from changes already happening; run a controlled trial; or hold off on changing practice. Here the records came first, then measuring soil, roots and grain at the same thirty spots through a season, with yield and bran fraction recorded too.
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.