Skip to content
Coordinates of ThoughtHome of Philosophy
Join
Debate

Thesis A correlation explains nothing until someone shows the mechanism

Wanjiru Kamau Contributor
· 6 replies · 228 views
Scientific methodAnalyticWesley Salmon#correlation#credit-scoring#mechanism#statistical-explanation
Follow

Where it stands

I build credit-scoring models for a mobile-money company. Our data say that people who charge their phones at regular times are more likely to repay small loans. The effect is real, stable and useful for prediction.

But when a manager says this explains repayment, I object. It explains nothing. Perhaps regular charging indicates regular work, or a stable home with electricity, or a personality that likes routine. Until we know which, we have a number, not an understanding.

There's a history to this in philosophy of science. Wesley Salmon first proposed that explaining an event means citing factors that are statistically relevant to it, that change its probability. Later he moved to a causal-mechanical account: you have explained something when you have shown the causal process that produced it. I think his second thoughts were right.

My thesis: statistical relevance on its own is never an explanation. It can justify prediction, sometimes action, but to explain you need a mechanism. Who disagrees?

LikeReply
Follow6 replies

Members reply here. Reading is always free.

Sign in to reply

6 replies

For the thesis 2

  1. Priya Raghavan

    For Fellow

    From cognitive science, I'm with Wanjiru. Brain imaging gives endless correlations between a region and a task. For years these were reported as if they explained the task. They don't. They tell you where to look. The explaining starts when you know what the region is computing and how.

    Helpful · 3
  2. Nur Aisyah Rahman

    For Fellow

    There's a fairness side to this. When a score relies on a correlation nobody can explain, people get judged on proxies they cannot see or change. If charging habits stand in for having electricity at home, the model penalises poverty while appearing to measure character. A demand for mechanism is also a demand for accountability.

Against the thesis 3

  1. Pieter de Jong

    Against Fellow

    I'll disagree. Aspirin relieved pain for most of a century before anyone discovered how it acts on prostaglandins. Were doctors before that unable to explain why a patient's headache went away? They could say “because she took aspirin”, and that's a perfectly good causal explanation without the mechanism. You're conflating causal with mechanistic.

    Helpful · 2
  2. Alistair Gow

    Against Contributor

    John Snow traced the 1854 Soho cholera outbreak to the Broad Street pump by mapping cases, and the pump handle came off before the medical establishment accepted any mechanism of waterborne infection. His map was not a full explanation, but it was enough of one to save lives. I'd say a correlation backed by the right comparisons can explain that something matters, even before anyone knows how.

    Helpful · 1
  3. Ignacio Fuenzalida

    Against Contributor

    Biologist's footnote: “mechanism” is also selective. Every mechanism we cite leaves out most of what's going on in the cell. We pick the parts relevant to the question. So even the mechanistic explanation Wanjiru wants is answering a particular why-question, not giving the whole story.

    Helpful · 3

Other replies

  1. Wanjiru Kamau

    Contributor

    Pieter and Alistair have moved me. I'll revise: a correlation alone explains nothing, but a correlation that survives the right comparisons, or that you can intervene on as Snow did, starts to identify a cause, and that is a real explanation even without the mechanism. My charging correlation is neither. We have never intervened, and nobody has tried to rule anything out. So I keep my objection to my manager, but on better grounds.

    Helpful · 2