← index

On causal noise

A short fragment on why correlation is loud and causation is quiet — and how to listen for the difference.

placeholder fragment — replace with your own words.

Most of what a model sees is noise wearing the costume of signal. Correlation is loud: it shows up everywhere, it is easy to measure, it flatters the dashboard. Causation is quiet. It hides under interventions you never ran.

The quiet test

A useful habit before trusting any effect:

  1. Imagine the intervention. What would you actually do to the system?
  2. Ask who is missing from the data — the unobserved confounder is the ghost in every regression.
  3. Only then look at the number.
# a sketch, not a library
effect = E[y | do(x=1)] - E[y | do(x=0)]

The interesting work is rarely in the estimator. It is in earning the right to write do(...) at all.

Why it matters here

These notes are a place to think out loud about uplift, identification, and the small machinery that makes decisions less superstitious. Expect fragments, not papers.