descriptive-evidence

Descriptive Evidence

Overview

Causal identification asks what an intervention did; structural estimation asks what a world we haven't seen would do; prediction asks what is likely true of a unit, so I can act on it. Beneath all three sits the work the discipline rushes past: describing what is actually in the data — the trend, the gap, the distribution, the summary-statistics table, the stylized facts that open almost every empirical paper. The deliverable is a faithful picture of the data — often the whole job, otherwise the thing that motivates the causal, structural, or predictive question that follows.

This is a legitimate destination, not a way-station: "just show me the trend" does not mean "skip the rigor" — it means the rigor is about the description itself.

The signature failure of this arm is the composition / aggregation artifact — a trend or gap that looks within-group but is really a shift in the mix (the aggregate wage rose because low-wage workers exited, not because anyone got a raise; a national rate can fall while it rises in every region — Simpson's paradox). It reconciles, reproduces, and plots beautifully while quietly describing the denominator or the changing sample, not the thing you named — the descriptive analog of leakage: a confident artifact.

The twin failure is the descriptive-to-causal slide: Y rises alongside X, and the write-up says "X raised Y." A described co-movement is a correlation; the moment it wears a causal verb it has left this arm for a claim it hasn't earned. A striking stylized fact's highest use is to motivate a causal question, which routes to causal-identification and earns the causal verb there — not here.

Core principle: a descriptive fact is trustworthy only when it survives the ordinary alternative explanations for the pattern — composition, deflation, scaling, selection, definition — and is stated as the correlation it is, not the cause it isn't. Everything below serves that one sentence.