zirui-song Power First
Diagnostics come before interpretation. Never write up a second stage as a finding until this pass has run and passed.
Report these four things first, in this order
- First-stage F (IV/2SLS only) — Kleibergen-Paap rk Wald F for the actual clustering used, not the homoskedastic default. Report the exact value and the number of instruments. Below ~10, say the first stage is weak and stop there.
- Effective N — not the raw row count. Report observations actually used after FE absorption and singleton dropping, plus the number of clusters in the clustering dimension. Few clusters is a power problem even with many rows.
- Minimum detectable effect at 80% power — in the paper's own units (e.g., basis points
of spread), given the estimated SE:
MDE ≈ 2.80 × SE(two-sided, α = 0.05, 80% power). State it as "this design cannot rule out effects smaller than X bps." - Two most likely violations of the identifying assumption — named concretely for this setting, not generic. For each, say what evidence in the data would reveal it and what test would falsify it.
Verdict, stated before any coefficient discussion
- Weak first stage (F < 10) → say so first. Do not report or interpret the second-stage coefficient as a finding.
- Underpowered (MDE larger than the effect size the theory predicts, or larger than comparable published estimates) → the null is uninformative. Say "underpowered," not "no effect."
- Informative null (MDE comfortably below the economically interesting effect) → this is a real finding. Write it up as one.
Conventions
- An honest null beats a reframed one. Do not go hunting for a subsample where it works and lead with that.
- Report coverage limits up front: sample window, which firms are in the panel, what fraction of the population that is.
- When the user asks "is this result real," this skill runs before any other answer.