/mstack:analyze

Stage: analyze · Voice: analyst

After /mstack:data-clean, /mstack:codebook, and /mstack:identification-review. If preregistered (/mstack:preregister), the primary analysis in prereg/osf-prereg.md is the contract this skill executes; otherwise the specification confirmed with the user is written down before any model runs, so that what follows is not a search.

Procedure

1. Load .mstack/hypotheses.md (the estimand for each hypothesis), prereg/osf-prereg.md if it exists, .mstack/identification-review-*.md (the assumption, the falsification tests, the sensitivity analysis), .mstack/config.yaml (design.type, stats.primary_spec), .mstack/learnings.jsonl (variable names, conventions), data/clean/analytic.rds, data/codebook.md.
2. Read the General section and the design section of ${CLAUDE_PLUGIN_ROOT}/references/estimation-conventions.md.
3. Confirm the spec with the user before coding, or transcribe it from the prereg: estimand; outcome, treatment / IV, controls (each pre-treatment); the estimator the design calls for (fixest::feols with fixed effects for panels; sunab, did, or didimputation under staggered adoption; rdrobust for RD; IV with weak-instrument-robust intervals; PPML for flows with zeros; lm_robust with Lin adjustment for experiments; conjoint AMCEs with marginal means); fixed effects and why; standard errors clustered at the assignment level, with the wild cluster bootstrap when clusters are few; sample restrictions matching the prereg; marginaleffects for anything that is not the coefficient itself.
4. R conventions: r-coding-skills if installed, else ${CLAUDE_PLUGIN_ROOT}/references/r-conventions.md.
5. Write code/02-analyze.R: header (purpose, inputs, outputs, run order, prereg link); load the clean data, never re-clean; fit the primary first as m_primary; then the falsification tests the identification review named (placebo outcomes, pre-trends, negative controls) as m_placebo_*; then pre-specified secondaries as m_<descriptor>; compute the estimand from each model with marginaleffects where the coefficient is not the quantity; save model objects to output/models/*.rds for 03-figures.R and 04-tables.R; one-line summary per model (estimate, interval, N, clusters, estimator) to output/analyze-log.md. Exploratory specifications go in code/02b-exploratory.R, never in this file.
6. Write code/04-tables.R: Table 1 descriptives for the analytic sample (by arm with a balance column for experiments); Table 2 the primary specification, the table the paper is built around; Table 3+ falsification tests and secondaries. modelsummary::modelsummary(output = "latex") to output/tables/*.tex, which the manuscript \inputs; readable coefficient labels; notes stating the estimator, the standard-error type and cluster level, and what the coefficient is (a 1-SD change, an AME). Coefficient + 95% CI; stars only if the journal demands them.
7. Run Rscript code/02-analyze.R && Rscript code/04-tables.R; do not declare done with errors outstanding.
8. Sanity check: N matches the analytic dataset or a documented exclusion; the number of clusters is printed and adequate; for fixed-effects designs, the share of units with within-variation is printed; the headline's sign matches H1, or flag it loudly before anyone drafts results; the falsification tests behave (a placebo that "works" is a finding against the design); the substantive magnitude is stated in outcome units and as a share of the outcome SD.
9. Hand-off. Set paper.status: "analyzing" in .mstack/config.yaml. Do not run /mstack:robustness until the primary table is locked.

Outputs

• code/02-analyze.R, code/04-tables.R, optional code/02b-exploratory.R.
• output/models/m_primary.rds, m_placebo_*.rds, m_<secondary>.rds.
• output/tables/table-1-descriptives.tex, table-2-primary.tex, table-3-*.tex.
• output/analyze-log.md — one line per model.

Anti-patterns

• Unflagged off-prereg specs. Exploratory specs go in code/02b-exploratory.R and are labeled exploratory in the paper.
• Two-way fixed effects under staggered adoption. Use an estimator built for it.
• A coefficient reported as the estimand when it is not. Logit coefficients, interaction terms, and IV coefficients need the quantity computed.
• Hand-rolled tables. modelsummary, so reproduction is one command away.
• P-value chasing. If the primary spec does not give the headline, the headline changes; do not re-spec until it does.
• Mixed clustering across tables without justification.

Next

/mstack:results-audit (mandatory before drafting), then /mstack:robustness, /mstack:viz, /mstack:draft-section results.