msom-data-analysis

Analysis & Replicability (msom-data-analysis)

When to trigger

• Proofs are drafted and you need a numerical study that earns its keep
• Estimates exist and you must defend identification and robustness
• A reviewer says "the result is an artifact of the assumptions/specification"
• You are preparing the data and code disclosure M&SOM requires

Analytical work: proofs plus a disciplined numerical study

M&SOM's dominant tradition is analytical/stochastic modeling, so rigor is judged first on the proofs and then on a numerical study that does real work:

• Proofs: state results as propositions/theorems; full proofs go in the online supplement (≤ 16 pages for a new submission). Verify monotonicity/convexity/threshold or base-stock structure rigorously.
• Numerical study: use realistic parameter ranges; quantify the magnitude of the structural effect, the value of the proposed policy versus benchmarks/heuristics, and the cost of relaxed assumptions. Report instance generation, seeds, and solver/settings so results reproduce.
• Sensitivity: show which assumptions bind and how conclusions move as primitives (demand variability, lead time, cost ratios) change.

Empirical work: identified effects, not correlations

• Defend the identification designed in msom-methods: parallel-trends/event-study evidence for DiD, first-stage strength and exclusion for IV, density/covariate-balance for RDD, model fit and identification for structural estimation.
• Treat operational decisions (capacity, inventory, staffing, routing) as endogenous; address selection and simultaneity explicitly.
• Robustness: alternative specifications, samples, and operational measures; cluster standard errors at the operational unit; report effect magnitudes in operational terms (units, hours, dollars), not just significance.

Replicability policy (M&SOM / INFORMS)

Manuscripts must contain enough detail and references to permit replication. You may be asked to provide raw data for editorial review, must be prepared to share it, and must retain it for a reasonable time after publication. For licensed data (Census, Compustat, CRSP, FactSet, WRDS) provide your own code plus detailed access/linking instructions so others can replicate without your redistributing the data. Any reuse of shared data/code beyond replicability verification must cite the paper and acknowledge the source.

Checklist

• Analytical: results stated as propositions; full proofs in the ≤16-page supplement
• Numerical study with realistic ranges; magnitudes vs. benchmarks; seeds/settings reported
• Empirical: identification defended (trends/first-stage/balance); endogeneity addressed
• Robustness across specifications/samples/measures; SEs clustered at the operational unit
• Effects reported in operational magnitudes, not p-values alone
• Data/code disclosure prepared; licensed-data access instructions included; retention noted

Anti-patterns

• "Numerical results show…" with no benchmark comparison or assumption stress-test.
• Proofs in the main text crowding out the operational insight (move them to the supplement).
• Treating capacity/inventory/staffing as exogenous in an empirical design.
• Reporting significance with no operational magnitude.
• Promising replication while withholding code/access instructions for licensed data.

Worked micro-example (illustrative)

Referees grade whether the analysis converts a model or dataset into a defensible operational claim: the analytical lane wants a structural result, a benchmark-anchored magnitude, the binding assumption stress-tested, and seeds/solver settings; the empirical lane wants an identified effect in operational units with clustered SEs, not stars alone.

Vignette: a data-calibrated inventory policy for an omnichannel retailer's fulfillment center, where ship-from-store substitutes for warehouse stock. Suppose the dual-index policy, benchmarked against the firm's base-stock rule, cuts expected backorders by 22% at a 1.3% inventory increase across 480 SKU-weeks (numbers illustrative). The discipline: make the 22%/1.3% trade the headline a manager weighs, not "the policy is optimal"; show the gain shrinks to an illustrative 9% once cross-store lead-time correlation is added, naming the assumption that carries the result; and report the fitting window, seed set, and solver tolerance so a referee could regenerate the table.

Referee-pushback patterns and the venue fix

• "The result is an artifact of the assumptions." → Add the sensitivity panel that varies the binding primitive, show where the conclusion flips, and state the operational regime in which it holds.
• "Capacity/inventory/staffing is treated as exogenous." → Instrument it, exploit a natural experiment, or model the choice; an OM empirical paper that ignores operational endogeneity is rarely identified. (For Census/Compustat/CRSP/WRDS data, ship your own code plus access instructions; confirm the current data-and-code requirement against the journal's author guidelines.)

Output format

【Analytical】propositions proven; supplement proofs: yes/no
【Numerical study】ranges / benchmarks / magnitude / seeds ...
【Empirical identification】trends / first-stage / balance ...
【Robustness】specs / samples / measures; SE clustering ...
【Replicability】data+code disclosure; licensed-data access instructions ...
【Next step】msom-contribution-framing