Data Analysis (crim-data-analysis)

Criminology reviewers are methodologically sophisticated and increasingly expect that your results can be reproduced from deposited materials (see crim-data-and-transparency). Analyze as if both are true. This skill covers execution and reporting norms; design decisions live in crim-research-design.

When to trigger

• Running main and supporting analyses; building the results section
• A reviewer asked for robustness, heterogeneity, or alternative specifications
• Fitting a trajectory model, fixed-effects panel, count model, or survival model
• Making the analysis reproducible before deposit

Analysis norms Criminology expects

1. Report uncertainty honestly. Confidence/credible intervals, not just stars; the magnitude and substantive meaning (e.g., incident-rate ratios, predicted counts, change in offending), not just significance.
2. Right model for crime data. Counts are over-dispersed and zero-heavy — prefer negative binomial / zero-inflated / hurdle over OLS on raw counts; rates need exposure offsets; rare-event cautions apply.
3. Within- vs. between-person. When the theory is developmental, isolate within-individual change (fixed effects / hybrid models); do not interpret a between-person association as a life-course effect.
4. Trajectory models with discipline. Report BIC across solutions, group shares, average posterior probabilities (AvePP ≥ 0.7), and odds of correct classification; do not over-interpret the group count.
5. Survival / recidivism. Handle right-censoring and competing risks; report the relevant hazard, not just a binary "recidivated."
6. Robustness that probes, not decorates. Show specs that could break the result (alternative crime measures, samples, estimators, fixed effects) and say what you learn.
7. Right inference. Cluster at the assignment/sampling level (often place or agency); randomization inference for experiments; few-cluster corrections when clusters are sparse.

Crime-measurement specifics

• State whether the outcome is reported crime, victimization, or self-report, and how the dark figure, reporting, and recording changes (e.g., UCR→NIBRS transition) could bias trends.
• Validate scales (self-report delinquency, legitimacy, collective efficacy); report reliability.

Reproducibility while you work (not at the end)

• One master script regenerates every table and figure from the (raw or constructed) data.
• Set and report seeds for bootstrap, randomization inference, EM-based trajectory fitting, simulation.
• Pin software/package versions (renv.lock, requirements.txt, recorded ssc/net/traj installs).
• Keep table/figure numbers matched to script outputs; document restricted-data steps that others can't rerun.

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. Criminology is observational — place/person panels where selection is pervasive; foreground DiD/IV/RDD and the selection objection.

• Many outcomes / specifications: romano_wolf (step-down FWER) or benjamini_hochberg — report the adjusted threshold.
• OVB sensitivity: oster_delta / sensemakr.
• Inference: wild_cluster_bootstrap (few clusters), twoway_cluster / conley; multilevel data → cluster at the right level.
• Re-fit off one handle: audit_result(result_id) lists the missing checks and the exact suggest_function for each.
• Exhibits: etable / did_summary_to_latex from the handle — no retyped numbers.

Keep the decisive checks in the body and the exhaustive battery in the supplement. See the executed chain in the JF execution walkthrough.

Anti-patterns

• OLS on raw, over-dispersed crime counts; ignoring exposure/offsets
• Stars-only tables with no rate ratios, effect sizes, or intervals
• Treating trajectory groups as literal offender types; cherry-picking the group count
• Reading a between-person coefficient as within-individual desistance
• "Robustness" that only reruns near-identical specs; p-hacking a significant interaction

Estimator choice keyed to the crime outcome (Criminology decision table)

Criminology reviewers are quantitatively literate and will name a mismatch between model and the data-generating process for offending. Use the outcome to pick the estimator, then defend the assumption the reviewer will probe.

Worked micro-example: reading a within-person estimate (illustrative)

Suppose a hybrid model returns an incident-rate ratio of 0.62 on the within-person "employed" indicator (illustrative): entering employment maps to roughly a 38% lower offending rate for the same person, 95% CI [0.49, 0.78], net of stable traits. The between-person column, IRR 0.80, is weaker and reading it as desistance would conflate selection (people prone to desist also find work) with the within-person change the life-course claim requires. Report both; tell the reader which identifies the mechanism.

Analysis-stage referee pushback (with the Criminology fix)

• "Official-records bias is unaddressed." Fix: state whether the outcome is arrest, victimization, or self-report, and how the dark figure and UCR→NIBRS recording shifts could bias it.
• "Between-person read as within-person." Fix: report and interpret only the fixed-effects/hybrid within column for developmental claims.
• "Robustness only decorates." Fix: show a spec that could break the result; say what held.

Output format

【Main estimate】magnitude (IRR / predicted count / hazard) + interval + substantive meaning
【Crime measure】reported / victimization / self-report + dark-figure caveat
【Within vs between】isolated correctly? [Y/N/NA]
【Model fit】counts: dispersion handled? trajectory: BIC/AvePP reported? [Y/N/NA]
【Robustness】specs that could break it → what held
【Reproducible】master script + seeds + pinned versions? [Y/N]
【Next】crim-tables-figures

Supplementary resources

• ../../resources/external_tools.md — count, trajectory, survival, and spatial packages
• ../../resources/official-source-map.md — transparency expectations and crime-data sources