Skills Data Science Criminology Research Design and Causal Inference

Criminology Research Design and Causal Inference

v20260724
crim-research-design
This skill guides authors through defending complex research designs in Criminology. It covers advanced quantitative methods (e.g., DiD, IV, RDD, fixed effects), longitudinal analyses, and qualitative case studies. It teaches how to establish causal identification, address selection bias, and argue against rival theories, ensuring methodological rigor for high-impact academic submissions.
Get Skill
340 downloads
Overview

Research Design (crim-research-design)

Criminology accepts many methodologies but is demanding about each. The design must credibly connect the mechanism (crim-theory-building) to crime evidence. This skill is mode-aware: pick the section that matches your work and defend it against the strongest rival explanation.

When to trigger

  • Specifying identification, a longitudinal design, case selection, or an experiment
  • A reviewer questioned causal claims, selection, the dark figure, or a confound
  • Choosing between a trajectory model, fixed-effects panel, or survival design
  • Justifying why your design adjudicates the rival theory from crim-literature-positioning

Quantitative / causal inference

  • Identification first. State the estimand and the assumptions that license a causal reading (ignorability, parallel trends, exclusion, continuity). Defend them; don't assert them.
  • Designs: experiments (incl. randomized policing/hot-spot trials), DID/event study (use modern staggered-adoption estimators, not naive TWFE), IV (first-stage strength, exclusion), RDD (density/manipulation tests, bandwidth robustness), matching/weighting with balance + sensitivity.
  • Inference: cluster at the level of treatment assignment (often place or agency); randomization inference for experiments; few-cluster corrections (wild-cluster bootstrap).
  • Crime-data validity: state which construct you measure — reported crime (UCR/NIBRS), victimization (NCVS), or self-report — and how the dark figure and reporting/recording bias affect inference.

Longitudinal / life-course / criminal careers

  • Within- vs. between-person: use fixed effects or hybrid models to isolate within-individual change when the theory is about turning points or desistance.
  • Trajectory / group-based models (GBTM, growth mixture): justify the number of groups (BIC, AvePP ≥ 0.7, group shares, classification odds); treat groups as a summary, not literal types.
  • Survival / recidivism: handle right-censoring and competing risks; distinguish timing from prevalence.
  • Criminal-career parameters: separate onset, frequency (λ), seriousness, and desistance; do not let prevalence masquerade as incidence.

Place-based & experimental

  • Randomized field trials (patrol, deterrence, reentry): report power/MDE, attrition, fidelity, ethics/IRB.
  • Spatial designs: address displacement vs. diffusion of benefits; near-repeat and hot-spot logic.

Qualitative / case-based

  • Case selection justified by design logic (typical, deviant, most/least-likely, paired comparison), not convenience. Say what the case is a case of.
  • Process tracing / life-history with explicit tests; state what evidence would have disconfirmed the argument. Plan source documentation (see crim-data-and-transparency).

The adjudication test (Criminology-specific)

For the single strongest rival theory, write one sentence: "If the rival mechanism were operating instead of mine, the crime data would look like ___; instead they look like ___." If you cannot, the design does not yet identify the contribution.

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. Criminology is observational — place/person panels where selection is pervasive; foreground DiD/IV/RDD and the selection objection.

  • detect_designrecommend → fit with as_handle=trueaudit_result.
  • Observational causal claims: staggered DiD (callaway_santanna / sun_abraham + bacon_decomposition + honest_did_from_result); IV (effective_f_test + anderson_rubin_ci); RDD (rdrobust + mccrary_test).
  • Experiments: randomization-based inference, romano_wolf for many-outcome family-wise control, and mediate for mediation (not naive controlling-away).
  • Sensitivity: oster_delta / sensemakr for observational claims.

Report the effect size in interpretable units; route the full battery to the appendix/supplement. A run end-to-end (synthetic data, real returns) is in the JF execution walkthrough.

Anti-patterns

  • Naive TWFE on staggered policy adoption; clustering below the assignment level
  • "Causal" language on a design that only supports association
  • Reading trajectory groups as real, fixed offender types
  • Ignoring the dark figure / reporting bias when using official counts
  • Convenience case selection dressed up as theory-driven

Identification expectations by design (Criminology calibration table)

A defensible Criminology design names the threat reviewers are trained to raise and the move that neutralizes it. Selection into offending and into treatment is the recurring worry.

Design Identifying assumption Threat a referee names Defensive move
Hot-spot / policing RCT randomization, no spillover displacement contaminates controls measure diffusion vs. displacement
Staggered deterrence-policy DID parallel trends across adopters bad-comparison TWFE staggered estimator + pre-trends
Life-course turning point within-person change isolates effect selection into marriage/work fixed-effects/hybrid + sensitivity
RDD at a sentencing threshold continuity at the cutoff manipulation at the line McCrary density + bandwidth robustness

Worked micro-example: a deterrence-policy quasi-experiment (illustrative)

A state raises a sentencing penalty in some counties before others. A naive TWFE gives a 9% drop (illustrative); a referee flags invalid comparisons among staggered adopters. Refit with a heterogeneity-robust staggered estimator: flat pre-trends and a credibly identified 4% first-year drop. Cluster at the county (assignment) level; with 14 treated counties add a wild-cluster bootstrap, and note a NIBRS transition could inflate pre-period UCR counts.

Design-stage referee pushback (with the Criminology fix)

  • "Selection into treatment/offending." Fix: isolate within-person change or use a quasi-experiment with a stated continuity/parallel-trends defense.
  • "Association, not causation." Fix: write the estimand and the licensing assumption; soften prose if the design only supports correlation.
  • "Official-records bias unaddressed." Fix: name reported vs. victimization vs. self-report and the dark-figure bias.
  • "Clustering below assignment." Fix: cluster at place/agency; few-cluster corrections when units are sparse.

Output format

【Mode】quant-causal / longitudinal-life-course / place-experiment / qualitative
【Estimand or claim】what is being identified/shown (and within- vs. between-person)
【Crime measure】reported / victimization / self-report + dark-figure note
【Key assumption(s)】and how each is defended
【Rival ruled out】the adjudication sentence
【Robustness/sensitivity】planned checks
【Next】crim-data-analysis

Supplementary resources

Info
Category Data Science
Name crim-research-design
Version v20260724
Size 7.74KB
Updated At 2026-07-28
Language