Aerj Research Design

Research Design (aerj-research-design)

AERJ accepts many methodologies but is demanding about each. The design must credibly connect the framework (aerj-theory-and-framework) to evidence and meet the relevant AERA reporting standards. This skill is mode-aware: name the dominant education-research lens and defend it against the strongest alternative explanation.

When to trigger

• Specifying sampling, measurement, identification, case selection, or an integration plan
• A reviewer questioned causal claims, generalizability, trustworthiness, or measurement validity
• Preparing a pre-analysis plan / preregistration for a prospective design
• Justifying how the design addresses the rival account from aerj-literature-positioning

Quantitative (the field's common designs)

• Nesting is the default. Students in classrooms in schools — use multilevel/HLM models; specify levels, random effects, and cluster-correct inference. Report the design effect / ICC.
• Measurement. Tie constructs to validated instruments; report reliability and, where relevant, IRT/factor evidence. Validity is a design issue, not an afterthought.
• Causal claims need a credible design: RCT (with power/MDE, attrition, fidelity), or quasi-experimental (DID/event study with modern estimators, RD, IV, matching) — defend identifying assumptions, don't assert them. Map to What Works Clearinghouse-style expectations when claiming effects.
• Large-scale assessment data require plausible values and replicate/survey weights.

Qualitative (judged on its own terms)

• Case/site/participant selection justified by design logic (typical, extreme, theoretical sampling), not convenience. Say what the case is a case of.
• Trustworthiness: prolonged engagement, triangulation, member checks, negative-case analysis, audit trail, researcher positionality/reflexivity.
• Data and analysis: how data were generated, how coding/interpretation proceeded, how themes were warranted by evidence (hand off to aerj-data-analysis).

Mixed methods

• State the design type (convergent, explanatory-sequential, exploratory-sequential, embedded) and the rationale for mixing — what integration buys you that one strand cannot.
• Plan the point and method of integration (e.g., joint displays); avoid two papers stapled together.

The adjudication test (AERJ-specific)

For the single strongest rival explanation, write one sentence: "If the rival were true rather than my account, the evidence would look like ___; instead it looks 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. AERJ is empirical education research — field experiments and observational school data; multilevel inference and many-outcome corrections are central.

• detect_design → recommend → fit with as_handle=true → audit_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

• Ignoring nesting (OLS on clustered data); clustering at the wrong level
• "Causal"/"effect" language on a descriptive or associational design
• Convenience sampling dressed up as theoretical sampling
• Mixed methods that never actually integrate
• Treating measurement validity or trustworthiness as boilerplate

Design-credibility matrix (what each tradition must defend)

AERJ judges each methodology on its own terms, so the credibility bar differs by mode. Use this matrix to locate the assumption a referee will press hardest.

Worked design vignette (illustrative)

An AERJ team evaluates a peer-tutoring program with a regression-discontinuity design on an eligibility test score. The credibility case states the estimand (effect at the cutoff), shows a density test with no manipulation, reports a bandwidth-robust estimate of an illustrative 0.21 SD on the outcome, and writes the adjudication sentence: if selection rather than the program drove the jump, covariates would also jump at the cutoff; instead they are smooth. That single sentence rules out the strongest rival. A weak version would assert "the program caused gains" with no continuity evidence — exactly the move a methodological referee rejects.

Referee pushback and the venue fix

• "Causal language on an associational design." → Either build the identification or downgrade the claim to description with a mechanism hypothesis.
• "Your sampling is convenience, not theoretical." → Justify case/site selection by design logic and say what the case is a case of.
• "The mixed design is two papers stapled together." → Specify the integration point and method; confirm method-specific expectations against the journal's current submission guidelines.

Output format

【Mode】quant / qualitative / mixed
【Estimand or claim】what is being identified/shown/understood
【Key assumption(s) / trustworthiness】and how each is defended
【Rival ruled out】the adjudication sentence
【Standards】which AERA reporting standard the design meets
【Next】aerj-data-analysis

Supplementary resources

• ../../resources/external_tools.md — multilevel/IRT/causal packages and CAQDAS for qualitative work
• ../../resources/official-source-map.md — AERA reporting standards + preregistration notes