| name | demog-research-design |
| description | Use when defending the research design of a Demography (PAA / Duke University Press) manuscript — choosing among demographic methods (life tables, decomposition, event-history/survival, age-period-cohort, multistate, microsimulation, projections) and, where the question is causal, defending identification. Demography judges each method on its own terms. Strengthens the design; it does not write code. |
Research Design (demog-research-design)
Demography accepts a wide variety of methodological approaches but is demanding about each. The design
must credibly connect the argument (demog-theory-building) to the demographic evidence. This skill is
method-aware: pick the section that matches your question and defend it against the strongest rival
explanation.
When to trigger
- Choosing the demographic method that actually answers the question
- A reviewer questioned the rate construction, the identification, or the projection assumptions
- Specifying an age-period-cohort, multistate, or microsimulation design
- Justifying why your design adjudicates the rival account from
demog-literature-positioning
Match the method to the question
- Life tables — for survival, life expectancy, and exposure: period vs. cohort, abridged vs.
complete; multiple-decrement (cause-specific) and multistate (healthy/disabled) where relevant.
- Decomposition — to attribute a difference or change in a rate to components: Kitagawa
(rate vs. composition), Arriaga (age contributions to e0), Horiuchi continuous, Das Gupta
(multi-factor). Say exactly what each component means.
- Event-history / survival — for timing and transitions: Cox, parametric, discrete-time, with
competing risks and multistate models when several destinations matter; check the
proportional-hazards assumption.
- Age-period-cohort — confront the identification problem head-on: APC effects are linearly
dependent, so state the constraint or modeling assumption (and its substantive justification) you
rely on; do not present a single "identified" APC partition as if it were assumption-free.
- Multistate / projections / microsimulation — make transition rates, the base population, and the
assumptions (closed/open, period/cohort) explicit; report sensitivity to key assumptions.
When the question is causal
- Identification first. State the estimand and the assumptions licensing a causal reading
(ignorability, parallel trends, exclusion, continuity); defend them, don't assert them.
- Selection and exposure are demographic hazards: mortality selection, migration selection, and
differential exposure can masquerade as effects — address them explicitly.
- Inference. Cluster at the right level (e.g., household, region, cohort); use survey weights and
design for complex samples; report uncertainty for derived demographic quantities.
- Sensitivity. How strong must an unobserved confounder (or a violated rate assumption) be to
overturn the result?
The adjudication test (Demography-specific)
For the single strongest rival explanation (e.g., compositional change, selection, tempo
distortion), write one sentence: "If the rival were true rather than my account, the age/cohort
pattern 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. Demography is formal + empirical demography; the causal chain serves its reduced-form lane, while formal demographic modeling uses its own tools — decomposition (oaxaca / gelbach) is often 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
- Running a regression when the question calls for a life table, a decomposition, or an event-history model
- Presenting an APC decomposition without naming the identifying constraint
- Period rates read as cohort experience (or vice versa) without justification
- Ignoring mortality/migration selection in a survival or panel design
- Projections whose assumptions are buried instead of varied and reported
Output format
【Method】life table / decomposition / event history / APC / multistate / microsim / projection / causal
【Quantity / estimand】what is being measured or identified
【Key assumption(s)】and how each is defended (name the APC constraint if used)
【Rival ruled out】the adjudication sentence
【Robustness/sensitivity】planned checks
【Next】demog-data-analysis
Supplementary resources
