技能 数据科学 调查设计与误差测量

调查设计与误差测量

v20260724
poq-survey-design-and-measurement
本指南旨在帮助用户全面审计调查研究的方法论,重点聚焦于“总调查误差”(TSE)框架。它提供了详细的步骤,指导用户严谨地论证调查数据的每一个方法学要素,包括覆盖面、抽样、非响应、权重和测量等,以达到顶级学术期刊(如POQ)的最高标准。适用于提升研究的科学性和可信度。
获取技能
307 次下载
概览

Survey Design & Measurement (poq-survey-design-and-measurement)

This is the POQ core. Reviewers are survey scientists who will probe every link in the Total Survey Error (TSE) chain. The design must credibly connect the hypotheses (poq-theory-and-hypotheses) to data that measure what you claim, and you must be able to disclose every methodological element to AAPOR standards. Defend each error source against the strongest alternative.

When to trigger

  • Specifying the sampling frame, questionnaire, mode, or weighting scheme
  • A reviewer questioned coverage, nonresponse, question wording, mode, or representativeness
  • Building the Appendix A: Disclosure Elements (do this here, not at the end)
  • Designing a survey experiment (wording/order split-ballot, conjoint, list experiment)

Walk the Total Survey Error chain

  • Coverage. Frame vs. target population; who is missed (online/RDD/ABS undercoverage). State the target population precisely.
  • Sampling. Probability vs. nonprobability; stratification, clustering, PSUs; selection probabilities. Nonprobability/online panels need an explicit representativeness argument.
  • Nonresponse. Report unit and item nonresponse; the response rate computed per AAPOR Standard Definitions (say RR1–RR6 and show the disposition-code calculation); assess nonresponse bias, not just the rate.
  • Measurement. Question wording, order, response scales, reference periods, social desirability, acquiescence. Validate constructs; pretest (cognitive interviews, behavior coding).
  • Mode. Single vs. mixed mode; quantify and adjust for mode effects; do not conflate a mode artifact with a substantive change.
  • Adjustment / weighting. Design weights, nonresponse adjustment, calibration/raking, post-stratification (incl. MRP). Report what the weights correct for and the resulting design effect.

AAPOR disclosure (build Appendix A now)

POQ requires you to disclose — for all data reported — or link to public documentation: funding; exact question wording; population under study; sample design; method and dates of collection; response rate and how it was calculated (AAPOR definitions); sample sizes and precision of findings; and any design effect due to clustering and weighting. Assemble these in "Appendix A: Disclosure Elements" as you design — see poq-transparency-and-data-policy.

Survey experiments

  • Split-ballot wording/order experiments: randomize, report the manipulation, isolate the measurement effect.
  • Conjoint / list experiments / vignettes: pre-specify estimands; address attention checks and attrition.
  • Preregister design and primary analyses; report power/MDE.

The artifact-vs-effect test (POQ-specific)

For the headline result, write one sentence: "If this were a survey artifact (coverage / nonresponse / wording / order / mode / weighting) rather than a real opinion signal, the data would look like ___; instead they look like ___." If you cannot, the design does not yet isolate the contribution.

Design audit table

Build a one-page audit before submission:

TSE component Design choice Residual risk Evidence or disclosure
Coverage Frame and eligibility rule Who is systematically absent? Benchmark comparison or limitation
Sampling Selection probabilities / panel recruitment Selection into the sample Weighting, calibration, or sensitivity
Nonresponse Contact protocol and disposition codes Nonresponse bias AAPOR RR calculation plus bias check
Measurement Wording, order, scale, translation Construct mismatch or satisficing Pretest/cognitive evidence and exact wording
Mode Web/phone/mail/mixed mode Mode-specific response pattern Mode controls, split test, or caveat
Weighting Design, nonresponse, calibration weights Inflated variance / model dependence Design effect and weighted/unweighted comparison

The final article should not merely say these issues were considered; it should point readers to the appendix row, table, or supplement where each was handled.

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. Public Opinion Quarterly is survey methodology and public opinion; the chain serves causal/experimental claims, while survey-design and measurement contributions use their own standards (sampling, weighting, measurement error).

  • 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

  • Reporting a response rate with no AAPOR definition or calculation shown
  • Treating a nonprobability online sample as representative with no argument or benchmark
  • Ignoring mode effects in a mixed-mode design; conflating mode artifacts with opinion change
  • Weighting the point estimate but ignoring weights/clusters in the variance (see poq-data-analysis)
  • Leaving Appendix A disclosure until submission
  • Naming the target population broadly while the frame covers only reachable or panelized respondents

Output format

【Target population & frame】coverage gaps named
【Sample design】probability/nonprobability; strata/clusters/PSUs
【Nonresponse】RR definition + value + bias assessment
【Measurement】wording/order/scale + validation + pretest
【Mode】single/mixed; mode effect handled?
【Weighting】what it corrects + design effect
【Artifact ruled out】the artifact-vs-effect sentence
【Design audit】TSE table complete; residual risks disclosed
【Appendix A started?】[Y/N]
【Next】poq-data-analysis

Supplementary resources

信息
Category 数据科学
Name poq-survey-design-and-measurement
版本 v20260724
大小 7.06KB
更新时间 2026-07-29
语言