科研 Skill

复杂抽样调查分析

alterlab-survey-analysis

Analyzes complex-sample survey data with design-based inference: weights, strata, PSUs, and calibration via Python svy or R survey/srvyr. Use for weighted surveys like GSS, ANES, or DHS.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Survey Analysis — Declare the Design Before You Estimate Anything

Skill type: ANALYSIS MODULE. Complex-sample surveys (GSS, ANES, ESS, DHS, Eurobarometer) are drawn with stratification, clustering, and unequal probabilities. Analyzing them as if they were a simple random sample underestimates standard errors and yields falsely narrow CIs and wrong tests. The discipline is design-based inference: a declared design object comes first, every estimate flows through it.

Core Mission

YOU MUST WEIGHT (AND DECLARE STRATA + PSUs) BEFORE QUOTING ANY NUMBER FROM A COMPLEX SURVEY.

When to Use This Skill

  • "Give me the weighted % who [X] from ANES/GSS/DHS, with correct standard errors."
  • "Why are my survey confidence intervals so narrow?" (← design ignored)
  • "Post-stratify / rake my sample to census margins."
  • "Fit a logistic regression on this weighted, clustered survey."

Does NOT Trigger

The request is really about…Route toWhy not this skill
Writing questionnaire items / choosing a sampling framealterlab-survey-designInstrument & sampling design, not weighted analysis.
Whether the sample size / frame is adequatealterlab-ssci-sampling-gateSampling-adequacy gate, upstream.
Causal identification (DiD/IV/RDD)alterlab-causal-inferenceDesign-based survey SEs ≠ causal identification.
Plain unweighted descriptive/inferential statsalterlab-statistical-analysisNo survey design to honor.

Hand-off rule. When the request matches a row above, open your answer by naming that skill by its exact id (for example alterlab-survey-design) and saying in one line why it fits better, then stop applying this skill's procedure. Name it even when you also need to ask for missing input. If only part of the request is out of scope, name the skill for that part and do only the in-scope part. A brief general pointer is fine; the named skill does the work.

Running the code. In Claude Code, run analysis code in a throwaway environment instead of installing packages: uv run --python ">=3.11" --with "svy>=0.29" --with polars python survey.py. This needs no install approval and leaves the user's environment untouched. If uv or the network is unavailable, say so and hand over the script; never report results you did not compute.

The design-object-first rule

Declare weights + strata + PSU/cluster + FPC before any estimate:

  • weights — the inverse-inclusion-probability weight; scales the sample to the population.
  • strata — variances are computed within each stratum and pooled. Dropping strata leaves point estimates unchanged but inflates SEs (you lose the variance reduction).
  • PSU / cluster — the unit of randomization. If whole districts were sampled, the district is the PSU; lower units are not independent. Declaring the PSU is what corrects the SE upward for the clustering.
  • FPC — finite-population correction when the sampling fraction is non-trivial.

Domain (subpopulation) estimation: subset the design object, never filter the data frame first — filtering discards the strata/PSU structure needed for correct domain SEs.

Weight-type discipline

Distinguish design weights (selection probability only) from post-stratification / calibration weights (also correct for sampling error and non-response). One or the other must always be used; report which. "Weight before quoting any percentage."

Variance estimation — support both families

  • Taylor linearization — the default analytic method.
  • Replicate weights — BRR, Jackknife (JKn), Bootstrap. Use these when the data provider ships replicate weights (many public files do); do not re-derive a design they already replicated.

Verified calls (pinned)

Python — svy (maintained successor to samplics; verified on svy 0.29):

import polars as pl, svy
design = svy.Design(stratum="strata", psu="psu", wgt="wt")   # FPC: pop_size=svy.PopSize(psu="N_psu")
sample = svy.Sample(data=pl.from_pandas(df), design=design)  # svy takes a polars DataFrame
sample.estimation.mean("trust", deff="wor")                  # Taylor linearization by default
sample.estimation.mean("trust", by="region")                 # domain estimation, design kept intact
sample.estimation.prop("trust01")                            # also .total() .ratio() .median()
sample.glm.fit("trust01", x=["age", svy.Cat("educ")], family="binomial")

Provider-shipped replicate weights: svy.Design(wgt="wt", rep_wgts=svy.JackknifeWgts(prefix="wtrep", n_reps=80)) (also BrrWgts, BootstrapWgts, SdrWgts), then method="replication". svy is still beta — pin the version and re-check signatures after upgrades. samplics (TaylorEstimator) still runs but is archived and warns on import; keep it for legacy code only.

R — survey / srvyr (field standard, fully verified):

library(survey)
des <- svydesign(ids = ~psu, strata = ~strata, weights = ~wt, fpc = ~fpc,
                 data = dat, nest = TRUE)
svymean(~trust, des, deff = TRUE)
svyby(~trust, ~region, des, svymean)                 # domain estimation (keeps structure)
svyglm(trust01 ~ age + educ, design = des, family = quasibinomial())
# replicate weights when provided:
rep <- svrepdesign(weights = ~wt, repweights = "wtrep[0-9]+", type = "JKn", data = dat)
# calibration:
des2 <- rake(des, sample.margins = list(~agecat, ~sex),
             population.margins = list(pop.agecat, pop.sex))

Full Taylor-vs-replicate math, calibration (post-stratification / raking / GREG), and the Python caveats vs the canonical R recipes: references/design_and_variance.md, references/python_vs_r.md.

Reporting checklist (put in every survey result)

DESIGN:     weights (type: design | post-strat/calibrated) + strata + PSU + FPC declared
VARIANCE:   Taylor linearization | replicate weights (BRR/JKn/Bootstrap)
N:          unweighted N  vs  weighted population estimate
DEFF:       design effect per key estimate (how much the design inflates variance)
DOMAINS:    subset of the DESIGN object, not a filtered data frame

References

  • references/design_and_variance.md — Taylor vs replicate variance, calibration math, DEFF, domain estimation.
  • references/python_vs_r.md — svy (and legacy samplics) caveats and the canonical R survey/srvyr recipes.

Part of the AlterLab Academic Skills suite.

示例提示词

我的调查数据带权重和分层,想做正确的设计权重估计和方差校正。

所属插件

社会学、政治学与城市研究面向社会科学研究:问卷与访谈设计、质性编码、文本即数据、社会网络、ABM、比较政治与选举数据、RCT 与 DID、城市与区域分析。

安装该插件会一并获得同领域的 58 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

复杂抽样调查分析能做什么?

Analyzes complex-sample survey data with design-based inference: weights, strata, PSUs, and calibration via Python svy or R survey/srvyr. Use for weighted surveys like GSS, ANES, or DHS。

如何安装复杂抽样调查分析?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「复杂抽样调查分析」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

复杂抽样调查分析需要哪些权限?

该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。

复杂抽样调查分析经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。