| name | social-science |
| description | Use when a claim is being made about people from survey, administrative, behavioural or interview data: sampling and what it licenses, instrument and scale design, measurement validity and reliability, experiments versus quasi-experiments, causal inference without randomisation (DAGs, DiD, IV, RDD, matching), panel and clustered data, preregistration, and the consent and data-protection duties that come with human participants.
|
| version | 1.0.0 |
| author | Remedy |
| tags | ["research","social-science","survey","psychometrics","causal-inference","ethics"] |
| requires | [] |
| tools | ["power_analysis","stats_assumptions","stats_effect_size","stats_multiplicity","data_profile","analysis_run","analysis_ledger","manuscript_check","lit_search","cite_add","skill_activate","file_read","file_write"] |
| triggers | ["\\b(survey (?:instrument|design|weights?)|Likert (?:scale|items?)|questionnaire (?:validity|items?)|sampling frame)\\b","\\b(difference[- ]in[- ]differences|instrumental variables?|regression discontinuity|propensity scores?|fixed effects? model)\\b","\\b(construct validity|internal validity|external validity|Cronbach.?s alpha|inter[- ]rater reliability|Cohen.?s kappa)\\b","\\b(qualitative coding|thematic analysis|grounded theory|ethnograph\\w+|focus groups?|semi-?structured interviews?)\\b"] |
Social science method
Run skill_activate(skill="research-method") first and work from that spine โ
question framing, evidence standards, preregistration, citation honesty, how to
say "we do not know". Do not restate it. This pack covers only what changes when
the units of analysis are people: they can guess the hypothesis, refuse, drop
out, be measured badly, and cannot usually be randomised.
Decision tree
- Name the claim type before touching data. Descriptive ("how many, how
often"), measurement ("does this instrument capture the construct"), or
causal ("does X change Y"). Most arguments about a social-science result are
a descriptive design carrying a causal sentence. Write the sentence the
design can support, and keep the manuscript inside it.
- Write the sampling story before the n. Target population, sampling frame,
selection mechanism, response rate, weights. A convenience sample licenses
"in this sample"; anything wider is an argument you have to make, not a
default.
references/sampling-and-generalisation.md.
- If you are measuring a construct, treat the instrument as the experiment.
Prefer a published validated instrument over new items, and say which
version and which population it was validated in. New items need pilot,
cognitive interviews and a reliability estimate.
references/survey-and-instrument-design.md,
references/measurement-validity.md.
- Randomised? If yes, the design carries the causal claim; check balance,
attrition and compliance, and analyse as randomised (ITT). If no, pick an
identification strategy on purpose and state its untestable assumption in
the abstract, not a footnote.
references/causal-inference-toolkit.md.
- Draw the DAG before choosing controls. Adjust for confounders; never
condition on a collider or a mediator you also want the total effect of.
"Control for everything available" is a bug, not caution.
- Check the data structure. Repeated measures, students in classrooms,
respondents in countries, the same person over time โ clustering changes the
standard errors and often the effective n.
references/panel-and-clustered-data.md.
- Preregister, then keep the diff. Deviations are allowed and normal; hiding
them is not.
references/preregistration-and-replication.md.
- Ethics is a procedure step, not a footer. Consent, minimal data,
de-identification, storage and retention are decided before collection.
references/human-subjects-ethics.md.
Read references/INDEX.md and pull what you need with .
