情境实验设计与分析规划
根据研究问题设计假设情境、因素分配与测量题项,给出样本量估算思路及效度风险。
完整提示词
<context>
Vignette experiments show respondents short descriptions of a person or situation in which chosen features are varied experimentally, then measure their judgement. They estimate the causal effect of each feature on stated judgements, and are widely used to study discrimination, professional decisions and norms. Design choices that experts weigh: between-subjects (one vignette each, cleanest but needs large samples) versus within-subjects factorial survey (several vignettes each, efficient but invites comparison and consistency effects); full factorial versus a fractional or D-efficient subset when the universe is large; avoiding implausible combinations; signalling features naturally (names to signal ethnicity or gender are only valid if pretested for the perception intended); keeping vignettes short and the text otherwise identical; ordering effects; and attention and manipulation checks. Analysis of within-subject designs needs multilevel models or clustered errors. The biggest limit is external validity: stated responses to hypothetical cases can differ from real behaviour, and social desirability is strong on sensitive features.
</context>
<task>
Design a vignette experiment for:
<question>
{{question}}
</question>
Respondents: {{respondents}}
1. Recommend vignettes or an alternative (conjoint, audit or field experiment, observational data) given the question, and between- or within-subjects with reasons.
2. Define factors and levels: three to six factors, levels per factor, which are of theoretical interest and which are controls, and how each is signalled in text. Flag signals that need pretesting.
3. Write a vignette template with slots, kept short and neutral, and one filled example. Exclude implausible combinations explicitly.
4. Plan allocation: full or fractional factorial, number of vignettes per respondent, deck construction and randomisation of decks and order.
5. Define outcomes (wording and scale), manipulation checks that do not cue the hypothesis, attention checks, and any follow-up items.
6. Plan power: the smallest effect of interest, a sample size approach (simulation for multilevel designs), and the number of respondents and vignettes needed, stating assumptions.
7. Plan analysis: models with respondent random effects or clustered errors, interactions to test pre-specified, handling of order effects, and a preregistration note.
8. List validity risks and mitigations: social desirability, demand effects, realism, generalisation to behaviour.
</task>
<constraints>
- Do not claim the design measures real behaviour; state the hypothetical nature in limits.
- Signals for protected characteristics must be pretested; do not assume a name or cue conveys a group membership.
- State all power assumptions; do not present a sample size without them.
- Keep the number of factors manageable and say if the user's list is too long for the sample.
</constraints>
<output_format>
## Design choice
Recommendation and reasons.
## Factors and levels
A table: factor | levels | role | how signalled | pretest needed.
## Vignette template
Template with slots and one example.
## Allocation design
Bullets.
## Measures and checks
Outcome items and checks.
## Power and analysis
Assumptions, sample size approach and models.
## Validity risks
A table: risk | mitigation.
</output_format>可替换内容
- {{question}}
- {{respondents}}
分类与标签
学习教育、情境实验、问卷研究、实验设计、统计分析