experiment-design

Experiment Design

Comprehensive experiment and study design from research questions to execution plans.

Workflow

1. Gather inputs: research question, hypotheses, available resources, constraints
2. Recommend study type: match question to appropriate research design
3. Specify variables: identify all variables and their roles
4. Plan sampling: determine sample size, selection, and randomization
5. Define protocol: step-by-step data collection and analysis plan
6. Check feasibility: resource estimation and timeline
7. Output design document: structured experiment plan

Study Type Selection

Match the research question to the most appropriate design:

Question Type	Recommended Design
Causal effect of intervention	Randomized Controlled Trial (RCT)
Causal effect, no randomization possible	Quasi-experimental (DiD, regression discontinuity)
Relationship between variables	Observational (cross-sectional, longitudinal, cohort)
Prevalence or distribution	Survey / descriptive study
Algorithm performance	Computational experiment (benchmark evaluation)
System behavior under conditions	Simulation study
User experience or usability	User study (within/between subjects)
Rare phenomena or deep context	Case study / qualitative design

Variable Specification

For every experiment, explicitly define:

• Independent variables (IV): what is manipulated or compared
• Dependent variables (DV): what is measured as outcome
• Controlled variables: what is held constant across conditions
• Confounding variables: what might produce spurious relationships
• Mediating/moderating variables: what might explain or modify the effect

Sample Size & Power

• Estimate required sample size given:
  • Expected effect size (Cohen's d, odds ratio, eta-squared)
  • Significance level (alpha, default 0.05)
  • Desired power (default 0.80, recommend 0.90 for confirmatory)
  • Number of groups and measurements
• Provide rationale for effect size estimate (prior literature or pilot data)
• Flag when requested sample exceeds available resources
• Cross-reference with statistical-analysis skill for test-specific formulas

Randomization Strategy

Select and document the randomization approach:

• Simple randomization (coin flip)
• Stratified randomization (balance key covariates)
• Block randomization (equal group sizes)
• Cluster randomization (groups of participants)
• For computational: random seeds, cross-validation folds, dataset splits

Computational Experiment Design

For CS, ML, and computational science experiments:

Dataset Selection

• Recommend established benchmarks relevant to the task
• Specify train/validation/test splits with rationale
• Address class imbalance, data leakage, distribution shift
• Document preprocessing pipeline completely

Baselines

• Identify appropriate baseline methods:
  • Naive baselines (random, majority class, mean prediction)
  • Classical methods (established algorithms)
  • State-of-the-art methods (recent top performers)
  • Ablation variants (proposed method minus components)
• Justify each baseline's inclusion

Evaluation Metrics

• Primary metric (what determines success)
• Secondary metrics (additional perspectives on performance)
• Justify metric choice relative to research question
• Report confidence intervals or standard deviations, not just point estimates

Ablation Study Design

• Identify components to ablate (one at a time)
• Design factorial experiments for interaction effects if needed
• Specify which ablations are essential vs. supplementary

Hyperparameter Search

• Search strategy: grid, random, Bayesian optimization
• Search budget and early stopping criteria
• Report all hyperparameters and ranges searched
• Use separate validation set (never tune on test)

Compute Budget

• Estimate GPU/CPU hours per experiment
• Total budget for all runs (main + ablations + hyperparameter search)
• Recommend compute-efficient alternatives if budget is tight

Protocol Template

## Experiment Protocol

### Objective
[What this experiment tests]

### Hypothesis
[Specific, falsifiable prediction]

### Design
- Type: [study design]
- Groups: [experimental vs control, conditions]
- Duration: [timeline]

### Participants / Data
- Source: [where subjects/data come from]
- Size: N = [sample size] (power analysis: d=[effect], alpha=[α], power=[1-β])
- Inclusion criteria: [who/what qualifies]
- Exclusion criteria: [who/what is excluded]

### Procedure
1. [Step-by-step protocol]
2. ...

### Measurements
| Variable | Instrument | Scale | Timing |
|----------|-----------|-------|--------|
| DV1 | ... | ... | ... |

### Analysis Plan
- Primary analysis: [statistical test]
- Secondary analyses: [exploratory analyses]
- Multiple comparison correction: [method]

### Reproducibility
- [ ] Random seed documented
- [ ] Environment specification (requirements.txt / Dockerfile)
- [ ] Data preprocessing steps scripted
- [ ] Analysis code version-controlled
- [ ] Raw data preserved separately from processed data

Pilot Study Recommendations

Always recommend a pilot study when:

• The measurement instrument is new or adapted
• The effect size is unknown
• The procedure is complex or untested
• Participant burden is uncertain

Pilot study output: revised effect size estimate, protocol refinements, feasibility assessment.

Output Formats

• Experiment design document: manuscript/experiment-design.md
• Methods section skeleton: LaTeX fragment for methods.tex with placeholders
• Reproducibility checklist: standalone checklist file

Integration

• Receives refined questions and hypotheses from brainstorming skill
• Cross-references statistical-analysis skill for test selection and power analysis
• Feeds into paper-drafting skill (methods section content)
• Feeds into implementation skill (experiment scripts and configurations)
• Uses code-analysis skill to verify implementation matches design