| name | analyze-quantitative-research |
| category | school |
| description | Analyze quantitative research with reproducible data checks, estimands, assumptions, uncertainty, sensitivity, and bounded interpretation. Use when testing numerical evidence against a research question. |
analyze-quantitative-research
Define the quantity being estimated before selecting a statistical procedure.
When to use
- Use for descriptive, comparative, causal, predictive, longitudinal, experimental, or observational analysis.
- Do not select tests after seeing results solely to obtain a preferred conclusion.
Procedure
- Restate the question, population, unit, design, variables, estimand, and preregistered decisions.
- Preserve raw data and create a reproducible analysis dataset with a transformation log.
- Check identifiers, duplicates, ranges, units, dates, missingness, exclusions, attrition, and design balance.
- Describe the sample and outcome distributions before modeling.
- Choose analysis based on design and estimand, then state model assumptions and dependence structure.
- Estimate effects with uncertainty intervals and meaningful units, not p-values alone.
- Examine missing-data mechanisms, influential observations, multiplicity, subgroup logic, and model diagnostics.
- Run justified robustness and sensitivity analyses without hiding unfavorable results.
- Distinguish association, prediction, and causation and compare effect magnitude with practical significance.
- Reproduce outputs from a clean environment and reconcile analysis counts to the study flow.
Worked example
A study finds a statistically significant two-point difference on a 100-point scale. The analysis reports the interval, attrition imbalance, clustered design, and sensitivity to missing outcomes. It concludes that the data support a small difference, not a large practical benefit.
Done
- A reproducible analysis package contains validated data, code, estimand, assumptions, diagnostics, uncertainty, and sensitivity results
- Counts, tables, and figures are verified from a clean run and every conclusion is bounded by design and missing-data evidence
