研究证据与方法评估

references/scientific_method.md

Use these principles as appraisal prompts. They are not an algorithm that proves a claim or substitutes for design-specific methods; see review_sources.md.

Fundamental Principles

1. Empiricism

• Knowledge derives from observable, measurable evidence
• Claims must be testable through observation or experiment
• Subjective experience alone is insufficient for scientific conclusions

2. Falsifiability (Popper's Criterion)

• A hypothesis should expose itself to observations inconsistent with its predictions; statistical hypotheses require probabilistic evaluation, not one decisive contrary datum
• Unfalsifiable claims are not scientific (e.g., "invisible, undetectable forces")
• Good hypotheses make specific, testable predictions

3. Reproducibility and Replication

• Distinguish reproducing computations with the same data/code from testing findings in new independent data; report the terminology used in the field
• Describe methods, data provenance, code, and environments sufficiently for checking
• Single studies are rarely definitive; replication strengthens confidence

4. Parsimony (Occam's Razor)

• Prefer simpler explanations over complex ones when both fit the data
• Don't multiply entities unnecessarily
• Extraordinary claims require extraordinary evidence

5. Systematic Observation

• Use standardized, rigorous methods
• Control for confounding variables
• Minimize observer bias through blinding and protocols

The Scientific Process

1. Question Formation

• Identify a specific, answerable question
• Ensure the question is within the scope of scientific inquiry
• Consider whether current methods can address the question

2. Literature Review

• Survey existing knowledge
• Identify gaps and contradictions
• Build on previous work rather than reinventing

3. Hypothesis Development

• State a clear, testable prediction
• Define variables operationally
• Specify the expected relationship between variables

4. Experimental Design

• Choose appropriate methodology
• Identify independent and dependent variables
• Control confounding variables
• Select appropriate sample size and population
• Plan statistical analyses in advance

5. Data Collection

• Follow protocols consistently
• Record all observations, including unexpected results
• Maintain detailed lab notebooks or data logs
• Use validated measurement instruments

6. Analysis

• Apply appropriate statistical methods
• Test assumptions of statistical tests
• Consider effect size, not just significance
• Look for alternative explanations

7. Interpretation

• Distinguish between correlation and causation
• Acknowledge limitations
• Consider alternative interpretations
• Avoid overgeneralizing beyond the data

8. Communication

• Report methods transparently
• Include negative results
• Acknowledge conflicts of interest
• Make data and code available when possible

Critical Evaluation Criteria

When Reviewing Scientific Work, Ask:

Validity Questions:

• Does the study measure what it claims to measure?
• Are the methods appropriate for the research question?
• Were controls adequate?
• Could confounding variables explain the results?

Reliability Questions:

• Are measurements consistent?
• Would the study produce similar results if repeated?
• Are inter-rater reliability and measurement precision reported?

Generalizability Questions:

• Is the sample representative of the target population?
• Are the conditions realistic or artificial?
• Do the results apply beyond the specific context?

Statistical Questions:

• Is the sample size adequate for the analysis?
• Are the statistical tests appropriate?
• Are effect sizes reported alongside p-values?
• Were multiple comparisons corrected?

Logical Questions:

• Do the conclusions follow from the data?
• Are alternative explanations considered?
• Are causal claims supported by the study design?
• Are limitations acknowledged?

Red Flags in Scientific Claims

1. Cherry-picking data - Highlighting only supporting evidence
2. Moving goalposts - Changing predictions after seeing results
3. Ad hoc hypotheses - Adding explanations to rescue a failed prediction
4. Appeal to authority - "Expert X says" without evidence
5. Anecdotal evidence - Relying on personal stories over systematic data
6. Correlation implies causation - Confusing association with causality
7. Post hoc rationalization - Explaining results after the fact without prediction
8. Ignoring base rates - Not considering prior probability
9. Confirmation bias - Seeking only evidence that supports beliefs
10. Publication bias - Only positive results get published

Standards for Causal Inference

Bradford Hill Viewpoints (adapted)

These are considerations, not necessary-and-sufficient criteria or a points score. Weak associations can be causal; strong ones can arise from bias, and specificity is not required. Mechanistic plausibility changes with knowledge.

1. Strength - Assess association size against plausible bias explanations
2. Consistency - Repeated observations by different researchers
3. Specificity - Specific outcomes from specific causes
4. Temporality - Cause precedes effect (essential)
5. Biological gradient - Dose-response relationship
6. Plausibility - Coherent with existing knowledge
7. Coherence - Consistent with other evidence
8. Experiment - Experimental evidence supports causation
9. Analogy - Similar cause-effect relationships exist

Appraising a Causal Estimate

• Define the intervention/exposure contrast, eligible population, time zero, follow-up, outcome, and effect measure; association, prediction, and causation are distinct targets.
• State assumptions that identify that target from the data: consistency, exchangeability, positivity, appropriate handling of selection/missingness, and interference when relevant.
• Use a causal diagram to reason about adjustment; confounders, mediators, and colliders have different roles. Do not select adjustment variables just by statistical significance.
• Randomization supports exchangeability at assignment; it does not guarantee balanced realized groups or remove later selection, non-adherence, measurement, or attrition bias.
• Observational designs may support causal inference under defended assumptions; acknowledge untestable assumptions and sensitivity to plausible alternatives rather than claim proof.
• Good prediction, small residual error, or a converged model does not establish identification.

Peer Review and Scientific Consensus

Understanding Peer Review

• Filters obvious errors but isn't perfect
• Reviewers can miss problems or have biases
• Publication and peer-review status are process information, not verification of every claim
• Retraction mechanisms exist for flawed papers

Scientific Consensus

• Emerges from convergence of multiple independent lines of evidence
• Consensus can change with new evidence
• Individual studies rarely overturn consensus
• Consider the weight of evidence, not individual papers

Open Science Principles

Transparency Practices

• Preregistration of hypotheses and methods
• Open data sharing
• Open-source code
• Preprints for rapid dissemination
• Registered reports (peer review before data collection)

Why Transparency Matters

• Reduces publication bias
• Enables verification
• Makes departures and selective reporting easier to detect; it does not prevent them
• Accelerates scientific progress