Series ID: EDKS-ADV-ZH-0018 · How to Learn Advanced English (Chinese Edition) · Lesson No.018 · C1 → C2
Write Methods for Reproducibility Without Writing a Lab Diary | 把研究方法写到可检查，而不是写成流水账
A Methods section is not a memory of everything that happened. It is a controlled account of the decisions another knowledgeable reader needs in order to understand, evaluate and—where the research design allows—reproduce or replicate the work.
Methods 不是“实验当天发生了什么”的回忆录。它是经过选择的研究记录：让另一个有专业能力的读者知道你怎样设计、怎样执行、怎样处理数据、怎样分析，以及哪些决定可能改变结论。
This distinction matters because two bad Methods sections can look very different.
One is too short:
We recruited students, gave them a test and analysed the results.
The other is too long:
At 8:47 a.m. the researcher opened the spreadsheet, checked the room, printed the forms, moved three chairs and then…
Neither helps the reader reconstruct the inferential chain.
The governing question is:
Which methodological details materially affect what the Results can mean?
哪些方法细节会实质性改变 Results 的解释？
Before we begin: Lesson 015 owns section boundaries; Lesson 018 owns Methods depth | 第015课管 section architecture；第018课深入 Methods
Lesson 015 established the core separation: Methods = what was done, Results = what was found, Discussion = what the findings may mean.
Lesson 018 goes much deeper. It asks how to make the Methods section sufficiently transparent without converting it into a procedural diary.
1. Reproducibility begins with transparent reporting | 可复现性从透明报告开始
Nature’s author guidance states that Methods should be concise but contain the elements necessary to allow interpretation and replication of results. It also encourages authors to share detailed protocols separately where appropriate.
Nature | Formatting Guide: Methods
That gives us the central balance:
enough to reproduce the logic and procedure, not every historical detail of the research process.
2. NIH links rigor to design, methodology, analysis, interpretation and reporting | Rigor 不是一个单独 section
NIH describes scientific rigor as the strict application of the scientific method to support unbiased and well-controlled design, methodology, analysis, interpretation and reporting. Current NIH reproducibility guidance also emphasises transparent reporting of protocols and analyses.
NIH | Enhancing Reproducibility through Rigor and Transparency
3. Reporting guidelines are design-specific | 不同 design 需要不同 reporting floor
The EQUATOR Network maintains hundreds of reporting guidelines and highlights major families including CONSORT for randomised trials, STROBE for observational studies, PRISMA for systematic reviews, SPIRIT for protocols, STARD for diagnostic studies, COREQ/SRQR for qualitative research, ARRIVE for animal studies, SQUIRE for quality improvement and CHEERS for economic evaluations.
EQUATOR Network | Reporting Guidelines
The lesson is not “memorise every checklist.”
The lesson is:
your Methods obligations depend on the kind of evidence you are claiming to produce.
4. A Methods section has three readers | Methods 至少有三个读者
- The evaluator — Can I trust the design?
- The replicator — Could I recreate the procedure sufficiently?
- The interpreter — Which limits must I remember when reading the Results?
Write for all three.
5. “Reproduce” and “replicate” can mean different things | reproduce / replicate 在不同领域定义不同
In computational work, reproduction may mean obtaining the same result from the same data/code.
Replication may mean collecting new data or independently repeating the study.
Other fields use the terms differently.
Do not build your Methods around vocabulary debates. Build it around inspectability.
6. The Methods section should expose the inferential chain | Methods 要暴露 inference chain
Readers need to see how you moved from:
research question → cases/data → measurement → procedure → processing → analysis → result.
If one link is invisible, interpretation becomes harder.
7. Start with the design | 先告诉读者你做的是什么研究
Examples:
- parallel-group randomised controlled trial;
- quasi-experimental pre–post study;
- prospective longitudinal cohort;
- cross-sectional survey;
- case-control study;
- qualitative interview study;
- ethnographic observation;
- corpus-based analysis;
- mixed-methods explanatory sequential design;
- systematic review and meta-analysis;
- simulation study.
8. Design labels compress expectations | Design label 是高密度信息
randomised tells the reader something about allocation.
longitudinal tells the reader time matters.
cross-sectional warns against causal/time-change inference.
But the label does not replace procedural detail.
9. Never upgrade the design name | 不要给 design 升级
A before–after study with no random assignment is not a randomised trial.
A convenience sample is not a representative population survey.
A literature search is not automatically a systematic review.
10. State where and when the study occurred when it matters | Setting / time 可能是方法变量
Setting changes:
- resources;
- participant behaviour;
- implementation;
- generalisation;
- measurement conditions.
Time matters when policy, software, curriculum or environment changes.
11. Population definition belongs before sample description | 先定义 target population
Who or what is the inference about?
Then:
Who or what actually entered the study?
12. Sampling is not administrative trivia | Sampling 决定外部解释
Random sample?
Convenience sample?
Volunteer sample?
Purposive sample?
Snowball recruitment?
Complete registry?
13. State recruitment route | 招募方式会产生 selection
Students volunteered after an email invitation.
is methodologically different from:
All eligible students were included.
14. Inclusion criteria need a research reason | Inclusion criteria 不能只列清单
What defines an eligible case?
Why does that boundary match the question?
15. Exclusion criteria need transparency | Exclusion 不能等看到 data 再随意决定
State:
- pre-specified exclusions;
- post-collection exclusions;
- reasons;
- numbers affected.
16. Exclusions can change the result | 删除谁，可能改变结论
Therefore exclusion logic is part of evidence, not housekeeping.
17. Sample-size planning belongs in Methods | Sample size 不是 Results 才解释
Depending on the design, report:
- power calculation;
- precision target;
- saturation logic;
- feasibility constraint;
- census/complete sample;
- simulation-based planning.
18. “We used 50 because 50 was available” may be honest | Honest constraint 比 fake power rationale 好
Then discuss what that constraint means for precision.
19. Do not justify sample size using the observed p-value | 不要 result-based sample justification
The sample was sufficient because the result was significant.
This reverses the logic.
20. Randomisation must be describable | Randomisation 不是一句 randomly assigned 就结束
Where relevant, state:
- sequence generation;
- allocation ratio;
- stratification/blocking;
- allocation concealment;
- who generated sequence;
- who enrolled participants;
- who assigned intervention.
21. “Randomly” is a technical claim | random 不是“随便分”
Alternating participants is not random.
Assigning by birth month may be systematic, not random.
22. Allocation concealment and blinding are different | allocation concealment ≠ blinding
Allocation concealment protects assignment before entry.
Blinding/masking protects later behaviour/measurement from knowledge of assignment.
23. State who was blinded | “double blind” often hides ambiguity
Participants?
Intervention providers?
Outcome assessors?
Data analysts?
Name the roles.
24. Some studies cannot blind participants | 无法 blind 不代表研究无效
Education interventions, behavioural programmes and surgery may make certain forms of blinding impossible.
State what was possible and how bias was reduced elsewhere.
25. Define the intervention/exposure precisely | Intervention 要可识别
Include:
- content;
- dose;
- duration;
- delivery;
- provider;
- timing;
- adaptation rules;
- fidelity checks.
26. “Received structured feedback” is often too vague | structured 到底是什么？
What structure?
What prompts?
What sequence?
What did control group receive?
27. Comparator detail matters | Control / comparator 不能写成“nothing” if it received something
Business-as-usual?
Waitlist?
Active comparison?
Placebo/sham?
Standard instruction?
28. Intervention fidelity can explain variable outcomes | 是否真的按设计执行？
Possible fidelity evidence:
- session completion;
- protocol checklist;
- recording review;
- dosage received;
- provider adherence.
29. Adaptation rules should be reported | 个性化 intervention 要写“如何变”
If tutors could alter the programme, what was fixed and what was flexible?
30. Variables need conceptual and operational definitions | Concept ≠ measure
Concept:
independent revision quality.
Operation:
score on an unseen revision task rated using a six-dimension rubric.
31. Outcomes need hierarchy | Primary / secondary outcomes 要分清
Primary outcome determines the study’s main test.
Secondary outcomes can enrich interpretation.
Do not decide primary outcome after seeing which result looks best.
32. Outcome timing is part of the outcome | 时间点不是附注
Immediate post-test.
One week.
Twelve weeks.
Each answers a different question.
33. Measurement instruments need enough detail | Instrument 需要可判断
Report, where relevant:
- instrument name/version;
- construct measured;
- scale/range;
- scoring;
- validation/reliability evidence;
- language/translation;
- adaptations.
34. Modified instruments require modified reporting | 改过量表，就不能只引用原版
What changed?
Why?
How might modification affect validity?
35. Translated instruments need translation logic | 翻译工具也有 method
Translation/back-translation?
Expert review?
Cognitive testing?
Local validation?
36. Rubrics need scoring transparency | Writing rubric 不只是“用 rubric 评分”
Who scored?
Training?
Blind to condition?
Disagreement handling?
Reliability?
37. Human judgement is part of the method | 人的判断不能假装不存在
Coding, diagnosis, classification, qualitative interpretation and subjective scoring all need procedural transparency.
38. Data collection environment matters when it can change behaviour | 环境会影响 data
Online or in-person?
Supervised?
Timed?
Device controlled?
Individual or group?
39. Order effects may matter | 顺序可能改变表现
If all participants see Task A before Task B, learning/fatigue may influence comparison.
40. Counterbalancing should be described | Counterbalancing 是 design decision
41. Missing data is part of the method | Missing data 不是 Results 才突然出现
State:
- how missingness was identified;
- how much;
- how handled;
- what assumptions the method requires.
42. Complete-case analysis is a choice | 丢掉 missing cases 也是 method
It can bias results if missingness is systematic.
43. Imputation needs method detail | Imputation 不能只写 “missing values were filled”
Which method?
Which variables?
How many datasets if multiple imputation?
44. Data cleaning is not invisible | Cleaning 可能改变结果
Duplicates.
Impossible values.
Outliers.
Bot responses.
Attention checks.
45. Preprocessing belongs in the inferential chain | Preprocessing 不是 technical appendix trivia
Tokenisation.
Normalisation.
Image resizing.
Signal filtering.
Feature engineering.
Winsorisation.
46. Report preprocessing order | 顺序有时会改变 output
Especially in computational pipelines.
47. Outlier rules should not be result-driven | Outlier 不能因为“不好看”就删
Prefer pre-specified or transparently justified rules.
48. Sensitivity analysis can test analytic fragility | 不同处理方式会不会改变 conclusion？
Methods should explain planned sensitivity analyses where relevant.
49. Analysis should answer the research question | Analysis 不是 software menu
Weak:
Data were analysed using SPSS.
Better:
We estimated the between-group difference in one-week revision score using linear regression adjusted for baseline score.
50. Software is implementation, not analytical logic | R / Python / SPSS 只是工具层
Still report software/version when reproducibility requires it.
51. Statistical models need variable roles | Model 要说 predictors/outcomes/covariates
Outcome?
Exposure?
Covariates?
Interaction?
Random effects?
52. Adjustment variables need rationale | Covariates 为什么进入 model？
Pre-specified theory?
Design factor?
Confounding control?
Not simply “because they were significant”.
53. Model assumptions need checking | 假设不能隐身
Depending on model:
- linearity;
- independence;
- distributional assumptions;
- proportional hazards;
- multicollinearity;
- residual structure.
54. Report what happened when assumptions failed | 不只是“assumptions were checked”
What criterion?
What alternative model?
55. Multiple testing needs transparency | 多次检验会增加 false positive 风险
Pre-specify primary comparisons.
Report adjustment strategy where appropriate.
56. Exploratory analysis should be labelled | Exploratory ≠ confirmatory
Do not rewrite post-hoc patterns as if they were pre-planned hypotheses.
57. Preregistration/protocol registration should be reported where relevant | Preregistration 提高 temporal traceability
Registration ID.
Date.
Protocol location.
Pre-specified outcomes/analyses.
58. Deviations from protocol need transparency | 改了计划，不等于研究失败
State:
- what changed;
- why;
- when the decision was made;
- whether change followed knowledge of results.
59. Protocol deviation is more informative than pretending it never happened | 透明比假整齐更可信
60. Code and analysis scripts can be part of reproducibility | 复杂 analysis 最好有 code trace
Where ethical/legal constraints allow, share:
- analysis scripts;
- environment/package versions;
- seed/random-state handling;
- pipeline instructions.
61. Random seeds do not make stochastic research perfectly reproducible | Seed 只是一个 control
Hardware, library versions and nondeterministic operations may still matter.
62. Computational environment can matter | Version drift 会改变结果
Report materially relevant versions.
63. Do not bury code-defined methods only in code | Methods prose 仍需解释 analytical logic
Code gives executable detail.
Prose gives conceptual rationale.
64. Data availability is related but separate | Data sharing ≠ Methods section
Still, Methods should identify data provenance and relevant access constraints.
65. Data provenance matters | 数据从哪里来？
Primary collection?
Administrative records?
Public dataset?
Web-scraped?
Synthetic?
66. Dataset version/date can matter | 动态 dataset 需要 version/time
67. Web-scraped data need collection rules | 网络数据不是“网上找的”
Source URLs/domains.
Dates.
Query/filter rules.
Deduplication.
Terms/ethical considerations.
68. Secondary data inherit original measurement limits | 二手数据不会因为样本大就自动适合新问题
State how variables were originally produced.
69. Qualitative Methods have different reproducibility goals | Qualitative 不能被硬塞进实验复制模型
The goal often becomes transparency of:
- sampling;
- context;
- researcher position;
- data generation;
- analytic process;
- reflexivity;
- credibility checks.
70. COREQ and SRQR are useful design-specific floors | Qualitative reporting 也有 checklist
The EQUATOR library includes COREQ and SRQR among major qualitative reporting guidelines.
EQUATOR | Qualitative Reporting Guidelines
71. Researcher position can influence qualitative data | Reflexivity 不是 autobiography
Relevant questions:
- What was the researcher–participant relationship?
- What assumptions or roles shaped interaction?
- How were interpretations challenged?
72. Do not list identity details with no analytic relevance | Reflexivity 要与 evidence production 有关系
73. Interview Methods need more than “semi-structured interviews” | Semi-structured 是起点，不是完整方法
Report:
- who interviewed;
- where/how;
- duration;
- guide development;
- recording/transcription;
- repeat interviews if any;
- field notes;
- sampling logic.
74. Translation in qualitative research needs transparency | 多语言 interview 如何转成 analysis？
Who translated?
At what stage?
Were original-language quotes retained?
How were meaning differences handled?
75. Coding Methods should describe analytic decisions | Coding 不是“we coded the data”
Inductive?
Deductive?
Hybrid?
Codebook?
Iterative?
Multiple coders?
76. Inter-coder agreement is not required in every qualitative tradition | 不要把一个 paradigm 的标准强加给所有 qualitative design
Follow the epistemological and methodological framework actually used.
77. Member checking is not universal | Member checking 也不是质量 stamp
Explain whether and why it was appropriate.
78. Saturation language needs definition | “Reached saturation” 不能像魔法词
What kind of saturation?
How assessed?
79. Mixed-methods Methods need integration architecture | Mixed Methods 不是 Quant + Qual 拼盘
State:
- design type;
- sequence;
- priority;
- connection point;
- integration method;
- how conflicting strands were handled.
80. Explain why both strands are necessary | 为什么一个 method 不够？
Example:
Quant estimates effect.
Qual investigates why implementation differs.
81. Integration can happen at design, data, analysis or interpretation | Integration 要说在哪里发生
82. Do not average unlike evidence into one fake score | Qual theme + test score 不是同一单位
Integration means relation, not arithmetic flattening.
83. Systematic-review Methods have a distinctive chain | Systematic review Methods
Report:
- eligibility criteria;
- information sources;
- search strategy;
- selection process;
- data collection;
- risk-of-bias assessment;
- synthesis methods;
- registration/protocol.
84. PRISMA is not a writing style | PRISMA 是 reporting checklist，不是 prose template
Use it to ensure necessary information exists.
85. Search strategy is reproducible methodology | Literature search 不是“we searched databases”
Databases.
Dates.
Search strings.
Language limits.
Grey literature.
Deduplication.
86. Screening decisions need traceability | 谁筛选？怎样解决分歧？
87. Risk-of-bias assessment is part of evidence weighting | 不是附加 decoration
88. Meta-analysis needs compatibility logic | 为什么这些 studies 可以合并？
Outcome definitions.
Effect measure.
Model.
Heterogeneity.
89. Observational-study Methods need confounding logic | STROBE-type transparency
Define:
- exposure;
- outcome;
- confounders;
- selection;
- follow-up;
- missingness;
- analysis.
90. Confounders should be chosen by causal/research logic | 不要只按 univariate p-value 选 covariate
91. Longitudinal studies need follow-up accounting | 谁开始、谁留下、谁失联？
Attrition can change the observed population.
92. Time-varying exposure/outcome needs timing clarity | 时间顺序是 causal inference 的基础
93. Case-control studies need case/control selection detail | Case/control 从哪里来决定 bias
94. Diagnostic studies need reference standard | 新 test 与什么“真值”比较？
95. Measurement studies need reliability/validity plan | Instrument paper 的 Methods 是核心贡献
96. Experimental animal studies need ARRIVE-type detail | Species/strain/sex/housing/randomisation/blinding 等可能影响 inference
97. Reporting guideline selection should match design | 不要用 CONSORT 检查 qualitative interview
EQUATOR explicitly provides tools for choosing appropriate reporting guidance by study type.
EQUATOR | Selecting the Appropriate Reporting Guideline
98. Ethics belongs where it affects conduct and permission | Ethics 不是一句 boilerplate
Where relevant, report:
- ethics approval;
- committee/identifier;
- consent;
- waiver;
- privacy protections;
- risk mitigation.
99. Ethical approval does not equal methodological quality | Approval ≠ design validation
Ethics and methodology answer different questions.
100. Consent process can affect participation | Consent 方式可能影响 selection/behaviour
101. Sensitive data need access and anonymisation detail | 数据保护也可能是 method
Especially if anonymisation transforms the data.
102. De-identification can remove analytic variables | Privacy preprocessing may affect analysis
103. AI-assisted research requires explicit methodological traceability | AI 工具进入 pipeline，就进入 Methods
If AI materially affected:
- data generation;
- translation;
- classification;
- coding;
- summarisation;
- analysis;
- stimulus generation;
report enough detail to understand its role.
104. Model name alone is not enough | AI Methods 不只是写 “used ChatGPT”
Depending on the research:
- model/system/version;
- date accessed;
- prompt/procedure;
- sampling/settings;
- number of runs;
- human review;
- validation;
- error handling.
105. Do not claim determinism when system is stochastic | 同 prompt 可能不产生同 output
106. Human validation belongs in Methods | AI output 如何被核实？
107. If AI generated study stimuli, preserve the stimuli | 刺激材料最好可访问
108. If AI transformed participant data, explain privacy controls | 个人数据处理需要额外透明度
109. Materials availability can shorten Methods | 把详细 protocol/materials 放 repository
Nature encourages protocol sharing platforms for detailed procedures.
110. But external repositories should not become missing prose | “See repository” 不能取代核心 method explanation
The paper must still explain what was done and why.
111. Cite established methods; describe modifications | 已发表标准方法可以引用，但变更必须写
Nature advises avoiding unnecessary repetition of already published methods while stating additions or variations.
112. A method citation is not enough if readers cannot know which version you used | Version/variation matters
113. Methods should distinguish planned from adaptive decisions | 哪些是事前，哪些是过程中调整？
114. Adaptive designs need decision rules | 什么时候扩样本？什么时候停止？
115. Stopping rules can bias evidence if hidden | Early stopping 必须透明
116. Data-dependent decisions deserve special attention | 看过 result 以后作出的决定，最容易改变 inference
Examples:
- outlier removal;
- subgroup creation;
- outcome switching;
- covariate choice;
- model switching.
117. Separate confirmatory from exploratory | Confirmatory / exploratory 是时间顺序信息
Not moral labels.
Both can be valuable if labelled honestly.
118. Blinded analysis can reduce analytic flexibility | 某些领域可在揭示 group labels 前锁定分析
119. Registered reports separate study value from result direction | 在适用领域，Registered Report 可减少 publication/result bias
120. Reproducibility is not achieved by word count | Methods 写得长不等于透明
Two thousand words of chronology can hide the one decision that mattered.
121. Use the “decision-impact” filter | Decision-impact 过滤器
For each detail, ask:
Could changing this detail plausibly change the result, interpretation or ability to reproduce the work?
If yes, keep it.
122. Use the “reader-needs” filter | Reader needs
Does the reader need this detail to:
- evaluate bias?
- reproduce procedure?
- understand measure?
- understand analysis?
- judge generalisation?
123. Use the “lab diary” filter | 流水账过滤器
Delete details that merely answer:
What happened next in real time?
unless chronology itself affects the method.
124. Chronology matters when order matters | 什么时候 diary-like chronology 合理？
Washout period.
Training sequence.
Repeated measures.
Intervention phases.
Time-sensitive sample handling.
125. Group by methodological function | Methods 常比纯时间顺序更清楚
Possible subheadings:
- Design;
- Participants;
- Materials;
- Procedure;
- Measures;
- Data Processing;
- Analysis;
- Ethics.
126. Subheading order should match the reader’s reconstruction task | 不是固定模板
127. Start each subsection with its job | Subsection 第一行告诉读者这段在干什么
Participants were recruited from…
The primary outcome was…
We estimated…
128. Avoid procedural throat-clearing | 删掉空开场
In order to conduct this study, several methodological steps were undertaken.
Delete.
129. Use direct verbs | Methods 需要动作清楚
We randomised…
Participants completed…
Two blinded raters scored…
130. Passive voice is useful selectively | Passive 不是 academic costume
Samples were stored at −80°C.
Actor irrelevant.
Two researchers independently coded transcripts.
Actor/process relevant.
131. Tense should follow function | Methods 常用过去时，但不是死规则
Completed action:
Participants completed…
Stable instrument property:
The scale ranges from 0 to 40.
132. Avoid empty precision | 过多小数不等于 reproducible
Report precision meaningful to the process.
133. Units must be explicit | 时间、剂量、距离、浓度都需要单位
134. Device specifications matter only when they can change measurement | 型号不是自动必须
Phone brand irrelevant for a paper survey.
Sensor model may be essential for physiological measurement.
135. Proprietary tools can create reproducibility barriers | 黑箱工具要多写验证与版本
136. If a tool cannot be shared, describe inputs/outputs enough for evaluation | 可复现性也可以通过透明边界提高
137. The Methods section should reveal degrees of researcher freedom | Researcher degrees of freedom
How many outcomes?
How many models?
How many exclusion options?
How many subgroup analyses?
138. Transparency makes flexibility interpretable | 灵活性本身不是错误，隐藏灵活性才危险
139. Sensitivity analyses can show conclusion stability | 如果不同合理 choices 都得出相似结论，confidence 提高
140. Report robustness checks in Results, define them in Methods | 计划/方法在 Methods，结果在 Results
141. Methods cannot contain future-perfect fantasy | 不要写实际没做的理想方法
Report what happened.
If the protocol planned something different, state the deviation.
142. “Standard procedures were followed” is often insufficient | Standard 对谁 standard？
Name the standard/protocol/version.
143. “As appropriate” hides decision rules | 根据什么 appropriate？
State the criterion.
144. “Outliers were removed” is incomplete | 定义 outlier + rule + timing
145. “Poor-quality responses were excluded” is incomplete | 什么叫 poor quality？
146. “Data were cleaned” is incomplete | Cleaning 做了什么？
147. “The model was optimised” is incomplete | Objective / search space / validation / stopping rule
148. “Themes emerged” hides analytic labour | Themes 怎样形成？
149. “Consensus was reached” hides disagreement process | How?
150. “Validated questionnaire” needs source/version | Which validation and population?
151. Methods should be written before memory fades | 不要等投稿前才重建
Maintain:
- protocol;
- decision log;
- analysis scripts;
- instrument versions;
- data dictionary.
152. Research log is not the Methods section | Research log 可以很详细，Methods 必须选择
Private/internal record: maximal trace.
Published Methods: reader-relevant trace.
153. Supplementary material can carry deep technical detail | Supplement 是第二层，不是垃圾场
Core methodological decisions stay visible in main text.
154. The “replicator test” | Replicator test
Give the Methods to a knowledgeable reader.
Ask them to list:
- population;
- sampling;
- design;
- intervention/exposure;
- outcomes;
- timing;
- analysis;
- exclusions.
Any major reconstruction error indicates missing information.
155. The “reviewer bias” test | Reviewer bias test
Could a reviewer identify major sources of:
- selection bias;
- measurement bias;
- confounding;
- attrition;
- analytic flexibility?
156. The “result lineage” test | 每个 Result 都有 Method parent 吗？
For every main Results paragraph, point to:
- measure;
- collection procedure;
- analysis.
157. The “unused method” test | Methods 里有没有完全没产生 Result 的内容？
Maybe secondary material is unnecessary—or perhaps a result is missing.
158. The “hidden decision” test | 哪个决定最可能改变结果，却没写？
Outliers?
Missingness?
Stopping?
Subgroups?
159. The “alternative analyst” test | 另一个 analyst 拿到同 data，会知道你怎么得到这个 estimate 吗？
160. The “version” test | 软件 / instrument / dataset / protocol 哪些版本真的重要？
161. The “timing” test | 哪些时间点改变 inference？
162. The “human judgement” test | 哪些步骤靠人决定？
Make those steps inspectable.
163. The “AI involvement” test | AI 在哪一层参与？
Generation?
Coding?
Analysis?
Translation?
What validation?
164. The “reporting guideline” test | Study design 对应哪个 checklist？
165. The “diary deletion” test | 删除所有“然后我们…”句子后，有没有真正信息损失？
If not, cut them.
166. Worked case: fictional advanced bilingual feedback trial | 完整案例：虚构试验
All details below are fictional teaching material.
Question:
Does structured action-oriented feedback improve independent revision one week later among advanced bilingual learners?
167. Weak Methods version | 弱版本
We recruited 96 students and randomly divided them into two groups. One group received structured feedback and the other received normal feedback. They did a writing task, and a week later they did another task. We then analysed the data using statistical software.
168. Problems in the weak version | 缺什么？
- recruitment/eligibility;
- randomisation method;
- feedback content;
- comparator definition;
- task equivalence;
- primary outcome;
- scoring;
- blinding;
- analysis;
- missing data/exclusions.
169. Stronger Design and Participants paragraph | 示例
We conducted a parallel two-group randomised trial with 96 advanced bilingual learners enrolled in upper-level English courses. Eligible learners had completed the programme’s prerequisite proficiency assessment and had not received the structured-feedback protocol previously. Participants were recruited by course-wide invitation and enrolled before group allocation. An independent script generated a 1:1 allocation sequence stratified by baseline revision score; the researcher who enrolled participants did not have access to the sequence.
170. Why this is stronger | 它暴露了 selection + allocation
The reader can now evaluate:
- population;
- volunteer recruitment;
- eligibility;
- allocation;
- concealment.
171. Stronger Intervention paragraph | 示例
Both groups revised the same 600-word argumentative task for 35 minutes. The structured-feedback group received comments organised into three required actions—clarify claim, integrate evidence and explain reasoning—while the standard-feedback group received evaluative comments identifying strengths and weaknesses without action prompts. Feedback length was capped at 180 words in both conditions.
172. Why comparator control matters | 如果 structured group 收到更多字，effect 可能来自 dose
173. Stronger Outcome paragraph | 示例
The primary outcome was independent revision quality on an unseen argumentative task completed seven days later without feedback. Two raters, blinded to group allocation, scored claim clarity, evidence integration, reasoning and organisation on a preregistered 0–5 rubric. Raters completed calibration training before scoring; disagreements greater than one point were reviewed by a third blinded rater.
174. This paragraph defines the construct | “independent transfer” 现在有 operational meaning
175. Stronger Analysis paragraph | 示例
We estimated the between-group difference in total revision score using linear regression adjusted for baseline revision score, with group allocation as the primary predictor. The primary analysis followed assigned groups. Secondary analyses examined rubric dimensions separately and were treated as exploratory. Missing outcome data were not imputed in the primary analysis; a sensitivity analysis used multiple imputation under a missing-at-random assumption. Analyses were conducted in R, with scripts archived before outcome interpretation.
176. What this exposes | 分析现在可检查
- estimand/comparison;
- adjustment;
- primary vs exploratory;
- missing-data strategy;
- sensitivity;
- software/code trace.
177. What is still not in Methods? | 哪些不该放这里？
Do not write:
The structured group scored substantially higher, confirming our hypothesis.
That belongs to Results/Discussion.
178. Practice A: delete diary detail | 练习 A
Which detail is probably unnecessary?
The researcher entered the room at 8:55, distributed pencils, then opened the timer app.
What would make any of these details methodologically relevant?
179. Practice B: repair vague randomisation | 练习 B
Participants were randomly placed into groups.
What additional information is needed?
180. Practice C: repair vague intervention | 练习 C
Students received structured feedback.
Write the minimum detail needed to distinguish it from control.
181. Practice D: repair vague analysis | 练习 D
Data were analysed in Python.
Rewrite as analytical logic.
182. Practice E: exclusion transparency | 练习 E
Seven participants were removed for “poor-quality data”.
What must Methods state?
183. Practice F: missing data | 练习 F
12% of delayed scores are missing.
List three Methods questions.
184. Practice G: qualitative coding | 练习 G
Repair:
Themes emerged from the interviews.
185. Practice H: AI classification | 练习 H
A language model classified 5,000 responses.
What Methods details matter?
186. Practice I: mixed methods | 练习 I
Scores and interviews are collected.
Write one sentence explaining integration.
187. Practice J: protocol deviation | 练习 J
Primary outcome measure changed after recruitment but before data analysis.
How should it be reported?
188. Practice K: reporting guideline | 练习 K
Choose likely reporting families for:
- randomised trial;
- observational cohort;
- systematic review;
- qualitative interviews.
189. Model answers | 示范解析
A: Room-entry time and pencil distribution are usually unnecessary unless timing/material standardisation could affect performance. Timer method may matter if task duration is an outcome/control condition.
B: Sequence generation, allocation ratio, blocking/stratification if used, allocation concealment and roles in enrolment/assignment.
C: The structured group received feedback organised into claim, evidence and reasoning prompts that each required a specific revision action; the control group received evaluative comments without action prompts.
D: We estimated the association between feedback condition and delayed revision score using linear regression adjusted for baseline score.
E: Define the quality criterion, whether it was pre-specified, who applied it, whether assessors knew outcomes/conditions, and how many cases were excluded per group.
F: Why are values missing? How will missingness be handled? What assumption does the chosen method require? A sensitivity analysis may also be needed.
G: Two researchers first coded transcripts inductively, compared code definitions, developed a shared codebook and grouped related codes into candidate themes, which were then reviewed against the full dataset. Exact wording must match the actual qualitative framework.
H: Model/version/date, prompt/instructions, preprocessing, number of runs/settings where relevant, human validation sample, performance/error evaluation, adjudication and privacy controls.
I: Interview themes were used to explain why classrooms with similar score gains differed in implementation burden, integrating the qualitative strand after quantitative analysis.
J: State the original outcome, revised outcome, rationale, timing of the change, whether outcome data/results were known, and update/identify the protocol or registration where applicable.
K: CONSORT; STROBE; PRISMA; COREQ/SRQR, with design-specific extensions where appropriate.
190. The 20-minute Methods drill | 20 分钟训练
| Time | Task |
|---|---|
| 3 min | identify design + inference |
| 4 min | map participants/intervention/outcome |
| 4 min | map preprocessing/analysis |
| 4 min | run hidden-decision audit |
| 5 min | delete diary detail + rewrite |
191. The 45-minute growth session | 45 分钟增长模式
| Time | Task |
|---|---|
| 8 min | read reporting guideline for design |
| 8 min | reconstruct one published Methods section |
| 12 min | draft design/participants/procedure |
| 10 min | draft measures/analysis |
| 7 min | replicator/reviewer audit |
192. The 90-minute deep session | 90 分钟深度模式
| Time | Task |
|---|---|
| 15 min | target-journal + guideline analysis |
| 15 min | inferential-chain map |
| 20 min | full Methods draft |
| 15 min | missingness/exclusion/preprocessing audit |
| 10 min | software/code/version audit |
| 15 min | replicator reconstruction + revision |
193. Seven-day Methods cycle | 七天训练循环
- Day 1: design + sampling.
- Day 2: intervention/exposure + comparator.
- Day 3: constructs + measures.
- Day 4: preprocessing + missing data + exclusions.
- Day 5: statistical/qualitative analysis.
- Day 6: reproducibility, code, protocol, ethics.
- Day 7: full Methods benchmark.
194. Twelve-week C1–C2 Methods progression | 12 周路线
| Weeks | Focus | Output |
|---|---|---|
| 1–2 | design, sampling, allocation | study-design sections |
| 3–4 | intervention, outcomes, instruments | procedure/measure sections |
| 5–6 | preprocessing, missing data, exclusions | data-processing sections |
| 7–8 | analysis, assumptions, sensitivity | analysis plans |
| 9–10 | qualitative/mixed/computational transparency | genre-specific Methods |
| 11–12 | reporting guidelines + independent transfer | submission-ready Methods portfolio |
195. Monthly benchmark | 每月基准任务
Choose one study design and produce:
- research question;
- design statement;
- population/sampling paragraph;
- intervention/exposure paragraph;
- measure/outcome paragraph;
- procedure paragraph;
- preprocessing/missing-data paragraph;
- analysis paragraph;
- ethics/registration statement where relevant;
- reporting-guideline checklist;
- replicator test;
- reviewer-bias test;
- hidden-decision audit;
- lab-diary deletion pass;
- Method → Result lineage map.
196. First weak-link diagnosis | 第一个薄弱环节
| Symptom | Likely weak link | Repair |
|---|---|---|
| Methods very short | inferential chain hidden | design → sampling → measure → analysis map |
| Methods extremely long | lab-diary detail | decision-impact filter |
| “random” unexplained | allocation transparency | sequence/concealment/roles |
| “structured intervention” vague | intervention identity missing | content/dose/delivery/comparator |
| software named but analysis unclear | tool–logic confusion | state estimand/model/question |
| outliers vanished | exclusion rule hidden | criterion/timing/numbers |
| qual themes “emerged” | analysis process hidden | coding/theme/reflexivity process |
| AI used but undocumented | pipeline opacity | model/procedure/validation/version |
197. What not to do | 不要这样写 Methods
- Do not write a chronological diary unless timing itself is methodologically important.
- Do not use “random” when allocation was not genuinely random.
- Do not hide recruitment, exclusions or missing-data decisions.
- Do not describe an intervention with a label that another researcher cannot reconstruct.
- Do not use software names as substitutes for analytical logic.
- Do not describe only the ideal protocol if actual conduct differed.
- Do not treat qualitative analysis as a black box.
- Do not hide AI or automated systems that materially shaped evidence.
- Do not force one reporting checklist onto the wrong study design.
- Do not confuse a longer Methods section with a more reproducible one.
198. Research and reference floor | 研究与参考基础
- Nature | Formatting Guide: Methods — concise Methods with elements necessary for interpretation and replication; protocol sharing where appropriate.
- NIH | Enhancing Reproducibility through Rigor and Transparency — rigor across design, methodology, analysis, interpretation and reporting.
- EQUATOR Network | Reporting Guidelines — design-specific reporting families including CONSORT, STROBE, PRISMA, SPIRIT, STARD, COREQ/SRQR, ARRIVE, SQUIRE and CHEERS.
- EQUATOR | Selecting the Appropriate Reporting Guideline — match the checklist to study design.
- EQUATOR | Procedure/Method Reporting Guidance — reporting-guideline resources specifically relevant to Methods.
- University of Manchester Academic Phrasebank | Describing Methods — method functions, procedural language and rationales.
- Purdue Writing Lab | Graduate Writers Guide — Methods as the analytical tools and processes that answer the research problem.
199. Canonical eduKate routes | eduKate 主页面路由
- Lesson 014 | Calibrate Certainty Without Becoming Vague
- Lesson 015 | Methods, Results and Discussion
- Lesson 016 | Abstract Compression
- Lesson 017 | Research Introduction
200. SEO language map | 本课自然覆盖的搜索意图
This lesson naturally serves learners searching for how to write methods section, research methods section, methodology section, reproducibility, replicability, transparent reporting, randomisation methods, blinding, sample size, sampling methods, missing data, data preprocessing, statistical analysis methods, qualitative methods, mixed methods, AI research methods reporting, CONSORT, STROBE, PRISMA, COREQ, SRQR, C1 academic writing, C2 academic English, Methods 怎么写, 研究方法怎么写, 可重复性, 可复现性, 随机分组, 盲法, 样本量, 数据预处理, 缺失数据, 学术英语 Methods, 中文母语学术英语 and C1 C2 research writing.
201. The one-page Methods operating system | 一页 Methods 操作系统
- Name the design accurately.
- Define population and sampling.
- Report recruitment and exclusions.
- Explain sample-size logic.
- Describe randomisation/concealment/blinding where relevant.
- Make intervention/exposure and comparator reconstructable.
- Operationalise constructs and outcomes.
- Identify instruments, versions and scoring.
- Describe procedure where sequence materially matters.
- Report preprocessing/cleaning/outlier rules.
- Report missing-data handling.
- Describe analytical logic, not just software.
- Separate confirmatory from exploratory decisions.
- Report protocol deviations honestly.
- Make human judgement inspectable.
- Document AI/automation that materially affects evidence.
- Use design-appropriate reporting guidelines.
- Run replicator, reviewer and lab-diary audits.
202. Final assignment | 最终作业
Choose one study design you understand well enough to describe honestly.
Complete this chain:
- Write the research question.
- Name the design.
- Write population and sampling logic.
- Write inclusion/exclusion rules.
- Write sample-size rationale.
- Describe allocation/blinding where relevant.
- Describe intervention/exposure and comparator.
- Define primary and secondary outcomes.
- Describe instruments/scoring.
- Write procedure.
- Write preprocessing and missing-data handling.
- Write analysis with model/variables/assumptions.
- Label exploratory analyses.
- State protocol/registration/deviations where relevant.
- State ethics/consent where relevant.
- Identify the reporting guideline for the design.
- Run the decision-impact filter on every methodological detail.
- Delete lab-diary details that do not affect reconstruction or inference.
- Give the Methods to a knowledgeable reader and ask them to reconstruct the study.
- Repair every major reconstruction error.
Then ask:
If another careful researcher obtained a different result, would my Methods give them enough information to determine whether the difference came from the phenomenon—or from a hidden methodological decision?
如果另一个认真研究者得到不同结果，我的 Methods 能不能让他判断：差异来自真正的现象，还是来自我没有写出来的方法决定？
If yes, your Methods is doing scientific work.
If no, add the missing decision—not the missing diary entry.
Continue | 继续
The next lesson will move to the other side of the Methods–Results handoff: how to write Results that make patterns visible without turning numbers into argument, hiding null findings or repeating every table cell.
下一课进入 Results 深层写作：怎样让 pattern 清楚可见，同时不把数字提前解释成 argument、不隐藏 null result，也不把每个 table cell 重新抄一遍。
Next: EDKS-ADV-ZH-0019 · Lesson No.019 · Write Results That Reveal the Pattern Without Arguing Ahead of the Evidence · 把研究结果写清楚，而不是提前替证据下结论
