---
name: data-analysis-chapter-development
description: >
  Presents quantitative or qualitative results at academic standard: results
  separated from discussion, complete statistical reporting with test statistic,
  degrees of freedom, exact p, effect size, confidence interval and n, assumption
  checks reported, themes evidenced by extracts with enough context to judge the
  interpretation, self-contained tables and figures, and null results reported
  honestly. Use for "results chapter", "chapter 4", "how do I report my
  statistics", "how much output should I include", "presenting themes with
  quotes", "my results and discussion have merged", "reporting a
  non-significant result".
category: 15 Academic University Research
ref: "15.10"
tier: 3
inherits: [K2, K3, K4, K5]
---

# Data Analysis Chapter Development

## 1. One-line description

A method for building the results chapter so that it is organised by the research questions rather than by the software output, reports every planned analysis completely including the ones that produced nothing, evidences every claim, and stops cleanly at the line where interpretation begins.

## 2. What this skill is used for

**The research problem it solves.** The results chapter fails in four characteristic ways, and rubrics penalise all four. **Collapse of results into discussion**: interpretation, comparison with the literature and speculation about causes appear inside the results, so the reader cannot tell what was observed from what is being argued, and the discussion chapter then repeats it. **Incomplete reporting**: a p value with no test statistic, no degrees of freedom, no effect size, no confidence interval and no n, which makes the result impossible to evaluate, impossible to include in any later synthesis, and impossible to defend. **The volume problem**: everything the software produced is included, so a reader must find the answer among forty tables, most of which answer nothing that was asked. **Under-evidenced qualitative claims**: themes asserted with one short extract each, stripped of the context that would let a reader judge whether the interpretation is reasonable. Two further failures are quieter and more damaging: assumption checks are performed and not reported, or not performed at all, and analyses that produced null results are silently dropped, which turns the chapter into a selective report of what worked.

**Where it sits in the research lifecycle.** After fieldwork and analysis, before the discussion. It is bounded on one side by the analysis plan in the methodology chapter, which it must execute and report against, and on the other by the discussion chapter, which it must not begin.

**Typical use cases.**
- Structuring a results chapter around research questions rather than around output.
- Reporting statistical results to the completeness standard an examiner expects.
- Presenting themes with extracts that carry enough context to be judged.
- Deciding what belongs in the chapter and what belongs in an appendix.
- Reporting a non-significant or unexpected result without apologising for it.
- Repairing a chapter where results and discussion have merged.
- Building tables and figures that stand alone.

**Who uses it.** Masters candidates writing chapter four; supervisors reviewing a results draft; students working quantitatively, qualitatively or in mixed designs, in any discipline where results are reported as a distinct chapter.

## 3. When to use it

- Analysis is complete or substantially complete and the chapter must now be written.
- The chapter currently follows the structure of the software output rather than the research questions.
- Results and discussion have merged and a supervisor has asked for them to be separated.
- You are unsure what a complete report of a statistical test includes.
- You have far more output than can be included and no principle for selecting.
- An analysis produced a null or unexpected result and you are unsure whether or how to report it.
- Themes exist but the evidence for them is thin or the extracts lack context.
- Tables and figures need numbering, captioning and checking for self-containment.

## 4. When NOT to use it

- **The analysis has not been done or is not sound.** This skill reports analysis; it does not perform or repair it. For descriptive work go to **05.01 Descriptive Analysis**, for tests to **05.02 Statistical Testing**, for regression and causal-claim control to **05.06**, and for qualitative coding and theme construction to **07.01 Thematic Analysis** and **07.03 Interview and Transcript Analysis**. Writing up an unsound analysis well is the worst available outcome, because it makes the fault harder to see.
- **The data has not been cleaned and validated.** Reporting from a dataset with unresolved duplicates, out-of-range values or undocumented exclusions produces numbers that will change, and every downstream sentence changes with them. Use **04.01 Data Validation** and **04.02 Data Cleaning** first, and log every change (K4 §4.4).
- **What you actually need is the discussion.** Interpreting results against the literature, answering the research questions in argument form, stating the contribution, handling unexpected findings and writing limitations and future research are **15.11 Discussion and Contribution Development**. This skill stops at the point where the finding has been stated and evidenced. In the other direction: 15.11 does not present results, does not report statistics, and assumes the results chapter has already established them.
- **The chapter is a combined results and discussion, which some disciplines require.** Where the department's convention genuinely is an integrated chapter, the separation rule here applies within each section rather than between chapters: state the finding, mark the transition into interpretation with an explicit signal, and keep the boundary visible at sentence level (K2 §3.3). Check the local convention before separating, because separating in a department that expects integration is also an error.
- **The results are being written for a non-academic audience.** Executive reporting compresses, leads with implication and drops the reporting apparatus. That is **12.05 Executive Research Reporting** or **11.03 Research Report Writing**. Applying academic reporting conventions to a business audience produces a document nobody reads, and the reverse produces a chapter that fails.
- **The findings would identify participants.** Where a quote, a characteristics table or a subgroup breakdown makes a participant identifiable, the reporting constraints from the ethics approval bind, and they override completeness. Return to the re-identification assessment in **15.09 Ethics Clearance Application**. `RESEARCHER DECISION REQUIRED` (K5 §2.4).
- **Your institution prohibits AI assistance for this task.** See §12 item 1. This chapter carries particular risk, because a fabricated number or extract here is indistinguishable from a real one to every reader.

## 5. Required inputs

**Required.**
- **The actual analysis output, from the actual data.** Tables, test results, model output, code frames, coded transcripts. Not summaries of them, not recollections of them. Nothing enters this chapter that is not in the output in front of you.
- **The research questions or hypotheses, final, in their submitted wording.** These are the chapter's structure and its selection rule.
- **The analysis plan from the methodology chapter.** The chapter must report what was planned, including the parts that produced nothing, and must distinguish planned from exploratory analysis.
- **The reporting constraints from the ethics approval.** What may be attributed, at what grain, and what must be generalised or withheld.

**Optional, and what each one adds.**
- **The departmental style guide and a recent accepted dissertation.** Determine table and figure numbering conventions, caption placement, statistical reporting format and how much output is conventional here. Conventions differ and the local one governs.
- **The conceptual framework's traceability matrix.** Gives the chapter a ready-made structure and ensures every construct is reported against.
- **The full participant or respondent characteristics.** Lets the chapter establish who the results describe before it describes anything else, which is required and often omitted.
- **The coding audit trail or a second coder's output.** Lets the chapter report how themes were arrived at and, where applicable, agreement.
- **Fieldwork notes on response, dropout and non-response.** Turns the sample description from a number into an account, and supplies the non-response information a reader needs.

## 6. Questions to ask before starting

1. **Does your department expect separate results and discussion chapters, or an integrated one?** Determines the whole architecture and the location of the interpretation boundary. Default if unanswered: separate, since it is the more common convention and the safer error, and mark the boundary explicitly either way.
2. **What did the analysis plan say would be done?** Determines what must be reported, including negatives. Default: report every planned analysis, and mark anything not planned as exploratory.
3. **Which results answer a research question, and which are merely available?** Determines what goes in the chapter and what goes in an appendix. Default: chapter for anything that answers a question or is needed to judge one, appendix for the rest, and nothing at all for output that answers nothing.
4. **What are the reporting constraints from ethics?** Determines quote handling, participant description and subgroup reporting. Default: apply the strictest interpretation of the approval and check with the supervisor before relaxing it.
5. **Were assumptions checked, and what did the checks show?** Determines whether the chapter can report the tests it reports. Default: if checks were not run, run them now, because a test whose assumptions were violated may need replacing, and this is far cheaper to find before submission.
6. **How many extracts per theme can the word count carry, and what is the selection rule?** Determines qualitative evidence density. Default: two to three extracts per theme from different participants, selected to show the range within the theme rather than only its clearest expression, with the rule stated.
7. **Is any analysis exploratory, conducted after seeing the data?** Determines how it must be labelled. Default: label it, because an unlabelled post hoc analysis presented as planned is a research integrity issue, not a presentational one.

## 7. Step-by-step methodology

**Step 1. Build the chapter's structure from the research questions, before writing anything.**
Write the research questions as the section headings, in the order they appear in the introduction. Under each, list the analyses or themes that answer it. Then look at what is left over: output that answers nothing goes to an appendix or is dropped, and any research question with nothing under it is a problem to solve now, not at submission. Two structures are defensible: organised by research question throughout, which is the safer default and the one most rubrics reward, or organised by method strand (quantitative then qualitative) with research questions as subsections, which suits mixed designs where the strands are analytically separate. What is not defensible is organisation by software output, which is how chapters end up with a section per variable and no section per question. State the chapter's structure in its opening paragraph so the reader knows what is coming. *Correct result: a section plan in which every research question has a section, every section has evidence beneath it, and no section exists without a question above it.*

**Step 2. Draw the results-discussion line explicitly and enforce it at sentence level.**
Everything to the left of the line is what the data show; everything to the right is what it means. In the results chapter you write findings and, where necessary, the immediately data-anchored observation that makes a finding legible, and you stop. Specifically, the following do not belong in the results: comparison with prior literature, explanation of why a result occurred, speculation about mechanism, implications for practice or theory, and evaluation of whether a result is good or surprising. The following do belong: the finding, its statistical or evidential support, its magnitude and direction, the base it rests on, and the descriptive observation of pattern (that an effect appeared in one subgroup and not another is a finding; why it might have is discussion). The enforcement test: scan for the words "because", "suggests", "indicates", "supports the view that", "consistent with the literature", "this may be due to". Each occurrence is either a boundary breach or needs the interpretation moved. Some disciplines permit a short closing paragraph per section summarising the finding without interpreting it; that is fine, and it is not an invitation to interpret (K2 §2.1, §3.3). *Correct result: a chapter through which a reader could not tell what you think the results mean, only what they are.*

**Step 3. Establish who and what the results describe, before describing them.**
Open with the sample and the data actually obtained. Quantitative: the number approached, the number responding, the response rate and how it was calculated (or a statement that it cannot be calculated and why, for open distribution routes), exclusions with reasons and counts, the final analysable n, and the demographic and other characteristics relevant to the questions. Report non-response information where you have any, since a sample of 200 from 1,000 approached is a different object from a sample of 200 from 220. Qualitative: the number of participants, the data volume (interviews conducted, duration range, total transcript length or equivalent), the characteristics relevant to the questions at a grain the ethics approval permits, and the recruitment outcome including who declined or withdrew. In both cases, state departures from the planned sample and their reasons. This section is short and it does a great deal of work: it is the base against which every subsequent claim is read, and its absence is why so many results chapters cannot be evaluated. *Correct result: a sample section from which a reader can state exactly who the findings describe and how many people they rest on.*

**Step 4. Report assumption checks and data conditions before the analyses that depend on them.**
Every inferential test rests on conditions, and reporting the test without the checks asks the reader to assume you did them. State which checks were run, how, and what they showed: distributional checks for tests that assume a distribution, variance homogeneity where the test requires it, independence of observations where the design could compromise it, linearity and residual behaviour for regression models, multicollinearity where predictors may overlap, and sphericity for repeated measures designs. Report what you found, including where an assumption was violated, and what you did about it: transformed the variable, used a robust or non-parametric alternative, adjusted the degrees of freedom, or proceeded with the violation noted and its consequence stated. Report missing data explicitly: how much, on which variables, whether the pattern was examined, and how it was handled, since a listwise deletion that silently drops a third of the sample changes what the results describe. For qualitative work, the equivalent is a short account of data adequacy: the range of participants who spoke to each theme, whether any theme rests on a small number of accounts, and any material that could not be used and why. *Correct result: an assumptions and data conditions section that precedes the analyses, reports violations, and states the remedy for each.*

**Step 5. Report quantitative results completely, every time.**
A complete report of an inferential result contains: the test name, the test statistic with its symbol, the degrees of freedom, the exact p value, an effect size with its name, a confidence interval where the statistic supports one, the n on which it rests, and the direction and magnitude in substantive terms. Written out, that is a sentence such as: an independent-samples comparison of the two groups showed a difference in mean score, t(148) = 2.31, p = .022, Cohen's d = 0.38, 95% CI [0.05, 0.71], n = 150, with the intervention group scoring 4.2 points higher. Report exact p values rather than thresholds, except where the value is smaller than the precision you can report, in which case use the conventional inequality. **The effect size is not optional**, and it is what a p value cannot tell you: with a large sample almost anything reaches significance, and with a small one a substantial effect may not. **The confidence interval is what turns a point estimate into a statement about precision**, and a wide interval is a finding about the study's power that a bare p value conceals. Report descriptive statistics with the dispersion measure, not the central tendency alone, and match the measure to the distribution. Round to the precision the method supports and no further (K3 §4.4). Never describe a difference as significant without a test behind it, and never describe an untested difference in language that implies one (K4 §3.1). And keep causal language out of the results entirely, whatever the design (K4 §3.2). *Correct result: every inferential result reportable in one sentence containing all seven elements, with no result stated as significant on inspection alone.*

**Step 6. Report the analyses that produced nothing, in the same detail as the ones that did.**
Every analysis specified in the analysis plan is reported, whatever it found. A non-significant result is reported with the same completeness as a significant one, including the effect size and the confidence interval, because those are what tell the reader whether the study found no effect or merely failed to detect one, which are entirely different findings. Write it accurately: "no statistically significant difference was found, t(88) = 0.94, p = .350, d = 0.20, 95% CI [-0.22, 0.62], n = 90" states the result; "there was no difference between the groups" overstates it; and dropping the analysis entirely misrepresents the study. Where an interval is wide and includes effects that would matter substantively, say so, because that is the honest reading of an underpowered null. The same rule holds qualitatively: a theme you expected and did not find is reported as expected and not found, and an interview question that yielded little is reported as yielding little. Interpretation of why belongs in the discussion; the reporting belongs here. This step is also where research integrity is most visibly at stake, because selective reporting of what worked is the single most common form of distortion in student and professional research alike. *Correct result: a chapter in which the count of analyses reported equals the count of analyses planned, plus any exploratory analyses, each labelled.*

**Step 7. Present qualitative results as themes evidenced with enough context to be judged.**
Each theme gets the same structure: a name that describes the content rather than labelling the topic ("Managing the gap between policy and caseload" rather than "Workload"); a definition of what the theme covers and what it excludes; a statement of its distribution across the dataset in appropriate language; two or more extracts from different participants; and a short analytical statement of what the theme comprises, which describes the pattern without interpreting its significance. On distribution: quantify carefully and in the register your tradition permits. Counts of participants ("14 of 24 participants described...") are informative and honest; percentages on small qualitative samples imply a precision the method does not have (K4 §7). Some interpretive traditions reject prevalence claims altogether, in which case use the tradition's own language of typicality and salience, and say which convention you are following. On extracts: they must be long enough to carry their context, since a six-word fragment can be made to support almost anything and an examiner reading it cannot judge your reading. Include the question or the turn that prompted it where the meaning depends on it. Attribute every extract with a participant identifier and the characteristics your ethics approval permits (K2 §4.2). State your transcription and editing conventions once, and keep edits within them (K4 §2.3). Select extracts to show the range within a theme, including the participant who fits it least comfortably, rather than only its clearest expression, because a theme evidenced only by its best example is a theme nobody can evaluate. *Correct result: each theme defined, distributed, evidenced by at least two contextualised and attributed extracts from different participants, and described without being interpreted.*

**Step 8. Solve the volume problem with a stated selection rule.**
Analysis produces far more output than any chapter can carry, and including all of it is not thoroughness, it is an unsorted transfer of work to the reader. Apply a three-way rule and state it in the chapter. **In the chapter**: any result that answers a research question, any result needed to judge such a result (sample characteristics, assumption checks, reliability figures), and any result that contradicts the study's overall pattern, which must never be exiled to an appendix. **In an appendix**: full model output, complete cross-tabulations, the full code frame, additional descriptive tables, and anything a sceptical reader might want to inspect but does not need in order to follow the argument. **Nowhere**: output that answers no question and informs no judgement. The discipline this imposes is useful in itself, since a result you cannot assign to a research question is usually a result you ran because the software offered it. Cross-reference every appendix from the chapter, because an appendix nobody is pointed to is not read and does not count. *Correct result: a chapter whose every element is assignable to a research question or to a judgement about one, with a stated rule and cross-referenced appendices.*

**Step 9. Build tables and figures that stand alone.**
The test is that a reader who opens the dissertation at that page can understand the table without reading the surrounding text. That requires: a number and a caption stating what is shown, in which population, with the n (tables captioned above, figures below, unless the local style says otherwise, and consistently either way); all variables and categories labelled in words rather than in variable names from the dataset; units stated; the base for every percentage, and a statement of what the percentage is a percentage of; notes defining any abbreviation or symbol, including the significance notation if used; and the source or the analysis that produced it. Two further rules. **Do not duplicate**: a table and a figure showing the same data is padding, and a table whose contents are also narrated in full in the text is padding twice. Use the text to state the finding and point to the table for the detail. **Do not decorate**: chart type follows the data structure rather than visual appeal, axes start at a defensible value and any truncation is stated, and colour is not the only carrier of meaning. Number sequentially by chapter, and check in the final draft that every table and figure is referred to in the text, in order, and that the numbers still match after edits, which is where late errors concentrate. See **11.04 Data Visualisation and Chart Selection** for chart choice. *Correct result: every table and figure numbered, captioned, self-contained, referenced in the text, and non-duplicative.*

**Step 10. Report mixed designs at the point of integration, not as two chapters in one.**
Where a study has both strands, decide and state where they meet. If integration occurs at interpretation only, present the strands separately and integrate in the discussion, saying so. If integration occurs in the results, present it explicitly: a joint display placing the quantitative result and the qualitative evidence side by side against the same research question is the clearest device, with a column stating whether the strands converge, diverge or complement. Where they diverge, report the divergence as a result rather than resolving it, since the divergence is frequently the study's most valuable finding and averaging it away destroys it (K2 §4.4). *Correct result: a stated integration point, and where integration is in the results, a joint display per research question with convergence explicitly characterised.*

**Step 11. Run the traceability and completeness audit.**
Check, mechanically: every research question has a section and an answer; every number in the text appears in a table or in the analysis output, and matches it exactly; every quote is traceable to a participant identifier and matches the transcript; every table and figure is cited in the text; every appendix is cross-referenced; every planned analysis is reported; every exploratory analysis is labelled; the total counts reconcile (participants in the sample section equal participants in the tables, and subgroup counts sum to the total). Then re-read for boundary breaches, since interpretation creeps back in during editing. Finally, read the chapter cold and ask the reader's question of each section: what is the answer to the research question this section sits under? If you cannot state it in one sentence from what is written, the section has reported without concluding. *Correct result: a completed audit with every check passing, and a one-sentence answer available for every research question.*

## 8. Analytical framework

**The reporting boundary.** The chapter occupies exactly three levels of the evidence chain and stops:

    Evidence → Analysis → Finding | Interpretation → Insight → Implication
                                  ↑
                       The chapter ends here.
                       Everything left is the results chapter.
                       Everything right is 15.11.

Applying it: the boundary is enforced at sentence level, not chapter level, because interpretation enters through single clauses rather than whole paragraphs. The reliable detector is the causal or comparative connective: "because", "suggests", "in line with", "this reflects". Each one is a crossing. A finding may be stated with its magnitude, direction, base and precision; the moment a sentence explains it or relates it to anything outside the dataset, it has moved chapters (K2 §2.1).

**The completeness unit.** Every reported result, quantitative or qualitative, is built from the same four components, and a result missing any of them cannot be evaluated:

    What was found → What it rests on (n, base, participants) →
    How strongly (effect size and interval, or prevalence and range) →
    Where to check it (table, figure, appendix, participant ID)

Applying it: the four components are what separate a reportable result from an assertion. A p value alone gives the first and none of the rest. A theme name alone gives the first and none of the rest. The discipline of writing all four for every result is what makes the chapter examinable, and it is also what makes the results usable by anyone who later wants to include the study in a synthesis.

## 9. Output format

**1. Chapter introduction.** What the chapter covers, its structure, and an explicit statement that interpretation is reserved for the following chapter.

**2. Sample and data obtained.** Approached, responded, response rate or why it cannot be calculated, exclusions with reasons, final n, characteristics, departures from plan, non-response information.

**3. Data conditions and assumption checks.** Missing data and its handling, distributional and other checks, violations found, remedies applied. Qualitative equivalent: data adequacy and coverage.

**4. Results by research question.** One section per question. Within each:

- The finding, stated plainly.
- The complete statistical report, or the theme with its definition, distribution and extracts.
- The table or figure carrying the detail, cross-referenced.
- Contradicting or unexpected results within that question, reported alongside.

**5. Statistical reporting format**, used consistently: test, statistic with symbol, degrees of freedom, exact p, effect size named, confidence interval, n, and the substantive magnitude and direction.

**6. Qualitative theme format**, used consistently: theme name, definition and boundary, distribution across the dataset, two or more attributed extracts from different participants, analytical description.

**7. Mixed designs.** Joint display per research question, with convergence, divergence or complementarity stated.

**8. Summary of results.** A short factual restatement, one entry per research question, with no interpretation.

**9. Appendices**, cross-referenced from the chapter.

**When results are thin, null or messy, the format must not force fabrication or inflation (K4 §1).** A null result is reported with its effect size and interval, which is more informative than a significant result reported without them. A theme supported by three participants is reported as supported by three participants. A response rate that cannot be calculated is reported as incalculable with the reason. A planned analysis that could not be run is reported as not run, with the reason. None of these is a failure of the study; concealing any of them is a failure of the researcher.

## 10. Quality checks

Run before the chapter goes to a supervisor. These sit on top of K4 §8.

1. Is the chapter organised by research question rather than by software output or by variable?
2. Does every research question have a section, and does every section answer its question?
3. Is there any interpretation, literature comparison, causal explanation or implication in the chapter?
4. Does the sample section state approached, responded, excluded and analysable, with reasons?
5. Are assumption checks reported before the analyses that depend on them, including violations and remedies?
6. Is missing data reported with its extent, pattern and handling?
7. Does every inferential result carry test statistic, degrees of freedom, exact p, effect size, confidence interval and n?
8. Is every planned analysis reported, including those that produced null results, with effect sizes and intervals?
9. Is every exploratory analysis labelled as exploratory?
10. Is the word "significant" used only where a test was run, and is any untested difference described as untested?
11. Is causal language absent, whatever the design?
12. Does every theme have a definition, a distribution statement, and at least two attributed extracts from different participants?
13. Do extracts carry enough context for a reader to judge the interpretation, and do they include the awkward case as well as the clear one?
14. Are transcription and editing conventions stated once and adhered to?
15. Is every table and figure numbered, captioned with its n, self-contained, cited in the text, and non-duplicative of the prose?
16. Does every number in the text match the output exactly, and do the counts reconcile across sections?
17. Does every quote carry a participant identifier, and does the attribution comply with the ethics approval?
18. Is anything in the chapter present because the format expected it rather than because it answers a question?

## 11. Common failure modes

| Failure | How to recognise it | How to prevent it |
|---|---|---|
| **Results-discussion collapse** | "Consistent with the literature" appears in chapter four | Scan for connectives, move every crossing (Step 2) |
| **Output-driven structure** | A section per variable, none per research question | Build the structure from the questions first (Step 1) |
| **Incomplete statistical reporting** | A p value with no statistic, effect size or n | Use the seven-element sentence every time (Step 5) |
| **Threshold reporting** | "p < .05" where the exact value is available | Report exact p values (Step 5) |
| **Silent assumption checks** | Tests reported with no mention of their conditions | Report checks before the analyses (Step 4) |
| **Disappearing null results** | Fewer analyses reported than the plan specified | Count planned against reported (Step 6) |
| **The overstated null** | "There is no difference" from a non-significant test | Report the interval and what it does not exclude (Step 6) |
| **The volume dump** | Forty tables, most answering nothing | Apply and state the three-way selection rule (Step 8) |
| **Fragment quoting** | Six-word extracts supporting large claims | Extracts long enough to carry context (Step 7) |
| **Cherry-picked extracts** | Every quote is the theme's clearest expression | Select for range, include the awkward case (Step 7) |
| **Percentages on tiny qualitative samples** | "67% of participants" where n is 9 | Use counts, or the tradition's own language (Step 7) |
| **Orphan tables** | A table nobody refers to in the text | Audit citations of every table and figure (Step 11) |
| **Duplicated data** | The same figures in a table, a chart and a paragraph | State the finding, point to the table (Step 9) |
| **Reconciliation failure** | Subgroup counts that do not sum to the total | Run the counts audit (Step 11) |
| **AI-fabricated statistics** | A number in the text that is in no output file | Nothing enters the chapter that is not in the output (§12) |
| **AI-fabricated or polished quotes** | An extract that reads too cleanly for speech | Every extract is copied from the transcript verbatim (§12) |
| **AI-invented themes** | A plausible theme with no coded data beneath it | Themes come from the coding, not from expectation (§12) |

## 12. AI guardrails

Skill-specific. The universal prohibitions in K4 apply in full and are not repeated. The human in the loop for every K5 marker in this skill is the candidate together with their supervisor.

1. **Academic integrity is a condition of use, not a footnote.** These skills assist a researcher's thinking, structure and rigour. They do not produce work to be submitted as the student's own unaided output. The user must comply with their institution's AI use policy and its declaration requirements, which vary by institution and by assessment. Where an institution prohibits AI assistance for a task, this skill must not be used for it. The skill never writes a passage for submission as though the student wrote it; it interrogates, structures, critiques and teaches. Operationally in this skill: it will supply the chapter structure, the reporting templates, the completeness checklists, the boundary audit and a critique of the student's draft, and it will tell them which of their reported results is incomplete. It will not write the results narrative, and it will not supply any number, extract, theme or table content.

2. **Never produce a number that is not in the output the student supplied.** Not a mean, not a percentage, not a test statistic, not a p value, not an effect size, not an n. This is the highest-consequence prohibition in the entire academic track, because a fabricated statistic in a results chapter is undetectable by every reader, is defended in a viva by someone who does not know it is false, and constitutes research misconduct. Where a value is needed and absent, the output is `[not available]` with a statement of what analysis would produce it (K4 §2.1).

3. **Never write, complete, tidy or reconstruct a participant extract.** A quote is copied from the transcript or it does not exist. Permitted edits are only those in K4 §2.3, and the convention must be stated in the chapter. Never merge two participants' words, never attribute an extract to a participant whose membership of the stated group has not been checked, and never smooth speech into readable prose while calling it verbatim.

4. **Never generate a theme, a code or a pattern the student has not found in their data.** A plausible theme offered as a suggestion becomes a theme in the chapter within one revision, and it will have no coded data beneath it. Assistance with qualitative results is limited to structure, definition sharpening, evidence adequacy critique and boundary checking against the coded material the student supplies.

5. **Never calculate a statistic from a description of data rather than from the data.** An effect size inferred from reported means without the dispersion, or a p value reconstructed from a test statistic without confirming the test and its degrees of freedom, produces numbers that look right and are not. Say what is needed to compute it.

6. **Never describe a result as significant, meaningful, strong or notable without the test and the effect size in front of you** (K4 §3.1), and never introduce causal language into a results chapter regardless of the design (K4 §3.2).

7. **Never help a chapter omit a planned analysis.** Where a student's draft reports fewer analyses than the plan specified, the correct response is to point at the gap. A missing null result is the most common form of selective reporting and the easiest to fix before submission.

8. **Never upgrade an exploratory analysis to a planned one, or accept a description that implies it was planned.** The distinction is fixed by the methodology chapter, and moving it retrospectively is an integrity failure rather than a wording choice.

9. **Never relax the ethics reporting constraints to make a result more vivid.** A more specific attribution, a richer participant description or an unedited quote may make the chapter better and the participant identifiable. The approval binds (K5 §2.4).

10. **Where a result contradicts the study's overall pattern, insist it stays in the chapter and out of the appendix** (K4 §4.1). Contradicting evidence is a result. An appendix is where it goes to be missed.

## 13. Best-practice principles

- **The research questions are the chapter's outline, its selection rule and its completeness test, all three.** Almost every structural problem in a results chapter dissolves when the questions become the headings.
- **The effect size is the finding; the p value is only whether you can rule out chance.** Reporting one without the other is half a result, and it is the half that says less.
- **A confidence interval is the most honest single number in a results chapter**, because it shows the reader the precision they are actually being offered.
- **Report the null result completely and it becomes informative.** Report it as "no difference" and it becomes wrong. Omit it and the chapter becomes selective.
- **A quote is evidence, and evidence needs enough context to be assessed.** The extract that needs the reader to trust you is not doing evidential work.
- **Show the range within a theme, not the best example of it.** The participant who fits least comfortably is usually the most informative one in the chapter.
- **Tables are for detail, prose is for the finding.** Narrating a table in full doubles the length and halves the clarity.
- **If you cannot say which research question a result answers, it does not belong in the chapter.** This single test removes most of the volume problem.
- **Write the sample section first.** It fixes the base for everything else and stops later claims drifting away from what the data can support.
- **Interpretation creeps back in during editing.** Run the boundary scan on the final draft, not only the first.
- **Every number in the chapter should be findable in an output file within thirty seconds.** If it is not, either the audit trail or the number has a problem.
- **The chapter is written for a reader who wants to disagree with you.** Give them what they need to check, and the ones who agree will trust it more.

## 14. Worked example

Generic fictional scenario, academic, in an education setting.

**INPUT**

A masters candidate in education has run a mixed-methods study on a peer-tutoring scheme in one secondary school: a pre-post measure of a validated confidence scale with 96 students, a comparison against a non-participating group of 84, and interviews with 11 participating students. Three research questions: whether participation is associated with a change in confidence, whether any change differs by prior attainment band, and how participating students describe the experience. The draft chapter has 31 tables, a section per statistical procedure, and interpretation woven throughout.

**PROCESS**

*Step 1.* The chapter is restructured to three sections, one per research question, with the sample and assumptions sections in front. Of the 31 tables, 6 answer a research question, 4 are needed to judge one (sample characteristics, scale reliability, assumption checks, missing data), 9 go to appendices, and 12 answer nothing that was asked and are dropped. The chapter loses a third of its length and gains an argument.

*Step 3.* The sample section is written for the first time. 180 students were invited, 96 participated and 84 formed the comparison group, and 12 of the participating group did not complete the post measure, leaving 84 analysable, which the draft had never stated. The dropout is reported with its reason where known, and the analysable n corrected throughout, which changes three of the reported results.

*Step 4.* Assumption checks are run properly for the first time. The change scores are acceptably distributed, but variance differs materially between the two attainment bands used in research question two, so the planned test is replaced with a variant that does not assume equal variances, and this is reported with the reason.

*Step 5.* Every result is rewritten to the seven-element standard. The draft's "the intervention group improved significantly (p<0.05)" becomes a full sentence with the paired statistic, degrees of freedom, exact p, effect size, confidence interval, n and the magnitude in scale points.

*The judgement call.* Research question two produced nothing: the difference in change between attainment bands was not statistically significant. The candidate's draft omitted it and the supervisor had not noticed. Three options are considered: leave it out, report it as "no difference", or report it fully. The first is selective reporting of a planned analysis. The second overstates: the confidence interval on the difference is wide and includes effects that would be educationally meaningful, so the study has not shown there is no difference, it has failed to detect one. The result is reported in full with its interval, and a single factual sentence notes that the interval does not exclude a difference of practical size. Why that might be, and what it implies, is left entirely for the discussion. This turns the chapter's apparent weak point into its clearest demonstration of statistical literacy, and it also gives the discussion something substantive to say about power.

*Step 7.* The qualitative section is rebuilt. Four themes are given content-describing names, definitions with boundaries, and distribution stated as participant counts rather than percentages on a base of 11. Extracts are lengthened to carry context, each attributed by participant identifier and year group at the grain the ethics approval permits. For the strongest theme, an extract is added from the one participant who described the scheme as unhelpful, which the draft had excluded, and which materially changes how the theme should be read.

*Step 10.* A joint display is built for research question one, placing the quantitative change alongside the qualitative accounts, with the convergence column noting that the two strands agree on direction while the interviews locate the change in a narrower domain (confidence in asking for help) than the scale measures.

**OUTPUT**

A chapter of roughly two-thirds the original length, organised by research question, opening with sample and assumptions, reporting all three planned analyses completely including the null, presenting four themes with defined boundaries and contextualised attributed extracts including a disconfirming case, with ten tables and figures that are self-contained and cited, nine appendices cross-referenced, and no interpretation anywhere.

`RESEARCHER DECISION REQUIRED` was raised at Step 7 on whether the disconfirming extract could be included without making the participant identifiable in a single-school study, resolved by generalising the year group (K5 §2.4). `RESEARCHER REVIEW RECOMMENDED` on the joint display's convergence characterisation, since judging whether two strands measuring adjacent constructs genuinely converge is an interpretive call (K5 §2.6).

## 15. Advanced usage

**Write the results shell before the data arrives.** Build the chapter structure, the empty tables with their captions and the reporting sentences with placeholders straight from the analysis plan. This makes the planned-versus-exploratory distinction automatic, removes the temptation to structure around whatever the output produced, and turns writing-up into filling in rather than composing.

**The reverse audit.** Take the finished chapter and try to reconstruct the analysis plan from it. Anything in the chapter that does not appear in the reconstruction was unplanned and should be labelled exploratory; anything in the plan missing from the reconstruction was dropped and must be restored. This takes twenty minutes and catches the failure examiners look for.

**Sensitivity reporting.** Where a result depends on a judgement (an exclusion rule, a transformation, a coding decision, a cut-point), report the main result and state what happens under the alternative. A result that survives the alternative is strengthened at almost no cost in words; one that does not is a fragility the discussion must address rather than a fact to be hidden.

**Negative case analysis as a reporting device.** In qualitative work, giving each theme a short explicit treatment of the cases that did not fit both strengthens the credibility claim made in the methodology and pre-empts the examiner's search for the counterexample. See **07.01 Thematic Analysis**.

**Where the standard approach does not fit.** Case study and narrative designs may present findings case by case rather than theme by theme, in which case the research questions are answered in a cross-case section that must still exist. Longitudinal designs organise by time point within research question rather than the reverse. Designs producing very small samples may report at participant level throughout, in which case the ethics reporting constraints bind hardest and the identifiability assessment must be revisited before writing. In every variant, the two non-negotiables survive: organisation that answers the questions, and a visible boundary before interpretation.

## 16. Skill chain

**Recommended previous skills:**
- **15.08 Research Design and Methodology Chapter.** Hands over the analysis plan this chapter must execute and report against, including the planned-versus-exploratory distinction.
- **05.02 Statistical Testing** and **05.01 Descriptive Analysis.** Hand over the quantitative output.
- **07.01 Thematic Analysis** and **07.03 Interview and Transcript Analysis.** Hand over the coded data, themes and evidence.
- **15.09 Ethics Clearance Application.** Hands over the reporting constraints on attribution and participant description.

**Recommended next skills:**
- **15.11 Discussion and Contribution Development.** Takes the findings and interprets them against the literature, which this chapter deliberately does not do.
- **15.12 Thesis Defence and Viva Preparation.** Takes the chapter's weak points, which are visible in the reporting, and prepares defences for them.
- **15.24 Journal Article Development**, where the results are being compressed for publication.

**Runs well alongside:**
- **11.04 Data Visualisation and Chart Selection**, for the figures.
- **07.04 Quote and Evidence Extraction**, for extract selection and verification.
- **13.03 AI Output Verification**, for auditing the chapter's numbers and quotes against source before submission.

---
A Yazi Supplied Skill and resource.