Chapter 7 — Dissertation writing and limitations

Chapter 7 — Dissertation writing and limitations

Lesson map

Lesson 07 published. Map, example, EX039.

flowchart TD
  N0["Learning outcomes and connection"]
  N1["7.1 Structure: one job per chapter"]
  N2["7.2 The argument and the results-to-claims register"]
  N3["7.3 Reporting results honestly"]
  N4["7.4 Number provenance"]
  N5["7.5 Reporting guidelines as checklists"]
  N6["7.6 Directional limitations"]
  N7["7.7 Accessible figures and tables"]
  N0 --> N1 --> N2 --> N3 --> N4 --> N5 --> N6 --> N7
  N7 --> EX["Worked example"]
  EX --> Q["Exercise EX039"]

Logical drawing

flowchart LR
  N0["Learning outcomes and connection"]
  N1["7.1 Structure: one job per chapter"]
  N2["7.2 The argument and the results-to-claims register"]
  N3["7.3 Reporting results honestly"]
  N4["7.4 Number provenance"]
  N5["7.5 Reporting guidelines as checklists"]
  N6["7.6 Directional limitations"]
  N7["7.7 Accessible figures and tables"]
  N0 --> N1 --> N2 --> N3 --> N4 --> N5 --> N6 --> N7 --> Q["EX039"]

[Learning outcomes and connection]
    |
    v
[7.1 Structure: one job per chapter]
    |
    v
[7.2 The argument and the results-to-claims register]
    |
    v
[7.3 Reporting results honestly]
    |
    v
[7.4 Number provenance]
    |
    v
[7.5 Reporting guidelines as checklists]
    |
    v
[7.6 Directional limitations]
    |
    v
[7.7 Accessible figures and tables]
    |
    v
[Worked example] --> [Exercise EX039]

Learning outcomes and connection

By the end of this chapter the learner will be able to:

1. Structure an empirical master's dissertation so that each chapter has one job and the argument runs visibly from question to conclusion.
2. Keep a results-to-claims register linking every claim in the abstract, discussion and conclusion to a specific result and its limits.
3. Report estimates with their uncertainty, avoid dichotomous "significance" language, and calibrate causal verbs to the design.
4. Make every number in the text traceable to analysis output, and check it with a script.
5. Use reporting guidelines for observational, routinely collected and qualitative data as checklists.
6. Write limitations that state the direction and likely size of each bias and what would resolve it.
7. Produce accessible figures and tables, and a data and code availability statement that respects information sensitivity.

This chapter turns Chapters 1–6 into a document an examiner can assess. Chapter 8 prepares for that assessment.

7.1 Structure: one job per chapter

A typical empirical master's dissertation has the following parts. Word budgets vary by institution; the proportions below are a common starting point for a 20,000-word dissertation.

Write methods first (they are already in the pre-analysis plan), then results, then discussion, and the introduction and abstract last, when you know what the dissertation actually shows. Many weak dissertations are weak because the introduction was written first and promised a finding the results did not deliver.

7.2 The argument and the results-to-claims register

Examiners follow the thread from the questions in Chapter 1 to the answers in the conclusion. Make it visible: every research question is answered explicitly in the discussion, using the same wording, and the conclusion repeats those answers in their scoped form.

A results-to-claims register is a table you keep while writing and attach as an appendix or use privately before submission. Each row is a claim that appears anywhere in the dissertation:

The register catches the commonest failure: a claim that grew while moving from results to abstract.

7.3 Reporting results honestly

7.3.1 Estimates and intervals, not verdicts

Report the estimate, its interval and what the interval rules out. "Alder Point's arrivals rose by 24.4% (95% interval 6.7% to 44.9%)" carries information; "a statistically significant increase (p < 0.05)" carries less and invites misreading. The American Statistical Association's statement on p-values (Wasserstein and Lazar, 2016) sets out principles including that p-values do not measure the probability that a hypothesis is true, that scientific conclusions should not rest only on whether a p-value passes a threshold, and that a p-value does not measure the size or importance of an effect. For an inconclusive result, report what the interval excludes: Calloway's interval excludes increases above about 12%; Brookfield's excludes neither zero nor a 20% increase.

7.3.2 Calibrated verbs

Causal language is not forbidden; a controlled ITS is a causal design. It must be accompanied by its assumptions in the same paragraph.

7.3.3 Multiplicity

The Spillover Study had one primary outcome (arrivals), three sites and one secondary outcome. State in the methods which estimates are primary, report all of them, and do not select the most striking for the abstract. When many outcomes are examined, some will look notable by chance; the prespecified primary outcome is the defence.

7.4 Number provenance

Every number in the text should come from the analysis output, not from memory or an earlier draft. Three practices help: generate results tables directly from scripts; keep a machine-readable results file from the final run; and check the text against it before submission. Lab 7 below automates a first pass.

Rounding deserves its own rule. The Spillover Study's main estimate is 24.4% in Table 4.2 and 24% in the rewritten abstract; both are correct only if the dissertation states that the abstract rounds to whole percentages. Keep the same precision for the same quantity in every chapter, and record units beside each value in the results file, so that a percentage cannot be confused with a count of minutes. When a figure changes in a late rerun, search the whole text for the old value, including captions, the abstract and the appendices, before resubmitting the draft.

7.5 Reporting guidelines as checklists

Reporting guidelines list what a reader needs to assess a study; using them as a writing checklist prevents omissions. Relevant ones for this field:

• STROBE for observational studies (von Elm et al., 2007).
• RECORD, its extension for studies using routinely collected health data (Benchimol et al., 2015) — relevant whenever data come from operational systems not designed for research, which is nearly always true in infrastructure research.
• SRQR (O'Brien et al., 2014) and COREQ (Tong, Sainsbury and Craig, 2007) for qualitative research and interviews.
• For interrupted time series there is no single reporting standard; Lopez Bernal et al. (2017) and the methodological review by Hategeka et al. (2020) identify what readers need: the impact model, the handling of autocorrelation, seasonality and trends, and the control series.

Attach the completed checklists as appendices with page or section references.

7.6 Directional limitations

A list of weaknesses ("small sample; one region; possible confounding") tells an examiner nothing they could not have guessed. A directional limitation states the problem, its likely effect on the main estimate (direction and rough size), what the study did about it, and what would resolve it.

7.7 Accessible figures and tables

Follow the book standard's expectations for your own work: every figure has a caption that states its message and a text alternative that conveys its content to a reader who cannot see it; colour is never the only carrier of meaning; tables have descriptive headers and a one-sentence plain-language summary when they are complex. An event-study figure's text alternative might read: "Daily log ratio of Alder Point to Wexmoor arrivals, days −180 to +89, with the fitted trend. The ratio is flat before day 0, rises during days 0–20 by about 0.2 on the log scale, and returns near the trend after day 34." Write numbers in text with the same precision as in tables.

7.8 Data and code availability under sensitivity

Infrastructure studies often cannot release their data. Say what is available, at what tier, and why the rest is not:

Code and synthetic test data are available in the project repository [link] under CC0. Daily ED aggregates were provided by the participating hospitals under data agreements that do not permit redistribution; researchers may request access from the custodians, whose contact routes are listed in Appendix C. Interview transcripts are not available because participants were promised confidentiality and some could be identified by role. Facility-level capacity figures are withheld because of their potential to aid targeting; regional summaries are reported in Chapter 4.

A statement like this meets the spirit of the FAIR principles for what can be shared and is honest about what cannot.

7.9 Worked example: rewriting the abstract (synthetic)

The candidate's first abstract ended: "This study proves that ransomware attacks significantly increase emergency-department crowding across the region, demonstrating the urgent need for cyber-resilience investment." The results-to-claims register flagged four problems: "proves" (the design supports estimation under assumptions); "ransomware" (the exposure was diversion); "across the region" (only the nearest ED showed a clear increase); "urgent need for investment" (the study did not evaluate any investment). Rewritten conclusion of the abstract:

During a 21-day ransomware-induced ambulance diversion in one synthetic region, ambulance arrivals at the nearest neighbouring emergency department rose by an estimated 24% (95% interval 7% to 45%) relative to an unaffected comparison region; estimates at more distant departments were small and inconclusive. Interviews indicated that diverted ambulances followed road geography and that neighbours learned of the diversion informally. Regional diversion plans may need explicit load-sharing and notification arrangements; whether these reduce strain was not tested.

7.10 Practical activity: number provenance

The lab holds the results from Lab 6 and a draft paragraph written partly from memory. It extracts every number from the text, checks whether it matches a recorded result (allowing rounding), and flags wording that overstates. Save it as lab7_provenance.py: the EX039 worked answer reads the draft from that file.

# Lab 7 - every number in the results text must trace to the analysis output (offline)
# RESULTS mirrors the Lab 6 output; DRAFT is a synthetic dissertation paragraph written from memory.
import re

RESULTS = {                      # value, as a percentage or a count, from the analysis log
    "AP_effect": 24.4, "AP_lo": 6.7, "AP_hi": 44.9,
    "BF_effect": 5.8, "BF_lo": -6.7, "BF_hi": 19.9,
    "CG_effect": 0.8, "CG_lo": -9.3, "CG_hi": 12.1,
    "AP_naive": 33.0, "placebo_n": 15, "event_days": 21,
}
DRAFT = """During the 21-day diversion, ambulance arrivals at Alder Point rose by 24.4%
(95% interval 6.7% to 44.9%) relative to the Wexmoor comparison hospitals, larger than any of
the 16 placebo windows. Brookfield rose by 5.8% (-6.7% to 19.9%) and Calloway by 0.8%.
A simple before-and-after comparison would have suggested a 33% rise. The attack caused
a significant increase in arrivals at Brookfield, and waiting times rose by 6 minutes."""
CAUSAL = ("caused", "proves", "significant increase", "due to the attack")

values = set(RESULTS.values())
def traced(x):
    return any(abs(x - v) <= 0.05 or (float(x).is_integer() and abs(x - round(v)) == 0) for v in values)

problems = []
for m in re.finditer(r"-?\d+(?:\.\d+)?", DRAFT):
    x = float(m.group())
    if x == 95:                  # the interval level, declared in the methods
        continue
    if not traced(x):
        problems.append(f"untraced number {m.group()!r}")
for phrase in CAUSAL:
    if phrase in DRAFT.lower():
        problems.append(f"check wording: {phrase!r}")
print("\n".join(problems) if problems else "all numbers traced")

Expected output:

untraced number '16'
check wording: 'caused'
check wording: 'significant increase'

The check found three problems: "16 placebo windows" (there were 15; 16 is the rank denominator), "caused" and "significant increase" applied to Brookfield, whose result is inconclusive. It missed two: "waiting times rose by 6 minutes" passed only because 6 matches the rounded Brookfield percentage (5.8%) — a false match across units — and the waiting-time result is not in RESULTS at all; and Calloway's estimate is given without its interval. A script is a first pass: it catches transcription errors cheaply, but only a reader with the results-to-claims register catches a claim that is wrong in substance. EX039 asks you to improve the check.

7.11 Common errors and limits

• Writing the introduction first and letting it promise results.
• Claims that grow between results, discussion and abstract.
• "Statistically significant" as a conclusion; "not significant" as "no effect".
• Causal verbs without assumptions, or causal timidity when the design supports a qualified causal estimate.
• Numbers typed from memory; tables pasted from an earlier run.
• A limitations section that lists weaknesses without direction or resolution.
• Figures without text alternatives; colour-only encodings.
• Data statements that promise availability the agreements do not permit, or that publish sensitive facility-level figures.

The limit of good writing is that it cannot rescue a weak design; its job is to make the design's strengths and weaknesses visible so that readers can judge the claims correctly.

Exercises — chapter 7

EX039 — Improve the provenance check. Objective: use and extend Lab 7. Inputs: the lab code. Constraints: offline. Steps: (1) run the lab and record the output; (2) change the check so that each number is matched only against results with the same unit (percent or minutes) using the word that follows the number; (3) add the waiting-time result (5.9 minutes, interval 2.6 to 9.2) to RESULTS with its unit; (4) add a check that every site estimate is followed by an interval; (5) rewrite the draft so the check passes and the text is accurate. Expected evidence: code, output and corrected paragraph. Rubric: 3 = unit-aware matching catches the minutes/percent false match, the missing Calloway interval is flagged, the corrected paragraph has accurate numbers, calibrated verbs and Brookfield described as inconclusive; 2 = one check missing; 1 = only the text corrected; 0 = none.

EX040 — Directional limitations. Objective: write limitations that inform. Inputs: a synthetic study estimating the effect of a payment-processor outage on pharmacy prescription fills using weekly data from one chain. Constraints: at least five limitations in the format of §7.6. Steps: for each, state the likely direction and rough size of bias, what was done and what would resolve it; rank them by threat to the main conclusion. Expected evidence: table and ranking. Rubric: 3 = directions reasoned (for example weekly resolution dilutes a three-day effect toward zero; one chain's customers may shift to competitors, which would overstate the drop in total fills), ranking justified; 2 = one direction missing; 1 = undirected list; 0 = none.

EX041 — Structured abstract. Objective: compress without overclaiming. Inputs: Lab 6's output and §6.6.4. Constraints: 300 words; background, objective, design, setting, outcomes, results with intervals, conclusions, main limitation; synthetic status stated. Steps: (1) draft; (2) check every claim against the results-to-claims register; (3) have a peer identify the main limitation from the abstract alone. Expected evidence: abstract and register excerpt. Rubric: 3 = numbers exact with intervals, exposure correctly named as diversion, distant sites described as inconclusive, limitation explicit, synthetic status stated; 2 = one overclaim; 1 = no intervals; 0 = none.

EX042 — Reporting checklist mapping. Objective: use guidelines as checklists. Inputs: the RECORD statement and the Spillover Study's methods as described in Chapters 3, 5 and 6. Constraints: map at least eight items. Steps: (1) list eight RECORD or STROBE items relevant to the study; (2) give the section that addresses each; (3) identify any item the study does not yet address and draft the missing text. Expected evidence: mapping table and drafted text. Rubric: 3 = items correctly chosen (data sources, validation of codes, linkage, data cleaning, access to the population), sections accurate, a genuine gap found (for example how the data custodians' definitions of an ambulance arrival were validated) and addressed; 2 = mapping superficial; 1 = checklist copied without mapping; 0 = none.

EX043 — Availability statement. Objective: be open within constraints. Inputs: the sensitivity tiers of Chapter 3 and a request from a journal that "all data be made publicly available". Constraints: one paragraph plus a two-sentence reply to the editor. Steps: (1) list what can be released, at what tier; (2) write the statement; (3) write the reply explaining what is withheld and why, and offering an alternative (for example access through the custodians, or synthetic data with the same structure). Expected evidence: statement and reply. Rubric: 3 = everything releasable is released, withheld items justified by agreements, consent or targeting risk, alternative access route concrete; 2 = vague on routes; 1 = promises release of restricted data; 0 = none.

EX044 — Accessible figure. Objective: communicate a result to every reader. Inputs: Lab 6's estimates for the three sites. Constraints: a figure description (you need not draw it), caption, text alternative and plain-language summary; no colour-only encoding. Steps: (1) choose a figure type (for example estimates with interval bars against distance); (2) write the caption stating the message; (3) write the text alternative giving every value; (4) write a one-sentence plain-language summary. Expected evidence: the four texts. Rubric: 3 = figure type suits the comparison, caption states the message and the inconclusive distant estimates, text alternative complete, summary accurate for a non-specialist; 2 = one text incomplete; 1 = caption describes rather than informs; 0 = none.

Chapter synthesis

Give each part of the dissertation one job and write it in the order that keeps promises honest: methods, results, discussion, then introduction and abstract. Keep a results-to-claims register so that claims cannot grow on the way to the abstract. Report estimates with intervals and what they exclude; avoid significance verdicts; calibrate causal verbs to the design and state assumptions alongside them. Make every number traceable and check it with a script, knowing that scripts catch transcription errors but not wrong claims. Use STROBE, RECORD, SRQR and COREQ as checklists. Write limitations with direction and resolution. Make figures accessible, and state data availability honestly within the constraints of agreements, consent and infrastructure sensitivity.

Continue studying — chapter 07 of 08

← Chapter 06 · Course chapter index · Chapter 08 →

Continue to the next authored lesson.