Case Study Design in Practice
A design that is right and a label that is wrong, in a study whose delivered file kept the numbers behind its charts. That is why three of its claims can be confirmed, why its safety headline cannot, and why a section missing from the body turns up in a stray table.
A controlled two-case comparison, named four different ways
One examined master's work — called Work L here — compares two near-identical repeat shutdown maintenance projects on one item of fixed plant, asking whether the lessons-learned framework in place is robust, effective and correctly used. The comparison runs on five deliverables: time, cost, quality, scope and safety.
ONE DESIGN, FOUR NAMES
| Where in the document | What the design is called |
|---|---|
| Abstract | "research via interviewing key stakeholders and key deliverable analysis"; "a combined approach" |
| Method chapter, first sentence | "a qualitative case study analysis of two extremely similar, typical" projects |
| Research design section | "a case study-based approach to data gathering and analysis" |
| Data collection section, last paragraph | "A mixed method approach confirmed the validity of the research" |
Seven paragraphs separate the second and fourth of these, and the thesis reconciles neither.
The method chapter goes further than mislabelling. It argues against a quantitative approach — data could be skewed by interdependencies and would not isolate the framework's effect — and the results chapter then reports safety counts, budget-against-actual costs, planned-against-actual durations and rework counts throughout.
Recruitment, bias controls and the instrument
Recruitment is the best-documented in its batch and it takes two sentences. The eligible population is defined by eight named stakeholder role categories; the sampling frame is named — the employer's own organisational databases; and the approach channel is named — an introductory email issued under the awarding institution's protocol, stipulating confidentiality and anonymity.
Three controls, each with a stated reason
Departmental separation
The study ran inside the researcher's own employer but in a different department, and participants had no direct link with the researcher's role. The reason is stated.
Interview timing
Interviews were held only after the second project completed, so stakeholders could not adjust conduct to satisfy an observation in progress — a Hawthorne control, named without the term.
No prior knowledge
Participants got neither the questions in advance nor any results during fieldwork, and both facts are stated twice. Nothing in the teaching material teaches a blinding procedure.
This is the strongest insider-research situation in this library and there is no positionality statement in it. What there is instead is two operational controls — not the same thing, and arguably more useful. Work L should be read as the worked case of insider research managed by procedure rather than by declaration. That framing is the library's.
The interview schedule is displayed: eleven questions behind a four-field respondent header, by phone, semi-formal, about thirty minutes each. It contains one good facilitation rule — respondents were "allowed to digress until starting to divert onto different (irrelevant) tangents" — and three faults. Items 3, 6 and 8 begin "If yes" with no branch instruction for a no; item 6's condition points at an item that is not a yes/no question; item 5 asks four questions in one. The header also collects names, in a study that promises anonymity four times. See interview design in project research.
Charts that carry their own data
Ten of this work's twelve figures are live charts whose plotted values survive inside the delivered file, and one embedded worksheet retains a source table. Nothing below is read off a picture — the values come from the chart objects' own stored series. That capability does not exist for most works in this library.
THREE CLAIMS THAT RECOMPUTE EXACTLY
| Figure | Values recovered from the chart | The thesis's claim | Check |
|---|---|---|---|
| Respondent experience | 2 / 3 / 3 / 2 — the chart's own title states the sum is 10 | "50% possessing at least 10 years industry experience" | Verified. 3 + 2 = 5 of 10 |
| Involvement in previous reviews | 8 and 2 — the chart's own labels read 80% and 20% | "a high portion (80%) … had direct involvement" | Verified. 8 + 2 = 10 |
| Rework items | 138 and 92 | "a significant reduction of rework … (33%)" | Verified. (138 − 92) ÷ 138 = 33.3% |
Recomputed for this library from the file's own chart data. One study's figures, not benchmarks.
Two of those figures independently confirm the sample size of ten, which the prose states three times. That is the only sample size in this library confirmed by two separate artefacts inside the document. Two other charts are less fortunate: one has no plotted values at all — and the thesis still draws a two-part finding from it — and one carries five response values that sum to eleven against a sample of ten.
The safety claim, and what the numbers actually show
The embedded worksheet retains the source table behind the two safety figures: hazards and events for each project at four severity levels, and the plotted series reconcile exactly against it. The thesis's claim is that the most notable difference was a 60% reduction in the highest-severity safety events.
EVERY CANDIDATE COMPUTATION FROM THE STUDY'S OWN DATA
| Quantity | Project 1 | Project 2 | Change |
|---|---|---|---|
| Highest-severity events | 4 | 2 | 50.0% reduction — the nearest candidate to the stated figure |
| Highest-severity hazards | 7 | 6 | 14.3% reduction |
| Highest-severity, hazards and events combined | 11 | 8 | 27.3% reduction |
| All events, all severities | 11 | 13 | an 18.2% increase |
| All hazards, all severities | 15 | 18 | a 20.0% increase |
| All items, all severities | 26 | 31 | a 19.2% increase |
Recomputed for this library. No computation yields 60%, and the nearest is the narrowest of the six.
Cost and time: three figures for one quantity
The cost section carries the study's largest commercial claim and its least support. The chart for the first project names a budget of $26m against an actual of $30m — a $4m overrun. The chart for the second carries a single unnamed series of two identical values on an axis labelled in thousands, which reads as $260m projected and $260m actual.
The second project reaped the benefits with a project saving of $6m dollars compared to the first project ($2m vs actual planned) and eventuated slightly delivered slightly under budget at $23.5m vs $24m projected budget…
- Project 1's overrun. The chart says $4m. The prose says $2m.
- Project 2's budget. The chart says $260m on its own axis units. The prose says $24m — a factor of about eleven.
- Project 2's actual. The chart shows projected and actual as equal; the prose says they differ and the project came in under budget.
- The saving. The stated $6m is consistent with the chart's Project 1 actual and the prose's Project 2 budget, and inconsistent with the prose's own Project 2 actual.
- The two charts use different scales — one raw dollars, one thousands — and neither says so.
The time finding runs the other way. Its section contains 203 words and no figure, no duration and no percentage: the entire quantitative content sits in a chart whose series are mislabelled. From the file, the second project's planned and actual durations are both 660 hours, confirming the prose's claim of "exact duration accuracy" — the study's only correct-by-construction quantitative claim is the one it declines to put a number on. The first project's two values are 475 and 660, and which is planned and which is actual cannot be determined from the file.
The missing section is an orphaned draft
The contents page is hand-typed, twenty-five lines with dot leaders, and it disagrees with the body in nine places — capitalisation, missing section numbers, appendix titles that differ from the appendices, a chapter numbered 6 with no chapter 5 before it. Eight of the nine are drift. The ninth is a numbered section listed in the contents that the document does not contain.
It does contain it. The heading exists once in the body, inside a single two-cell table that sits outside the reading order, holding 485 words: a repeated sentence from the research background, the missing heading, four paragraphs of an earlier literature review, and a ten-item list of knowledge areas printed with nine bullets.
What the orphan contains that the thesis does not
The only direct disagreement with a cited author
The draft weighs one author's claim against two others and concludes that the notion "is disproved". Nothing in the printed thesis disagrees with any source.
The researcher's own career
A line citing over ten years' experience in the field. It is a career figure and never this study's n — and in the printed thesis it does not appear at all.
Cited nowhere the reader can see
All four entries an automated uncited-reference check flags are accounted for by this cut block — they are cited inside the orphaned table and nowhere else.
The mechanism behind all of this is one decision. The document carries no heading styles at all — the twenty-four bold pseudo-headings are formatting, not structure. No generated contents field was possible, so the contents page had to be typed, so it drifted; and a block of cut text inside a table stayed invisible to every navigation tool the writer had.
What to take into your own case study
Design, evidence and delivery
- Justify your cases by what makes them comparable — near-identity is the argument for a controlled comparison
- Name the design once and use that name everywhere else
- If you collect both counts and accounts, say so rather than arguing against one of them
- Name the population by role, the sampling frame and the approach channel — this study does all three in two sentences
- State your invitation count so a response rate exists (library addition)
- Put branch instructions on every conditional interview item, and check each condition points at the right item
- Print the counts behind every chart in your text, not only in the picture
- Draw two charts of one quantity on the same scale, or say in the caption that you have not
- Compute your headline percentage from your own table, and state which quantities it is over
- Distinguish leading from lagging measures — a rise in reported hazards and a rise in incidents mean opposite things
- Apply heading styles, so your contents page is generated and cut text cannot hide
What to carry forward
- Two near-identical cases is a real design. Justify it by the near-identity, and concede the construction differences your own results reveal.
- Charts that carry their data make a study auditable. Everything verified here — and everything falsified — follows from that.
- Compute the headline number from your own table. Six candidate computations exist here and none is the one printed.
- A rise in recorded hazards may be better reporting, not worse safety. Apply the reading consistently.
- Name your design once. Called qualitative in one section and mixed method in another, it defends neither strand.
- An uncited reference can be the fossil of a deleted paragraph. Here, four of four were.
Frequently asked questions
Is comparing two projects a valid research design?
It can be, if the comparability is argued rather than assumed. This study justifies its pair by their near-identity, which is the correct argument for a controlled comparison. Its own results then concede construction differences between the two, which is the assumption being contradicted in the findings — so the design is sound and the caveat has to travel with every number.
Why does it matter whether the charts carry their data?
Because it is the difference between a claim and a checkable claim. Ten of this work's twelve figures kept their plotted values in the delivered file, which is why three percentage claims could be recomputed and confirmed, and why the headline safety claim could be shown to match no computation from the same data. Deliver the same figures as flat images and none of that is available.
How should I report safety performance across two projects?
Separate leading measures from lagging ones and report both totals. In this study, high-severity events fell while total recorded items rose from 26 to 31, and the rise is concentrated in the lowest severity level, which the thesis never mentions. Either reading can be defended; what cannot be defended is claiming improvement without printing the totals that move the other way.
What is an orphaned draft and why should I care?
It is a block of superseded text still sitting in the delivered file, usually somewhere the writer's navigation could not see — here, inside a two-cell table outside the reading order. It carried 485 words, the study's only direct disagreement with a cited author, and citations to four references that appear uncited in the printed document. Search your own file for text outside the flow before you submit.
Do I need a positionality statement if I study my own employer?
This work has none and instead has two operational controls: a different department from the researcher's own role, and interviews held only after the work completed. That is insider research managed by procedure rather than declaration, and it is arguably more useful than a statement. What it never does is name the insider position as a limitation, and a reader has to assemble that themselves.
Can a mixed method approach confirm the validity of my research?
No — a design cannot confirm its own validity, and this thesis says it did. What it actually describes is triangulation across two data sources, the word for which never appears in the document. Say which two sources you are comparing, what agreement between them would mean, and what disagreement would mean.
References and source attribution
- An examined master's work supplied as student work: a two-case comparison of repeat shutdown maintenance projects assessed on time, cost, quality, scope and safety, combining ten stakeholder interviews with project performance metrics, with its interview schedule displayed as an appendix. Researcher, supervisor, institution, employer, place and operation scrubbed. Used as observed practice, not as a model answer.
- The delivered file's own chart objects and embedded worksheet, from which the experience distribution, the review-involvement counts, the rework counts, the four-level safety table and the cost and duration series were recovered and recomputed for this library. No value is read from the appearance of a figure.
- A 2015 methods text cited by the thesis for three stated benefits of case study as a design, and a cited model of accident causation used to support its confounding caveat. Both are named in the thesis; neither is reproduced here.
- The supplied teaching source: weekly study notes, slide decks and assessment activities for a master's-level research methods subject in project management, in which case study is named repeatedly and justified nowhere. Author, institution and year not stated in the supplied files.
Suggested questions for Ask KEVOS
- How do I justify selecting two cases for a comparison study?
- What bias controls can I use when researching inside my own organisation?
- How do I check that a percentage in my results matches my own data?
- What is the difference between a leading and a lagging safety indicator?
- Why do uncited references appear in a reference list, and what should I do about them?
- How should I label a design that uses both interviews and project metrics?