Research Integrity in Examined Work
None of what follows is misconduct. All of it is the ordinary decay of a long document assembled over months against a deadline — and almost all of it sits in the last link of the chain, where nobody thinks to look.
What this page is not saying
Start here, because the rest of the page is useless without it. Every work discussed was submitted, examined and passed. Nothing in the material shows data invented, results altered, or another person's work presented as the writer's. Those are the things integrity policies exist to catch, and none of them is on the record here.
What is on the record is smaller and more ordinary: a number retyped and not re-checked, a caveat written once and not carried forward, a reference list that stopped being maintained in parallel with the text, a hypothesis carried through a document that no instrument item could ever have tested. These are the failure modes of a long document made under time pressure by one person doing several jobs at once. They are worth studying precisely because they are ordinary.
Where the failures actually are
Ask a careful reader where a thesis is most likely to go wrong and they will usually say the data — miscoding, a mishandled spreadsheet, a tabulation nobody checked. One project in this material makes that testable, because it arrives with everything behind it: the survey form as issued, the raw response file, a ten-sheet workbook with eight pivot tables and eight charts, and the finished document.
Six reported figures were traced through every link. Five survive intact. One fails. And the failure is not where the expectation puts it.
- Instrument — sound enough to field
- Raw file — complete, no blanks
- Tabulations — all eight correct
- Charts — all correct
- Sentence — where both errors are
The practical consequence is uncomfortable. A data-integrity check that opened the spreadsheet and stopped there would have found nothing wrong with this project. The apparatus was better than the write-up, and the only place the two could be compared was the last link.
Four failure modes, and what each one costs
The recorded modes across eight works
Citations that do not resolve
Twenty distinct in-text sources across seven works cannot be reached from the documents that cite them, and in six of the seven at least one carries a load-bearing claim: the source of a focus-group size rule cited four times; three sources on which a central finding rests, one of them cited six times; the citation carrying one move of a gap argument. The reader loses the source, not merely the tidiness.
Quantities that contradict
One questionnaire item reported at 87% and at 99% one page apart, the second offered as support for the first. One headline quantity given at 75%, 42% and roughly 40% in three consecutive chapters. One survey reported twice with eleven disagreeing values, one pair of which reverses the finding. A lessons-learned register counted at 162 and at 161 two pages apart.
A variable never measured
Two hypotheses about entrepreneurial success, a conceptual framework that leaves that variable in an unconnected box, and an instrument with no item measuring it in any form — not revenue, not growth, not survival, not self-assessment. Both hypothesis failures in that work trace to this single root, and it is visible in two separate artefacts.
A statistic set up and never reported
Correlation introduced, the formula printed as a figure, a five-line interpretation key printed, a scoring scheme constructed and stated exactly, two plots produced, strength asserted twice — and no value of the coefficient anywhere in the document. Both stated results also carry the wrong sign against the work's own printed key.
Each of the four breaks the chain between evidence and claim rather than the evidence itself, which is what makes them findable by a reader and fixable by a writer without touching the data.
The uncited reference is not the interesting number
Reference-list health is usually reported as one figure: what proportion of listed entries are never cited. Across these works that figure runs from zero to 45.7%, and the spread is mostly telling you what kind of document you are holding rather than how carefully it was made.
WHAT AN UNCITED-REFERENCE RATE MEASURES, BY DESIGN (findings of these documents; the metric should not be reported as one number across them)
| Work | Rate | What the uncited entries actually are |
|---|---|---|
| A hybrid Delphi study | 5.6% | One definition looked up and not used. Tidiness. |
| A two-project case study | 17.4% | Four of the entries are fossils of a deleted paragraph still sitting inside the file, invisible in the printed document. Drafting residue. |
| A systematic quantitative review | 45.7% | Most plausibly a review corpus that was listed and not quoted. A property of the genre, not a fault — and the work never marks which papers were retained, so the set cannot be recovered. |
| A focus-group study | 0% | Every listed entry resolves to an in-text citation — and four in-text sources carrying the study's method are absent from the list. A perfectly maintained list can still be the wrong list. |
Why these findings exist at all
Almost nothing on this page could have been established from the printed pages alone. Four different kinds of survival made it possible, and only one of them was deliberate.
WHAT SURVIVED, AND WHAT IT MADE VISIBLE
| What survived | Deliberate? | What it made checkable |
|---|---|---|
| A complete raw dataset printed in an appendix — twenty questions by sixteen respondents | Yes | Eighteen of the work's own claims about it, of which twelve are fully correct and four are wrong; and three mis-reported modes diagnosed as a spreadsheet's tie-breaking rule |
| Chart data left inside live chart objects, and one embedded worksheet | No — an artefact of the file format | A headline safety claim shown to match none of six candidate computations, and a cost section shown to disagree with itself four ways |
| A deleted section still present inside a table in the file | No | Four apparently uncited references explained, and a missing contents entry accounted for |
| A complete apparatus: instrument, consent sheet, raw file, workbook and schedule | Yes — supplied with the work | Six figures traced end to end; a dependent variable shown to be unmeasured; a schedule compared against the artefacts it planned |
Five checks that catch nearly all of it
Run these in the last fortnight, in this order
The transcription check
Take every number in your results, discussion and conclusions and walk it back to the tabulation that produced it. Read the tabulation, not your memory of it. This is the check that would have caught both failures in the one fully traceable project, and it is the only check that catches an error which has already propagated into a second chapter.
The one-quantity check
For each headline quantity, search the whole document for it and confirm that every occurrence gives the same value on the same base. Where a figure appears in a results chapter, a summary and a discussion, it is being retyped twice, and the works here disagree with themselves at exactly those seams.
The citation reconciliation, both ways
List to text: every entry should resolve to at least one citation. Text to list: every in-text citation should resolve to an entry. Do the second one first — it is the direction that costs a reader a source, and it is the direction almost nobody runs.
The measurement check
For every variable named in a hypothesis, a research question or a conclusion, name the instrument item that measured it. If you cannot point at an item, you have found the failure that produced Mode 3 above, and you have found it while you can still write around it.
The apparatus check
Search for the name of every statistical technique you mention. Each one should have a value, an n and, where your design supports it, a measure of uncertainty within a paragraph or two. Apparatus without an output is either an unreported analysis or a leftover from a plan.
What integrity looks like when it is working
The same eight documents contain several passages that do the opposite of everything above, and they deserve naming as precisely as the defects do.
- A source audit performed rather than asserted. One work takes a widely quoted failure rate to the four sources most often given for it, reads each in the original, quotes what each actually says, names the transformation in each case — an observation becoming a statistic, an explicitly unscientific estimate becoming a point estimate — and enumerates what a substantiated claim would have carried.
- A validity threat named, the data kept, and the use bounded. "These results have not been omitted from analysing … however the data will not be a single source to draw conclusion." That sentence states the threat, states the decision, and states the mitigation.
- Disconfirmations of the researcher's own earlier round, reported three times. A second-round survey overturns three findings from the first-round interviews, and the work says so rather than quietly dropping them.
- A confounding caveat carried through three chapters, each time naming a specific alternative cause, rather than stated once beside the finding and forgotten.
- A failed data-collection attempt recorded in the method chapter — a first focus-group sitting abandoned at two attendees — and an instrument abandoned mid-session reported three separate times, with the reason.
- A scope reduction reported in the introduction with the discarded items named: five models intended, three analysed.
- A withdrawal clause that is honest about its own limits — it grants withdrawal during completion, refuses it after submission, and gives the reason, which is that anonymous collection makes retrieval impossible.
Every one is a writer making their own work harder to believe uncritically. That is what integrity looks like in practice, and it is cheaper than any defect above.
The line between a defect and misconduct
How to classify what you find in your own draft
For the citation audit in detail, see citation health in examined work. For the contradiction audit, when a thesis disagrees with itself. For the transcription check run link by link, tracing a figure back to its source. For claims that travel past their evidence, overreach in conclusions. For where the formal integrity obligations sit, plagiarism and research integrity and unethical behaviour in research.
What to carry forward
- In the one project traceable end to end, the instrument, the raw file, the tabulations and the charts are all sound, and every error is a writing-up error. Check the last link hardest.
- Four modes account for nearly everything recorded here: citations that do not resolve, quantities that contradict, a variable never measured, and a statistic set up and never reported.
- Run the citation reconciliation from text to list, not only list to text. That is the direction that costs a reader the source.
- For every variable in a hypothesis or conclusion, name the item that measured it. If you cannot, you have found a design defect while you can still write around it.
- Defect counts are counts of what could be checked. The works that show the most are the works that can be criticised the most, and that is an argument for disclosure rather than against it.
- None of this is misconduct. It is what a long document does under deadline — which is exactly why it has to be scheduled for, not intended.
Frequently asked questions
Is a mistyped percentage in a thesis a research integrity issue?
It is a defect in the document rather than misconduct, and the distinction matters. Misconduct involves invented, altered or suppressed data. A number retyped wrongly from a correct tabulation is a transcription error — real, worth catching, and categorically different. Where it becomes serious is when it propagates into a derived total or a conclusion, which is what happened in the traced project here.
Which check finds the most defects for the least effort?
Tracing every number in your discussion and conclusions back to its tabulation. It catches transcription errors, propagated errors, base changes and figures that exist nowhere else in the document. On a document of this size it is an afternoon, and it can be done by someone who is not the author.
Why does publishing more raw material make my work look worse?
It does not make it worse; it makes it checkable. Every finding on this page exists because an instrument, a dataset or a whole apparatus survived. Works that publish less show fewer defects without having fewer. A reader who understands this will read a displayed dataset as a sign of confidence, which is what it is.
How do I audit my reference list quickly?
Two passes, mechanically. First, extract every in-text citation as author plus year and confirm each has an entry — this is the pass that matters most, and it is the one usually skipped. Second, take each listed entry and confirm the surname appears somewhere in the body. Anything unresolved in either direction is a fix, not a judgement call.
What should I do if I find a hypothesis I never actually tested?
Say so, and report what your descriptive evidence supports instead. A sentence stating that no inferential test was performed costs you nothing and is far stronger than a claimed rejection a reader can see was never computed. The failure to avoid is carrying the hypothesis silently into the conclusions.
Do these findings mean examined theses are unreliable?
No, and the material does not support that reading. Eight works are not a population, all eight passed examination, and most of the defects recorded are visible only because those particular works published enough to be checked. Read them as worked examples of where documents fail, and use them to build your own checking routine.
References and source attribution
- Eight examined master's works in project management, supplied as student work and used here as observed practice rather than as model answers. Researchers, supervisors, institutions, employers, jurisdictions, industries and communities scrubbed. All eight were examined and passed.
- The complete research apparatus of one of those projects — survey instrument as issued, participant information sheet, raw response file, ten-sheet analysis workbook and project schedule — from which six reported figures were traced from question to sentence and five were confirmed intact.
- A complete printed raw dataset from a second project, twenty questions by sixteen respondents, recomputed mean by mean and mode by mode for this library; and chart values recovered from the chart objects and embedded worksheet of a third.
- The reference lists of seven of the works, audited entry by entry against every in-text citation in both directions, recording twenty in-text sources that cannot be reached and forty-one further resolution defects.
- The supplied teaching source: weekly study notes, slide decks and assessment activities for a master's-level research methods subject in project management, including its plagiarism and unethical-behaviour material. Author, institution and year not stated in the supplied files.
Suggested questions for Ask KEVOS
- Where do errors actually occur in a research document — the data or the write-up?
- How do I audit my reference list in both directions before submission?
- Is a contradictory figure in two chapters a research integrity problem?
- How do I check that every variable in my hypotheses was actually measured?
- What checking tasks should I put in my thesis schedule and how long do they take?
- Why do theses that publish their raw data appear to contain more errors?