When a Thesis Disagrees With Itself
Both sets of numbers are printed in the same document, a few pages apart, and nothing in either work acknowledges that they differ. Here is how one survey becomes two incompatible accounts, and the twenty-minute check that catches it before an examiner does.
One survey, two accounts, eleven disagreements
One work reports a 46-response survey twice: once in a Survey Results sub-section that goes figure by figure, and once in a Summary of Findings a few pages later. A third account appears in the Discussion and agrees sometimes with one and sometimes with the other. All three sit inside one 10,783-word document, and the work nowhere acknowledges that the accounts differ. It is one of the works mapped at Eight Examined Theses Compared.
THE SAME SURVEY, REPORTED TWICE
| Quantity | Survey Results | Summary of Findings | Discussion |
|---|---|---|---|
| Has created a risk management plan | 87% | 86% | — |
| Risk management always part of the plan | 54% | 56% | "only half" |
| Risk management sometimes part of the plan | 44% | 42% | — |
| Starts at project initiation | 17% | 21% | 21% |
| Starts during planning | 63% | 60% | "the majority" |
| Starts after planning, before the event | 18% | 16.6% | "a quarter" |
| Managed during the event | 2% | omitted | — |
| Plan is reviewed after the event | 11% | 29% | — |
| Plan is not reviewed after the event | 28% | 12.5% | 12.5% |
| Plan is sometimes reviewed | 61% | 58% | 58% |
| Training received on the job | 57% | 59% | "mostly on the job" |
| More training would help | 18% | 29% | "the number one request" |
| Events organised per year | 46% organise 1–9 | 47% organise 1–20 | — |
Every figure above is printed in the document. None is corrected anywhere. These are one study's own numbers and describe one survey of 46 people.
Some of the gaps are trivial in size and one is not a data difference at all: 87% of 46 is 40.0, and 40 of 46 is 86.96%, so "87%" and "86%" are two roundings of the same count. That is the useful reference case, because it shows what an innocent disagreement looks like — and by contrast, what the others are.
The one that reverses the finding
Post-event review of the risk plan is this work's own recommendation area: its conclusions recommend "reviewing the plan after the event to consider what can be done better next time". It is reported twice.
Survey Results
- Reviewed: 11%
- Not reviewed: 28%
- Sometimes: 61%
- → Not-reviewed outnumbers always-reviewed by more than two to one.
Summary of Findings
- Always reviewed: 29%
- Not at all: 12.5%
- Sometimes: 58%
- → Always-reviewed outnumbers not-reviewed by more than two to one.
Two accounts of one survey item report the opposite finding. The Discussion then quotes the second set — "12.5% admit to not reviewing … and 58% sometimes reviewing" — and draws a conclusion consistent with the first: that risk "continues to be treated as a one-time activity". The conclusion is right; the numbers printed under it are the ones that contradict it.
One quantity, three values
A second work has the same problem in a different shape: not two accounts of everything, but one headline quantity stated three ways. Its central empirical question is how many construction workers have experienced pressure to work unsafely. For the conventions this section is measured against, see Writing the Results Section.
The first two figures are a paragraph apart in the same section. None of the three can be checked against the appendix the work cites for one of them — the appendix prints 75 numbered rows against a reported sample of 100, and 46 of those rows are a bare "No", so the item was not one only affirmative respondents completed. Counting the rows gives 29 of 75, or 38.7%, which rounds to the Discussion's "roughly 40%" against a denominator of 75 rather than 100. That count is this library's, not the work's, and the coincidence of "75 rows" and "75%" is noted and must not be presented as an explanation.
How it happens
Nothing in either case is fabrication, and nothing in either case is arithmetic incompetence. Both are the signature of a result written twice, at different times, from different working notes — and the corpus shows five specific mechanisms doing the damage.
Five mechanisms, all visible in these documents
Two passes, two sources
A results section written from the charts and a summary written later from notes, or from a different version of the tabulation. Every one of the eleven disagreements is one or two respondents wide, which is exactly what a re-derivation from a slightly different working file produces.
No absolute counts printed
Every result in that work is a percentage of 46. At that sample size each respondent is 2.17 percentage points, so 43%, 44% and 46% can be the same twenty people. Printing counts alongside percentages makes both accounts checkable in one glance and makes drift impossible to miss.
Two bases in one sentence
Percentages of the whole sample and percentages of a sub-group reported side by side without saying so. The clearest case: "Those who selected 'other', 62% stated that nothing gets in the way" — 62% of an unstated number, printed beside four percentages of the whole sample.
An undeclared repair of the instrument
The survey offered overlapping bands — 1-10, 10-20, 20-30 — so a respondent organising exactly ten events could choose either. The results are then reported with non-overlapping bands, 1-9 and 10-19, which are not the categories anyone was offered. The repair is sensible and undeclared, and a third band, "1-20 at 47%", exists in neither the instrument nor the results.
Ranking on one statistic, narrating another
In a third work, the survey items are ranked by mode and described by mean, and the rule is never stated. Three separate claims — an "equal ninth", a sub-ranking among mode-5 items, and a "three worst" selection — are internally consistent with a mode-first ordering and inconsistent with the means printed beside them. One sentence declaring the ordering rule would make all three correct.
The same failure in three more shapes
Two accounts of one survey is the cleanest instance, but the corpus supplies three more that a reader should recognise.
- Four figures in one sentence, two of them contradicting the chart beneath it. One work's headline commercial result is a project saving of "$6m", stated in a sentence that also gives a $2m variance the chart puts at $4m, a $24m budget the chart puts at $260m, and a $23.5m actual the chart shows as identical to its projected value. The two cost charts are drawn on different scales — raw dollars and thousands of dollars — and neither says so, so two adjacent figures comparing the same quantity cannot be read against one another as printed.
- A count of one register given as 162, then 161, then 161 again, two pages apart, described first as covering both positives and improvements and then as improvements only. If both descriptions are right the register holds 161 improvements and one positive, which contradicts the same work's own list of four named positive learnings. Neither count can be checked, because the register itself is a set of unreadable images.
- A breakdown that does not close. One work reports 17 participants and then 13 in one unit and 5 in another. The very next paragraph explains an over-count on a different item — 24 role responses from 17 people, because participants held more than one role — and the same explanation is not applied one paragraph earlier. In the same section, a reported 42% is not attainable from a base of 17 at all: 7 of 17 is 41.2% and 8 of 17 is 47.1%. Nine of the ten reported figures in that section do reconcile to 17. The defect is the missing sentence, not the arithmetic.
The check that catches it
The reconciliation table is the whole method. One row per reported quantity, one column per place in your document where it appears, filled in from the document rather than from your working file. On a small survey it takes about twenty minutes, and it catches every failure on this page. Presenting Data in Tables covers how the surviving figures should then be laid out.
Seven passes over your own results
Build the reconciliation table
List every quantity you report. Add a column for each location it appears in — results, summary of findings, discussion, conclusions, abstract, figure captions. Fill it in by reading the document, not by re-deriving from your data. Anything that differs across a row is either an error or a rounding you must justify.
Print counts beside percentages
"28 of 46 (61%)" cannot drift the way "61%" can, and it lets a reader check your arithmetic without your file. Neither of the two works on this page prints a single absolute count.
Check that every percentage is attainable from your n
With 17 respondents the attainable values are 0, 5.9, 11.8, 17.6, 23.5, 29.4, 35.3, 41.2, 47.1 and so on. A figure that is not on that list did not come from that denominator, and finding out why is always worth the five minutes.
Sum every set to 100 and to n
A set that sums to 97.6 has a category missing. A set that sums to 99.5 has a rounding you should state. A set whose members are not integer counts of your sample came from somewhere else.
Name the base in every sentence
If a percentage is of a sub-group, say which sub-group and how many people it contains. Two bases in one sentence with no labels is the most common way a correct number becomes an incorrect statement.
Declare any repair to your instrument
If you merged, split or re-drew response bands after collection, say so and say why. An undeclared repair means your reported categories are not the ones anyone answered.
State your ordering rule before you rank anything
If you rank on mode and narrate on mean — or on any two different statistics — one sentence saying so makes the whole passage correct. Without it a reader checking your ranks against your own printed means will find them wrong.
What to carry forward
- Report each quantity once, in one place, and cross-reference everywhere else. Every contradiction on this page comes from writing the same result twice.
- Print absolute counts beside percentages. At n = 46 each respondent is 2.17 points, and a percentage alone cannot be checked or defended.
- Build a reconciliation table before you submit: one row per quantity, one column per location, filled in from the document rather than from your data.
- Check that every percentage you print is attainable from your denominator, and that every set sums both to 100 and to your sample.
- State the rule you ranked on, and declare any repair you made to your instrument after collection. One sentence in each case removes the contradiction entirely.
Frequently asked questions
Are these errors of fabrication?
No, and nothing on this page suggests it. Every one is the signature of a result written twice from different working notes: disagreements one or two respondents wide, percentages of different bases in one sentence, undeclared repairs to response bands, and a ranking rule that was applied but never stated. All the documents concerned were examined and passed.
Which set of numbers should a reader trust?
In the case set out here, only one of the two sets reconciles to the stated sample of 46 — the three values sum to 46 people and to 100%, and the other set contains a percentage that is not an integer count of 46. That is an inference from arithmetic, drawn by this library. The work itself never says the accounts differ, let alone which is authoritative.
How do I avoid reporting the same result twice?
Report each quantity once, in the section where it belongs, and cross-reference it everywhere else rather than restating it. Where a summary section is required, build it by copying figures from the results section rather than by re-deriving them from your data file — re-derivation is where the drift comes from.
Should I print counts as well as percentages?
Yes, at any sample size where a single respondent moves the percentage noticeably. At 46 respondents each person is 2.17 percentage points, so 43%, 44% and 46% can describe the same twenty people. Neither of the two works on this page prints a single absolute count, which is why their disagreements cannot be resolved from the documents themselves.
What if I need to change my response categories after collecting data?
You can, and one work here plainly did — its overlapping bands were replaced with clean ones in the reporting. The defect is that the repair is undeclared, so the reported categories are not the ones respondents were offered. Say what you changed and why, in the results section, in one sentence.
Would an examiner catch this?
The documents on this page were examined and passed with both sets of figures printed, so the honest answer is that you cannot rely on it. That is the practical argument for the reconciliation table: it costs about twenty minutes, and it is the only stage of the process where all your figures sit in one place and can be compared.
References and source attribution
- Examined master's works in project management, supplied as student work, whose reported quantities were tabulated location by location and recomputed against the stated sample sizes for this library's record. Researchers, supervisors, institutions, jurisdictions, employers and programmes scrubbed. Used as observed practice, not as model answers.
- The consolidated extract of four of those works, which records eleven numeric disagreements between two accounts of one 46-response survey, the arithmetic checks that reconcile one account and not the other, and a headline quantity stated at three values in three consecutive chapters.
- The consolidated extract of the three works examined afterwards, which records a cost section disagreeing with its own charts in four places, a register counted at 162 and 161 two pages apart, a demographic breakdown of 18 against a stated 17, and a survey ranked on mode and narrated on mean.
- The supplied teaching source: weekly study notes, slide decks and assessment activities for a master's-level research methods subject in project management, which sets no rule on restating results, on reporting counts alongside percentages, or on declaring changes to response categories. Author, institution and year not stated in the supplied files.
Suggested questions for Ask KEVOS
- How do I build a reconciliation table for my results?
- Should I report percentages, counts, or both?
- How do I check that a percentage is possible given my sample size?
- What do I do if my results and my summary of findings disagree?
- How do I declare a change to my response categories after collecting data?
- How should I state the rule I used to rank survey items?