Data Screening: A Worked Decision
The library's screening page records that no worked example exists anywhere. One now does, and its lesson is not the one you would expect: the exclusion was correct, complete, and nobody made it.
What is actually in the file
The raw data for this project is a single comma-separated file of 2,901 bytes with four columns and 204 physical rows. It belongs to one examined master's survey project supplied with its instrument, its workbook and its thesis, and everything below was read from the file and the workbook themselves.
THE RAW FILE, ROW BY ROW
| Rows | Content |
|---|---|
| 1 | Header — the four demographic question texts, each prefixed with its number and a tab character |
| 2-51 | Fifty complete responses. No blank cell anywhere in the block |
| 52-199 | 148 entirely empty rows |
| 200-204 | Five further rows, carrying some demographic values and no gender |
Verified: all fifty rows in 2-51 are complete on all four columns, and the gap between row 51 and row 200 is exactly 148 rows.
The file holds only the four demographic questions. The four substantive questions the findings rest on are not in it — they exist only inside the analysis workbook, which matters for screening because half the dataset has no raw copy to be screened against. That absence is examined at building an analysis workbook.
THE FIVE ROWS, TRANSCRIBED EXACTLY
| Row | Age | Gender | Experience | Enterprise size |
|---|---|---|---|---|
| 200 | >50 years | (empty) | >40 years | <10 |
| 201 | 18-30 years | (empty) | 0-5 years | (10-50) |
| 202 | 31-40 years | (empty) | 10-20 years | 50-100 |
| 203 | 41-50 years | (empty) | 20-40 years | 100-250 |
| 204 | (empty) | (empty) | 5-10 years | >250 |
Four rows carry an age; row 204 does not. All five carry an experience band and a size band. None carries a gender.
Three facts that can be established
Before any interpretation, here is what the artefacts prove. All three are verifiable by anyone holding the same two files.
What the files show
Each column contains each of its categories exactly once, with no repeats
The age column holds all four age bands across rows 200-203, with row 204 blank. The experience column holds all five experience bands, one per row. The size column holds all five size bands, one per row. The gender column, which has three categories, is empty in all five rows.
The workbook holds the same block, in a different order
The workbook's demographic sheet carries five rows at exactly the same positions, 200 to 204. Its size column is in the identical order to the raw file's. Its age and experience columns are in a different order — both ascending, where the raw file's are not. The two files agree about the contents of these five rows and disagree about their arrangement.
They sit outside every analysis range
All eight pivot tables in the workbook declare their source ranges. Four read A1:D51 on the demographic sheet, three read A1:H51 on the dataset sheet, and one reads E1:E51. Every range stops at row 51, and every pivot cache records fifty records.
That last point is the finding. A range boundary and a screening decision produce identical output and are completely different things. If the ranges had been drawn one row wider — a single keystroke, or a table converted to a different format — the tabulations would have changed silently and nothing in the project would have flagged it.
Three readings, and why none of them gets chosen
The source explains the five rows nowhere, so the honest position is to set out what fits the evidence and stop. The strongest reading is the first, and it is still a reading.
WHAT COULD ACCOUNT FOR THE FIVE ROWS
| Reading | What supports it | What it would require |
|---|---|---|
| (a) A category legend or scratch block — a list of the category set, typed out to drive or check counting | Fits every observed feature. Five rows in which every column holds each category once and none twice is the shape of a lookup list, not of five people's answers. The blank gender column fits: gender has three categories and would not fill five rows. So does the reordering between the two files — response data is not silently resorted, a scratch list is | Nothing that is not already observed. This library's reading, and no more than that |
| (b) Five partial responses | Cannot be excluded on the evidence | Five people each answering three or four of four compulsory questions — which the form's required setting on every question makes impossible through the form — and their answers containing no duplicate in any column, which for five draws from four, five and five categories is a coincidence rather than an impossibility |
| (c) A test or template block left in the file | Also consistent with everything observed | Nothing more than an ordinary working habit, and no way to confirm it from the artefacts |
Reading (a) is this library's, attributed here and not to the source. The source settles none of the three.
How to tell the difference in your own data
This is where the example earns its place. You will meet a block of rows you did not expect, and you will have to decide whether it is data. The supplied teaching material offers four screening checks and two prompts and no procedure of this kind; the sequence below is this library's, generalised from what settled the question here and what did not.
Seven moves on an unexplained block of rows
Establish what a legitimate response row looks like
Which questions were compulsory? Can your platform accept a partial submission at all? If every question was required, a row with blanks did not come from a respondent through the form, and you can stop treating it as one.
Look for the signature of a legend
Does every column hold each of its categories exactly once, with no repeats? Is a column with fewer categories than rows simply empty? Both are shapes that data does not naturally take, and lookup lists do.
Check the position
Is the block contiguous with the responses, or separated by a gap? A gap of 148 empty rows is not how an export stacks late responses. It is how someone types something below the data.
Compare every copy of the file
If a second copy exists, is the block in the same order? Response data is not silently resorted between copies; scratch material is.
Read the analysis ranges, not the sheet
Open every pivot table, chart series and formula and write down the range each one declares. Compare the record counts against the number you intend to publish. This is what turned an assumption into a fact here.
Look for an identifier and a timestamp
This export has neither, so no submission order and no field period can be recovered. Their absence is itself evidence about what the file is — and a reason to keep the platform's full export, not a reduced copy.
Write the decision down before you draw the range
One paragraph in the method chapter: what you found, what you concluded, what you excluded, and the count before and after. If you cannot settle the provenance, say so and report the count both ways.
Turning the evidence into a rule
What a screening record should contain
The gap this project illustrates is not an arithmetic gap. Every one of the eight tabulations was recomputed from the fifty raw rows for this extract and every one matches, category by category, totalling fifty. What is missing is the paragraph that says what was checked and what was decided — the record that planning data management exists to produce.
The screening paragraph your method chapter needs
- How many records the platform reported, and how many rows the file holds.
- What you checked for: completeness, valid categories, duplicates, outliers and out-of-range values, and anything the instrument made impossible.
- What you found, including anything you could not explain.
- What you excluded, on what rule, and the count before and after.
- Where the analysis ranges start and stop, so a reader can see the base is what you say it is.
- The order in which you did entry, screening and cleaning — the supplied material gives that order two different ways, so state yours.
Why the exclusion was right, and why that is not the point
Whatever those five rows are, they do not belong in the counts. On reading (a) they are not responses at all. On reading (b) they are incomplete records that any defensible rule would exclude from a demographic tabulation. On reading (c) they are debris. Excluding them was correct, the published counts are correct, and the base of fifty is correct.
The defect is that no one recorded the decision, because the ranges made it unnecessary to take one. That is a different failure mode from the one screening pages usually describe, and it is worth naming: the analysis boundary did the screening. It will do it silently, it will do it again next time the workbook is rebuilt, and it will do it differently if a range is dragged one row further. The same project supplies a worked check on the other end of the chain at tracing a figure back to its source.
One limit on all of this. The finding exists only because the raw file, the workbook and the thesis all survived together — the ranges could be read, the copies compared, the counts recomputed. Most published research supplies none of that, which is why this is a worked example rather than evidence about how often the pattern occurs. It is one project, and it is not a population.
What to carry forward
- A range boundary and a screening decision produce the same output and are not the same thing. Check which one you have.
- Read the declared source range and record count of every pivot table, chart and formula before you publish a base.
- A block in which every column holds each category exactly once is shaped like a lookup list, not like data — but that is a reading, and it should be labelled as one.
- Never describe unexplained rows as discarded participants. Describe what the evidence shows and stop.
- Write the screening paragraph even when nothing was excluded. The absence of a decision is what makes an exclusion unverifiable later.
Frequently asked questions
Were five participants excluded from this study?
That is one of three readings and it is not the strongest. The five rows contain each category exactly once with no repeats, carry no gender at all, sit 148 empty rows below the data, and appear in a different order in the workbook copy — a pattern that fits a category legend better than five people's answers. Every question on the form was compulsory, which makes a partial submission through the form impossible. The evidence does not settle it, and calling them excluded participants goes beyond what can be shown.
If the exclusion was correct, why does it matter that nobody decided it?
Because an undocumented exclusion cannot be checked, defended or repeated. Here the tabulations stop at row 51 because that is where the source ranges were drawn, not because a rule was applied. Drag a range one row further and the published counts change with no warning and no trace.
How would I have caught this in my own workbook?
Open each pivot table and read its declared source range and record count, then compare both against the number you intend to publish. In this workbook all eight ranges stop at row 51 and every cache records fifty records, which is what makes the exclusion demonstrable rather than assumed.
What should I write in my method chapter if I cannot explain a block of rows?
Say what you found, say you could not establish its provenance, state the rule you applied, and give the count before and after. If a finding might turn on the choice, report it both ways. An honest unresolved paragraph is stronger than a silent range boundary.
Does this mean pivot tables are a poor way to tabulate survey data?
No. The eight tabulations in this workbook were recomputed from the raw rows and all eight are exactly right. The lesson is about the declared range, which is a property of any tool that reads a block of cells — a formula, a chart series or a statistics package import will all do the same thing quietly.
Can I generalise from this example?
Only as a worked case. It is one project, and it is visible only because its raw file, workbook and thesis all survived together, which is not usual. Use it to practise the checks, not as evidence about how common the pattern is.
References and source attribution
- A complete research dataset supplied as student work: one examined master's thesis with its raw response file (four columns, 204 rows), its ten-sheet analysis workbook including eight pivot tables with declared source ranges, its survey instrument and its participant information sheet. Researcher, supervisor and institution scrubbed. Used as observed practice, not as a model answer.
- The supplied teaching source: weekly study notes and slide decks for a master's-level research methods subject in project management, including the data-preparation week that defines screening and cleaning, offers four checks and two prompts, and states no threshold. Author, institution and year not stated in the supplied files.
Suggested questions for Ask KEVOS
- How do I decide whether an unexplained block of rows in my data file is real data?
- What should a data screening paragraph in a method chapter contain?
- How do I check that my tabulations cover every case in my dataset?
- What is the difference between excluding a case and never including it in the range?
- How should I report an exclusion I cannot fully explain?