Missed prediction
P1 missed once
- Predicted: 0,3 or less
- Measured: 0,4
Eerste monteur service W · bbl · months to a first job
The service W route took 0,4 months to a first job. Reported as a miss, not rounded away.
Krapte per vak
Is it me, or the market?
Why do employers in techniek and zorg get so few applicants? I built a study from official data, retracted my own first headline after a validity audit, and confirmed the finding with a preregistered test: the starter you're looking for already has a boss.
- Official sources
- 10
- Headline retracted
- 1
- Preregistered test
- 9 of 9
- Research reports
- 6
DUO · diploma year 2025
Already had a boss on diploma day
New electricians 9 of 10
New designers 0 of 10
Confirmed with a preregistered test · 9 of 9 routes pass
In short
One headline, three versions
Each version keeps its colour further down the page, so you can jump to the part you need.
v1 · Sun 08:36 · Draft headline
3,8
new designers per starter job, from my own count of 153 ads. It agreed with my hunch.
- 153 ads
- 1 day
- 1 coder
How the number was made
v2 · Sun 12:16 · Retracted
3,8
Counted honestly, design and techniek overlap. I took the headline off the page.
Strict starter · 90% interval · overlap in yellow
v3 · Sun 18:40 · Published
9 of 9
routes in techniek and zorg start faster than design, in a test I froze before looking.
9 pass · 1 route not published by ROA
Part 1 of 3 · v1 Draft
Where did the first headline come from?
Own sampleShort answer
From a hunch. UWV doesn't measure designers, so I counted 153 job ads myself. The count agreed with me.
01 · The question
Is it me, or the market?
Half 1 · the market
Is it the market?
Answered with official data per trade
- RQ1
- RQ2
- RQ3
Half 2 · their own hiring
What should I change?
Answered with job ads and published research
- RQ4
- RQ5
In the news
The market is cooling, on average. Per trade, it isn't.
Q2 2026 95
2022 142
Open vacancies per person on WW, by job
UWV tension index, Q2 2026 · capped at 16
- Plumbers
- 14,7
- Machine fitters
- 10,6
- Electricians
- 9,3
- Care assistants
- 8,9
- Marketing advisers
- 2,4
The argument to test
My own hypothesis, not a finding- Big brand Gets a queue of applicants
- Big brand Can afford slow, picky hiring
- Small firm In a shortage trade, copies those habits
- Result Loses applicants
The page exists to earn trust, so readers ask for the free Afhaakcheck. One overclaim a reader sees through costs more than a weaker headline.
Research questions
- RQ1 "Is it my trade, or is it me?" Answered
- RQ2 "How hard is it for starters in my trade?" Answered · test H1
- RQ3 "Who fills my vacancies, then?" Answered
- RQ4 "Do we ask for too much experience?" Running until 7 Dec
- RQ5 "Does a long application cost me people?" Partly answered
02 · Measures
Measures, and why I chose them
Four checks for every measure. A rejected measure failed where the orange mark is.
| Measure | Official source UWV, CBS, DUO or ROA | Per job One figure per job | Comparable Same method, every job | Direct Measures the thing itself |
|---|---|---|---|---|
| Used · 6 | ||||
| Shortage per job UWV tension index Same formula for 93 job groups. WW only, so read the order. | ||||
| How hard the start is ROA school-leaver survey Measures the outcome itself, 6 cohorts pooled. | ||||
| Who fills the jobs bbl share of mbo diplomas (DUO) Shows who already had an employer before the diploma. | ||||
| Supply against demand Diplomas (DUO, CBS) against new vacancies (UWV) My own request. Diplomas aren't supply, and vacancies aren't entry jobs. | ||||
| The national picture CBS vacancies per 100 unemployed What readers hear in the news. One average for all trades. | – | |||
| Experience asked Own job-ad study (H4) No public source splits experience per job. Running until 7 Dec. | ||||
| Rejected · 4 | ||||
| Supply against demand "Graduates per starter vacancy" A ratio of three uncertain parts, from ads on different sites. | ||||
| Vacancies before 2023 The pre-2023 vacancy estimate A constant share of the sector, applied backwards. It looked like data. | – | |||
| Vacancy rate Vacancies per 100 workers A fair rate, but its job set and denominators didn't match. | ||||
| Popular figures Two figures from the news Only second-hand, or from a vendor case study. Neither was used. | – | – | – | |
- Yes
- Partly
- No
- Why it was rejected
- – Not relevant
03 · Exploration
Three rounds in one night
The brief came in on Saturday at 23:03: show job demand per trade against qualified people, from credible sources, with design as the example of an employer's market. I explored in three rounds that night.
- Round 1 Sat 23:03–23:28
Shortage per job (UWV, CBS, ROA)
UWV doesn't measure designers, and the creative class as a whole is still "krap" (2,15). So "more designers than jobs" can't be shown with UWV's own measure.
- Round 2 Sat 23:44 → Sun 00:24
Diplomas against new vacancies
4,564 design diplomas, 20254,600 new design vacancies, 2025Design comes out about 1 to 1.
- Round 3 Sun 00:34–01:08
Own sample of 153 job ads
One coder. 3,8 new designers per starter vacancy in design, against 2,0 starter vacancies per new electrician.
Became the draft headline at 08:36
Sun 12 Oct, 00:34 · after round 2
The hunch that shaped round 3
"I don't want to agree on the design part yet. Is there a way to find out the years of experience needed for each graphic design job opening?" Part 2 of 3 · v2 Retracted
Why did I take it back?
Derived, method givenShort answer
Before publishing, I audited my own study. It found a problem on every type of validity. Counted honestly, design and techniek overlap.
04 · The retraction
The headline that didn't hold
Before publishing, I re-derived every number from the raw files and scored 29 page items keep, fix or drop. The official numbers held. The headline didn’t.
Drafted Sun 08:36 · never public
Retracted after the audit
New designers per starter vacancy
3,8
2025 · DUO, UWV and own sample
Retracted after the audit
Starter vacancies per new electrician
2,0
2025 · DUO, UWV and own sample
Graduates per starter vacancy, 2025
Median and 90% interval · Monte Carlo, 40,000 draws
- As published
- Strict starter, all quantifiable uncertainty
- Where the ranges overlap
Design 3,8 (2,4–7,0) 5,6 (2,9–13,9)
Electricians 0,50 (0,34–0,84) 1,1 (0,5–3,3)
Plumbers 0,29 (0,22–0,44) 1,1 (0,5–3,1)
The audit opened 7 issues against the headline
Sun 12 Oct, 12:03–12:16 · every number re-derived from the raw files
7 Open 0 Closed
Validity 6
- #1 “Starter” counted ads that ask 2 years Construct Every starter share too high
- #2 Design ads came from other sites Selection Pushed design up
- #3 29 ads were really about 8 posters Independence Intervals far too narrow
- #4 A one-day snapshot over-counts long-open ads Sampling Pushed design up more
- #5 One coder, and a rule that favoured the story Reliability Favoured the story
- #6 One day’s shares applied to three years Time Same share for 2023–2025
Found later 1
- #7 93 pages fetched that robots.txt disallowed Ethics Logged in public
Verdict
Exploratory only. Off the page.
The method note on the live page
"In een eerdere versie stond een eigen telling van 153 vacatures. Die was te klein en kwam van verschillende sites, dus ik heb hem eruit gehaald." Part 3 of 3 · v3 Published
What holds up?
Verified at the sourceShort answer
A test I froze before looking: 9 of 9 routes pass. The literature backs the advice, and two more tests report in December.
05 · The test
A test that could fail
The design outcomes and some techniek outcomes had already been seen, so they could only describe. The test used routes I hadn't seen. I froze the predictions and the pass rule with a hash before opening the data.
How the test routes were picked
- 1 Unseen routes in the jobs readers hire for, plus 1 contrast 11
- 2 At least 90 respondents 10
- 3 Outcomes published by ROA 9
- 4 Beat both design benchmarks 9 of 9
Preregistration · frozen
SHA-256 e0aa1c03…Sun 12 Oct, 12:29
Written down before I opened the data.
- Hypotheses
- H1 to H4, each with a rule for when it fails
- Selection rule
- The jobs readers hire for most, plus one declared contrast. At least 90 respondents per route.
- Benchmark
- The best design figures: 7,2% unemployed and 1,9 months to a first job
- Also fixed
- Sources, coding scheme, sample size and analysis
Predictions, written down before looking
- P1 1 miss
bbl routes
17 of 18
checks held. The miss: service W at 0,4 months.
≤ 2,0% unemployed · ≤ 0,3 months · ≥ 65% own field
- P2 Held
Care routes from school
All held
No misses.
≤ 3,0% unemployed · ≤ 0,5 months · ≥ 70% own field
- P3 Direction only
Techniek routes from school
In direction
The intervals overlap per programme, so it’s not a proven gap.
≤ 1,0 months · unemployment above its bbl route
- Main Held
The main test
9 of 9
evaluable routes beat both design figures.
Beats 7,2% unemployed and 1,9 months
The rule for the main test, also written first
- 6 or fewer of 10 Not supported The design contrast leaves the headline
- 7–8 of 10 Partly supported A softer line that names the difficulties of techniek starters
- 9–10 of 10 Supported The page may say: the start is much harder in design than in techniek and zorg
9 of 9 routes start faster than design, on both measures
About 1 year after the diploma · ROA, 6 cohorts pooled · routes I hadn’t looked at
- Route, with 95% interval
- Best design figure
Unemployed, %
Care · 4 Techniek · 5
Months to a first job
Care · 4 Techniek · 5
Wilson 95% intervals · scored Sun 12 Oct, 12:58
Deviation, logged
The tenth route had no data
Missing counts as a fail 9 of 10 · supported
Missing left out 9 of 9 · supported
Supported needs 9 or 10 of 10. Both readings land there.
ROA publishes no outcomes for it, and the plan hadn't said how to treat missing data. So both readings are reported.
Counter-signal, kept on the page
Electricians from school
- Electricians from school
- Design benchmark
Unemployed 1 year after the diploma · n = 102 · 95% interval 3,1–12,9%
Close to design. ROA rates this mbo-3 route "slecht", like mbo-4 media design. It fits a quieter thesis: techniek employers prefer starters who already have practice.
06 · Testing the argument
The data compare jobs. The argument compares employers.
So each claim in the essay needed its own evidence, traced to the source. A claim I couldn't verify was dropped or reworded.
19 studies traced to the source, by DOI
Rules, written before checking
- Primary sources only
- Population, setting and effect recorded
- Abstract-only marked
- Vendor data labelled
- No paywall workarounds
7 claims in the essay
- 4 safe
- 2 reworded
- 1 dropped
Safe 4
-
Well-known employers get more applicants
Say “research in the US”.
Turban & Cable 2003; Collins & Han 2004
-
How you treat applicants affects whether they want you
With no numbers.
Chapman et al. 2005; Uggerslev et al. 2012 (meta-analyses)
-
A fast offer gets more yeses
With no “within X hours”.
Becker et al. 2010 (n = 3,012)
-
An honest picture of the job keeps people longer
With no percentage.
Earnest et al. 2011; Phillips 1998
Reworded 2
-
Show the pay and you get more applicants
Good pay shown draws people; a low wage shown can push them away.
Banfi & Villena-Roldán 2019; Balgova et al. 2025; Belot et al.
-
Long applications lose applicants
Mostly vendor data, and one large study contradicts it. Say “keep the first step short”.
Hardy et al. 2017 (222,772 job seekers)
Dropped 1
-
“60% abandon long applications”, “365% more completions”No primary source.
—
Closest support for the thesis
The data can't test the employer half. The closest study found that the recruiting that works depends on how well-known the firm is.
Collins & Han (2004)
Which recruiting works better, by employer profile · abstract only, US firms
| Low-effort recruiting | High-effort recruiting | |
|---|---|---|
| Well-known firm | – | Works better |
| Low-profile firm | Works better | – |
07 · What the evidence supports
How strong each claim on the page is
The official data carry the claims; my own data is described in a plain method note. Every claim on the page gets the strength it earned. The one I can't support yet isn't on the page.
| Strength | Claim | Source |
|---|---|---|
| Confirmed | In techniek and zorg a starter has work within about 2 weeks; a design graduate searches 8 to 12 | ROA, 6 cohorts · preregistered H1 |
| Descriptive | 9 of 10 new electricians and installatiemonteurs already had a boss at their diploma; designers 0 of 10 | DUO 2025 · official |
| Descriptive | 6,5 new vacancies per new electrician in 2025; design about 1 (with the caveats in 02) | DUO, CBS, UWV |
| Descriptive | Monteurs in Brabant: 9 to 16+ open vacancies per person on WW | UWV Q2 2026 · official |
| Descriptive | The market cools nationally, but not in techniek and zorg | CBS, UWV · out-of-sample check H3 pending |
| Direction only | School leavers in techniek without practice have a slightly harder start | ROA · the intervals overlap |
| Argument | Don't copy big-brand hiring; make it easy for the applicant | Verified research where stated (06) |
| Not claimed | Ads in design ask for more experience | Measured until 7 Dec · H4 |
Known limits
What the data can’t tell you, so the page doesn’t claim it.
- ROA Intervals are approximate
- UWV Counts only people on WW, capped at 16
- UWV Flows read off a dashboard by hand, each checked
- DUO 2025 figures are provisional
- Diplomas Aren’t supply; vacancies aren’t entry jobs
- Sales Least precise; left out of the diplomas comparison
- Applicants No first-hand drop-off data yet
- Literature Brand studies are from the US
How the data was handled
Public sources only. What the data collector for H4 does and doesn't do:
Respected
- robots.txt (RFC 9309, every redirect)
- Text-and-data-mining opt-outs
- The requested crawl delay
- An honest user agent
Never used
- Logins
- CAPTCHAs
- Paywall workarounds
Not stored
- Contact names
- Phone numbers
- Addresses
Not published
- Anything from the ads
Client data
- None. Future checks ask consent for anonymous, pooled use (5+ per group)
Corrections are public in the plan, including the robots.txt mistake.
08 · Still running
Two more tests that can still fail
Both were written down before any data came in, and both report in December. The results go on the page whichever way they fall.
H3 Waiting for UWV's Q3 figures
Shortage per job holds while the national market cools
- Q2 2026, published
- 4,0 = "zeer krap"
- Fail zone
Plumbers 14,7
Machine fitters 10,6
Electricians 9,3
Care assistants 8,9
- Prediction
- Electricians, plumbers, machine fitters and care assistants stay at 4,0 or higher ("zeer krap"). Sales reps stay below 4,0, and the creative class below techniek.
- Falsified if
- Two or more of the four fall below 4,0. The page then gets a dated update.
- Reports
- When UWV publishes its Q3 2026 figures, around December.
H4 Running · 13 Oct → 7 Dec
Design ads ask for more experience than ads in the reader's trades
Ads needed per job, out of a planned 150
Planned: 150 = about ±10 points. H4a needs ~65. H4b needs ~180.
- Prediction
- H4a: design ads ask 3+ years at least 20 points more often than each techniek job. H4b: their strict-starter share is at least 10 points lower.
- Falsified if
- The 95% interval of the difference includes 0. Reported either way, also if it contradicts the retracted sample.
- Reports
- A fixed end on 7 Dec. No extension, whatever n design reaches.
How H4 closes the old issues
The ad study, done properly · 13 Oct → 7 Dec · design fixed before any data
0 Open 9 Closed
From the audit 7
- #1 “Starter” counted ads that ask 2 years Construct Strict starter defined in advance; vague asks get their own bars
- #2 Design ads came from other sites Selection One source for all seven jobs: EURES
- #3 29 ads were really about 8 posters Independence At most 2 ads per employer per job; intervals clustered by employer
- #4 A one-day snapshot over-counts long-open ads Sampling Flow: only new ads, each weekday, for a fixed 8 weeks
- #5 One coder, and a rule that favoured the story Reliability Fixed codebook, masked ads, three coders, kappa ≥ 0,70
- #6 One day’s shares applied to three years Time H4 measures its own window; nothing is applied to past years
- #7 93 pages fetched that robots.txt disallowed Ethics robots.txt checked to the standard (RFC 9309), every redirect too
Added before the start 2
- #8 Cut-off texts tied to the job Pilot Completed from the ad’s own page, only if the texts match
- #9 Room for optional stopping Plan A fixed end on 7 Dec, whatever n design reaches
09 · Learnings
What I learned
-
A number that agrees with you deserves the hardest check. The sample was designed after a result I didn't like, and it found what I hoped. That was exactly the moment to audit.
-
Write down what you expect before you look. Then a pass means something, and a fail would have changed the headline. I said so in advance.
-
Measure the reader's question directly. "How fast do starters find work?" beat a ratio built from three uncertain parts.
-
Validity is more than sample size. The old sample's biggest problems were a label, a source and a sampling method, not n.
-
A market fact needs the part the reader controls. Otherwise "it's the market" becomes an excuse not to act.
-
How to check it
- Scripts re-derive every public number from the raw files
- The page's data file is generated, never typed
- The freeze log holds a hash of every plan version
- Raw downloads are kept in their own folders
Publication
Published as a data story
The findings went live as a Dutch data story at talentbezorgd.com/inzichten/krapte: one measure per chart, the reader picks their trade once and every chart follows it, every chart has a table, and a method note names the sources, the preregistered test and the retracted sample.
How the findings reach an employer
-
Every market fact ends with the part the reader controls: "Zie je een goede starter? Wacht dan niet."
-
Official data carry the claims. My own data sits in a plain method note.
-
One frame for numbers: counts out of 10, like "9 van 10 nieuwe elektriciens hadden al een baas".
-
The page opens on the reader's own trade; design only appears "ter vergelijking".
More work
More case studies
Vacaturescan
A free job-ad checker that marks problems while you type. Testing showed free AI does the same job, so I stopped it on day 3.
Built in 1 day
Test rounds 2
Stopped on Day 3
Stichting Join Us
Nobody signed up after 200,000+ ad views. I found out why and redesigned the page: from 0 to 20 sign-ups.
Ad views 203,613
Clicks 1,309
Sign-ups 0 → 20