Methodology

Every response is published, and each belongs to exactly one of three groups: full-time employees, part-time employees, and contractors and B2B. Each has its own figures, its own count and its own threshold. Nothing is left out for being the wrong kind of contract.

They are never averaged into one number, because they are not one quantity. A contractor’s gross carries social contributions an employer would otherwise pay, so it is far higher for the same take-home; mixing it with employed gross is the single most distorting mistake available in European pay data. A part-time salary is published exactly as it is paid and never scaled up to full time, because extrapolating a 60% contract invents a salary nobody receives.

Contractors are measured per day. A day rate is the number that actually describes self-employed work: it is a single negotiated price, with no bonus, equity, thirteenth month or part-time fraction in it. Where somebody quoted a rate per day, it is used exactly as given. Where they quoted an annual or monthly figure, it is divided by a standard 220-day year; an hourly rate is multiplied by a standard 8-hour day.

The standard year is the important part. Dividing by the days somebody personally billed would fold their holidays, parental leave and dry spells into the median, which would report time off as though it were a discount on the rate. Taking six weeks off does not make you cheaper. The raw amount and its period are what we store, so anyone who prefers different standards can recompute the whole thing from the dataset.

Included

Gross base pay, plus bonus actually paid in the last twelve months, plus annualised equity.

However you are paid

Base pay can be quoted per year, month, day or hour. Asking everyone for an annual figure forced a conversion in their head, worst of all for freelancers and B2B contractors, who think in day rates and would have had to guess a working year.

Not included

Pension contributions, benefits, and anything not asked for. Employer social contributions are excluded, which is part of why employee and contractor figures are published separately.

The multiplier is asked for, never assumed

A monthly figure needs the number of payments (12, 13 or 14, the last two being normal in Spain, Portugal, Italy, Austria and Greece). A day rate needs the days actually billed last year; an hourly rate, the hours. €600 a day is €120,000 over 200 days and €138,000 over 230, and picking between those on someone's behalf would be publishing a figure nobody supplied. Where the count is missing, the response is left out of the figures rather than annualised on a guess.

Everything is converted to euro using the European Central Bank’s daily reference rates. Rates are stored per day, so re-running the aggregation reproduces the same numbers rather than drifting with today’s exchange rate. The raw amount and its original currency are what we store; the converted figure is derived.

No cost-of-living adjustment is applied to any published figure. €70,000 in Zurich and €70,000 in Lisbon are different lives, and pretending one number captures that would be worse than leaving you to make the comparison yourself.

A figure is checked once it has been annualised and roughly converted to euro, because neither check works alone: 200 a year is nonsense while 200 a day is an ordinary contractor, and 3,000 is a fine monthly salary in euro but about €7.50 in forint.

Anything that lands outside roughly €1,000 to €3,000,000 a year is refused at submission. That is a typo, and one of them in a small sample moves a median. Merely unusual figures are accepted and flagged for review: the lowest-paid junior in Ukraine and the highest-paid principal in Zurich are both real people whose answers we want.

The rates used for that check are approximate and never touch published figures, which convert at the European Central Bank’s stored daily rates.

Before any figure is computed, the lowest 1% and highest 1% of a sample are dropped. Trimming rather than capping: a mistyped €10,000,000 should leave the sample entirely, not be pulled down to the 99th percentile where it still drags the average.

The rule is symmetric, so it cannot be tuned to push a median in a preferred direction, and samples below twenty are left alone, because cutting the tails off a small sample removes a meaningful share of it and does more harm than the outliers would.

Fewer than 5 responses

Withheld entirely. This is the privacy rule: it is what stops "principal engineer, Ljubljana, 15 years" from being one identifiable person.

Fewer than 60 salaries

No median published for that slice, for full-time or part-time employees. This is a statistical rule, deliberately much higher than the privacy one: a median over six people is not a median worth printing.

Fewer than 25 contractor day rates

No day-rate median for that slice. Lower than the 60 a salary median needs, because a day rate is a single negotiated price with no bonus, equity, thirteenth month or part-time fraction in it. Less of that variance means the median settles on fewer answers; it is not a lower standard applied to get something on the page sooner.

Suppression is applied when the published files are generated, so a withheld figure is genuinely absent rather than hidden by the interface.

Figures are rebuilt nightly, never live. That is partly cost and partly defence: with no immediate feedback, somebody submitting invented salaries cannot tell whether it moved anything, which removes most of the incentive to try.

Releasing one row per response would contradict the 5-response rule, because a single row is a slice of one. So the file is released under k-anonymity: cities are dropped, experience is banded into five-year ranges, pay is rounded to the nearest €500 so an exact figure cannot fingerprint a row, and any combination of characteristics appearing fewer than 5 times is withheld.

Licensed CC BY 4.0. Plot it, audit it, argue with it.

Self-selection

People who answer salary surveys are not a random sample of engineers. They skew towards the engaged, the online, and often the better paid. Treat every figure as a description of who responded, not of the profession.

No verification

Nothing here is checked against a payslip. The defences are against automated and repeated submission, not against one person being untruthful once.

No demographics

Gender is not collected, so no pay-gap analysis is possible. That is a real cost of keeping the dataset as thin as it is.

Thin cells

Small countries and senior levels will stay unpublished for a long time. That is the design working, not a bug.

A standard working year

Contractor rates derived from an annual figure assume a 220-day year for everybody. It is the same assumption for everybody, which is what makes the medians comparable, but it is still an assumption: somebody who bills 180 days a year and somebody who bills 240 are not described equally well by it.

Every threshold on this page is read from the same module the aggregation uses, so this description cannot drift from the behaviour. All of it is in the public repository.