Getting and using market compensation data

1. A benchmark is a defined comparison population

“The market pays €70,000” is incomplete. Specify the work, level, geography, employer segment, pay measure, working time, effective date and statistic. A national occupation median, an employer-weighted survey median, a job advertisement range and a modelled salary prediction are different products.

Market evidence informs a pay policy. It does not determine the morally correct wage, an individual's legal entitlement or the value of their work. Cullen, Li and Perez-Truglia study access to salary benchmarks using payroll data and a difference-in-differences design, finding reduced salary dispersion in that setting. This supports taking information frictions seriously, not treating an observed market rate as an immutable price. Cullen, Li & Perez-Truglia (2026), Review of Economic Studies 93(4), 2531–2573. Online 2025.

Choose a policy reference such as a matched market median only after deciding the relevant labour market, affordability, internal equity and employment proposition. P75 is a percentile, not a synonym for “competitive”. Setting every employee to P75 cannot put every employer below that point in an unchanged market distribution.

2. Public sources: what each can actually support

The following access points are public. Some detailed microdata or restricted products require applications, conditions or payment. Access to an aggregate table does not grant unrestricted redistribution of all underlying material. Check the specific dataset's licence and suppression rules.

These sources are valuable even when they cannot supply a ready-to-use visual benchmark. A precise-looking number with the wrong compensation boundary is less useful than a documented absence.

3. A practical Germany-first acquisition workflow

1. Define a benchmark job. Record purpose, principal tasks, responsibility, required skill, autonomy, organisational scope and level. Match job content before title. “Manager” can describe an individual contributor or a large department head.
2. Choose the labour market. Local operational roles and internationally recruited specialists may have different relevant markets. Record location policy for remote workers. Do not combine residence and employer location without explanation.
3. Find public context. Use BA occupation/requirement-level and regional results, Destatis tables and an applicable collective agreement. Preserve the displayed occupation aggregate and any fallback. Do not encode a censored value such as “greater than X” as an exact median X.
4. Assess whether context is sufficient. If job level, incentives or executive equity materially affect the decision, broad public data may be inadequate. Seek a properly documented specialist survey or an independent compensation analysis. Do not manufacture missing P10/P90 values from a median.
5. Acquire under appropriate rights. For a purchased/participation survey, ask for the pay-code definition, job catalogue, participant/employer count, incumbent count, weighting, fieldwork dates, effective date, eligibility treatment, suppression, quality indicators and permitted internal/report redistribution. For public data, download the table and its metadata together.
6. Build a comparison record and obtain a second review. A reward specialist should approve the match and the pay boundary before the analyst loads it. Keep unresolved matches separate from accepted ones.
7. Load only compatible benchmarks. Use a unique reference key and repeat its identical values on relevant employee rows in the one-table input. Validate that the join did not duplicate employees.
8. Freeze a dated version. Retain the source file, selection/filter specification, transformation, reviewer and licence. Reassess when the job, pay policy or source definition changes, not merely when the date becomes old.

For a public-source-only starting point, deliver a short evidence note for each benchmark job rather than asserting a fully matched salary range. Suggested categories are “directly comparable”, “context only” and “not usable”. These are our governance labels, not statistical confidence levels.

4. What to record with every market reference

The visual has source/date and reference-key fields, plus flexible tooltip fields, but it is not a complete survey-governance database. Keep the full record in controlled source documentation. A concise reference key connects it to the employee table. You may retain additional audit columns in that same table. The optional-column adapter preserves extra columns, although authors must bind additional tooltip fields to display them.

5. Transformations, ratios and statistical traps

Time alignment. If an approved, relevant wage index is used, aligned value = original value × index(new date) / index(old date). With an explicitly assumed constant annual movement g, an alternative is value × (1+g)^(months/12). Both are transformations, not new observed salaries. Preserve the original and adjusted values and identify the assumption. An inflation index measures prices. It need not describe wage movement in the target labour market. BLS ECI.

Currency. Use a documented FX convention for the decision: spot, period average or budget rate. A currency conversion does not make different labour markets identical. Do not use PPP-adjusted cross-country wages as the euro amount promised in an employment contract. This visual currently supports EUR. Conversions must be performed upstream.

Market ratio. employee comparable pay / matched market P50. €66,000 / €60,000 = 1.10, or 110%. That is 10% above the matched median. It is not the 60th or 75th percentile. A percentile rank cannot be inferred from one median.

Compa-ratio. employee salary / internal range midpoint. The denominator is an internal policy midpoint, which may differ from market P50. Range penetration is (pay − range minimum)/(range maximum − range minimum) when the range has positive width. A value outside 0–100% is possible and informative. Do not clamp it simply for cosmetic presentation. These are recommended explicit definitions. Confirm any provider's use of similarly named indicators.

Quantiles cannot be pooled from quantiles. Job A pays [40, 50, 60] and Job B pays [100], in thousands. Their medians are 50 and 100. The average of medians is 75, while the pooled median is 55. Even a headcount-weighted average of medians, 62.5, is wrong. A grade benchmark therefore needs a directly supplied grade distribution or appropriately governed underlying observations. This project's separate grade-market fields prevent a synthetic “grade median” formed by averaging job medians.

Employee versus employer weighting. A firm with 1,000 incumbents can dominate an employee-weighted sample. Equal employer weighting answers a different question. Retain the survey's method. Do not describe either as universally unbiased. Combining two survey results can also double-count overlapping participants.

Eligibility and selection. Bonus-recipients-only data differ from all eligible employees including zero awards. Paid salaries differ from job-advertised ranges and self-reported offers. Voluntary survey participation, occupational coding, unobserved job level and missing high earners can affect interpretation. Ask how the source addresses them. Do not “repair” them with an arbitrary uplift.

Band and market are distinct. Internal range minimum/midpoint/maximum are design parameters. Market P10/P50/P90 are distribution estimates. A market P10 is not a contractual salary floor. An internal band constructed around P50 remains a policy choice.

6. How market mode has been simplified

The normal internal menu has eight pay totals. In the prepared market view, only three have matching benchmark sets: Base Salary, Total Cash · Actual, and Total Cash · Target. Total direct/remuneration are not compared to a cash benchmark simply because their currency is the same. You can explicitly configure a properly matched reference for another measure in the author settings.

Job and grade market references remain separate. Partial percentile coverage is retained. Absent percentiles stay absent. Inconsistent references for the same key and incompatible actual/target bases are excluded from the comparison. Company/location distributions remain available internally. The market grouping choices are intentionally restricted to the reference relationships the model can support. Selecting a broader grouping must not invent a matching survey cut.

The shipped market values are invented fixtures for testing appearance, missingness and calculations. No real public or commercial data have been inserted into them. A real benchmark import should follow the acquisition and approval procedure above.

7. Questions a reviewer should be able to answer

Can another analyst reproduce the source cut and match? Are actual and target distinct? Are effective dates aligned? Is the statistical population relevant? Were censored or suppressed observations handled honestly? Are all transformations documented? Is the result suitable for the decision, and what changes if a plausible alternative match is used?

Where answers are weak, present a range of plausible decisions or seek better evidence. Extra decimal places cannot compensate for an uncertain job match. A well-governed “insufficient evidence” outcome is preferable to an unsupported market claim.