Outliers and Data Cleaning: The Rule We Use
Large administrative datasets contain rows that no one should average. In LCA disclosure files the most obvious examples are wage fields that were plainly entered as placeholders or in the wrong unit: annual salaries of a few dollars, rates that appear to be hourly figures typed into an annual filing, and round, extreme values far beyond any plausible salary for the occupation. A single such row barely moves a median — which is one reason this site leads with medians — but clusters of templated rows from a single filing pattern can distort means, ranges and percentile tails, and they make minimum and maximum figures meaningless if reported raw.
Our rule is deliberately simple and published in full. Before any median, mean or percentile is computed, an annual-wage filing is excluded from wage statistics if its offered wage is below $15,080 or above $1,000,000. The lower bound is the annualized federal minimum wage: $7.25 an hour multiplied by 2,080 hours equals $15,080, and no full-time professional filing under these programmes can lawfully offer less, so a lower annual figure is a data problem rather than a salary. The upper bound reflects that offers above one million dollars in these filings are, in the observed data, dominated by placeholder-style entries — repeated identical extreme values inconsistent with the occupation and employer — rather than by genuine seven-figure professional offers.
Excluded rows are not deleted from history. They still count in case totals where they are certified filings, they are counted separately in the site manifest, and this page and the methodology both state the rule so any reader can replicate or challenge it. What they do not do is enter a median, a mean, a percentile, a minimum or a maximum. The 'lowest included offer' and 'highest included offer' labels in our wage tables are worded that way deliberately: they describe the data after cleaning, and pretending otherwise would smuggle the placeholders back in through the range figures.
We chose fixed, principled bounds instead of statistical trimming for a reason. A rule based on percentiles of the data itself — trimming the top and bottom one percent, for example — would remove genuine high salaries in high-paying occupations and genuine low salaries in lower-paying ones, and its effect would differ invisibly from page to page. Fixed bounds apply identically to every employer, occupation, state and city on the site, they can be explained in two sentences, and a reader with the raw files can reproduce every figure on this site from them. Reproducibility is worth more here than a marginally smoother distribution.
Cleaning has limits, and we state them. The rule cannot catch a plausible-looking but wrong wage — an annual figure that is really a monthly figure multiplied incorrectly, for instance, can sit comfortably inside the bounds. It cannot correct an employer that mapped a role to the wrong SOC code, and it cannot repair a worksite recorded as a headquarters instead of a client site. Those are reasons to read distributions rather than single rows, and to treat any one filing as an attestation with known error modes rather than as audited fact. The methodology page lists the remaining known limitations alongside this rule.
How These Figures Were Computed
Every figure on this page was computed by our script directly from the U.S. Department of Labor OFLC LCA disclosure files for FY2024 Q4 and FY2025 Q1–Q4, retrieved in October 2026. We kept only cases whose CASE_STATUS is exactly “Certified” — denied, withdrawn and certified-withdrawn cases are excluded. Headline medians use filings whose wage unit is a year, because annual salaries compare cleanly; hourly, weekly and monthly filings are counted in case totals but are not converted into the headline median, and our methodology page explains that choice in full. Wages below $15,080 a year (the annualized federal minimum wage at $7.25 an hour) or above $1,000,000 a year are excluded from medians as likely data-entry placeholders, and the number excluded is reported in the manifest that accompanies this site. Medians are shown alongside means because a handful of very large offers can drag a mean far above what a typical filing actually offered.
Frequently Asked Questions
Where does the data behind this guide come from?
All figures referenced in this guide come from the U.S. Department of Labor OFLC LCA disclosure files for FY2024 Q4 and FY2025 Q1–Q4, retrieved in October 2026, filtered to 662,198 cases with a status of exactly Certified. The methodology page documents the full cleaning process.