Methodology
How the figures on this site are calculated, and why they may differ from other sources.
What the grant rate counts
Grant rate = granted ÷ (granted + denied). Administrative closures, withdrawals, abandonment and "not adjudicated" are excluded from the denominator: the court never ruled on the merits, and recording a denial where none occurred would be wrong. Counting them as denials drops the percentage by tens of points and makes it stop meaning what people read it to mean.
The unit is a decision, not a person
We count decisions per proceeding. One person may go through several proceedings and will then be counted more than once. Counting people would be a fairer answer to "what are my chances", but it would be comparable to no public source.
Withholding and CAT are counted separately
Applicants almost always file asylum, withholding of removal and Convention Against Torture protection together. A decision that denies asylum but grants withholding is simultaneously a denial and a protection from removal: the person stays, but with no refugee status, no right to bring family, no path to a green card, and the protection can later be lifted. In the asylum grant rate such a decision therefore counts as a denial, and the cases are shown as their own figure.
U.S. fiscal year
October 1 through September 30. FY2025 began on October 1, 2024. This is the first reason figures fail to match sources that count calendar years.
An interval, not a bare percentage
Every percentage is shown with a 95% Wilson interval. It shows how reliable the figure is at that sample size: for a judge with eleven decisions the interval spans more than thirty percentage points, and "91% granted" there means something entirely different than it does for a judge with three thousand.
The display threshold, and the dash
A percentage is not shown where there is nothing to compute it from: for a judge with a dozen decisions it is noise, not a characteristic. The judge list therefore holds those with fifty asylum decisions or more, and each block of a page has its own threshold, named right under the table. Where the volume fell short the cell holds a dash, and the dash means "too little to count", not "no such cases".
Why months are not available for every year
On judge and court pages the period can also be taken by month — but only for the last three fiscal years. Anything earlier is counted a whole year at a time. The reason is page weight, not missing data: a month-by-month breakdown of a large court's whole history weighs three times as much as the rest of the page, and nobody asks it of 2011. The boundary falls exactly on the start of a fiscal year so that no year is cut in half.
Elapsed time: a median, not a mean
A proceeding's elapsed time runs from the date of the charging document (NTA) to the date it was completed, and is shown as a median with quartiles. A mean would be useless here: a handful of cases that drag on for a decade pull it to where nobody's case actually sits. Unfinished proceedings are counted differently — from the NTA to the release's data cutoff: that is not a forecast, it is how long the person has already waited.
In absentia decisions
The share of proceedings completed in absentia is counted over removal cases only: other kinds of proceeding almost never end in absentia, and a combined share would quietly mix two different streams. The figure speaks to appearance, not to how strict the judge is, and proceedings with no appearance recorded count in neither the numerator nor the denominator.
Three dates, and they mean different things
A release is named by the year and month EOIR published it. “Data through” is the last day the figures hold for — the release contains nothing after it. “Built” is the day we assembled the database from that release, and it falls weeks later: EOIR publishes with a lag, and the site is rebuilt as soon as it does. So a page carries a build date weeks after the day its numbers stop, and that distance belongs to the source, not to the site.
The data is rewritten every month
EOIR publishes no changes, only the whole database again, and months already closed arrive in the new copy with records added. The practical consequence is one: the most recent month is always incomplete and cannot be compared with earlier ones. How much the source rewrites the past is visible month by month.
Gaps in the source
The release is known to lose records: some entries, applications for protection among them, disappear from it — researchers and members of Congress have both raised this with EOIR. That is a defect of the source, not of the calculation. Data missing from your case here does not mean it is missing at the court.
The codes in a case record
Court data is made of codes: one letter for "denied", two for "individual hearing". Where EOIR publishes a lookup table — about twenty of them ship with the release — we substitute the plain wording. Where it publishes none (hearing medium, motion decision, charge status) the letter stays, with a note beside it saying what is known about it and how. We will not invent a plausible label: read as fact, an invented one is worse than an opaque letter.
The asylum clock is an estimate
The clock is computed from EOIR's own asy_clk_stat markings on adjournment codes: the reason for each adjournment determines whether the clock runs or stops. Some codes EOIR resolves manually — those days are reported separately rather than assigned to either side. For the official figure ask the court or check ACIS.
The short version that travels with the data
Grant rate = granted / (granted + denied). Administrative closures, withdrawals, abandonment and "not adjudicated" are excluded from the denominator. What is counted is a decision per proceeding, not per application: asylum, withholding and CAT protection are almost always asked for together, so withholding and CAT grants are counted separately from asylum. The U.S. fiscal year runs October 1 – September 30.
The same methodology in one paragraph. This text is served through the API next to every figure: the caveat has to travel with the numbers wherever they appear — other people's integrations included.