Asylum and immigration data methodology

How to read a statistic

Start with the population, period, counting unit and denominator shown beside a figure. A proceeding is not a unique person or case. Compare only like definitions and complete periods; historical shares describe recorded groups, not personal probabilities. Detailed calculations and audits follow below.

• Definitions
• Published data editions

Recorded grant share · M1.1.0

Grants ÷ (grants + denials) among eligible asylum proceeding records. One proceeding event is the unit; this is not a count of unique people. Removal, deportation, exclusion and asylum-only proceedings. ASYL grant codes G/I/P; denial D. Other outcomes are excluded from this denominator. Conflicting applications and inconsistent case links are excluded.

Periods and sample size

The standard comparison period is shown in Published data editions on the Sources page. Judge and court histories can cover additional years. A fiscal year ending after the source cutoff is labeled partial wherever it is shown; it is not included in complete-year comparison windows. A year without records adds zero to complete-source totals. Separately sourced measures retain their own availability rules.

• Published data editions

Intervals and uncertainty

This interval describes binomial sampling variability only; it does not account for administrative errors, case mix, attribution uncertainty, or selection bias.

Record linkage, selection and fiscal year

The unit is an eligible IDNPROCEEDING event, not a unique case, person or raw application row. We group ASYL rows in tbl_Court_Appln by proceeding, require a single linked IDNCASE matching B_TblProceeding.IDNCASE, and keep only one nonconflicting application result. Duplicate rows with the same result count once; conflicting or missing results do not enter the grant-or-denial denominator. Eligible proceeding types are RMV, DEP, EXC and AOC. The proceeding’s recorded court, assigned judge code and COMP_DATE determine the court, judge and fiscal year; FY2025 spans 2024-10-01 through 2025-09-30. We do not select only the last proceeding or last application in a case, so separate eligible proceedings can contribute separately. The source has no independent application decision date or verified decision-author field. COMP_DATE therefore attributes the linked outcome to a proceeding date, not a proven standalone asylum-decision date; the recorded court may differ from the historical hearing venue.

• EOIR source release and SHA-256
• Santa Ana integer audit and pinned source hashes

What a judge or court label means

Judge attribution uses the proceeding’s IJ_CODE assignment. It does not establish who personally authored every decision. Court attribution uses the proceeding’s recorded court, which may differ from historical hearing venue.

How judge attribution works

EOIR defines B_TblProceeding.IJ_CODE as the immigration judge assigned to the case. Asylum grants and denials come from the joined tbl_Court_Appln.APPL_DEC: G, I and P count as grants, and D as a denial. B_TblProceeding.COMP_DATE is documented as the date an immigration judge rendered a decision on the proceeding; it is the date used to group these records by fiscal year. The available application table has no separate decision date or judge-author field, so neither IJ_CODE nor COMP_DATE establishes who personally made each APPL_DEC. Some saved outcomes are associated with judge codes on proceedings dated outside independently documented service periods. We keep those records because service dates cannot identify the decision author or justify changing the source population. Judge figures therefore describe recorded outcomes on eligible EOIR proceedings carrying a judge code, not verified personal decisions.

• EOIR Case Data Code Key

Coverage before filters

Every judge uses the same six coverage dimensions: identity linkage, associated source proceedings, asylum linkage, counting eligibility, date availability and result availability. The edition-wide summary is independent of dated details and current filters. Missing dates and missing outcomes can overlap and must not be added as separate proceedings. A historical judge-code association does not verify decision authorship.

• Missing, empty and failed states

Where outcome totals come from

Judge × court grant and denial totals come from the dated judge records, court totals from the population history, and nationality and its court/judge intersections from a complete aggregate of the pinned public EOIR CASE source. These series show a period with no record as no records and flag small N. The original counts.json stays unchanged as a historical artifact. Representation and other measures retain their separately documented source and publication rules.

Sample and disclosure rules

Every verified public EOIR court, judge and nationality outcome count is shown when at least one eligible record exists, including N = 1 and one-sided outcomes. N below 100 is flagged as a small sample; a recorded grant share for a few proceedings can swing sharply and is not an individual forecast. Difference and ranking measures may require N ≥ 100 for interpretability. USCIS and other source tables can have values the source itself did not publish; those remain unavailable.

Legacy masks and current publication

The historical counts.json release used a count-reconstruction suppression model. Those masks are retained only to preserve that release. The current EOIR court, judge, nationality, representation, bond, duration, entry and absence views use separately validated complete public-source editions, so a legacy withheld cell can be visible on the site. BIA editions have separate disclosure scopes: the v1 companions retain certified small-cell and complementary masks, while earlier exact editions remain separate historical downloads and context. The old certificate never guaranteed non-reconstruction across later public outputs and is not a privacy claim for the current site.

Complete public judge result records can reconstruct counts hidden in legacy aggregate tables. Those tables no longer provide a non-reconstruction guarantee. The timeline omits case identifiers and personal attributes.

Historical completed-proceeding duration · M57.1.0

One eligible grant-or-denial proceeding record contributes one interval from B_TblProceeding.INPUT_DATE to B_TblProceeding.COMP_DATE, grouped by completion fiscal year and recorded court. INPUT_DATE must be on or after 1980-01-01; both dates must be usable and COMP_DATE must be on or after INPUT_DATE. For complete FY2021–FY2025 and proceeding types RMV, DEP, EXC and AOC, the displayed 25th, 50th, 75th and 90th percentiles use the valid completed intervals. They use continuous linear interpolation at position p × (valid N − 1) in the sorted intervals, between floor and ceil positions; excluded date records are counted separately. A court-year distribution appears with at least one valid interval, and fewer than 100 valid intervals are marked as a small sample. This measures the historical interval of eligible proceeding records, not the total elapsed time of an entire immigration case or the remaining wait for an unfinished case.

• EOIR source and field definitions

Annual grant-share variation · M04

For the selected outcome series, take each visible annual grant share with N ≥ 100, omit the series’s partial fiscal year, and, for a population-history series, use only FY2021–FY2025. Sort by fiscal year; require at least three eligible years. Each annual percentage has equal weight, regardless of its proceeding count. The population standard deviation is sqrt(sum((share − mean)^2) / number of eligible years), not the sample standard deviation. MAD is the median of |share − median(eligible shares)|; for an even number of values, each median averages the two middle values. The largest adjacent change is the maximum absolute percentage-point difference only where two eligible years are consecutive fiscal years. Missing or ineligible intervening years are skipped without treating a cross-gap difference as an annual change; if no consecutive pair exists, that change is unavailable. These summaries describe variation in recorded shares, not judge quality.

• Historical grant-share population and source

Court distribution across judge-code groups · M05

For a court in complete FY2025, take each visible judge-code × court grant-or-denial group with N ≥ 100. At least three groups are required. Each eligible judge-code group contributes its own grant share once and is weighted equally, regardless of its number of proceedings; this is not the pooled court grant share. After sorting shares ascending, min and max are the ends. The mean is the unweighted sum divided by group count. The population standard deviation divides the sum of squared differences from that mean by group count before taking the square root. Q25, median and Q75 use linear interpolation: at p = 0.25, 0.5 or 0.75, position x = p × (group count − 1), interpolate between the shares at floor(x) and ceil(x). Missing, unavailable and N < 100 groups are excluded. The partial FY2026 is not used. The source IJ_CODE is an assignment association, not a verified decision author or a quality score.

• Judge-code attribution and source fields

Saved in-absentia flag share · M18

The unit is one eligible proceeding in RMV, DEP or EXC with nonblank DEC_CODE and COMP_DATE in a complete fiscal year. N counts those proceedings; the numerator counts ABSENTIA = Y, and the share is 100 × flagged / N. Blank and other flag values remain in N, without being called confirmed attendance. The current complete edition includes every observed national, court, nationality and court-nationality group, including N = 1; N < 100 is marked as a small sample near the displayed share, not suppressed. The flag is not an official removal-order count, a unique-person count or evidence that an order remains in force.

• Detailed in-absentia method and official source

Saved future schedule lead · M26

The population is schedule records classified as future_observation after the source linkage, case match, completion, calendar, creation-date, slot-time and unique-open-slot checks. The reference is the 2026-08-31 snapshot cutoff month, 2026-08. For each saved future calendar month, lead = 12 × (year − 2026) + (month − 8). Count observations in bands 1–3, 4–6, 7–12 and 13–120 months, through the published 2036-08 horizon. The displayed middle month is the first chronological month whose cumulative observation count reaches ceil(N / 2); it is a count-weighted lower middle month when N is even. If N = 0 it is unavailable. The lead panel uses all future months for the selected court and hearing type, regardless of the console’s month-range filter. Missing, uncertain and excluded records are outside future_observation, while repeated schedule records can remain separate; this is not a unique case or person count, a verified hearing, or an estimate of an individual’s wait.

• Saved schedule source and cutoff
• Explore saved hearing schedules

Historical judge-code ranking · M42

For the selected fiscal year or complete multi-year period, first limit judge-code × recorded-court groups by any selected state or court. A group is eligible only when its grant-or-denial cell is visible and N = grants + denials is at least 100. The current FY2026 option is partial through the source cutoff and is labeled partial; it is not a full-year comparison. Sort eligible groups from highest to lowest exact grant share, comparing two groups by grants(A) × N(B) versus grants(B) × N(A), so rounded display percentages cannot change their order. Equal exact shares receive competition ranks such as 1, 1, 3. Name, then code, breaks ties only for deterministic display order; it does not change tied ranks. Searching hides rows without reranking, and clicking table headings changes display order without changing the assigned grant-share rank. These are selected historical recorded outcomes associated with EOIR IJ_CODE and court, not personally authored decision rates, judge quality scores, recommendations or future probabilities.

• Grant-share denominator and judge attribution
• EOIR source and field definitions

Methods for each topic

Each topic explains its population and counting rules beside its records.

• How lawyer-record comparisons are calculated
• Court, judge and nationality selection methodology
• Entry-cohort completion methodology
• BIA outcome methodology
• In absentia methodology
• Workload methodology

Additional historical measures

Lawyer-record comparisons, BIA case-appeal dispositions, court in-absentia flags and entry-year completion snapshots are available under their own population definitions. Active counsel at the exact asylum decision, causal effects and personal forecasts are not claimed.

• Explore these measures

Geography

EOIR ZIP routing · retrieved 2026-09-14. The source does not state an effective date. This map illustrates ZIP routing with 2020 Census areas. Use the court result and official links for routing details. Court points are matched conservatively to official addresses. Unverified historical directory records remain labeled as such.

• EOIR operational directory and ZIP lookup

Known data defects

The source has known malformed lines: 357 physical lines were rejected; accepted logical rows are 262 below EOIR’s count file. Statistics describe accepted records. The official asylum report uses a different unit and will not exactly match these totals. The 2026-08 snapshot contains 4,526 proceedings with conflicting ASYL outcome codes across application rows; those are excluded. A prior-schedule attribution audit found disagreement with proceeding assignment in about 3.10% of judge records and 0.33% of court records in the five-year cohort. These uncertainties are not fixed by large sample sizes.

Why grant rates can differ

Compare the outcome field, counting unit and date before comparing percentages. AsylumCompass joins an ASYL application outcome to the same proceeding and uses that proceeding’s completion date. A case-level dataset can instead pair the last asylum application across the case with its last proceeding date. A formula counting Relief Granted versus Remove measures recorded case dispositions; neither label alone identifies an asylum grant or denial. These choices can change both the numerator and denominator, even for the same court and fiscal year.

• DDP field definitions and processing rules

Santa Ana FY2025: 12.93% and 11.7%

Our 2026-09-27 audit reproduced 12.93% from 688 ASYL grants and 4,633 denials in the 2026-08 EOIR source. On a pinned DDP case dataset, counting last ASYL outcomes gave 896 grants and 4,756 denials (15.85%). Applying a third-party site’s published case-outcome formula to cases with an asylum application instead gave 1,000 Relief Granted and 7,546 Remove outcomes (11.70%, displayed as 11.7%). Those 7,546 included 2,485 cases whose asylum result was Abandonment; 106 of the 1,000 relief outcomes did not record an asylum grant. This reproduces the displayed precision, not the site’s undisclosed full query: its lifetime total and current FY2024 figure did not reconcile. The audit documents the record-selection steps and these limits. None of these historical shares predicts an individual case.

• Integer counts, source hashes and audit limits
• Reproduce the three formulas with SQL
• Third-party formula checked on 2026-09-27

Calculation audits

These audits reproduce selected published metrics from the specified 2026-08 EOIR source edition and calculation rules. They do not independently verify that the source dataset captures every real-world event. Full source reproduction requires the pinned EOIR package; the public artifacts contain formulas, hashes and expected results without raw case records.

• Representation calculation audit
• Duration calculation audit
• In-absentia calculation audit
• Hearing schedule calculation audit
• Motion record calculation audit