Interpretation and source limitations
The unit is an application record in a processed USCIS source snapshot. A person may have multiple applications and repeated records. Neither all rows nor approved rows measures distinct people currently holding DACA or TPS. Counts of decisions by receipt year and counts by decision year answer different questions; unresolved applications and right censoring affect comparisons.
The source uses pseudonymous/redacted applicant identifiers. This pipeline makes no verified cross-form, cross-release or EOIR/ICE linkage claim. _source_row identifies only the physical row within a specific pinned table; it is not a persistent application identifier. file_original and row_original preserve the publisher's original-input provenance. The pair (file_original, row_original) is nonmissing and unique within each full pinned table; no cross-release identity guarantee follows.
Publisher duplicate flags are retained. duplicate_likely identifies repeated applicant/receipt-date groups where both values are available; the likely flag is NULL exactly when either applicant ID or receipt date is missing in the pinned processing rule. duplicate_last uses applicant (a unique placeholder per missing ID), receipt date, then input-file/input-row ordering for DACA; TPS additionally sorts decision date before input file/row. Later DACA reconstruction changes physical order, so _source_row does not reconstruct last-in-group. duplicate_drop_row is coalesce(duplicate_likely,FALSE) AND NOT duplicate_last. The importer checks this relationship but does not drop any flagged row. Unknown likely flags lead to false drop flags through coalescing; false does not establish absence of duplication. Analysts may choose a sensitivity analysis with that publisher recommendation; it does not establish a universally correct person count.
DACA form_type_filled_in copies original form_type for every ELIS row and every missing-ID row (including C3). Only data_source='C3' AND unique_identifier IS NOT NULL is reconstructed: 1,326,373 rows in this snapshot, versus 3,200,000 copied rows. Within applicant, the publisher orders by decision_date, then computes cumulative decision == "Approved" including the current row. Counts at most one yield DACA - Initial, larger counts DACA - Renewal, and unknown cumulative counts yield NULL. This replaces any original C3 label. A denied/pending renewal following the first approval can therefore remain Initial. Duplicate approvals also enter the cumulative sum; equal-date duplicate approvals can change the label of the last retained copy. Treat this as a heuristic, not verified application history. Compare original form type, source system and filled value. The exact pinned processing code is retained in process_daca.R and process_tps.R.
C3/CLAIMS3 and ELIS missingness differs sharply. All 1,326,378 DACA C3 and 3,402,409 TPS C3 birth-year values are NULL; only 1 DACA and 2 TPS ELIS birth years are NULL. DACA field office is NULL in 3,199,992 of 3,199,995 ELIS rows, and TPS in all 3,277,763 ELIS rows, while C3 has no field-office NULLs. These are system-coverage differences, not substantive negatives.
Country of birth is not residence. Processing field offices and service centers also do not identify applicant residence and should not be joined to ICE offices/AORs as a residence proxy. These files do not provide verified residence geography.
The publisher resolves legacy C3 birthplace codes with a lookup. DACA failed lookups become NULL (15 C3 NULLs; another 37 ELIS NULLs). TPS failed lookups default to Unlisted Entry (Any foreign country not listed). The processed C3 subset contains 698 rows with that label; the original-code breakdown cannot be recovered, and the label can include invalid/unresolved or missing raw codes as well as any explicitly mapped category. There are no stored TPS C3 country NULLs, while ELIS has 2,637 NULLs and 51,691 Unlisted entries. DACA ELIS also has 506 Unlisted entries. The processed files cannot recover the original-code breakdown. Do not count Unlisted as a known country or equate stored NULL counts with underlying missingness.
Dates and anomalies remain unchanged. The full dataset contains 15,853 DACA and 297,594 TPS rows with decision dates before receipt dates. All 149 DACA and 2,603 TPS rows flagged duplicate_drop_row=true remain present. There are 19,135/604 missing applicant identifiers, 19,135/612 unknown duplicate-likely flags, and 0/8 missing receipt dates respectively. DACA receipt dates reach 2012-08-05 and LPR dates 1987-08-22. TPS RFE dates reach 1999-11-16, earlier than its first receipt date 2000-01-01. TPS includes an upstream LPR date as early as 1918-04-12; this is retained as an unusual source value, not interpreted as a valid modern program event. The full manifest reports exact counts and ranges for every date column. Do not silently repair dates or calculate elapsed-time outcomes without assessing such records.
The snapshot has information through February 2026. A later Git commit, download or database build timestamp is not a newer observation period. Status fields describe the released snapshot, not current legal status or a complete longitudinal event history. Source coverage, redaction, FOIA selection, administrative recording and publisher processing can all affect analyses.
The dictionary's full-file NULL counts come from pinned Parquet footers and must match a full import. This importer preserves zero, empty string, redaction labels and NULL as stored in the processed source. Upstream transformations can already have merged original categories, including TPS birthplace fallback. The publisher removed fully redacted rows/blank columns, repaired ELIS UTF-8, recoded F/M/U to Female/Male/Unknown (Unknown is distinct from NULL), and harmonized TPS form labels. No imputation, general string cleanup or deduplication is added by this importer. The included pilot is the first rows in source order and is unsuitable for prevalence estimates.
References: publisher codebook, source catalog, and the retained pinned DACA processing code and TPS processing code.
The unchanged source lock retains the pre-acquisition schema_evidence statement "full hash remains to be checked on acquisition". That historical preflight statement is superseded by the completed manifest's full_source_hashes_verified: true; the lock bytes are retained for build provenance. In the manifest, null_duplicate_flags refers only to NULL duplicate_drop_row; consult null_counts.duplicate_likely for unknown likely flags.