New Mexico Health Data: Check 4 Fields Before Statewide Release

42.5% is the first reason not to trust a single statewide completeness display. In RAND's Oregon Medicaid claims data, race and ethnicity was missing for that share of records; in the companion electronic health records, missingness was 4.9%. Those findings are not New Mexico rates, but they expose a release-control problem: completeness can vary sharply by source, and a statewide average can conceal the exact cell that fails a cap.

Before release, the required fields need row-level usability checks, crosswalk validation, plan-by-county completeness checks, and threshold decisions. The supplied evidence does not identify the four field names or provide a New Mexico numerator, denominator, missingness rate, benchmark, or statewide result, so none can be declared passing. A displayed 99% should not be treated as approved until the underlying rows and cells are evaluated.

That distinction matters because a handful of unusable keys can be misclassified as members or drop people from person-level rates, distorting counts and outcomes. The Oregon study also shows why imputation is not a neutral repair: surname-based estimates were available for more than 99% of records, yet BISG and incomplete self-report analyses produced substantially different disparity estimates in the claims data. Use such methods only with documented validation, and release only after field, row, and cell gates all pass.

Four Fields, 33 Counties

The release test is not “mostly complete.” It is a denominator-first audit of exactly four field groups, with any ID or date failure quarantined and any failed race/ethnicity or county rate held. I first pin the applicable 2026 New Mexico Health Care Authority technical specification—title, version, effective date, and cell definition—in the run log. The thresholds below are my internal preflight. If the 2026 state requirement is stricter in any scope, that stricter requirement controls.

The supplied SOURCE DATA review identifies no documented 2026 New Mexico reporting mandate, New Mexico-specific numerator, denominator, missingness rate, or statewide result. I therefore do not present this internal preflight as a verified legal mandate or observed state performance. According to that review, RAND’s Oregon/OCHIN estimates are not New Mexico baselines and are not imported as such.

For every group, I classify null, blank, and placeholder entries as syntactic missingness; “Unknown,” “not documented,” and other noninformative labels as semantic missingness; and present-but-invalid values as a third class. “Unknown” is not rescued because it is a string, and a county label is not rescued because it is spelled correctly. For field g, missingness is U/E, where E is the eligible-record denominator fixed at the stated scope and U is the unusable count. I also report C/E, where C counts rows for which all four groups are usable; that diagnostic never cancels a failed field gate. Every displayed rate remains attached to its numerator and denominator.

The paired race–ethnicity element retains race and ethnicity as distinct source values before any approved categorization. According to JMIR Public Health and Surveillance, federal standards provide limited guidance for ambiguous or multi-select responses; I document the mapping rather than silently forcing a response. County means the patient’s residence, never clinic or laboratory location. A valid service/result date outside the applicable window controls eligibility under the measure’s period rule; it is not mislabeled as missing.

The patient/member-ID audit runs inside the controlled source environment. Only an approved crosswalk key, pseudonym, or aggregate may leave it; direct identifiers and row-level audit traces remain inside. Failure extracts carry field labels, unusable counts, eligible counts, and scope labels—not names or medical record numbers.

At close, I evaluate statewide results and every plan-by-county-by-period cell independently. A statewide pass cannot offset a cell breach. If ID/date is nonzero, I quarantine the failures; if race/ethnicity or county misses either its statewide or cell limit, I hold the failed rate. I do not impute. The file is release-ready only when every gate passes and the all-four-complete count preserves the numerator/denominator trail.

Federal Benchmark Is Not a 2026 New Mexico Result

A federal completeness benchmark is not a release certificate. The supplied evidence does not establish a benchmark figure for this file. Any federal comparator would have to be verified separately and still could not establish its ID, date, county, or plan-by-county-by-period results. Similarity to a federal reference cannot replace the article’s measured release tests.

According to the U.S. Census Bureau’s 2020 Decennial Census, New Mexico’s resident count establishes the scale of the state, not the denominator of an ambulatory or laboratory extract. The eligible denominator must be built from records eligible under each field’s audit rule. Combining residents with healthcare records would answer a population-coverage question, not a field-completeness question.

The Census ethnicity and race-alone results also require separate treatment. Hispanic or Latino is an ethnicity, while American Indian or Alaska Native is reported as a race alone; those categories can overlap. I preserve both subvalues rather than forcing a mutually exclusive classification that the source did not create. Recoding one into the other could alter both the measured numerator and its eligible denominator.

The State of New Mexico’s managed-care contract award identifies a narrow plan universe, but statewide pooling still can conceal a plan-specific failure. Each plan’s records need the same denominator-first audit, followed by county and period-specific cell checks. A reassuring statewide result cannot rescue a failed required cell.

I place the external references in this context panel only:

The release table is a different evidentiary object. For every required field group, it must expose the measured unusable numerator, eligible denominator, source-system result, and plan-by-county-by-period result. Because no verified extract measurements were supplied, those cells must read “not evaluated,” not “pass.” Populate them from the current file and apply the article’s statewide and cell-level gates. Quarantine ID or date failures, hold failed race/ethnicity or county rates, and never impute. That procedure—not an external benchmark—determines release readiness.

Five Missingness Methods

The most defensible missingness method is also the least inferential: preserve what was reported, quarantine unsafe records, and hold noncompliant cells. I choose a release method with three tests: whether it preserves person and geographic denominators, keeps observed values distinct from inferred values, and exposes missingness by source and plan-by-county-by-period cell. A method that fails any test is not release-ready, regardless of how complete its statewide average appears.

The gate wins because the failure costs are not interchangeable. A bad key can corrupt linkage; a bad date can move a service into the wrong measurement period; demographic or geographic gaps can erase small groups. One averaged completeness score conceals those different mechanisms. The external evidence also cautions against promoting estimates to observations: according to RAND publication EP69080, BISG estimates had concordance of 0.87–0.95 with self-reported identities in the studied data, yet BISG estimates and incomplete self-report produced substantially different disparities in almost half of 22 claims measures. That research is not New Mexico-specific, so I use it only as methodological evidence—not as validation of this extract or its fields.

Race/ethnicity collection design can also make missingness nonrandom. A 2025 analysis in JMIR Public Health and Surveillance identifies ambiguous “other race” options, select-all-that-apply directions, and open-ended fields followed by a request specification as problematic response-design features. The authors note that such designs can contribute to nonrandom missingness and misclassification; recoding those responses as valid categories would therefore conceal rather than solve the defect.

I assign status at two levels. A row passes or is quarantined for an ID or date failure. A cell passes or is held when its applicable race/ethnicity or county cap is breached. A race/ethnicity or county defect is not repaired by deleting the row, and the literal value “Unknown” remains preserved in the source while counting as unusable.

With every decision, I archive the metric version, denominator definition, source-system list, extraction timestamp, exception rationale, and applicable 2026 specification version. Before release, I rerun the row, statewide, and cell checks from that archived extract; a statewide average cannot override a quarantine or hold.

What the Data Doesn't Tell You

A populated cell is not necessarily a usable cell. That is the failure mode a raw completeness percentage can hide: a copied member ID, a date shifted across year-end, race or ethnicity inferred from a name, or a county set to the laboratory location all look complete. I treat those entries as unusable under the release rule, not as a quality bonus. ID and date defects are quarantined; failed race/ethnicity and county rates are held, and no imputation rescues them.

Conversely, an apparently complete, zero-missing result can be the warning. If one source writes the same default code whenever a value is absent, nulls have become false precision. Before trusting a rate, I examine distinct-value counts, code frequencies, and source-level distributions, then test whether each populated code has valid provenance and meaning. Completeness is necessary, but it does not establish that the values are distinct or correct.

Statewide averaging can conceal where the process actually fails. An apparently low rate may be concentrated in one managed-care plan, EHR, laboratory interface, rural service area, or demographic group, leaving affected records systematically undercounted. The statewide result therefore cannot replace the plan-by-county-by-period checks. The release gate fails at the offending cell rather than being repaired by a reassuring aggregate.

These are illustrative comparisons, not New Mexico estimates. They show the denominator trap: matching rates can carry very different exposure to an additional failure. Every rate should be released with its count, denominator, and an uncertainty interval; rounding a small cell does not remove its instability.

If absence is related both to the measured value and the source process, reweighting or multiple imputation based on missing-at-random assumptions can remain biased. The assumption is substantive, not a diagnosis produced by software. The external evidence supports that caution: according to RAND publication EP69080, self-reported race and ethnicity values were missing in 42.5% of Oregon Medicaid claims and 4.9% of electronic health records; BISG-based Black-White disparities were generally larger than disparities based on the incomplete claims data. These are not New Mexico thresholds, but they demonstrate source and measurement dependence. Modeling belongs only in a separately labeled sensitivity analysis and cannot clear the release gate.

Finally, a pass is deliberately narrow. It covers only crosswalkable patient/member ID, service/result date, race/ethnicity, and county of residence under the schema pinned for the 2026 extract. It does not validate provider identifiers, clinical values, eligibility, full schema conformance, HCA acceptance, or legal compliance. The concrete action is to label a passing result as report-ready for these four fields under the pinned schema—not as a general claim that the file is clinically correct, formally accepted, or legally compliant.

An HCA Drill

The drill uses a hypothetical quarterly planning extract. Its row-level classifications are a worked simulation, not observed HCA performance or New Mexico results.

In the worked extract, patient or member ID is classified as usable across the simulated rows, meeting the internal identifier rule, while service or result date contains unusable entries and therefore violates the internal date rule.

Race or ethnicity contains both null and Unknown entries. The simulation classifies the statewide result as passing and Contract A–County A as failing, so the failed cell is held.

County of residence contains blank and Unknown entries. The simulation classifies the statewide result as passing and Contract B–County C as failing, so the failed cell is held.

The worked simulation is classified as not report-ready: date failures are quarantined, both offending cells—Contract A–County A for race/ethnicity and Contract B–County C for county—are held, all Unknown values are preserved, and no imputation is performed; release depends solely on corrected field-and-cell results, not statewide averages.

Five 2026 Rules

For the New Mexico Health Care Authority's 2026 ambulatory/laboratory extract, my release decision is binary. I run the gates in a fixed order and treat null, blank, placeholder, unknown, and invalid entries as unusable. Clean statewide totals never excuse a defective row or failed cell.

Rule 1 — I choose fail if one eligible row lacks a usable, crosswalkable patient or member ID. An identifier that cannot resolve to the intended person or member through the pinned crosswalk leaves the row unlinkable, so it cannot enter a valid person-level statewide numerator or denominator. I quarantine the row and preserve its exception reason; I never create a surrogate identifier.

Rule 2 — I choose fail if one eligible row lacks a usable service or result date or falls outside the pinned 2026 period definition. I apply the date role declared by the schema and measure: a collection date cannot silently replace a required result date. No inferred, backfilled, or ambiguous date is admitted, so the row is quarantined.

Rule 3 — I choose hold for race or ethnicity when unusable values exceed an applicable ceiling in a verified specification. A source-reported value of Unknown remains Unknown and counts as unusable; it must not be overwritten from name, language, geography, or another member record. A rate exactly at its applicable ceiling passes.

Rule 4 — I choose hold for county when unusable values exceed an applicable ceiling in a verified specification. The qualifying field is patient residence, not service location: neither a facility county nor a provider address can fill a missing residence county. I do not geocode an address or convert service geography into residence evidence.

Rule 5 — I choose report-ready only when all four field gates and every plan-by-county-by-period cell pass on the pinned 2026 schema. The release package must store the numerator, denominator, source list, extraction timestamp, and exception log, together with a versioned field map, contract roster, county crosswalk, period rule, and unusable-value dictionary that reproduce the decision. Any ID/date failure is quarantined, any failed race/ethnicity or county rate is held, and no value is imputed. Otherwise the status is not report-ready—not provisionally passed.

According to RAND publication EP69080, the study assessed 22 quality and utilization measures. That breadth does not certify a 2026 extract: however many measures an extract supports, release approval still depends on the same reproducible field-level and cell-level gates.

What to do next