The operational question
A home care agency serving Medicaid members works from two pictures of the same person. One is the agency's census: who is receiving services, under which plan, with which authorization. The other is the plan's eligibility roster: who the plan says is enrolled for the month. When the pictures disagree, the agency may serve someone the plan does not recognize, miss a new member, bill the wrong plan, or continue service after coverage ended.
This design asks how an agency can measure roster-to-census discrepancies and their operational impact without overstating what a monthly file can prove. The unit is the client-plan-month, not the client alone, because a person can change plans, gain retroactive eligibility, or appear in two files during a transition. The primary outcome is a classified discrepancy. Secondary outcomes include resolution time, service hours exposed, and claims held or denied.
The design is descriptive and audit-oriented. It cannot determine Medicaid eligibility, adjudicate a claim, or prove a denial's cause without claim-level evidence.
Enrollment, eligibility, and authorization are different
Enrollment
Enrollment is the plan's record that a person is a member for a period. It may be reported monthly with effective and termination dates, or as a snapshot. A person can be eligible for Medicaid but not enrolled in a plan, and can appear on a roster while a redetermination is pending.
Eligibility
Eligibility is the state's determination that a person meets the rules for a Medicaid category. The state agency, not the plan or the home care agency, makes that determination. Enrollment usually follows eligibility, but timing can lag, and a retroactive determination can change the picture after services are delivered.
Authorization
Authorization is the plan's approval of specific services, units, dates, and providers. It is not enrollment. A member can be enrolled and still lack authorization, or have authorization that does not match the schedule. Treating a roster as an authorization list misreads both.
Keeping the concepts separate is the foundation of the study. A discrepancy in one is not evidence of another, and the corrective action differs. A missing roster record may require a plan call, a missing authorization a service request, and an eligibility question the state or plan channel.
The roster as a data source
Plans typically provide member lists on a schedule set by contract or state requirements. The file may arrive as a spreadsheet, secure transfer, portal download, or API feed. It may include member identifier, name, date of birth, plan, effective date, termination date, aid category, and address. Fields vary by plan and state.
The roster is an administrative file, not a clinical record. Its quality depends on the plan's enrollment processing, the state's data exchange, and extract timing. A file dated the first may not reflect a change processed on the fifth, and a current-member file may omit people whose coverage ended retroactively. Record the file date, extract time, period, record count, and field layout.
Data.Medicaid.gov and the managed care enrollment report provide context about program enrollment and plan participation, but they are not substitutes for plan-level files. They help an analyst understand the landscape, not reconcile a member.
Matching keys and normalization
Deterministic and probabilistic matching
Deterministic matching uses exact agreement on a key, such as the state Medicaid identifier plus the plan member identifier. It is transparent and auditable but fails when an identifier is reformatted or a member changes plans. Probabilistic matching weights agreement across name, date of birth, address, and identifier fragments. It finds more matches but requires calibration and produces false positives.
A practical design uses deterministic matching first, then reviews unmatched records with a documented probabilistic or manual step. The rule should be written before analysis, with examples of accepted and rejected pairs, and should specify how nicknames, suffixes, hyphenated names, transposed dates, and leading zeros are handled.
What a match does not prove
A match means two records likely describe the same person. It does not prove eligibility for the whole month, authorization of the service, or correct billing, nor does it prove an unmatched record is ineligible. The person may be in another plan, have a new identifier, or appear under a different legal name.
False matches are a risk in households with similar names, among twins, and when identifiers are reused. A false match can hide a genuine discrepancy. Sample matched pairs and verify them against a second source.
Retroactive eligibility and plan changes
Medicaid eligibility can be determined retroactively, and plan enrollment can change mid-month. A roster snapshot may show a member active at month end while the census shows services under another plan earlier in the month. Neither record is necessarily wrong. The study needs a time dimension.
Represent each client-plan-month with effective and termination dates, and distinguish the snapshot date from the coverage period. If a member changes plans on the fifteenth, split the month into two plan periods. If eligibility is granted retroactively, earlier service dates may become covered after the fact. A month-end comparison will misclassify both situations.
This is a form of left truncation and right censoring. Services before the roster's first reliable date may have incomplete coverage history, and coverage continuing past the study window may be unresolved. Report client-months affected by mid-month change or retroactive determination, and analyze them separately rather than forcing them into matched or unmatched categories.
A discrepancy taxonomy
A useful taxonomy separates the type of divergence from its cause and consequence. Classify each client-plan-month as one of the following:
- Confirmed match: roster and census agree on identity, plan, and coverage month.
- Missing from roster: the agency has an active client-month but the plan file has no record.
- Not in census: the roster lists a member with no agency service that month.
- Plan mismatch: the member appears in both sources under different plans.
- Identifier mismatch: the person appears to match but the member identifier differs.
- Date mismatch: coverage or service dates do not overlap as expected.
- Authorization gap: enrollment matches but no valid authorization is recorded.
- Duplicate: the same client-plan-month appears more than once in one or both sources.
- Unresolved: evidence is insufficient to classify the record.
Each category needs an owner and resolution path. A missing roster record may be a plan lag, a disenrollment, or a keying error. A plan mismatch may reflect an election, assignment, or retroactive change. Do not collapse these into one error count, because the operational response differs.
Denominators, lag, and measures
The denominator is the count of client-plan-months in the frame. Report distinct clients separately so one long-stay client does not dominate. If service hours are used, report hours exposed and their share of total hours.
A discrepancy rate is the number of classified discrepancies divided by the denominator, usually per 100 client-months. Report numerator and denominator for every rate. A rate without a denominator is not a finding.
Lag is the time from the first observable signal to resolution. Define the start event, such as the roster file date or identification date, and the end event, such as plan confirmation or claim reprocessing. Report the median, 75th percentile, and the number still open at follow-up. Censoring matters here as in any time-to-event study.
Measure impact with care. Count service hours delivered while a discrepancy was open, claims held, and claims denied. Do not attribute every denial to the roster without reading the reason. A denial can reflect a missing authorization, coding error, or timely filing even when the roster also disagrees.
Governance, validation, and audit trail
Roster reconciliation involves identifiable health and enrollment information. Use the minimum necessary data, limit access to staff with a reconciliation role, and avoid placing full files on shared drives or personal devices. The HIPAA Privacy Rule governs use and disclosure, and the Security Rule's safeguards apply to the systems that store the files. The NIST Cybersecurity Framework helps organize access control, logging, and incident response; the GAO Green Book provides control language for ownership and corrective action.
Store each roster with its source, receipt time, reporting period, and version marker. Preserve the original file, match rule, query, classification, reviewer, and every correction. If a plan sends a corrected file, keep both versions and record which one supported each decision.
Sampling and validation
Reconciliation can be census-level for matching, but field accuracy and classification need validation. Draw a stratified sample by plan and category. Include confirmed matches, not only exceptions, so false or misclassified matches can be detected, plus records near month boundaries and mid-month changes.
Have a second reviewer independently classify the sample with the same taxonomy. Measure agreement, adjudicate disagreements, and record the reason for every change. If the sample estimates a population rate, report a confidence interval and account for stratification.
Validate the roster against a second source where possible, such as the plan portal, a state eligibility verification, or a remittance advice. Do not treat the agency's census as ground truth. It may contain stale plan assignments, duplicate clients, or services recorded under the wrong payer. The study measures agreement between sources, not conformity to an external truth.
Limitations and what the evidence can prove
The design can show the size and composition of roster-to-census disagreement in a period, the plans and categories most affected, and the operational exposure while records are unresolved. It can support a conversation with a plan about file timing or identifier quality, and identify internal steps that need better evidence, such as authorization checks or plan assignment at intake.
It cannot determine whether a member is eligible, whether a service is covered, or whether a claim should be paid. It cannot establish that a plan's file is wrong, only that it differs. It cannot prove a discrepancy caused a denial without claim-level review, and it cannot generalize to other agencies, plans, or states. One month is a snapshot, not a trend.
Those limits belong in the report, not a footnote. The design replaces anecdotal complaints about roster quality with a reproducible count, a clear taxonomy, and a documented resolution path. For related work, see Medicaid document collection and authorization aging. To discuss a controlled reconciliation workflow, contact us.

