Evidence-first guide · Workforce intelligence
How providers use analytics to improve retention
Educational update October 9, 2026
- First control
- Stable metric definitions
- Decision owner
- Named human manager
- Review unit
- Closed employee cohort
- SeniorCRE status
- Pre-production
Repeatable process
A method a team can run and review
- 01
Establish baseline definitions
Freeze the period, employee identity, role, unit, voluntary separation, transfers, average active headcount and closed-cohort rules. Keep agency workers outside employee retention denominators.
- 02
Reconcile payroll and scheduling
Join schedule, time, payroll and agency invoices by worker and covered shift. Separate worked hours from paid leave, corrections and premiums; retain unresolved mismatches rather than averaging them.
- 03
Identify exposure
Compare reconciled coverage with the approved demand rule. Inspect overtime concentration, open shifts, pay errors and early-tenure continuity; a signal is not an employee judgment or a fitted prediction.
- 04
Assign human review
A named manager confirms, defers or rejects each material exception. Suppress artifacts with a reason. HR authorizes employment-affecting decisions; clinical leaders accept coverage changes.
- 05
Measure interventions
Retain the accepted forecast before close. Compare it with reconciled actuals; log action completion, driver movement and closed-cohort retention. Record concurrent changes and unattributed results; before-and-after movement alone is not causation.
Synthetic worked example · not measured outcomes
Paid hours are not automatically care coverage
All records and counts below are invented for one week in a fictional 80-unit community. Payroll and scheduling initially show the same total, but paid leave changes what the total means. No employee records, operator observations or validated predictions are represented.
| Record / definition | Synthetic value | Authority and reconciliation |
|---|---|---|
| Posted employee schedule | 1,140 hours | Fictional schedule export, same closed week |
| Employee worked hours | 1,100 hours | Fictional approved time record; 40 fewer than scheduled |
| Payroll paid hours | 1,140 hours | 1,100 worked + 40 paid leave; paid leave is not care coverage |
| Agency covered shifts | 60 hours | Fictional invoice matched to shifts; outside employee headcount |
| Approved demand assumption | 1,260 hours | 72 occupied units × 7 days × 2.5 hours; not a mandated staffing level |
| Uncovered demand | 100 hours | 1,260 − (1,100 + 60); manager checks missing records before acting |
Review, not automatic intervention. Payroll confirms the 40 leave hours; the scheduler checks missing punches and agency matches; the clinical owner reviews the demand rule. If the 100-hour gap is real, operations assigns a coverage action with an owner and due date. HR handles any employee-level decision under policy.
Retention arithmetic is separate from evidence of an effect
- Voluntary separations: 2 ÷ 40 average active employees = 5.0% for the defined period.
- Closed day-90 cohort: 18 ÷ 20 hires retained = 90.0% survival.
- Completed actions: 6 ÷ 10 confirmed signals = 60.0% coverage—not intervention effectiveness.
These are independent fictional counts. Preserve cohort eligibility and small denominators; record concurrent wage, census, supervisor and schedule changes. No before/after comparison or causal effect is asserted.
2027 planning · assumptions, not results
2027 labor-budget worksheet
Six sheets separate assumptions, monthly coverage, employee wages and burden, agency invoices, blank reconciled actuals and retention review. Calendar demand uses all 365 days of 2027. This forward budget is distinct from the closed-week example above.
Illustrative inputs: 72 occupied units; 2.5 care hours per resident day; 1,140 scheduled hours weekly; 4% call-off allowance. Wage, agency rate, growth, burden and gap-fill mix remain editable assumptions. Scheduled hours exclude paid leave and gap-fill overtime; 40 weekly paid-leave hours add separate wages, not care coverage.
Download the 2027 labor-budget worksheet (XLSX)Formula, sources and limits
Required hours = resident days × approved staffing rule; effective employee hours = scheduled hours − rounded call-off allowance; coverage gap = max(0, required − effective); price regular, overtime, paid leave and agency separately; employer burden applies to employee wages only. Annual budget is the sum of monthly scenarios; blank actuals are not zero outcomes.
Synthetic teaching scenario, October 9, 2026. Wage is an illustrative value within the BLS-linked base-hourly range above; agency rate, burden, growth, staffing rule and coverage shares are invented assumptions, not market estimates. Overtime uses simplified regular-rate arithmetic; consult DOL Fact Sheet 54 and applicable state law before budgeting payroll.
One fictional direct-care role only. Not a full enterprise payroll model, staffing mandate, compliance calculator, forecast-accuracy result or savings claim. Excludes recruitment, orientation, bonuses, shift differentials, benefits outside the chosen burden, agency extras and non-care departments unless separately added. Regular rate, overtime eligibility, daily/state rules and any 8-and-80 arrangement require payroll/legal review.
Published exposure-method registryEducational reference points
- U.S. Bureau of Labor Statistics: JOLTS concepts and definitions
Defines hires, quits and other separations for aggregate labor statistics. A local closed-cohort survival metric is a separate operator definition; national rates do not predict individual exits.
- U.S. Department of Labor: Fact Sheet 54: Health care overtime
Explains regular-rate inclusion of differentials and nondiscretionary bonuses, and qualifying 8-and-80 arrangements. The worksheet is a simplified budget, not an overtime compliance calculation; state law can impose additional requirements.
- Centers for Medicare & Medicaid Services: Payroll-Based Journal staffing data submission
Explains staffing submissions based on payroll and other auditable records for applicable nursing facilities. PBJ obligations must not be generalized to every independent-living or assisted-living community.
These sources support definitions and reporting rules, not SeniorCRE outcomes. Structured markup describes the visible article; it cannot guarantee a citation.
Measurement design
Start with definitions that survive a board review
Retention analytics fails when numerator, denominator, cohort, and event rules change between HR, finance, and operations. Freeze the definitions first, version changes prospectively, and preserve the source behind every reported result.
Voluntary turnover
Voluntary separations divided by average active headcount, with involuntary exits and same-owner transfers separated.
90-day survival
A closed hire-date cohort measured at day 90, rather than a moving snapshot that changes for unrelated reasons.
Intervention coverage
The share of confirmed risk signals that received a named, logged action—not the share merely shown on a dashboard.
Override rate
Manager overrides reviewed as a data- and model-quality signal, never as a manager-compliance score.
Leading indicators
Measure operating friction before the resignation
Lagging turnover rates explain what already happened. Leading indicators should be observable, job-related, and tied to conditions a manager can inspect. Protected characteristics, private communications, health information, and personal-device data do not belong in the feature set.
Schedule strain
Overtime concentration, consecutive days, clopen patterns, swaps, no-shows, and open-shift exposure by role and shift.
Early-tenure stability
Orientation completion, preceptor continuity, unit consistency, and day-14, day-30, and day-90 milestones.
Workload and continuity
Acuity-informed workload and primary-caregiver continuity, with the demand definition and clinical boundary visible.
Earnings consistency
Pay-period variance, premium concentration, and differential errors compared with the employee’s own recent history.
Intervention governance
The steps that stand between a score and an employee
Governance is the part of retention analytics that survives audit. Each step below is a recorded decision with a named owner, so an intervention can be explained months later without reconstructing it from memory.
1. Signal raised
The contributing conditions, the source records, and the timestamp are shown with the signal. A score with no visible inputs is not actionable.
2. Suppression check
Known data artifacts — a payroll correction, a swapped badge, a unit re-map — are suppressed and the suppression is logged rather than silently dropped.
3. Human confirmation
A named manager confirms, defers, or rejects. Confirmation is required before any intervention, and deferral and rejection are first-class outcomes.
4. Cause classification
The confirmed signal is classified as schedule strain, early-tenure instability, workload, pay error, supervisor change, or unclassified. Unclassified stays unclassified.
5. Authorization
Anything affecting pay, schedule, assignment, or employment status is authorized under the provider’s HR policy, with legal review where policy requires it.
6. Action recorded
The intervention, its owner, its date, and its completion state are written back to the record alongside the signal that prompted it.
7. Structural escalation
When several signals share one driver, a single structural staffing item is opened with an accountable owner, kept separate from the employee-level decisions.
8. Review and disclosure
Override rate, suppression rate, coverage, and cohort results are reviewed on a fixed cadence, and unattributed cases are reported as unattributed.
Senior housing example
When three caregivers on one shift are flagged
Do not start three generic retention conversations. First test whether the unit has the same overtime, open-shift, acuity, or supervisor-change pattern. If the common driver is confirmed, open one structural staffing item with an accountable owner, while retaining each employee-level decision separately.
Rolling 13-week workflow
A 13-week retention-analytics cycle, run alongside the labor forecast
Retention work and payroll forecasting share a horizon so the same reconciled record serves both. One quarter of the cycle is shown below; the horizon then re-extends to a full thirteen weeks.
| Weeks | What the team does |
|---|---|
| Weeks 1–2 | Freeze definitions: voluntary versus involuntary separation, the active-headcount denominator, transfer handling, and the cohort close rule. Publish them to HR, finance, and operations together. |
| Weeks 3–4 | Reconcile the workforce record across HRIS, scheduling, time, and payroll: employee identity, unit assignment, tenure, hours, earnings, and separation reason. |
| Weeks 5–6 | Establish the leading-indicator baseline on the provider’s own history — overtime density, consecutive days, swaps, no-shows, open-shift exposure, earnings variance, early-tenure continuity, supervisor changes. |
| Week 7 | Stand up the confirmation path before any score reaches a manager: who confirms, who defers, who rejects, what is suppressed, and where each disposition is recorded. |
| Weeks 8–11 | Run weekly confirmed-signal review. Match each intervention to the observed operating cause, open structural staffing items separately from employee-level decisions, and log every action with an owner and a date. |
| Week 12 | Review driver movement and intervention coverage, and review override rate as a data- and model-quality signal rather than a manager score. |
| Week 13 | Close the cohorts due this quarter and report day-14, day-30, and day-90 survival, voluntary separation, and agency share — with attributed, unattributed, suppressed, and overridden cases shown separately. |
Practical checklist
Retention analytics readiness checklist
Work through this before a risk signal reaches a manager. An unchecked item belongs in the disclosure, not in a dashboard footnote.
- Voluntary and involuntary separation are defined separately, and transfers are preserved as their own event.
- The headcount denominator and the cohort close rule are written down and dated.
- Employee identity reconciles across HRIS, scheduling, time, and payroll.
- The feature set is documented, and protected characteristics, health information, private communications, and personal-device data are excluded.
- Leading indicators were tested against the provider’s own separation history, not assumed from a benchmark.
- A named manager confirms, defers, or rejects every signal before any intervention.
- Data artifacts are suppressed, and suppression is logged rather than silently dropped.
- Each intervention type has a defined trigger, owner, and record of completion.
- Structural staffing causes are opened as their own item, separate from employee-level decisions.
- Override rate is reviewed as a model- and data-quality signal, never as a compliance score for managers.
- Employment-affecting actions pass the provider’s HR policy and legal review.
- Reported results separate attributed, unattributed, suppressed, and overridden cases, and avoid claiming causation from retention movement alone.
Metric definitions
Definitions that keep retention analysis comparable
Freeze each formula, numerator and denominator rule, and known error before comparing communities or cohorts. These are the same definitions provided in the downloadable methodology.
| Metric | Formula | Numerator and denominator rule | Most common error |
|---|---|---|---|
| Annualized voluntary turnover rate | (voluntary separations in period / average active headcount in period) × (12 / months in period) | Numerator excludes involuntary terminations, RIFs, and internal transfers. Denominator is average active headcount, not period-end headcount. | Mixing voluntary and involuntary separations, or using period-end headcount, which inflates the rate in a shrinking portfolio. |
| 90-day new-hire survival | hires still active on day 90 / hires in the cohort | Cohort is defined by hire date, not by month of report. Rehires within twelve months are flagged separately. | Reporting a rolling snapshot instead of a closed cohort, which makes the number move for reasons unrelated to retention. |
| Regrettable turnover | voluntary separations flagged regrettable / total voluntary separations | Regrettable requires a documented manager designation at separation, not a retrospective judgment. | Backfilling the flag after the fact, which makes the metric unauditable. |
| Agency share of worked hours | agency worked hours / (agency worked hours + employee worked hours) | Agency hours come from invoices reconciled to the schedule, not from the schedule alone. | Using scheduled rather than invoiced agency hours, which understates share whenever cancellations are billed. |
| Overtime density | OT hours / total worked hours, by role and shift | Computed at role-and-shift level; portfolio averages hide the concentration that drives exits. | Reporting a single portfolio OT percentage. |
| Vacancy duration (time-to-fill) | median days from requisition open to first worked shift | Ends at first worked shift, not at offer acceptance. | Measuring to offer acceptance, which omits the no-show and orientation-attrition window. |
| Primary-caregiver continuity | resident-days covered by the top-two assigned caregivers / total resident-days | Measured per resident, then rolled up. Agency coverage counts against continuity. | Measuring at the unit level, which averages away the residents with the worst continuity. |
| Cost-to-fill an open shift | incremental premium (agency rate differential + OT premium + guarantee + travel/orientation) for one covered shift | Incremental cost only — base hours the operator would have paid regardless are excluded. | Charging the full agency invoice as incremental, which overstates avoidable cost. |
| Fully loaded cost per separation | separation + vacancy coverage premium + recruiting + onboarding + ramp productivity loss | Built from the operator’s own payroll, invoice, and recruiting cost lines; industry ranges are used only to sanity-check. | Adopting a published per-separation figure as if it were the operator’s own cost. |
13-week forecast example · illustrative
How a labor forecast can surface retention pressure before an exit
This fictional example follows one caregiver role at one community. It does not predict an employee exit. It shows how forecasted coverage pressure can become a reviewable operating signal while the rolling horizon extends each week.
| Week | Required hours | Scheduled hours | Open hours | Forecast review |
|---|---|---|---|---|
| 1 | 1,260 | 1,214 | 46 | Freeze the baseline; confirm census and staffing-rule versions |
| 2 | 1,260 | 1,206 | 54 | Review overtime concentration by employee, role, and shift |
| 3 | 1,278 | 1,208 | 70 | Confirm whether census, vacancy, or schedule changes explain the gap |
| 4 | 1,278 | 1,198 | 80 | Separate data corrections from operating variance |
| 5 | 1,296 | 1,202 | 94 | Review consecutive-day and open-shift exposure |
| 6 | 1,296 | 1,216 | 80 | Confirm, defer, or reject any retention signal with named ownership |
| 7 | 1,296 | 1,228 | 68 | Record approved schedule, preceptor, or coverage actions |
| 8 | 1,314 | 1,236 | 78 | Compare the retained forecast with actual worked and paid hours |
| 9 | 1,314 | 1,248 | 66 | Reconcile agency invoices and overtime by shift |
| 10 | 1,314 | 1,260 | 54 | Review whether the targeted operating driver moved |
| 11 | 1,332 | 1,270 | 62 | Escalate shared unit-level causes as structural items |
| 12 | 1,332 | 1,284 | 48 | Review forecast bias, overrides, suppressions, and action coverage |
| 13 | 1,332 | 1,292 | 40 | Close due cohorts; retain unattributed outcomes and re-extend the horizon |
Illustrative only. Every number here is invented to show the arithmetic; none is drawn from an operator, a SeniorCRE deployment, or a benchmark. Required and scheduled hours are planning inputs, not promised staffing levels or a measured SeniorCRE result. No savings or retention outcome is inferred from the declining open-hour assumption.
How operators evaluate retention analytics
Questions to put to any retention-analytics vendor
Care operators evaluating retention analytics should test the method, not the dashboard. The questions below apply to any vendor, including SeniorCRE.
| Question | A strong answer | A warning sign |
|---|---|---|
| How is turnover defined? | Voluntary and involuntary separated; transfers kept distinct; denominator stated | One blended turnover percentage |
| Which inputs drive a risk signal? | Named job-related inputs shown beside every signal | An opaque score with no visible inputs |
| Are protected characteristics excluded? | Documented exclusion of age, race, pregnancy, disability, and proxies, with a bias review | No answer, or "the model decides" |
| Who acts on a signal? | A named manager confirms, defers, or rejects before any action | Signals trigger actions automatically |
| How is accuracy reported? | On the operator’s own closed cohorts, with false positives and small denominators disclosed | A headline accuracy figure from other customers |
| How are outcomes attributed? | Only when the intervention and changed condition are documented; unattributed cases reported | Retention gains credited to the tool by default |
| What is live today? | Capabilities labeled production, tested, or planned, with dates | Plans described as features |
Staffing data requirements
What data a forecast or retention review needs, and where it usually lives
Before evaluating any tool, confirm each field below exists, has one declared source, and reconciles with the others. A missing field is a disclosure in the output, not something to model around.
| Data | Typical source | Needed for | Minimum history | Common problem |
|---|---|---|---|---|
| Employee identity and role | HRIS | Forecast and retention | Current plus 12 months | Duplicate IDs across communities |
| Scheduled hours by shift | Scheduling system | Forecast and retention | 12 months | Schedule edited after the fact |
| Worked hours (punches) | Time and attendance | Forecast and retention | 12 months | Missed punches, manual edits |
| Paid hours and pay components | Payroll | Forecast | 12 months | Corrections posted in later periods |
| Pay rates and differentials | Payroll / HRIS | Forecast | Current plus rate history | Differentials coded inconsistently |
| Agency hours and invoices | Agency invoices / AP | Forecast | 12 months | Invoices not matched to shifts |
| Census and resident days | Clinical record or property system | Forecast | 12 months | Two systems report different census |
| Approved staffing rules | Operator policy | Forecast | Current, versioned | Unwritten or community-specific |
| Hire, transfer, and separation events | HRIS | Retention | 24 months | Transfers recorded as separations |
| Separation reason (voluntary / involuntary) | HRIS | Retention | 24 months | Reason field left blank |
Minimum history is a starting guide, not a rule; small communities may need longer windows before a pattern means anything.
Measurement plan
How to measure retention analytics without overstating results
Agree the measures and the attribution rule before the first signal reaches a manager.
| Measure | Definition | Cadence | Owner | Guardrail |
|---|---|---|---|---|
| Voluntary separation rate | Voluntary separations ÷ average active headcount | Monthly | HR | Report by community; disclose small denominators |
| 90-day survival | Share of a closed hire cohort still employed at day 90 | Per cohort | HR | Never use a moving snapshot |
| Intervention coverage | Confirmed signals with a logged action ÷ confirmed signals | Weekly | Operations | Count actions, not signals shown |
| Override rate | Signals rejected by managers ÷ signals raised | Monthly | Data owner | Read as a model-quality signal, never a manager score |
| Driver movement | Whether the targeted condition (e.g., overtime density) changed | Per intervention | Operations | Report unattributed outcomes separately |
These measures describe method only. SeniorCRE reports no measured operator retention outcome.
Open methodology
Download the definitions and method
Use the same published definitions in analysis, review, and procurement. The files are ungated and machine-readable.
- SeniorCRE Workforce Intelligence methodology (JSON)
JSON
Machine-readable methodology: model inputs and exclusions, the 30/60/90-day intervention workflow, retention metric definitions, ROI formulas with variable definitions, sensitivity treatment, limitations, and the cited case-study format.
- Retention & workforce metric definitions (CSV)
CSV
Metric, formula, numerator/denominator rules, and the most common measurement error for each retention and labor metric used in the methodology.
Evidence boundary
SeniorCRE® describes pre-production workforce capabilities and does not publish operator-production outcomes, proven turnover reduction or model accuracy. The operating record is designed and not yet implemented in any community. Synthetic arithmetic teaches a method; it is not evidence that the product performs it. Methods here can be used independently; reliance requires source reconciliation, security and HR review, workflow validation and written operator acceptance.
Written by SeniorCRE, LLC. Review accountability: Clinical Operations Lead · Workforce Analytics Lead · Data Governance Owner. Named individuals and credentials are pending approval; no independent clinical or HR review is claimed. Editorial update: October 9, 2026.
Read the Workforce Evidence RecordSources
Where the public reference points come from
Public sources below describe reporting requirements, definitions, and labor context. They are cited for method and definition, not as evidence of any SeniorCRE outcome. ReviewedOctober 9, 2026; each publisher may revise its material after that date.
- Staffing Data Submission Payroll-Based Journal (PBJ)
Centers for Medicare & Medicaid Services
States that payroll-derived staffing data, combined with census, is used to report staffing levels as well as employee turnover and tenure. A useful anchor for why turnover definitions must be reconcilable to payroll.
- Electronic Staffing Data Submission Payroll-Based Journal policy manual, version 2.8 (August 2026)
Centers for Medicare & Medicaid Services
Defines employee records, staffing hours, labor and job codes, and census for federal submission — the same fields a retention cohort depends on.
- Job Openings and Labor Turnover Survey (JOLTS)
U.S. Bureau of Labor Statistics
National quits, hires, and separations series including health care and social assistance. Context for external labor conditions; it does not describe any single provider.
- Employment tests and selection procedures
U.S. Equal Employment Opportunity Commission
Federal guidance on the use of tests and selection procedures in employment decisions. Cited because a predictive score that influences an employment action is subject to employment law, not analytics preference.
Natural-language questions
Frequently asked questions
How do healthcare providers use analytics to improve retention?
They define retention consistently, reconcile scheduling and payroll records, monitor job-related leading indicators, require a manager to confirm any signal, match interventions to the observed operating cause, and compare closed cohorts over time. Analytics supports the decision; it does not replace accountable human judgment.
What are the best leading indicators of healthcare turnover?
Useful operational indicators include overtime density, consecutive-day and clopen patterns, shift swaps and no-shows, earnings variance, early-tenure continuity, open-shift exposure, workload, and supervisor-change events. Their usefulness must be tested against the provider’s own history.
Should a turnover-risk score automatically trigger an employee action?
No. A named manager should confirm, defer, or reject the signal, inspect the contributing conditions, and document the decision before any intervention. Employment actions require appropriate HR policy and legal review.
How should a senior housing operator measure whether an intervention worked?
Track whether the targeted driver changed, whether an action was completed, and what happened to the closed employee cohort. Report attributed, unattributed, suppressed, and overridden cases separately; do not infer causation from retention alone.
Commercial readers
Bring the definitions, exceptions and accepted budget
The Workforce Intelligence hub separates method, capability, evidence, and what remains unproven.