Census, ADC and patient days: getting the basics right

Census is the number every agency leader asks about first, and the one most often reported three different ways in the same meeting. Before any forecast, dashboard or AI model can help, the organization needs to agree on what it's counting. Here's how I define the core volume measures, and how they behave differently in Home Health, Hospice and Personal Care.

Three numbers that sound alike

  • Census is a snapshot: the number of patients on service at one moment, such as midnight on the last day of the month.
  • Patient days add up every day each patient was on service during a period. Two patients on service for 10 days each is 20 patient days.
  • Average daily census (ADC) is patient days divided by days in the period. It describes the whole period, not one moment of it.
One week, six patients
PatientDay 1Day 2Day 3Day 4Day 5Day 6Day 7Patient days
Patient 17
Patient 23
Patient 34
Patient 43
Patient 51
Patient 61
Census that day333222419
4End-of-week census (Day 7 snapshot)
19Patient days (sum of every day on service)
2.7Average daily census (19 / 7 days)
A Day 7 snapshot says census is 4. The week actually averaged 2.7. Three patients were admitted late in the week and two were discharged early, so the snapshot overstates the week by 47%.

Snapshots are easy to pull and easy to misread. A month-end census swings with the timing of admissions and discharges in the last few days. ADC smooths that out, which is why it's the better basis for staffing, trend lines and budgets.

The relationship that ties them together

Over a stable period, average daily census is roughly admissions per day multiplied by average length of stay. This is Little's Law from queueing theory, and it's the most useful sanity check in post-acute analytics:

It also tells you where growth has to come from. To raise ADC, you either admit more patients or keep them on service longer. In Hospice, length of stay is largely clinical and regulatory. In Home Health, it's bounded by need and by how 30-day periods are paid. In practice, admissions are the lever, which is why growth planning starts with the referrals and conversion rates needed to produce them.

How each service line counts

Home Health

Census is the number of patients with an open episode of care. Revenue under PDGM is paid per 30-day period, so periods per admission, recertifications and LUPA rates matter as much as headcount. Two agencies with the same census can have very different revenue if one discharges after a single period and the other recertifies appropriately.

Hospice

ADC drives staffing for the interdisciplinary team, and length of stay drives exposure to the aggregate cap. Average length of stay is skewed by a few very long stays, so I always report the median alongside it. Live discharges belong in their own measure: they reduce census like deaths do, but they carry very different clinical and compliance meaning.

Personal Care and CDS

Census counts active clients with a current authorization, but revenue follows hours delivered, not heads. The measures that matter are authorized hours, scheduled hours, worked hours and billed hours, and the capture rate between them. A client on census who receives no visits looks fine in a census count and costs you in every other measure.

Common mistakes that skew the numbers

  • Averaging ratios across branches. The average of ten branches' visits-per-admission isn't the organization's rate. Rebuild ratios from their components: total visits divided by total admissions.
  • Counting holds as active. Patients on hold for a hospital stay are often still "open" in the EMR. Decide whether they count and apply the rule everywhere.
  • Double-counting transfers. A patient moving between branches or from Home Health to Hospice can appear twice on the same day without a clear transfer rule.
  • Mixing day-count conventions. Counting both the admission day and the discharge day adds a day per patient. Pick a convention, such as counting the admission day but not the discharge day, and document it.
  • Comparing snapshots to averages. A month-end census compared with last year's ADC will show growth or decline that never happened.

Why this matters for AI

A forecasting model trained on a census that sometimes includes holds and sometimes doesn't will learn the inconsistency, not the business. Every agent I build starts from agreed definitions in a governed data model. That's less exciting than the model itself, and it's the part that decides whether the model can be trusted.