Prevalence (Point, Period, and Annual) in RWE

In plain language

Prevalence answers the question 'what share of a population has this condition right now (or at some point during a time window)?' You pick everyone in your defined group, count who among them is a case, and divide. Whether you count people who are cases on one specific date (point prevalence) or anyone who is a case at any moment during a whole year (period prevalence) changes the number — so the window must be stated up front. One honest caveat: prevalence mixes people who got the disease recently with people who have had it for years, so it cannot tell you whether the disease is becoming more or less common.

Prevalence

is the proportion of a defined population that has a condition, in contrast to incidence, which counts new onset. In real-world data the headline number is governed almost entirely by two specifications that protocols routinely leave implicit: the time frame of the numerator/denominator and the case-finding rule. Getting either wrong changes the estimate by a factor of two or more, so both must be written into the estimand before any code is run.

Core conceptual distinction

. Three time frames produce three distinct estimands, and they are not interchangeable. (1) Point prevalence is a cross-sectional snapshot on a single index date: the numerator is people who are prevalent cases on that date and the denominator is people observable (continuously enrolled) on that date. (2) Period prevalence counts anyone who is a case at any moment during an interval; the denominator is people observable at some point in the interval, and because the window admits both surviving and incident cases it is always ≥ point prevalence for the same population. (3) Annual prevalence is period prevalence with the interval fixed to a calendar year — the operational standard for CMS Chronic Conditions Warehouse flags, BRFSS, and NHIS, chosen because it aligns with enrollment files and benefit years and is comparable across years. A second, orthogonal axis is how a case is identified: diagnosed (administrative) prevalence requires qualifying diagnosis codes; treated prevalence requires a qualifying dispensing/procedure and so measures the treated pool, not the diseased pool; diagnosed-and-treated intersects both. Treated prevalence systematically undercounts undiagnosed and untreated disease, which is exactly what makes it the right numerator for a budget-impact "treated population" but the wrong numerator for disease burden. Prevalence is a proportion (dimensionless, bounded 0–1), not a rate per person-time; if you find yourself dividing by person-years you are estimating incidence density, not prevalence.

Pros, cons, and trade-offs

. - vs incidence rate (`incidence-rate-calculation-rwe`): Prevalence is cheap, needs only a cross-sectional or windowed snapshot, and directly sizes a market or care burden. Cost: it confounds onset with survival — a therapy that prevents death raises prevalence, so prevalence is useless for etiologic questions and treacherous for "is the disease getting more common?" Prefer incidence for causation and trend-in-onset; prefer prevalence for resource planning and current burden. - vs treated prevalence as a proxy for true prevalence: A pharmacy- or procedure-based numerator is fully observable in claims and unambiguous, but it conflates disease occurrence with diagnosis and treatment access. In conditions with large untreated fractions (early CKD, mild OA, undiagnosed AF) treated prevalence can be a small and biased fraction of diagnosed prevalence. Use treated prevalence only when the treated pool is the quantity of interest. - vs longer diagnostic lookback to raise sensitivity: Extending the "ever-diagnosed" lookback (1 → 2 → 3 years) captures chronic prevalent cases whose only coded encounter predates the window, but it mechanically inflates the estimate, ages the prevalent pool, and makes cross-study comparison meaningless unless the lookback is held fixed. The single-claim vs ≥2-claims (Klabunde-style) rule is the same trade-off in reverse: one claim maximizes sensitivity but admits rule-out/coding-artifact diagnoses; two claims ≥X days apart improves PPV at the cost of sensitivity.

When to use

. Sizing an eligible or treated population for a budget-impact or commercial forecast; describing disease burden for an HTA submission or epidemiology section; CMS/HEDIS-style chronic-condition surveillance; any descriptive denominator where the question is "how many people currently have / are treated for X" rather than "how many newly developed X." Point prevalence for a snapshot ("on 1 July"), period/annual prevalence for a reporting interval, treated prevalence when the deliverable is the treatable market.

When NOT to use — and when it is actively misleading or dangerous

. - As a stand-in for incidence or risk. Prevalence ÷ duration ≈ incidence only under steady-state and never when survival or treatment differs across groups; using prevalence to infer onset or to compare arms is confounded by survival (prevalence–incidence / Neyman bias). For causal contrasts use incidence or cumulative incidence (`cumulative-incidence-risk-rwe`). - Comparing prevalence across data sources, years, or studies with different case rules or lookbacks. A "rising prevalence" can be pure artifact of a lengthened lookback, a coding-intensity change (ICD-9→ICD-10), or a shift in the diagnosed-vs-treated definition. Hold the operational rule fixed or the comparison is void. - Reporting a crude prevalence across populations with different age/sex structure. Crude prevalence is a weighted mash-up of stratum-specific prevalences; differences may be pure composition. Age/sex-standardize (`direct-standardization-rwe`, `indirect-standardization-smr-sir-rwe`) before any cross-population claim. - Treated prevalence presented as disease prevalence in a condition with substantial undiagnosed/untreated burden — this understates need and can mis-size a market by an order of magnitude.

Data-source operational depth

. - Claims (FFS vs MA): The denominator must be tied to observable person-time. For point prevalence, require active enrollment on the index date; for annual prevalence, require a minimum coverage fraction of the year (e.g., ≥11 of 12 months, or full-year continuous enrollment) so the numerator opportunity is comparable across people. The dominant failure mode is Medicare Advantage / capitated person-time that lacks fee-for-service claims: MA enrollees generate few or no FFS diagnosis/pharmacy claims, so they look disease-free and deflate prevalence — restrict the denominator to FFS-observable (Parts A/B and, for treated prevalence, Part D) person-time, or use encounter data where complete. Plan switching truncates the lookback and can drop the qualifying diagnosis; claims adjudication lag and reversals mean recent months are incomplete; the single-claim vs ≥2-claim rule and lookback length must be pre-specified and varied in sensitivity analysis. - EHR: Capture is encounter-driven, so prevalence is conditioned on contact with the system — patients who are well, who get care elsewhere (external-care leakage), or who churn out are differentially missing. Structured problem lists undercount; resolved/historical flags and copy-forward inflate. Diagnosed prevalence from EHR alone is a lower bound on a panel that actually visits. - Registry: Often the gold standard for the numerator (adjudicated cases, completeness within catchment) but the denominator (source population at risk) is frequently external census/enrollment data, so the numerator–denominator mismatch and catchment definition drive validity. - Linked claims–EHR–registry: EHR/registry sharpens case ascertainment while claims supply a clean, enumerable denominator and full pharmacy fills for treated prevalence — but only the linkable subset is analyzable, introducing selection that must be characterized before generalizing.

Worked claims example (annual diagnosed and treated prevalence of type 2 diabetes, calendar year 2024). (1) Denominator: every person with ≥11 months of continuous FFS Parts A/B enrollment in 2024 (add Part D if treated prevalence is required), restricted to FFS-observable person-time so MA-only members do not deflate the estimate. (2) Diagnosed numerator: a member is a prevalent diagnosed case if they have ≥2 claims with a T2DM diagnosis (ICD-10 E11.x) on ≥2 distinct dates ≥30 days apart, looking back from 31 Dec 2024 over a 24-month lookback (2023-01-01 through 2024-12-31) — the two-claim rule and 24-month lookback are the judgment-dependent thresholds and are each varied in sensitivity analysis (1-claim; 12-month vs 36-month lookback). (3) Treated numerator: a member is a prevalent treated case if they have ≥1 fill (`fill_date` in 2024, `days_supply` ≥ 0) of a glucose-lowering NDC during 2024. (4) Annual prevalence = (cases meeting the rule) ÷ (eligible denominator), reported as diagnosed and treated separately; treated < diagnosed by construction. (5) Because age structure differs across plans/years, report crude and directly age/sex-standardized prevalence with exact-binomial 95% CIs, and re-estimate point prevalence on 1 July 2024 (active enrollment that day) to show the point-vs-period gap. Preserve raw claim dates alongside the derived prevalent flag and audit pre/post counts.

Interpreting the output

A small health district with 200 enrolled residents reports point prevalence of hypertension on July 1, 2024 as 18.0% (36 / 200) and annual period prevalence for all of 2024 as 25.0% (50 / 200). The two estimates use the same 200-person denominator but different numerator windows.

(1) Formal interpretation. Point prevalence (18.0%) counts only residents who were active hypertension cases on the specific index date of July 1; the denominator is all 200 residents enrolled and observable on that date. Period prevalence (25.0%) counts anyone who was a case at any moment during calendar year 2024; it is always greater than or equal to point prevalence for the same denominator because it accumulates incident cases who were diagnosed after July 1 or who had prior diagnoses recorded at other points in the year. The 7-percentage-point gap (25.0% vs 18.0%) reflects cases that were prevalent in the annual window but not on the snapshot date — some diagnosed before July 1 in remission, some newly diagnosed after. Prevalence is a dimensionless proportion (cases / observable population), not a rate per person-time. Both estimates are sensitive to the case-finding rule: a two-claim vs one-claim definition or a 12-month vs 24-month lookback will change the numerator by a meaningful margin.

(2) Practical interpretation. A disease-management program sizing its target population should use annual period prevalence (25.0%, 50 patients) as the denominator for coverage calculations, since it captures everyone with the condition at any point in the plan year. A point-in-time prevalence (18.0%) is appropriate for a cross-sectional intervention or a budget snapshot tied to a specific date. Always state which measure is reported — presenting 18.0% as "the prevalence" in a dossier when 25.0% is the better operational figure for a full-year program is a common undercount error.

Worked example

Scenario

A small rural health district wants to understand how much hypertension affects its population. The district has records for 200 residents who were enrolled in its health plan for the entire year 2024. A clinic nurse runs two prevalence calculations from the same enrollment list: (1) point prevalence on July 1, 2024 — who is an active hypertension case on that exact date — and (2) period (annual) prevalence for all of 2024 — who had a hypertension diagnosis recorded at any point during the year. The two numbers will differ because some people were only diagnosed earlier or later in the year.

Dataset

Patient records for the 200-person enrolled population. Columns show whether each person had a hypertension diagnosis active on July 1, 2024 (the point-prevalence date) and whether they had any hypertension diagnosis recorded at any time during the full year 2024 (the period/annual window). Rows shown are a representative sample of 10 patients.

Steps

• Define the denominator: all 200 people who were enrolled for the full year 2024 — these are the people we can reliably observe.

• Compute point prevalence (July 1, 2024): count only the people who are active hypertension cases on that one date. From the full population, 36 out of 200 residents meet this criterion.

• Point prevalence = 36 cases on July 1 / 200 enrolled residents = 0.180, or 18.0%.

• Compute period (annual) prevalence (all of 2024): count anyone who had a hypertension diagnosis recorded at any time during the year — this picks up people diagnosed in January, people diagnosed in November, and everyone in between. From the full population, 50 out of 200 residents had at least one hypertension diagnosis during 2024.

• Period prevalence = 50 cases during 2024 / 200 enrolled residents = 0.250, or 25.0%.

• Note the direction: period prevalence (25.0%) is higher than point prevalence (18.0%) because period prevalence adds the 14 people who had a diagnosis recorded at some point in 2024 but were not flagged as active cases on July 1 specifically — for example, someone diagnosed in March whose record was not flagged as active by mid-year, or someone newly diagnosed in October.

• This difference is expected and not an error: point prevalence is a narrow snapshot; period prevalence casts a wider net over the full year.

Result

Point prevalence on July 1, 2024 = 36 / 200 = 0.180 (18.0%). Annual (period) prevalence for calendar year 2024 = 50 / 200 = 0.250 (25.0%). Period prevalence is always greater than or equal to point prevalence for the same population and window because it captures everyone the point estimate does plus anyone who was a case at any other moment in the year.

Runnable example

import pandas as pd
from scipy.stats import beta

def _ci(x: int, n: int):
    # Exact (Clopper-Pearson) 95% CI for a proportion.
    lo = 0.0 if x == 0 else beta.ppf(0.025, x, n - x + 1)
    hi = 1.0 if x == n else beta.ppf(0.975, x + 1, n - x)
    return x / n, lo, hi

def denominator_point(enroll: pd.DataFrame, index_date: pd.Timestamp) -> set:
    # Active, FFS-observable enrollment on the index date.
    m = (enroll["enroll_start"] <= index_date) & (enroll["enroll_end"] >= index_date) & enroll["ffs_observable"]
    return set(enroll.loc[m, "person_id"])

def denominator_year(enroll: pd.DataFrame, year: int, min_months: int = 11) -> set:
    # FFS-observable coverage for >= min_months of the calendar year (overlap-month count).
    ys, ye = pd.Timestamp(year, 1, 1), pd.Timestamp(year, 12, 31)
    e = enroll[enroll["ffs_observable"]].copy()
    e["ov_start"] = e["enroll_start"].clip(lower=ys)
    e["ov_end"] = e["enroll_end"].clip(upper=ye)
    e = e[e["ov_start"] <= e["ov_end"]]
    e["months"] = ((e["ov_end"].dt.to_period("M").astype("int64") -
                    e["ov_start"].dt.to_period("M").astype("int64")) + 1)
    cov = e.groupby("person_id")["months"].sum()
    return set(cov[cov >= min_months].index)

def diagnosed_cases(dx: pd.DataFrame, denom: set, asof: pd.Timestamp,
                    lookback_days: int = 730, min_claims: int = 2, min_gap_days: int = 30) -> set:
    # >= min_claims diagnosis dates spanning >= min_gap_days within the lookback, among the denominator.
    win = dx[(dx["person_id"].isin(denom)) &
             (dx["svc_date"] <= asof) &
             (dx["svc_date"] >= asof - pd.Timedelta(days=lookback_days))]
    dates = win.groupby("person_id")["svc_date"].agg(["nunique", "min", "max"])
    ok = dates[(dates["nunique"] >= min_claims) &
               ((dates["max"] - dates["min"]).dt.days >= min_gap_days)]
    return set(ok.index)

def treated_cases(rx: pd.DataFrame, denom: set, win_start: pd.Timestamp, win_end: pd.Timestamp) -> set:
    # >= 1 qualifying fill inside the window, among the denominator.
    m = (rx["person_id"].isin(denom)) & (rx["fill_date"] >= win_start) & (rx["fill_date"] <= win_end)
    return set(rx.loc[m, "person_id"])

def annual_prevalence(enroll, dx, rx, year=2024):
    denom = denominator_year(enroll, year)
    ys, ye = pd.Timestamp(year, 1, 1), pd.Timestamp(year, 12, 31)
    dxc = diagnosed_cases(dx, denom, asof=ye, lookback_days=730)
    rxc = treated_cases(rx, denom, ys, ye)
    n = len(denom)
    return {
        "n_denominator": n,
        "diagnosed": _ci(len(dxc), n),                 # (prevalence, lo, hi)
        "treated": _ci(len(rxc), n),
        "diagnosed_and_treated": _ci(len(dxc & rxc), n),
    }

library(data.table)

pt_ci <- function(x, n) {
  bt <- binom.test(x, n)              # exact (Clopper-Pearson) 95% CI
  c(prev = x / n, lo = bt$conf.int[1], hi = bt$conf.int[2])
}

denom_year <- function(enroll, year, min_months = 11L) {
  setDT(enroll)
  ys <- as.Date(sprintf("%d-01-01", year)); ye <- as.Date(sprintf("%d-12-31", year))
  e <- enroll[ffs_observable == TRUE]
  e[, ov_start := pmax(enroll_start, ys)]
  e[, ov_end   := pmin(enroll_end, ye)]
  e <- e[ov_start <= ov_end]
  mn <- function(d) as.integer(format(d, "%Y")) * 12L + as.integer(format(d, "%m"))
  e[, months := mn(ov_end) - mn(ov_start) + 1L]
  cov <- e[, .(months = sum(months)), by = person_id]
  cov[months >= min_months, person_id]
}

diagnosed_cases <- function(dx, denom, asof, lookback_days = 730L,
                            min_claims = 2L, min_gap_days = 30L) {
  setDT(dx)
  win <- dx[person_id %chin% denom & svc_date <= asof &
            svc_date >= asof - lookback_days]
  agg <- win[, .(nd = uniqueN(svc_date),
                 span = as.integer(max(svc_date) - min(svc_date))), by = person_id]
  agg[nd >= min_claims & span >= min_gap_days, person_id]
}

treated_cases <- function(rx, denom, win_start, win_end) {
  setDT(rx)
  unique(rx[person_id %chin% denom & fill_date >= win_start & fill_date <= win_end, person_id])
}

annual_prevalence <- function(enroll, dx, rx, year = 2024L) {
  denom <- denom_year(enroll, year)
  ys <- as.Date(sprintf("%d-01-01", year)); ye <- as.Date(sprintf("%d-12-31", year))
  dxc <- diagnosed_cases(dx, denom, asof = ye, lookback_days = 730L)
  rxc <- treated_cases(rx, denom, ys, ye)
  n <- length(denom)
  list(n_denominator = n,
       diagnosed = pt_ci(length(dxc), n),
       treated   = pt_ci(length(rxc), n),
       diagnosed_and_treated = pt_ci(length(intersect(dxc, rxc)), n))
}