Prevalent New-User Design

In plain language

The prevalent new-user design is a way to study how two drugs compare in the real world when most patients reach the second drug by switching from the first one rather than starting it fresh. It lets those switchers into the study, but only at the exact moment they switch, and carefully matches each switcher to a patient who stayed on the original drug and had been on it for the same amount of time. This time-on-drug matching step is what keeps the comparison fair by ensuring both groups have survived the same duration on the comparator drug before the clock starts.

The prevalent new-user (PNU) design (Suissa, Moodie & Dell'Aniello, 2017) is a hybrid between the strict new-user (incident-user) design and an analysis that would naively pool prevalent users. A pure new-user design restricts both arms to patients initiating treatment with no prior use, which cleanly aligns time zero and removes prevalent-user bias but discards the large pool of patients who were already on the comparator drug and then switch to (or add) the study drug. When the study drug is typically reached by switching from an established comparator — as is common in chronic disease, oncology lines of therapy, and many add-on indications — the strict new-user design can throw away the majority of the real-world treated population, leaving the study underpowered and unrepresentative. The PNU design recovers those patients: a comparator user becomes eligible to enter the study-drug arm at the time they switch/initiate, and is matched to a comparator user who remains on the comparator and who has the same accumulated treatment history (same time since comparator initiation) and the same time-conditional propensity score for switching at that moment.

Core conceptual mechanism

The design solves two coupled problems that arise when you allow prevalent users in. (1) Aligned, well-defined time zero: each study-drug initiator's follow-up starts at their switch date, and the matched comparator's follow-up starts at the same calendar/treatment-history point, so the two arms share a comparable "now" rather than the study-drug arm being a survivor cohort observed from an arbitrary point. (2) Comparable prior exposure: because patients are matched on time since comparator initiation (the "prevalent" clock), the design balances depletion of susceptibles — the phenomenon whereby people who tolerate a drug and survive on it for years are a healthier, selected subset. Matching on this clock and on the time-conditional propensity score (TCPS) — the probability of switching at a given moment given covariates measured up to that moment — means the comparator group represents the counterfactual "what if this prevalent user had stayed on the comparator." A single prevalent user can serve as a potential match (a "prevalent-user moment") at several time points, mirroring the risk-set sampling logic of a nested case-control / matched cohort, with the analysis run as a matched cohort (e.g., Cox with a frailty or robust variance for the matched sets).

What it is and is not

The PNU design is an extension of and complement to the active-comparator new-user design, not a replacement for it. It does not license pooling arbitrary prevalent users; the validity hinges on (a) being able to model the switching propensity well from longitudinal covariates, (b) having enough comparator users at each treatment-history time to form matches, and (c) the study drug being reached predominantly by switching rather than by de-novo initiation. It is also distinct from a target-trial emulation: PNU is a matching-and-time-alignment strategy for the specific switching structure, whereas target-trial emulation is a general framework for specifying the protocol; the two are compatible and PNU can be viewed as one way to operationalize the eligibility/time-zero rules of an emulated switching trial.

Pros, cons, and trade-offs

- vs the strict `new-user-design`: PNU recovers the (often large) population of switchers that the incident-user design discards, restoring power and external validity for drugs that are reached by switching, while still controlling prevalent-user/depletion bias through matching on time since comparator initiation and the TCPS. Cost: it requires longitudinal data rich enough to estimate the switching propensity over time and is more complex to implement and explain. Prefer the strict new-user design when de-novo initiation is the real-world reality and a clean incident cohort is large enough; prefer PNU when most patients reach the study drug by switching and the new-user cohort would be small or unrepresentative. - vs the `active-comparator-new-user` design: Active-comparator new-user fixes confounding by indication by comparing two initiators with the same indication; PNU keeps that spirit but extends the eligible population to comparator-to-study-drug switchers, matching at the switch moment. Prefer active-comparator new-user when two drugs are genuinely interchangeable first-line options; prefer PNU when the study drug is a later-line or add-on/switch therapy so that "new use" of the study drug coincides with prevalent use of the comparator. - vs accepting `prevalent-user-bias` (naive pooling of prevalent users): Naive prevalent-user comparisons mix aligned and misaligned time zeros, condition on survival, and bias estimates unpredictably. PNU is the principled way to include prevalent users without that bias. Never prefer naive pooling; PNU exists precisely to avoid it.

When to use

Comparative effectiveness or safety of a drug that real-world patients predominantly reach by switching from or adding to an established comparator (second-line antidiabetics, switched antihypertensive or antidepressant classes, later-line oncology regimens, add-on therapies); when a strict new-user cohort would be too small or would represent an atypical de-novo-initiating subpopulation; when longitudinal claims/EHR data support estimating a time-conditional switching propensity from covariates measured up to each potential entry time.

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

- When de-novo initiation is the actual clinical pattern. If patients genuinely start the study drug as a first treatment, a strict new-user design is cleaner and PNU adds complexity with no gain; forcing a switching structure where none exists misrepresents the population. - When the time-conditional switching propensity cannot be credibly modeled. PNU's bias control rests entirely on the TCPS and on matching the prevalent clock; if longitudinal covariates are sparse or switching is driven by unrecorded factors (e.g., unmeasured disease progression), the matched comparator is not a valid counterfactual and the estimate is confounded — reporting it as bias-controlled is misleading. - When there are too few comparator users at each treatment-history time to match. Sparse risk sets force coarse matching or dropped switchers, reintroducing imbalance; check match rates and balance by the prevalent clock before trusting the result. - When the prevalent clock is ignored. Matching only on a baseline propensity score while allowing prevalent users in reintroduces depletion-of-susceptibles bias; the time-since-comparator-initiation match is not optional.

Data-source operational depth

- Claims (FFS vs MA): The natural substrate — dispensing records give exposure start, switch dates, and the longitudinal covariate stream (diagnoses, prior fills, utilization) needed for the TCPS. Build the prevalent clock from the first observable comparator fill within a continuously enrolled, FFS-observable window; Medicare Advantage enrollees generate no fee-for-service claims, so MA-only spans corrupt both the prevalent clock and the switching-propensity covariates — restrict to FFS-observable person-time. Define switching with explicit grace/gap rules so an add-on is not misread as a switch. - EHR: Medication orders, problem lists, labs, and vitals give richer time-varying covariates for the switching model (disease severity, response markers), sharpening the TCPS, but encounter-driven capture and out-of-system care mean the switch date and prior-exposure clock may be incomplete; require demonstrable in-system activity and reconcile order vs fill. - Registry / linked: Disease registries supply adjudicated severity and line-of-therapy information that materially improves the switching propensity and the time-since-initiation clock; linked claims-EHR is the strongest substrate but introduces linkage selection that must be reported.

Worked example

Scenario

We want to compare drug B (a newer oral diabetes medication) against drug A (the established standard). Drug B is only prescribed to patients who tried drug A first and then switched. We identify a patient, Pat, who started drug A on 2022-01-01 and switched to drug B 9 months later on 2022-10-01. Under a strict new-user design, Pat would be excluded because Pat is not a fresh starter of drug A. Under the prevalent new-user design, Pat enters the study-drug arm at the switch date (2022-10-01). We then find a match for Pat: another patient, Casey, who started drug A on 2022-01-01 (so also 9 months into drug A as of 2022-10-01) and did not switch. Both Pat and Casey begin their study follow-up on 2022-10-01.

Dataset

Key dates and measurements for the switcher (Pat) and the matched comparator continuer (Casey) as they would appear in a linked claims file.

Steps

• Pat starts drug A on 2022-01-01. The prevalent clock begins: month 1, month 2, ... month 9.

• On 2022-10-01 (9 months in), Pat switches to drug B. This date becomes Pat's index date and the start of follow-up.

• To find a match for Pat, we look for drug-A patients who were also exactly 9 months into drug A on 2022-10-01 and had not yet switched. Casey qualifies: Casey also started drug A on 2022-01-01, is still on it, and has the same estimated propensity to switch as Pat.

• Both Pat and Casey begin their follow-up on 2022-10-01. Any events (hospitalizations, lab changes) from 2022-10-01 forward are counted for both in their respective arms.

• Matching on time-since-comparator-start (9 months for both) ensures neither arm has more drug-A survivors selected into it than the other, keeping the comparison of outcomes fair.

Result

One matched pair: Pat (9 months on drug A then switched to drug B) versus Casey (9 months on drug A, continuing). Follow-up for both starts 2022-10-01. The design recovers Pat for the analysis instead of discarding this common real-world switcher.

Timeline Spec

Title

Prevalent new-user design: switcher Pat matched to comparator continuer Casey at 9 months

Window

Start

2022-01-01

End

2022-10-01

Label

Comparator period (9 months on drug A) leading to index date

Events

• Label

  Pat: starts drug A

  Start

  2022-01-01

  Length Days

  273

  Quantity

  9 months on comparator

• Label

  Pat: switches to drug B (index date)

  Start

  2022-10-01

  Length Days

  1

  Quantity

  switch event

• Label

  Casey: starts drug A

  Start

  2022-01-01

  Length Days

  273

  Quantity

  9 months on comparator

• Label

  Casey: matched at same time point (continues drug A)

  Start

  2022-10-01

  Length Days

  1

  Quantity

  match event

Spans

• Kind

  exposed

  Start

  2022-01-01

  End

  2022-09-30

  Label

  Pat: 9 months on drug A (comparator period)

• Kind

  followup

  Start

  2022-10-01

  End

  2022-10-01

  Label

  Pat: index date (switch to drug B) — follow-up starts here

• Kind

  exposed

  Start

  2022-01-01

  End

  2022-09-30

  Label

  Casey: 9 months on drug A (same duration as Pat)

• Kind

  followup

  Start

  2022-10-01

  End

  2022-10-01

  Label

  Casey: matched index date — follow-up starts here on drug A

Result

Label

Both Pat and Casey begin follow-up 2022-10-01 with identical time-on-comparator (9 months). Match valid.

Value

matched pair, 9 months on comparator each

Caption

Pat (top row) accumulates 9 months on drug A then switches to drug B on 2022-10-01. Casey (bottom row) also has exactly 9 months on drug A on 2022-10-01 and continues. Both enter follow-up on the same calendar date with the same treatment history length, making the comparison of outcomes valid.

Alt Text

Two parallel horizontal timelines. The top row shows Pat starting drug A on January 1 2022, an arrow at October 1 2022 labeled switch to drug B which becomes the index date, then a follow-up bar extending right. The bottom row shows Casey starting drug A on January 1 2022, a match marker at October 1 2022 (same date), then a follow-up bar continuing on drug A. Both follow-up bars start at the same point, illustrating aligned time zero with 9 months of comparator history behind each patient.

Runnable example

import numpy as np
import pandas as pd
import statsmodels.formula.api as smf

# Illustrative person-period data: one row per person per month since comparator initiation.
rng = np.random.default_rng(3)
rows = []
for pid in range(2000):
    comorb = rng.binomial(1, 0.3); age = rng.normal(60, 10)
    switched = False
    for t in range(1, 25):                       # months on the comparator
        if switched:
            break
        p_switch = 1/(1+np.exp(-(-4 + 0.03*(age-60) + 0.8*comorb + 0.04*t)))
        on_study = rng.binomial(1, p_switch)     # switch to study drug this month?
        rows.append(dict(person_id=pid, t=t, on_study=on_study, age=age, comorb=comorb))
        switched = bool(on_study)
pp = pd.DataFrame(rows)

# Time-conditional propensity score: probability of switching at month t given covariates up to t.
tcps = smf.logit("on_study ~ age + comorb + t", data=pp).fit(disp=0)
pp["ps"] = tcps.predict(pp)

# Within each prevalent-clock time t, nearest-neighbor match switchers to comparator continuers on the TCPS.
matches = []
for t, g in pp.groupby("t"):
    switchers = g[g.on_study == 1]
    controls  = g[g.on_study == 0].copy()
    for _, s in switchers.iterrows():
        if controls.empty:
            continue
        j = (controls["ps"] - s["ps"]).abs().idxmin()      # closest TCPS at the same t
        matches.append((s["person_id"], controls.loc[j, "person_id"], t))
        controls = controls.drop(j)                        # match without replacement within t
matched = pd.DataFrame(matches, columns=["study_id", "comparator_id", "match_time"])
print(f"switchers matched: {len(matched)} (each pair shares t and TCPS)")
print(matched.head())

library(MatchIt)
set.seed(3)

# Illustrative person-period data: one row per person per month since comparator initiation.
mk <- function() {
  comorb <- rbinom(1, 1, 0.3); age <- rnorm(1, 60, 10); out <- list(); sw <- FALSE
  for (t in 1:24) {
    if (sw) break
    p  <- plogis(-4 + 0.03*(age-60) + 0.8*comorb + 0.04*t)
    on <- rbinom(1, 1, p)
    out[[t]] <- data.frame(t = t, on_study = on, age = age, comorb = comorb)
    sw <- on == 1
  }
  do.call(rbind, out)
}
pp <- do.call(rbind, lapply(1:2000, function(i) cbind(person_id = i, mk())))

# Time-conditional PS via pooled logistic; exact-match the prevalent clock t, NN-match the TCPS.
m <- matchit(on_study ~ age + comorb + t,
             data = pp, method = "nearest", distance = "glm",
             exact = ~ t, ratio = 1)        # exact on time since comparator initiation
summary(m)                                  # balance of the TCPS and covariates within t
matched <- match.data(m)                    # feed to a downstream Cox on the matched cohort
cat(sprintf("matched rows: %d\n", nrow(matched)))