Medication Adherence in RWE

Medication adherence in RWE

is a time-ordered process, not one percentage. The process begins with a prescription or treatment decision, moves through initiation, continues through day-to-day implementation, and ends at discontinuation. Persistence is the time from initiation until the last dose before discontinuation. This parent card separates the two electronic-data branches that are routinely mixed together: primary non-adherence, in which a new prescription is never dispensed or started within a prespecified window, and secondary non-adherence, in which a patient who has initiated treatment does not obtain or use medication as intended during follow-up. The distinction determines the denominator, the data needed, and the question answered.

Core conceptual distinction

Primary adherence is an order-to-first-dispensing event. Its denominator is newly ordered prescriptions, so it requires an EHR or e-prescribing order linked to dispensing or administration data. Claims alone cannot identify people who were prescribed a drug but never filled it. Secondary adherence is conditional on initiation and uses a post-initiation treatment history. PDC and MPR quantify medication availability over a fixed observation window; persistence quantifies duration until discontinuation; refill-gap and trajectory methods preserve more timing information. These are indirect measures of possession or administration, not proof of ingestion. A paid pharmacy claim shows that supply was dispensed, while a medication order or reconciled medication list does not show that the medication was obtained.

Pros, cons, and trade-offs

A family framework prevents a common category error: reporting PDC among first-fill patients as though it described adherence among everyone who was prescribed treatment. It also keeps implementation and persistence separate, so a patient with intermittent short gaps is not automatically treated as equivalent to a patient who permanently discontinues. The cost is that a complete adherence pathway needs linked order, dispensing, administration, enrollment, and sometimes clinical data. A single database rarely observes every stage. Prefer the full hierarchy when describing treatment uptake or real-world strategy performance; use a narrower child measure when the estimand is explicitly limited to one stage and the denominator is stated.

When NOT to use - and when it is actively misleading

Do not collapse primary and secondary non-adherence into one rate unless every patient has a captured prescribing event and longitudinal dispensing follow-up. Do not call a prescription order an exposure, a dispensing an ingestion, or an absent claim non-adherence when pharmacy benefit capture is incomplete. Do not use an arbitrary PDC >= 0.80 threshold as a universal biological boundary; thresholds and observation windows are drug- and question-specific. Do not classify adherence over a future window and then use that classification from baseline in an outcome model, because the patient must survive and remain observable long enough to earn the label. Use a landmark, time-varying exposure, or a properly weighted per-protocol strategy.

Data-source operational depth

Claims and pharmacy dispensing data are strong for secondary adherence because `fill_date`, `days_supply`, reversals, and enrollment can be used to reconstruct coverage, but they cannot measure primary non-adherence without an order denominator and they do not prove ingestion. EHR and e-prescribing data capture orders and can define primary adherence when linked to fills; medication lists and refill requests are weaker because they may be stale. Medication administration records and procedure/J-code claims are preferable for clinician-administered products, where days_supply-based PDC may be undefined. Registries can capture reasons for stopping and patient-reported use but often miss complete dispensing histories. Linked EHR-order, pharmacy-claim, administration, and enrollment data provide the strongest continuum, provided order cancellations, claim reversals, cash fills, benefit carve-outs, inpatient stays, and out-of-network fills are reconciled.

Worked example

A health system writes 1,000 new antihypertensive prescriptions. Eight hundred are dispensed within 30 days, so primary adherence is 80% and 200 orders are primary non-adherent under the prespecified rule. Among the 800 initiators, 600 have PDC >= 0.80 over the next 180 days, 90 have lower PDC but remain persistent, and 110 discontinue after exceeding the permissible gap. The correct report keeps the denominators visible: 60% of all newly prescribed patients are both initiated and PDC-adherent (600/1,000), while 75% of initiators meet the PDC threshold (600/800). Reporting only 75% hides the 200 people who never initiated.

Decision diagram

flowchart LR
  A[New prescription] --> B{First dose or dispensing?}
  B -->|No| C[Primary non-adherence]
  B -->|Yes| D[Secondary adherence / implementation]
  D --> E[PDC or MPR]
  D --> F[Persistence and discontinuation]
  D --> G[Refill-gap or trajectory methods]

Index definitions

Worked example

Scenario

A health system links new medication orders to pharmacy dispensings and follows each initiator for 180 days. The study needs separate counts for initiation, post-initiation coverage, and discontinuation.

Dataset

Cohort-level counts at each adherence stage.

Steps

Result

Primary adherence is 80%; PDC adherence is 75% among initiators but 60% across all newly prescribed patients.

Trade-offs

Runnable example

import pandas as pd

WINDOW_DAYS = 30
valid_fills = fills.loc[~fills["reversed"].fillna(False)].copy()
linked = orders.merge(valid_fills[["person_id", "fill_date"]], on="person_id", how="left")
linked["days_to_fill"] = (linked["fill_date"] - linked["order_date"]).dt.days
eligible = linked.loc[linked["days_to_fill"].between(0, WINDOW_DAYS, inclusive="both")]
first = eligible.groupby("order_id", as_index=False)["fill_date"].min()
out = orders.merge(first, on="order_id", how="left")
out["initiated"] = out["fill_date"].notna()
out["primary_nonadherent"] = ~out["initiated"]

library(data.table)
setDT(orders); setDT(fills)
window_days <- 30L
valid <- fills[is.na(reversed) | reversed == FALSE]
linked <- merge(orders, valid[, .(person_id, fill_date)], by = "person_id", all.x = TRUE)
linked[, days_to_fill := as.integer(fill_date - order_date)]
first <- linked[days_to_fill >= 0L & days_to_fill <= window_days,
                .(fill_date = min(fill_date)), by = order_id]
out <- merge(orders, first, by = "order_id", all.x = TRUE)
out[, initiated := !is.na(fill_date)]
out[, primary_nonadherent := !initiated]

%let window_days = 30;
proc sql;
  create table adherence_stage as
  select o.person_id, o.order_id, o.order_date,
         min(case when f.reversed ne 1 and
                       f.fill_date-o.order_date between 0 and &window_days
                  then f.fill_date end) as first_fill format=date9.
  from orders o
  left join fills f on o.person_id=f.person_id
  group by o.person_id, o.order_id, o.order_date;
quit;
data adherence_stage;
  set adherence_stage;
  initiated = not missing(first_fill);
  primary_nonadherent = not initiated;
run;

Citations

1. [1]Vrijens B, De Geest S, Hughes DA, et al. A new taxonomy for describing and defining adherence to medications. Br J Clin Pharmacol. 2012;73(5):691-705.
2. [2]Raebel MA, Schmittdiel J, Karter AJ, et al. Standardizing terminology and definitions of medication adherence and persistence in research employing electronic databases. Med Care. 2013;51(8 Suppl 3):S11-S21.
3. [3]Cramer JA, Roy A, Burrell A, et al. Medication compliance and persistence: terminology and definitions. Value Health. 2008;11(1):44-47.

1. [4]Steiner JF, Prochazka AV. The assessment of refill compliance using pharmacy records: methods, validity, and applications. J Clin Epidemiol. 1997;50(1):105-116.