Secondary Medication Non-Adherence

Secondary medication non-adherence

begins only after treatment has been initiated. It describes incomplete implementation of the prescribed regimen, early discontinuation, or both during a defined follow-up period. In electronic RWE, the observed behavior is usually medication acquisition or administration rather than ingestion: pharmacy claims show when supply was dispensed, medication administration records show doses given in a care setting, and EHR medication lists show what clinicians believe is active. The card therefore treats every metric as an operational proxy tied to a data source and denominator.

Core conceptual distinction

PDC and MPR are fixed-window coverage-intensity measures; persistence is a duration-to-discontinuation construct; refill-gap measures preserve the timing and size of uncovered intervals; trajectory models classify longitudinal patterns such as consistently high, declining, intermittent, or rapid discontinuation. PDC counts the union of covered days and is bounded at 1.0. MPR sums acquired days_supply and can exceed 1.0 when fills overlap unless capped. Persistence asks how long therapy continues before a prespecified permissible gap is exceeded and can be high even when several smaller gaps reduce PDC. No one metric is the definition of secondary non-adherence. The estimand should determine which child measure is primary.

Pros, cons, and trade-offs

A secondary-adherence hub makes the measurement choice explicit and prevents PDC, MPR, and persistence from being used as synonyms. PDC is usually preferable for bounded quality measurement; MPR can preserve information about oversupply or acquisition volume; persistence is more intuitive for time-to-discontinuation; gap sequences and trajectories expose patterns hidden by one summary number. The cost is analytic multiplicity. Reporting several metrics without prespecifying one primary rule creates threshold shopping. Choose the primary measure, observation window, stockpiling rule, inpatient rule, permissible gap, switch rule, and censoring strategy before outcomes are examined.

When NOT to use - and when it is actively misleading

Do not apply secondary-adherence metrics to patients who never initiated; that is primary non-adherence. Do not calculate PDC or MPR when days_supply is absent or meaningless, as for many infused or procedure-administered therapies. Do not interpret a paid claim as confirmed ingestion, or an EHR order as a dispensing. Do not count missing benefit data, plan disenrollment, cash fills, or out-of-network dispensings as uncovered days. Do not define a 12-month adherence group with future data and then start outcome follow-up at initiation; this creates immortal-time and reverse-causation bias. Use a landmark before outcome follow-up, a time-varying adherence history, or a properly specified per-protocol analysis.

Data-source operational depth

Pharmacy claims support PDC, MPR, refill gaps, and persistence when `fill_date`, `days_supply`, reversals, drug identity, and continuous pharmacy enrollment are reliable. Early refills require an explicit stockpiling/carryover rule, hospitalizations require denominator removal or bridging, and switches within a class require an interchangeable-product policy. EHR data may add prescribed dose, stop orders, reasons for discontinuation, and clinical response, but orders and medication lists are not proof of supply. Medication administration records support dose-completion and appointment-gap measures for clinician-administered therapies. Registries and patient reports add reasons and actual-use context but may be incomplete longitudinally. Linked sources are strongest when their date semantics are reconciled.

Worked example

One patient initiates a chronic oral medication on January 1 and receives three consecutive 30-day fills on January 1, January 31, and March 1. The 180-day observation window runs from January 1 through June 28. Under a carryover rule, the fills provide 90 unique covered days, so PDC is 90/180 = 0.50. MPR is also 90/180 = 0.50 because no oversupply remains beyond the window. With a 60-day permissible gap, discontinuation is assigned to March 31, the first uncovered day after the last supply; persistence is 90 days. A patient with early overlapping fills could have MPR above PDC, and a patient who restarts after a long gap could have moderate PDC but short persistence to first discontinuation. Those differences are the reason the measures remain separate children.

Decision diagram

flowchart TD
  A[Initiated treatment] --> B[Secondary adherence]
  B --> C[PDC: unique covered days]
  B --> D[MPR: acquired supply ratio]
  B --> E[Persistence: time to discontinuation]
  B --> F[Gap sequence / CMG]
  B --> G[Adherence trajectory classes]

Index definitions

Worked example

Scenario

A patient has three consecutive 30-day fills and then no further dispensing during a 180-day window. The analysis compares fixed-window coverage measures with time to discontinuation.

Dataset

Three pharmacy dispensings for one initiated patient.

Steps

Result

PDC = 0.50, MPR = 0.50, and persistence = 90 days under a 60-day permissible-gap rule.

Trade-offs

Runnable example

import pandas as pd

GAP_DAYS = 60
rows = []
for person_id, person_fills in fills.sort_values("fill_date").groupby("person_id"):
    start = index.loc[index.person_id.eq(person_id), "index_date"].iloc[0]
    end = index.loc[index.person_id.eq(person_id), "window_end"].iloc[0]
    window_days = (end - start).days + 1
    covered = set()
    supply_end = start
    discontinuation = pd.NaT
    total_supply = 0
    for fill in person_fills.itertuples():
        total_supply += int(fill.days_supply)
        fill_start = max(fill.fill_date, supply_end)
        if (fill.fill_date - supply_end).days > GAP_DAYS and pd.isna(discontinuation):
            discontinuation = supply_end
        fill_end = fill_start + pd.Timedelta(days=int(fill.days_supply))
        covered.update(pd.date_range(fill_start, fill_end - pd.Timedelta(days=1), freq="D"))
        supply_end = max(supply_end, fill_end)
    covered_days = sum(start <= day <= end for day in covered)
    rows.append({"person_id": person_id, "pdc": covered_days / window_days,
                 "mpr": total_supply / window_days,
                 "persistence_days": ((discontinuation if pd.notna(discontinuation) else supply_end) - start).days})
adherence = pd.DataFrame(rows)

library(data.table)
setDT(fills); setDT(index)
gap_days <- 60L
daily <- fills[, .(covered_date = seq(fill_date, by = "day", length.out = days_supply)),
               by = .(person_id, fill_date)]
daily <- unique(daily)
daily <- merge(daily, index, by = "person_id")
daily <- daily[covered_date >= index_date & covered_date <= window_end]
pdc <- daily[, .(covered_days = uniqueN(covered_date)), by = person_id]
supply <- fills[, .(total_supply = sum(days_supply)), by = person_id]
out <- Reduce(function(x, y) merge(x, y, by = "person_id", all = TRUE), list(index, pdc, supply))
out[, window_days := as.integer(window_end - index_date) + 1L]
out[, `:=`(pdc = pmin(covered_days / window_days, 1), mpr = total_supply / window_days)]
setorder(fills, person_id, fill_date)
fills[, previous_end := shift(fill_date + days_supply), by = person_id]
first_gap <- fills[as.integer(fill_date - previous_end) > gap_days,
                   .(discontinuation = min(previous_end)), by = person_id]
out <- merge(out, first_gap, by = "person_id", all.x = TRUE)
out[, persistence_days := as.integer(fifelse(is.na(discontinuation), window_end + 1L, discontinuation) - index_date)]

%let gap_days = 60;
proc sql;
  create table fill_windows as
  select f.person_id, f.fill_date, f.days_supply,
         i.index_date, i.window_end,
         max(f.fill_date, i.index_date) as cover_start format=date9.,
         min(f.fill_date + f.days_supply - 1, i.window_end) as cover_end format=date9.
  from fills f inner join index i on f.person_id=i.person_id;
quit;
data covered_days;
  set fill_windows;
  do covered_date=cover_start to cover_end;
    if cover_start <= cover_end then output;
  end;
  keep person_id covered_date;
run;
proc sort data=covered_days nodupkey; by person_id covered_date; run;
proc sql;
  create table pdc_summary as
  select i.person_id,
         count(distinct c.covered_date)/(i.window_end-i.index_date+1) as pdc
  from index i left join covered_days c on i.person_id=c.person_id
  group by i.person_id, i.index_date, i.window_end;
  create table mpr_summary as
  select i.person_id,
         sum(f.days_supply)/(i.window_end-i.index_date+1) as mpr
  from index i left join fills f on i.person_id=f.person_id
  group by i.person_id, i.index_date, i.window_end;
  create table adherence as
  select p.person_id, p.pdc, m.mpr
  from pdc_summary p left join mpr_summary m on p.person_id=m.person_id;
quit;

Citations

1. [1]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.
2. [2]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.
3. [3]Andrade SE, Kahler KH, Frech F, Chan KA. Methods for evaluation of medication adherence and persistence using automated databases. Pharmacoepidemiol Drug Saf. 2006;15(8):565-574.

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.