Medication Adherence Measurement Methods: Why Pill Counts Lie

Pill counts and dosing diaries usually produce excellent adherence numbers. Plasma drug-level testing from real trials tells a different story, with a far smaller share of participants showing any drug in their system than the counts and diaries claimed. Both measurements describe the same people. Only one of them describes what those people did.

Raising adherence is one problem. Knowing whether your adherence number is real is another, and it gets far less attention. Our companion guide on how to improve medication adherence covers the behavioral and operational levers that move the rate. This post is about measurement validity: whether the rate you write into the clinical study report describes what participants did, or only what they were willing to report. For sponsors, CROs, and clinical operations leaders, those are separate liabilities.

The systematic bias hiding in “good” adherence numbers

Self-report and pill counts do not fail randomly. They fail in one direction: they overstate. Participants want to be good study subjects, they forget accurately less often than they forget conveniently, and the moment a count or a diary becomes a performance, it stops being a measurement.

Two HIV-prevention trials made this impossible to ignore.

In the VOICE trial (MTN-003), end-of-study retention and self-reported product use were both high. When investigators tested plasma drug levels, tenofovir was detected in only about 29% to 30% of random samples from participants assigned oral tenofovir-containing PrEP, according to analyses summarized by FHI 360 and the PrEP adherence-monitoring literature. Fewer than one in three participants showed biochemical evidence of taking a drug that nearly all of them said they were using.

The FEM-PrEP trial told the same story. An analysis of self-report and pill-count accuracy found fewer than 40% of women had drug detected in plasma despite high reported adherence, and the positive predictive value of every self-report and pill-count measure assessed was under 45%. Both trials reported no efficacy. The objective biomarker, not the questionnaire, explained why.

The lesson is not that participants lie maliciously. Any method built on recall or inference drifts upward, and it drifts most among the participants who are not taking the drug. That is the worst place to lose accuracy.

How pill counts go wrong, mechanically

Pill counts estimate adherence by subtracting returned tablets from dispensed tablets. The arithmetic assumes every absent tablet was swallowed, on time, by the right person. Each assumption is breakable.

• Dumping. A participant who forgot to dose for a week can remove a week of tablets before a visit, and the count looks perfect. In a study of HIV-positive adolescents, median pill-count adherence ran at about 98.7% while electronic monitoring sat lower, and the subgroup whose pill counts exceeded 100% (a signature of dumping) went on to have virologic failure roughly 33% of the time versus 13% in everyone else. The electronic and viral measures caught what the count hid (PMC).
• Over-adherence artifacts. Counts above 100% are common and physically meaningless; they signal dumping or miscounting, not diligence.
• Timing blindness. A count cannot tell whether 28 tablets were taken once daily or in three frantic catch-up days. For drugs with adherence-sensitive pharmacokinetics, timing is the endpoint.

A scoping review comparing electronic monitoring with other methods found pill counts overstated median adherence by roughly 8% versus electronic caps, with individual discrepancies spanning about negative 25% to positive 50%. An 8% median bias sounds survivable until you remember it sits on top of self-report bias and lands hardest on your non-takers.

The five measurement methods, compared

Every method trades cost against rigor, and rigor itself splits into distinct properties: how easily the measure can be gamed, how quickly it surfaces a problem, and how strong the resulting evidence is in front of a reviewer. Scoring them on those axes is more useful than ranking them on a single line.

Method	Relative cost	Gameability	Timeliness	Evidentiary strength
Patient diary / self-report	Very low	Very high (recall bias, social desirability)	Lagging, periodic	Weak: reports intent, not ingestion
Pill count	Low	High (dumping, miscount)	Lagging, visit-bound	Weak: infers from inventory
Smart bottles / MEMS caps	Moderate to high	Moderate (opening is not swallowing)	Near real-time per opening	Moderate: proves access events, not ingestion
PK / blood-level testing	High	Low (hard to fake a biomarker)	Intermittent snapshots	Strong for “drug was present,” blind between draws
AI-verified video dosing	Moderate	Low (timestamped recording of the act)	Near real-time per dose	Strong: observed ingestion in a tamper-evident record

A few trade-offs deserve naming directly.

Diaries and pill counts are cheap because they measure the wrong thing

Their low cost is real, and for low-stakes, non-adherence-sensitive endpoints they may be defensible. But they measure willingness to report, not exposure. Treat their output as a participant-engagement signal, not as adherence evidence you would defend in an inspection.

MEMS caps timestamp openings but cannot see the swallow

Electronic monitoring is a step up: it timestamps each access event and resists the crude dumping that defeats pill counts. Its ceiling is that an opened cap is not a swallowed dose, and dedicated participants can curate openings. It is strong on timeliness, partial on ingestion.

Pharmacokinetic testing is rigorous but episodic

A drug level is the closest thing to ground truth that the drug was in the body. That strength is also its limit: it captures a point in time, it is invasive and costly to run often, and “white-coat dosing” (taking the drug only right before a scheduled draw) can flatter the result. PK is excellent for periodic validation, poor for continuous management.

AI-verified video dosing captures the ingestion itself

Asynchronous video dosing asks the participant to record a short clip of the ingestion (pill on the tongue, swallow), which software then verifies under fixed rules. It captures the ingestion event itself, timestamps it, and writes it into an audit-grade record, approximating directly observed therapy without a clinician watching live. It does not replace PK for biochemical confirmation, and it depends on capture quality, but it closes the ingestion-evidence gap that pill counts, diaries, and even MEMS leave open. This is the approach behind Trialsights participant compliance.

What to do with this in your monitoring plan

The practical takeaway is not “switch to the most expensive method.” It is to match the method to the stakes and stop pretending a cheap proxy is evidence.

• Triangulate. No single method is sufficient. Pair an objective per-dose measure with periodic PK validation, and treat agreement between them as your confidence signal. Where they diverge, believe the biomarker.
• Score gameability explicitly. During vendor and protocol selection, ask of each measure: what is the easiest way for a non-adherent participant to look adherent? If the answer is “dump the bottle” or “tick the box,” you do not have evidence.
• Demand timeliness for management, rigor for proof. Real-time signals let coordinators intervene before a participant becomes a withdrawal; tamper-evident records let you defend the final number. You need both, and they are not the same system requirement.
• Write the audit trail into the method, not around it. A measurement is only as defensible as the record it lands in. Tamper-evident capture matters as much as the measurement itself, a theme we develop in the clinical trial compliance tools buyer’s guide.

This is the same uncomfortable pattern that shows up elsewhere in trial oversight: a comfortable headline metric that quietly fails to mean what teams assume. We make the parallel argument about verification coverage in why 100% SDV is a myth, and about attrition assumptions in our clinical trial retention benchmarks. In each case the fix is the same: interrogate how the number was produced before you trust what it says.

Adherence is also tightly coupled to retention. The participants whose pill counts look suspiciously perfect are often the ones drifting toward dropout, which is why lab and site surveillance that surfaces both signals early tends to pay for itself.

The number you can defend

A high adherence rate is not an asset if you cannot say how it was measured. The VOICE and FEM-PrEP failures were expensive lessons in this: the trials did not lack adherence data, they lacked adherence data that was true. Self-report and pill counts gave a reassuring picture; plasma levels gave the real one, after the studies had already enrolled, dosed, and concluded.

The teams that avoid that outcome do something unglamorous. They decide, up front, which measurement methods produce evidence they would put in front of a regulator, and they stop accepting the rest as anything more than a hint.

---

Want to see objective, AI-verified dose evidence on your own protocol? Book a demo and we will walk through video dosing, surveillance, and the tamper-evident audit trail on a live demo trial.