Medication adherence data is any recorded information about whether a patient takes their prescribed medications as directed, including when, how often, and in what quantity. It comes from pharmacy refill records, insurance claims, electronic health records, patient surveys, pill-counting devices, blood and urine tests, and even ingestible sensors. Health plans, clinicians, pharmaceutical companies, and researchers all rely on this data, but they collect it differently and use it for different purposes, which means the same patient can look “adherent” by one measure and “non-adherent” by another.
The Most Common Source of Adherence Data Is Pharmacy Claims
The workhorse of medication adherence measurement is pharmacy claims data. When you fill a prescription, your pharmacy submits a record to your insurer that includes the drug, dose, quantity dispensed, and the date. From these records, analysts calculate how consistently you’re picking up your medications over time. The two dominant metrics are the medication possession ratio (MPR) and the proportion of days covered (PDC). Both aim to estimate the fraction of time you had medication on hand, but their calculation methods differ enough that comparing adherence rates across studies or health plans is surprisingly difficult.
MPR divides the total days’ supply dispensed by the number of days in a measurement period, which means it can actually exceed 100% if someone refills early. PDC, by contrast, caps each day as either “covered” or “not covered,” so it tops out at 100%. The Centers for Medicare and Medicaid Services (CMS) adopted PDC as its preferred metric, which pushed most U.S. health plans toward that standard. Still, the specific rules around how to handle overlapping fills, medication switches, and gaps in coverage vary from one organization to another, making cross-plan comparisons messy.
Claims data has real advantages: it’s collected automatically, covers millions of patients, and doesn’t depend on anyone remembering to report anything. But it also has a fundamental blind spot. A filled prescription tells you the patient picked up the medication. It does not tell you the patient swallowed it. Pharmacy refill data can also miss prescriptions filled at out-of-network pharmacies, mail-order services not linked to the insurer, or drugs paid for out of pocket. One study found that linking electronic health records to pharmacy dispensing networks captured fills that insurance claims alone missed, and that adherence estimates could differ depending on which data source was used.
What Happens Before the First Fill
Most adherence metrics assume the patient has at least started the medication. But a meaningful number of prescriptions are written and never filled at all. Researchers call this primary non-adherence, and it’s largely invisible to systems that rely on refill data, because there’s no initial claim to trigger monitoring. Studies in primary care settings have examined how often patients simply never pick up a new prescription, and the rates can be substantial depending on the drug class and the patient population.
This gap matters because most adherence interventions target patients who have already begun filling a medication and then fall off. If the data infrastructure doesn’t capture the patients who never started, those patients never enter the intervention pipeline. Some health systems have started comparing e-prescribing records against pharmacy claims to flag unfilled new prescriptions, but this kind of cross-system matching is still far from routine.
Self-Report Measures
Asking patients directly how well they’ve been taking their medications is the simplest and cheapest way to collect adherence data. Clinicians do it in office visits, and researchers use standardized questionnaires. The evidence suggests that self-report measures show moderate agreement with more objective methods and can predict clinical outcomes to a meaningful degree.
The catch is that patients tend to overestimate how adherent they’ve been. This isn’t necessarily dishonest; people genuinely misremember. A validated three-item self-report scale, tested against electronic pill bottle monitoring for both HIV and non-HIV medications, showed correlations in the range of 0.47 to 0.59 with the electronic standard. That’s decent but far from perfect, and it means self-report is better at identifying patients who freely admit to missing doses than at catching those who don’t realize or don’t disclose their gaps.
For certain populations, self-report gets even trickier. In pediatric conditions, a caregiver often administers the medication, but neither the caregiver’s nor the child’s self-assessment reliably matches electronically measured adherence, especially when the child is the one responsible for taking the drug. One study of pediatric glaucoma patients found that when caregivers administered eye drops, the caregiver’s report correlated with measured adherence, but the child’s report did not. When children self-administered, neither report was significantly associated with actual use.
Biochemical Verification
The most direct way to confirm someone took a medication is to detect it in their body. Urine and blood tests using liquid chromatography-tandem mass spectrometry (a lab technique that identifies specific drug molecules at very low concentrations) have become the preferred method for objectively detecting non-adherence. Researchers have used this approach to check whether patients with type 2 diabetes were actually taking their oral antidiabetics, blood pressure medications, and statins by analyzing baseline urine samples.
Serum drug level testing has similarly been validated as an effective way to assess adherence in heart failure patients. The appeal of biochemical testing is that it’s binary in a useful way: either the drug or its metabolite is present, or it isn’t. But it has limitations. It generally tells you about recent intake, not long-term patterns. A patient who skips doses all month and then takes one the day before a clinic visit will test positive. The tests are also more expensive and logistically demanding than claims-based measures, so they tend to be used in research settings or targeted clinical scenarios rather than as routine population-level screening.
One reassuring finding is that variations in urine concentration (how dilute or concentrated a sample is) do not appear to affect the reliability of these biochemical tests, which removes one potential source of false negatives.
Ingestible Sensors and Digital Monitoring
A newer category of adherence technology embeds a tiny sensor directly inside a pill or capsule. When the sensor reaches the stomach and dissolves, it transmits a signal to a wearable patch or nearby device, which relays a time-stamped record to a server. This creates a direct, real-time log of each ingestion event. Two digital pill systems have received FDA clearance, and pilot studies have established the bioequivalence and stability of digitized versions of medications including HIV pre-exposure prophylaxis and antiretroviral therapy.
Accuracy in controlled pilot studies has been high, with the sensor successfully detecting ingestion events at rates ranging from 68% to 100%. In real-world clinical settings, accuracy drops to the 68% to 90% range, largely because patients don’t always wear the relay device or follow the system’s protocol correctly. In other words, the technology works well when used as intended, but non-adherence to the monitoring system itself becomes a new problem layered on top of medication non-adherence.
Another approach uses radio-frequency identification (RFID) tags embedded in gelatin capsules. Once the capsule dissolves, the tag activates and transmits a unique signal to a relay device, which sends a time-stamped message to a cloud server. This creates an objective, real-time record. Researchers developing these systems have noted that optimizing connectivity and the design of the relay device is important for maximizing patient acceptance, and that concerns about gut retention of metallic sensor components and drug dissolution inside capsule shells need continued attention.
How Health Plans Use Adherence Data for Star Ratings
In the United States, adherence data directly affects health plan finances through the Medicare Star Ratings program. CMS grades Medicare Advantage plans on a one-to-five-star scale across dozens of quality measures, and medication adherence for diabetes medications, blood pressure medications, and statins accounts for a heavily weighted portion of the overall score. Plans that score well earn bonus payments and are more attractive to enrollees; plans that score poorly can lose both.
A decade-long analysis of Star Ratings performance found that among plans achieving at least 4 stars on the medication adherence measures, roughly 70% to 74% also achieved at least a 4-star overall summary rating. Among plans that scored a perfect 5 stars on any adherence measure, 85% to 90% also reached at least 4 stars overall. The relationship between adherence scores and overall ratings has been strong and consistent year after year, which is why health plans invest heavily in adherence improvement programs.
Health system-level factors also play a role. Research within large integrated delivery systems has identified modifiable predictors of adherence performance on Star metrics, giving clinicians and plans concrete targets for improvement. These include features of how care is organized and delivered, not just patient-level behaviors.
Adherence Data at the Point of Care
Clinicians are increasingly seeing adherence information inside electronic health records. The idea is straightforward: if a doctor can see that a patient hasn’t filled their blood pressure medication in two months, they can address it during the visit instead of assuming the treatment isn’t working and escalating to a stronger drug. Most clinicians who have been surveyed on this prefer a system that integrates claims data directly into the medication list, using color-coded categories to flag adherence status at a glance.
Whether this actually changes outcomes is a harder question. A randomized trial of over 5,400 patients tested a clinical decision support tool that alerted primary care clinicians when patients appeared non-adherent to cardiometabolic medications. After 12 months, patients in the intervention group had better adherence to blood pressure medications compared with usual care, but the tool did not produce a statistically significant improvement in adherence to statins or non-insulin diabetes medications. The results suggest that decision support can help, but the effect may depend on the medication class and how clinicians respond to the alerts.
Adherence Data in Drug Development
Clinical trials depend on participants actually taking the study drug. If adherence is poor and unmeasured, the trial may underestimate a drug’s true effect, leading researchers to conclude a medication doesn’t work when the real problem is that people weren’t taking it consistently. Variable underdosing during trials has been recognized as a source of adverse consequences that can be mitigated by reliably measuring adherence, managing it during the trial, and using appropriate statistical methods to account for it.
Accurate adherence measurement during trials also helps researchers distinguish between minor dosing errors, such as taking a pill a few hours late, and significant deviations that could affect the drug’s safety or efficacy profile. This distinction matters for regulatory submissions, because a drug’s labeled dose-response relationship is only as reliable as the adherence data behind it.
Machine Learning and Predicting Who Will Stop Taking Medications
Rather than waiting for a patient to become non-adherent and then reacting, researchers are building predictive models that flag patients at risk before they fall off. These models draw on demographic, clinical, and sometimes behavioral data to generate a risk score. In one study of patients with type 2 diabetes, the best-performing algorithm achieved an area under the curve of 0.87, using nine variables including age, gender, fasting blood glucose control, duration of the current treatment regimen, diet adjustment, daily medication cost, blood glucose levels, cholesterol status, and body mass index.
Other researchers have pushed this further by incorporating cultural and belief-related factors. A study of Malaysian patients with chronic diseases found that adding beliefs about complementary and alternative medicine to the model improved prediction, achieving an AUC of 0.82 with a stacked ensemble model. The significant predictors included holistic health beliefs, race, number of daily doses, number of prescribed medications, religion, education level, and treatment duration. These findings highlight how adherence is shaped by context that purely clinical models miss.
Predictive models are promising, but deploying them raises practical questions. A risk score is only useful if someone acts on it, and the interventions triggered by that score need to be effective and not merely burdensome for already-stretched clinical teams.
Where Adherence Data Gets It Wrong
Every method for measuring adherence introduces its own errors. Claims data can’t confirm ingestion. Self-report is biased upward. Pill counts, once considered a reasonable objective measure, have their own problems: one study using unannounced telephone pill counts found that the most common source of error was overcounted doses sitting in pillboxes. Electronic monitoring devices record when a bottle is opened, not whether a pill was removed or swallowed, and patients may open the bottle out of curiosity or to consolidate pills without actually taking a dose.
Prescription claims databases have been found to be inaccurate for medications that don’t come in discrete dosage forms, like topical creams or inhalers, or for drugs prescribed on an “as-needed” basis. If someone has a prescription for a rescue inhaler they use only during asthma attacks, a low refill rate doesn’t mean non-adherence; it might mean they’re doing well.
When claims data is used to identify non-adherent patients for outreach, the results can be surprisingly imprecise. One study found that among patients flagged as non-adherent based on a gap of more than 30 days in refill history, pharmacist interviews revealed a range of reasons for the gap, not all of which represented true non-adherence. Some patients had received samples, switched to a different pharmacy, or experienced insurance-related disruptions.
Social Determinants Shape the Data
Adherence data doesn’t exist in a vacuum. Whether someone fills and takes their medications is heavily influenced by factors outside the clinic. A systematic review and meta-analysis found that food insecurity was associated with significantly lower odds of adherence, as was housing instability. Overall, the presence of adverse social determinants was associated with about 25% lower odds of being adherent.
Analysis of U.S. national survey data spanning nearly a decade found significant differences in medication adherence based on ethnicity, gender, socioeconomic class, alcohol consumption, disability status, ability to afford balanced meals, insurance coverage, marital status, and whether the person had a usual place for healthcare. These aren’t just confounders to adjust for in a regression model. They represent barriers that adherence data, taken at face value, can obscure. A patient flagged as “non-adherent” by a claims-based metric may be rationing doses because they can’t afford the copay, or skipping refills because they lack transportation to the pharmacy.
This creates an equity problem. If health plans use adherence data to allocate resources or assess quality, and if adherence is systematically lower in populations facing structural barriers, the data can end up penalizing the patients and communities that need the most support. Evidence on copayment policies and their effects on adherence has been shown to vary across countries and health systems, reinforcing that adherence patterns are not just about individual behavior but about the systems people navigate.
Digital Tools for Improving Adherence
Smartphone apps, text message reminders, and connected pill bottles are increasingly used both to measure and to improve adherence. A systematic review and meta-analysis of mobile app interventions found a modest but statistically significant improvement in adherence, with the most impactful app features being documentation of medication-taking, medication reminders, and data sharing with providers. The quality of the evidence, however, was rated low, and the long-term durability of app-driven improvements remains unclear.
Pharmacist-led medication therapy management, which often incorporates adherence data as both a trigger and an outcome measure, has shown stronger results in some settings. A randomized study of patients with type 2 diabetes found that adherence in the intervention group rose from about 9% at baseline to 61% at six months, compared with a rise from about 13% to 30% in the control group. The intervention group also had roughly half as many hospital admissions related to poorly controlled blood glucose.
Privacy and the Ethics of Knowing When Someone Takes a Pill
As adherence monitoring becomes more granular, from refill records to real-time ingestion tracking, ethical questions intensify. Ingestible sensors that transmit data to cloud servers raise concerns about informed consent, data confidentiality, and the power dynamics between patients and the institutions monitoring them. For patients, having both a clinical informed consent process and a technology user agreement creates confusion about what they’re agreeing to and who can see their data. For providers, the ability to verify whether a patient took their medication fundamentally changes the trust relationship: a clinician who can see that a patient skipped three doses faces new questions about how to respond without being punitive.
Third-party monitoring adds another layer. If an insurer or employer has access to real-time adherence data, the potential for that information to influence coverage decisions, premium pricing, or employment is not hypothetical. Researchers have flagged affordability as an ethical concern in its own right: if the cost of a digital monitoring system is passed to the patient, it could paradoxically worsen adherence by increasing the financial burden of treatment. And if the sensor or relay device malfunctions, the resulting false data, either false positives suggesting adherence that isn’t there or false negatives triggering unnecessary interventions, creates liability questions that current legal frameworks haven’t fully addressed.