Much nonadherence research focuses on implementation once treatment is started despite evidence showing that non-initiation of medication (primary nonadherence) is associated with poorer health outcomes including higher mortality rates (Jackevicius et al., 2008) and emergency department visits (Lee et al., 2016). The limited research that exists exploring reasons for non-initiation suggest that the factors that influence patients’ decisions to initiate a medication or not are similar to the factors influencing whether a patient continues to take a medication long-term or stops it prematurely (discontinuation or non-persistence). Part of non-initiation is primary nonadherence, whereby a medication is newly prescribed but then the prescription is not filled at a pharmacy (Fischer et al., 2010). Cheen et al. (2019) systematically reviewed 33 studies and estimated that 17% of patients with six long-term conditions did not collect a newly prescribed medication with rates highest in osteoporosis and hyperlipidaemia (both 25%) and lowest in diabetes mellitus (10%). Studies of patients with a mean age under 65 years old had significantly higher primary nonadherence rates. But, lower primary nonadherence was found in patients aged 19–44 than in children or patients aged over 45 in a recent analysis of 34,243 Canadian primary care patients (Zeitouny et al., 2023) but older patients with polypharmacy were also at increased risk of primary nonadherence. Rates of primary nonadherence are also likely to vary with treatment type and healthcare system factors; Anaba and Arabambi (2022) examined rates at which dermatology patients collected a prescribed medication from a hospital in Lagos, Nigeria, and found 72% topical medications were not collected compared to 23% of oral medications, with more than half of patients who had not collected a medication saying that lack of availability and cost were the reasons for their nonadherence (Anaba and Arabambi, 2022).

Many patients who collect a medication (or have it delivered to them after dispensing) may still not initiate treatment (i.e., take the first dose) (Fischer et al., 2010). Estimates of the number of patients who obtain a prescription but then do not take the first dose are not widely available. Few adherence measures specify whether any dose is taken. Where digital adherence monitors (e.g., MEMs caps) are used to monitor adherence with oral medication in newly treated patients over time, there appear to be low rates of patients with 100% nonadherence (Hebing et al., 2022) but, people participating in research studies in which adherence is monitored may be more likely to initiate treatment than people who are not monitored. Studies of medication waste, such as analysis of medications returned to Dutch community pharmacies (Bekker et al., 2018), report returns of unopened packets, perhaps hinting that not all collected medications may be started. With the difficulties with capturing medication initiation, the true rates of non-initiation may be to accurately measure. Triangulating different data sources such as linking prescribing and dispensing records, along with electronic adherence monitoring, may help provide useful estimates of rates of non-initiation (Figure 2).

Implementation nonadherence is the most frequently assessed and commonly known form of nonadherence, with rates varying widely across patient groups, contexts and medication (Gast and Mathes, 2019; Foley et al., 2021). Implementation nonadherence is most often assessed in relation to number of doses taken, but can encompass timing, amount of medication taken, overuse, and adherence to other instructions (e.g., combination with food/fluid) (Nieuwlaat et al., 2014). Helmy et al. (2019) estimated that 37.9% of patients receiving immunosuppressants after heart transplant did not implement their medication as prescribed, within this 26.2% of patients took their immunosuppressant at a different time from that prescribed, while 17.3% did not take all of their immunosuppressant doses.

Rates of nonadherence are also likely to vary depending on the cut-off used to classify participants as nonadherent and the time period evaluated. Davis et al. (2010) estimated that 61% of patients with Parkinson’s disease took less than 80% of their medication (based on prescription refill data) over a 7 year period. Whereas, Buh et al. (2023) found 37.7% of patients with HIV at a clinic in Cameroon had missed one or more dose of their medication in the last month. When following up people taking antiretroviral treatment for 20 months using electronic monitoring, Wagner et al., found implementation nonadherence rates increased as time progressed (Wagner et al., 2020).

Rates may vary systematically across different contexts or healthcare systems, for example, Mahmood et al. (2022) reviewed 66 studies assessing implementation nonadherence to antihypertensives in Asia, and estimated an overall prevalence of 48% nonadherence, but found wide variation across regions (Mahmood et al., 2021). Relating to healthcare system factors, rates of antihypertensive implementation nonadherence in one cohort from Islamabad were lower in tertiary care patients than primary and secondary care patients and lower in those who had access to free medical care than those who did not (Mahmood et al., 2020). As with all factors that contribute to medication nonadherence, it is important not to overgeneralise or assume simple causation when considering associations between healthcare system and context factors and implementation nonadherence rates. For example, cost-related implementation nonadherence may occur because of medication unaffordability, but may also occur because groups who experience cost-related medication nonadherence may also be at increased risk of depression, which itself is linked to nonadherence (Briesacher et al., 2007; Gonzalez et al., 2011).

Implementation nonadherence is also likely to be higher for treatments that are more difficult to take or access. Okada et al. (2021) review studies estimating the prevalence of nonadherence to intravitreal ocular therapy for macular degeneration, which requires attendance at regular appointments for injections into the eye, and found rates of implementation nonadherence as high as 95.6% (Okada et al., 2021). There is evidence that treatments that involve multiple doses, or are involve complicated dosing instructions also achieve poor implementation rates (Ingersoll and Cohen, 2008).

Interestingly, there is emerging evidence that suggests there may be time-of-day effects on medication adherence with morning doses achieving greater adherence than evening doses. Phillips et al. conducted a study with electronic medication monitors in patients on twice-daily dosing for type 2 diabetes over 1 month and found that patients overall missed fewer morning pills (Phillips et al., 2021). However, the authors did not find that variability in dose timing differed between morning compared to evening. Thus, better morning adherence may not be due to consistency in the timing per se of the medication taking, but perhaps the linking of the morning adherence with a particular consistent routine such as morning coffee, which could vary in timing across different days. In contrast, the evenings may be more disrupted where the medication is either not taken at all, or if remembered, was taken at roughly the same time each evening. More research into the role of behavioural patterns and routines on routine medication taking is warranted to explore time-of-day and seasonal effects. Overall, implementation nonadherence can be said to be common but prevalence estimates are highly variable given the variation in conditions, seasonal and timing effects.

Non-persistence, whereby patients stop taking a medication before the time agreed with a healthcare professional is generally believed to increase over time. For example, Joret et al. (2022) found that non-small cell lung cancer patients took an estimated 98% of doses at the beginning of tyrosine kinase inhibitor treatment, but that at around 2 months, nearly half of patients had discontinued treatment and at 4 months more than 60% had discontinued. Hardtstock et al. (2022) evaluated non persistence to long-acting asthma treatments over a 12 months follow-up period and estimated non-persistence at 86.7%. There is some suggestion that non persistence increases with experience of adverse effects, with Fleming et al. (2022) finding that a majority of studies included in a review of adjuvant breast cancer treatment persistence found that patients who reported more adverse effects were more likely to discontinue. Alefan et al. (2022) found that adverse effects were particularly strongly linked to rates of discontinuation when the adverse effects were not anticipated by the patient.

Taking the estimated rates of initiation, implementation and persistence together, it is hypothesised that the often-quoted estimate of 30%–50% nonadherence is likely to be an underestimate. Measurement of all three components of nonadherence and longer follow-up times might demonstrate that nonadherence is more common than adherence in many patient populations.

Despite the potential for harnessing new technologies to map nonadherence, arguably self-report measures are still the dominant technique used to evaluate the extent to which someone is following the recommendations of their healthcare provider (Kamusheva et al., 2024). Often, they are relatively low cost, can be more feasible to use in routine care, can provide an immediate picture of adherence to facilitate intervention and can give insight into elements of nonadherence that may not be accessible from other measurements.

Common critiques of self-report measurement include that it risks over-estimation due to social desirability bias, is reliant on accurate memory of nonadherence, and may be dependent on patients having an accurate understanding of the recommendations that they have been given about how to take their medication (Stirratt et al., 2015). There is reasonably strong evidence that some patients who self-report good adherence are not accurately reporting their behaviour, for example, a recent US study (Hebel et al., 2020) used urine testing for biomarkers for antiretrovirals for pre-exposure prophylaxis of HIV and found that 12%–15% of patients self-reporting full adherence had nonadherence indicated through urine testing.

The COSMIN checklist (Mokkink et al., 2010), arose from a consensus exercise focusing on how to evaluate patient reported outcome measures such as adherence self-report measures. It highlights internal consistency, content validity, hypothesis testing for construct validity, criterion validity and responsiveness as key dimensions on which to evaluate new measures. These criteria have been increasingly applied to evaluate or develop self-report adherence measures. Kwan et al. (2020) found most self-report measures of (non)adherence had good evidence of construct validity, structural validity and content validity. However, there was weak evidence of the test-retest reliability perhaps unsurprising, given that adherence can be a dynamic behaviour, minimal evidence relating to cross-cultural validity of measures, and poor reporting of how measures how been developed. Tegegn et al. (2022) reviewed self-report measures for medication (non)adherence in cardiovascular disease against the COSMIN criteria; no measure assessed all elements of initiation, implementation, and persistence/discontinuation, with most focused on implementation and none on initiation. Few (non)adherence self-report measures have been validated across all target conditions or groups, or in a wide range of languages/cultures, with implications for relevance. For example, Vianna et al. (2021), reviewed the use of self-report measures to assess adherence to warfarin therapy and highlighted that generic measures had been used but that these did not capture adherence to some of the medication-taking recommendations (e.g., changing dose if experiencing bleeding) which patients taking warfarin are asked to follow.

Overall medication nonadherence self-report measures are increasingly robustly validated. For example, Chan et al. (2020), reported on the development of a five-item self-report scale, the Medication Adherence Report Scale (MARS-5) which included items probing intentional and unintentional nonadherence and considered properties including internal reliability, construct validity and hypothesis testing. The Morisky Medication Adherence Scale (MMAS-8) is another commonly used eight-item structured, self-report measure that assesses medication adherence (Morisky et al., 2008). There are also disease specific questionnaires. Wilson et al. (2016) developed a three-item measure for use in patients with HIV based on reported doses taken/missed over the previous 30 days and validated against objective measures. Dima et al. (2017) validated a measure of nonadherence in asthma against dispensing records and probed overuse of treatment as well as underuse.

With the rise of ‘open science’ and increasing concern for access to scientific tools and outputs, there has been an increasing focus on legal and cost implications of self-report adherence measure use Tesfaye and Peterson (2022) highlight that many measures may be infeasible for clinicians and researchers in low resource settings to use due to legal and cost restrictions. In clinical practice there may be additional barriers of time, and uncertainty regarding whether the measure is validated for use that fits with clinical practice (e.g., verbal delivery). Garfield et al. (2011) highlighted that there is limited available data on key factors relevant for clinicians assessing adherence such as how long a measure takes to complete, and suitability for carer completion.

Selection of a self-report adherence measure therefore needs consideration of the psychometric properties of the measure, the aspects of adherence that need to be assessed, consideration of use restrictions and cost, and the available data on relevance to the particular research/clinical context, patient group and medication. The use of validated self-report measures can therefore be used to provide key insights into the behavioural drivers of medication nonadherence.

As our understanding of nonadherence as a behaviour has advanced, so too have technologies to assess nonadherence. These have been developed to better capture the complexity of nonadherence and medication-related behaviours, whilst at the same time providing opportunities for nonadherence promoting interventions. Electronic medication monitors (EMM) have existed since the 1990s in the form of smart inhalers (Julius et al., 2002), electronic dispensing ‘smart’ pill boxes/bottles and smart pills. These devices in its simplest form record the number of doses taken over time, though current available devices now routinely capture date/time stamps of each medication dose. A recent meta-analysis showed that individuals receiving EMM has significantly reductions in nonadherence with a large magnitude of effect though this did not always translate to clinical benefit in studies which reported both outcomes (Chan et al., 2022). Whilst EMM capture one aspect of medication taking – which is opening of the medication container or inhaler actuation, EMMs still cannot confirm actual medication consumption, which may explain why EMM studies of adherence do not always correlate with clinical outcomes (Chan et al., 2013). How nonadherence relates to clinical outcomes is a question that requires further exploration outside this review but is worth acknowledging that nonadherence alone is only an intermediate outcome and that changes in clinical outcomes are possible even without associated increases in nonadherence.

More sophisticated EMM can also capture location of dosing, allowing linkage with GPS-related data such as environmental factors, and linkage with other datasets. One example is the Propeller Health adherence monitoring inhaler device which can record the location of reliever inhaler use (Merchant et al., 2016). The ability to track and map the location of medication use has provided insights into where ‘hot spots’ of asthma attacks are occurring and allowed further investigation into linkage with environmental triggers such as weather and pollen. This is likely to have important benefits as the effects of climate change increase in years to come, with geographic mapping of medication nonadherence offering new insights to inform resource planning, medication access policies and population health management. EMM can link with wearables, health provider portals, patient apps and be used with AI in predictive analytics to see how changes in patterns of medication use can predict outcomes. For example, changes in reliever medication use alone without input from any other predictors has been shown to predict the onset of an asthma exacerbation 5 days before the attack occurs (Lugogo et al., 2022). The availability of real-time medication use data can thus be used to inform early-warning systems and alerts for patients and providers of negative health outcomes.

Another method of nonadherence assessment that has exploded in the last 20 years is evaluation of prescription and pharmacy databases to establish patterns of prescription redemption as a proxy for medication-taking. A range of indicators can be calculated. These include whether a prescription is redeemed, indicating primary nonadherence or non-initiation (Cheen et al., 2019). Medication possession ratio whereby the number of doses of redeemed treatment are evaluated against the number of doses prescribed can also be calculated to indicate potential implementation nonadherence in terms of missed doses (Vink et al., 2009). Finally, the date of last prescription refill can indicate discontinuation or persistence with treatment (Gershon et al., 2021).

Prescription refill data has been validated against nonadherence biomarkers and self-report [e.g., (Osula et al., 2022)], with studies showing moderate correlations between medication possession ratio and other outcomes that would be associated with nonadherence (Hood et al., 2018). Unlike self-report data it is ‘objective’ and less likely to be influenced by social desirability bias. But, prescription refill rates are known to be affected by factors such as oversupply, and prescribing duration patterns (Galozy et al., 2020). In addition, prescription refill data can only indicate whether a medication is dispensed but not whether or how it is taken such as timing, storage or use (e.g., inhaler technique), so may not correlate with some nonadherence outcomes (Pattock et al., 2024).

Another potentially useful tool to provide system- or population-level analysis of nonadherence is the utilization of patient records. For example, nonadherence discussions and support provided by healthcare providers and recorded in electronic patient notes may provide insight into patterns of nonadherence (Insani et al., 2023). Healthcare records systems may have specific codes or processes for healthcare professional logging of nonadherence, but this data is yet to be widely used in research. Healthcare records are being linked to pharmacy and other data to gain additional insights (Xu et al., 2023). The growth of large language models and artificial intelligence offer potential for data mining of electronic healthcare and pharmacy records to gain insights into nonadherence (Turchin et al., 2024). Offering nonadherence support automatically to certain patients based on healthcare records has been piloted (Bosl et al., 2013) but is under-explored. As dispensing data are often routinely collected, and, depending on access rights and availability, accessible to healthcare professionals they may be a useful cue for provision of nonadherence support within daily practice. Another use of prescription or dispensing records is to track longitudinal medication use and examine how trajectories of treatment initiation and discontinuation relate to outcomes (Hommel et al., 2017).