Patient Dropouts

Abstract

Patient dropout, in the usage of the United States National Library of Medicine, names patients who leave a treatment program or a research study before it is complete. It is the discontinuation pole of treatment adherence and compliance: a failure of persistence rather than of initiation, and it is pervasive. Roughly one in five adults leaves psychotherapy prematurely, and longitudinal studies routinely lose a fifth or more of their samples. Dropout matters twice over, clinically, because a treatment stopped early seldom works, and methodologically, because those who leave rarely resemble those who stay, so the data left behind are biased. This article covers how dropout is defined and counted, why the count depends on the definition, how differential attrition distorts study results, the cognitive decision that underlies leaving, and what reduces it.

Keywords: treatment adherence, no-show patients, attrition, premature termination, missing data

Patient dropout is the point at which a person who entered a course of treatment or a study stops taking part before it has run its course. The event is simple to name and surprisingly hard to pin down: a patient who misses one appointment may return, one who stops a medicine may restart, and a research participant lost at the third wave may reappear at the fourth. What unites these cases is a decision, often unspoken and sometimes barely deliberate, to discontinue, and that decision sits at the end of the same behavioral chain that governs adherence. A dropout is a patient who accepted a plan, perhaps initiated it, and then did not persist. The question this article addresses is why that happens, how it is measured, and why it damages both the patient who leaves and the evidence base built on those who remain.

What Patient Dropouts Are

Patient dropout, as the United States National Library of Medicine defines it, refers to patients who leave a treatment program or a research study before it is complete. The term spans two settings that share a structure but differ in what is at stake. In the clinical setting, a dropout is a patient who stops attending or stops treatment before the clinician judges the course finished; the cost falls on that patient, whose half-completed treatment may do little good. In the research setting, a dropout is a participant lost to follow-up before the study ends; the cost falls on the study, whose remaining sample is no longer the one that was enrolled.

In the behavioral language of treatment adherence and compliance, under which MeSH classifies patient dropout, leaving treatment is a failure of persistence — the last of the phases through which a recommendation is carried out, after acceptance, initiation, and implementation (Vrijens et al., 2012). This placement matters because it separates dropout from the other ways a treatment can fail. A patient who refuses a plan never accepts it; a patient who forgets doses implements it poorly; a dropout does neither of these, but having begun, stops. The terminology of medication-taking makes the same distinction, reserving discontinuation for the end of persistence and separating it from the gaps and lapses of ongoing implementation (Cramer et al., 2008). Discontinuation is not a rare edge case: in long-term regimens for chronic conditions, adherence averages only about 50%, with much of the shortfall owing to patients who stop entirely (Osterberg & Blaschke, 2005).

Because persistence unfolds over time, dropout is best pictured not as a single event but as the thinning of a cohort. Of every hundred patients who begin, some number are still present at each later point, and the curve that traces them downward — a retention or survival curve — is the natural description of dropout in both the clinic and the study. Its shape carries information: a steep early drop points to problems with the treatment's acceptability or first experience, while a slow, steady decline points to the ordinary erosion of persistence over a long regimen.

How Dropout Is Counted

The single most important fact about dropout rates is that they depend on how dropout is defined, and the definitions in use do not agree. A patient can be counted as a dropout for failing to complete a fixed number of sessions, for stopping before the therapist judged treatment complete, for being lost to the study's final assessment, or for a clinician's global rating of premature termination — and these criteria classify different patients. An early meta-analysis of psychotherapy dropout, pooling studies that used varied definitions, arrived at a mean dropout rate of roughly 47%, a figure inflated by lenient criteria such as counting anyone who attended fewer than a set number of sessions (Wierzbicki & Pekarik, 1993). When a later and larger meta-analysis restricted itself to a consistent, defensible definition of premature discontinuation, the rate fell to about one in five — 19.7% of clients — a far more credible estimate of how often adults leave therapy early (Swift & Greenberg, 2012).

One fixed cohort of 1000 clients, sorted by how many sessions each attended. Slide the rule for what counts as dropping out. Bars left of the line are counted as dropouts; not one client changes, yet the reported rate swings widely with the definition alone.

Counted dropouts: 230 of 1000 = 23.0%

A lenient rule (counting anyone below many sessions) pushes the rate toward the ~47% of early pooled estimates; a strict, defensible rule brings it near the ~one-in-five of the consistent-definition meta-analysis (Wierzbicki & Pekarik, 1993; Swift & Greenberg, 2012). A reported dropout rate is uninterpretable without the definition that produced it. Illustrative distribution, not data from a specific trial.

The lesson generalizes beyond psychotherapy. Whenever a dropout rate is quoted, the operational definition behind it governs the number, and comparing rates across studies that defined dropout differently is comparing different quantities. A reported rate is therefore only interpretable alongside its definition, the length of the intended course, and the window over which patients were observed. This is the same measurement problem that afflicts adherence generally, where the instrument shapes the estimate (DiMatteo, 2004), sharpened here because the event being counted — what it means to have left early — admits so many reasonable operationalizations.

Dropout in Psychotherapy and Over Time

Psychotherapy is the setting in which dropout has been studied most carefully, because the treatment is delivered in countable units — sessions — and its premature end is consequential and visible. The consistent-definition estimate of about 19.7% conceals real variation: dropout is higher among children and adolescents than adults, higher for some presenting problems than others, and higher in community and training settings than in tightly controlled trials (Swift & Greenberg, 2012). A related analysis separated two distinct events that are often conflated — refusal, declining treatment before it begins, and premature termination, leaving after it starts — and showed that both vary systematically across psychotherapy, pharmacotherapy, and their combination, so the choice of treatment itself shapes how many patients never start and how many leave (Swift et al., 2017).

Follow 1000 patients over 8 intervals. The first-interval rate sets the early drop; the steady rate sets the gradual erosion after it. Because patients must remain at every interval, retention is a running product, so modest rates compound into a small survivors' club.

Still present at the end: 383 of 1000 — overall dropout 61.7%

A low first-interval rate with high steady retention gives an early-steep curve (a problem with the treatment's acceptability or first experience); a high first rate with lower steady retention gives a gradual decline (ordinary erosion of persistence). The shape tells a clinician where persistence is breaking down (Cramer et al., 2008). Illustrative rates, not a specific cohort.

Viewed over time, dropout compounds. If a cohort retains some fraction of its remaining members at each successive session or follow-up, the fraction still present after several intervals is the product of the per-interval retention rates, not their average, so a sequence of individually unalarming rates can leave a minority of the original cohort by the end. This is why a treatment that loses only a modest share of patients per session can nonetheless deliver a full course to fewer than half who began, and why the shape of the decline, early and steep versus late and gradual, tells a clinician where in the course persistence is breaking down (Cramer et al., 2008).

Attrition and the Missing-Data Problem

In research, dropout has a second cost that is easy to miss: it does not merely shrink the sample, it distorts it. Participants who leave a study are seldom a random subset of those who entered; they tend to differ in ways tied to the very outcomes under study, so the data that remain give a biased picture of the population that was enrolled. A long-running population study that tracked its attrition directly found that those lost over fifteen years differed systematically from those retained, and that the resulting bias threatened the generalizability of estimates drawn from completers alone (Gustavson et al., 2012).

The standard framework for reasoning about this is the taxonomy of missing data. Data are missing completely at random (MCAR) when the probability of dropping out is unrelated to anything, observed or not; missing at random (MAR) when it depends only on data already observed; and missing not at random (MNAR) when it depends on the unobserved outcome itself — the patient who stops because they are getting worse, precisely the value the study wanted to record. Only MCAR leaves a completer analysis unbiased. MAR can be addressed with appropriate methods that use the observed data; MNAR, the common case for clinical dropout, cannot be fully corrected by any statistical technique, because the information needed to correct it left with the patient (Little et al., 2012).

A cohort of 1,000: 600 doing well (would score 80) and 400 doing poorly (would score 50), so the true mean is fixed at 68.0. Set how many of each group stay. Equal retention keeps the estimate honest; losing the poorly-doing group faster inflates it.

740 completers · observed 71.9 · bias +3.9 · dropout 26% · MNAR (tied to outcome)

Set both retentions equal and the observed mean returns to 68.0 however high the dropout — attrition unrelated to the outcome (MCAR) costs sample but not accuracy. Drop the poorly-doing group faster and the estimate drifts upward even at modest overall loss: the harm tracks the selectivity of dropout, not its size (Little et al., 2012; Gustavson et al., 2012). At the default 90% / 50% this reproduces the Worked Example: 740 completers, 71.9, bias +3.9.

This is why a National Research Council panel convened on the problem concluded that the most reliable remedy is not a cleverer analysis but a better-designed study: minimize dropout while the trial is running, define the target estimand in a way that acknowledges those who leave, and treat large amounts of missing data as a threat to a trial's validity rather than a nuisance to be patched afterward (Little et al., 2012). Prevention dominates correction because no method recovers what was never measured.

Worked Example

Consider a one-year study that enrolls 1,000 patients. Suppose 600 are in fact doing well and would score 80 on the outcome measure, and 400 are doing poorly and would score 50. The true mean the study is trying to estimate is therefore (600 × 80 + 400 × 50) / 1,000 = 68.0.

Now let the patients drop out. In the realistic case, those doing poorly leave at a higher rate — they are discouraged, or sicker, or both — so the missingness is tied to the outcome (MNAR). Say 90% of the well group is retained but only 50% of the poorly-doing group. The completers and the mean computed from them are:

The observed mean among the 740 completers is (540 × 80 + 200 × 50) / 740 = 53,200 / 740 = 71.9. The study would report 71.9 against a true value of 68.0 — an upward bias of about 3.9 points — even though only 26% of patients dropped out, because the ones who left were disproportionately the low scorers. The bias comes not from the amount of dropout but from its selectivity.

Contrast the same 26%-ish loss applied without respect to outcome: retain 70% of each subgroup (MCAR). Completers are 420 well and 280 poorly, and the observed mean is (420 × 80 + 280 × 50) / 700 = 47,600 / 700 = 68.0 — exactly the true value. The arithmetic makes the framework's central claim concrete: identical overall dropout can leave an estimate perfectly unbiased or materially wrong, and what decides which is whether leaving was tied to the outcome. No amount of the missing-at-random case hurts the mean; a modest amount of the not-at-random case does.

Discussion

Patient dropout occupies an unusual position: it is at once a clinical problem, a measurement problem, and a methodological one, and progress has come from treating it as all three. The clinical reframing, visible in the psychotherapy literature, moved dropout from a mark against the patient toward a signal about the treatment and its fit — a patient who leaves may be telling the clinician that the plan's perceived burden has come to outweigh its perceived benefit, the same necessity-concerns balance that governs adherence more broadly (Swift & Greenberg, 2012). The measurement reframing established that a dropout rate is meaningless without its definition, so that the apparent disagreement between a 47% figure and a 20% figure dissolves once the operational criteria are made explicit (Wierzbicki & Pekarik, 1993) (Swift & Greenberg, 2012).

The methodological reframing has been the most consequential for how evidence is produced. Recognizing that dropout is typically not random, and that the most damaging kind cannot be fixed after the fact, shifted the emphasis in trial design from statistical repair toward prevention, and made the handling of missing data a standard criterion for judging a trial's credibility (Little et al., 2012) (Gustavson et al., 2012). The common thread is that dropout is informative. Whether in a clinic or a cohort, the fact that a particular patient left, and the way they resemble or differ from those who stayed, is data about the treatment and the population, not merely an absence to be lamented.

Cognitive Implications

The decision to drop out is a cognitive act, and locating it in the architecture of goal pursuit explains both its timing and its predictors. A patient who has accepted and begun a treatment holds a goal to complete it, but that goal must be re-endorsed at every juncture — each appointment, each refill, each follow-up — and dropout is what happens when a re-evaluation comes out against continuing. Because the evaluation recurs, dropout is a hazard distributed over the whole course rather than a single hurdle at the start, which is why retention curves decline continuously rather than stepping down once.

What tips a given re-evaluation toward leaving is, in large part, the same balance that predicts adherence: when a patient's concerns about a treatment — its burden, its side effects, its slow or absent benefit — come to outweigh their sense of its necessity, the rational response from the patient's point of view is to stop, and they do. This is why dropout is so weakly a matter of patients not understanding their treatment and so strongly a matter of their evaluation of it, and why interventions that merely supply information rarely retain patients, while those that address the felt burden and the perceived benefit can (Swift et al., 2017). Framed this way, premature termination is not the failure of a passive patient to comply but the outcome of an active patient's cost-benefit judgment, repeatedly made — a judgment a clinician can inform, but cannot make for them.

Current Directions

The most consequential recent movement is in clinical-trial methodology, where the recognition that not-at-random dropout cannot be analyzed away has reorganized how trials are planned. Guidance now frames the question around a clearly specified estimand — exactly what quantity the trial means to estimate, including how it treats patients who discontinue — and treats the prevention of missing data as a design obligation rather than an afterthought, with analysis strategies and sensitivity analyses chosen in advance to probe how far conclusions depend on untestable assumptions about those who left (Little et al., 2012). The practical effect is that a modern trial is judged partly on how little it lost and how honestly it reasons about what it did.

In the clinic, the active front is matching retention efforts to the patient and the moment. Because dropout reflects a recurring cost-benefit judgment rather than a fixed trait, the work has turned toward identifying which patients are at risk and when — the determinants that predict discontinuation span the patient, the condition, the therapy, and the health system rather than any single cause (Kardas et al., 2013) — and toward addressing the specific concern or barrier driving a given patient's re-evaluation rather than exhorting persistence in general. This is the same tailored, phase-specific logic that the broader adherence-intervention literature has converged on after large syntheses found generic reminders wanting: a Cochrane review of interventions for medication adherence identified no approach that was both highly effective and broadly applicable (Nieuwlaat et al., 2014), and meta-analyses pooling hundreds of trials have found that even the better interventions yield only modest, hard-won improvements (Conn & Ruppar, 2017) (Kini & Ho, 2018).

Common Misconceptions

A high dropout rate always means a bad treatment.

Not necessarily, and not comparably across studies: because the measured rate depends on how dropout is defined, a study using a lenient definition can report a far higher rate than one using a strict definition of the same treatment, so a rate is uninterpretable without its definition (Wierzbicki & Pekarik, 1993) (Swift & Greenberg, 2012).

If only a small fraction drops out, the results are safe.

The harm done by dropout depends on its selectivity, not only its size: a modest amount of outcome-related (not-at-random) attrition can bias an estimate substantially, while even heavy attrition that is unrelated to the outcome leaves the estimate unbiased (Little et al., 2012).

Statistical methods can fix dropout after the fact.

Only partly. Methods exist for data missing at random, but for data missing not at random — the usual case, where patients leave because of the outcome itself — no technique can fully recover the lost information, which is why prevention is emphasized over correction (Little et al., 2012).

Patients who drop out simply do not care.

Dropout typically reflects an active re-evaluation in which a patient's concerns about a treatment come to outweigh its perceived necessity or benefit, not indifference; it is a cost-benefit judgment a clinician can inform rather than a lapse of will (Swift et al., 2017).

Glossary

Attrition bias.

The distortion of a study's results that arises when the participants lost to follow-up differ systematically from those retained, so the completing sample no longer represents the enrolled one.

Attrition.

The loss of participants from a study over its course; the research-setting form of dropout, measured as the thinning of the enrolled cohort over time.

Completer analysis.

An analysis restricted to participants who finished the study; unbiased only when dropout is unrelated to the outcome, and otherwise a source of attrition bias.

Dropout.

A patient or participant who leaves a treatment program or research study before it is complete; in adherence terms, a failure of persistence after a treatment has begun.

Estimand.

A precise statement of the quantity a trial sets out to estimate, including how it treats patients who discontinue; specifying it in advance forces an explicit decision about what dropout means for the result.

Intention-to-treat analysis.

An analysis that includes every enrolled participant in the group to which they were assigned, regardless of whether they completed, limiting the bias that dropping completers would introduce.

Loss to follow-up.

The research-setting form of dropout: a participant who cannot be assessed at a scheduled point because contact has lapsed, whether or not they formally withdrew.

Missing at random (MAR).

Missingness whose probability depends only on data already observed, not on the unobserved value itself; it can be addressed with appropriate statistical methods.

Missing completely at random (MCAR).

Missingness unrelated to any variable, observed or not; the only case in which a completer analysis remains unbiased.

Missing not at random (MNAR).

Missingness whose probability depends on the unobserved outcome itself, such as a patient leaving because they are deteriorating; the common case for clinical dropout, and one no method can fully correct.

Persistence.

The phase of adherence describing how long a patient continues a treatment before discontinuing; dropout is the failure of this phase.

Premature termination.

Leaving a treatment after it has begun but before the provider judges it complete; distinguished from refusal, which is declining treatment before it starts.

Retention.

The complement of dropout: the fraction of an enrolled cohort still participating at a given point, traced over time as a retention or survival curve.

Survival curve.

A plot of the fraction of a cohort still present as a function of time; its shape distinguishes early, steep dropout from slow, steady attrition.

Key Researchers

Roger P. Greenberg

(SUNY Upstate Medical University). Co-author of the psychotherapy-dropout meta-analyses and of work comparing psychotherapy, pharmacotherapy, and their combination, grounding the account of why patients leave treatment early in treatment expectations and experience. Faculty

Roderick J. A. Little

(University of Michigan). Built, with Donald Rubin, the modern taxonomy of missing data — completely at random, at random, and not at random — and chaired the National Research Council panel whose report reframed how clinical trials should prevent and analyze dropout. Wikipedia · Faculty

Joshua K. Swift

(Idaho State University). His meta-analyses quantified premature discontinuation across adult psychotherapy at about one in five clients, and established the preferences-and-expectations work that frames dropout as a decision rather than a failure. ORCID · Scholar · Faculty

Bernard Vrijens

(AARDEX Group and University of Liège). A biostatistician who led the consensus that split adherence into initiation, implementation, and persistence, so that discontinuation — a failure of persistence — is named as its own phase rather than blurred into generic non-compliance. Repository

Frequently Asked Questions

What is a patient dropout?

A patient dropout is a person who leaves a treatment program or a research study before it is complete. In a clinic the cost falls on the patient whose treatment is cut short; in a study the cost falls on the evidence, because the remaining sample no longer matches the one that enrolled (Swift & Greenberg, 2012).

How is dropout different from refusing or forgetting treatment?

Dropout is a failure of persistence: the patient accepted and began the treatment and then stopped. Refusal is declining before starting, and poor implementation is executing an ongoing treatment incorrectly; dropout is specifically leaving after having started (Vrijens et al., 2012) (Cramer et al., 2008).

How common is dropping out of psychotherapy?

Using a consistent definition, a large meta-analysis estimated that about one in five adults, 19.7%, leaves psychotherapy prematurely, though rates are higher among younger patients and in some settings. Older estimates reaching nearly half reflect more lenient definitions of what counts as dropping out (Swift & Greenberg, 2012) (Wierzbicki & Pekarik, 1993).

Why do reported dropout rates vary so much?

Because the measured rate depends on how dropout is defined. Counting anyone below a set number of sessions, versus stopping before the therapist judged treatment complete, versus loss to final assessment, classifies different patients, so rates from studies with different definitions are not directly comparable (Wierzbicki & Pekarik, 1993).

Why does dropout bias research results?

Because those who leave usually differ from those who stay. When leaving is tied to the outcome being studied, patients dropping out because they are doing worse, the completing sample overstates how well the treatment worked, and the bias can be large even when overall dropout is modest (Gustavson et al., 2012) (Little et al., 2012).

What do MCAR, MAR, and MNAR mean?

They classify why data are missing. Missing completely at random means dropping out is unrelated to anything; missing at random means it depends only on observed data; missing not at random means it depends on the unobserved outcome itself. Only the first leaves a completer analysis unbiased, and the last cannot be fully corrected (Little et al., 2012).

Can statistics correct for dropout after a study ends?

Only partly. Methods exist for data missing at random, but for data missing not at random, the usual case for clinical dropout, no technique recovers the information that left with the patient, which is why minimizing dropout during the study is emphasized over repairing it afterward (Little et al., 2012).

Does a patient who drops out simply not care about getting better?

Usually not. Dropping out typically reflects an active re-evaluation in which the patient's concerns about a treatment come to outweigh its perceived benefit or necessity, a cost-benefit judgment rather than indifference, which is why addressing the specific barrier works better than urging persistence in general (Swift et al., 2017).

References