Dropout is one of those problems the field has documented thoroughly enough that the basic numbers no longer cause much surprise. Across adult individual psychotherapy, approximately one in five clients drops out before agreed termination, with the figure varying by setting, presentation, and how strictly "dropout" is defined, but with a remarkably consistent floor around that point.
This matters not just for the dropped client. It also propagates upward into the trial literature — intent-to-treat analyses average over the dropouts and pull effect sizes downward; per-protocol analyses exclude them and inflate effect sizes in the opposite direction. The meta-analyses then average over both, and the field receives a slightly fuzzier picture of what its interventions actually do. Dropout is a systematic problem with downstream consequences across the evidence base.
The therapeutic alliance is the variable the dropout literature consistently identifies as the most important predictor of who leaves and who stays. This is not a new finding and it is not contested.
The Sharf meta-analysis — alliance and dropout
Sharf, Primavera and Diener's 2010 meta-analysis in Psychotherapy synthesised the literature linking therapeutic alliance to dropout across 11 studies and roughly 1,400 participants. The conclusion was unambiguous. Weak alliance, measured by standard instruments early in treatment, is a significant predictor of subsequent dropout. The effect was robust across modalities, settings, and presenting problems. It was not driven by any single study. It is the kind of finding that meta-analyses confirm rather than discover — the alliance-dropout link was already well-established at the level of individual studies; the Sharf paper made the size of the effect more precise.
There is a temptation, on first encountering this evidence, to read it as a claim about therapist warmth or rapport. That reading is incomplete. Alliance, in the Bordin framework that most of these studies operationalise, has three components: the bond between client and therapist, agreement on the goals of treatment, and agreement on the tasks of treatment. A rupture in any of the three is an alliance problem, and the bond component is not necessarily the most important. A client who finds their therapist personally pleasant but does not understand why they are being asked to do behavioural experiments, or whose target goals are not the goals the therapist is treating, is a client with an alliance problem in the technical sense. The bond may register as fine; the dropout risk is still elevated.
This will become relevant for what the early signals of rupture actually look like.
Safran and Muran — what ruptures actually are
The most useful framework for thinking about alliance ruptures in practical clinical terms comes from Jeremy Safran and J. Christopher Muran's body of work, anchored in their 2000 Guilford volume Negotiating the Therapeutic Alliance. The framework has stood up across two decades of subsequent research, and it gives the field a language for what is happening when an alliance shows signs of strain.
Safran and Muran's central claim is that ruptures are not failures of the alliance — they are ordinary moments within it. Any alliance of any duration will encounter strains; what matters is whether the therapist notices the strain and whether the strain is repaired. The framework is therefore not about preventing ruptures, which is not really possible, but about identifying them and handling them.
The taxonomy distinguishes two broad patterns:
Withdrawal ruptures. The client disengages, becomes superficially compliant, smiles too much, agrees readily, brings less material to sessions, completes homework in a perfunctory way. The surface signal is cooperation; the underlying signal is that the client has stopped genuinely working with the therapist and is going through the motions. Withdrawal ruptures are particularly important because they look like good therapy from the outside. A compliant, polite client who completes the worksheets is not, on first impression, a client in difficulty.
Confrontation ruptures. The client expresses dissatisfaction directly — challenging the therapist's approach, expressing doubt about whether therapy is helping, becoming irritated or hostile, raising concerns about the treatment plan. Confrontation ruptures are easier to spot because the signal arrives as overt content. They are not necessarily easier to repair; they can be uncomfortable for the therapist to engage with and the path of least resistance is to defuse the content without acknowledging the rupture it points to.
The clinical implication of the taxonomy is sobering. Withdrawal ruptures are simultaneously the more common pattern and the harder pattern to detect, because the relevant signal is the absence of something — engagement, challenge, genuine collaboration — rather than the presence of overt difficulty. Therapists are trained to respond to what is in the room; ruptures of the withdrawal type partly consist of what has quietly stopped being in the room.
The Eubanks meta-analysis — repair matters
Eubanks, Muran and Safran's 2018 meta-analysis in Psychotherapy took up the question of what happens when ruptures are repaired. The finding, across 11 studies and roughly 1,300 participants, was that rupture repair is associated with improved outcomes — the size of the effect modest in absolute terms but consistent across studies.
The mechanism that matters here is the explicit one. Outcomes improve not because the ruptures are absent but because they are noticed and named. The repair process, in the studies that captured it well, involved the therapist identifying that something had shifted, raising it collaboratively with the client, exploring what had happened, and re-establishing alignment around the goals or tasks that the rupture had disturbed. The repair is conversational work, not a technique applied to the client; the client is a participant in it.
This means the field has a reasonably clear position on what helps. Alliance ruptures are common, weak alliance predicts dropout, repaired ruptures predict better outcomes than unrepaired ones. The chain is intact. What the chain depends on is the therapist's ability to detect the rupture in time to do something about it — and that detection is, in practice, the weak link.
The detection problem
Withdrawal ruptures, in particular, are harder to detect in real time than in retrospect. This is not a failing of any particular therapist. The signal is subtle — a slight increase in surface compliance, slightly less spontaneous content brought to session, a small drop in homework engagement, a marginally more agreeable manner. Each of these on its own is within normal session-to-session variation; the rupture signal is the pattern across them, and pattern recognition requires comparison against the client's baseline.
The therapist holds the client's baseline implicitly, and that implicit baseline is often inaccurate. The clinician's sense of "the client has seemed a bit different over the past few weeks" tends to crystallise around the third or fourth session of the changed pattern, by which point the rupture has had time to compound and the dropout decision may already have formed in the client's mind.
The retrospective clarity is striking. After a client drops out, the therapist can usually identify, on reflection, the session at which something shifted. The compliance had become slightly hollow. The homework had been technically completed but without the engagement of previous weeks. A particular suggestion had been received with politeness that on reflection had been the polite end of dismissal. The signal was available; it was just not visible at the time, against the noise of the caseload.
This is not a problem that resolves itself through more careful clinical attention. The information density in a session is high; the therapist's attentional bandwidth is finite; the comparison data needed to spot the pattern is held in memory across multiple sessions for multiple clients simultaneously. Asking the therapist to do better pattern detection by hand is asking them to do something the cognitive load of their work makes structurally difficult.
What signals actually matter
The detection problem is partly resolved by externalising the comparison data — making the relevant signals visible alongside each other in a form the therapist can read at a glance before the session opens.
Several signals carry useful information about emerging ruptures. None is decisive on its own; the pattern across them is what carries diagnostic weight.
ROM trajectory deviation. A client whose session-by-session ROM scores have plateaued or deteriorated after an initial improvement is a client who may be experiencing the kind of stalling that often precedes withdrawal. The session-by-session ROM piece takes this up in its own right; the link to rupture detection is that the not-on-track signal is itself diagnostically interesting for alliance, not only for treatment response.
ROM engagement drop. A client who has been completing pre-session measures reliably and then misses one, completes a measure perfunctorily, or skips items they previously answered is exhibiting a different kind of signal from a deteriorating score. The score change reflects symptom trajectory; the engagement change reflects the relationship to the measure, which is part of the relationship to the therapy. Engagement drops, in clinical experience, are often earlier signals than score deteriorations.
Homework engagement drop. This connects to the homework evidence literature. The Kazantzis effect on homework quality is not just about treatment response; the quality of homework engagement is itself a window into the working alliance, because homework is the joint task whose execution depends on continued alignment on the work being done. A drop in homework engagement, particularly when symptom severity has not improved enough to justify it, is one of the more reliable withdrawal-rupture signals available.
Attendance reliability change. A client who has been reliable and starts arriving late, requesting reschedules, missing sessions without notice, or asking to space sessions further apart is exhibiting behaviour that often signals an alliance issue. The signal is noisy at a single occurrence; the pattern across two or three weeks is more informative.
Within-session content change. The harder signal to instrument but the most diagnostically useful one for an experienced clinician — the client who used to bring spontaneous material now waits for prompts, the client who used to disagree with the formulation now agrees with everything, the client who used to question the rationale now does not.
The first four can be aggregated and surfaced ahead of the session by infrastructure. The fifth remains a within-session observation, but the other four set up the within-session observation by telling the therapist what to attend to as the session opens.
What good rupture-detection infrastructure looks like
The instrumentation problem here is similar in shape to the ROM problem and the homework problem. The data the therapist needs in order to spot the rupture is, in principle, capturable. It is currently scattered across chart notes that take minutes to retrieve mid-session, measure spreadsheets that may or may not be up to date, mental tallies of homework engagement that depend on therapist memory, and attendance records that live in the booking system.
What good infrastructure does is aggregate this into a single pre-session view. Trajectory line for the relevant ROM measure. Engagement-with-measure line over time. Homework completion rate over the past four sessions. Attendance pattern. Any notable changes in the past two or three weeks flagged for the therapist's attention before they open the session.
The point is not that infrastructure replaces clinical judgement on alliance. It is that clinical judgement on alliance operates much more accurately when the relevant data is visible. A therapist who opens the session knowing that the client has missed the past two ROM completions, that homework engagement has halved, and that their score has plateaued for three sessions is in a different position from a therapist who realises in week six that things have felt slightly different for a while. The rupture conversation, when it happens, has data behind it rather than being an unsupported clinical hunch the client can readily disagree with.
This is the practical answer to the detection problem the Sharf and Eubanks evidence implies. The literature has established that alliance matters, that ruptures matter, that repair matters, and that detection is the weak link. The detection-end of the chain is mostly an instrumentation problem, and instrumentation problems are tractable.
Supervisia Companion surfaces these signals before each session.
Companion aggregates the between-session data — ROM scores and engagement, homework engagement and content, attendance patterns — into a pre-session view that flags the patterns the rupture literature identifies as important. The not-on-track flag for treatment response and the engagement-drop flag for relational signal sit alongside each other, in the form the session opening can actually use. The detection end of the alliance-rupture-repair chain is not a problem the therapist should be solving from memory across a caseload.
References
- Sharf, J., Primavera, L. H. & Diener, M. J. (2010). Dropout and therapeutic alliance: A meta-analysis of adult individual psychotherapy. Psychotherapy, 47(4), 637–645. DOI: 10.1037/a0021175.
- Eubanks, C. F., Muran, J. C. & Safran, J. D. (2018). Alliance rupture repair: A meta-analysis. Psychotherapy, 55(4), 508–519. DOI: 10.1037/pst0000185.
- Safran, J. D. & Muran, J. C. (2000). Negotiating the Therapeutic Alliance: A Relational Treatment Guide. New York: Guilford Press.
- Kazantzis, N., Whittington, C., Zelencich, L., Kyrios, M., Norton, P. J. & Hofmann, S. G. (2016). Quantity and quality of homework compliance: A meta-analysis of relations with outcome in cognitive behavior therapy. Behavior Therapy, 47(5), 755–772. DOI: 10.1016/j.beth.2016.05.002. PubMed: 27816086.
- Lutz, W., De Jong, K., Rubel, J. A. & Delgadillo, J. (2022). Measuring, predicting, and tracking change in psychotherapy. World Psychiatry, 21(2), 213–214. DOI: 10.1002/wps.20977.
Last updated: May 2026