RESEARCH DOSSIER
Adverse Effects, Deterioration, and Psychological Treatment Safety
Abstract. This fundamental research monograph examines negative effects, reliable symptom deterioration, treatment-related harms, patient perceptions, unethical clinical conduct, and failures of monitoring in psychological care. It reconstructs historical concepts, compares primary empirical evidence from randomized trials, cohort and survey research, psychometric analyses and reviews, and evaluates causal attribution, measurement, treatment fit, feedback systems, digital care and clinical governance. The synthesis proposes a multidomain evidence framework for assessing psychological treatment safety, emphasizes the differences between adverse event occurrence and demonstrated treatment causation, and identifies priorities for comparative intervention research and responsible Artificial-era decision support.
Keywords: psychotherapy safety; negative effects of psychotherapy; adverse events; treatment deterioration; reliable change index; treatment-attributable harm; outcome monitoring; patient-reported effects; clinical governance; digital psychological interventions
Introduction: Safety as a Dimension of Psychological Treatment
Psychological treatment safety concerns the probability, form, severity, duration, preventability, and attribution of unfavorable changes occurring in the course of psychological interventions. The object of this dossier encompasses individual psychotherapy, structured psychological therapies, guided and unguided digital interventions, and clinical systems through which treatments are selected, delivered, monitored, or discontinued. A treatment can improve the average symptom score of a randomized group while an identifiable minority of participants deteriorate, encounter burdensome events, or experience damaging conduct. Safety therefore belongs to the same evidentiary description of a treatment as efficacy, acceptability, and durability. The relevant scientific question is how patients fare across multiple outcomes and time points under a specified intervention compared with credible alternatives.
The literature contains a decisive methodological tension: the occurrence of an unwanted event during therapy is observable, whereas its causal origin usually remains uncertain. A depressive episode may intensify during appropriate care because of its natural course, an external crisis, concurrent treatment, changes in measurement, or consequences of the intervention. Symptom increases during exposure-based work may be temporary and anticipated, while sustained impairment, coercive behavior, violations of confidentiality, or exploitation implicate different safety categories. The distinction is central to the event-to-attribution framework of Linden (2013) and to the critical assessment of adverse events by Klein (2024).
This monograph organizes six questions. What counts as a negative treatment outcome? Which historical traditions produced the relevant evidence? How can symptom deterioration and harm be measured with adequate sensitivity? What do trials, service data, and patients' accounts establish about frequency and phenomenology? Which mechanisms, clinical contexts, and structural inequalities influence risk? And what evaluative designs permit prevention, responsible disclosure, and comparison of treatment strategies? The synthesis draws on primary outcome and feedback studies, scale-development research, national surveys, systematic reviews, qualitative synthesis, clinical guidelines, and methods standards. This is a critical narrative research synthesis, with explicitly identified source and inferential limits; it is neither a prospectively registered systematic review nor an original meta-analysis.
A safety-oriented synthesis also needs positive evidence about treatment benefit. Evidence that many psychological treatments work for many patients provides the correct clinical denominator for evaluating their risks. The existence of adverse outcomes does not invalidate comparative treatment efficacy, and the existence of efficacy does not settle adverse-event attribution. The 2025 contextual review by Rosendahl and colleagues emphasizes that definitions, informants, study designs, and comparison with symptom trajectories determine what a published estimate means. This framework connects patient-level outcomes to institutional accountability and the conditions under which treatment is delivered.
Conceptual Boundaries: Events, Effects, Harm, and Deterioration
An unwanted event is an unfavorable occurrence during or around treatment, identified without presuming a causal link. A treatment-emergent event is temporally associated with starting or continuing the intervention. An adverse treatment reaction is an unwanted effect judged plausibly caused by the treatment, and a side effect is an unwanted consequence of a correctly delivered treatment. Deterioration denotes worsening on an outcome domain relative to a defined baseline and can be classified statistically, clinically, or both. Nonresponse denotes failure to improve by a specified criterion; it need not imply a decline. A serious adverse event is a distinct severity and regulatory category, whose operational definition must be established in the study protocol. Malpractice and unethical misconduct introduce failures of care, professional competence, and ethical responsibility that should be evaluated separately from an unwanted effect of competent intervention.
These terms have no interchangeable prevalence. A patient may experience emotional distress during a session while achieving lasting improvement; another may report lower symptoms but serious relational or financial damage; a third may show reliable symptom worsening whose cause cannot be confidently attributed to the treatment. The taxonomic separation introduced by Linden (2014), the later framework of Klein (2024), and the synthesis of Rosendahl et al. (2025) support a multidimensional outcome classification rather than a single omnibus harm rate.
Treatment acceptability requires its own interpretation. Premature discontinuation may follow harm, symptom recovery, economic hardship, transportation problems, dissatisfaction, competing responsibilities, inaccessible scheduling, or improvement that makes further sessions unnecessary. Swift (2012) demonstrates substantial variation in dropout by setting, definition, and patient characteristics. Consequently, dropout is a vital implementation and tolerability indicator but not a validated substitute for adverse-event measurement.
A defensible safety estimate must state the event definition, whether causation was adjudicated, the period of observation, the unit of analysis, the population, the treatment, the comparator, the ascertainment method, and the denominator. In the absence of these details, apparent contradictions between studies frequently reflect different measurement targets. Safety researchers should also distinguish event counts from the proportion of patients with one or more events, and separate transient experiences from enduring consequences. These are methodological conditions for meaningful interpretation, not merely reporting preferences.
Historical Foundations of Treatment Harm Research
The emergence of contemporary psychotherapy safety research followed several research traditions that were initially separated. Efficacy trials compared treated and untreated groups, clinical-significance research investigated individual outcomes, service research described early termination and worsening, and clinical ethics addressed harmful practitioner conduct. These programs gradually converged because group-average improvement concealed clinically meaningful dispersion within treatment arms. Early interest in negative treatment outcomes also appeared in the psychotherapy process and therapist-effect literature, where responsibility for differences in patient trajectories became an empirical question.
The 2007 analysis by Lilienfeld (2007) placed potentially harmful psychological treatments within the broader problem of evidence-based clinical decision-making and warned that apparent plausibility can conceal adverse effects. Barlow (2010) emphasized the methodological blind spots created by a near-exclusive emphasis on average benefit and urged patient-focused assessment of negative outcomes. Subsequently, Jonsson (2014) systematically reviewed the reporting of harms in randomized controlled trials of psychological interventions, showing how frequently adverse events were poorly monitored or described. These contributions supply complementary historical foundations: distinguishing potentially harmful techniques, assessing individual deterioration, and improving transparent trial reporting.
Jacobson (1991) supplied an influential statistical approach to clinically meaningful change, linking reliable individual differences to a clinical cutoff. Their contribution was not a general theory of treatment-induced harm. It made worsening detectable against error in repeated measurement, providing an essential component of later outcome monitoring. The Lambert (2001) then tested whether formal monitoring could improve results among patients predicted to respond poorly. Hannan (2005) investigated the discrepancy between clinician prediction and algorithmic warning signals. These studies established a practical research line on recognizing deterioration before termination.
Another historical line developed from the consequences of seemingly plausible early interventions following trauma. The Rose (2002) found that single-session psychological debriefing did not prevent PTSD; one included trial suggested increased PTSD risk at a later follow-up. The result illustrates why theoretical plausibility and compassionate intent cannot replace comparative evidence. Current National Institute for Health and Care Excellence. (2018). Post-traumatic stress disorder (NICE guideline NG116). https://www.nice.org.uk/guidance/ng116/chapter/recommendations (2018) advises against psychologically focused debriefing for PTSD prevention or treatment. The clinical lesson concerns a specific procedure and indication, rather than a claim that discussing trauma in evidence-based therapy is intrinsically harmful.
The third historical line focused on classification and active detection. Duggan (2014) audited adverse-event recording in publicly funded psychological-treatment trials. Linden (2013) described the unwanted-event-to-adverse-reaction assessment procedure. An expert Rozental (2014) urged routine investigation of deterioration and subjective negative effects. The Rozental (2016) offered a structured patient-reported vocabulary, later examined psychometrically by Rozental (2019).
Contemporary research incorporates large patient surveys, multi-study qualitative analyses, and reviews of adverse-event reporting in randomized trials. The Crawford (2016) evaluated perceived lasting adverse effects, while Vybíral (2024) synthesized clients' negative experiences. Two systematic reviews of trial harms appeared in the January 2025 issue of Psychotherapy Research: Klatte (2025) and Honkalampi (2025). Their existence marks a move from incident anecdotes toward comparable safety evidence, although definitional and reporting heterogeneity continue to constrain pooled inference.
The Scientific Problem of Causal Attribution
Let an outcome under psychotherapy be compared with the outcome the same person would have had at the same time under a specified alternative. This counterfactual contrast defines the causal effect conceptually, but an individual can reveal only one of those outcomes. A patient's worsening during treatment thus supplies evidence about trajectory, not by itself proof that treatment produced the deterioration. Random assignment helps identify differences in the distribution of outcomes between groups, while within-person measurements help identify change and potential warning signs. The two designs answer different questions.
Causal attribution requires an explicitly defined comparator: waitlist, usual care, another psychotherapy, attention control, pharmacotherapy, or no intervention. Each has interpretive consequences. A waitlist comparison can illuminate average differences during a defined interval but may introduce expectancy and help-seeking contrasts. An active-treatment comparison answers a relative safety question between specified interventions rather than whether either produces harm. A naturalistic service sample provides ecological validity but brings selection effects, co-interventions, clinical heterogeneity, and differential follow-up. The taxonomy of Klein (2024) makes attribution a separate step after observing an unwanted event.
Attribution is further complicated by delayed and context-dependent effects. A therapeutic intervention may produce early discomfort followed by symptom improvement, create a temporary relapse with longer-term recovery, or coincide with deterioration caused by an unrelated event. Patients can experience consequences across domains that standard symptom questionnaires never register: relational conflict, loss of trust, diminished occupational functioning, financial burden, or compromised autonomy. Longitudinal designs therefore require pre-specified measurement intervals, qualitative event narratives, concurrent-treatment records, and adverse-event follow-up, including after treatment termination where feasible.
Independent assessment can limit conflicts of interest, although a reliable attribution decision still needs transparent rules and uncertainty categories. The ethically relevant distinction between ordinary treatment side effects, avoidable misapplication, and abusive behavior additionally requires review of the actual process of care, not only outcomes. Investigators should record the clinician's technique, treatment indication, expected and unexpected reactions, adherence, patient priorities, and contextual disruptions without assuming that a manualized treatment is safe for every person or circumstance. These distinctions remain central to modern psychotherapy harm contextualization.
The Unit of Observation and Multidomain Outcomes
Safety can be analyzed at the level of the event, patient, session, course of treatment, clinician, service, or treatment program. Patient-level proportions and session-level incidence can differ materially when individuals attend unequal numbers of sessions. A treatment period with more intensive contact provides more opportunities for detection and documentation, just as a longer follow-up increases the probability of recording an event. Time at risk, exposure duration, and missing visits therefore belong in the estimand alongside severity and attribution.
Outcome domains should include targeted symptoms, generalized distress, functioning, quality of life, adverse experiences, self-harm and suicidality where relevant, and patient-defined consequences. Patient reports capture experiences inaccessible to symptom scales. Clinical interviews and independent records can clarify severity and immediate safety implications. Informants may disagree because they observe different domains or because of unequal incentives to report events. Vybíral (2024) show how patient testimony widens the domain beyond diagnostic scales, whereas Rozental (2016) provide an instrumented approach to unwanted consequences.
Time should be represented as a sequence rather than a single pre-post contrast. The minimum useful structure for many research designs includes baseline, an early-course assessment, a later treatment assessment, endpoint, and follow-up. Each observation may have distinct sources of error and missingness. A person whose treatment ends after a crisis or a rupture may be systematically absent from endpoint analyses. Treating only completers as the safety population can therefore produce a falsely reassuring estimate. Recording reasons for missingness and performing sensitivity analyses are essential.
Interpreting improvement requires both the intended target and the wider lived consequences of treatment. A narrowly successful symptom reduction can coexist with impaired functioning or an undesirable change in social relationships; conversely, a temporarily elevated symptom score during an effective intervention can coexist with gains in autonomy. These outcomes should be reported separately before any composite assessment is made. The general principle is that psychological treatment safety is an outcome architecture involving multiple evaluative dimensions.
Methods and Evidence Hierarchy for This Dossier
The research strategy follows the evidentiary hierarchy appropriate to each question. Randomized studies and controlled comparisons offer the strongest conventional design for average relative treatment effects when randomization, follow-up, and measurement are sound. Cohort studies of routine services reveal real-world trajectories, rare exposures, and implementation variation, but causal inference requires additional adjustment and careful sensitivity analysis. National surveys estimate self-reported experiences among respondents. Qualitative studies explain mechanisms and classifications visible to participants but do not alone provide population prevalence. Psychometric validation establishes whether instruments represent the intended constructs, while clinical guidelines translate bodies of evidence and professional judgment into service recommendations.
This monograph uses an interpretive, critical narrative synthesis rather than a reproducible database search with a registered eligibility protocol. It places accessible original reports, peer-reviewed systematic reviews, research-methods papers, and current official guidance in direct dialogue. Where a source is accessible through an abstract alone, the treatment of its results is bounded by that abstract. Bibliographic details are supplied in the final References section, and the narrative links are placed adjacent to attributed claims. These safeguards distinguish evidence review from the appearance of exhaustive coverage.
The 2025 review by Klatte (2025) and the Honkalampi (2025) receive particular weight on trial reporting because they directly examine this topic. The extensive patient survey of Crawford (2016), the qualitative synthesis by Vybíral (2024), and the measurement articles by Rozental (2016) provide complementary perspectives. Current National Institute for Health and Care Excellence. (2022). Depression in adults: Treatment and management (NICE guideline NG222). https://www.nice.org.uk/guidance/ng222/chapter/recommendations (2022) and National Institute for Health and Care Excellence. (2018). Post-traumatic stress disorder (NICE guideline NG116). https://www.nice.org.uk/guidance/ng116/chapter/recommendations (2018) supply a dated and jurisdiction-specific reference for patient safety duties.
The underlying principle is evidentiary proportionality. Numerical estimates are retained only with the corresponding measure and sample. A treatment-induced effect is distinguished from an event observed during treatment. A statistical signal is distinguished from a patient-centered account, and a guidance recommendation is distinguished from the trial results on which it was based. Later chapters carry those distinctions through research on risk monitoring, intervention differences, professional conduct, digital treatment, implementation, and the prospective science of safer psychological care.
Reliable Deterioration and the Logic of Clinical Significance
A raw increase in a symptom score is an observable measurement difference. Whether it is reliable depends on the instrument's measurement error, population norms, and the prespecified threshold for change. The widely used Jacobson (1991) divides an individual change score by the standard error of the difference. Under the usual classical-test assumptions, the standard error of measurement can be approximated as the baseline standard deviation multiplied by the square root of one minus reliability, and the standard error of the difference as the square root of two times that measurement error. A difference exceeding the commonly used absolute threshold of 1.96 is conventionally described as statistically reliable. Reliability estimates, reference populations, the direction of scoring, and the timing of measures must all be reported.
Reliable deterioration and clinically significant change are related but separate. Reliability asks whether the observed change exceeds an error-based threshold; clinical significance asks whether the change matters relative to a clinical distribution, a normative cutoff, a patient-important threshold, or functional consequences. A patient can cross a clinically meaningful threshold with an uncertain measurement change, and a patient can worsen reliably on a scale while remaining in the same broad diagnostic range. Neither classification demonstrates that psychotherapy caused the change. The original Jacobson (1991) framework was designed to illuminate patient-level outcomes rather than to substitute for causal comparisons.
The classical reliable-change index assumes a reasonably stable measurement scale and an appropriate error model. Practice effects, floor and ceiling effects, regression to the mean, repeated observations, and differences in the population used to estimate reliability can undermine naive classification. If clients enter with exceptionally high symptoms, some apparent improvement may follow ordinary fluctuations; if scores are near the scale floor, deterioration is easier to register than further improvement. The interpretation of change should therefore incorporate the actual metric, symptom history, and clinical context. Multilevel approaches and repeated-measurement models offer additional ways to represent error and temporal structure, but their assumptions are also substantive and should be tested.
A robust outcome report describes the full distribution: reliable improvement, no reliable change, and reliable deterioration, alongside clinically defined recovery and follow-up outcomes. Reporting only the mean treatment effect can mask different trajectories. Conversely, reporting a deterioration fraction without the denominator or instrument can suggest a level of precision that the data do not support. The safety value of reliable-change analysis lies in flagging meaningful within-person deviation for further assessment, not in pronouncing a cause or moral judgment about the practitioner.
Patient-Reported Negative Effects and Psychometric Instruments
Direct patient questioning captures adverse experiences that may never manifest in a symptom-total score. The Rozental (2016) emerged from efforts to define and empirically organize unwanted experiences of psychological treatment. Its 2016 exploratory factor analysis examined dimensions that included worsening symptoms and perceived problems with treatment quality or the therapeutic relationship. The 2019 Rozental (2019) continued evaluation of the instrument. These studies provide a research resource for identifying patient-perceived events, with the qualification that instrument validation in one population does not automatically prove equivalence across languages, diagnoses, age groups, or delivery formats.
A questionnaire's scope determines what it can find. Asking whether treatment made a person feel worse is different from asking whether a particular unfavorable event occurred and then asking whether the person attributes it to treatment. Instruments may conflate occurrence, distress, duration, severity, and perceived causal attribution unless they keep those dimensions apart. The distinction affects quantitative interpretation: a frequently reported but brief and manageable reaction differs materially from a rare, severe, persistent adverse experience. A good research protocol therefore records discrete domains and allows narrative clarification. Patient attribution is itself a scientifically relevant perception, but its accuracy is not guaranteed by sincerity.
The patient is the primary witness to certain categories of experience, including shame, alienation, perceived pressure, unwanted dependence, or feeling misunderstood. Therapist ratings may add process detail, and independent assessment may address serious clinical events. These perspectives require triangulation. The Vybíral (2024) illustrates the wide variety of events that emerge when researchers ask participants to describe treatment experiences in their own words. Structured instruments and narrative elicitation are therefore complementary methods rather than mutually exclusive alternatives.
Measurement can itself influence reporting. Patients may hesitate to criticize a clinician on whom they depend for treatment, fear loss of access, wish to avoid conflict, or lack language to describe boundary violations. Conversely, retrospective reports after a disappointing outcome may be colored by current distress and reconstructed expectations. Independent anonymous assessment can reduce some pressures, while prospective collection reduces some recall problems. Researchers should record who asked the questions, where and when they were asked, whether confidentiality was credible, and whether patients could report concerns after leaving the service.
What Randomized Trials Reveal—and Conceal
Randomized controlled trials are uniquely useful for estimating average differential harm between interventions when their allocation and ascertainment are credible. Their value depends on reporting all relevant unfavorable outcomes in both arms, including events unrelated to the treatment assignment. Trials that record only clinician-judged, treatment-related events risk overlooking harms whose attribution is still unsettled. Trials that suppress adverse outcomes because the protocol emphasizes a primary efficacy measure make subsequent risk–benefit interpretation incomplete.
Klatte (2025) systematically examined adverse events in psychotherapy randomized trials. Their review reported that adverse events or symptom deterioration were explicitly reported in 60% of reviewed studies. Reporting approaches varied substantially, and association between the events and the study treatment remained unclear for most events. The review described adverse events in more than one in ten participants and serious adverse events in more than one in twenty-one participants within its reviewed trial evidence. These are observational frequencies of recorded events under heterogeneous protocols, not a universal rate of psychotherapy-induced harm or proof that serious events were caused by treatment.
The separate review by Honkalampi (2025) addresses negative effects across randomized psychological-intervention studies. Together these reviews underscore the cost of inconsistent definitions and collection procedures. In particular, failing to mention adverse effects in a paper cannot be interpreted as demonstrating that no such events happened. The literature also risks differential surveillance: trials with active questioning can appear less safe than trials that collect almost nothing. Future pooled analyses require alignment of ascertainment, exposure, event definitions, follow-up, and denominators before the reported fractions can be meaningfully compared.
An important causal detail is that some serious adverse events reflect the underlying disorder or broader life circumstances. Hospitalization, a suicide attempt, or a major deterioration during a depression trial requires immediate clinical attention and full reporting, but randomization and case review are necessary for causal inference. In trials lacking a credible adverse-event collection strategy, even large treatment arms cannot compensate for missingness. The Junqueira (2023), published with explanatory material in 2023, offers a detailed reporting framework applicable to trial methodology, with adaptation to the distinctive risks of psychological interventions.
Survey Evidence and the Meaning of Prevalence
National and service-based surveys document the range of adverse experiences that patients recognize after receiving treatment. In a cross-sectional survey of users of 184 services in England and Wales, Crawford (2016) reported 763 respondents, or 5.2% of 14,587, who endorsed lasting bad effects of psychological treatment. The endpoint was a respondent's report of lasting harm, and the design cannot determine whether each event was caused by treatment. The sample includes treatment recipients from services and represents the respondents captured by this specific survey, not all psychological-treatment recipients worldwide.
Crawford and colleagues also reported associations with several respondent characteristics and with treatment information. Participants indicating that they had received enough information before treatment were less likely to report lasting negative effects, with an adjusted odds ratio of 0.65 and a 95% confidence interval of 0.54 to 0.79. Those uncertain about the type of treatment received were more likely to report such effects, with an odds ratio of 1.51 and a 95% confidence interval of 1.22 to 1.87. The observational design does not establish that information provision itself prevented harm. This association may also reflect therapeutic quality, service organization, communication, or participant characteristics.
The fraction of people reporting any unwanted experience can be substantially higher than the fraction reporting enduring harm or reliable symptom worsening. The Klein (2024) summarizes patient-reported adverse events under broad definitions; the Vybíral (2026) notes extreme between-study variation attributable to different definitions, methods, and populations. Comparing such percentages as though they measured the same clinical object creates a false impression of disagreement. A prevalence estimate is meaningful only together with its measurement question, population, observation period, and response rate.
Surveys also make visible a distinctive missing-data problem: people with the worst experiences may disengage and never respond, while people motivated by particularly negative experiences may be more likely to answer a retrospective questionnaire. Some experiences remain unreported because of shame or concern about the consequences of speaking. Prospective population-linked follow-up, anonymous patient-report opportunities, and systematic collection from both treatment completers and early leavers can improve representativeness, although none eliminates selection bias by itself.
Qualitative Evidence: The Experience of Harm
The subjective form of a negative event is part of its scientific significance. Vybíral (2024) synthesized 936 statements from 51 qualitative primary studies into 21 categories. Their broad clusters concerned therapist misbehavior, hindering dimensions of the therapeutic relationship, poor treatment fit, and negative treatment impacts. This meta-synthesis supplies an organized description of the vocabulary and domains of negative experience. It does not establish the population prevalence of each cluster because qualitative recruitment, sampling, coding, and publication differ from the design of incidence studies.
Therapist misbehavior can encompass disparaging conduct, inappropriate boundary crossing, coercion, or disregard for the patient's stated values. A hindering relationship may involve invalidation, pressure, persistent misattunement, or a rupture that remains unresolved. Poor fit involves mismatch between intervention methods and the person's goals, cultural context, capabilities, preferences, or current safety needs. Negative impacts may include symptom increase, deterioration in personal relationships, perceived dependency, or practical disruption. These descriptions should be treated as analytic categories rather than universally established causal pathways.
Qualitative accounts offer access to mechanisms that quantitative scales only imply. For example, a patient may identify the moment a therapeutic interpretation was experienced as shame-inducing or a standardized procedure conflicted with an important cultural meaning. Such accounts suggest testable process hypotheses and potential improvements in consent and monitoring. They also show the limits of interpreting the same clinical indicator across all settings: a high distress score may reflect emotional activation, unaddressed safety risk, or worsening functioning, each calling for different research and clinical interpretation.
The analytical challenge is to honor experience while maintaining careful causal claims. An interview can accurately document a participant's perception and detailed event sequence without establishing that no alternative explanation exists. Independent clinical records, temporally structured measures, and clinician interviews can supplement the story. Researchers should not make access to validation or care dependent on the participant's ability to establish scientific causation. That is an ethical concern distinct from the question of how a trial attributes effects.
Attrition, Missingness, and Selective Safety Evidence
A trial may report excellent average outcomes among completers while losing the participants who experienced the greatest distress. Missing outcome data are consequently a substantive threat to both efficacy and safety inference. When missingness depends on an unobserved worsening outcome, conventional complete-case estimates can be biased. Even intention-to-treat analysis requires an explicit strategy for missing outcomes and assumptions about the missing-data mechanism. Sensitivity analyses should explore plausible departures from those assumptions rather than hiding them behind one imputed value.
Swift (2012) reviewed 669 adult psychotherapy studies involving 83,834 clients and reported a weighted dropout estimate of 19.7% with a 95% confidence interval of 18.7% to 20.7%. The authors found important moderation by study characteristics and dropout definitions. Those figures establish the practical significance of attrition, not the proportion who left because of harm. Failure to discriminate adverse termination from recovery-related or logistical ending risks confusing access, tolerability, implementation, and safety.
Negative effects are also susceptible to selective publication and outcome reporting. A trial's public abstract may highlight symptom benefit even when the protocol anticipated adverse-event recording that is missing from the final report. A negative event may be classified under a broad category or omitted because it was judged unrelated. Trials with inadequate harm reporting can distort a systematic review irrespective of how carefully its authors search. Registry protocols, dated analysis plans, published supplementary material, and transparent denominator accounting help expose these distortions.
A mature safety review makes clear which information was actively collected, which was passively observed, which was unavailable, and which was judged causally connected. Researchers can then distinguish lack of evidence from evidence of a low event rate. That difference is essential when comparing protocols across clinical disciplines and when making treatment decisions using incomplete literatures.
Treatment Mechanisms, Emotional Activation, and Temporal Dynamics
An intervention may change how a person attends to threatening situations, remembers distressing experiences, behaves in relationships, or interprets personal difficulties. Any method that deliberately changes psychological functioning can generate experiences that require monitoring. The possibility of harm is especially relevant when therapeutic demands temporarily exceed available coping resources, when the patient's situation changes, or when the chosen technique is poorly fitted to current needs. Mechanistic evaluation must distinguish the intended process of change, the accompanying discomfort, and a sustained adverse outcome. Treatment-specific mechanisms should be evaluated in the literature for that treatment rather than inferred solely from generic psychological theory.
Exposure-based approaches provide a clear example of why timing matters. They may invite a patient to confront feared cues under conditions intended to support new learning. Distress during such work can be a predictable component of the intervention and does not establish deterioration. Persistent symptom escalation, increased avoidance outside sessions, destabilization, or impaired everyday functioning would warrant a different evaluation. A rigorous safety study would identify the target mechanism, define permissible and concerning trajectories before data collection, monitor outcomes between sessions, and compare those trajectories with an appropriate alternative. Neither a single moment of distress nor the clinician's expectation that distress is useful should settle the causal question.
The same temporal principle applies to cognitive work, interpersonal interventions, behavioral activation, and trauma-focused care. Challenging a deeply held belief may be experienced as clarifying or invalidating depending on formulation, timing, consent, and therapeutic alliance. An intervention that temporarily changes relationships may yield beneficial autonomy, unwanted conflict, or both. Research should examine symptom trajectories alongside the person's own evaluation of the change and independent functional outcomes. The reviews by Klein (2024) and Rosendahl et al. (2025) reinforce the need to contextualize negative effects before attributing them to treatment.
Mechanisms should be treated as hypotheses that can be tested. Investigators can assess temporal ordering, mediator change, treatment fidelity, dose, individual moderators, and the effects of modifying an intervention component. Cross-sectional associations between a process variable and later harm may reflect confounding or reverse causality. Even a statistically significant mediator in a randomized intervention trial may not identify a causal mediator if the mediator itself was not experimentally controlled. Safety science benefits from theoretically informed mechanisms while preserving the independence of the causal evidence.
The Therapeutic Relationship, Ruptures, and Repair
The therapeutic relationship has multiple dimensions: agreement about goals, agreement about tasks, and the interpersonal bond. Ruptures in collaboration, trust, or understanding may be associated with disengagement and negative experiences. Their meanings depend on context. A disagreement that is recognized and worked through may be part of constructive collaboration, whereas persistent invalidation, coercion, intimidation, or disregard for patient concerns can undermine care. The Vybíral (2024) identifies relationship difficulties as one major family of patient-reported negative experiences.
A rupture is not equivalent to malpractice. Research must distinguish ordinary conflict over treatment tasks from professional misconduct, and it must distinguish an alliance score from the patient's actual freedom to disagree or terminate care. These distinctions matter in high-dependency relationships, where patients may be reluctant to disclose dissatisfaction or may reinterpret coercive interactions as necessary treatment. Clinician accounts of adequate rapport can differ from the patient's report of pressure or disregard. Independent and confidential feedback channels may reduce that mismatch.
The Eubanks (2018) examines alliance rupture repair and treatment outcome. It is relevant to the safety question because responsiveness to strain may influence whether an adverse interaction persists, although favorable outcome associations alone do not prove that rupture repair causally prevents harm. Measurement should capture when a rupture occurred, how therapist and patient perceived it, whether an attempt at repair followed, and what happened to functioning and the relationship over time. Studies that code recorded sessions have advantages for process observation, while patient interviews can identify harms occurring outside the observable session.
A stronger research design would combine repeated alliance measures, event-specific patient reports, session-level coding, reliable symptom change, and longitudinal outcomes. It would examine whether the effect of conflict differs with therapist behavior, severity of symptoms, and the patient's prior experiences with authority and care. The relevant goal is to understand interactional pathways by which collaborative work becomes supportive or harmful and to distinguish these pathways from the natural course of the clinical problem.
Treatment Fit, Choice, and Informed Consent
Treatment fit concerns whether the intervention's rationale, demands, setting, pace, goals, and resource requirements align with the person's clinical presentation and values. It includes indication and contraindication, but also comprehension, cultural meaning, accessibility, prior treatment experience, and the patient's ability to participate in decision-making. Fit can change during treatment as symptoms evolve or as new information emerges. A method supported by strong average efficacy may still be a poor choice for a specific patient at a particular time. That proposition requires case-specific assessment, not a global ranking of named therapies as inherently safe or unsafe.
Informed consent has both ethical and empirically researchable dimensions. Patients require information about likely benefits, alternative approaches, relevant burdens, uncertainty, confidentiality, and how concerns can be communicated. Merely collecting a signature does not show that risks were understood. Crawford (2016) found that sufficient pretreatment information was associated with fewer self-reported lasting negative effects, but the cross-sectional association cannot determine whether improved consent directly reduced these outcomes. A future intervention trial could test consent processes and examine understanding, autonomy, retention, and adverse experiences in parallel.
Patient preferences are relevant not only to satisfaction but to exposure and adherence. A person who finds a technique unacceptable may terminate early, participate under pressure, or disengage in ways that make treatment ineffective. Researchers should study the reasons for such responses rather than treating unwillingness as evidence of pathology. Clinical choice also depends on what is available: long waits, lack of language-concordant care, or financial barriers may turn an otherwise sound intervention into an inaccessible option. Service-level safety extends beyond the technical content of the psychotherapy session.
A transparent consent procedure can acknowledge uncertainty without exaggerating risk. It should distinguish ordinary challenges that may occur during treatment from severe or unacceptable events, provide a means of reviewing the treatment plan, and identify how a patient can seek independent help if necessary. Current National Institute for Health and Care Excellence. (2022). Depression in adults: Treatment and management (NICE guideline NG222). https://www.nice.org.uk/guidance/ng222/chapter/recommendations (2022) emphasizes shared choice, review of treatment response, adverse effects, and access to support. These provisions offer a model of clinically governed decision-making rather than evidence that any one consent format eliminates treatment-related harm.
Therapist Conduct, Professional Boundaries, and Misconduct
Research into psychological treatment safety must explicitly include professional behavior. Unethical sexual conduct, exploitation, breaches of confidentiality, discriminatory treatment, intimidation, and other boundary violations are harms of care delivery that cannot be explained away as ordinary side effects of a correctly administered treatment. Their occurrence raises questions about professional oversight, reporting routes, power asymmetry, and organizational response. The conceptual framework of Linden (2013) and Klein (2024) separates unwanted effects of competent therapy from malpractice and unethical behavior.
The observational unit differs between treatment technique and misconduct. Technique-related harm is often studied by comparing protocols or trajectories. Misconduct may require confidential complaint records, independent investigations, disciplinary documentation, and qualitative research with affected clients. Such evidence has its own selection effects. Official complaints are likely to capture only some violations, and allegations, substantiated findings, and retrospective recollections must be clearly distinguished. A study can report the existence and characteristics of a complaint without assuming its adjudicated truth or completeness.
Professional power can shape the event itself and the conditions of detection. Some clients depend on a single clinician for access to specialized treatment; others fear that disagreeing will damage their records or future care. A treatment system that relies exclusively on the accused clinician to solicit and classify adverse-event reports is structurally vulnerable to underdetection. Independent routes, transparent escalation procedures, secure records, and separation of complaint review from clinical authority are therefore important system-design considerations. Their effectiveness should be empirically evaluated rather than presumed.
Clinical training and supervision also belong in the safety architecture. National Institute for Health and Care Excellence. (2022). Depression in adults: Treatment and management (NICE guideline NG222). https://www.nice.org.uk/guidance/ng222/chapter/recommendations (2022) specifies regular clinical supervision and monitoring of competence for personnel delivering psychological interventions. Competence is not a guarantee against harm; it is a modifiable condition of treatment delivery. Safety studies should examine actual supervision quality, fidelity to indicated treatment, the capacity to recognize deteriorating outcomes, and whether institutions act appropriately when patients report concerns.
The Distinctive Example of Psychological Debriefing
Single-session psychologically focused debriefing after trauma is a historically important case because it was adopted in some settings before compelling evidence of benefit was established. The Rose (2002) found no evidence that such intervention prevented PTSD in the reviewed trials, and one trial reported increased PTSD risk at a one-year follow-up. These findings cannot be generalized to every trauma-focused therapy, and the evidence base should not be reduced to a claim that early psychological support universally causes PTSD.
The scientific importance of the case is the inferential error it reveals. An intervention can possess a persuasive rationale, high acceptability among providers, and a compassionate institutional purpose without outperforming a credible comparison condition. If a treatment is administered immediately after trauma, ordinary patterns of distress and spontaneous recovery make before–after impressions especially misleading. The design must establish what would likely have happened in the absence of that specific intervention.
The National Institute for Health and Care Excellence. (2018). Post-traumatic stress disorder (NICE guideline NG116). https://www.nice.org.uk/guidance/ng116/chapter/recommendations (2018) states that psychologically focused debriefing should not be offered for prevention or treatment of PTSD. This is a recommendation about a defined intervention and clinical indication, developed from guideline evidence review, not a blanket proscription on talking with trauma survivors or providing appropriate early support. In scholarship, clinical policy statements must be represented at their actual level of specificity.
The wider lesson is that intervention safety depends on empirically demonstrated outcomes, temporal context, and the appropriate comparator. Historical popularity cannot be used as proof of effectiveness. Where a method shows no established benefit and plausible harm, its benefit–risk profile demands re-examination even if the proposed psychological mechanism appears reasonable.
Symptom Deterioration, Self-Harm, and Serious Events
Psychological treatment trials involving severe depression, suicidal crises, trauma-related disorders, or other conditions with meaningful baseline risk need a plan for identifying serious clinical deterioration. This requirement concerns study design and professional responsibility rather than a suggestion that a particular therapy usually causes such events. A safety protocol must specify seriousness, urgency, duration, adverse-event grading, reporting responsibility, and relevant follow-up. It should record events in all randomized groups irrespective of initial attribution.
The causal challenge is especially acute for suicide attempts, emergency presentations, hospitalization, or serious self-injury. These events can occur during treatment because the underlying disorder is dangerous, because of external circumstances, or because care was inadequate or counterproductive. Clinical assessment must address the event immediately, while research attribution should follow a transparent process using temporality, treatment exposure, comparator information, and independent review. Safety conclusions should avoid implying that a reported event was treatment-induced solely because it appeared in the treatment arm. The Klatte (2025) explicitly notes the uncertain treatment association of most reported adverse events.
For adults with depression, National Institute for Health and Care Excellence. (2022). Depression in adults: Treatment and management (NICE guideline NG222). https://www.nice.org.uk/guidance/ng222/chapter/recommendations (2022) recommends direct assessment of suicidal ideation and intent, appropriate support when risk increases, monitoring of symptoms and harm, and early review of treatment response. These are jurisdiction-specific clinical recommendations that integrate multiple treatment modalities. The present dossier analyzes them as institutional safeguards and does not convert them into individualized medical instructions.
Studies of treatment safety should report both the number of people affected and the number and severity of events, preserving confidential details and meaningful case characterization. Independent monitoring may be necessary when the treating team faces conflicting incentives. Comparative evaluation must account for baseline risk, treatment arm, observation time, intervention changes, and incomplete follow-up. Serious-event collection is an essential component of a credible treatment trial, but it is not sufficient by itself to determine causation.
Internet-Delivered Psychological Treatments and Digital Safety
Digital interventions introduce delivery conditions that affect both risk and its detection. Structured online cognitive behavioral interventions may be guided by clinicians or used with little ongoing human contact. The content may be standardized, but actual engagement is uneven, and clinical oversight may vary greatly between platforms. Safety assessment should therefore record the intervention's degree of human support, its screening procedures, frequency of contact, alert mechanisms, privacy protections, and escalation options. Results from one guided program cannot automatically be extended to every app or conversational agent described as mental-health support.
A Rozental (2014) called for systematic monitoring of negative outcomes in internet interventions, including both symptom deterioration and patient-perceived adverse experiences. This was a methodological consensus developed using a Delphi process, not an empirical proof of a uniform adverse-effect rate. The Rozental (2015) explored patients' experiences of negative effects in online care. Together these studies establish that delivery mode and technical design are relevant explanatory factors rather than merely neutral channels for established treatments.
Comparative individual-participant evidence helps distinguish symptom deterioration during an intervention from the trajectory under a control condition. In a Rozental (2017), 29 internet-based cognitive behavioral therapy trials included 2,866 participants. Reliable deterioration on primary outcomes was reported for 122 participants (5.8%) in treatment conditions and 130 (17.4%) in control conditions. The control group showed higher odds of deterioration (OR 3.10, 95% CI 2.21–4.34). These figures describe the included trial populations and the study's outcome definition, and should not be translated into a universal incidence of therapy-related harm. This analysis demonstrates that deterioration can occur in both trial arms and that comparative designs can reveal net protection despite observable worsening in some treated individuals.
The lack of direct clinical observation can complicate interpretation of rapidly worsening symptoms, suicidal crises, or misunderstanding of therapeutic exercises. On the other hand, digital systems can enable frequent questionnaires and longitudinal detection of changes that are difficult to obtain in episodic in-person care. The relevant question is whether collected information is valid, reaches a responsible reviewer, and triggers an appropriate response. A notification that no trained professional receives cannot be treated as a completed safety intervention. A systematic research account should map both the data flow and the decision rights.
Digital safety includes confidentiality, data retention, secondary uses of sensitive information, transparency about automation, misleading claims of clinical competence, and dependency created by continuous interaction. These are distinguishable from clinical symptom deterioration even when they influence mental health. In the Artificial Era, AI-mediated interventions also raise questions about anthropomorphic trust, perceived authority, and the relation between simulated conversational continuity and qualified therapeutic responsibility. The scientific object is the human user's observable experience, the implemented system behavior, and empirically testable outcomes. Assertions about the inner psychological status of artificial systems belong to separate theoretical inquiry.
Children, Adolescents, and Developmentally Sensitive Monitoring
Treatment safety in younger populations requires developmentally appropriate measures, consent and assent procedures, confidentiality rules, involvement of appropriate caregivers, and accessible reporting mechanisms. Changes in symptoms and behavior may reflect development, school context, family stress, neurodevelopmental differences, and treatment exposure. Research should distinguish caregiver observations from the young person's own experience, particularly when they disagree about the meaning of an intervention or about whether the patient felt harmed.
A recent Munir (2026) directly examined how studies of self-help digital mental health interventions for young people monitored and reported adverse events and harms. Its importance for this dossier is methodological: it makes monitoring and reporting themselves the research endpoints. A scoping review maps the evidence and its gaps rather than estimating the causal effect of any particular intervention. The date of this review is especially relevant for capturing contemporary digital-delivery practices.
Child and adolescent safety cannot be reduced to adult symptom thresholds. Outcomes may include school functioning, peer relationships, family dynamics, identity-related distress, and the ability to understand or dispute treatment goals. Reporting systems must consider the possibility that the source of harm lies within a family, educational institution, or treatment relationship itself. Overly broad disclosure to caregivers can also create confidentiality risks. Local legal and professional requirements remain crucial because consent, safeguarding, and confidentiality conditions differ by jurisdiction.
Trials should define adverse outcomes in advance, apply validated age-appropriate scales, obtain the young person's perspective wherever feasible, and document crisis pathways. When family or school informants contribute data, disagreements should be preserved as information rather than mechanically averaged away. Longitudinal studies are especially important because developmental transitions and treatment effects can unfold over longer periods than standard short follow-up windows.
Cultural Context, Marginalization, and Differential Exposure
Risks and negative experiences are not distributed solely by diagnosis. Language differences, discrimination, minority stress, treatment expectations, and unequal access can influence the probability of an adverse interaction, the interpretation of symptoms, and the possibility of safely reporting harm. Cultural context can affect whether a method is understood as supportive, intrusive, stigmatizing, or inconsistent with a person's values. A well-designed safety study therefore specifies the population and social setting instead of assuming that the experiences of one dominant treatment sample represent all patients.
In the cross-sectional national survey by Crawford (2016), sexual and ethnic minority respondents reported lasting negative experiences more frequently than relevant comparison groups. The evidence indicates an association in that sample; it does not identify one mechanism, and it does not warrant treating minority status as an intrinsic susceptibility to harm. Possible pathways include discrimination, reduced treatment fit, communication barriers, and service-level inequality, each requiring targeted research and careful distinction from confounding variables.
Cross-cultural measurement is itself a safety issue. An instrument developed in one language may not preserve the same factor structure, item thresholds, or meaning in another. Definitions of autonomy, family obligation, emotional expression, and psychological distress differ across communities and contexts. Comparative studies require culturally grounded construct definitions, translation and adaptation procedures, tests of measurement invariance when feasible, and qualitative assessment of unanticipated categories. Otherwise, a numerical difference in reported adverse effects may reflect instrument behavior rather than a real difference in risk.
Organizations can study differential safety through stratified reporting and appropriate multilevel analysis while protecting patient privacy. Underrepresentation and small subgroup sizes limit precision, and the pursuit of subgroup comparisons should not encourage unreliable rankings. Equity-oriented safety research concerns both clinical outcomes and the conditions under which people are informed, believed, supported, and able to choose or change care.
Comorbidity, Severity, and Complex Clinical Presentations
Many psychological-treatment studies recruit populations with defined inclusion and exclusion criteria, whereas real clinical services treat people with co-occurring psychiatric and medical conditions, concurrent medications, unstable living circumstances, and differing levels of impairment. This gap complicates generalization of safety findings. A trial may accurately characterize a standardized intervention under close monitoring and still offer limited evidence about patients whose circumstances differ substantially. Extrapolation should consider baseline severity, co-interventions, available crisis support, and the ability to complete the assigned treatment.
Comorbid conditions may influence the clinical meaning of treatment-emergent change. A person receiving psychotherapy for anxiety may experience changes in sleep or substance use, while someone with depression may have a fluctuating physical illness or a contemporaneous medication adjustment. The observed event should be classified across relevant domains, and the analysis should record co-treatment as a potential confounder or effect modifier. A single disorder-specific scale cannot represent the full safety picture for such patients.
Risk stratification also has an ethical dimension. Identifying a clinical feature that predicts worsening should help improve monitoring and treatment fit; it should not automatically exclude a person from effective treatment. Predictive models are particularly vulnerable to poor calibration when applied to populations beyond their training samples. The ability to predict a higher event probability does not establish that withholding treatment improves outcomes. Comparative treatment strategies and careful causal evidence are needed before prediction is translated into a decision rule.
The National Institute for Health and Care Excellence. (2022). Depression in adults: Treatment and management (NICE guideline NG222). https://www.nice.org.uk/guidance/ng222/chapter/recommendations (2022) addresses treatment planning, monitoring, and coordination across severity levels and comorbidities. Its recommendations can guide evaluation of service safeguards in applicable settings. This dossier retains the distinction between studying a clinical guideline and issuing individual treatment recommendations; decisions about severe or urgent illness belong to responsible qualified services.
Group, Family, and Institutional Interventions
Group therapies involve relational dynamics beyond the clinician–patient dyad. The possibility of unwanted disclosure, conflict between group members, perceived rejection, coercion, or uneven facilitation requires a safety analysis tailored to the intervention format. Participants may have limited control over the actions of other members, and confidentiality is partly dependent on group behavior. Adverse-event protocols should document incidents within and outside sessions where relevant, including the patient's understanding of confidentiality and grievance procedures.
Family and couple interventions likewise change interpersonal systems in which there may be unequal power, coercion, or risk of violence. A treatment designed to improve communication can be hazardous if it presupposes equal bargaining power or fails to recognize intimidation. This is a conceptual reason to evaluate screening, consent, session structure, and confidentiality carefully. Actual comparative risk for a specified intervention must be established empirically; a general theoretical possibility does not quantify prevalence or causality.
Institutional settings may combine psychological interventions with mandatory attendance, surveillance, restricted autonomy, or consequences for declining participation. The meaning of consent differs where treatment is tied to detention, employment, education, or other institutional authority. Research in these environments should separate the effect of the psychological technique from harms arising from coercive conditions, administrative decisions, or inadequate safeguarding. Forensic and Legal Psychology provides a complementary framework for professional accountability and consent under unequal power.
The central scientific insight is that the unit of treatment may be an interpersonal or organizational system. Safety assessments restricted to symptom change in the individual participant can miss significant interpersonal, legal, and social consequences. Qualitative research, process observation, and institutional event registers can complement randomized comparisons to describe these distinct dimensions.
The Problem of Withdrawing, Switching, or Ending Care
Safety includes the termination of a treatment episode. An abrupt end may follow a breach of trust, access barriers, symptom remission, worsening, administrative decisions, or patient preference. An outcome recorded at the last session can conceal subsequent relapse or delayed benefit. Studies that compare therapies only at termination may favor shorter or longer formats depending on how termination is defined and who remains under observation.
The Swift (2012) meta-analysis supports the need to treat early ending as a measurable but heterogeneous phenomenon. The Klein (2024) similarly calls for more nuanced descriptions of dropout in adverse-event research. A patient who changes therapist after an unproductive course is not equivalent to a patient who stops because of a serious adverse interaction, nor to someone who improves rapidly and chooses to end. Studies should collect these reasons prospectively where possible.
A strong longitudinal design follows patients across the end of treatment and, when feasible, across referral or switching to another form of care. This allows estimation of the consequences of a pathway, not only the outcome of a single program. It also reduces the temptation to interpret the disappearance of a patient from the dataset as resolution of a problem.
In service evaluation, transparency about discontinuation protects both scientific validity and patient agency. Research questions include whether a clinical transition was planned, whether information was transferred with consent, whether relevant support remained accessible, and whether deterioration after treatment was recognized. Such process indicators require explicit measurement and a defined jurisdictional context.
Routine Outcome Monitoring and Early-Warning Systems
Routine outcome monitoring (ROM) aims to detect departures from expected progress while treatment is still under way. Instead of waiting for a final assessment, services collect repeated symptom or functioning measures and may compare observed trajectories with reference patterns. An early-warning flag can indicate that further assessment is needed, but it does not establish the cause of worsening, and it should not replace a clinician's immediate response to serious risk. ROM is a measurement and decision-support infrastructure whose value depends on the decisions it changes.
In a Lambert (2001), 609 patients were allocated across feedback and control groups. Feedback to clinicians was associated with better outcomes among patients predicted to be treatment failures, including fewer classifications of deterioration in the reported study. The Lambert (2003) brought early patient-focused feedback studies together. These results suggest a potentially useful mechanism: objective indications of poor progress can challenge overly optimistic clinical judgments and prompt reconsideration of care. Their estimates remain contingent on the monitoring system, sample, outcome definition, and implementation.
Hannan (2005) examined clinician judgment and algorithm-based detection of patients at risk of failure. Their report described greater sensitivity of formal monitoring than unaided prediction in the studied setting. The result should not be interpreted as guaranteeing the performance of any algorithm in other services. Detection systems are vulnerable to false alarms, missed events, unstable norms, and differential data quality. A warning algorithm trained on one cohort may be poorly calibrated for another diagnosis, population, language, or service model.
A later Shimokawa (2010) evaluated outcome-feedback systems for patients at risk of treatment failure. The research focus was whether supplying progress information improved outcomes, rather than whether the technology itself caused all improvements. Subsequent controlled work by Delgadillo (2017) and a Delgadillo (2018) extended feedback-informed research to routine depression and anxiety services. Together these studies justify testing monitoring as a safety strategy, with local evaluation of feasibility, accuracy, response, cost, and downstream outcomes.
Feedback Is a Clinical Process, Not Merely a Score
Collecting a questionnaire becomes a safety intervention only when the information reaches a person or system capable of acting on it. The relevant pathway involves measurement completion, score computation, identification of a meaningful change, communication to an accountable professional, patient discussion, and an appropriate subsequent decision. Missing any link in this chain can turn an apparently comprehensive dashboard into passive documentation.
Different feedback systems establish different decision rules: absolute thresholds, deviations from expected change, reliable deterioration indicators, or combinations of risk and response measures. The threshold determines the tradeoff between missed problems and excessive alerts. The reference distribution determines what counts as expected progress. When the baseline distribution differs by setting or population, calibration can fail. The best threshold for predicting nonresponse may be inappropriate for detecting acute self-harm risk, which requires a distinct clinical protocol and appropriate urgency.
Patients' own reports should remain an essential part of interpreting quantitative feedback. A person who crosses a risk threshold may have experienced an unrelated crisis, difficulties with the measurement items, a temporary increase associated with therapeutic work, or significant worsening attributable to treatment. The signal provides a reason for investigation, not an automatic finding. The same principle applies to patients whose scores appear stable but who report feeling coerced, invalidated, or violated: broad safety oversight cannot be reduced to statistical trajectories.
Future implementation trials should randomize the addition or format of feedback, measure whether treatment plans actually changed, and follow patient-important outcomes. They should compare measured deterioration, treatment uptake, changes in functioning, adverse events, and cost, including unintended effects of monitoring itself. The National Institute for Health and Care Excellence. (2022). Depression in adults: Treatment and management (NICE guideline NG222). https://www.nice.org.uk/guidance/ng222/chapter/recommendations (2022) incorporate routine outcome monitoring as a consideration alongside monitoring for harm and suicidality, not as an independent replacement for clinical responsibility.
Comparative Risk–Benefit Evaluation
A benefit–risk evaluation requires the same population, intervention, comparator, and time horizon for both favorable and unfavorable outcomes. Comparing the efficacy from one highly controlled trial with adverse-event reports from a different clinical setting is vulnerable to confounding by severity, surveillance, and treatment exposure. Safety differences between approaches should be evaluated with a prespecified set of outcomes and consistent ascertainment methods. When a treatment improves one domain but worsens another, the report should preserve that multidimensional profile instead of collapsing it into an unexplained single score.
Some outcomes can be expressed as proportions or risk differences, while others require scales, time-to-event models, or qualitative interpretation. Serious but infrequent events may need larger samples and pooled data than ordinary efficacy endpoints. Small trials can fail to observe rare events even when the underlying risk is meaningful. Effect estimates should be accompanied by uncertainty intervals and assessed for selective ascertainment. The unit of reporting should remain clear: persons affected, incidents, episodes, or exposure-time rates.
The clinical meaning of risk also depends on alternative options. If a beneficial intervention has a moderate rate of transient manageable discomfort and an alternative offers less benefit, their comparative value cannot be determined solely by counting experiences labeled negative. Conversely, if two treatments have similar expected benefit, differences in serious or enduring harm, burden, access, and acceptability can become decisive. Klein (2024) explicitly discuss this future possibility for comparing psychological treatments.
Safety estimates should remain transparent about patient preferences and heterogeneity. Some people prioritize rapid symptom reduction, others place greater weight on autonomy, functional recovery, continuity of relationships, or avoiding a particular kind of adverse experience. A rigorous publication can describe distributions and uncertainty while identifying the values that affect the interpretation of tradeoffs. Scientific reporting supplies evidence for shared decisions; it does not erase genuine variation in the outcomes patients consider important.
Reporting Standards and the Architecture of a Safety Trial
A credible randomized trial protocol should specify what counts as an adverse event, who will collect it, how often collection occurs, who reviews possible seriousness, and how relatedness to intervention will be assessed. It should define the treatment exposure period, the follow-up period, the expected categories of events, and a pathway for unanticipated events. All randomized arms need sufficiently comparable ascertainment. Trial reports should include denominators, participant flows, event severity, timing, and the extent of unresolved causal attribution.
Junqueira (2023), explained in 2023, updates harms reporting for randomized trials. Its broader reporting logic applies to psychological interventions even when intervention-specific adverse-event instruments need further development. Duggan (2014) and the Klatte (2025) demonstrate why advance definitions and transparent reporting matter: a trial emphasizing psychological benefit can otherwise leave the reader unable to assess harms on comparable terms.
Researchers should prespecify how they will handle adverse events not considered causally related to the experimental treatment, since exclusion before presentation can bias the evidence. Adjudication categories such as possible, probable, and unlikely relatedness require justified operational criteria and preferably independent reviewers. Distinct concepts of severity, seriousness, preventability, and relatedness should have separate fields. Longitudinal event logs make it possible to examine the timing and persistence of reactions without pretending that every self-reported event is confirmed treatment-induced damage.
Protocols should also describe the response to emerging risk, treatment discontinuation rules, and ethical oversight. Monitoring a trial is an ongoing duty, and the integrity of a randomized comparison does not justify withholding necessary urgent care. After trial completion, methods, outcomes, supplementary results, and the full event taxonomy should be available to allow independent interpretation, subject to legitimate privacy protections.
Evidence Synthesis Under Heterogeneous Harm Definitions
A conventional systematic review estimates the evidence using explicit eligibility criteria, comprehensive search methods, selection, extraction, risk-of-bias assessment, and a specified synthesis plan. Where events differ fundamentally by measurement and timing, statistical pooling may have limited interpretability even if numerous studies are retrieved. A reported pooled proportion can become a mixture of questions and procedures rather than an estimate of one coherent phenomenon. The Klatte (2025) review specifically identifies heterogeneity in conceptualization and reporting as a barrier to comparability.
A future evidence synthesis should stratify by event definition, source of report, active versus passive surveillance, treatment format, comparator, severity, observation period, and whether causal attribution was made. It should distinguish symptom deterioration from subjective negative experiences and serious adverse events. Risk of bias may arise through missing outcome data, selective publication, differential ascertainment, inadequate randomization, or the institutional incentives of those collecting events. Meta-regression cannot reliably repair fundamentally incompatible constructs or observational bias.
When comparative trials cannot be pooled coherently, a structured narrative synthesis should state the coverage of available literature, identify conflicting results, describe their designs and populations, and explain why quantitative combination is unwarranted. Qualitative evidence synthesis can provide an integrated category system, as Vybíral (2024) demonstrated, while acknowledging that these categories do not determine incidence. Scoping reviews, such as Munir (2026), can map what has and has not been measured. Each method contributes a distinct kind of evidence.
A single universal percentage of psychological-treatment harm is scientifically unhelpful when studies ask different questions. More interpretable outputs are outcome-specific and transparent: for example, the proportion with reliably worsened symptoms over a defined interval, the proportion who retrospectively reported lasting bad effects, or the number with a serious reported event during a trial. Each still requires a description of causal uncertainty. This approach preserves the complexity necessary for clinically useful scientific comparison.
Clinical Governance, Transparency, and Patient Agency
The organization responsible for treatment safety extends beyond the individual therapist. Competency requirements, supervision, incident-reporting systems, accessible complaints procedures, clinical referral paths, privacy safeguards, and independent oversight contribute to the reliability of care. A program may have technically effective interventions but fail to identify deterioration or respond to patient complaints. Its safety should therefore be evaluated as a property of the treatment process and the service setting, not merely the therapy manual.
Transparency includes the ability to communicate uncertainty to patients without dismissing either potential benefits or adverse experiences. Patients should be able to report problems through channels that do not require confrontation with the treating clinician and to participate in reassessing the goals and method of care. Crawford (2016) connects patient-reported lasting effects with the informational conditions of treatment, while the qualitative synthesis of Vybíral (2024) documents difficulties arising from the therapeutic relationship and poor treatment fit.
Institutional auditing requires more than a single numerical deterioration target. Services should examine whether event reporting is accessible, whether high-risk groups are underrepresented in routine measures, whether transfers or early terminations conceal adverse events, and whether incident adjudication is independent of financial or reputational pressures. Safety oversight can learn from individual cases without using them to make unwarranted population-level claims. Protecting confidentiality is essential, particularly where serious clinical incidents are documented.
Guidance such as National Institute for Health and Care Excellence. (2022). Depression in adults: Treatment and management (NICE guideline NG222). https://www.nice.org.uk/guidance/ng222/chapter/recommendations (2022) establishes expectations for monitoring treatment response, harm, suicidal ideation, clinician competence, and regular supervision within its jurisdiction. General scientific principles must be adapted to the legal requirements and professional standards of each service. The appropriate level of action depends on the seriousness of the event; this monograph addresses comparative research and system governance rather than offering personal instructions or replacing qualified clinical review.
Artificial Era: Decision Support and New Safety Questions
The Artificial Era introduces new arrangements for psychological treatment in which digital systems participate in screening, triage, outcome monitoring, psychoeducation, therapeutic exercises, and conversational support. Their effects on safety must be studied at the level of actual human experience and the full care pathway. A machine-generated response may be delivered directly to a person, reviewed by a licensed professional, or incorporated into an organizational decision. These architectures differ in who interprets warnings, who exercises clinical judgment, what records persist, and who is accountable when the interaction fails. Treating all such arrangements as one intervention category obscures their divergent mechanisms and risks.
Algorithmic monitoring can make longitudinal measurements available, identify patterns in symptom reports, and help a responsible clinician notice an unexpected trajectory. Its promise resembles the earlier evidence on structured feedback reviewed by Shimokawa (2010), while its implementation raises additional questions about model calibration, data quality, automation bias, and human oversight. An alert model can be evaluated for discrimination and calibration against a clearly defined outcome, but a statistically accurate alert has clinical value only when the surrounding system reliably interprets and responds to it. Prediction of deterioration and prevention of deterioration remain distinct research problems.
Conversational systems can generate language that resembles empathy, authority, or professional understanding. A user may regard such responses as personally meaningful, including when the underlying system has no verified capacity to provide medical diagnosis or clinical accountability. Safety inquiry should examine the reliability of the advice, the effects of apparent personal continuity, user expectations, disclosure of system limitations, and the consequences of prolonged interaction. These are questions about observable interaction and outcomes, answerable through controlled evaluation, longitudinal observation, qualitative interviewing, and safety audits. The philosophical questions of Artificial agency and Artificial identity belong to an adjacent research horizon and should be distinguished from empirically demonstrated clinical effects.
The use of automated systems creates new dependencies between measurement and institutional action. If an application detects an elevated risk report but offers no clinically accountable pathway, data collection may create a misleading assurance of protection. If an algorithm prompts a change in treatment without intelligible human review, a false-positive alert may cause unnecessary disruption and a false-negative result may delay recognition of deterioration. Research should measure downstream decisions and their outcomes, not merely model-level performance. Privacy protection, informed understanding of data use, accessibility, independent auditing, and documented escalation responsibilities form part of the evaluation.
Artificial systems also make comparative design particularly important. Studies should compare a specified AI-supported intervention with appropriate human-led or digital alternatives, maintain consistent adverse-event ascertainment across arms, and report clinically meaningful differences with uncertainty. Where a system is frequently updated, investigators must record the actual model and version delivered to participants, relevant prompts or safety policies, and the dates of changes. Without stable intervention identification and reproducible exposure records, a safety result cannot be reliably transported to later systems.
Competing Interpretations and Limits of the Evidence
The psychotherapy harm literature contains several recurring interpretive conflicts. One concerns whether any unwanted experience should be classified as an adverse effect or whether causal attribution should be required. The broad definition is sensitive to experiences that clinicians might miss; the narrower definition supports causal comparisons but may undercount poorly understood treatment harms. A defensible approach records occurrence first, then severity, persistence, patient perception, and attribution separately. This preserves patient testimony while allowing different research questions to be answered with appropriate denominators. Linden (2013), Klein (2024), and Rosendahl and colleagues (2025) illuminate this definitional problem.
A second disagreement concerns the boundary between temporary treatment-related distress and sustained deterioration. A symptom increase may be part of exposure to a difficult subject, an indication of emerging crisis, or ordinary disorder fluctuation. A retrospective judgment made after recovery may differ from a concurrent report made during treatment. Neither timing automatically settles clinical significance. Comparative research should therefore prespecify temporal windows and distinguish transient negative experiences from persistent impairment, serious adverse events, and clinically significant symptom change.
The third conflict concerns the transfer of adverse-event rates across treatments and settings. A survey of perceived lasting effects, an active-monitoring psychotherapy trial, and a registry of serious clinical incidents estimate different quantities. Their percentages cannot be combined as if they described one underlying risk distribution. The quantitative findings in Crawford (2016), the Klatte (2025) trial review, and the qualitative taxonomy of Vybíral (2024) are complementary because each captures a different aspect of safety.
A fourth issue is whether a successful early-warning system is necessarily protective. Evidence that formal feedback improves outcomes for some at-risk patients is important, but the causal pathway requires attention to clinician response, organizational incentives, patient consent, and possible adverse consequences of monitoring. Feedback can fail through omission, excessive alerts, or inappropriate clinical action. The Lambert (2001) and Delgadillo (2018) studies show why monitoring belongs in comparative evaluation; their findings should not be turned into claims of universal algorithmic protection.
Research limitations remain considerable. Harm reporting is heterogeneous, many trials were designed principally for efficacy, patient groups and treatment formats vary, severe events are often rare, incomplete follow-up is consequential, and observational attribution is vulnerable to confounding. Prospective adverse-event collection may reveal more events than passive systems, generating apparent differences caused by surveillance intensity. These limitations call for improved research methods rather than a single sweeping conclusion about the safety of psychotherapy as a whole.
A Research Framework for Evaluating Treatment Safety
A useful general model organizes safety evidence through seven linked stages: intervention definition, baseline vulnerability, event observation, measurement and classification, causal evaluation, clinical response, and longitudinal outcome. This is an analytical framework for comparing the existing literature and specifying future studies, not an assertion of an independently validated treatment theory. Every stage corresponds to a potential source of error or an actionable research question.
Intervention definition requires a reproducible description of the treatment, its dose and duration, relevant therapist qualifications, setting, co-interventions, and degree of clinical supervision. Baseline vulnerability includes symptoms, functioning, history, external stressors, relevant medical conditions, and the patient's own treatment priorities. Event observation needs repeated assessment and accessible routes for patient reporting. Measurement must distinguish reliable symptom change from experiential consequences and serious clinical incidents. Causal evaluation compares the observed trajectory with credible alternatives and acknowledges uncertainty at the individual level.
Clinical response concerns what actually happens after an event is recognized: whether it is acknowledged, independently assessed when appropriate, and incorporated into a documented change in care. Longitudinal outcome extends observation beyond a single session or termination point. The whole model can be investigated through randomized comparisons, prospective cohorts, patient interviews, process analysis, and transparent adverse-event registries. Each design adds information that the others cannot supply by themselves.
This framework implies a minimal reporting architecture. Every event record should identify its source, date, treatment exposure, domain, severity, seriousness, duration, patient assessment, clinician assessment where appropriate, relatedness judgment, concurrent circumstances, action taken, and outcome. Aggregate reports should give the relevant denominators, missing-data patterns, time horizons, and uncertainty estimates. Such records enable later systematic reviews to determine which studies actually measured comparable events. They also permit institutional evaluation of whether an event was recognized and addressed.
The normative rationale is patient safety and informed agency; the empirical question is whether implementing such a framework produces better outcomes than the applicable alternative. Its introduction may increase recorded events simply because reporting improves. A higher detected event count can therefore accompany better observation rather than greater underlying harm. Controlled implementation and independent auditing are needed to interpret changes over time.
Priorities for the Next Generation of Safety Research
The first priority is prospective, standardized, multidomain measurement. Trials should record patient-perceived negative experiences and symptom deterioration alongside serious events, quality of life, functioning, and reasons for ending treatment. Definitions and observation windows should be specified before analysis. Where validated instruments exist, investigators should report their properties and cultural applicability rather than assuming equivalence across populations. The work on the Rozental (2016) and the trial-reporting analysis of Klatte (2025) provide foundations for further harmonization.
The second priority is comparative causal research. Studies should be designed to test differences in specific harms between treatments, rather than treating adverse events as unstructured notes alongside efficacy outcomes. Adequate sample sizes, balanced surveillance, serious-event review, predefined estimands, and longer follow-up are particularly important. When randomization is infeasible, observational studies should state causal assumptions, address selection and confounding, and preserve appropriate comparators. Investigators should avoid labeling a treatment-related event as a proven treatment-caused event without supporting evidence.
A third priority is mechanistic work linking treatment process and adverse outcomes. Repeated measurement of therapeutic alliance, intervention fidelity, patient understanding, contextual stressors, and symptom trajectories can identify plausible pathways. Randomized modifications of specific safety practices may then test whether those pathways are modifiable. Therapist effects, institutional practices, cultural adaptation, and patient participation deserve attention as potential determinants of both benefits and harms. The Vybíral (2024) is especially valuable for generating hypotheses that conventional symptom measures overlook.
A fourth priority is early-warning evaluation in real services. Research should establish not only whether a monitoring model predicts negative outcomes, but whether feedback changes care in a way that improves safety. Investigators should track the entire response pathway, including false alerts and ignored warnings. The historical feedback research of Lambert (2001) supplies a starting point for rigorous comparative work on human and automated decisions.
Finally, emerging digital and AI-supported interventions require intervention-specific reporting, privacy and accountability studies, and long-term observational evidence. The standards of scientific inference remain the same: define the exposure, establish the comparator, measure outcomes, evaluate alternative explanations, and document the uncertainty. The Artificial Era increases the range of systems participating in psychological care and therefore expands the forms of responsibility that safety research must examine.
Conclusion
Psychological treatment safety is a fundamental dimension of evidence-based care. Its object includes clinically meaningful symptom deterioration, patient-experienced adverse effects, serious unwanted events, treatment misfit, harmful professional conduct, and failures of monitoring or response. These outcomes should be studied as distinct constructs with specified definitions, denominators, observation periods, and evidentiary levels. No single questionnaire, case narrative, deterioration index, or trial average captures the entire safety profile.
The scientific literature supports several firm conclusions. Patient-level trajectories contain clinically important variation that group-average efficacy estimates can conceal. Patients report adverse experiences extending beyond measured symptoms. Some widely adopted interventions have required revision after controlled evidence challenged their rationale. Routine outcome feedback can improve results in specified clinical settings, but the effectiveness of detection depends on responsive professional and institutional systems. The existing trial evidence also contains substantial gaps in adverse-event ascertainment, reporting, causal attribution, and long-term follow-up.
The central methodological conclusion is that temporal association, reliable worsening, perceived harm, and treatment-caused damage are related but separate objects of investigation. Their careful distinction makes research more useful to patients, clinicians, reviewers, and institutions. A mature safety science integrates randomized comparative evidence, longitudinal measurement, independent adverse-event assessment, patient testimony, clinical governance, and transparent reporting of uncertainty.
As psychological care extends into digital and Artificial systems, these principles retain their force. The organization delivering a treatment, the system observing outcomes, and the authority deciding how to respond become increasingly interdependent. Scientific accountability therefore requires evidence about the actual care pathway and its consequences. Psychological treatment becomes safer through the disciplined detection, interpretation, prevention, and responsible management of unfavorable outcomes.
References
- Barlow, D. H. (2010). Negative effects from psychological treatments: A perspective. American Psychologist, 65(1), 13–20. https://doi.org/10.1037/a0015643
- Crawford, M. J., Thana, L., Farquharson, L., Palmer, L., Hancock, E., Bassett, P., Clarke, J., & Parry, G. D. (2016). Patient experience of negative effects of psychological treatment: Results of a national survey. The British Journal of Psychiatry, 208(3), 260–265. https://doi.org/10.1192/bjp.bp.114.162628
- Delgadillo, J., Overend, K., Lucock, M., Groom, M., Kirby, N., McMillan, D., Gilbody, S., Lutz, W., Rubel, J. A., & de Jong, K. (2017). Improving the efficiency of psychological treatment using outcome feedback technology. Behaviour Research and Therapy, 99, 89–97. https://doi.org/10.1016/j.brat.2017.09.011
- Delgadillo, J., de Jong, K., Lucock, M., Lutz, W., Rubel, J., Gilbody, S., Ali, S., Aguirre, E., Appleton, M., Nevin, J., O’Hayon, H., Patel, U., Sainty, A., Spencer, P., & McMillan, D. (2018). Feedback-informed treatment versus usual psychological treatment for depression and anxiety: A multisite, open-label, cluster randomised controlled trial. The Lancet Psychiatry, 5(7), 564–572. https://doi.org/10.1016/S2215-0366(18)30162-7
- Duggan, C., Parry, G., McMurran, M., Davidson, K., & Dennis, J. (2014). The recording of adverse events from psychological treatments in clinical trials: Evidence from a review of NIHR-funded trials. Trials, 15, Article 335. https://doi.org/10.1186/1745-6215-15-335
- Eubanks, C. F., Muran, J. C., & Safran, J. D. (2018). Alliance rupture repair: A meta-analysis. Psychotherapy, 55(4), 508–519. https://doi.org/10.1037/pst0000185
- Hannan, C., Lambert, M. J., Harmon, C., Nielsen, S. L., Smart, D. W., Shimokawa, K., & Sutton, S. W. (2005). A lab test and algorithms for identifying clients at risk for treatment failure. Journal of Clinical Psychology, 61(2), 155–163. https://doi.org/10.1002/jclp.20108
- Honkalampi, K., Urhonen, H.-R., & Virtanen, M. (2025). Negative effects in randomized controlled trials of psychotherapies and psychological interventions: A systematic review. Psychotherapy Research, 35(1), 100–111. https://doi.org/10.1080/10503307.2024.2301972
- Jacobson, N. S., & Truax, P. (1991). Clinical significance: A statistical approach to defining meaningful change in psychotherapy research. Journal of Consulting and Clinical Psychology, 59(1), 12–19. https://doi.org/10.1037/0022-006X.59.1.12
- Jonsson, U., Alaie, I., Parling, T., & Arnberg, F. K. (2014). Reporting of harms in randomized controlled trials of psychological interventions for mental and behavioral disorders: A review of current practice. Contemporary Clinical Trials, 38(1), 1–8. https://doi.org/10.1016/j.cct.2014.02.005
- Junqueira, D. R., Zorzela, L., Golder, S., Loke, Y., Gagnier, J. J., Julious, S. A., Li, T., Mayo-Wilson, E., Pham, B., Phillips, R., Santaguida, P., Scherer, R. W., Gøtzsche, P. C., Moher, D., Ioannidis, J. P. A., & Vohra, S. (2023). CONSORT Harms 2022 statement, explanation, and elaboration: Updated guideline for the reporting of harms in randomized trials. Journal of Clinical Epidemiology, 158, 149–165. https://doi.org/10.1016/j.jclinepi.2023.04.005
- Klatte, R., Strauss, B., Flückiger, C., & Rosendahl, J. (2025). Adverse events in psychotherapy randomized controlled trials: A systematic review. Psychotherapy Research, 35(1), 84–99. https://doi.org/10.1080/10503307.2023.2286992
- Klein, J. P., Rozental, A., Sürig, S., & Moritz, S. (2024). Adverse events of psychological interventions: Definitions, assessment, current state of the research and implications for research and clinical practice. Psychotherapy and Psychosomatics, 93(5), 308–315. https://doi.org/10.1159/000540212
- Lambert, M. J., Whipple, J. L., Smart, D. W., Vermeersch, D. A., Nielsen, S. L., & Hawkins, E. J. (2001). The effects of providing therapists with feedback on patient progress during psychotherapy: Are outcomes enhanced? Psychotherapy Research, 11(1), 49–68. https://doi.org/10.1080/713663852
- Lambert, M. J., Whipple, J. L., Hawkins, E. J., Vermeersch, D. A., Nielsen, S. L., & Smart, D. W. (2003). Is it time for clinicians to routinely track patient outcome? A meta-analysis. Clinical Psychology: Science and Practice, 10(3), 288–301. https://doi.org/10.1093/clipsy.bpg025
- Lilienfeld, S. O. (2007). Psychological treatments that cause harm. Perspectives on Psychological Science, 2(1), 53–70. https://doi.org/10.1111/j.1745-6916.2007.00029.x
- Linden, M. (2013). How to define, find and classify side effects in psychotherapy: From unwanted events to adverse treatment reactions. Clinical Psychology & Psychotherapy, 20(4), 286–296. https://doi.org/10.1002/cpp.1765
- Linden, M., & Schermuly-Haupt, M.-L. (2014). Definition, assessment and rate of psychotherapy side effects. World Psychiatry, 13(3), 306–309. https://doi.org/10.1002/wps.20153
- Munir, S., Dallison, S., Padley, H., Fung, H. M., Perry, G., Hall, C. L., & Loades, M. E. (2026). How have contemporary studies evaluating self-help digital mental health interventions for young people monitored and reported on adverse events and harms? A systematic scoping review. Frontiers in Psychology, 17, Article 1679968. https://doi.org/10.3389/fpsyg.2026.1679968
- National Institute for Health and Care Excellence. (2018). Post-traumatic stress disorder (NICE guideline NG116). https://www.nice.org.uk/guidance/ng116/chapter/recommendations
- National Institute for Health and Care Excellence. (2022). Depression in adults: Treatment and management (NICE guideline NG222). https://www.nice.org.uk/guidance/ng222/chapter/recommendations
- Rose, S., Bisson, J., Churchill, R., & Wessely, S. (2002). Psychological debriefing for preventing post traumatic stress disorder (PTSD). Cochrane Database of Systematic Reviews, 2002(2), Article CD000560. https://doi.org/10.1002/14651858.CD000560
- Rosendahl, J., Klatte, R., Frenzl, D., Castonguay, L. G., Vybíral, Z., & Strauss, B. (2025). Contextualizing the negative effects of psychotherapy. Nature Reviews Psychology, 4(9), 559–575. https://doi.org/10.1038/s44159-025-00472-8
- Rozental, A., Andersson, G., Boettcher, J., Ebert, D. D., Cuijpers, P., Knaevelsrud, C., Ljótsson, B., Kaldo, V., Titov, N., & Carlbring, P. (2014). Consensus statement on defining and measuring negative effects of Internet interventions. Internet Interventions, 1(1), 12–19. https://doi.org/10.1016/j.invent.2014.02.001
- Rozental, A., Boettcher, J., Andersson, G., Schmidt, B., & Carlbring, P. (2015). Negative effects of Internet interventions: A qualitative content analysis of patients' experiences with treatments delivered online. Cognitive Behaviour Therapy, 44(3), 223–236. https://doi.org/10.1080/16506073.2015.1008033
- Rozental, A., Kottorp, A., Boettcher, J., Andersson, G., & Carlbring, P. (2016). Negative effects of psychological treatments: An exploratory factor analysis of the Negative Effects Questionnaire for monitoring and reporting adverse and unwanted events. PLOS ONE, 11(6), Article e0157503. https://doi.org/10.1371/journal.pone.0157503
- Rozental, A., Magnusson, K., Boettcher, J., Andersson, G., & Carlbring, P. (2017). For better or worse: An individual patient data meta-analysis of deterioration among participants receiving Internet-based cognitive behavior therapy. Journal of Consulting and Clinical Psychology, 85(2), 160–177. https://doi.org/10.1037/ccp0000158
- Rozental, A., Kottorp, A., Forsström, D., Månsson, K., Boettcher, J., Andersson, G., Furmark, T., & Carlbring, P. (2019). The Negative Effects Questionnaire: Psychometric properties of an instrument for assessing negative effects in psychological treatments. Behavioural and Cognitive Psychotherapy, 47(5), 559–572. https://doi.org/10.1017/S1352465819000018
- Shimokawa, K., Lambert, M. J., & Smart, D. W. (2010). Enhancing treatment outcome of patients at risk of treatment failure: Meta-analytic and mega-analytic review of a psychotherapy quality assurance system. Journal of Consulting and Clinical Psychology, 78(3), 298–311. https://doi.org/10.1037/a0019247
- Swift, J. K., & Greenberg, R. P. (2012). Premature discontinuation in adult psychotherapy: A meta-analysis. Journal of Consulting and Clinical Psychology, 80(4), 547–559. https://doi.org/10.1037/a0028226
- Vybíral, Z., Ogles, B. M., Řiháček, T., Urbancová, B., & Gocieková, V. (2024). Negative experiences in psychotherapy from clients' perspective: A qualitative meta-analysis. Psychotherapy Research, 34(3), 279–292. https://doi.org/10.1080/10503307.2023.2226813
- Vybíral, Z., Řiháček, T., Borovička, P., & Ogles, B. M. (2026). Clients' negative experiences in psychological treatment. Current Opinion in Psychology, 67, Article 102175. https://doi.org/10.1016/j.copsyc.2025.102175