The practice is not the treatment

Imagine a paper reporting that “meditation increased activity in attention-related regions”. Before interpreting the result, replace the word meditation with a precise action. Was the participant sustaining attention on the breath, monitoring whatever arose, repeating a phrase, cultivating compassion, scanning bodily sensations, visualising an image, or examining the apparent boundary of self? These operations can share posture and silence while placing different demands on attention, affect and representation.

A meta-analysis of 78 functional neuroimaging investigations found dissociable activation and deactivation patterns across focused attention, mantra recitation, open monitoring and compassion practices. Some regions appeared across techniques, but convergence was the exception rather than the rule.[4] Lutz and colleagues therefore proposed a multidimensional phenomenological matrix rather than a single definition of mindfulness, tracking properties such as object orientation, meta-awareness, stability, clarity and effort.[5] Bottom-up work with experienced practitioners likewise catalogues many basic techniques rather than one canonical act.[6]

The configured-exposure model

The smallest useful unit of analysis is not “meditation”. It is a configuration with six fields:

These fields do more than add independent effects. They interact. Forty minutes of open monitoring after a rested week may differ from forty minutes during severe insomnia. The same bodily sensation can be stabilising when titrated with skilled support and destabilising when intensified in isolation. An expert may need less overt control effort than a novice to produce a more stable state. A participant who expects calm may rate agitation as failure, while another interprets it as useful exposure.

The causal object is a trajectory, not a label

Picture the first five minutes of one session. A participant begins with narrow attention on the breath, notices chest tightness, shifts into monitoring the whole body, becomes anxious, opens their eyes, receives reassurance from an instructor and resumes with a softer anchor. The protocol may still record one twenty-minute dose of focused-attention meditation. The enacted exposure was a sequence of changing operations, bodily states and adaptations.

This matters because the six fields are time-varying. Let C_t denote the configuration at moment t: the operation being attempted, the skill available, the expectation active at that moment, recent dose, current person-state and available support. The observed outcome, Y_t, is produced by the path through those configurations, not by the intervention name alone. This notation is not a claim that the six fields can already be reduced to one numerical formula. It simply prevents the programme label from replacing the causal history.

The distinction creates a cleaner replication target. Two studies can use the same branded programme but instantiate different exposures because instructors adapt differently, home practice is concentrated differently, or participants enter with different expectations and vulnerabilities. Conversely, two traditions may use different vocabulary while asking participants to enact a similar operation. The unit worth replicating is the configured operation and its trajectory, not the prestige or familiarity of its label.

A minimal study can record this trajectory without turning every session into surveillance. At session start, capture sleep, distress and expected benefit. During the task, use a few unpredictable probes and one mild perturbation. At session end, record enacted operation, unusual experience, valence and functional carry-over. The resulting record is small enough to analyse, yet rich enough to distinguish “the instruction was delivered” from “the intended exposure occurred”.

flowchart LR
  A["The same instruction enters<br/>two different nervous systems"] --> B["The practice is not the treatment"]
  B --> C["Expertise changes what the<br/>scanner receives"]
  C --> D["The expectancy mirror"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#7B4CD8,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

Expertise changes what the scanner receives

Researchers often operationalise expertise as lifetime hours. That is convenient but incomplete. Hours are retrospective, traditions count them differently, and equal totals may contain different techniques, retreat density and recent practice. More importantly, hours describe exposure history, not whether the requested state was produced during the measured trial.

The distinction is visible in early work on focused-attention experts. Brefczynski-Lewis and colleagues reported an inverted U-shaped association: practitioners averaging roughly 19,000 hours showed greater recruitment of attention-related regions than novices, while practitioners averaging roughly 44,000 hours showed less.[7] A plausible interpretation is that intermediate expertise recruits more control, while very high expertise achieves stability with less effort. Yet the study was cross-sectional. Selection, lifestyle and tradition could also explain group differences.

Recent studies shift the question from average activation towards connectivity and state occupancy. Kral and colleagues compared forty long-term practitioners with 124 meditation-naive controls and found group differences in resting connectivity and subjective attention, although one behavioural orienting measure did not differ by group.[9] Panitz and colleagues reported that experienced practitioners spent more resting time in certain sensory and attention-related network states.[8] These results are suggestive, not proof that practice caused the difference. Their deeper value is methodological: expertise may alter how often a state occurs, how quickly it is entered and how much effort sustains it.

Expertise has at least four coordinates

A useful expertise record separates four questions that lifetime hours merge. The first is exposure history: what was practised, for how long, how recently and with what retreat density. The second is enactment skill: how quickly and reliably the practitioner can produce the requested operation under measurement. The third is flexible control: whether the practitioner can change objects, broaden or narrow monitoring and recover from interruption when the task demands it. The fourth is phenomenological resolution: whether the practitioner can distinguish distraction, dullness, effort, affective tone and bodily strategy without merely repeating a tradition’s preferred vocabulary.

These coordinates can diverge. Someone may have accumulated many hours in mantra practice yet be a novice at open monitoring. Another person may stabilise attention quickly but describe the episode coarsely. A third may provide highly articulate reports while performing no better on a perturbation task. Calling all three “expert” turns a useful difference into noise.

The smallest improvement is to replace a binary novice-expert variable with an expertise manifest. It should record technique-specific hours, recent frequency, retreat exposure, teacher or lineage where relevant, prior scientific-task experience and in-session performance. Report history and performance separately. Do not average them into one prestige score. The history identifies candidate learning; the performance tests whether that learning entered the experiment.

This also changes causal interpretation. A cross-sectional neural difference may reflect training, prior disposition, lifestyle, social selection or familiarity with the scanner task. A longitudinal change narrows those rivals but still mixes maturation, repeated testing and concurrent life changes. Within-person perturbations are often more revealing: vary interruption, object, duration or required effort and ask whether the supposed expertise predicts the changed response. Expertise becomes mechanistic only when it changes a specified control problem.

From instruction fidelity to state fidelity

A protocol can be delivered exactly while the intended state varies widely. Call the first property instruction fidelity: the audio, timing and wording were correct. Call the second state fidelity: the participant actually enacted the specified operation with measurable stability. A study that verifies only instruction fidelity can compare two different mental processes under one label.

State fidelity does not require pretending that private experience is directly observable. It requires converging indicators. Intermittent probes can ask where attention was and what operation was occurring. Respiration and motion can reveal bodily strategy. A surprise distractor can test recovery. A post-trial micro-description can distinguish broad monitoring from narrow fixation. Repeated sessions can estimate within-person reliability. None of these measures proves the experience, but together they narrow the space of rival explanations.

No measure gets to be the referee

Self-report is indispensable because the operation is partly defined by what the participant attempted and noticed. It is also shaped by vocabulary, memory, demand and training in what counts as a reportable event. Behaviour is observable, but the same score can arise from different strategies. Respiration, pupil size and heart-rate variation track bodily regulation, yet none is specific to meditation. Brain activity offers spatial and temporal structure, but a difference does not identify the mental operation by itself.

The answer is a measurement braid. Before collecting data, predict what should happen across at least three strands. If focused attention becomes more stable, mind-wandering probes should decline, recovery after a distractor should accelerate and trial-to-trial neural dynamics should become less variable in a specified network. If only the retrospective rating improves, expectancy or reporting calibration remains a serious rival. If physiology changes without better state fidelity, breathing strategy may be carrying the effect.

Disagreement among strands is therefore not failed convergence. It is often the most informative result. The study should preserve the discordance long enough to explain it rather than compressing report, performance and physiology into a single wellness score. A preregistered pattern of expected agreements and informative disagreements turns triangulation from a slogan into a mechanism test.

The expectancy mirror

A meditation trial rarely hides the intervention from participants. People know whether they are sitting quietly, receiving a prestigious programme or waiting for treatment. Teachers know what they teach. Researchers may have invested careers in the intervention. This does not invalidate the work. It changes what a control condition must accomplish.

Across forty-four meta-analyses representing 336 randomised trials and 30,483 participants, mindfulness-based interventions generally showed larger effects against passive controls than against active controls.[3] In MacCoon and colleagues’ validation of the Health Enhancement Program, both MBSR and the active comparator improved several participant-reported outcomes, while MBSR did not outperform the comparator on those measures.[10] That result does not mean MBSR was inert. It means time, instructor contact, group membership, health education, movement and credible expectation can produce real change too.

Expectancy is a route, rather than only a reporting nuisance

Expectation can enter the experiment before the first instruction. Recruitment language attracts people who already believe the programme will help or who hope to access a valued identity. Once enrolled, credibility affects attendance and home practice. During a session, a promise of calm can direct attention towards signs of relief, while a warning about difficulty can make the same bodily event more salient. Afterwards, participants decide whether agitation counts as failure, purification, exposure or harm. Each decision can alter both experience and report.

Teacher and researcher allegiance add another layer. An instructor who knows the hypothesis may give warmer explanations, adapt more skilfully or encourage persistence differently across arms. An assessor may probe one group with more interest. An analyst may prefer outcomes that fit the programme’s theory. Research on allegiance in mindfulness-based interventions is a reminder that the people who design, deliver and evaluate the treatment belong inside the causal diagram, not outside it.[22]

None of this makes expectancy unreal. If belief changes adherence, breathing, threat appraisal or pain regulation, it is a genuine part of the configured exposure. The error is to claim that an operation-specific mechanism caused the whole outcome. A clean study measures expected benefit, treatment preference and intervention credibility before allocation, after the rationale is delivered and after participants have direct experience. It also records guesses about group assignment and uses assessors who are blind where practical.

A good comparator is therefore not merely busy. It must be credible for the population, match contact and demand, and avoid quietly teaching the target operation. Matching every common factor perfectly is usually impossible. The aim is to state which routes were matched, measure those that were not, and limit the conclusion accordingly.

A small numerical example

Suppose stress is scored from zero to twenty, where lower is better. A waitlist begins at fourteen and ends at thirteen. A credible active programme begins at fourteen and ends at ten. The meditation programme begins at fourteen and ends at nine. Against waitlist, the apparent improvement is four points. Against the active comparator, the incremental improvement is one point. Both comparisons are true. They answer different questions.

The active-control contrast estimates the added value of the specified mental operation under those conditions. The passive-control contrast estimates the value of receiving the whole configured programme rather than waiting. Clinical decisions may care about the whole package. Mechanism claims require the narrower contrast.

The control follows the decision

There is no universally correct control because different decisions ask for different causal quantities. A health service deciding whether to offer an eight-week programme may want the effect of the whole package against usual care. A scientist claiming that repeated attentional return changes conflict monitoring needs an active comparison that matches group contact, movement, education and expectation. A designer choosing between two contemplative operations needs a head-to-head comparison with comparable dose and credibility.

This is the estimand problem in ordinary language: what exact difference would change the decision? “Meditation versus nothing” estimates access to a programme, not the unique effect of its mental operation. “Meditation versus a credible active programme” subtracts more common factors, but it can also answer a less clinically relevant question. A factorial design can estimate operation, expectancy and their interaction, yet may feel artificial and reduce adherence.

A positive result permits only the conclusion supported by the contrast. A passive-control advantage supports offering the package under similar conditions. It does not establish a distinctive contemplative mechanism. An operation advantage over a credible comparator supports a narrower mechanism claim, but only for the measured population, dose and outcome. A null result may mean the operation adds little, the control contained an active ingredient, state fidelity was poor or the measure missed the relevant timescale. The design must name these alternatives before results are known.

stateDiagram-v2
  [*] --> Claim
  state "The configured-exposure model" as Claim
  state "The causal object is a<br/>trajectory, not a label" as Test
  state "Expertise has at least four<br/>coordinates" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Safety is part of causal identification

Safety discussions often polarise into two claims: meditation is generally safe, or meditation can cause serious harm. Both can be compatible with the data because they refer to different practices, populations, definitions and methods of asking.

A systematic review of 83 studies estimated an overall adverse-event prevalence of about 8.3 per cent, while warning that trials likely under-reported events and lacked a standard assessment method.[12] In an actively probed sample of ninety-six people completing variants of an eight-week mindfulness-based programme, Britton and colleagues found high rates of meditation-related side effects, with 37 per cent reporting negative impact on functioning and 6 to 14 per cent reporting lasting bad effects depending on the duration threshold.[13] A 2025 population-based survey of 886 United States meditators produced very different rates depending on the instrument, with 9.1 per cent reporting functional impairment.[17]

These numbers should not be averaged into one master risk rate. Spontaneous reports miss experiences that structured interviews find. Broad checklists capture unusual or unpleasant events that may not impair life. Retrospective surveys cannot establish whether distress preceded practice. Intensive retreats differ from manualised clinical programmes. Measurement is not a window onto a fixed prevalence; it helps create the category being counted.

Four categories that should not be collapsed

Some researchers caution that expected transient discomfort should not automatically be labelled an adverse event, particularly when the training deliberately exposes participants to difficult sensations.[20] That is a useful correction, but it cannot justify interpreting every deterioration as progress. Functional impairment, duration, participant goals and capacity to consent to continued exposure matter. A painful physiotherapy exercise can be expected and still require modification when it damages tissue. “Expected” and “safe” are not synonyms.

Risk changes the data-generating process

Safety is often treated as a report placed beside the efficacy analysis. In contemplative research it can change the efficacy estimate itself. A participant who becomes distressed may reduce practice, switch technique, seek outside care, miss a scan or leave the study. If those changes are not recorded, the analysis compares survivors of different exposure paths while pretending that allocation alone defined treatment.

Consider a programme in which ten participants intensify practice rapidly. Three develop persistent sleep disruption. One pauses with support and recovers, one continues after being told the reaction is normal, and one withdraws without giving a reason. A per-protocol analysis may retain only the people who tolerated the exposure. An intention-to-treat analysis preserves allocation but still cannot explain which configuration produced deterioration or recovery. The missing object is the post-allocation trajectory: dose changes, adaptations, support, symptoms and functioning over time.

This produces a form of survivor bias with practical consequences. The people visible at the final scan may be those whose nervous systems, expectations and circumstances fit the programme. Their neural pattern cannot safely be generalised to those who stopped. Attrition should therefore trigger a structured follow-up rather than disappear into a flow diagram. The reason for stopping is both a safety outcome and a clue about the intervention’s boundary.

A safety event is not only an ethical endpoint; it is evidence that the causal configuration changed. Shortening a session, grounding through movement or referring for clinical care may be exactly the right action. The analysis should not hide the adaptation. It should model what happened before and after it. This makes the study more humane and more scientifically honest.

Why apparently conflicting safety studies can all be informative

Hirshberg and colleagues analysed symptom change in community MBSR classes and randomised comparisons, finding no evidence that MBSR produced higher rates of harm than waitlist on the measured symptom indices.[15] Van Dam and Galante argued that this approach could underestimate harms not represented by those broad scales or not elicited through active questioning.[16] Britton’s structured interviews detected many more events because they asked item by item about occurrence, valence, duration and functional impact.[13]

The disagreement is not resolved by choosing one camp. It reveals a measurement hierarchy. Group-level symptom deterioration answers whether the programme worsens common symptom scales on average relative to no treatment. Active adverse-effect interviews answer whether individuals experienced specific meditation-related problems, including domains not covered by those scales. Both belong in a serious trial.

Galante and colleagues add another complication. In a randomised university study, a mindfulness programme increased some altered-state experiences up to a year later, although several trial-arm comparisons lost significance after correction and formal practice dose effects were more persistent.[18] Unity or disembodiment is not inherently therapeutic or harmful. Its significance depends on appraisal, context, duration and functioning. A safety system therefore needs phenomenology as well as symptom counts.

When interpretation must yield to function

Contemplative traditions provide rich languages for difficult experiences. Those languages can help a practitioner make sense of fear, bodily change or altered self-boundaries. They can also become hazardous when one interpretation has authority over all others. A research protocol cannot assume that distress is purification, resistance, trauma release or evidence of progress. Those are hypotheses about meaning, not substitutes for observing sleep, judgement, self-care and ordinary functioning.

The governing rule is simple. Meaning may guide inquiry while function governs escalation. If a participant reports an unusual but welcome shift, sleeps normally and continues daily life, the study can document it without pathologising it. If the same description accompanies worsening insomnia, impaired judgement or inability to work, independent assessment is required regardless of its contemplative interpretation. The threshold concerns consequences and capacity, not whether the experience sounds spiritually elevated.

This rule also protects against the opposite error. Not every difficult sensation requires termination. A brief, expected and consented period of discomfort may be compatible with learning when the person remains able to choose, recover and function. The task is titration: preserve challenge while keeping the participant inside an agreed recovery envelope. That envelope should be specified before the study, not improvised by the person most invested in continuation.

A minimum safety and causal protocol

A research protocol should make six commitments before recruitment:

Specify the exposure. Record technique, instructions, session length, frequency, retreat density, teacher contact and concurrent practices. Assess baseline state. Capture sleep, current distress, recent medication or substance changes, prior adverse meditation experiences and relevant clinical history through appropriately qualified staff. Monitor repeatedly. Ask about unusual experiences, valence, function and duration, not only serious events or drop-out. Separate roles. Let an assessor who is not the teacher conduct at least some safety interviews. Predefine adaptation. State when to shorten, ground, switch practice, pause, withdraw or refer. Follow after exposure. Some effects appear or persist beyond the final session.

This is not a universal clinical algorithm and it should not be used by individuals to diagnose themselves. Acute danger, suicidal intent, mania, psychosis, severe dissociation or inability to care for oneself requires prompt assessment through local emergency or qualified clinical services. The research contribution is narrower: a study must know who holds clinical responsibility and what happens when a threshold is crossed.

Define the configured exposure before reading the brain

Do not approve, interpret or communicate a contemplative neuroscience study because it contains a meditation group and a brain measure. Ask whether the study has identified the configured exposure.

The decision sequence is concrete. Name the mental operation. Define expertise through both history and in-session performance. Measure expectancy and match credible common factors. Record temporal dose rather than total minutes alone. Track current person-state, especially sleep and functioning. Predefine support, adaptation and clinical escalation. Analyse reports, behaviour, physiology and daily functioning as distinct outcomes before looking for convergence.

A brain difference is interpretable only after the study shows what state was produced, which rival routes were controlled and whose deterioration could have been missed. This standard will often make claims smaller. It will also make them more useful. A narrowly identified mechanism can be reproduced, adapted and challenged. A safety signal tied to dose, state and support can change a protocol before it becomes a crisis.

The practical decision is therefore not “meditation: yes or no”. It is: which operation, for this person and purpose, at this density, under this monitoring and recovery capacity? That question respects contemplative expertise without romanticising it, takes neural evidence seriously without turning correlation into revelation, and treats safety as part of understanding what the practice actually does.