Can Neurofeedback Prove Causality?

This review by Kvamme, Ros, and Overgaard (2022) tackles a deceptively simple question with big implications for both neuroscience and clinical practice: can neurofeedback provide evidence of direct brain–behaviour causality? Because neurofeedback involves measuring brain activity in real time and feeding it back to a person so they can learn to change it, many authors have assumed that if behaviour changes after training, then the trained brain signal must be a direct cause of that behaviour.

The authors push back on that assumption. Drawing on modern theories of causality and interventionism, they argue that what neurofeedback tells us is more subtle—and, in some ways, more interesting. Neurofeedback clearly shows that people can self-regulate their brain activity and that this learning can alter perception, emotion, attention, and symptoms. But whether the specific trained signal is the primary cause of those changes, or just one node in a larger causal network, is much harder to prove.

In simple terms, neurofeedback (and biofeedback more broadly) are methods where we make internal physiology visible—brainwaves, blood flow, heart rate, skin conductance—and then use that feedback to help people learn to regulate it. The promise is twofold: better clinical outcomes and deeper insight into how mind and brain are linked.

This paper sits right at that junction. It argues that understanding how neurofeedback exerts its effects—directly on a targeted brain signal, indirectly through upstream networks, or via broader learning and expectation processes—matters for both science and practice. And it frames neurofeedback not as a magic causal lever, but as a powerful tool to distinguish between mere correlations and genuinely non-spurious links between brain activity and cognition.

Methods: how do you study causality with neurofeedback?

This is a conceptual review rather than a single experimental study, so the “methods” are really a set of logical tools and examples the authors use to examine causal claims in neurofeedback.

First, they adopt an interventionist framework of causality, especially Woodward’s and Pearl’s formulations. In this view, to say that brain signal X causes behaviour Y, we must be able to intervene on X (and only X, not on hidden upstream variables Z), and see a systematic change in Y while other causes are held constant. That sounds simple, but the brain is messy.

The authors then contrast four main approaches that link brain and behaviour (illustrated in a diagram in the article):

1. Conventional neuroimaging – mostly correlational: measure variability in brain activity and correlate it with behaviour.

2. EMG biofeedback – provide feedback on muscle activity and observe behavioural or symptom change (e.g., ADHD hyperactivity).

3. Neurofeedback (EEG/MEG/fMRI) – provide real-time feedback from a specific brain signal and train participants to up- or down-regulate it.

4. Brain stimulation (e.g., TMS) – apply an exogenous current or magnetic pulse to induce changes in brain activity and behaviour.

By using causal diagrams and analogies (like the famous “ice cream sales and shark attacks” example), they clarify the difference between:

• Spurious correlations (two things change together but are driven by a third factor).

• Non-spurious correlations (two things are linked via a shared causal network, even if one is not the primary driver).

• Direct causality (intervening on X alone is sufficient to change Y).

They also devote an important section to implicit multivariate pattern analysis (MVPA) neurofeedback. Instead of training the average signal in a region of interest, these protocols decode fine-grained multivoxel or multichannel patterns associated with a specific cognitive state, and reward the brain whenever that pattern reappears—often without telling the participant what is being trained. This technique is presented as the current best candidate for making stronger causal claims.

Throughout, the “method” is essentially to take concrete neurofeedback studies—alpha-band attention training, amygdala down-regulation, EMG biofeedback in ADHD, decoded fear reduction protocols—and analyze what kinds of causal inferences are actually justified when you apply a strict interventionist lens.

Results: what does neurofeedback really tell us about cause and effect?

Because this is a review, the “results” are a set of conceptual conclusions and case studies rather than one data table. Several key threads emerge:

1. Neurofeedback can distinguish spurious from non-spurious brain–behaviour correlations.

   • If repeated, well-controlled neurofeedback studies successfully train a target signal and do not change behaviour, that is evidence that the original correlation between that signal and the behaviour was likely spurious.

   • If successful training does change behaviour, especially relative to strong control conditions (e.g., sham, inverse feedback, non-specific regions), then the target signal is part of a causal network influencing the behaviour—even if it is not the primary cause.

2. Indirect causality is often the most plausible explanation.

   • The EMG biofeedback example in ADHD shows this nicely: reducing muscle activity can lead to ADHD symptom improvements similar to EEG neurofeedback. That does not mean muscles “cause” ADHD; rather, EMG is tapping into a motor–cortical network that indirectly influences symptoms.

   • By analogy, when we train alpha power in parietal cortex or BOLD signal in the amygdala, we may actually be conditioning upstream control systems, connectivity patterns, or meta-cognitive processes that are the deeper drivers of behavioural change.

3. Alpha and attention: a cautionary tale.

   • The paper scrutinises state-of-the-art MEG neurofeedback studies that trained alpha lateralisation and then concluded that alpha oscillations cause attentional shifts. The authors point out that covert attention strategies may still have been used, and that other work shows alpha and attentional markers (like SSVEPs) can change independently.

   • Their more cautious conclusion: alpha is very likely involved in attentional networks, but current neurofeedback evidence doesn’t prove it is the single causal lever for attention.

4. Amygdala neurofeedback supports a network model of emotion regulation.

   • Real-time fMRI studies targeting amygdala activity in anxiety and PTSD consistently find changes not only in the amygdala, but also in prefrontal regions and connectivity between them.

   • Authors of these studies tend to talk about regulation networks rather than one causal region, which fits the indirect/network causality view.

5. Implicit MVPA neurofeedback offers the strongest causal case—so far.

   • By identifying fine-grained patterns that encode a specific perceptual or emotional state and reinforcing them without explicit strategies, decoded neurofeedback can more cleanly “intervene on X while holding Z constant”.

   • Additional analyses, like information transmission mapping, can show that induced patterns are spatially confined to target regions plus reward circuits, strengthening causal claims.

   • Even so, fMRI’s indirect nature and dynamic changes in patterns over training keep the picture complex.

Overall, the “result” is a reframing: neurofeedback is extremely useful for testing whether a brain–behaviour link is real and part of a causal network, but by itself it rarely justifies strong claims of direct, unidirectional causality.

Discussion: from causal purity to clinical usefulness

This review invites a shift in how we talk and think about neurofeedback. Instead of asking, “Does training signal X prove that X causes behaviour Y?”, it suggests we ask, “Does training signal X show that X belongs to a network that reliably influences Y—and how can we leverage that clinically?”

For people seeking help, this is actually a reassuring message. You don’t need your neurofeedback protocol to be a single, precise causal laser pointed at the one “faulty” brain region. What matters is that the training interacts with the right networks—those that support attention, emotion regulation, sensory integration, or self-control—and that these networks adapt in healthier directions. Whether the protocol works through direct modulation of the targeted signal, through upstream control systems, or via broader learning and expectancy effects, the brain is still learning new, more flexible response patterns.

For clinicians and referrers, the paper is a gentle warning against overconfident causal slogans. Statements like “up-training SMR at C3 directly causes better impulse control” or “down-training beta at Fz directly causes reduced anxiety” are too strong when viewed through an interventionist lens. A more accurate description is that such protocols engage networks associated with behavioural inhibition, arousal regulation, and sensorimotor integration—and that repeated, contingent feedback can nudge those networks toward more adaptive regimes.

This network view fits nicely with existing clinical observations:

• Training sensory–motor rhythm (SMR) around 12–15 Hz at sites like C3, Cz, or C4 is often associated with calmer motor output, reduced hyperactivity, and improved sleep. Rather than SMR being a single “cause” of self-control, it likely reflects improved thalamo-cortical gating and more efficient coordination between sensory, motor, and frontal systems.

• Increasing posterior alpha at POz in anxious or over-aroused clients often aligns with lower somatic tension and improved relaxation. The review’s discussion of alpha and attention suggests that alpha is part of a broader inhibitory/gating system; clinically, we are tapping into that system rather than flipping a single “relaxation switch”.

• Frontal protocols (e.g., reducing excessive high beta at Fz/FCz or promoting midrange beta or alpha–theta balance) can support executive control and emotion regulation, again through large-scale prefrontal–limbic networks that rarely change in isolation.

The authors also emphasise that neurofeedback should be seen alongside other methods—lesion studies, TMS/tDCS, conventional neuroimaging, and behavioural interventions. No single method can carry the full “burden of causality”. Instead, converging evidence across methods helps us triangulate which brain dynamics are most central, which are compensatory, and which are epiphenomenal.

Clinically, this encourages a more multimodal mindset: using neurofeedback not as a stand-alone proof of mechanism, but as one component in a network of interventions—psychotherapy, medication when appropriate, lifestyle changes, body-based work, and traditional biofeedback (HRV, EMG, EDA, breathing). Each modality perturbs the system in slightly different ways; together, they help us map out and reshape the relevant networks.

An important interpretive theme in the paper is humility about what our current signals actually represent. EEG rhythms, BOLD activity, and decoded patterns are emergent summaries of vast neural populations. From a causal perspective, it is legitimate to treat these as “entities” that can be intervened upon, but we should remember that they stand in for many interacting mechanisms—self-organisation, redundancy, degeneracy, and feedback loops.

This has two practical consequences:

1. When a neurofeedback protocol “works”, we should assume that multiple mechanisms—learning, expectation, strategy use, network reconfiguration, physiological coupling—are involved.

2. When a protocol “fails” in a trial, we should be cautious about declaring the targeted signal irrelevant; compensatory mechanisms or poorly matched strategies may have masked its role.

In that light, the authors propose that the true strength of neurofeedback, especially in research, is its ability to move us beyond simple correlation. By actively perturbing brain activity endogenously—using the brain’s own learning machinery—we can test whether a putative marker is at least part of the causal web. That alone is a substantial upgrade over static imaging studies.

Brendan’s perspective

Reading this paper, I found myself nodding along and thinking of many clients whose progress would completely mystify a strict, one-signal-one-symptom causal model.

Take the teenager with “classic” ADHD: restless, impulsive, struggling to finish tasks. We might start with SMR training at C3, Cz, or C4 and a bit of frontal beta regulation. After 20–30 sessions, the parents report calmer evenings, fewer school calls, and better sleep. On paper, we could say: “See? Increasing SMR caused reduced hyperactivity.” But in the room, what I actually witness is a network changing: posture relaxing, breathing slowing, a new curiosity about internal sensations, a growing sense of agency—“I can do something with my brain.” One of the major elements that shifts on the qEEG is an increase in posterior alpha (note: it was Sterman showed us this mechanism back in his initial work with cats; SMR training = increase in posterior alpha). 

Kvamme and colleagues give us language for that intuition. When we reward a narrow EEG band at a specific site—say 12–15 Hz at C3, Cz or C4—we are almost certainly not acting on a single causal culprit. Instead, we are nudging a distributed system: thalamic gating, basal ganglia loops, cerebellar timing, prefrontal monitoring, even body tension. The EMG biofeedback example from the paper is a perfect metaphor: you can train the muscle and still meaningfully influence the brain.

In my own practice, I increasingly think in terms of entry points rather than single causes. For a client with anxiety and sensory overload, we might work on:

• Posterior alpha up-training at POz to support inhibition of irrelevant sensory noise.

• High-beta down-training at Fz/FCz to calm cognitive overdrive and worry loops.

• HRV biofeedback to deepen parasympathetic tone.

Each of these is an entry into the same underlying regulatory network. If the client’s panic attacks diminish, it’s tempting to crown one protocol the hero. But from a causality perspective, what we’ve done is repeatedly perturb and reinforce an entire self-regulating system until it finds a more resilient configuration.

This paper also highlights something many clinicians feel but rarely name: research designs can be both essential and misleading for practice. Trials often lock onto a single region or frequency (e.g., “parietal alpha for attention”) and then treat it as the causal mechanism. In real life, we are constantly adapting—we can move reward thresholds, shift electrode placements, tweaking frequency bands, adding or removing tasks—based on how the client responds. The research world wants clean interventions on X; the clinical world operates on adaptive interventions on a family of Xs, Ys, and Zs.

Rather than seeing that as a flaw, I think we should embrace it and describe it more transparently. When I adjust a protocol—say, moving from SMR at C3 to a blend of SMR and low beta at C4, or adding an alpha–theta protocol at POz—I am running quasi-experiments in a single brain. I am asking: “If I perturb this piece of the network, do sleep, mood, attention, or body tension shift?” Over time, patterns emerge that are clinically meaningful even if they never map neatly onto a single causal diagram.

Where does this leave us in terms of concrete EEG protocol thinking?

• For behavioural inhibition and impulse control, I still favour SMR-focused training (12–15 Hz) at central sites (C3, Cz, C4), sometimes combined with inhibiting excessive theta (around 4-8Hz) gentle down-training of excessive high beta (20–30 Hz). I now describe this to clients as supporting “motor brakes and sensory filters” rather than claiming SMR directly causes impulse control.

• For anxiety and hyperarousal, I often target posterior alpha (8–12 Hz) at POz, alongside reducing fast beta where it is clearly overexpressed. The review’s discussion of alpha and attention helps me explain that alpha is part of the brain’s “noise gate”, not a simple relaxation volume knob.

• For emotion regulation and trauma-related symptoms, I sometimes think in terms of “front–back” and “top–bottom” networks: supporting prefrontal stability (e.g., F3, Fz, F4) while calming hyper-reactive posterior and limbic-linked rhythms. The amygdala fMRI work reviewed in the paper reassures me that we don’t have to hit the amygdala directly with EEG to influence that network.

Another important takeaway is how we talk about evidence with colleagues and clients. Instead of promising that a protocol proves a direct causal link, we can say:

• There is good evidence that this brain pattern is consistently involved when people struggle with X.

• Neurofeedback trials show that when people learn to change this pattern, their symptoms of X reliably improve more than in control conditions.

• That means this pattern is part of the causal web of X, even if it’s not the whole story.

Clinically, that’s enough to act on—and it keeps us honest.

Finally, the paper’s discussion of decoded and implicit neurofeedback hints at the future. As tools become more precise—combining EEG with fMRI, using adaptive decoders, training connectivity patterns rather than single channels—we may be able to target networks in a more principled way. Even then, though, I suspect the story will remain one of systems rather than single switches.

For now, this article gives us a useful compass: use neurofeedback as a way to probe and shape causal networks, not as a hammer looking for one nail. It encourages humility in our explanations and creativity in our protocols—both of which, in my experience, make for better outcomes and better conversations in the therapy room.

Conclusion

Kvamme and colleagues invite us to grow up a little in how we talk about neurofeedback. Rather than claiming that training one signal proves a direct cause–effect relationship between that signal and behaviour, they argue that neurofeedback is best understood as a tool for identifying and training causal networks. It can tell us whether a brain–behaviour link is real and influenceable, but not usually whether it is the unique primary cause.

Far from diminishing neurofeedback, this perspective actually strengthens its place in neuroscience and clinical work. By actively perturbing brain activity through learning, neurofeedback goes beyond simple correlation and helps map which patterns and networks are genuinely involved in perception, emotion, and behaviour. For practitioners, it legitimises the flexible, network-based way protocols are already being used in the clinic.

The take-home message is simple: neurofeedback doesn’t need to prove single-variable causality to be powerful. Its real value lies in helping brains discover healthier ways of organising themselves—and that is precisely where its clinical and scientific promise shines.

References

Kvamme, T. L., Ros, T., & Overgaard, M. (2022). Can neurofeedback provide evidence of direct brain–behavior causality? NeuroImage, 258, 119400. https://doi.org/10.1016/j.neuroimage.2022.119400