Anxiety is not a disease. It is a normal and useful response to threat — the alertness that makes you check the road before crossing, prepare for an exam, or notice that something is wrong. A system that produces no anxiety at all would be dangerous to own.
An anxiety disorder is that system firing out of proportion, persistently, and at a cost. The fear or worry is excessive relative to the actual threat, it does not switch off when the threat passes, and it interferes with work, relationships or daily activity. The boundary is not the presence of anxiety but its proportionality, its persistence and its consequences — which is why the diagnosis is a clinical judgement and not a measurement, a point this review returns to.
It is a family of conditions rather than one. They share the underlying machinery and differ in what triggers it:
- Generalised anxiety disorder — pervasive, hard-to-control worry across many domains, with restlessness, tension and poor sleep [1].
- Panic disorder — recurrent, abrupt surges of intense fear with physical symptoms (racing heart, breathlessness, chest tightness, a sense of impending death), plus persistent fear of the next attack [2] [3].
- Social anxiety disorder — fear of scrutiny and negative evaluation, leading to avoidance of social or performance situations [4] [5] [6].
- Specific phobias — intense fear of a particular object or situation.
- Agoraphobia — fear of situations where escape might be difficult, which often follows panic attacks and produces the most restrictive avoidance of all [7] [8].
They are among the commonest mental-health conditions. In the United States, lifetime prevalence of specific phobia is 15.6 percent and of social phobia 10.7 percent, with panic disorder at 6.8 percent, agoraphobia at 2.5 percent and obsessive-compulsive disorder at 2.3 percent [9]. Globally, after adjusting for methodological differences across 87 studies in 44 countries, current prevalence is 7.3 percent, ranging from 5.3 percent in African cultures to 10.4 percent in Euro/Anglo ones [10]; a separate systematic review put pooled 12-month prevalence at 10.6 percent and lifetime at 16.6 percent, with women consistently higher than men [11] [12].
They start early. The median age of onset for anxiety disorders in the National Comorbidity Survey Replication was 11 years — earlier than substance use disorders (20) or mood disorders (30) — with half of all lifetime mental disorders beginning by age 14 and three quarters by 24 [13] [14] [15]. Within the anxiety family the split is wide: phobias and separation anxiety have median onsets at 15 to 17, while panic disorder and generalised anxiety onset at 23 to 30 [9]. Behavioural inhibition in childhood has been proposed as a precursor to later panic disorder and social phobia [16].
They are disabling, and they are undertreated. Social phobia impairs a broad spectrum of outcomes from school dropout to quality of life, and depressive comorbidity contributes only modestly to that impairment [17]. Generalised anxiety produces disability comparable to major depression even in its pure form [1]. And in a primary-care study of 965 patients, 19.5 percent had at least one anxiety disorder — and 41 percent of those reported no current treatment at all [18]. Treatment rates have risen over time [19], and a global return-on-investment analysis estimated benefit-to-cost ratios of 2.3 to 3.0 to 1 for scaling up treatment of depression and anxiety on economic grounds alone, rising to 3.3–5.7 to 1 when health gains are valued [20].
And they travel with depression to an extent that challenges the idea they are separate. Among people meeting 12-month criteria for generalised anxiety disorder, 59.1 percent also met criteria for major depression and 55.9 percent for another anxiety disorder [21]. The genetics goes further, and is covered below.
A note on tone, because this literature invites overstatement in both directions. Anxiety disorders are real, common and impairing, and effective treatments exist. They are also a field where the size of treatment effects depends heavily on what the treatment was compared against, where most trials are of modest quality, and where the honest summary is more equivocal than either enthusiasts or sceptics usually allow. That is the subject of the centerpiece.
Centerpiece: what you compare a treatment against decides how well it appears to work
The cleanest dataset in this field is a network meta-analysis of social anxiety disorder: 101 trials, 13,164 participants, 41 interventions or control conditions in 17 classes [22]. It is unusually informative because it reports the same interventions against two different kinds of comparator.
Against a waiting list, everything works. Individual CBT scores a standardised mean difference of 1.19, MAOIs 1.01, benzodiazepines 0.96, group CBT 0.92, SSRIs and SNRIs 0.91, exposure with social skills training 0.86, self-help with support 0.86, anticonvulsants 0.81, self-help without any support 0.75, and psychodynamic psychotherapy 0.62 [22]. Ten classes, all clearing the line, most of them in the "large effect" range.
That left panel should provoke suspicion rather than optimism. Self-help with no therapist contact at all scores 0.75 — higher than psychodynamic psychotherapy and within reach of the anticonvulsants. A comparator that cannot distinguish an unsupported booklet from an active drug class is not measuring what it appears to measure. A waiting list controls for the passage of time and for regression to the mean, and for nothing else: not expectation, not attention, not the act of being enrolled in something.
Against a real placebo, most of it goes. The same analysis found that individual CBT versus psychological placebo (−0.56) and SSRIs/SNRIs versus pill placebo (−0.44) were the only two classes that still beat their appropriate control [22].
The check the arithmetic was never given. The paper prints both comparators for both classes and never divides one by the other. Doing so:
- Individual CBT: 0.56 / 1.19 = 0.471. Fifty-three percent of the apparent effect disappears.
- SSRIs and SNRIs: 0.44 / 0.91 = 0.484. Fifty-two percent disappears.
The two surviving fractions agree to within 0.013. A talking therapy compared against a psychological placebo and a drug class compared against a pill placebo — two entirely different modalities judged against two entirely different kinds of control — lose almost exactly the same share of their measured effect. That is strong evidence that the shrinkage is a property of the comparator design, not of the treatment. Whatever a waiting list fails to control for, it fails to control for it by about the same amount in both literatures.
An independent meta-analysis reaches the same conclusion by a different route. Across 144 trials and 184 comparisons, CBT produced effects that looked large — g = 0.80 for generalised anxiety, 0.81 for panic and 0.88 for social anxiety [23]. But more than 80 percent of the anxiety trials used waiting-list controls, and the few using other controls "pointed at much smaller effect sizes"; only 17.4 percent of trials were high quality, and restricting to those dropped panic to 0.61 and social anxiety to 0.76. The authors' own conclusion is worth quoting in substance: CBT is probably effective, effects are large against waiting list but small to moderate against care-as-usual or pill placebo, and because so few trials are high quality "these effects are still uncertain and should be considered with caution" [23].
And a second comparison the papers never make. That same analysis adjusted CBT in generalised anxiety from g = 0.80 to 0.59 for publication bias — a 26 percent loss. Swapping a waiting list for a placebo costs 53 percent. The choice of control condition is roughly twice as large a distortion as publication bias, and publication bias is the one that gets the attention.
A third, independent estimate sits between the two. A meta-analysis restricted to the 27 randomised placebo-controlled trials of CBT for adult anxiety disorders found a pooled Hedges' g of 0.73 for continuous anxiety severity, with no significant difference in attrition between CBT and placebo [24]. Higher than the 0.56 from the network analysis, well below the 1.19 against waiting list.
What this does and does not mean. It does not mean the treatments do not work. Two of them clear a genuine placebo, which is a real result, and 0.44 to 0.73 is a clinically meaningful range — comparable to much of general medicine. It means that the headline effect sizes quoted for anxiety treatment are inflated by roughly a factor of two by the comparator, that most of the apparent difference between rival treatments dissolves when they are tested properly, and that claims of one modality's superiority over another should be treated with scepticism unless the comparison was direct.
Placebo response in anxiety is substantial and is not nothing. In panic disorder, patients who responded to placebo differed from those who responded to sertraline on quality-of-life measures across all seven definitions of response examined — placebo responders "may show symptom relief but may not experience improvement in quality of life" [25]. The placebo effect in anxiety and depression has been reviewed in its own right [26]. So the placebo response is real, is large, and is not identical to drug response.
Pillar 1: measurement and diagnosis
There is no biomarker, and this is the central fact
No blood test, scan or genetic assay diagnoses an anxiety disorder. Diagnosis is clinical: a specified set of symptoms, present for a specified duration, causing distress or functional impairment, and not better explained by something else. Everything below is a way of making that judgement more consistent, not of replacing it.
Functional neuroimaging finds reproducible group differences — a meta-analysis of emotional processing across PTSD, social anxiety disorder and specific phobia found greater amygdala and insula activity than in matched comparison subjects, and a similar pattern during fear conditioning in healthy people [27], with amygdala hyperactivity to threatening faces tracking social anxiety severity [28] [29] [30] [31] [32]. Those are group-level findings with overlapping distributions. None of them diagnoses an individual, and the finding that healthy people show the same pattern during fear conditioning is precisely why: the circuitry is normal circuitry, operating out of proportion.
The scales, and what they are actually for
The GAD-7 is the standard brief measure of generalised anxiety: seven items, each scored 0–3 [33]. In its validation on 2,740 primary-care patients, a cut point optimised sensitivity at 89 percent and specificity at 82 percent, rising scores tracked functional impairment on all six SF-20 scales and on disability days, and — importantly — factor analysis confirmed anxiety and depression symptoms as distinct dimensions with differing but independent effects on impairment, despite frequently co-occurring [33]. It was subsequently standardised in the general population, where internal consistency was 0.89 across all subgroups, women scored higher than men (3.2 versus 2.7), and about 5 percent of the population scored 10 or more and 1 percent scored 15 or more [34]. Its one-factor structure and psychometrics have held up in large clinical samples [35], and its sensitivity to change and minimal clinically important difference have been examined [36]. A two-item version, GAD-2, performed well as a screen for all four common anxiety disorders in primary care, with areas under the curve of 0.80 to 0.91 [18].
Disorder-specific instruments exist because the disorders differ: the Panic Disorder Severity Scale [37] [38], the Liebowitz Social Anxiety Scale [39] [40] and social phobia scrutiny/interaction measures [41], the Penn State Worry Questionnaire for generalised anxiety [42], the Yale-Brown Obsessive Compulsive Scale [43], the Body Sensations and Agoraphobic Cognitions Questionnaires and the Mobility Inventory for agoraphobia [44] [45], and the Clinician-Administered PTSD Scale and PTSD Checklist [46] [47]. General instruments — the Hospital Anxiety and Depression Scale [48] [49] [50], the Beck Anxiety Inventory [51], the Depression Anxiety Stress Scales [52] [53] [54] and the PHQ anxiety scales [55] — cover the ground more broadly.
What all of them are for is severity and change, not diagnosis. A GAD-7 of 12 does not diagnose generalised anxiety disorder; it quantifies symptom burden in someone whose diagnosis has been made clinically, and lets you tell whether treatment is working. Defining what counts as a meaningful change is itself a methodological problem with a formal literature [56].
Ruling out the mimics
Several medical conditions produce the physical symptoms of anxiety, and several produce panic attacks specifically. Thyroid disease is the classic: hyperthyroidism causes palpitations, tremor, sweating, heat intolerance and restlessness, and is excluded with one blood test. Cardiac disease matters in the opposite direction — chest pain and palpitations in panic disorder are frequently investigated for cardiac causes, and should be, because the symptoms genuinely overlap. Substances cut both ways: caffeine, stimulants, and alcohol or benzodiazepine withdrawal all produce anxiety states, and withdrawal in particular is a treatable cause that is easily mistaken for the primary disorder [57] [58].
The related clinical skill is distinguishing disordered from normal anxiety — and the epidemiology shows the boundary is genuinely graded rather than sharp. Subthreshold generalised anxiety, defined below the diagnostic threshold, is more common than the disorder itself and carries its own impairment [21].
The overlap with depression, and how deep it goes
This is the finding that most complicates the categorical picture. Bivariate twin modelling in 1,033 female twin pairs found that genetic factors were important for both major depression and generalised anxiety disorder and were completely shared between them; familial environment played no role in either [59]. A follow-up in the same cohort found a genetic correlation of unity between the two, with the environmental correlation +0.70 under a non-hierarchical diagnosis and zero under a hierarchical one — leading to the title "same genes, (partly) different environments" [60].
If two diagnoses share their genetic liability entirely, they are not two diseases in any biological sense. The tripartite model offers the standard reconciliation: a large shared component of general distress, plus physiological hyperarousal specific to anxiety and anhedonia specific to depression [61]. The GAD-7 data are consistent with it — anxiety and depression separate as factors and have independent effects on impairment, even while co-occurring constantly [33]. Comorbidity of mood and anxiety disorders has been documented since the field began measuring it [62] [63], and sleep disturbance is bidirectionally related to both [64].
Pillar 2: management, with the uncertainty stated
Two first-line options have real evidence behind them, and — this is the honest headline — they perform about as well as each other, and better than placebo by a moderate amount.
Cognitive behavioural therapy, and exposure in particular
CBT for anxiety disorders is a family of techniques, and the most specific of them is exposure: systematic, planned, repeated contact with the feared situation, without the escape or safety behaviour that normally follows.
The rationale is a cycle. Anxiety prompts avoidance; avoidance produces immediate relief; the relief reinforces the avoidance; and because the person never stays in the situation long enough to learn that the feared outcome does not occur, the fear is preserved intact. Avoidance is the mechanism by which anxiety disorders sustain themselves. Exposure breaks the cycle by removing the escape, so that the fear response has the opportunity to extinguish. Agoraphobic avoidance is the clearest case, and the treatment literature is built on it [65] [66] [67].
The evidence that exposure is doing specific work, not general work. The most striking demonstration is a trial in panic disorder with agoraphobia comparing alprazolam and exposure, alone and combined. Both were effective, but exposure had twice the effect size of alprazolam; during taper and follow-up, gains after alprazolam were lost while gains after exposure were maintained; combining the two "marginally enhanced gains during treatment, but impaired improvement thereafter"; and — the honest detail — panic itself improved as much with placebo as with alprazolam or exposure, with the drug/exposure difference showing up on phobias and disability [68]. In obsessive-compulsive disorder, intensive exposure and ritual prevention produced treated and completer response rates of 62 and 86 percent, against 42 and 48 percent for clomipramine and 8 and 10 percent for placebo, with the combination no better than exposure alone [69].
CBT for adult anxiety disorders has been meta-analysed against placebo at g = 0.73 [24] and reviewed across disorders [70] [71] [72] [22]. Specific protocols have been tested for generalised anxiety — applied relaxation against CBT [73], and a treatment targeting intolerance of uncertainty in which 20 of 26 participants (77 percent) no longer met diagnostic criteria after treatment, with gains maintained at 6 and 12 months [74] — for panic [75] [76] [77], and for social phobia [78] [79]. It works in children and adolescents as well [80].
The honest caveats. The waiting-list problem above applies in full. And a multidimensional meta-analysis that looked past effect sizes to clinical outcomes found something less comfortable: a substantial proportion of panic patients improve and stay improved, but most patients treated for depression and generalised anxiety do not improve and remain improved at clinically meaningful follow-up intervals — and there was "a systematic relation across studies between exclusion rates and outcome", meaning the trials that screened out the most complex patients reported the best results [70].
SSRIs and SNRIs
These are the first-line drugs, acting on the serotonin transporter [81]. They work across the anxiety disorders — paroxetine in panic disorder [82], fluvoxamine in social phobia [83], SSRIs in OCD against placebo [84] and against clomipramine [85], and sertraline in PTSD [86] [87]. Venlafaxine extended release outperformed both placebo and buspirone in generalised anxiety, and the trial shows the time course clearly: it was significantly better than placebo at all time points except weeks 1 and 2 [88]. That delay is the single most important practical fact to tell a patient — these drugs do not work immediately, and stopping at two weeks because nothing has happened is the commonest avoidable failure.
Against pill placebo the effect is SMD −0.44 in social anxiety [22] — real, moderate, and about half what the waiting-list comparisons suggest. Network meta-analysis of pharmacological treatments for generalised anxiety exists [89], as do class comparisons [81] [90].
Side effects and stopping. Sexual dysfunction and weight gain are the common long-term problems [91]. Discontinuation symptoms are real and vary by drug: all three antidepressants studied produced more discontinuation symptoms than placebo, with escitalopram producing significantly fewer than paroxetine or venlafaxine XR — and, usefully, no evidence that longer treatment increases discontinuation burden [92] [93]. Tapering matters; duration of treatment apparently does not make tapering worse.
Benzodiazepines: effective, and the reason they are not first-line
Benzodiazepines act on the GABA-A receptor and they work — in the social anxiety network analysis they scored 0.96 against waiting list, statistically indistinguishable from the SSRIs at 0.91 [22], and one meta-analysis of social phobia found benzodiazepines and SSRIs equally effective and more effective than controls [72]. In panic disorder they are superior to control for panic attacks [94].
The problem is what happens next. In the alprazolam-versus-exposure trial, gains during treatment were largely as previously reported, but relapse was usual after alprazolam was stopped, whereas gains persisted to six-month follow-up after exposure ceased [68]. In a controlled comparison, panic-free rates on completion were 87 percent for behavioural panic control treatment, 50 percent for alprazolam, 36 percent for placebo and 33 percent for waiting list — and alprazolam differed significantly from neither the behavioural treatment nor placebo [95].
Discontinuation is difficult and well characterised. In a study of long-term therapeutic users, relapse back onto benzodiazepines occurred in 27 percent of those on long half-life drugs and 57 percent of those on short half-life drugs; short half-life and higher daily dose predicted more severe withdrawal; and patients who managed five weeks free of benzodiazepines reached lower anxiety levels than before discontinuation [57] [96] [58]. That last finding is the important one: for some long-term users the drug had become a cause of anxiety. CBT helps people get off them — 76 percent successful discontinuation with a cognitive-behavioural programme against 25 percent with slow taper alone, and 77 percent of those still benzodiazepine-free at three months [97].
Add the harms in older people — benzodiazepines and hip fractures [98], and a broader review of adverse outcomes [99] — and the position is clear and not moralistic: benzodiazepines are genuinely effective in the short term, they do not deliver durable benefit, they are hard to stop, and they can impair the exposure-based learning that does produce lasting change [68].
Other options, combination and stepped care
Buspirone was significantly better than placebo in generalised anxiety on global improvement at weeks 6 and 8, but less effective than venlafaxine XR [88]. Anticonvulsants scored 0.81 against waiting list in social anxiety [22]. MAOIs are effective — phenelzine beat cognitive behavioural group therapy on some measures at 12 weeks, and worked faster [78] — but their dietary and interaction constraints keep them in reserve.
Combination therapy is not reliably better than the parts. In panic disorder, a Cochrane review of 23 comparisons and 1,709 patients found combined psychotherapy plus antidepressant superior to antidepressant alone (RR 1.24) and to psychotherapy alone (RR 1.17) in the acute phase — but with more dropouts due to side effects [100]. In the largest single panic trial, imipramine and CBT were each superior to placebo (response 45.8 and 48.7 percent versus 21.7 percent on the Panic Disorder Severity Scale), the combination reached 60.3 percent but was not significantly better than either alone, and among responders imipramine produced a higher-quality response [101]. In OCD, exposure plus clomipramine was no better than exposure alone [69]. And in agoraphobia, combining alprazolam with exposure actively impaired improvement after treatment stopped [68].
Physical activity has a modest, real effect: a meta-meta-analysis across 306 study effects and 10,755 participants found physical activity reduced anxiety in non-clinical populations, with depression reduced by a medium effect (SMD −0.50) [102] [103]. That is a supplement to treatment, not a substitute for it.
Pillar 3: progress
Access is the largest solvable problem, and digital delivery is the main lever. Forty-one percent of primary-care patients with an anxiety disorder were receiving no treatment [18], and there are nowhere near enough trained therapists. Computerised CBT for anxiety and depressive disorders achieved a mean effect size of 0.88 (number needed to treat 2.13) across 22 controlled comparisons, with benefit across all four disorders, maintained a median 26 weeks, and good adherence [104] — though, consistent with the centerpiece, effect sizes were non-significantly higher against waitlist than against active controls. Internet CBT works for panic and agoraphobia [105] [106], transdiagnostically across three anxiety disorders with effect sizes of d = 0.76–1.44 [107], and — the finding with the most policy weight — guided self-help and face-to-face psychotherapy produced a difference of d = −0.02, essentially identical, across 21 studies [108] [109] [110]. Therapist time in these programmes runs to about 75 minutes per patient [106] [107]. Conversational-agent delivery has been tested in young adults [111].
Translational fear-extinction work is the most mechanistically ambitious direction. Because exposure is extinction learning, agents that enhance extinction should enhance therapy — and D-cycloserine added to exposure produced significantly less social anxiety than exposure plus placebo, with medium-to-large controlled effect sizes in a 27-participant pilot [112]. The underlying biology is well developed: fear memories require protein synthesis in the amygdala for reconsolidation after retrieval [113], amygdala circuitry bidirectionally controls anxiety [114] [115], medial prefrontal stimulation suppresses central amygdala output [116], and the phasic-versus-sustained distinction maps onto fear versus anxiety in the extended amygdala [117] [118] [119]. Targeting these circuits from laboratory to clinic is an explicit programme [120], and novel GABA-A subtype-selective agents aim to separate anxiolysis from sedation and dependence [121] [122] [123].
Precision matching remains aspirational. The serotonin-transporter promoter polymorphism 5-HTTLPR was associated with anxiety-related traits [124] and with amygdala reactivity [125] [126] — a much-cited line of work whose clinical yield has been limited. Threat-related attentional bias is reliably present in anxious individuals across paradigms but with an effect size of only d = 0.45 [127], which is a fair summary of where cognitive markers stand: real, replicable, and too small to allocate treatment with.
Careful re-evaluation of long-term drug use is the quieter progress. The discontinuation literature is now specific enough to inform tapering [92], the benzodiazepine withdrawal syndrome is characterised in detail [57], and CBT-supported discontinuation has a randomised evidence base [97].
What is not progress, stated plainly. Novel psychoactive agents attract attention disproportionate to their current evidence in anxiety disorders specifically. This substrate contains MDMA-assisted therapy trials in PTSD [128] [129] [130], LSD-assisted therapy for anxiety with and without life-threatening illness [131] [132], cannabidiol in a simulated public-speaking test [133] and reviews of psychedelic-assisted psychotherapy [134]. These are small trials, mostly in populations distinct from the common anxiety disorders this review covers, and none of them supports a claim of established efficacy in generalised anxiety, panic or social anxiety disorder. The review notes their existence and stops there.
Dig deeper in lmmol
Depression is the companion condition and the two should be read together: 59.1 percent of people with 12-month generalised anxiety also meet criteria for major depression [21], twin modelling puts the genetic correlation between them at unity [60] [59], and the tripartite model is the standard account of how two conditions sharing all their genetic liability can still be distinguished — by physiological hyperarousal in anxiety and anhedonia in depression [61]. The centerpiece applies to both literatures, and the same meta-analysis that found CBT's anxiety effects shrink against active controls found the same for major depression [23]. Thyroid disease is the first medical mimic to exclude — hyperthyroidism reproduces the physical symptoms of anxiety and one blood test settles it. Coronary artery disease is the mimic in the other direction, since the chest pain and palpitations of panic attacks overlap genuinely with cardiac symptoms and warrant investigation. Obstructive sleep apnea belongs here because insomnia is bidirectionally related to anxiety and depression [64], and unrefreshing sleep is on both symptom lists. Migraine and epilepsy are the neurological neighbours where anxiety comorbidity is high and where episodic, unpredictable attacks generate anticipatory fear and avoidance of their own — the same cycle exposure therapy targets. On the molecular side, the serotonin transporter is the SSRI target and the site of the most-studied candidate polymorphism in anxiety [124] [125], and the GABA-A receptor alpha-1 subunit is where benzodiazepines act and where subtype-selective successors are being sought [121] [122]. The full collection is at health.