Respiratory therapy for the treatment of anxiety: meta-analytic review and regression

Abstract

Objective:

Respiratory abnormalities are a hallmark of anxiety symptomatology and may serve as clinically useful modifiers for alleviating anxiety symptoms. However, gold-standard anxiety treatments (e.g., cognitive-behavioral interventions) often do not directly address respiratory components despite their theoretical utility and clinical accessibility. This review examined the clinical effectiveness of respiratory interventions, interventions that directly target respiration abnormalities and processes, in treating trait anxiety symptoms.

Methods:

The final analysis included 40 randomized controlled trials including at least one measure of trait anxiety, a respiratory-focused intervention group, and a non-respiratory control-group (active or inactive treatment). Overall effects of respiratory focused interventions were examined, as well as the effect of hypothesized moderators.

Results:

Respiratory component interventions yielded significantly greater improvements (moderate to large effect) in anxiety symptoms than controls, with the stronger effects observed in comparison to inactive, rather than active, control conditions. Significant heterogeneity in findings suggests that variability in intervention design, population, and control comparison may obfuscate interpretation of findings.

Conclusions:

Evidence supports the clinical utility of respiratory interventions as either an independent anxiety treatment, or as an adjunct to other interventions. Clinical and research implications of findings along with recommendations for ongoing investigations in this domain are discussed.

Anxiety and related disorders are estimated to affect 11.6% (95% CI 7.6–17.7%) of people worldwide (Baxter et al., 2013) and 21.3% of the U.S. adult population ages 18–64 annually (Kessler et al., 2012). Given anxiety disorders’ global health burden and associated impairment (Baxter et al., 2014), extensive efforts have been made to better characterize their etiology and course and to refine extant evidence-based interventions utilized to treat them (e.g., Baxter et al., 2014; Lau & Rapee, 2011; Olatunji et al., 2010). Yet, even among those who receive what is regarded as ‘gold standard’ anxiety treatment (i.e., cognitive-behavioral therapy; CBT), effect sizes are small to medium (Carpenter et al., 2018). In addition, there is a notable treatment gap where, according to one report, just 21.7% of individuals diagnosed with an anxiety disorder have sought treatment from a health care provider in the prior 12 months (Wang et al., 2005). There is thus a need to decrease barriers to treatment (Wang et al., 2005) and to develop easily disseminable and effective treatment packagess that target symptoms central to anxiety pathology and maintenance.

Cognitive-behavioral interventions that have received the most robust empirical support among non-pharmacological interventions for anxiety disorders largely focus on addressing maladaptive beliefs and behaviors that serve to maintain anxiety and associated impairment (Carpenter et al., 2018; Cuijpers et al., 2016; Hofmann et al., 2012; Hofmann & Smits, 2008; Montero-Marin et al., 2018; Olatunji et al., 2010; Olatunji & Wolitzky-Taylor, 2009; Tolin, 2010). However, for many individuals, cognitive-behavioral interventions do not result in clinically significant changes, with smaller effect sizes obtained for posttraumatic stress disorder (PTSD)¹, social anxiety disorder (SAD), and panic disorder (PD; Carpenter et al., 2018). One possible explanation for this treatment gap is that CBT may fail to adequately address hyperarousal often observed in these disorders. Exposure, particularly interoceptive exposure, which addresses fear of bodily sensations associated with anxious arousal (Boettcher et al., 2016), tends to serve as the primary empirically supported means by which somatic symptoms are targeted in CBT interventions. Yet, use of exposure is low even amongst therapists who report practicing cognitive-behavioral interventions (Becker et al., 2004; Hipol & Deacon, 2013; Whiteside et al., 2016; Wolitzky-Taylor et al., 2015). This is likely due to a lack of therapist training and/or confidence, or of perceived discomfort on the part of clients who may report that exposures are aversive (Becker et al., 2004; Gunter & Whittal, 2010; van Minnen et al., 2010). Moreover, interoceptive exposures are designed to expose the individual to sympathetic arousal, thus the effective target is psychological distress rather than arousal itself, suggesting that therapists and clients may benefit from alternative intervention approaches (Boettcher et al., 2016).

Respiration is one viable target that may serve as a more direct means of targeting anxiety symptoms, and in particular, those related to anxious arousal (Paulus, 2013). State and clinical anxiety have been found to be closely related to an elevated perception of respiratory dysregulation (Leander et al., 2014), altered respiratory features (Grassi et al., 2013; Stein et al., 1995), and elevated responsivity to respiratory challenges (Papp et al., 1997). Hyperventilation, for example, defined as an arterial blood pH above 7.4 (i.e., respiratory alkalosis) and blood CO₂ (i.e., PaCO₂) levels less than 30 mmHg (i.e., hypocapnia; Balke & Lillehei, 1956; Cohen, 1988; Gibson, 1978; Lickel et al., 2008), is a core symptom of panic disorder and a diagnostic specifier in other anxiety disorders (American Psychiatric Association [APA], 2013). Importantly, meta-analytic work examining respiratory dysfunction in panic disorder has found that dysfunctional breathing, a term we will use to describe the various respiratory patterns that produce hyperventilation (Boulding et al., 2016; Courtney, 2017; Courtney, Greenwood, & Cohen, 2011; Courtney, van Dixhoorn, et al., 2011), appears to be chronic in this population, even in the absence of a panic attack (Grassi et al., 2013). Also, sensations of dyspnea (i.e., breathlessness or the inability to get an adequate breath) are frequently observed in anxious individuals, such that anxiety appears to precede the onset of, and perhaps contribute to, dyspnea (Neuman et al., 2006). Thus, it seems probable that chronic dysfunctional breathing and related sensations could increase vulnerability to anxiety symptoms when stress or other respiratory stimulants cause blood levels of carbon dioxide (PaCO₂) to reach symptom levels.

Although the relation between dysfunctional breathing and anxiety appears to have particular specificity to panic, trait anxiety is also related to respiratory dysregulation (Ben-Amnon et al., 1995; Leander et al., 2014), and through complex processes affects affective and cognitive dysregulation. The characteristic overuse of accessory muscles of the thorax and shoulders during dysfunctional breathing, versus abdominal breathing (Courtney, van Dixhoorn, et al., 2011), may contribute to increases in sympathetic nervous system activity (Delaney et al., 2010; Fatouleh & Macefield, 2013; Hellström et al., 2005; Radovanovic et al., 2015) and increased respiratory drive. Ultimately, this cascade can result in a positive feedback loop where increases in muscle tension lead to increased sympathetic arousal producing afferent feedback to the central nervous system that can compound anxiety and further increase ventilation (Ritz et al., 2013).

The effects of this forward feedback loop may exacerbate dysfunctional breathing and in turn lead to subsequent blood vessel constriction in the brain and body (Bindslev et al., 1986; Malatino et al., 1992), thus decreasing availability of oxygen (Bindslev et al., 1986; Malatino et al., 1992). Resultant increased affinity of hemoglobin to oxygen, i.e., the ‘Bohr effect’ (Bohr et al., 1904) compounds oxygen deprivation to body tissues and the brain, which may impact cognitive impairment (Van Diest et al., 2000) and contribute to dissociation (Michel et al., 1966; Simeon et al., 2003), in addition to its effects on physiological symptoms of anxiety (Van den Bergh et al., 2013). In some cases, these uncomfortable and sometimes frightening symptoms may further contribute to feelings of anxiety and ultimately panic. Therefore, targeting dysfunctional breathing across anxiety pathology may be beneficial.

Additional support for targeting dysfunctional breathing in anxious individuals comes from research on anxiety sensitivity, defined as the fear of anxiety relevant sensations and their potential negative social, mental, and physical outcomes (Reiss et al., 1986). Anxiety sensitivity, the purported target of interoceptive exposures, is an established risk for the onset and maintenance of anxiety and its disorders, in particular those characterized by autonomic arousal, such as panic disorder and posttraumatic stress disorder (Olatunji & Wolitzky-Taylor, 2009). The physical concerns subscale of the anxiety sensitivity index, as compared to mental and social concerns, is correlated with greater bodily vigilance (Zvolensky et al., 2007; Zvolensky et al., 2002). This could magnify the salience of uncomfortable respiratory-related symptoms and further potentiate an anxiety spiral. In support of this, anxiety sensitivity has been found to predict self-reported ratings of dyspnea and dyspnea-related avoidance (Simon et al., 2006). Moreover, the use of interoceptive exposure in the context of cognitive-behavioral therapy has been linked to significant reductions in anxiety sensitivity (Boswell et al., 2013; Collimore & Asmundson, 2014; Wald & Taylor, 2008), suggesting that changes in anxiety sensitivity may mediate treatment effects. It is therefore plausible that among anxious individuals, subjective reports of respiratory distress in the absence of pulmonary pathology may contribute to misinterpretation and avoidance of physiological symptoms. This avoidance may also generalize to other various behaviors that can generate these sensations through increases in respiratory drive, including exercise, exposure to stress, altitude, or elevations in ambient carbon dioxide that might occur in crowded settings. Whereas interoceptive exposures may promote inhibitory learning by altering maladaptive beliefs about anxiety, interventions that target respiratory processes directly may offer an alternative approach toward diminishing anxiety interference.

Together, given the close connection between dysfunctional breathing and anxiety symptoms, therapies that directly target respiration might be expected to have a clinically significant therapeutic effect, both via direct effects on the central and parasympathetic nervous systems, and indirectly via secondary effects on cognitive vulnerabilities. For example: (1) diaphragmatic breathing can reduce the use of thoracic muscles; (2) controlled or paced breathing can regulate dysfunctional breathing via timed inhalation and expiration; (3) heart rate variability biofeedback, whereby individuals are taught to breathe at the resonance frequency of the baroreflex system (approximately 6 breaths/minute), can improve alveolar gas exchange efficiency (Bernardi et al., 1998; Giardino et al., 2004), thereby preventing hyperventilation (Bernardi et al., 2007; Bernardi et al., 1998; Spicuzza et al., 2005), and increase parasympathetic activation (Lehrer et al., 2003); and (4) PaCO₂ biofeedback (as approximated by end-tidal CO₂ [ETCO₂]; Meuret & Ritz, 2010; Meuret et al., 2008) can decrease hypocapnia and improve availability of oxygen to the tissues through the Bohr effect (See Table 1 for descriptions of respiratory intervention types).

Table 1.

Search Terms

Respiratory Search Terms	breath* retrain*, respirat* retrain*, HeartMath, freeze framer, breathing and meditat*, respir* meditation, ETCO2 biofeedback, zen, tai chi, gigong, chi-gong, qigong, mindful*, emwave, jan van dixhoorn, pranya*, pranaya*, vinyasa*, buteyko, respiratory exercises, stress eraser, cardiosignal*, resp-e-rate, Whole body breathing, capnometry assisted respiratory training, diaphragm* breath*, paced breathing, paced respiration, breathing exercises, slow breathing, slow respiration, abdominal breathing, abdominal respiration, relaxed breathing, relaxed respiration, resonance frequency biofeedback, resonant frequency biofeedback, HRV*, heart rate variability, RSA*, respiratory sinus arrhythmia
Anxiety Search Terms	phobi*, panic*, fear*, psychological stress, PTSD, Post-traumatic stress, posttraumatic stress, anxi*, ataque*, taijin kyofusho, khyal, shenkui, grisi sicknis, effort syndrome, neurocirculatory ischemia, ataca de nerviosa, emotional stress, Hypocapn*, Alkalosis, alkylosis
Treatment Search Terms	interven*, treat*, feedback, biofeedback, retrain*, train*, therapy, therapeutic, therapies, management*

^*indicate truncated versions of each term

Despite the strong theoretical utility of respiratory interventions for the treatment of anxiety, availability of these interventions, and their common use in clinical intervention, to date no one has systematically reviewed literature on their independent clinical effectiveness. Notably, some CBT interventions have included relaxation and breathing retraining as additional techniques to target anxious arousal and comparison of cognitive-behavioral interventions with versus without these components has found no significant difference in clinical outcomes (Siev & Chambless, 2007), or additive value (Hofmann et al., 2012; Olatunji et al., 2010). Yet, breathing retraining is just one means by which to target dysfunctional breathing and these findings do not negate the potential clinical value of respiratory interventions for anxiety. Moreover, we do not anticipate that respiratory interventions should function as standalone treatments for anxiety given the importance of cognitive (e.g., worry) and behavioral (e.g., avoidance) symptoms that may play a more central role in impairment (Stapinski et al., 2010), and the experience of psychophysiological symptoms that are less directly related to dysfunctional breathing (McKinnon et al., 2015; Pollard et al., 2017; Saigo et al., 2014). While respiratory interventions may not adequately address these various symptom domains, a clearer understanding of their utility as a treatment component or adjunct is warranted.

The current meta-analytic review examined the effect size of interventions that target various manifestations of dysfunctional breathing for the treatment of anxiety. In addition to using DSM diagnostic categories (APA, 1994, 2013), we approached anxiety symptoms dimensionally, a perspective that has gained wide acceptance in recent research (Kliem et al., 2016; Kotov et al., 2017; Kotov et al., 2015; Uliaszek et al., 2015), and is supported by extant findings linking dysfunctional breathing to various anxiety symptoms independent of DSM disease categories (Crockett et al., 2016; Hagman et al., 2008; Paulus, 2013; Simon et al., 2006). This dimensional approach allowed us to examine the effects of respiratory interventions on specific symptoms that have a more psychophysiological basis as compared to disorder categories.

Methods

Literature Search

A combination of respiratory, anxiety, and treatment search terms were employed (see Table 3) to capture all interventions that may have used a respiratory component and measured trait anxiety as a primary or secondary outcome of interest. Specifically, search terms within each of the three categories were linked with the Boolean operator ‘or,’ and categories of search terms were linked using the Boolean operator ‘and,’ resulting in capture of manuscripts that included at least one ‘Respiratory,’ ‘Anxiety,’ and ‘Treatment’ search term. We used PsycInfo, Medline, and PubMed search engines to identify potentially relevant studies (date range: through November, 2019). We additionally selectively searched for studies by authors who have published substantially in the area of respiration and psychopathology and cross-checked reference sections of identified articles, with an emphasis on review and theoretical articles. The initial literature search was independently conducted by MV and MJY, yielding identical results. Screening of manuscripts for inclusion in the analyses was completed by MV, MJY, and HB, following training and supervision by TML and PL.

Table 3.

Characteristics of Randomized Controlled Trials of Respiratory Interventions for the Treatment of Anxiety (N=40)

Author(s), year	Sample Characteristics	Intervention	Control Comparison(s)	Sessions		Treatment Duration (weeks)	Followup (weeks)	Anxiety Measures	Anxiety as Primary Study Outcome?
				Number	Duration (min)
Berger, 2000	N=21; PD; Outpatient	Diaphragmatic Breathing (single)	CBTa	6	30–60	6	4	PDSS; PA frequency	Yes
Bonn et al., 1984	N=12; Agoraphobia; Outpatient	Diaphragmatic Breathing with Exposure (combined)	Exposurea	2	60	2	4 & 24	PA frequency; Global Phobia Score; Agoraphobia Score	Yes
Bratton, 2008	N=20;Pregnant Women; Outpatient	Diaphragmatic Breathing (single)	Mindfulness meditationa	1	90	1	N/A	STAI	Yes
Chandl a et al., 2013	N=96; Medical Students; Non-Clinical	DiaphragmaticBreathing1 (single)	Yoga Asanaa	42	60	6	N/A	HAM-A	Yes
Chang et al., 2020	N=35; Clinical; Acute Ischemic Stroke Patients	HRV Biofeedback (single)	Usual Carei	4	20	1	4 & 12	HADS	Yes
Chen et al., 2017	N=30; Clinical; Outpatient	Diaphragmatic Breathing (single)	Breathing Relaxation Practice with Biofeedback Monitoring a	12	30	8	N/A	BAI	Yes
Craske et al., 1997	N=38; PD; Outpatient	Diaphragmatic Breathing w/ cognitive restructuring & in-vivo exposure (combined)	Interoceptive exposure w/ cognitive restructurin g, & invivo exposurea	12	90-120	12	N/A	ADIS-R HAM-A; Situational Fear Question naire	Yes
Colgan et al., 2016	N=102; PTSD; Outpatient	HRV Biofeedback 2 (single)	Body Scan Meditation a; Mindful Breathinga; Sitting Quietlyi	6	20	6	N/A	PCL_C	Yes
Dhruv a et al., 2012	N=16; Chemotherapy patients; Outpatient	Diaphragmatic Breathing1 (single)	Waitlisti	7	60	7	N/A	HADS	Yes
Enrigh t et al., 2004	N=19; Cystic Fibrosis; Outpatient	High Effort Inspiratory Muscle Training (single)	Low Effort Inspiratory Muscle Traininga; No Treatmenti	24	45	8	N/A	HADS	No
Fasko et al., 1992	N=10; Undergrad uate; Non- Clinical	Controlled Breathing3 (single)	Saying ‘ 1’ upon exhalationa	1	30	1	N/A	STAI	Yes
Gatche l & Procto r, 1976	N=36; Undergrad uate; Non- clinical	HR Biofeedback (single)	Light trackingp	2	20	2	N/A	Personal Report of Confide nce as a Speaker Self Reported Anxiety	Yes
Gruzel ier et al., 2014	N=32; Undergrad uate; Non- Clinical	HRV Biofeedback (single)	Alphatheta neurofeedbacka; Choreology Instructionp; No Treatmenti	10	20	10	N/A	DASS	Yes
Hayam a & Inoue, 2012	N=23; Gynecological cancer; Outpatients	Diaphragmatic Breathing (single)	Chemother apy and Nursing Carei	4	10	2	N/A	POMS	Yes
Hakke d et al., 2017	N=27; Competitive Swimmers; Non-Clinical	Yogic Breathing (single)	Waitlisti	20	30	4	N/A	SAS-2	Yes
Henriq ues et al., 2011	N=30; Undergrad uate; Non- Clinical	HRV Biofeedback (single)	Waitlisti	4	15	20	N/A	MASQ	Yes
Hibber t & Chan, 1989	N=42; PA; Outpatient	Paced Breathing (single)	Obesrving physical anxiety symptomsp	5	30–60	6	N/A	HAM-A	Yes
Lee et al., 2015	N=10; Undergrad uate; Non- Clinical	HRV Biofeedback (single)	No Treatmenti	4	45	8	N/A	STAI	Yes
Lin, 2018	N=82; Online recruited healthy adults; Non-Clinical	HRV Biofeedback (single)	Relaxation Training Groupa; No Treatment i	4	60	4	N/A	BAI	Yes
Lin et al., 2019	N=48; MDD; Outpatient	HRV Biofeedback (single)	Waitlisti	6	60	6	N/A	BAI	Yes
Lickel, 2010	N=57 Undergrad uate; Non- Clinical	Diaphragmatic Breathing with Psychoeducation (combined)	Psychoeducationa	1	Not specified	1	N/A	BAI	Yes
Meuret et al., 2008	N=37 PD; Community sample	ETCO2 Biofeedback (single)	Waitlisti	5	60	4	N/A	PDSS	Yes
Meuret et al., 2010	N=41 PD; Outpatient	ETCO2 Biofeedback (single)	Cognitive Skill Traininga	5	60	4	N/A	PDSS	Yes
Paul & Garg, 2012	N=30 Basketball Players; Non-Clinical	HRV Biofeedback (single)	viewing motivation al filmp; No Treatmenti	10	20	1.5	9.5	STAI	Yes
Penzli n et al., 2015	N=48 AUD; Inpatient	HRV Biofeedback +CBT, MI, and Occupationa lTherapy (single)	CBT, MI, and Occupation al Therapya	6	20	2	5 & 8	Symptom Checklist-90-Revised	No
Pham et al., 2016	N=63 Online recruited adults; Non-Clinical	Diaphragmatic Breathing (single)	Waitlisti	Variable	Atleast 1	4	N/A	GAD-7; PDSS	No
Smith et al., 2019	N=169; Technology Service Employees; Non-Clinical	Mindful Breathing (single)	Waitlisti	Variable	Atleast 1	4	N/A	MASQ; CDC Healthy Days-Days Anxious	Yes
Stefanaki et al., 2015	N=46 PCOS; Outpatient	Diaphragmatic Breathing (single)	No Treatmenti	9	30	8	N/A	DASS-21	Yes
Tan et al., 2011	N=20; PTSD; Outpatient	HRV Biofeedback (single)	Treatment as Usuala	8	30	8	N/A	PCL-S; CAPS	Yes
Thoma s et al., 2009	N=183; Asthma; Outpatient	Diaphragmatic Breathing (single)	Psychoeducationa	3	60	7	4 & 24	HADS	No
Thurbe r, 2007	N=14; Undergrad uate; Non- Clinical	HRV Biofeedback (single)	No Treatment i	4	40	4	N/A	STAI	Yes
Valenz a et al., 2014	N=46; COPD; Inpatient	Controlled Breathing (single)	Standard MedicalCarea	20	30	2	N/A	HADS	Yes
van der Zwan et al., 2015	N=75; Community Sample; Non- Clinical	HRV Biofeedback (single)	Mindfulness Meditation a Physical Activitya	5	20	1	6	DASS	Yes
van der Zwan et al., 2019	N=50; Pregnant and NonPregnant women; Outpatient	HRV Biofeedback (single)	Waitlisti	5	60–90	5	5 & 11	DASS	Yes
Van Dixho orn et al., 1990	N=137; Cardiac Patients; Outpatient	EMG Biofeedback for Inspiratory control+Physical Activity (single)	Physical Activitya	6	60	6	N/A	STAI	Yes
Wahbeh et al., 2016	N=102; PTSD; Combat Veterans	Slow Breathing2 (single) Slow Breathing2 & Mindfulness Meditation (combined)	Mindfulness Meditation a; Sitting Quietlyi	6	20	6	N/A	PCL-C	Yes
Wang et al., 2017	N=55; COPD; Outpatient	Inspiratory Muscle Training + Cycle Ergometer Training (combined)	Cycle Ergometer Traininga	24	44	8	N/A	HADS	Yes
White, 2008	N=36; PTSD; Inpatient	HRV Biofeedback (single)	Progressiv e Muscle Relaxationa	5	20	5	N/A	DAPS	Yes
Yekta et al., 2017	N=60; Clinical; Mastecto my Surgery Patients	Rhythmic Breathing (single)	Benson’s Relaxation Techniquea; No Treatmenti	2	40	1	N/A	CSAQ	Yes
Zucker et al., 2009	N=38; PTSD; Inpatient	HRV Biofeedback (single)	Progressiv e Muscle Relaxationa	28	20	4	N/A	DAPS; PCL-C	Yes

Note. N/A=Not Available; AUD=Alcohol Use Disorder; COPD=Chronic Obstructive Pulmonary disease; MDD= Major Depressive Disorder; MI=Motivational Interviewing; PA=Panic Attack; PD=Panic Disorder; PCOS=Polycystic Ovary Syndrome; PTSD=Posttraumatic Stress Disorder; CBT=Cognitive Behavioral Therapy; HRV=Heart Rate Variability; BAI (Beck Anxiety Index; (Beck, Epstein, Brown, & Steer, 1988)); CAPS (Clinician-Administered PTSD Scale: (Weathers & Litz, 1994)); CDC Healthy Days-Days Anxious (Center for Disease Control’s Healthy Days Core and Symptoms Modules (CDC HRQOL-14; Centers for Disease Control and Prevention, 2000); CSAQ (Cognitive-Somatic Trait Anxiety Questionnaire: (Schwartz, Davidson, & Goleman, 1978); DAPS (Detailed Assessment of Posttraumatic Stress; (Briere, 2001)); DASS (Depression Anxiety Stress Scale: (Lovibond & Lovibond, 1995); DASS-21 (Depression Anxiety Stress Scale: (Henry & Crawford, 2005)); GAD-7 (Generalized Anxiety Disorder Scale;(Spitzer, Kroenke, Williams, & Löwe, 2006)); HADS= (Hospital Anxiety and Depression Scale, all studies utilized anxiety subscale; (Zigmond & Snaith, 1983)); HAM-A (Hamilton Anxiety Scale: (Hamilton, 1960)); MASQ (Mood and Anxiety Symptom Questionnaire, all studies utilized arousal and distress scale; (Keogh & Reidy, 2000)); PCL-C or S (PTSD Checklist-Civilian or Specific: (Tan et al., 2011)); PDSS (Panic Disorder Severity Scale; (Shear et al., 2001)); POMS (Profile of Mood State; (McNair, Lorr, & Droppleman, 1971)); STAI (State-Trait Anxiety Inventory: (Spielberg, Gorsuch, & Lushene, 1983)); SAS-2 (Sport Anxiety Scale-2: (R. E. Smith, Smoll, Cumming, & Grossbard, 2006)

^aActive control comparison

ⁱInactive control comparison

^PPlacebo control comparison

¹Pranayama Yoga

²RESPeRATE

³Comeditation

Inclusion/Exclusion Criteria for this Study

Primary inclusion criteria required that manuscripts: (1) included a treatment with at least one component targeting respiration; (2) employed a control comparison group that allowed for isolation of the respiratory intervention component; and (3) included an anxiety index as an outcome. In addition, while there were no explicit inclusion restrictions regarding the study design employed, included studies must have indexed anxiety symptoms after the administration of a respiratory intervention. As such, studies that reported pre- and post-intervention trait anxiety symptoms, pre- to post-intervention change in trait anxiety symptoms, or post- intervention trait anxiety symptoms were eligible for inclusion. All manuscripts were required to be written in English due to author language limitations. Lastly, no inclusion restriction was placed based on the year of publication of a manuscript. To minimize the ‘file drawer effect’, whereby studies without significant findings often remain unpublished, we also screened and included unpublished dissertations, when relevant.

We screened 7,518 unique abstracts to determine initial eligibility. Following initial abstract review, 995 manuscripts were selected for full-text examination. Articles whose abstracts did not provide adequate information to determine if they met the primary inclusion criteria were also retained for full-text review (see Figure 1 for the consort chart of screened articles). Next, full-text articles were retrieved and independently read and coded by MJY, MV, and HB. Where initial reviewers disagreed or were unclear about inclusion of particular studies, the decision was made by TML or PL. We then determined whether studies provided sufficient information to compute effect size estimates (see Data Synthesis). Where studies did not provide this information, authors were directly contacted.

Figure 1.

Consort Chart of Screened Articles

Next, the remaining articles (n=49) were coded for inclusion of appropriate anxiety measurement, composition of experimental and control group and interventions, duration of treatment and assessment, and presentation of data in a manner interpretable for meta-analysis. In total, n=40 studies were identified as meeting our inclusion criteria and provided sufficient data for analyses.

In order to obtain the largest possible pool of studies we included some studies with methodological flaws, such as failing to blind of statistical analysts, lack of behavioral placebo conditions, or absence of double blinding approximations. To partially compensate for this, we included all anxiety measures included in each study, whether or not they were identified as primary outcome measures.

Assessment of Study Quality

Risk of Bias (RoB) of assignment to intervention (i.e., the ‘intention-to-treat, effect) was assessed based on the Cochrane Handbook (Higgins et al., 2019), and supported using available tools (e.g., signaling questions and algorithm; Risk of bias, 2020). RoB as ‘high,’ ‘low,’ or having ‘some concerns’ was assessed across five domains: (1) Domain 1: Bias arising from the randomization process; (2) Domain 2: Bias due to deviations from intended intervention; (3) Domain 3: Bias due to missing outcome data; (4) Domain 4: Bias in measurement of the outcome; and (5) Domain 5: Bias in selection of the reported result. First, raters (HRB, DLH, MJY) were trained (TML) on how to classify studies across the five domains indicated. Next, they completed independent ratings of 12.5% (k=5) studies. Discrepancies were then discussed with TML, adjustments were made to the tool, and ratings were again completed independently, to ensure rater consensus. The remaining studies were randomly assigned for independent review by the three raters with stratification for dissertation studies (i.e., gray literature); a total of 20% (k=8) of all the studies were duplicate reviewed by two raters. Minor discrepancies in RoB for duplicate reviews were noted, consensus was reached, and ratings were adjusted independently according to that consensus. If relevant, conclusions reached during consensus discussions were re-applied to all studies to ensure consistency. When rating domains, sub-criteria are rated as ‘Yes,’ ‘Probably Yes,’ ‘No,’ ‘Probably No,’ and ‘No Information.’ Given the Cochrane criteria were developed with pharmaceutical trials in mind, we found several criteria within domains to be overly conservative for the purposes of the current review. Therefore, several adjustments to the signaling questions and tool algorithm were made following initial discrepancy discussions. First, we opted to remove Domain 2 from our assessment, which is informed by blinding of participant (Criterion 2.1) and interventionist (Criterion 2.2); an impossibility based on the behavioral basis of interventions reviewed such that all studies would have been rated as high RoB in this domain, thus resulting in a high overall RoB assessment. Also, the majority of retained studies were not pre-registered (k=39), which typically results in a ‘some concerns’ RoB rating for Domain 5. To account for this, in the case of no pre-registration of an analytic plan, we instead opted to code RoB in Domain 5 based upon match between the proposed analytic plan and reported results within the manuscript.

Data Synthesis

Data were analyzed using Comprehensive Meta-Analysis, Version 3, using Hedges’ g as a correction to Cohen’s d (Hedges & Olkin, 1985). Hedges’ g values of 0.0 to 0.3 are interpreted as no or minimal difference between the respiratory condition and control conditions, while 0.3 to 0.5, 0.5 to 0.7, and > 0.7 as small to medium, medium to large, and large to very large effect sizes, respectively (Cohen, 1988). Because the studies we reviewed are independent, we used a random effects model (Borenstein et al., 2017; Hedges & Vevea, 1998) and report pooled mean effect size estimates, allowing us to make population-based inferences (Berkeljon & Baldwin, 2009). We additionally examined the influence of outlying studies that may either reflect bias or the presence of unique population or procedural variables, by examining a funnel plot, which allows for greater between-study variability among studies by accounting for sample size. In studies with smaller n’s where the influence of sampling error is expected to be greater (Sterne & Harbord, 2004), we inspected them for unique procedural or population characteristics.

Several indices were used to assess for heterogeneity. We report Cochran’s Q-statistic (i.e., the weighted sum of squared differences between individual study effects and the pooled weighted effect across studies) based on the n for all studies, where Q is distributed as χ² with k (i.e., number of studies), minus 1 degree of freedom. Here, significant values suggest we reject the null hypothesis of homogeneity, indicating dispersion in treatment effects among various populations. However the Q-statistic is highly sensitive to number of studies and to the cumulative number of participants, so a significant Q with a large sample size, as in the current review, can yield statistical significance for inconsequential heterogeneity (Higgins et al., 2003). We additionally interpret the I² statistic, which estimates the proportion of the variance reflecting true effect size differences (i.e., the percentage of the variance that would remain if it were possible to remove all within-study sampling error; Borenstein et al., 2017). Finally, we use tau-(τ) and tau-squared to estimate dispersion of true effect sizes between studies (Borenstein et al., 2017). Tau-squared is the variance of effect size parameters across the population of studies (i.e., variance in true effect sizes). Tau represents the standard deviation of true effects across all studies. Tau can also be used to calculate a 95% prediction interval, wherein 95% of new studies would expect to yield g. Because the actual population effect size and variance are unknown a priori, we assumed a non-normal distribution and used an adjusted calculation for the corresponding prediction interval (Borenstein et al., 2017).

Where multiple anxiety outcomes were reported (n=9), an equal weight was assigned to each measure, calculated by dividing the Hedges’ g for each measure by the number of measures used (Rosenthal & Rubin, 1986). This ensured that each sample contributed just one overall anxiety effect size in each study, weighted by number of participants. Where there was more than one ‘control’ comparison (i.e., active and inactive controls; n=7), we averaged effects across control conditions, and later, used both separate meta-analyses and moderator analysis to characterize differences in effect size estimates as a function of treatment comparison. Where available, we used mean and standard deviation values to calculate Hedges’ g effect size estimates, weighted by sample size. When these values were not available (n=8), we calculated g from change scores and SDs (n=1), or from F-, t- or p-values to estimate effect sizes (n=7). Where studies reported both post-test and follow-up data, we averaged the two in our initial analysis. This was done, primarily, because the term “follow-up” has different meanings among studies, where follow-up time in one study sometimes was shorter than post-test time in other studies. Separate analyses were also conducted on post-test or follow-up data separately.

We separately analyzed the effects of respiratory interventions on dimensional anxiety symptom categories (e.g., panic, general anxiety, and post-traumatic stress) population (i.e., non-psychiatric/medical; psychiatric), and diagnostic category of the sample treated. For dimensional anxiety symptoms, we looked separately at inactive control comparisons to evaluate the independent effectiveness of respiratory interventions. Given respiratory biofeedback interventions provide highly specific training using objective physiological assessment of breathing relevant parameters, we used mixed-effects analyses (i.e., random effects) within subgroups to estimate and compare effect sizes of interventions including a biofeedback component and non-biofeedback interventions. Mixed-effects analyses were also used to compare active versus inactive control effects. Random effects meta regression analyses, using weighted least squares to account for variability in sample size, were used to investigate several moderator effects on overall effect size estimates (Borenstein et al., 2017; Stanley & Doucouliagos, 2012; Stanley & Jarrell, 1989). These included treatment duration and number of treatment sessions. Finally, year of publication was included in moderator analyses to estimate bias that may occur in the experimental and peer review processes over time. In all studies, we coded Hedges’ g in the negative direction indicating greater decreases in scores on anxiety scales in the respiratory treatment than in the control groups, or greater increases in scores on tests reflecting freedom from anxiety. For our primary analysis, however, we calculated the combined effect size for all comparisons.

Results

Systematic Review

The final sample of k=40 studies had an aggregated n of 3,047 participants, with treatment group sizes ranging from five to 92. The most common anxiety outcomes were the State Trait Anxiety Index (n=6), Hospital Anxiety and Depression Scale (n=6) Beck Anxiety Inventory (n=4), Panic Disorder Severity Scale (n=4), and the Depression Anxiety and Stress Scale (n=4). There was great variability in the populations examined, with k=10 studies examining a psychiatric population, k=7 focused on various physical conditions, including chronic obstructive pulmonary disease (Valenza et al., 2014), polycystic ovary syndrome (Stefanaki et al., 2015), pregnancy (Bratton, 2008), and chemotherapy patients (Dhruva et al., 2012), and k=12 on various non-psychiatric populations (Table 3), including participants with specific fears (e.g., fear of public speaking; Gatchel & Proctor, 1976), and nonclinical samples with elevated scores on various anxiety inventories, (e.g., Pham et al., 2016) that did not necessarily meet for a clinical diagnosis or were not diagnostically assessed.

The majority of active interventions examined the effects of diaphragmatic breathing (k=12) and heart rate variability (HRV) biofeedback (k=16) on anxiety outcomes. The remaining (k=12) interventions examined breathing modification techniques that did not specify diaphragmatic techniques or employed techniques to specifically affect other physiological processes. These included ETCO₂ biofeedback, EMG biofeedback of the frontalis muscle, inspiratory muscle training, paced breathing at frequencies differing from those in HRV biofeedback, deep breathing, slow breathing, slow breathing, and co-meditation² (Table 3).

Risk of Bias Findings

Our RoB assessment using our adjusted algorithm indicated k=35 studies classified as ‘high’ with the remaining k=5 classified as ‘some concerns’ (Table 4). Here, it is notable that k=33 studies were coded as ‘high’ risk in Domain 4, Outcome Measurement. In this domain k=16 included an inactive control comparison condition and k=4 including active comparator conditions that were not adequately matched to the active condition. For k=20 studies, this was the only domain coded as ‘high’ risk. Another k=6 studies were coded as ‘some concerns’ in Domain 4 due to lack of explicit details regarding how the conditions were presented to participants (e.g., neither inferior to the other). Together this indicates that a major limitation for extant studies of respiratory interventions is the failure to use an appropriate control comparison condition. We also noted that just k=1 study (Valenza et al., 2014) included a pre-registered analytic plan, suggesting that without editing the algorithm used for Domain 5, Selection of Reported Result, the majority would have been classified as ‘some concerns’ bias risk. Our assessment underscores the importance of considering control comparison type in between-group effects and mixed-effects analyses.

Table 4.

Risk of Bias Assessments by Domain and Overall

Author(s), year	Domain 1: Randomization Process	Domain 3: Missing Outcome Data	Domain 4: Outcome Measurement	Domain 5: Selection of Reported Result	Overall
Berger, 2000	S	S	H	L*	H
Bonn et al., 1984	H	L	H	L*	H
Bratton, 2008	H	H	H	L*	H
Chandla et al., 2013	S	L	S	L*	S
Chang et al., 2020	L	S	H	L*	H
Chen et al., 2017	S	H	H	L*	H
Craske et al., 1997	S	S	S	S*	S
Colgan et al., 2016	H	L	H	L*	H
Dhruva et al., 2012	L	H	H	L*	H
Enright et al., 2004	S	L	H	L*	H
Fasko et al., 1992	S	L	S	L*	S
Gatchel & Proctor, 1976	S	S	H	L*	H
Gruzelier et al., 2014	S	S	H	L*	H
Hayama & Inoue, 2012	S	S	H	L*	H
Hakked et al., 2017	S	H	H	L*	H
Henriques et al., 2011	S	L	H	L*	H
Hibbert & Chan, 1989	S	L	L	L*	S
Lee et al., 2015	S	L	H	L*	H
Lin, 2018	S	H	H	L*	H
Lin et al., 2019	H	H	H	L*	H
Lickel, 2010	S	S	H	L*	H
Meuret et al., 2008	S	S	H	L*	H
Meuret et al., 2010	S	L	S	L*	S
Paul & Garg, 2012	S	L	H	L*	H
Penzlin et al., 2015	S	L	H	H*	H
Pham et al., 2016	S	L	H	L*	H
Smith et al., 2019	S	H	H	L*	H
Stefanaki et al., 2015	L	L	H	L*	H
Tan et al., 2011	S	L	H	L*	H
Thomas et al., 2009	S	H	H	L*	H
Thurber, 2007	H	H	H	L*	H
Valenza et al., 2014	L	L	H	L	H
van der Zwan et al., 2015	L	H	S	L*	H
van der Zwan et al., 2019	S	L	H	L*	H
Van Dixhoorn et al., 1990	S	S	H	L*	H
Wahbeh et al., 2016	L	S	H	L*	H
Wang et al., 2017	L	S	H	L*	H
White, 2008	H	L	S	L*	H
Yekta et al., 2017	S	L	H	L*	H
Zucker et al., 2009	H	S	L	L*	H

Note. Risk of Bias as assessed using the Cochrane Risk of Bias Tool; Domain 2 was removed as the impossibility of blinding in behavioral interventions would have resulted in every study being ranked as High risk of bias in this domain; H, high risk of bias; S, some concerns regarding risk of bias; L, low risk of bias.

^*Trial did not have a pre-registered analytic plan, thus the assessor’s judgment for this domain was changed from the suggested algorithm assessment of “some concerns” to an assessment based on what the algorithm would display if signaling questions were answered based upon match to the analytic plan as specified within the manuscript.

Meta-analysis

Overall, averaging post-treatment and follow-up ratings of anxiety across both active and inactive control conditions, therapies that included a respiratory component yielded significantly greater decreases in anxiety symptoms compared to controls, indicating a medium to large symptom improvement effect (Hedges’ g= −.678, 95% CI = −0.932, −0.424, p<.001) (see Table 5 and Figure 2). Significant heterogeneity suggests a large proportion of variance is estimated to reflect true variance rather than sample error (I²). Inspection of the funnel plot (Figure 3) identified Chandla et al. (2013) as an outlier. This trial exhibited unique characteristics wherein medical students were randomized to practice either slow paced pranayamic breathing or yoga asanas, led by expert instructors, each morning for about 40 minutes over the course of 6 weeks, totaling 42 sessions. Conversely, the treatment dose for all other studies included ranged from 1 to 28 sessions (Table 3). Due to its outlier status and unique study characteristics, Chandla et al. (2013) was removed from subsequent analyses.

Table 5.

Within-Group Respiratory Intervention Effect Sizes

Author(s) (Year)	Hedge’s g	Standard Error	Variance	Lower Limit	Upper Limit	Z-Value	P Value
Berger (2000)	−0.151	0.423	0.179	−0.979	0.678	−0.356	0.722
Bonn et al. (1984)	−2.880	0.939	0.881	−4.720	−1.040	−3.067	0.002
Bratton (2008)	−0.330	0.431	0.186	−1.176	0.515	−0.766	0.444
Chandla et al. (2013)	−4.487	0.382	0.146	−5.236	−3.739	−11.749	<.001
Chang et al. (2020)	−0.802	0.345	0.119	−1.478	−0.125	−2.322	0.020
Chen et al. (2017)	−1.538	0.407	0.166	−2.335	−0.740	−3.778	<.001
Colgan et al. (2016)	0.342	0.279	0.078	−0.204	0.889	1.228	0.219
Craske et al., 1997	0.842	0.338	0.114	0.179	1.504	2.490	0.013
Dhruva et al. (2012)	−0.495	0.481	0.231	−1.438	0.447	−1.030	0.303
Enright et al. (2004)	−0.684	0.453	0.205	−1.572	0.204	−1.510	0.131
Fasko et al. (1992)	0.491	0.582	0.338	−0.649	1.631	0.845	0.398
Gatchel et al. (1976)	−1.415	0.366	0.134	−2.132	−0.697	−3.865	<.001
Gruzelier et al. (2014)	−0.583	0.353	0.125	−1.276	0.109	−1.651	0.099
Hakked et al. (2017)	−0.468	0.379	0.144	−1.211	0.274	−1.236	0.216
Hayama et al. (2012)	−0.756	0.417	0.174	−1.574	0.062	−1.811	0.070
Henriques et al. (2011)	−1.523	0.407	0.166	−2.322	−0.725	−3.738	<.001
Hibbert et al. (1989)	−0.678	0.312	0.097	−1.289	−0.067	−2.174	0.030
Lee et al. (2015)	−1.917	0.714	0.510	−3.316	−0.517	−2.684	0.007
Lickel (2010)	0.075	0.277	0.077	−0.468	0.618	0.270	0.787
Lin (2018)	−0.319	0.269	0.072	−0.846	0.208	−1.185	0.236
Lin et al. (2019)	−0.834	0.296	0.088	−1.415	−0.253	−2.814	0.005
Meuret et al. (2008)	−2.475	0.432	0.187	−3.322	−1.627	−5.724	<.001
Meuret et al. (2010)	0.201	0.307	0.094	−0.401	0.803	0.654	0.513
Parsa-Yekta et al. (2017)	−0.681	0.265	0.070	−1.201	−0.161	−2.566	0.010
Paul et al. (2012)	−3.175	0.662	0.438	−4.472	−1.878	−4.798	<.001
Penzlin et al. (2015)	−0.595	0.291	0.085	−1.165	−0.025	−2.046	0.041
Pham et al. (2016)	−0.110	0.249	0.062	−0.598	0.379	−0.440	0.660
Smith et al. (2019)	−0.254	0.157	0.025	−0.562	0.054	−1.614	0.107
Stefanaki et al. (2015)	−1.417	0.363	0.132	−2.129	−0.705	−3.900	<.001
Tan et al. (2011)	−0.552	0.439	0.193	−1.412	0.308	−1.257	0.209
Thomas et al. (2009)	−0.215	0.148	0.022	−0.504	0.075	−1.452	0.147
Thurber (2007)	−0.146	0.501	0.251	−1.128	0.837	−0.290	0.771
Valenza et al. (2014)	−0.979	0.307	0.094	−1.582	−0.377	−3.187	0.001
van der Zwan et al. (2015)	0.318	0.281	0.079	−0.233	0.869	1.132	0.258
van der Zwan et al. (2019)	−0.244	0.284	0.080	−0.800	0.312	−0.860	0.390
Van Dixhoorn et al. (1990)	−0.071	0.170	0.029	−0.404	0.262	−0.418	0.676
Wahbeh et al. (2016)	−0.312	0.278	0.077	−0.856	0.232	−1.124	0.261
Wang et al. (2017)	−1.030	0.289	0.083	−1.595	−0.464	−3.568	<.001
White (2008)	−0.064	0.328	0.108	−0.707	0.579	−0.194	0.846
Zucker et al. (2009)	−0.247	0.319	0.102	−0.873	0.379	−0.774	0.439
Average Effect Size	−0.678	0.129	0.017	−0.932	−0.424	−5.234	<.001

Figure 2.

Forest Plot of Breathing Interventions Pre- to Post-Treatment Change in Anxiety (n=40)

Note. Negative values of Hedges’ g indicate better treatment outcome.

Figure 3.

Funnel Plot of Standard

Error by Hedges’ g for Measures of Anxiety (n=40)

Upon removal of Chandla et al. (2013), we still found a medium effect size in favor of respiratory interventions compared with control interventions (Hedges’ g= −0.550, 95% CI = −0.747, −0.352, p<.001). Heterogeneity also remained significant, with the majority of variance (reflected in I²) reflecting true variance. Borenstein’s (2017) correction of τ resulted in a prediction interval of −1.613 to 0.513 suggesting that another study with a comparable population and method would have a 95% chance of falling between a very large effect size favoring respiratory therapies and a medium effect size favoring a control condition (see Table 6). To characterize whether therapeutic effects were maintained over time, we also examined effect size estimates at follow-up time points, where data were available (k=8) and found a medium to large effect (Hedges’ g= −0.710, 95% CI = −1.251, −0.168, p<.001)). Examination of post-test data (k=37) only, yielded a significant medium effect (Hedges’ g= −0.535, 95% CI = −0.740, −0.331, p<.001).

Table 6.

Main Study Outcomes

	k	Hedge s’ g	SE	Variance	95%CI	95%CI	Test of Null	p-value	Q-value	I2	Prediction Intervals
					LL	UL	Z
All studies	40	−0.678	0.129	0.017	−0.932	−0.424	−5.234	<.001	257.206	84.837	−2.176 to 0.820
Removing Chandla et al. (2003)	39	−0.550	0.101	0.010	−0.747	−0.352	−5.458	<.001	144.9231	73.779	−1.613 to 0.513
Follow-Up Data Only	8	−0.710	0.276	0.076	−1.251	−0.168	−2.570	.010	39.245	82.163	−2.472 to 1.052
Post-Test Data Only	37	−0.535	0.104	0.011	−0.740	−0.331	−5.128	<.001	139.971	74.280	−1.613 to 0.543
Anxiety Symptom Category (dimensional)
Panic Symptoms	6	−0.784	0.555	0.308	−1.872	0.303	−1.414	.157	61.285	91.841	−4.586 to 3.018
Inactive control comparis ons only	2	−1.220	1.229	1.511	−3.629	1.189	−0.993	.321	24.292	95.883	n/a
General Anxiety Symptoms	28	−0.521	0.099	0.010	−0.715	−0.328	−5.279	<.001	76.681	64.789	−1.3603 to 0.407
Inactive control comparis ons only	20	−0.668	0.121	0.015	−0.905	−0.432	−5.544	<.001	54.884	65.382	−1.566 to 0.318
Posttraumatic Stress Symptoms	5	−0.114	0.148	0.022	−0.405	0.177	−0.767	.443	4.378	8.644	−0.675 to 0.447
Sample Characteristics
Nonpsychiatric and Medical	26	−0.572	0.108	0.012	−0.784	−0.359	−5.279	<.001	78.058	67.973	−1.483 to 0.339
Psychiatric sample	13	−0.513	0.230	0.053	−0.963	−0.062	−2.229	0.026	66.432	81.936	−2.208 to 1.182
Panic disorder, only	4	−0.375	0.664	0.441	−1.677	0.927	−0.565	0.572	38.823	92.273	−6.555 to 5.805

Note. k = Number of Studies, SE=Standard Error, CI=Confidence Interval, LL=Lower Limit, UL=Upper Limit

Dimensional Anxiety Symptoms

Between-group effect size estimates on dimensional measures of panic symptoms (k=6) were large, but did not reach statistical significance, with significant heterogeneity (Hedges’ g= −0.784, 95% CI = −1.872, 0.303, p>.05). Between-group effect size estimates on general anxiety symptoms (k=28) were significant and medium (Hedges’ g= −0.521, 95% CI = −0.715, −0.328, p<.001), with a prediction interval of −1.360 to 0.407, suggesting that another study with comparable population and method would have a 95% chance of falling between a very large effect size in favor of respiratory therapies and a small effect size in favor of control interventions (Borenstein et al., 2017). The same approach was used for general anxiety symptoms among studies with inactive control conditions (k=20), only, and yielded a medium to large effect size estimate (Hedges’ g= −0.668, 95% CI = −0.905, −0.432, p<.001). and a prediction interval of −1.566 to 0.318 suggesting that another study with comparable population and method would have a 95% chance of falling between a very large effect size favoring respiratory therapies and a small effect size favoring a an inactive control condition (Borenstein et al., 2017). Effect size estimates for posttraumatic stress symptoms were small and non-significant (Hedges’ g=−0.114, 95% CI = −0.405, 0.177, p>.05). All analyses evidenced significant heterogeneity, with a large portion of variance reflecting true variability in effect sizes (Table 6).

Analysis by Diagnostic Category

Between-group effect size estimates on anxiety outcomes in non-psychiatric and medical populations (k=26) were medium to large and significant (Hedges’ g= −0.572, 95% CI = −0.784, −0.359, p<.001). with a prediction interval of −1.483 to 0.339 suggesting that another study with a comparable population and method would have a 95% chance of falling between a very large effect size favoring respiratory therapies and a small effect size favoring a control condition. We also examined effect size estimates for anxiety outcomes in studies with psychiatric participants (k=13). Here, we found between-group effect size estimates that were medium and significant (Hedges’ g= −0.513, 95% CI = −0.963, 0.062, p<.05) with a prediction interval of −2.208 to 1.182 suggesting that another study with comparable population and method would have a 95% chance of falling between a large effect size favoring respiratory therapies and a large effect size favoring a control condition. A more granular examination of intervention effects in specific diagnostic groups yielded a small to medium non-significant, pooled effect size estimate for panic disorder (k=4; Hedges’ g= −0.375, 95% CI = −1.677, 0.927, p>.05). We did not examine effects in individuals diagnosed with PTSD (k=4) given redundancy with aforementioned effect size estimates for posttraumatic stress symptoms. All analyses evidenced significant heterogeneity, with a large portion of variance reflecting true variability in effect sizes (Table 6).

Mixed-effects and Moderator Analyses: Treatment modality, control comparison, intensity, duration, publication year

Results from mixed-effects and moderator analyses can be found in Tables 7 and 8, respectively. We identified 13 distinct types of respiratory interventions, but the small number of studies for each intervention precluded examination of mean effects within treatment type or examination of effects on overall Hedges’ g via meta-regression. Instead, mixed-effects models were used to estimate and compare random effects within interventions that employed a biofeedback component (k = 19) versus those that did not (k = 20). We first examined this in trials with inactive control comparisons (k=25), which yielded a large effect size for biofeedback studies (B=−0.974, 95% CI = −1.428, −0.520, p<.001) and a medium effect size for other studies (B=−0.476, 95% CI = −0.711, −0.240, p<.05; see Table 7). In this inactive control comparisons analysis, differences between biofeedback and non-biofeedback interventions were marginally significant, suggesting the overall effect of respiratory interventions on anxiety was related to intervention modality (Q(1)=3.648, p=.056). This same analysis, including all available studies (i.e., inactive and active control conditions; k=39), again yielded a medium to large effect size for biofeedback interventions (B=−0.631, 95% CI = −0.961, −0.301, p<.001) and a medium effect size for other studies (B=−0.489, 95% CI = −0.731, −0.247, p<.001). Here, differences between biofeedback and non-biofeedback studies were not significant (Q(1)=0.462, p=.497). Notably, in both the inactive control analysis and analysis with all available studies, significant unexplained variance in true effects (I²) within each subgroup was observed, suggesting most of the within-subgroup variance reflecting true differences in study effects (see Table 7).

Table 7.

Mixed-Effects Subgroup Analyses of Dichotomous Moderators

Description	k	B	SE	95%CI	95%CI	Z	p-value	Between Study Heterogeneity
				LL	UL			Q	df	p-value	I2
Intervention Type
Biofeedback	19	−0.631	0.168	−0.961	−0.301	−3.749	<.001	87.823	18	<.001	79.504
Non-Biofeedback	20	−0.489	0.123	−0.731	−0.247	−3.966	<.001	57.088	19	<.001	66.718
Combined between- group effects								0.462	1	0.497
Biofeedback (Inactive control comparisons only)	13	−0.974	0.232	−1.428	−0.520	−4.205	<.001	59.068	12	<.001	79.684
Non-Biofeedback (Inactive control comparisons only)	12	−0.476	0.120	−0.711	−0.240	−3.969	<.001	23.823	11	0.013	53.826
Combined between- group effects								3.648	1	0.056
Control Type3
Active	25	−0.279	.132	−0.537	−0.021	−2.118	.034	55.595	19	<.001	66.047
Inactive	20	−0.709	.127	−0.958	−0.460	−5.586	<.001	85.433	24	<.001	71.908
Combined between-group effects								5.526	1	0.019

Note. k = Number of Studies, SE=Standard Error, CI=Confidence Interval, LL=Lower Limit, UL=Upper Limit

¹Number of studies listed in dichotomous analyses of control type is greater than analyzed sample of k=39, due to the following studies containing both active and inactive conditions: Colgan et al., 2016; Enright et al., 2004; Gruzelier et al., 2014; Lin, 2018; Wahbeh et al., 2016; Yekta et al., 2017

Table 8.

Results of Moderation Analyses

Description	No Studies	B	SE	95%CI	95%CI	Z	p-value	R2 analogue
				LL	UL
Treatment Duration (# weeks between pre- and post-treatment anxiety ratings)
All available studies	34	0.003	0.042	−0.078	0.085	0.08	0.937	0.00
Inactive comparisons only	24	−0.015	0.056	−0.125	0.095	−0.27	0.789	0.00
Treatment Intensity (# of sessions)
All available studies	37	−0.020	0.015	−0.050	0.009	−1.34	0.181	0.05
Inactive comparisons only	23	−0.020	0.024	−0.067	0.027	−0.83	0.407	0.01
Publication Year
All available studies	39	−0.000	0.010	−0.021	0.020	−0.02	0.983	0.00
Inactive comparisons only	25	0.016	0.013	−0.009	0.041	1.28	0.201	0.02

Note. k = Number of Studies, SE=Standard Error, CI=Confidence Interval, LL=Lower Limit, UL=Upper Limit

Mixed-effects models were also used to examine whether estimated effects differed by inactive (k=25) or active (k=20) treatment comparison type. Inactive controls lacked a specific intervention component or were given a sham intervention component not specifically designed to target anxiety. These included treatment as usual (TAU) as well as placebo, attention, waitlist and no treatment control condition (Table 2). Analyses indicated a medium to large and significant effect for inactive control comparisons (B=0.709, 95% CI = −0.958, −0.460, p<.001) and a small but significant effect for active treatment control comparisons (B=0.279, 95% CI = −0.537, −0.021 p<.001). Differences between inactive and active control comparison conditions were significant (Q(1)=5.526, p<.05), indicating that effect sizes are related to control comparison type. Again, a significant unexplained variance in true effects (I²) within inactive and within active control comparisons was observed suggesting most of the within-subgroup variance reflects true differences in study effects (see Table 3).

Table 2.

Overview of Respiratory Interventions

Types of Respiratory Intervention	Description
Diaphragmatic Breathing	Use of the diaphragm muscle to move air.
Controlled or Paced Breathing	Breathing at a given respiratory frequency (i.e., pace).
Breathing Retraining	Induction of hyperventilation, which is then contrasted with diaphragmatic breathing, to promote the practice of diaphragmatic breathing.
Heart Rate Variability Biofeedback	Breathing at objectively measured individual’s resonance frequency aiming to increase their heart rate variability. Feedback of physiological parameters is provided in real time to promote accuracy.
PaCO2 Biofeedback	Breathing that targets a specific PaCO2 level, approximated via end-tidal (i.e., expired) CO2, to optimize gas exchange. Feedback of relevant physiological parameters is provided in real time to promote accuracy.

In addition, we examined the effect of treatment duration as reflected by the number of weeks between pre- and post-treatment (range = 1 to 20; see Table 3). We found no significant relation between treatment duration and outcomes in the full sample (k=34; B=0.003, 95% CI = −0.078, 0.085, p>.05) or inactive control comparisons only (k = 24; B=−0.015, 95% CI = −0.125, 0.095, p>.05). Similarly, a meta regression analysis of treatment intensity, as reflected by number of treatment sessions (range: 1–28; k=37), did not account for significant variance in effect size estimates in the full sample (B=−0.020, 95% CI = −0.050, 0.009, p>.05) or inactive control comparisons, only (k = 23; B=0.020, 95% CI = −0.067, 0.027, p>.05; Table 8).

Finally, in an attempt to account for biases in a changing peer-review process over the past several decades as well as intervention quality, year of publication was entered into model as a covariate. This did not account for a significant portion of variance in the entire sample including all studies sample (B=−0.00, 95% CI = −0.21, 0.020, p>.05), or just those comparing to an inactive control condition sample (B=−0.016, 95% CI = −0.009, 0.041, p>.05).

Discussion

Respiratory abnormalities are a hallmark of anxiety symptomatology (Grassi et al., 2013; Leander et al., 2014; Papp et al., 1997; Stein et al., 1995) and may provide a direct means by which to target symptoms such as interoceptive sensitivity (Paulus, 2013). Yet, gold standard interventions do not emphasize intervention components that directly target respiration. The aim of this meta-analysis was to estimate the efficacy of respiratory interventions on trait anxiety symptoms. A total of 40 randomized controlled trials met our inclusion criteria. Overall, a medium to large and significant effect size was observed, indicating that, on average, respiratory interventions are better at reducing anxiety symptoms when compared with control conditions. However, a majority of included studies were classified as ‘high’ risk of bias, many due to the inclusion of an inactive control comparison, which contributed to a bias of outcome measurement. When examining respiratory interventions’ effects as compared to active controls, only, effects were small. Significant heterogeneity in our findings suggests that variability in the intervention ingredients and delivery, population, and control comparison may have obscured the clarity of our findings. Yet, given some promising observations, we make several recommendations for ongoing investigations in this domain.

Respiratory interventions yielded the most promising effects for dimensional anxiety symptoms with medium, and medium to large, effects found in comparison to any control condition and inactive control conditions alone, respectively. Very large, but non-significant, effects were found for panic symptoms in a likely under-powered analysis with just six studies reporting panic outcomes. Observed effect sizes for generalized anxiety and panic symptoms are not surprising given that somatic symptoms, in particular cardiopulmonary and musculoskeletal symptoms, are prominent in generalized anxiety and panic (Bekhuis et al., 2015), and subtypes of individuals with panic are in fact characterized by respiratory dysfunction (Boulding et al., 2016; Courtney, 2017; Courtney, Greenwood, et al., 2011; Courtney, van Dixhoorn, et al., 2011). It is possible that respiratory interventions reduce arousal via correcting breathing abnormalities, thereby preventing the onset of panic distress. Interestingly, dismantling studies of CBT for panic (Pompoli et al., 2018; Schmidt et al., 2000) as well as experimental studies targeting panic relevant processes (Lickel et al., 2013) have found that the inclusion of respiratory interventions does not improve clinical outcomes, despite being associated with treatment acceptability. Yet, others have found that anxiety interventions with respiratory components (i.e., relaxation training) are superior to CBT (in older adults; Thorp et al., 2009), and our meta-analytic review indicates that respiratory interventions indeed hold clinical value for the treatment of anxiety, including panic symptoms. We did not identify any trials on individuals diagnosed with generalized anxiety disorder. Given evidence that these interventions reduce general anxiety symptoms, such a line of inquiry is warranted. Conversely, respiratory interventions yielded small and non-significant effects on posttraumatic stress symptoms. Although posttraumatic stress symptoms include physiological components (e.g., hyperarousal), these symptoms are conditional on the type of trauma. Additionally, heterogeneity of posttraumatic symptoms, relative to panic and generalized anxiety (Asmundson et al., 2000; King et al., 1998; Pai et al., 2017), may further dilute, or make it more difficult to detect, the effects of a respiratory based intervention among those with trauma symptoms.

The effects of respiratory interventions were medium to large in non-psychiatric and medical populations, and medium in the populations presenting with a psychiatric condition, suggesting broad intervention utility. Of the trials including psychiatric populations, just two (Meuret et al., 2008; Lin et al. 2019) used an inactive control comparison. Our findings are relatively consistent with previous meta-analyses of CBT for dimensional anxiety that have found low to medium effects in comparison to other forms of psychotherapy (e.g., Hedges’ g= 0.43; Tolin, 2010). Just three of eleven interventions conducted on psychiatric samples included the respiratory intervention as a treatment adjunct (Bonn et al., 1984; Craske et al., 1997; Penzlin et al., 2015). This indicates that respiratory interventions may offer a promising anxiety intervention approach in psychiatric and non-psychiatric samples who do not have access to, are unwilling to engage in, or have not experienced symptom alleviation from other gold-standard interventions. Moreover, respiratory interventions are highly disseminable and can easily be incorporated into primary care settings, and/or administered remotely. For example, many respiratory and respiratory biofeedback interventions are readily accessible via smartphone apps (Pham et al., 2016).

We did not find support for the perspective that treatment effects vary based on modality (e.g., biofeedback), duration, or intensity. This is surprising given biofeedback interventions provide immediate data regarding changes in physiology that are occurring in the context of a respiratory manipulation, which may enhance fidelity, participant engagement, and accuracy of directly targeting mechanisms purported to promote and maintain anxiety interference. Moreover, some respiratory interventions are purported to require extensive practice and mastery in order to obtain an effect (Lehrer & Vaschillo, 2004; Lehrer et al., 2000). Our current findings challenge this notion and elicit further questions regarding mechanisms of treatment and how to enhance intervention effects. However, given the heterogeneity in study methods, our analysis of treatment modality was limited, and should be re-examined in the future.

One possible, but not yet tested, explanation of observed effects is that respiratory interventions improve cognitive processes implicated in the onset and maintenance of anxiety, such as intolerance of uncertainty (i.e., unwillingness to tolerate the possibility of something bad happening in the future) which are posited to be a central feature of various types of anxiety (Gentes & Ruscio, 2011; Holaway et al., 2006). Anxious individuals who are unable to withstand the anxiety that accompanies the possibility of something bad happening in the future may experience respiratory interventions as a means by which to control their physiology. This may generalize to a greater sense of anxiety control and self-efficacy in managing symptoms (Gallagher et al., 2014), while also providing them with a skill to use when feeling anxious. Respiratory interventions may also act upon similar mechanisms to interoceptive exposure, such as anxiety sensitivity. For example, heart rate variability biofeedback requires individuals to dramatically slow their breathing pace, which, while increasing relaxation when done correctly, may initially result in worry that one is not getting enough oxygen, and therefore promote anxiety. Similarly, many respiratory interventions promote increased awareness of respiration and other related physiological changes, which may increase sensitivity to sensations, and initially increase anxiety. Finally, interventions such as breath awareness training may increase tolerance and acceptance of anxiety relevant discomfort. This is consistent with hypothesized mechanisms underlying Acceptance and Commitment Therapy (Hayes et al., 2006) that ask patients to practice attending to present experiences (such as respiration) as a means of more broadly strengthening present-focused awareness of negative thoughts, feelings, and sensations. This in turn allows patients to become more aware of the role that anxiety plays in their behaviors and provides an opportunity for patients to volitionally choose to engage in desired activities despite the presence of anxiety. While assessing the potential mediating role of cognitive processes in explaining the benefits of respiratory interventions is beyond the scope of the current investigation, future trials should index these processes as a means of explicating mechanisms of change in these interventions.

Our meta-analytic review is limited by considerable variability amongst studies with regard to sample characteristics, intervention type, duration and intensity of treatment, and use of clinical or non-clinical (e.g., STAI) outcome measures, as reflected in immense effect size dispersion. Despite this, included interventions have a common target (i.e., respiration), and their heterogeneity may be used to propel several important lines of inquiry. First, although some previous research on clinical samples (e.g., Pompoli et al., 2018; Schmidt et al., 2000) has found that respiratory interventions are inferior to other treatment packages, they present a viable alternative that is acceptable and highly disseminable, and perhaps optimal for certain populations (e.g., older adults; Thorp et al., 2009). Additional research should also determine what type of anxiety symptoms are most effectively addressed by respiratory interventions. Such a line of inquiry may also help researchers and clinicians understand when, and for whom, respiratory adjuncts may be most useful. Future research also is needed to examine possible mechanisms explaining favorable outcomes for respiratory interventions, including arousal reduction, extinction, and modifying appraisal or expectancies. Finally, anxiety is commonly comorbid with a variety of health conditions, and often complicates treatment. Future work would benefit from examining the efficacy of respiratory interventions in other health risk behaviors, such as illicit substance use, alcohol (e.g., ClinicalTrials.gov ID: NCT02579317), and smoking (e.g., ClinicalTrials.gov ID: NCT03972137), where comorbid anxiety is highly prevalent.

Collectively, the current review provides support for respiratory interventions serving as effective treatments for anxiety disorders. In particular, evidence suggests respiratory interventions are efficacious in treating dimensional symptoms of anxiety, notably generalized anxiety and panic, while also being acceptable anxiety interventions for a range of clinical and non-clinical populations. Yet, findings from this review also reflect the predominant use of non-active comparators when evaluating respiratory intervention efficacy, and the need for future trials to incorporate more adequately matched active controls. Moreover, there is a need for increased standardization across methodological designs in addition to more transparent reporting of trial outcome data that is consistent with Cochrane Database assessment tools. In efforts to reduce risk of bias future trials would benefit from pre-registering data analytic strategies, as well as incorporating active comparators matched in treatment intensity and dose (i.e., frequency and duration of in-person sessions, assignment of at-home practice, etc.). Greater standardization across methodological designs may also assist in identifying common and distinct therapeutic elements across respiratory modalities the current review was unable to detect. This may also facilitate in identifying how respiratory elements may differentially relate to and operate on underlying processes of change implicated in anxiety pathology. Given the promise respiratory interventions hold in aiding current anxiety disorder treatments, ongoing research initiatives that address these methodological barriers are highly encouraged.

Highlights.

• Respiratory abnormalities are a hallmark of anxiety symptomatology

• The effectiveness of respiratory interventions for anxiety is not well established

• Respiratory interventions yield significant improvements in anxiety symptoms

• Respiratory interventions hold clinical value for the treatment of anxiety

Author Bios:

Teresa M. Leyro, Ph.D.: Dr. Teresa Leyro is an Assistant Professor in the Doctoral Program in Clinical Psychology in the Department of Psychology at Rutgers University-New Brunswick.

Mark Versella, M.S.: Mark Versella is a student in the Doctoral Program in Clinical Psychology in the Department of Psychology at Rutgers University-New Brunswick.

Min-Jeong Yang, Ph.D.: Min-Jeong Yang is a post-doctoral fellow in the Department of Health Outcomes and Behavior at Moffitt Cancer Center and former student in the Doctoral Program in Clinical Psychology in the Department of Psychology at Rutgers University-New Brunswick.

Hannah R. Brinkman, M.S.: Hannah Brinkman is a student in the Doctoral Program in Clinical Psychology in the Department of Psychology at Rutgers University-New Brunswick.

Danielle L. Hoyt, M.A.: Danielle Hoyt is a student in the Doctoral Program in Clinical Psychology in the Department of Psychology at Rutgers University-New Brunswick.

Paul Lehrer, Ph.D.: Dr. Paul Lehrer is a Professor, emeritus, in the Department of Psychiatry at Rutgers Robert Wood Johnson Medical School