Introduction
In 2019, an estimated 301 million people worldwide were living with anxiety disorders, representing a 54.64% increase between 1990 and 2019 (GBD 2019 Mental Disorders Collaborators, 2022). Moreover, among mental disorders, anxiety disorders account for 22.9% of disability-adjusted life years (DALYs), second only to depression, and are the eighth leading cause of years lived with disability (GBD 2019 Mental Disorders Collaborators, 2022). In 2020 alone, anxiety disorders caused 44.5 million (30.2–62.5) DALYs globally (COVID-19 Mental Disorders Collaborators, 2021). In 2012, the estimated costs derived from anxiety in Europe were approximately €74,380 million, of which 62.2% were direct medical costs, 0.2% were direct nonmedical costs, and 37.6% were indirect costs (Olesen, Gustavsson, Svensson, Wittchen, & Jönsson, Reference Olesen, Gustavsson, Svensson, Wittchen and Jönsson2012).
Although there are many effective treatments available for anxiety disorders (Bandelow et al., Reference Bandelow, Reitt, Röver, Michaelis, Görlich and Wedekind2015), often people do not have access to them for a variety of reasons, such as diagnostic errors, poor treatment adherence, or inadequate treatment (Chapdelaine, Carrier, Fournier, Duhoux, & Roberge, Reference Chapdelaine, Carrier, Fournier, Duhoux and Roberge2018; Fernández et al., Reference Fernández, Haro, Martinez-Alonso, Demyttenaere, Brugha, Autonell and Alonso2007). This is particularly problematic because, while efforts to reduce the burden of anxiety disorders have largely focused on closing the ‘treatment gap’, recent analyses suggest that this alone may not be sufficient. Beyond increasing access to treatment, addressing a ‘quality gap’ – ensuring treatments meet clinical guidelines and reach those in greatest need – is also crucial. Additionally, a ‘prevention gap’ may exist, where resources for reducing incidence through prevention have lagged behind treatment efforts (Jorm, Patten, Brugha, & Mojtabai, Reference Jorm, Patten, Brugha and Mojtabai2017). There seems to be a lack of awareness of the importance of prevention programs and mental health promotion, leading to a disproportionate allocation of funding and resources toward treatment rather than prevention in most countries (WHO, 2021).
Different psychological and educational interventions have proven to be effective in preventing anxiety disorders, yielding a small but significant effect, standardized mean difference (SMD) = −0.31 (95% confidence interval [CI]: −0.40 to −0.21; p < 0.001), according to a previous meta-analysis that included data from 29 randomized controlled trials (RCTs) (Moreno-Peral et al., Reference Moreno-Peral, Conejo-Cerón, Rubio-Valera, Fernández, Navas-Campaña, Rodríguez-Morejón, Motrico and Bellón2017). However, to effectively reduce the incidence of anxiety, these interventions must be accessible to a broad population. Information and communication technology is emerging as a reliable solution to address some of these issues. In 2023, a total of 5.16 billion people were Internet users, equivalent to 64.4% of the world’s total population (We Are Social & Meltwater, 2023). Computerized interventions offer a more cost-effective way of scaling up preventive interventions. They also offer several additional advantages over traditional methods such as anonymity, enhanced flexibility and accessibility, allowing users to access them at any time and from virtually any location, lower costs compared to face-to-face interventions, the ability to bridge geographic distances, and the option to revisit therapy material as needed (Khanna, Aschenbrand, & Kendall, Reference Khanna, Aschenbrand and Kendall2007; Schuster, Topooco, Keller, Radvogin, & Laireiter, Reference Schuster, Topooco, Keller, Radvogin and Laireiter2020).
As a result, the number of digital mental health interventions is rapidly accelerating. This method has already been proven effective for the treatment of anxiety disorders, showing that online interventions can be as effective as face-to-face treatments, with a combined effect size (Hedge’s g) of approximately 0.80, and disorder-specific effect sizes between 0.62 and 1.31 (Andrews et al., Reference Andrews, Basu, Cuijpers, Craske, McEvoy, English and Newby2018; Eilert et al., Reference Eilert, Enrique, Wogan, Mooney, Timulak and Richards2020; Pauley, Cuijpers, Papola, Miguel, & Karyotaki, Reference Pauley, Cuijpers, Papola, Miguel and Karyotaki2023). A recent meta-analysis by Pauley et al. (Reference Pauley, Cuijpers, Papola, Miguel and Karyotaki2023) also performed subgroup analyses comparing guided versus unguided online interventions for the treatment of anxiety and found no differences in effectiveness between these delivery methods, suggesting that even self-administered digital interventions can be as effective as face-to-face therapy.
Evidence on the effectiveness of digital interventions for the prevention of anxiety remains limited, with few meta-analyses specifically addressing these types of programs. One of these, by Pennant et al. (Reference Pennant, Loucas, Whittington, Creswell, Fonagy, Fuggle and Kendall2015), reviewed the evidence for all types of computerized interventions for anxiety and depression in children and young people (5–25 years of age). The results highlighted potential benefits for the general population of young people (SMD = −0.15), although the effect sizes were smaller compared to those observed in participants with mild to moderate anxiety or those considered ‘at risk’ (SMD = −0.77). Later, Sander, Rausch, and Baumeister (Reference Sander, Rausch and Baumeister2016) conducted a systematic review and meta-analysis to evaluate the effectiveness of existing Internet-based preventive interventions for mental disorders, supporting effective preventive interventions for subthreshold anxiety. Another meta-analysis conducted by Deady et al. (Reference Deady, Choi, Calvo, Glozier, Christensen and Harvey2017), which assessed eHealth interventions for the prevention of depression and anxiety in the general population (18–64 years old), found a similar small but significant effect size for both outcomes, with an overall mean difference of 0.31 for anxiety symptoms. Most recently, a meta-analysis conducted by Edge, Watkins, Limond, and Mugadza (Reference Edge, Watkins, Limond and Mugadza2023) on self-guided computerized interventions for the prevention of anxiety and/or depression in adults (>16 years of age) found a small but significant effect of these interventions on reducing anxiety symptomatology (overall SMD = −0.21, p < 0.001). Nevertheless, these studies were limited to specific age ranges (Deady et al., Reference Deady, Choi, Calvo, Glozier, Christensen and Harvey2017; Edge et al., Reference Edge, Watkins, Limond and Mugadza2023; Pennant et al., Reference Pennant, Loucas, Whittington, Creswell, Fonagy, Fuggle and Kendall2015), had very specific inclusion criteria (such as including only self-guided interventions) (Edge et al., Reference Edge, Watkins, Limond and Mugadza2023), did not limit inclusion solely to studies on the prevention of anxiety but also encompassed its treatment, or only reported mean scores without clearly stating that participants did not exceed clinical cut-offs at baseline (Edge et al., Reference Edge, Watkins, Limond and Mugadza2023; Pennant et al., Reference Pennant, Loucas, Whittington, Creswell, Fonagy, Fuggle and Kendall2015). For these reasons, a meta-analysis should be carried out with the most recent data, including only purely preventive studies, that is, those that include participants without a diagnosis of anxiety disorder at baseline/before enrolling in the study. Thus, we aimed to conduct a meta-analysis of RCTs assessing the effectiveness of digital psychological interventions for the prevention of anxiety disorders in all types of populations.
Materials and methods
This study followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines for reporting systematic reviews and meta-analyses (Page et al., Reference Page, McKenzie, Bossuyt, Boutron, Hoffmann, Mulrow and Moher2021). This meta-analysis is also registered with the International Prospective Register of Systematic Reviews (registration number: CRD42022307194).
Selection criteria
Among all the experimental designs that can be used to measure effectiveness, we selected RCTs since they have the lowest risk of bias (Piantadosi, Reference Piantadosi2005). We focused on psychoeducational and/or psychological interventions. Studies or arms that compared other types of interventions were excluded. Psychoeducational interventions consist of providing information about anxiety through lectures or fact sheets, whereas psychological interventions attempt to change how people think, their behaviors, and their learning skills by using a variety of strategies or therapeutic approaches (e.g. cognitive behavioral therapy [CBT] or systemic therapy).
We only included those studies that excluded participants with an anxiety disorder at baseline, or those that provided separate results for non-anxious participants at baseline, by using a standardized interview (e.g. Structured Clinical Interview for DSM Disorders), or validated self-reports with standard cut-off points (e.g. Beck Anxiety Inventory II). In these cases, only the non-anxious participants were included in the analysis. We included studies in all types of populations, regardless of age, sex, clinical, or health condition (e.g. pregnancy) and regardless of where they are recruited (general population, primary care, mental health clinics, etc.). The allowed comparators were care-as-usual, no intervention, waiting list for intervention, or attention control. The intervention had to be entirely digital; therefore, combined digital and face-to-face sessions (‘blended interventions’) were excluded. Outcomes included the incidence of new cases of any DSM-5 anxiety disorder and/or the reduction in anxiety symptomatology measured by standardized interviews or validated symptom scales. There were no restrictions on the language or setting of the studies.
Search strategy
In this meta-analysis, we searched six electronic databases including PubMed, PsycINFO, EMBASE, Web of Science, OpenGrey (System for Information on Grey Literature in Europe, which links to DANS Easy Archive), and CENTRAL (Cochrane Central Register of Controlled Trials) from inception to December 12, 2024. The specific search strategies used in each database are provided in Appendix A. Additionally, the reference list of existing systematic reviews on the topic (Andrews et al., Reference Andrews, Basu, Cuijpers, Craske, McEvoy, English and Newby2018; Deady et al., Reference Deady, Choi, Calvo, Glozier, Christensen and Harvey2017; Edge et al., Reference Edge, Watkins, Limond and Mugadza2023; Eilert et al., Reference Eilert, Enrique, Wogan, Mooney, Timulak and Richards2020; Linardon et al., Reference Linardon, Torous, Firth, Cuijpers, Messer and Fuller‐Tyszkiewicz2024; Moreno-Peral et al., Reference Moreno-Peral, Conejo-Cerón, Rubio-Valera, Fernández, Navas-Campaña, Rodríguez-Morejón, Motrico and Bellón2017; Newby, Twomey, Yuan Li, & Andrews, Reference Newby, Twomey, Yuan Li and Andrews2016; Noh & Kim, Reference Noh and Kim2023; Pauley et al., Reference Pauley, Cuijpers, Papola, Miguel and Karyotaki2023; Pennant et al., Reference Pennant, Loucas, Whittington, Creswell, Fonagy, Fuggle and Kendall2015; Sander, Rausch & Baumeister, Reference Sander, Rausch and Baumeister2016; Seegan, Miller, Heliste, Fathi, & McGuire, Reference Seegan, Miller, Heliste, Fathi and McGuire2023) was manually revised, in order to find additional studies. Online trial registers were also consulted, specifically ClinicalTrials.gov and Australia New Zealand Clinical Trials Register. We also consulted experts in the field to identify any new studies that met our inclusion criteria.
Study selection
After removing duplicates, all studies were initially reviewed based on their title and abstract by two pairs of reviewers (PMP and CGH, SCC, and CMV). Studies that did not meet the eligibility criteria were excluded. Potentially eligible studies then underwent a full-text review for final inclusion, using the same method. Any discrepancies were discussed and resolved by consensus.
Data extraction
The data extracted from each selected study were compiled in an evidence table. Specifically, we extracted data concerning: (a) bibliographic information (such as author, year, country); (b) characteristics of the participants (such as mean age, symptoms of anxiety); and (c) characteristics of the RCTs (such as sample size, type of intervention, type of comparator, or outcome). This process was also conducted by two reviewers (PMP and SCC) and replicated by another two (CGH and CMV). Any discrepancies were discussed and resolved by consensus.
Risk of bias
The quality of each included RCT was assessed in accordance with the Cochrane Collaboration’s Risk-of-Bias (RoB) 2.0 tool, based on five dimensions: (1) bias arising from the randomization process; (2) bias due to deviations from the intended interventions; (3) bias due to missing outcome data; (4) bias in measurement of the data; and (5) bias in selection of the reported result. This tool classifies each RCT as having low risk of bias, if all dimensions are categorized as low risk; some concern of bias, if one or more dimensions are categorized as some concerns, but none are classified as high risk; or high risk of bias, if at least one dimension is categorized as high risk, or multiple domains are categorized as some concerns in a way that substantially reduces confidence in the results (Sterne et al., Reference Sterne, Savovic, Page, Elbers, Blencowe, Boutron and Higgins2019). Each of the five dimensions was assessed both qualitatively (categorized as low risk of bias, some concerns, or high risk of bias) and quantitatively (receiving zero, one, or two points, respectively). This assessment was conducted in duplicate by two researchers (PMP and SCC), and any discrepancy was resolved through consensus.
Statistical analysis
All analyses were carried out using the STATA statistical package (version 14.2) and Comprehensive Meta-Analysis (CMA version 2.2.064).
When the outcome was differences in anxiety symptoms, the mean scores and standard deviations were extracted, and SMDs between the intervention and the control groups were calculated and used to estimate the pooled effect size. When an RCT only provided data on the incidence of anxiety, the CMA package was used to convert it into an SMD. We used the SMD because most of the RCTs included in our meta-analysis reported differences in anxiety symptoms. For each RCT, SMD was calculated by combining this parameter from the different post-intervention follow-up assessments into a mean estimated difference, and its 95% CI. For any RCT that included two different intervention groups and a single control group, standard errors in nested comparisons in the same RCT were inflated, following the recommendation of Cates (Rücker, Cates, & Schwarzer, Reference Rücker, Cates and Schwarzer2017). Negative SMDs indicated an improved outcome (reduction of anxiety symptoms) in the intervention group. Cohen proposed the following interpretation of effect sizes as −0.2 small, −0.5 medium, and −0.8 large (Lachenbruch & Cohen, Reference Lachenbruch and Cohen1989). A priori, a random-effects model was used to estimate the pooled effect size on the assumption that the RCTs included in our meta-analysis were conducted in heterogeneous populations. Moreover, RCTs with disproportionately high effect sizes were excluded a priori to minimize the impact of extreme outliers on the pooled results.
Heterogeneity was assessed using the I 2 statistic, where an I 2 of 0%–40% indicates not important heterogeneity; 30%–60% moderate heterogeneity; 50%–90% substantial heterogeneity; and 75%–100% considerable heterogeneity, according to the Cochrane Handbook (Higgins et al., Reference Higgins, Thomas, Chandler, Cumpston, Li, Page and Welch2019). To determine whether differences in the effect sizes of the individual studies exceeded those that would be expected due to chance, we used the Q-test, considering p > 0.10 as nonsignificant heterogeneity.
We evaluated publication bias by assessing funnel plot asymmetry using the Duval and Tweedie (Reference Duval and Tweedie2000) trim-and-fill procedure. This procedure yields an adjusted pooled effect size after accounting for missing studies due to publication bias. To objectively assess this asymmetry, we also performed the rank correlation test (Begg & Mazumdar, Reference Begg and Mazumdar1994) and the Egger test.
Subgroup analyses were performed using a mixed-effects model according to:
(a) Characteristics of the sample: continent, mean age, recruitment setting, and sample size.
(b) Characteristics of the intervention: type of prevention (universal, selective, or indicated), therapeutic approach (e.g. CBT, psychoeducation), presence of guidance, number of sessions, and intervention format (e.g. web-based, videoconference).
(c) Methodological characteristics: type of outcome (primary/secondary), type of outcome measure (symptom scale vs. standardized diagnostic interview), comparator, adherence rate, risk of bias, and duration of follow-up.
We conducted sensitivity analyses at the first and last follow-up, using Hedge’s g, and excluding from the analysis the RCT that caused the greatest increase in heterogeneity.
We also performed bivariate random-effects meta-regressions (with only one moderator included in each model), which enables the estimation of robust standard errors for random effects (Knapp & Hartung, Reference Knapp and Hartung2003). The post hoc analysis strategy to explain the maximum heterogeneity consisted of obtaining the most parsimonious meta-regression model (including the least number of variables) with the best goodness of fit. We used the Higgins and Thompson permutation-test approach to calculate p values, taking into account multiplicity adjustments (Higgins & Thompson, Reference Higgins and Thompson2004).
The quality of evidence
The quality of the evidence was assessed using the ‘Grading of Recommendations Assessment, Development, and Evaluation’ (GRADE) working group methodology. This method evaluates the following domains: risk of bias, consistency, directness, precision, and publication bias (Balshem et al., Reference Balshem, Helfand, Schünemann, Oxman, Kunz and Guyatt2011).
Results
Study selection
After concluding the search in six databases, and consulting experts and references from previous research, a total of 6793 articles were retrieved. After removing 1116 duplicates, 5677 studies were reviewed based on their titles and abstracts. Of these, 208 underwent full-text review, ultimately resulting in the selection of 11 RCTs that met the inclusion criteria for the meta-analysis. In addition, four RCTs that met all criteria except for excluding participants with anxiety at baseline were included after the author(s) provided data specific to participants who did not exceed the anxiety threshold at the beginning of the study. This resulted in a final sample of 15 RCTs (Bendtsen, Müssener, Linderoth, & Thomas, Reference Bendtsen, Müssener, Linderoth and Thomas2020; Calear, Christensen, Mackinnon, Griffiths, & O’Kearney, Reference Calear, Christensen, Mackinnon, Griffiths and O’Kearney2009; Christensen et al., Reference Christensen, Batterham, Mackinnon, Griffiths, Kalia Hehir, Kenardy and Bennett2014; Cukrowicz & Joiner, Reference Cukrowicz and Joiner2007; Fledderus, Bohlmeijer, Pieterse, & Schreurs, Reference Fledderus, Bohlmeijer, Pieterse and Schreurs2012; Fonseca, Alves, Monteiro, Gorayeb, & Canavarro, Reference Fonseca, Alves, Monteiro, Gorayeb and Canavarro2020; Garcia-López et al., Reference Garcia-López, Jimenez-Vazquez, Muela-Martinez, Piqueras, Espinosa-Fernandez, Canals-Sans and Ehrenreich-May2024; Howell, Rheingold, Uhde, & Guille, Reference Howell, Rheingold, Uhde and Guille2019; Lokman et al., Reference Lokman, Leone, Sommers-Spijkerman, van der Poel, Smit and Boon2017; Mak et al., Reference Mak, Tong, Fu, Leung, Jung, Watkins and Lui2024; Monteiro, Pereira, Canavarro, & Fonseca, Reference Monteiro, Pereira, Canavarro and Fonseca2020; Schotanus-Dijkstra et al., Reference Schotanus-Dijkstra, Drossaert, Pieterse, Boon, Walburg and Bohlmeijer2017; Topper, Emmelkamp, Watkins, & Ehring, Reference Topper, Emmelkamp, Watkins and Ehring2017; Vivas-Fernández et al, Reference Vivas-Fernandez, Garcia-Lopez, Piqueras, Muela-Martinez, Canals-Sans, Espinosa-Fernandez and Ehrenreich-May2023; Zarski et al., Reference Zarski, Weisel, Berger, Krieger, Schaub, Berking, Gorlich, Jacobi and Ebert2024). The selection process is detailed in a flowchart presented in Figure 1.
Flowchart of the inclusion of records in the systematic review and meta-analysis, according to the PRISMA guidelines.
Study characteristics
Table 1 describes the characteristics of the 15 RCTs included in the systematic review. All were published between 2007 and 2024; four were conducted in the Netherlands, two in Spain, two in Australia, two in Portugal, two in the United States, one in China, and one in Sweden. The total number of participants included in the studies was 6093, and the sample size ranged between 68 and 1239 (median = 275, interquartile range = 410) and was comprised of adolescents and adults. Two studies focused specifically on postpartum women as their target population. Five of the studies focused on universal prevention, eight on selective prevention, and two on indicated prevention. The majority of studies (11) were based on CBT strategies for their interventions, two of them in positive psychology, one in acceptance and commitment therapy, and one combined CBT with other therapies. Comparators were a waiting list in eight RCTs, active control in six, and care as usual in one. Anxiety assessment was the primary outcome for eight of the RCTs, and the secondary outcome for another seven. Guidance was present in nine of the interventions and absent in the other six. The intervention format was web-based in ten studies, video conference in two, e-mail in two, and text messages in one. Anxiety outcomes were measured by symptomatology scales in all RCTs except one, which used a standardized diagnostic interview. The follow-up periods ranged between 8 and 52 weeks. Recruitment settings in these RCTs were general population in seven, educational in seven, and medical in one. Exclusion of anxiety cases at baseline was performed using symptom scales in seven trials, diagnostic interviews in four, and in four cases, the authors provided the data for participants without anxiety at the beginning of the study (those with a score below the cutoff in a validated symptom scale).
Characteristics of the studies included
Note. HADS = Hospital Anxiety and Depression Scale; MHC-SF = Mental Health Continuum Short Form; CES-D = Center for Epidemiologic Studies of Depression; RCMAS = Revised Children’s Manifest Anxiety Scale; MINI = Mini-International Neuropsychiatric Interview; GAD-7 = Generalized Anxiety Disorder questionnaire (7 items); BAI = Beck Anxiety Inventory; BDI = Beck Depression Inventory; PANAS = Positive and Negative Affect Schedule; STAI-S = State–Trait Anxiety Inventory (State anxiety scale); HADS-A = Hospital Anxiety and Depression Scale (Anxiety subscale); PDPI-R = Postpartum Depression Predictors Inventory-Revised; EPDS = Edinburgh Postnatal Depression Scale; HADS = Hospital Anxiety and Depression Scale; PHQ-9 = Patient Health Questionnaire-9; IDS-SR = Inventory of Depressive Symptomatology Self-Report; HADS-D = Hospital Anxiety and Depression Scale (depression subscale); FS = Flourishing Scale; MASQD-30 = Mood and Anxiety Symptoms Questionnaire (30 items); BDI-II: Beck Depression Inventory version 2; GADQ-IV = Generalized Anxiety Disorder Questionnaire IV; SDQ = Strengths and Difficulties Questionnaire; ACT = Acceptance and Commitment Therapy; CBASP=Cognitive-Behavioral Analysis System of Psychotherapy; Strengths and Difficulties Questionnaire; RCADS-30 = Revised Children’s Anxiety and Depression Scale; PSWQ = Penn State Worry Questionnaire; RRS = Rumination Response Scale; PSS = Perceived Stress Scale; STAIX-I = State–Trait Anxiety Inventory (State anxiety scale, X version); WEMWBS = Warwick-Edinburgh Mental Wellbeing Scale; MW-S = Mind Wandering Spontaneous; DERS = Difficulties in Emotion Regulation Scale; RS-14 = Resilience Scale (14 items); ACS-S = Attention Control Scale (Shifting scale); ACS-D = Attention Control Scale (Distraction scale); ADIS-5 C/P = Anxiety and Related Disorders Interview Schedule for DSM-5; HAM-A = Hamilton Anxiety Rating Scale; SIGH-A = Structured Interview Guide for the Hamilton Anxiety Scale; QIDS-C = Quick Inventory of Depressive Symptomatology.
a Type of prevention: Indicated: patients with subthreshold anxiety; Selective: patients with a risk factor for anxiety; Universal: general population.
b Subsample of participants without clinical anxiety at baseline proportioned by the authors.
c These data are from two different articles, both from Vivas-Fernández et al., (Reference Vivas-Fernandez, Garcia-Lopez, Piqueras, Muela-Martinez, Canals-Sans, Espinosa-Fernandez and Ehrenreich-May2023).
Study risk of bias
Of the 15 studies, 13 presented an overall high risk of bias, and only two (Garcia-López et al., Reference Garcia-López, Jimenez-Vazquez, Muela-Martinez, Piqueras, Espinosa-Fernandez, Canals-Sans and Ehrenreich-May2024; Zarski et al., Reference Zarski, Weisel, Berger, Krieger, Schaub, Berking, Gorlich, Jacobi and Ebert2024) presented a low risk of bias. Regarding the randomization process, two studies had some concerns of bias, and the rest presented a low risk of bias. All the RCTs had a low risk of bias derived from deviations from the intervention. With respect to missing data, five studies presented a low risk of bias, whereas the rest presented a high risk. The measurement of the outcome led to a low risk of bias in five studies, and a high risk in the remaining 11. Finally, the risk of bias associated with the selection of the results was low in most of the studies, with some concerns in two studies (Table B.1, Appendix B).
Primary analysis
Figure 2 contains the forest plot. Although 15 studies were included in the systematic review, one (Lokman et al., Reference Lokman, Leone, Sommers-Spijkerman, van der Poel, Smit and Boon2017) was excluded from the meta-analysis because it reported an excessively large effect size and was identified as an outlier. After removing this study, the primary analysis shows a small preventive effect size SMD = −0.325 (95% CI: −0.44 to −0.20, p < 0.001). Between-outcomes heterogeneity was high I2 = 72.4% (95% CI: 56%–83%) and statistically significant (Q-test p < 0.001).
Forest plot.
Note: SMD = standardized mean difference.
Sensitivity analysis
Sensitivity analyses were conducted to assess the robustness of the results. Hedge’s g indicates a small preventive effect size of −0.324 (95% CI: −0.447 to −0.200, p < 0.001). The effect size suffered a minimal decrease when using the results from the first post-intervention evaluation SMD = −0.311 (95% CI: −0.437 to −0.185, p < 0.001). When using data from the last follow-up, the pooled effect size remained practically unchanged SMD = −0.326 (95% CI: −0.453 to −0.200, p < 0.001) as compared to the primary analysis. Regarding age groups, smaller effect sizes were found in studies focusing exclusively on adolescents (SMD = −0.265), as well as in those in which the mean age of participants fell within the young adult range (SMD = −0.242). For studies targeting adult populations (mean age ≥ 30 years), the effect size increased compared to the primary analysis (SMD = −0.420). These results are presented in Table 2.
Primary and sensitivity analyses of the effectiveness of digital interventions in preventing anxiety
Publication bias
Regarding publication bias, two RCTs were imputed to enhance funnel plot symmetry. As a result, the adjusted pooled effect size increased, SMDadj = −0.367(CI: −0.492 to −0.241; p < 0.001). Publication bias was statistically nonsignificant (Egger test, p = 0.978; Begg test p = 0.909). Figure C.1 (Appendix C) provides the adjusted funnel plot.
Subgroup analysis
Subgroup analyses revealed statistically significant between-group differences based on several factors including continent, sample mean age, sample size, recruitment setting, type of prevention, intervention format, therapeutic approach, guidance, type of anxiety outcome, and adherence rate. Further information can be found in Table D.1 (Appendix D).
Meta-regression
The results obtained by performing bivariate regressions can be found in Table E.1 (Appendix E). The final meta-regression model (Table 3) included only one variable, the country (The Netherlands) [β = −0.538 (95% CI: −0.708 to −0.368); p < 0.001], which was associated with a higher preventive effectiveness. This model explained 100.0% of the variance, and its goodness of fit was good (Figure F.1, Appendix F).
Final meta-regression model
a I2 residual = 0.99%; adjusted R2 = 100%.
b Knapp and Hartung method for the estimation of standard error and 95% CI.
Quality of evidence
The initial quality of the evidence was rated as high, as we only included RCTs in this meta-analysis. Heterogeneity was considerable, and although this was fully explained by meta-regression, we reduced the rating from high to moderate. We further reduced the rating from moderate to low because only one study was assessed as having a low risk of bias. Conversely, no statistical evidence of publication bias was identified. Although the number of studies included was small, it was adequate to ensure the precision of the meta-analysis. Indirectness was very low because, although we searched for all types of populations, we found no trials targeting mature adults or the elderly; therefore, the results do not represent all types of populations, which was our primary interest. In summary, the quality of evidence according to GRADE was very low.
Discussion
Main findings
This systematic review identified 15 RCTs, including a total of 6093 participants from America, Asia, Europe, and Oceania, with most interventions grounded in CBT principles. One study was excluded from the quantitative synthesis because it reported an excessively large effect size and was considered an outlier. The subsequent meta-analysis of the remaining 14 RCTs showed that digital psychological and psychoeducational interventions exert a small preventive effect on anxiety. Publication bias was nonsignificant, between-study heterogeneity was high, and the certainty of evidence according to GRADE was very low.
Strengths
To our knowledge, this is the first meta-analysis of RCTs to assess the effectiveness of digital interventions for the prevention of anxiety in non-anxious and varied populations. Our strict selection criteria ensured that we only evaluated preventive psychological interventions, and not treatments, in diverse populations. A broad selection of complementary databases, combined with a wide range of search terms, allowed a highly sensitive screening process, thus maximizing the inclusion of all studies meeting the selection criteria. This process was carried out by trained, independent reviewers. Moreover, no language or population restrictions were applied. PRISMA criteria were followed throughout the entire development of this meta-analysis, and GRADE criteria were used to assess the quality of the evidence. We also performed sensitivity analyses, which support the robustness of a small but still statistically significant pooled SMD. Finally, the meta-regression was able to explain 100% of the heterogeneity.
Limitations
This meta-analysis has some limitations: (i) the overall high risk of bias of the RCTs included in this study demonstrates the need for further evaluation of computerized interventions for the prevention of anxiety following more rigorous methodologies; (ii) only one RCT used a standardized diagnostic interview, a more valid method to measure anxiety outcomes, with the remainder using symptom scales; (iii) four of the RCTs represent a subsample of participants without anxiety at baseline, as provided by the authors of the studies that did use this as an exclusion criterion; (iv) the included RCTs used four different therapeutic approaches, with CBT being the most common; this concentration of evidence on CBT limits the generalizability of our findings, as the effectiveness of alternative interventions remains underexplored and warrants further investigation; (v) older populations are not represented in these studies due to the nature of the interventions and the existence of a digital divide; (vi) all RCTs were conducted in high-income countries, which does not allow us to assess whether these results may differ in low- and middle-income regions; (vii) only two studies provided incidence data, however, even when incidence data are not provided, symptom assessment using clinical scales is a reliable predictor of the incidence of new cases in depression (Institute of Medicine (US) Committee on Prevention of Mental Disorders, Reference Mrazek and Haggerty1994; Cuijpers & Smit, Reference Cuijpers and Smit2004); (viii) follow-up periods were short in the majority of the studies, with a maximum of 1 year; and (ix) only 14 RCTs were included in this MA, due to the exclusion of an outlier RCT.
Comparison with previous research
These results are in line with previous research indicating that digital preventive interventions have a small but significant effect on preventing anxiety symptoms (Edge et al., Reference Edge, Watkins, Limond and Mugadza2023; Sander et al., Reference Sander, Rausch and Baumeister2016). However, evidence on their impact on the incidence of anxiety disorders remains limited, as only two RCTs provided information on new anxiety cases (Christensen et al., Reference Christensen, Batterham, Mackinnon, Griffiths, Kalia Hehir, Kenardy and Bennett2014; Topper et al.) in addition to symptomatology scores. This is also consistent with previous meta-analyses of online interventions, in which most of the results were provided in terms of symptomatology levels (Edge et al., Reference Edge, Watkins, Limond and Mugadza2023; Pauley et al., Reference Pauley, Cuijpers, Papola, Miguel and Karyotaki2023; Sander et al., Reference Sander, Rausch and Baumeister2016). In this line, the overall aim of the three types of preventive intervention – universal, selective, and indicated – is the reduction of the occurrence of new cases. Usually, this is done through a risk-reduction model, and even though outcomes are in the distant future and the goal of fewer cases has not yet been established, the decrease in risk and/or increase in protective factors can be documented (Institute of Medicine (US) Committee on Prevention of Mental Disorders, Reference Mrazek and Haggerty1994), even including estimations of the individual probability of suffering anxiety in the future (Moreno-Peral et al., Reference Moreno-Peral, Conejo-Ceron, Motrico, Rodriguez-Morejon, Fernandez, Garcia-Campayo, Roca, Serrano-Blanco, Rubio-Valera and Bellon2014). Depressive symptoms are a good predictor of future incidence of depression (Cuijpers & Smit, Reference Cuijpers and Smit2004), and their reduction can be seen as an indicator of decreased risk. Additionally, the aims of indicated preventive interventions might be to reduce the duration of early symptoms and to halt progression of severity so that the individuals do not meet, nor come close to meeting, DSM diagnostic levels (Institute of Medicine (US) Committee on Prevention of Mental Disorders, Reference Mrazek and Haggerty1994). Therefore, from this conceptual framework, using differences in anxiety symptoms as an outcome does not, in itself, imply that the term ‘anxiety prevention’ cannot be used.
Other meta-analyses of anxiety prevention, such as the one conducted by Moreno-Peral et al. (Reference Moreno-Peral, Conejo-Cerón, Rubio-Valera, Fernández, Navas-Campaña, Rodríguez-Morejón, Motrico and Bellón2017), reported a small effect size (SMD = −0.31), including both digital and face-to-face interventions. In contrast, a meta-analysis by Pauley et al. (Reference Pauley, Cuijpers, Papola, Miguel and Karyotaki2023) of computerized interventions for anxiety disorders, covering treatment and prevention, found a larger effect size (g = −0.80).
Our findings show no differences in effect sizes between the first posttreatment assessment and the last follow-up, and both were similar to the overall effectiveness of the interventions. This is contrary to previous research on anxiety prevention, which showed that preventive effects have a tendency to diminish over time (Moreno-Peral et al., Reference Moreno-Peral, Conejo-Cerón, Rubio-Valera, Fernández, Navas-Campaña, Rodríguez-Morejón, Motrico and Bellón2017; Zalta, Reference Zalta2011). Nonetheless, this may be affected by the short duration of the follow-up assessments of the included studies.
In terms of therapeutic approach, no differences were found regarding the effectiveness of the interventions. There is mixed evidence on the role of the intervention model in relation to the effectiveness of the intervention, with some meta-analysis finding CBT to be associated with a major effect size when assessing apps to treat generalized anxiety symptoms as the primary outcome (Linardon et al., Reference Linardon, Torous, Firth, Cuijpers, Messer and Fuller‐Tyszkiewicz2024), while others find no significant differences in psychological interventions for the prevention of anxiety (Moreno-Peral et al., Reference Moreno-Peral, Conejo-Cerón, Rubio-Valera, Fernández, Navas-Campaña, Rodríguez-Morejón, Motrico and Bellón2017). This highlights the need for further studies testing other orientations and combined orientation interventions.
No significant effect was found in terms of intervention format. This is in line with the results of a meta-analysis on internet-based psychotherapeutic interventions conducted by Barak, Hen, Boniel-Nissim, and Shapira (Reference Barak, Hen, Boniel-Nissim and Shapira2008). Although video conferencing interventions differ from other digital approaches to anxiety prevention, particularly by providing real-time relational experiences that more closely mimic face-to-face sessions, they also share several key features and advantages with other digital formats. Like web-based or self-guided interventions, video therapy reduces geographical barriers, increases accessibility for individuals with limited mobility or those in underserved areas, and offers greater scheduling flexibility compared to in-person therapy (Nalongo-Bina, Reference Nalongo-Bina2024). Both formats rely on digital infrastructure and remote delivery, enabling scalable implementation and potential cost-effectiveness. Despite offering a more synchronous and familiar therapeutic environment, video conferencing still falls within the broader category of digital interventions. Evidence suggests that video therapy can be as effective as in-person therapy (Stubbings, Rees, Roberts, & Kane, Reference Stubbings, Rees, Roberts and Kane2013), whereas the effectiveness of other digital interventions may vary depending on the users level of engagement (Gan, McGillivray, Han, Christensen, & Torok, Reference Gan, McGillivray, Han, Christensen and Torok2021). Given the distinct characteristics of each format, further research is needed to refine comparisons between digital intervention formats and to better understand their specific strengths, limitations, and optimal contexts for application. This could help determine, for instance, whether certain delivery channels are more suitable for universal prevention of anxiety disorders, while others are better for indicated or selective prevention, or whether the effectiveness of different delivery formats varies across age groups.
Selective prevention appears to be the most common preventive intervention for anxiety, with nine RCTs included in this review, a tendency that does not differ from previous literature (Edge et al., Reference Edge, Watkins, Limond and Mugadza2023; Moreno-Peral et al., Reference Moreno-Peral, Conejo-Cerón, Rubio-Valera, Fernández, Navas-Campaña, Rodríguez-Morejón, Motrico and Bellón2017). The type of prevention, however, did not show a significant effect on the effectiveness of the intervention. This aligns with previous literature, which offers inconclusive evidence regarding the association between the type of prevention and effect size. Some meta-analyses have found no association (Deady et al., Reference Deady, Choi, Calvo, Glozier, Christensen and Harvey2017; Fisak, Richard, & Mann, Reference Fisak, Richard and Mann2011; Zalta, Reference Zalta2011), whereas others suggest that selective and/or indicated interventions tend to be more effective (Teubert & Pinquart, Reference Teubert and Pinquart2011). Conversely, some studies report universal prevention as more effective (Stockings et al., Reference Stockings, Degenhardt, Dobbins, Lee, Erskine, Whiteford and Patton2016).
Our findings on adolescents and young adults showed small effect sizes in studies exclusively with adolescents and in those in which the mean age corresponded to young adults. There are not many meta-analyses on the effectiveness of preventive interventions for anxiety exclusively in these populations. However, these results are consistent with those of previous meta-analyses focusing on the treatment of anxiety in similar samples. For example, a meta-analysis of 10 studies found an overall positive effect of digital interventions on reducing anxious symptoms (SMD = 0.440; 95% CI: 0.20–0.67; I2 = 82.9%) (Fischer-Grote, Fössing, Aigner, Fehrmann, & Boeckle, Reference Fischer-Grote, Fössing, Aigner, Fehrmann and Boeckle2024) in children, adolescents, and young adults. Although this effect size is slightly higher than that observed in our analysis, the high heterogeneity suggests considerable variability between studies, reinforcing the need to interpret these results in terms of intervention type, population, and methodological design. Unlike previous meta-analyses on the prevention of depression and anxiety (Campos-Paíno et al., Reference Campos-Paíno, Moreno-Peral, Gómez-Gómez, Conejo-Cerón, Galán, Reyes-Martín and Bellón2023; Moreno-Peral et al., Reference Moreno-Peral, Conejo-Cerón, Rubio-Valera, Fernández, Navas-Campaña, Rodríguez-Morejón, Motrico and Bellón2017) and on their treatment (Andrews et al., Reference Andrews, Basu, Cuijpers, Craske, McEvoy, English and Newby2018; Newby, Twomey, Yuan Li, & Andrews, Reference Newby, Twomey, Yuan Li and Andrews2016), the comparator had no significant effect on the results. Similarly, the effectiveness of the intervention was not affected by the duration of the program, which is consistent with findings from a recent meta-analysis by Seegan et al. (Reference Seegan, Miller, Heliste, Fathi and McGuire2023) of applications for anxiety and depression.
Even though the majority of the studies included had a high risk of bias, this variable proved to be nonsignificant in relation to the effectiveness of the interventions. This is in line with previous literature (Campos-Paíno et al., Reference Campos-Paíno, Moreno-Peral, Gómez-Gómez, Conejo-Cerón, Galán, Reyes-Martín and Bellón2023; Pauley et al., Reference Pauley, Cuijpers, Papola, Miguel and Karyotaki2023; Rigabert et al., Reference Rigabert, Motrico, Moreno-Peral, Resurrección, Conejo-Cerón, Cuijpers and Bellón2020), showing that risk of bias is not associated with the effect sizes of various types of interventions.
Similarly, the duration of the interventions had no impact on the effectiveness, which is consistent with a recent meta-analysis of applications for anxiety and depression (Seegan et al., Reference Seegan, Miller, Heliste, Fathi and McGuire2023). Finally, differences in effect sizes were not significantly related to the measurement method, the presence of guidance in the intervention, or the adherence rate.
Conclusions
The growing importance of developing and assessing the effectiveness of preventive interventions stems from the high incidence and costs associated with anxiety disorders. Recognizing digital interventions as a plausible solution paves the way for reducing the high disease burden linked to anxiety disorders.
This meta-analysis provides evidence supporting the preliminary effectiveness of digital interventions for the prevention of anxiety, although with a small effect and very low quality of evidence. There is a need to develop innovative digital interventions targeting anxiety prevention and to conduct new RCTs to assess their effectiveness using rigorous methodologies to ensure the validity of the results.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0033291725102262.
Acknowledgments
The authors would like to thank Dr. Calear, Dr. Fonseca, Dr. Lokman, and Dr. Monteiro for providing them with the data they required to include their studies in their MA. The authors would also like to thank Dr. Ebert and Dr. García-López for suggesting studies that could be included in this meta-analysis.
Author contribution
Cristina García-Huércano: conceptualization, investigation, formal analysis, writing – original draft. Patricia Moreno-Peral: conceptualization, methodology, investigation, writing – review and editing, project administration, supervision. Sonia Conejo-Cerón: conceptualization, investigation, writing – review and editing. Carmela Martínez-Vispo: conceptualization, investigation, writing – review and editing. Olaya Tamayo-Morales: writing – review and editing. Alberto Rodríguez-Morejón: writing – review and editing. Juan Ángel Bellón: conceptualization, writing – review and editing.
Funding statement
Funding for this study was provided by the Spanish Ministry of Health, the Institute of Health Carlos III, and the European Regional Development Fund ‘Una manera de hacer Europa’ and State Investigation Agency (grant references: PID2020-119652RA-100 and CP19/00056); as well as by the Prevention and Health Promotion Research Network ‘redIAPP’ (RD16/0007/0010) and Chronicity, Primary Care, and Health Promotion Research Network ‘RICAPPS’ (RD21/0016/0012). Funding for open access charge: Universidad de Málaga / CBUA.
Open access
