This study strictly adheres to the guidelines of the Cochrane Handbook for Interventions [30] and is reported according to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 guidelines [31] (the complete PRISMA checklist is available in Checklist 1). The research protocol has been registered in the PROSPERO system: CRD42024534389.

To comprehensively and systematically collect relevant literature, this study searched five major databases up to August 27, 2025, including Web of Science, PubMed, PsycINFO, Embase, and the Cochrane Controlled Trials Register. The search strategy combined controlled vocabulary (eg, MeSH terms) and free-text keywords using Boolean operators (“AND” and “OR”). The main search terms included the following: (“digital intervention” OR “internet intervention” OR “e-health” OR “m-health”) AND (“drug abuse” OR “substance use disorder” OR “illicit drugs”) AND (“psychotherapy” OR “psychoeducation” OR “psychodynamic”) AND (“randomized controlled trial” OR “single blind procedure” OR “random sample”). The complete search strategy for each database is provided in Multimedia Appendix 1.

Inclusion criteria were as follows: (1) Individuals aged 18 years and above with illicit drug use behavior. Illicit drugs refer to controlled substances used for nonmedical or nonscientific purposes, including but not limited to cannabis, cocaine, amphetamines, and opioids [1]. (2) Digital psychosocial intervention is the primary treatment. Operationally defined as structured psychological intervention primarily delivered through digital platforms, including mobile applications, web-based programs, or digital communication tools, with or without minimal human support. (3) The article must report sample size and dropout rates. (4) Randomized controlled trials. Exclusion criteria were as follows: (1) treatment involving only face-to-face therapy. (2) mixed samples with insufficient proportion of illicit drug users (less than 80%) or without independent subgroup data (eg, alcohol and tobacco users). (3) non-English studies. (4) unpublished reports, study protocols, meta-analyses, reviews, doctoral theses, or other gray literature.

To ensure the accuracy of literature screening, a dual-screening process was adopted. First, two researchers independently screened the titles and abstracts of retrieved literature to exclude those clearly not meeting inclusion criteria. Subsequently, the full texts of the literature were reviewed for further evaluation. Finally, manual searches were conducted on the reference lists of included studies and related reviews to identify additional studies meeting inclusion criteria. Any disagreements were resolved through discussion.

Two researchers independently extracted data using a predesigned data extraction form. Disagreements were resolved through discussion or consultation with a third researcher. This data extraction form has been piloted in some studies and adjusted according to the recommendations and structured framework of the GRADE manual. For articles that met the inclusion criteria but lacked important data, we contacted the corresponding author via email, and studies that could not provide sufficient data to calculate effect sizes were excluded.

To assess the bias risk of the included studies, two researchers independently scored each study in five aspects using the revised Cochrane Risk of Bias tool ROB 2.0 [45]: randomization process, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of the reported result. Any disagreements were resolved through discussion.

We used Comprehensive Meta-Analysis software (CMA 4.0) to synthesize dropout rates across studies [46]. For each trial, dropout counts and total sample sizes were extracted separately for the intervention and control groups, from which group-specific dropout proportions were calculated. To stabilize variances and account for the bounded nature of proportions, these proportions were transformed into logit event rates with corresponding standard errors, which served as the primary effect size metric. Pooled estimates were calculated separately for intervention and control groups and subsequently back-transformed into raw proportions and expressed as percentages for interpretability, an approach that has been widely applied in meta-analyses of proportion-type outcomes [47]. Subsequently, between-study heterogeneity was examined using the Q statistic [48,49] and quantified with the I² statistic [50]. Given the significant heterogeneity among included studies in outcome measures and moderators [51,52], all analyses were conducted under a random-effects model [53]. Publication bias was assessed using funnel plots, Egger’s, Duval and Tweedie’s trim and fill, and Classic fail-safe N tests [54], while sensitivity analyses were conducted to evaluate the robustness of the results. To explore potential influencing factors, meta-regression and subgroup analyses were further employed to examine the association between moderators in the intervention group and dropout rate.

After screening relevant articles based on predefined inclusion and exclusion criteria, a total of 41 studies were finally included (see Figure 1), involving 9693 participants with an age range of 19 to 50 years. The selection characteristics of the included studies are shown in Table 2. The studies included 82 intervention groups, with a total of 48 dropout rate data points, including 18 posttreatment dropout rates and 30 follow-up dropout rates, showing different data results between the two measurement points.

The risk of bias in the included studies was assessed using the Cochrane Risk of Bias tool (ROB 2.0). Detailed results and percentage plots are presented in Multimedia Appendix 2. The results showed that approximately 90% of the included studies had a low risk in terms of the randomization process (D1), measurement of the outcome (D4), and selection of the reported result (D5). Approximately 55% of the included studies had some concerns about deviations from intended intervention (D2). About 50% of the studies had a high risk of missing outcome data (D3), which is a key focus of our research.

Follow-up analysis of the intervention group was based on 30 studies, with an average dropout rate of 28.2% (95% CI 0.19‐0.39) (Figure 5), while the rate in the control group was 27.8% (95% CI 0.20‐0.37). However, heterogeneity testing again indicated high variability among the studies (Q=1293.13, df=29, P=.000, I²=98%). Further analysis revealed that the variance of the true effect size reached 1.79 (logit units) with a standard deviation of 1.34 (logit units).

Meta-regression analysis and subgroup analysis (Table 4) revealed that this extreme variability is primarily due to 4 variables among three types of characteristics: (1) participant characteristics: dropout rate showed a negative correlation with single status (OR 0.95, 95% CI 0.91‐0.99; P=.01); (2) clinical characteristics: significantly positive correlation with baseline medication frequency (OR 1.18, 95% CI 1.05‐1.32; P=.004); (3) intervention characteristics: participants recruited via website showed a positive correlation with dropout rate (OR 5.74, 95% CI 1.85‐17.76; P=.002), while participants recruited via campus showed a negative correlation with dropout rate (OR 0.28, 95% CI 0.12‐0.66; P=.003); The association between the degree of digitalization and dropout rates varied depending on whether studies with unreported digitalization status (not reported [NR] group) were included. When all studies, including the NR group, were analyzed, the overall model reached statistical significance (Q=28.13, df=2, P<.001), with the NR group showing a strongly significant negative effect (OR 0.16, 95% CI 0.06‐0.41; P<.001). However, when the NR group was excluded and only studies explicitly reporting “fully digital” or “partially digital” were considered, the results were not statistically significant (Q=0.24, P=.62). The other 32 factors showed no significant correlation with dropout rate.

The funnel plot showed some studies beyond the expected range (see Figure 6). Combined with Egger test results (P=.023), publication bias was further confirmed. After trimming the six missing studies on the right side, the effect size changed from 28% to 37% after correction, without crossing the clinical threshold. Further leave-one-out sensitivity analysis showed that 1244 unpublished studies would need to be included to make the current results statistically insignificant, supporting the stability of the research conclusion. Meanwhile, sensitivity analysis (see Figure 7) indicated that the results of this study were robust and not dependent on individual studies.

This meta-analysis systematically evaluated the treatment retention effect of digital psychosocial interventions among adult illicit drug users. The pooled dropout rate was 22%, slightly lower than the approximately 30% reported for face-to-face psychosocial interventions [37], suggesting potential advantages of digital formats for treatment retention. Nevertheless, the substantial heterogeneity across studies limits the generalizability of these findings. Dropout rates also varied across settings and populations. For instance, adults with co-occurring severe mental disorders and substance use had an average dropout of 27% [95], whereas clinical samples of opioid users showed rates as high as 41% [73]. Beyond dropout, adherence constitutes another key indicator of engagement, with evidence showing that participants completed, on average, 60% of digital intervention modules, and only about half finished the full program [96]. Taken together, these results underscore the importance of considering both dropout and adherence when evaluating intervention effectiveness. Building on this, our moderator analyses further revealed complex interactive effects. To ensure clarity, we retained the classification system established during data extraction, presenting results separately across four major categories of characteristics as well as between short-term and longest intervention stages.

At the posttreatment stage, dropout was significantly influenced by participants’ demographic, intervention, and clinical characteristics. Regarding demographics, unemployment did not predict dropout, whereas higher employment was unexpectedly associated with greater attrition. This suggests that unstable or high-intensity work may interfere with regular participation. In addition, the short-term income from employment may reduce some patients’ motivation for treatment, especially when symptoms temporarily improve, leading them to discontinue prematurely due to “feeling better” [97]. For intervention characteristics, intervention frequency showed a negative correlation with dropout, indicating that more frequent contact may help consolidate behavior change, strengthen the therapeutic alliance, and enhance commitment [98-100]. Future studies should explore the optimal intervention frequency under different conditions [101], balancing treatment intensity with patient burden [102].

The results of baseline clinical characteristics indicated that both baseline clinical diagnosis and baseline cocaine use were significantly positively associated with dropout rates. Specifically, patients with a clear baseline diagnosis were at greater risk of dropout due to challenges such as dependency, withdrawal symptoms, and impaired cognitive or emotional functioning [42]. For this population, the integration of adjunctive pharmacological or behavioral therapies is recommended to reduce dropout [103]. Furthermore, consistent with previous findings [32], participants with baseline cocaine use were more likely to discontinue treatment. Cocaine use disorder is often closely linked to impulsive behavior and diminished adherence [37]. These substance-specific risks highlight the importance of developing differentiated intervention strategies tailored to distinct types of substance use in future research [104]. Nevertheless, the small sample size of drug-use subgroups (k≤5) remains a limitation, which could be addressed through multi-institutional collaborations to expand subgroup samples.

During the longest follow-up, dropout was significantly influenced by demographic, clinical, and intervention characteristics. In demographics, a higher proportion of single participants was linked to lower dropout. This may be related to reduced drug exposure in family environments [105-107]. In addition, single participants with low social support were more likely to continue seeking health information online. Future research could involve non–drug-using significant others in monitoring the intervention process and integrate peer support modules [108]. In clinical characteristics, participants with higher baseline drug use frequency faced markedly greater dropout risk. This finding is consistent with recent studies [109]. For this high-risk group, we recommend the implementation of multistage intensive intervention programs [110], together with the development of immediate-response modules (eg, crisis management tools, real-time consultation functions) to reduce early dropout [81].

In terms of intervention characteristics, participants recruited through websites exhibited higher dropout rates, whereas those recruited from campus showed lower dropout rates. This may be explained by the lack of intensive treatment services typically provided in clinical settings, as well as the relative stability of campus environments [6,111]. Based on this finding, we recommend adopting a mixed online–offline recruitment strategy [112]. In addition, intervention content should be optimized for online recruits [113], including simplifying operational procedures, providing regular reminders, and offering personalized feedback. The study also analyzed the association between the degree of digitalization and dropout rates. During data processing, studies that did not report their digitalization status (23/30, 77%) were categorized separately as a “Not reported” group for analysis rather than being directly excluded. The analysis revealed a significant association: compared to the nonsignificant negative correlation between fully digital interventions and dropout rates, interventions with unreported digitalization status showed a significant negative correlation, while non-fully digital interventions demonstrated a significant positive correlation with dropout rates. However, the reliability of these subgroup comparisons is constrained by the prevalent issue of poorly reported data. When we excluded the “Not reported” studies and repeated the analysis, no significant differences were found between fully digital and partially digital interventions. This suggests that the initial findings were likely confounded by nonrandom reporting bias rather than reflecting true effects, making definitive evaluation difficult. Therefore, these results primarily highlight the urgent need for future research to standardize the reporting of specific digital intervention details in order to more reliably explore the role of digitalization degree and human support in improving retention rates [114].

This study systematically evaluated the dropout rate and its predictive factors among adult illicit drug users in digital psychological interventions, thereby addressing a critical research gap in the field. Unlike previous studies that primarily focused on demographic characteristics, this analysis incorporated multidimensional predictive variables—including clinical features, therapist-related factors, and intervention characteristics—to establish a more systematic theoretical framework. The identification of eight key predictive factors provides valuable insights for personalized interventions, guiding the development of tailored digital tools for patients at high risk of dropout. Optimization strategies derived from this evidence are expected to substantially reduce dropout rates and enhance intervention effectiveness [30].

This study has several limitations. First, few of the included trials provided detailed information on software quality or reasons for dropout, which limited our ability to assess the reasons why participants stopped treatment [115]. Future studies could combine machine learning methods to predict dropout risk [116] and use participant-centered questionnaires to collect data on perceived barriers. Previous research [117-120] emphasized common reasons for dropout, including technical difficulties, lack of engagement, and perceived ineffectiveness of the intervention. Collaboration with software engineers may help optimize the digital experience and reduce technical-related attrition [121]. Additionally, methodological improvements, such as combining intention-to-treat analysis with run-in phase dropout screening [15,122], may provide more refined methods for managing early dropout.

Second, most of the digital interventions included in the studies adopted limited forms, such as videos, virtual characters, or text messages, and lacked interactive features. Incorporating gamification elements may enhance user engagement [123], especially when personalized to individual preferences [86,121]. Emerging evidence suggests that well-designed therapeutic video games can improve cognitive and mental health outcomes [121], even inducing neurobiological changes, including alterations in white matter microstructure [124-128].

Finally, many studies did not clearly report key methodological details, such as the degree of digitalization or level of human support. Although we conducted analyses including and excluding the “Not reported” category, the lack of such information led to inconsistent findings, preventing definitive conclusions regarding the impact of digitalization on dropout rates. Future studies should standardize reporting of intervention details, including digitalization and human support, to better understand active components and optimize strategies [121,129]. Another limitation is the high heterogeneity in the meta-analysis (I²>90%), which may reduce robustness. Despite sensitivity and moderator analyses, some variability remained unexplained, suggesting pooled effects may not apply equally across interventions, populations, or outcomes. Future research should adopt rigorous methodologies, including detailed reporting, preregistration, data sharing, and large-scale RCTs. Individual participant data meta-analyses can further clarify subgroup effects and sources of heterogeneity, improving generalizability [130].

In summary, this meta-analysis systematically examined dropout rates and their predictive factors in digital psychosocial interventions for adult illicit drug users, aiming to provide a comprehensive picture of the research landscape in this field. The results indicate that both short-term and long-term adherence to interventions are characterized by considerable complexity. In the short term, dropout rates were primarily associated with employment status, baseline clinical diagnoses, baseline primary substance use, and intervention frequency. Over longer follow-up periods, marital status, baseline drug use frequency, and recruitment source emerged as key predictors. These findings suggest the need for further investigation into factors that contradict common assumptions or remain insufficiently reported in the literature, as well as greater standardization in the design, measurement, and reporting of randomized controlled trials to improve research quality. Moreover, more attention should be given to tailoring interventions for specific populations, particularly through the design of intervention functions and modules. Continued exploration in these areas will contribute to better supporting patients’ long-term recovery.