Predictors of study dropout in cognitive-behavioural therapy with a trauma focus for post-traumatic stress disorder in adults: An individual participant data meta-analysis

Abstract

Background

Available empirical evidence on participant-level factors associated with dropout from psychotherapies for post-traumatic stress disorder (PTSD) is both limited and inconclusive. More comprehensive understanding of the various factors that contribute to study dropout from cognitive-behavioural therapy with a trauma focus (CBT-TF) is crucial for enhancing treatment outcomes.

Objective

Using an individual participant data meta-analysis (IPD-MA) design, we examined participant-level predictors of study dropout from CBT-TF interventions for PTSD.

Methods

A comprehensive systematic literature search was undertaken to identify randomised controlled trials comparing CBT-TF with waitlist control, treatment-as-usual or another therapy. Academic databases were screened from conception until 11 January 2021. Eligible interventions were required to be individual and in-person delivered. Participants were considered dropouts if they did not complete the post-treatment assessment.

Findings

The systematic literature search identified 81 eligible studies (n=3330). Data were pooled from 25 available CBT-TF studies comprising 823 participants. Overall, 221 (27%) of the 823 dropped out. Of 581 civilians, 133 (23%) dropped out, as did 75 (42%) of 178 military personnel/veterans. Bivariate and multivariate analyses indicated that military personnel/veterans (RR 2.37) had a significantly greater risk of dropout than civilians. Furthermore, the chance of dropping out significantly decreased with advancing age (continuous; RR 0.98).

Conclusions

These findings underscore the risk of premature termination from CBT-TF among younger adults and military veterans/personnel.

Clinical implication

Understanding predictors can inform the development of retention strategies tailored to at-risk subgroups, enhance engagement, improve adherence and yield better treatment outcomes.

WHAT IS ALREADY KNOWN ON THIS TOPIC.

WHAT THIS STUDY ADDS

• This study provides evidence that military personnel and veterans are at a significantly higher risk of dropping out of CBT-TF compared with civilians.

• Additionally, the likelihood of dropout decreased with advancing age.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE, OR POLICY

• The findings highlight the need for tailored retention strategies for younger adults and military personnel/veterans in CBT-TF to enhance engagement and adherence, improving treatment outcomes for these at-risk groups.

Background

Post-traumatic stress disorder (PTSD) is considered a global public health priority due to the significant burden it places on individuals and society.¹ Several psychotherapies have demonstrated efficacy for PTSD with moderate to large effect sizes but high dropout rates are concerning.²

A recent meta-analysis reported dropout rates ranging from 14% to 22% across psychotherapy trials for PTSD.³ Studies focusing on military veterans have reported considerably higher dropout rates, ranging from 31% to 39%.4,6

Psychotherapeutic interventions embedded in a framework of cognitive-behavioural therapy with a trauma focus (CBT-TF) have been extensively researched and proven highly effective for PTSD.⁷ CBT-TF is defined as a range of therapeutic approaches that are designed to assist individuals with PTSD and early traumatic stress symptoms by targeting and modifying thoughts, beliefs and behaviours.⁸ By employing exposure, cognitive restructuring and anxiety management techniques, CBT-TF aims to create a safe environment for patients to confront their traumatic memories and modify dysfunctional thoughts and emotions.⁸ Among these are prolonged exposure therapy,⁹ cognitive processing therapy,¹⁰ narrative exposure therapy¹¹ and brief eclectic psychotherapy,¹² which share common elements.

Predictors of dropout from CBT-TF trials have been inconsistent and underpowered. One study found male gender predicted dropout,¹³ whereas another found female gender did.¹⁴ Dropout risk decreased with advancing age in some studies^{6 15} but increased in another.¹⁶ Lower education^{17 18} and shorter military service duration⁴ are associated with higher dropout.

Some evidence suggests that marital status,^{13 17} medication use¹⁹ and exposure to multiple traumas¹⁹ are not related to dropout. Some studies found higher baseline PTSD severity associated with higher dropout^{20 21}; others have not.^{19 22} Similarly, elevated baseline depression scores predicted higher dropout in a few studies^{21 23} but not in others.^{17 18} Findings on comorbid alcohol and substance use disorder and dropout are similarly inconsistent.²⁴

Randomised controlled trials (RCTs) often have small samples and, consequently, low statistical power to detect meaningful associations between predictors and outcomes.²⁵ This can undermine the value of predictor analyses in individual studies, ultimately rendering them unreliable. Aggregating data from RCTs typically provides only summarised study-level information. Individual Participant Data Meta-Analysis (IPD-MA) combines harmonised data from multiple studies into a single dataset, offering enhanced statistical power for more precise estimates and better detection of significant associations.²⁵ The enhanced statistical power allows more precise predictor estimates and improves the detection of significant associations. IPD-MA is gaining popularity with the growing focus on data sharing.²⁶

Objective

The primary objective of this study was to investigate sociodemographic and clinical predictors of study dropout in CBT-TF interventions for PTSD using an IPD-MA approach.

Hypotheses

Based on the preponderance of findings from prior individual studies, we hypothesised that higher risk of dropout would be associated with (1) military service/veteran (vs civilian) status, (2) lower educational attainment and (3) younger age. Analyses of other predictors reported in individual investigations were considered exploratory, given the mixed results for those variables across those studies.

Methods

Eligibility criteria

Study inclusion was restricted to RCTs comparing a CBT-TF intervention to any comparison group (eg, waiting list control, treatment-as-usual or another psychological intervention). We excluded pharmacotherapy-based comparison groups. The interventions were required to be individual therapy and delivered in person. Studies comprised adults (>17 years old) with PTSD. In each study, a minimum of 70% of the study sample had to be diagnosed with PTSD according to any version of the Diagnostic or Statistical Manual of Mental Disorders²⁷ or the International Classification of Diseases.²⁸ Due to the substantial prevalence of comorbidity in persons with PTSD, we placed no restriction on the co-occurrence of psychiatric or physical conditions.²⁹

Search strategy and study identification

An existing psychotherapy trials database for PTSD, which included studies published from conception until 1 May 2018, was obtained from the Cardiff University Traumatic Stress Research Group. We subsequently updated the search to include articles published until 11 January 2021. The screened academic databases included PubMed, EMBASE, PsycINFO, PTSDpubs and CENTRAL (refer to online supplemental appendix 1). Searches incorporated terms related to PTSD, trauma and psychotherapy. Authors contacted for their participant-level datasets were asked if they had any additional studies that might meet study inclusion. Furthermore, we searched past systematic reviews for unidentified articles that met our study inclusion criteria.^{30 31} The titles and abstracts of all hits identified in the academic search were independently examined by two reviewers (SW and DP for the initial screening and SW and ABW for the update). Both reviewers then independently screened the included full texts. A third team member (MS) resolved any uncertainties surrounding study inclusion. This review was reported in line with the Preferred Reporting Items for Systematic Review and Meta-Analyses (PRISMA) IPD statement.³² Refer to online supplemental appendix 2 for the PRISMA study selection process.

Data collection

Primary authors of studies that met the final inclusion criteria were emailed to request access to their anonymised, participant-level datasets. We used the email contact details available in the study’s publication and a more recent email address when available online on academic websites or in more recent publications. We sent six reminder emails before reaching out to another two coauthors.

Data sharing and storage of the participant-level data complied with the European General Data Protection Regulation (Regulation (EU) 2016/679).

Data extraction

The outcome of interest was study dropout from CBT-TF s for PTSD. A list of the definitions of these CBT-TF protocols is provided in online supplemental appendix 3. Participants were considered study dropouts if they did not complete the post-treatment assessment, allowing us to retain data from the maximum number of available participants. This made it possible to keep the dropout definition consistent across the included studies. Post-treatment assessment refers to the assessment conducted directly after completion of the treatment protocol. Putative predictors included years of age (continuous), years of education, gender (male or female), marital status (married or unmarried), divorced versus other (single/married/cohabiting), using psychotropic medication (no or yes), PTSD total severity score, intrusion severity score, avoidance severity score, hyperarousal severity score, comorbid substance use problem (abuse or dependence including alcohol; no or yes), comorbid major depressive disorder (MDD; no or yes), psychiatric comorbidity (substance use, mood or anxiety disorder; no or yes) and population (military personnel/veterans or civilian).

Risk of bias

The Cochrane Risk of Bias 2 tool³³ was used to assess the risk of bias within the available studies included in these analyses. Two reviewers (DP and SW) rated the risk of bias related to the randomisation process (D1), deviations from the intended intervention (D2) and measurement of the outcome (D4). The domains missing outcome data³ and selection of the reported result⁵ were not relevant because the analyses made use of participant-level data. The risk of bias for each domain was rated as low, high or some concerns.

Data analysis

Data were only extracted for the CBT-TF intervention groups because we focused on examining predictors of dropout within these intervention groups. Data analysis was conducted in STATA V.17.³⁴ Once having received the datasets, we extracted sociodemographic and clinical participant-level characteristics. Datasets were then harmonised by converting key variables to a uniform format. Baseline PTSD scores were standardised (transformed into z scores) within each available study before combining the individual datasets into one large, pooled dataset. Sociodemographic data were inconsistently reported across trials. Imputation of data that were completely missing in a study would require estimates based entirely on data from other samples that could not be assumed to be comparable to those from which the data were missing. Therefore, the probability of data being missing in studies for which a sociodemographic variable was not reported could not be assumed to be related to the outcome variable, and listwise deletion was used rather than imputation.³⁵

We analysed the effects of predictors on dropout accounting for the clustering of participants within studies using multilevel analyses. We conducted the analyses in three steps: (1) we conducted a series of multilevel bivariate analyses to assess the RR of each factor at a time (the so-called ‘bivariate model’), (2) we repeated the analyses entering all factors simultaneously into the multilevel multivariate model (the ‘complete model’) and (3) we simplified the complete model, retaining only those factors in the model that were statistically significant (the ‘parsimonious model’).

Post hoc power analyses were conducted to assess the statistical power of the predictor analyses. The power calculations were based on the sample sizes, effect sizes (expressed as the natural log of the relative risk), number of included studies and a significance level of 0.05 for each predictor.

Sensitivity analyses were run to examine the robustness of our findings and potential interaction effects. First, we explored the effect of removing any studies rated to have a high risk of bias on any domain. We then examined the risk of dropping out by age category (18–29, 30–39, 30–39, 40–49, 50–59, 60–69, 70 years and older). We also re-examined the population variable while independently controlling for baseline PTSD symptom severity as well as by publication year.

We explored the possible interaction effect of MDD. We did not include MDD as a predictor in main bivariate and multivariate analyses due to the small sample. We first ran a multilevel bivariate model for MDD. Second, we included it in the parsimonious model. Finally, we also looked at the risk of dropping out for MDD while controlling for age (continuous) and population independently.

Findings

Study selection

A total of 81 eligible studies (n=3330) were identified. Fifty-six eligible datasets were not accessible and could not be included in this IPD-MA (see online supplemental appendix 4 for the list of eligible studies).

Study characteristics

Participant-level data from 28 CBT-TF intervention groups from 25 studies(n=823) were pooled. Of these studies, three provided data from two different CBT-TF intervention groups for analysis, and both were included in the current study. Study characteristics and intervention details are presented in online supplemental appendix 5.

Overall, 221 (27%) participants of 823 dropped out. Of 581 civilians, 133 (23%) dropped out, as did 75 (42%) of 178 military personnel/veterans. The mean (SD) number of treatment sessions completed for participants who dropped out of the study without completing the post-treatment assessment was 4.91 (4.28; n=164) compared with 13.66 (5.16; n=348) for study completers. On average, 13 treatment sessions were offered between the baseline and first post-treatment assessment. Participants who dropped out attended 35% of treatment sessions offered, whereas study completers attended 88% of treatment sessions offered. A summary of participant characteristics is presented in online supplemental appendix 6.

Risk of bias

Some concerns arose in (D1) randomisation process and (D2) deviations from intended interventions. In D1, 10 studies were rated with some concerns because of baseline differences. Bias in D2 was attributed to some uncertainty due to slight deviations in protocol adherence or insufficient information in seven studies. No studies were rated as high risk of bias on these two domains. All studies were rated low risk of bias on (D4) measurement of the outcome because the outcome was study dropout (see online supplemental appendix 7 for risk of bias ratings).

Predictors of dropout in CBT-TF treatment for PTSD

The results of the multilevel bivariate, multivariate and parsimonious analyses of participant-level predictors and dropout are presented in table 1. Results from bivariate analyses indicated that the chance of dropping out decreased with increasing age (Risk ratio [RR] 0.98; p=0.010). The chance of dropping out was significantly higher for military personnel/veterans (RR 1.94; p=0.045) than for civilians (population).

Table 1. Baseline predictors of dropout in CBT-TF.

	Multilevel bivariate model				Multivariate modelN (k)=36311			Parsimonious modelN (k)=73022
Participant level predictors	n (k)	RR	95% CI	P value	RR	95% CI	P value	RR	95% CI	P value
Age (continuous)	794 (25)	0.98	0.97 to 1.00	0.010*	0.98	0.96 to 0.99	0.003*	0.98	0.97 to 0.99	0.001*
Education (years)	429 (10)	0.97	0.91 to 1.04	0.386	–	–	–	–	–	–
Male gender	822 (28)	0.95	0.66 to 1.37	0.778	1.10	0.66 to 1.83	0.726	–	–	–
Marital status (married)	535 (19)	1.20	0.83 to 1.72	0.328	1.15	0.80 to 1.65	0.463	–	–	–
Divorced vs other (single/married/cohabitating)	548 (19)	0.85	0.56 to 1.29	0.439	–	–	–	–	–	–
Using psychotropic medication	528 (15)	1.09	0.76 to 1.56	0.627	–	–	–	–	–	–
PTSD total severity	793 (27)	1.04	0.91 to 1.20	0.539	1.03	0.86 to 1.23	0.769	–	–	–
PTSD intrusion severity	499 (20)	1.15	0.96 to 1.37	0.119	–	–	–	–	–	–
PTSD avoidance severity	485 (19)	0.93	0.78 to 1.11	0.413	–	–	–	–	–	–
PTSD hyperarousal severity	486 (18)	0.98	0.82 to 1.16	0.782	–	–	–	–	–	–
Substance use problem (abuse/dependence)	434 (12)	1.01	0.59 to 1.73	0.962
Psychiatric comorbidity (anxiety, depression, substance use problem)	566 (17)	1.34	0.82 to 2.22	0.246	1.01	0.55 to 1.88	0.964	–	–	–
Population (military personnel/veterans)	759 (25)	1.94	1.01 to 3.73	0.045*	1.87	1.12 to 3.12	0.017*	2.37	1.51 to 3.71	0.000*

^*p < 0.05

knumber of included study armsnnumber of participantsPTSDpost-traumatic stress disorderRRrate ratio

Other predictors were not significant. Under the multivariate model, age (RR 0.98; p=0.003) and population (RR 1.87; p=0.017) remained statistically significant predictors of study dropout. Under the parsimonious model, both age (RR 0.98; p=0.001) and population (RR 2.37; p=0.000) remained statistically significant predictors of study dropout.

We explored the effect of removing the one study rated with a high risk of bias.³⁶ Results from this multilevel multivariate analysis found both age (RR 0.98; 95% CI 0.97 to 0.99; p=0.001) and population (RR 2.37; 95% CI 1.51 to 3.71; p=0.000) remained statistically significant predictors of dropout. Post hoc power analyses uncovered varying degrees of statistical power for the predictors examined. Notably, continuous variables showed lower power, while dichotomous variables demonstrated higher power (see online supplemental appendix 8).

Table 2 presents the risk of dropout within specific age cohorts, where participants aged 18–29 years served as the reference group for the analysis. Compared with participants in the 18– 29 year age group, the 30–39 (p=0.399), 40–49 (p=0.264) and 50–59 (p=0.098) year age group did not differ significantly. The risk of dropping out only differed significantly between the 18–29 year age group and the 60–69 (p=0.043) year age group. There was a slight increase in risk of dropping out for participants aged 70 years and older (p=0.186).

Table 2. Risk of dropping out by age category (n=794, k=25).

Variable	RR	95% CI	Z	P value
18–29 vs 30–39 years	0.86	0.61 to 1.22	−0.84	0.399
18–29 vs 40–49 years	0.81	0.56 to 1.17	−1.12	0.264
18–29 vs 50–59 years	0.68	0.44 to 1.07	−1.65	0.098
18–29 vs 60–69 years	0.41	0.17 to 0.97	−2.02	0.043*
18–29 vs 70+ years	0.38	0.09 to 1.60	−1.32	0.186

^*p < 0.05

RRrate ratio

When controlling for baseline PTSD symptom severity, population (RR 1.94; 95% CI 1.04 to 3.64; p=0.038) remained a statistically significant predictor, as well as when controlling for publication year (RR 2.01; 95% CI 1.02 to 3.95; p=0.043).

We explored the risk of dropout between patients with and without baseline comorbid MDD. We did not include comorbid MDD in the main predictor analyses due to the small sample (n=360). The results of the sensitivity analysis appear in table 3. Multilevel bivariate analysis found a higher risk of dropout in patients with comorbid MDD compared with those without (RR 1.91; p=0.015). When including comorbid MDD in the parsimonious model, both age (RR 0.97; p=0.004) and population (RR 1.87; p=0.003) remained significant, but comorbid MDD no longer did (RR 1.55; p=0.110). When we removed population, both age (RR 0.98; p=0.006) and comorbid MDD (RR 1.77; p=0.032) were significant. However, when we removed the age variable, only population (RR 1.73; p=0.009) was significant, and MDD (RR 1.70; p=0.052) was marginally significant.

Table 3. Sensitivity analysis for the effects of baseline comorbid MDD.

	Bivariate*n (k)=360 (9)			Parsimonious*n (k)=321 (8)			Multivariate* excluding populationn (k)=359 (9)			Multivariate* excluding ageN (k)=322 (8)
	RR	95% CI	P value	RR	95% CI	P value	RR	95% CI	P value	RR	95% CI	P value
Comorbid MDD†	1.91	1.14 to 3.21	0.015*	1.55	0.91 to 2.65	0.110	1.77	1.05 to 2.99	0.032§	1.70	0.99 to 2.90	0.052
Age‡	–	–	–	0.97	0.96 to 0.99	0.004§	0.98	0.96 to 0.99	0.006§	–	–	–
Population	–	–	–	1.87	1.24 to 2.84	0.003§	–	–	–	1.73	1.15 to 2.62	0.009§

^*Models were multilevel.

^†Major Ddepressive Ddisorder;.

^‡cContinuous.

^§p < 0.05

RRrate ratio

Discussion

This study examined factors predicting study dropout in CBT-TF interventions for PTSD using data from 25 RCTs. The overall dropout rate (27%) is consistent with previous findings but significantly higher among military personnel and veterans (42%) compared with civilians (23%).³

Consistent with earlier research, this study identified limited predictors of treatment dropout.^{24 37} This aligns with the broader literature, which has found that commonly examined predictors have limited utility in explaining why patients discontinue therapy.^{24 37}

Bivariate and multilevel analyses indicated a significantly greater likelihood of study dropout among military personnel and veterans than civilians. This result remained robust after controlling for PTSD severity, publication year and comorbid MDD. Post hoc power analysis confirmed the strong predictive value of population. Higher dropout in military personnel may stem from the unique challenges of combat trauma, lack of support, therapist factors or comorbid conditions.^{38 39}

Initial analysis showed individuals with comorbid MDD were at a higher risk of dropout. This suggests that patients with PTSD with comorbid MDD may face additional challenges. However, when controlling for age and population, the effect was non-significant. Caution is needed in interpreting these results due to the small MDD subgroup, and further replication in future studies is necessary.

Findings showed that dropout risk decreased with increasing age. While this predictor yielded low power, it remained significant in all analyses, suggesting robustness. Sensitivity analysis revealed no significant differences in dropouts among participants aged 30–59 years compared with those aged 18–29 years. There was a significant drop in risk among individuals aged 60–69 years, with a small uptick after 70 years, which might be attributed to relatively few participants in this age category. However, we did observe a decreasing trend in p values across age groups, from 30 to 70 years. These findings align with PTSD^{4 6 17 18} and depression research.⁴⁰ Younger adults may face more challenges in managing various life responsibilities such as childcare, education and work demands, obstacles that might affect their ability to complete treatment or effectively cope with treatment-related distress.²

None of the other baseline covariates achieved statistical significance in predicting dropout. Removing studies rated with high risk of bias did not substantially alter the results.

Statistical power was high for variables like marital status, being divorced, PTSD intrusion severity, psychiatric comorbidity and population. However, power was low for predictors, such as age, education, PTSD severity, avoidance severity, hyperarousal severity and substance use problems, reducing our ability to detect significant effects. Future RCTs with larger sample sizes are required to increase power in pooled analyses and clarify these relationships.

On average, participants who dropped out in this study only attended 35% of available treatment sessions, whereas post-treatment assessment completers attended 88% of available treatment sessions. This aligns with previous research that found most participants who dropped out did so before the halfway point of the intervention protocol.^{2 41} While some participants may drop out due to symptomatic improvement, it is unlikely that only a third of the treatment course would suffice.⁴¹

Interpreting these findings requires acknowledging the study’s limitations. A common issue in IPD-MA is inconsistent reporting across studies. Several trauma-related predictors (eg, frequency, severity, chronicity of trauma; childhood vs adult-onset trauma) were not included due to inconsistent or absent data. Most studies provided no data on living status, although marital status was available. In contemporary society, these concepts often diverge due to changing social norms, financial factors or personal choices. Similarly, most studies measured sex/gender with limited options preventing analysis of sexual orientation.

Including pharmacotherapy comparisons in our meta-analysis would have added value, but due to the small number of relevant studies, we excluded them during the design phase of this project to ensure a focused and robust analysis. There was an insufficient number of the different CBT-TF treatment types to compare the dropout rate among them reliably. While examining participant-level predictors in conventional meta-analysis is not possible, this design would be better suited for examining the risk of dropout among different CBT-TF protocols in all the eligible CBT-TF studies. Finally, we only had MDD diagnosis data for a limited proportion of the total sample, restricting our ability to investigate its effect on dropout.

Another area for improvement is the variation in randomisation timing across studies. Some studies randomised participants after the baseline assessment, counting those who missed the first session as dropouts. Others deferred randomisation until after the first session, potentially lowering dropout rates.

Most studies were not designed to evaluate dropout reasons and rarely reported why patients discontinued. Even when asked, patients may not disclose the true reasons. As a result, this study could not examine dropout reasons. Similar to other IPD-MA studies, obtaining participant-level data provided a significant challenge. Despite the substantial time and effort spent on data collection, lost datasets, data storage in outdated formats (eg, paper charts or floppy disks), lack of willingness to share, restrictive institutional policies and other regulatory hurdles, were impediments. As a result, many eligible studies could not be included.

While dichotomous variables demonstrated greater power than continuous variables, overall power for detecting significant moderators was still constrained despite combining participant-level data from 28 CBT-TF intervention groups. This highlights the need for larger sample sizes within individual trials.

This study has notable strengths that contribute to its reliability and comprehensiveness. First, the relatively large overall sample size compared with individual RCTs and conventional meta-analyses is a strength. By using participant-level data, we could delve into individual-level factors, such as age, gender and other sociodemographic characteristics. While we could not investigate all desired variables, employing an IPD-MA approach enabled us to conduct participant-level predictor analyses with statistical power tailored for such investigations.

Further research is required to explore predictors of non-trauma-focused therapies for PTSD, factors contributing to dropout in military veterans/personnel and younger adults, and therapist factors that may be associated with study dropout including, training in retention and logistical factors (eg, in-person therapy vs telehealth).

Our definition of dropout aimed to maximise available participant data for a more comprehensive analysis. However, this approach may not fully account for all true completers. Differences in dropout definitions across studies present a significant challenge when seeking to make direct comparisons, complicating efforts to synthesise findings and draw broader conclusions. Therefore, our study highlights the urgent need for standardised definitions of dropout in future PTSD treatment studies to enhance the comparability and reliability of results.

Clinical implications

Given the higher dropout rates among military personnel and veterans, clinicians may consider additional support measures tailored to their unique challenges. Our findings also suggest that younger adults are more likely to drop out of CBT-TF. Clinicians should explore age-specific engagement strategies to address the particular life challenges that younger adult’s face, such as work, education and childcare responsibilities.

These results highlight the importance of adequate power in study design to identify meaningful predictors of treatment dropout which can inform better intervention strategies. Despite several limitations, this study contributes to a more granular understanding of the factors associated with study dropout in studies investigating CBT-TF interventions for PTSD.

Data availability statement

Data may be obtained from a third party and are not publicly available.