A Meta-Analytic Review of Prolonged Exposure for Posttraumatic Stress Disorder

We selected well-controlled randomized trials of PE for PTSD using a comprehensive search strategy. We searched the following databases: PsycINFO (1840 to March 2009), MEDLINE (1966 to March 2009), Scopus (1869 to March 2009), and the Cochrane Central Register of Controlled Trials up to March 2009. The searches included the following terms: “prolonged exposure,” “cognitive behavioral,” “clinical trial,” “random,” “randomly,” “randomize,” “randomise,” “randomized,” or “randomised” alone and in combination with “posttraumatic,” “post-traumatic,” “posttraumatic stress disorder,” “post-traumatic stress disorder,” or “PTSD”. These words were searched as key words, title, abstract, and Medical Subject Headings. Also, we examined citation maps and used the “cited by” search tools. These findings were cross-referenced with references from reviews. In addition, we contacted authors of PE randomized trials for emerging publications. These initial search strategies produced 127 potential articles. Examination of the abstracts identified 23 relevant articles. Inclusion criteria were as follows: (a) participants who met full DSM-III-R, DSM-IV, or DSM-IV-TR criteria for posttraumatic stress disorder; (b) random assignment; (c) adequate control condition (psychological placebo or wait-list control); (d) adult or adolescent participants; and (e) more than one session of PE during the acute phase of treatment. Treatments were classified as PE if they included both imaginal and in vivo exposure. Exposures were classified as in vivo if they involved confronting trauma cues in real life. All included studies used exposure therapy that was based on the manualized treatment developed by Foa et al. (1991). Exclusion criteria were: (a) single case studies; (b) studies focused on acute stress disorder; and (c) studies with insufficient data, unless study authors were able to provide such data. Of the 23 identified studies, ten were excluded. Four studies had inadequate trials of PE based on the above definition (Bryant et al., 2003; Cloitre, Koenen, Cohen, & Han, 2002; Glynn et al., 1999; Ironson, Freund, Strauss, & Williams, 2002). Four studies did not have an adequate control condition (Bryant et al., 2008; Ironson et al., 2002; Lee, Gavriel, Drummond, Richards, & Greenwald, 2002; Paunovic & Öst, 2001); in one study medication was the only treatment provided during the acute phase of treatment (Rothbaum et al., 2006); and in one study over one third of the participants did not meet full DSM-IV-TR criteria for PTSD (Difede et al., 2007). Thirteen studies with a total sample size of 658 participants met the final inclusion criteria and were included in the meta-analysis.

All analyses were completed with Comprehensive Meta-Analysis (Borenstein & Rothstein, 1999). Comprehensive Meta-Analysis is a program funded by the National Institutes of Health’s SBIR program.

Data on the following variables were collected: treatment conditions, treatment dose (number of sessions and total hours), number of participants, and year of publication. Dependent variables were classified into categories including primary (PTSD/domain-specific subjective distress) and secondary (general subjective distress: Table 1). In some of the included comparisons, PE was modified and combined with other treatments such as cognitive restructuring (Foa et al., 1999,2005; Marks, Lovell, Noshirvani, Livanou, & Thrasher, 1998; Power et al., 2002); results will be reported with and without these comparisons included.

Control conditions were classified into three categories: active treatment, psychological placebo or wait-list. Active treatment was defined as those treatments that were purported to contain specific ingredients to alleviate PTSD symptoms. Treatments that were categorized as psychological placebo included: supportive counseling (SC), relaxation (R), Present Centered Therapy (PCT), Time Limited Psychodynamic Therapy (TLDP), and treatment as usual (TAU). Wait-list (WL) was defined as a control condition in which participants did not receive any treatment for PTSD symptoms for a specified amount of time. Six of thirteen studies utilized psychological placebos as the control condition, 5 of the 13 studies had wait-list as the control condition and 2 of the 13 studies included both a psychological placebo condition and a wait-list condition (see Table 2).

The quality of the included studies was then rated using a modified version of Jadad et al.’s quality assessment guidelines (Jadad et al., 1996). Jadad et al. included three criteria for the quality assessment: 1) random assignment to condition; 2) double-blinding; and 3) description of withdrawals and dropouts. Because this meta-analysis reviews a psychological treatment, the second criterion cannot be applied here because patients cannot remain blind to their treatment condition; therefore we modified this criterion to require that the evaluators were blind to the therapeutic condition (i.e., single blind).

Between-group effect sizes for each study were computed using Hedges’s g (Rosenthal, 1991). Studies with multiple outcomes were categorized as above (see Table 1) and then combined within each domain. When the necessary data were available, Hedges’s g was calculated directly using the following formula: where is the mean of the treatment group, is the mean of the comparison group, and S_P is the pooled standard deviation. If these data were not provided, Hedges’s g was estimated using conversion equations for significance tests (e.g., t, F) (Rosenthal, 1991). All effect sizes were corrected for small sample sizes according to Hedges and Olkin (1985). Therefore, a smaller sample size reduces the estimated effect size helping control for different sample sizes across studies. These controlled effect sizes may then be interpreted conservatively with Cohen’s convention of small (0.2), medium (0.5), and large (0.8) effects (Cohen, 1988). Hedges’s g also may be computed directly from Cohen’s d with the following formula: . When there were multiple outcomes per domain they were combined according to Borenstein, Hedges, and Rothstein (2007). The overall mean effect size for all of the studies combined was computed using the following formula: where w_j is the weight for each study and g_j is the effect size for each study. Effect sizes were calculated with random effects models. The random effects analysis estimates the overall effect size assuming the studies included are only a sample of the entire population of studies and/or when the studies are heterogeneous.

PE will outperform the control conditions on primary outcome measures at post-treatment.

This analysis included 13 studies with 675 participants. Consistent with prediction, prolonged exposure outperformed control conditions on primary outcome measures at posttreatment showing a large effect size (Hedges’s g = 1.08 [SE = 0.20, 95% CI: 0.69 to 1.46, p < 0.001)]). Thus, the average participant receiving prolonged exposure fared better than 86% of the control participants at posttreatment on primary outcome measures (Fig. 1).

PE will outperform the control conditions on secondary outcome measures at post-treatment.

This analysis included 13 studies with 666 participants. Consistent with prediction, prolonged exposure outperformed control conditions on secondary outcome measures at posttreatment showing a large effect size (Hedges’s g = 0.77 [SE = 0.12, 95% CI: 0.53 to 1.01, p < 0.001]). Thus, the average participant receiving prolonged exposure fared better than 79% of the control participants at posttreatment on secondary outcome measures.

PE versus wait-list will produce a larger effect than PE versus psychological placebo at post-treatment on primary outcome measures.

This analysis included 13 studies with 675 participants. Consistent with prediction, the overall effect size of PE compared to wait-list (Hedges’s g = 1.51 [SE = 0.20, 95% CI: 1.12 to 1.90]) was significantly greater than the overall effect size of PE compared to psychological placebo (Hedges’s g = 0.65 [SE = 0.19, 95% CI: 0.29 to 1.01]) at post-treatment (p = .001).

PE will not significantly outperform other active treatments on primary outcome measures.

This analysis included 6 studies with 262 participants in PE, EMDR, CPT, CT, and SIT. Consistent with prediction, there was not a significant overall difference between PE and active treatment conditions on primary outcome measures at posttreatment (Hedges’s g = −0.07 [SE = 0.18, 95% CI: −0.42 to 0.28, p = .69]).

PE will outperform the control conditions on primary outcome measures at follow-up

This analysis included 7 studies with 348 participants. Follow-up assessments ranged between one and 12 months. Consistent with prediction, PE outperformed control conditions at follow-up on primary outcome measures showing a medium to large effect size (Hedges’s g = 0.68 [SE = 0.21, 95% CI: 0.27 to 1.10, p = .001]). Thus, the average participant receiving prolonged exposure fared better than 76% of the control participants at follow-up on primary outcome measures (Fig. 2).

PE will outperform the control conditions on secondary outcome measures at follow-up.

This analysis included 7 studies with 368 participants. Consistent with prediction, PE outperformed control conditions at follow-up on secondary outcome measures showing a medium effect size (Hedges’s g = 0.41 [SE = 0.19, 95% CI: 0.03 to 0.78, p = .03)]). Thus, the average participant receiving prolonged exposure fared better than 66% of the control participants at follow-up on secondary outcome measures.

A heterogeneity analysis was conducted to test the assumption that the effect sizes were from a homogeneous sample (Hedges & Olkin, 1985). For this analysis all 13 studies were included with all time points (post-treatment and follow-up) on primary outcome measures. The test was significant, Q(12) = 59.90, p<0.001, suggesting that the random effects analyses were most appropriate for this study.

The following analyses were completed using unrestricted maximum likelihood meta regressions. There was no significant relationship between effect size and time since trauma (β = 0.55, p = 0.61), publication year (β = −0.00, p = 0.86), dose (β = −0.01, p = 0.71), or study quality (β = −0.30, p = 0.07). There was also no significant difference in effect sizes across types of trauma (combat/ terror, childhood sexual abuse, rape, mixed: p = 0.14).

Several authors suggest there may be a potential discrepancy between the number of published trials and the total number that are completed (Bakan, 1967; McNemar, 1960; Smart, 1964; Sterling, 1959). Therefore, any meta-analysis of published studies may be missing non-significant studies and therefore overestimate the overall effect size. Rosenthal and others have called this confound “The File Drawer Problem” (Rosenthal, 1991). A conservative method of addressing this problem is to assume that the effect sizes of all current or future unpublished studies are equal to 0 and compute the number of such studies it would require to reduce the overall effect size to a non-significant level (Rosenthal & Rubin, 1988). This value may be referred to as the “fail-safe N”.

Rosenthal suggested the following equation to compute a fail-safe where K is the number of studies in the meta-analysis and is the mean Z obtained from the K studies (Rosenthal, 1991). Rosenthal also suggested that findings may be considered robust if the required number of studies (X) to reduce the overall effect size to a non-significant level exceeded 5 K + 10 which in this study would be 75 (Rosenthal, 1991). Analyses revealed that it would require more than 446 current or future unpublished studies with an effect size of 0 to bring the overall effect size of the primary analyses within the non-significant range, suggesting that the findings in this meta-analysis are robust.

This meta-analysis of 13 (N = 658) randomized controlled PE trials generally supported the hypotheses. As predicted, PE performed significantly better than control conditions on measures of PTSD both at post-treatment (g = 1.08) as well as at follow-up (g = 0.68). Similarly, PE treatment was associated with significantly better outcomes on secondary outcome measures, both at post-treatment (g = 0.77) and at follow-up (g = 0.41). These effect sizes compare quite favorably to those of other empirically supported psychotherapies; for example, cognitive therapy for depression has a reported effect size at post-treatment of d = 0.82 versus wait-list or placebo (Gloaguen, Cottraux, Cucherat, & Blackburn, 1998). In addition, the large effect sizes associated with PE in the present analyses are in contrast to the relatively smaller effects seen when “psychotherapeutic treatments” were lumped together (d = 0.52 post-treatment, d = 0.64 at follow-up) (Sherman, 1998). The current effect sizes were very robust and clearly not attributable to publication bias, given that it would require more than 400 PE trials with effect size equal to 0 to render the current results nonsignificant (Rosenthal, 1991). Taken together these results demonstrate that PE is a highly effective treatment for PTSD that confers lasting benefits across a wide range of outcomes.

Consistent with previous meta-analyses, there was no significant difference between PE and other active treatments (CPT, EMDR, CT, and SIT). This meta-analysis cannot answer the question of why these different treatments show similar efficacy. It is possible that these treatments work through separate mechanisms that are equally efficacious. Alternatively, it may be that the application of exposure routinely employed in all of these treatments is responsible for improvement. Future research may help further answer this question. Consistent with our hypothesis, the effect size for PE versus wait-list (g = 1.06) was significantly greater than the effect size for PE versus psychological placebo (g = 0.65). This is consistent with previous meta-analyses of psychological approaches (e.g. Feske & Chambless, 1995). There was significant effect size heterogeneity among the 13 studies included in the meta-analysis, indicating greater variability in effect sizes across studies than would be expected from sampling error alone (Hedges & Olkin, 1985). There was no significant relationship between effect size and dose (number of sessions). This finding could be due to restricted range (8 to 17 sessions) or flexible dosing (4 studies included flexible dosing). However, when the four studies with a flexible dose were removed we also found no relationship between effect size and dose (β = − 0.09, p = 0.61). It is also encouraging to note that effect sizes were not moderated by time since trauma, publication year, study quality, or type of trauma.

Results of the current meta-analysis have meaningful implications for mental health professionals who treat patients with PTSD. A therapist can expect that his/her average PE treated patient will fare better than 86% of patients treated with supportive counseling and similar unstructured talk therapies. The present analyses are also of interest to patients with PTSD who are weighing their treatment options, as well as to policy makers responsible, for example, for setting training priorities.

Several limitations deserve comment. First, the current analyses cannot definitively answer the question of the relative effectiveness of PE compared to other active psychological treatments for PTSD as there were an insufficient number of eligible studies comparing PE with CPT (one study), EMDR (three studies) or SIT (two studies), to obtain a stable estimate of these effect sizes. Nevertheless, we presented this analysis by combining the six comparison treatments. There also are too few dismantling studies to examine the relative efficacy of the individual components of PE (imaginal and in vivo exposure). To our knowledge there is only one study that included an imaginal exposure only condition and met the other inclusion criteria (Bryant et al., 2003). Although there was only one study, it is interesting to note that the controlled effect size for imaginal exposure only was similar in magnitude to our overall findings (g = 0.83).

Second, the current analysis did not test the effectiveness of PE compared to pharmacological treatments. However, existing results suggest that the effect of pharmacotherapy for PTSD is significantly less than the effects observed here for PE. For example, a meta-analysis of nine randomized placebo-controlled pharmacotherapy trials found an effect size of 0.41 for PTSD symptoms. Even the medication (the selective serotonin reuptake inhibitor fluoxetine) with the highest effect size at 0.65 was not of the same magnitude as the result found for PE (Penava et al., 1996).

Third, several of the studies were carried out by the same research group (i.e., Foa et al.). This raises the question of whether findings are similar when this type of treatment is implemented by different groups of researchers. We chose to examine this question empirically. Analysis showed that effect sizes were not significantly different between Foa’s group and other researchers (p = 0.47). Unfortunately data were not available to further determine the effects of gender or chronicity.

Finally, the follow-up analyses included fewer studies (7 vs. 13) and patients (348 vs. 675 for primary outcome measures) than the posttreatment analyses because many of the studies did not include a follow-up assessment. It is crucial to know whether the substantial posttreatment gains are maintained over time, and therefore it is encouraging to see a substantial effect size at follow-up. However, analyses that include fewer patients result in a loss of statistical power and therefore a less reliable estimate of the treatment effect size. Hopefully, future studies will include follow-up assessments such that subsequent meta-analyses can confirm the long-term effectiveness of PE for PTSD.

In sum, the results show that PE is highly effective in treating PTSD, and significantly more effective than inactive (waitlist) and active (psychological placebo) control conditions. In fact, the average PE treated patient fared better than 86% of patients in the control conditions. These findings are in marked contrast to the conclusions of Benish et al. (2008) and suggest that the efficacy of exposure treatments for PTSD cannot be attributed solely to non-specific factors. Consistent with prediction (and with the Benish et al. findings), there was no significant difference between PE and other active treatment conditions (CPT, EMDR, CT, and SIT). PE effectively treats both PTSD-specific distress as well as more general trauma-related distress (e.g., depressed mood), and the benefits are dramatic and enduring. The large effect sizes support the status of PE as the first line treatment-of-choice for PTSD.