A recent meta-analysis by Bloch and Hannestad (B&H)1 examining the efficacy of omega-3 fatty acids (ω3-FA) in major depressive disorder (MDD) claims no significant overall benefit compared with placebo; standard mean difference (SMD) 0.11 (95% CI, –0.04, 0.26) with no specific influence of eicosapentaenoic acid (EPA) dose on meta-regression analysis. These findings contrast with those from other recent meta-analyses.2, 3, 4, 5 As potential reasons for this disparity, we question the validity of their results on the basis of inclusion/exclusion criteria, study subgroup selection, strategy for selecting outcome measures, SMD estimates and choice of effect modifiers. Finally, we present a reanalysis yielding different conclusions.
Trial inclusion and exclusion
B&H's stated inclusion criteria were ‘randomized and placebo-controlled trials examining the efficacy of ω3 FAs in adults with MDD… [, trials] examining the efficacy of ω3 FAs in treating MDD in subjects with medical comorbidity (that is, cardiac disease and Parkinson's disease) or pregnancy… [and trials examining] the efficacy of omega-3 FAs to target depressive symptoms (as a primary outcome) in patients who may not have received a formal psychiatric diagnosis (that is, in a primary care setting)’. However, B&H's reported aims, results and conclusions appear to be somewhat misleading as they comment exclusively on MDD when, by virtue of the third inclusion criteria, their meta-analysis included the study by Rogers et al.6 examining individuals with mild-to-moderate depression without formal psychiatric diagnosis. It appears, therefore, that their meta-analysis did not exclusively examine MDD as implied, but a broader categorization that included subclinical depressive symptoms that would not fulfill diagnostic criteria for a major depressive episode (MDE) or MDD.
B&H's search strategy included trials to May 2010, but two relevant earlier trials satisfying their inclusion/exclusion criteria were not included.7, 8 Although the trial by Bot et al.7 (epub date May 2010) might not have been available at the time B&H ceased data collection, the trial by Rondanelli et al.8 (epub date February 2010) should have been included as it examined ω3-FA monotherapy in elderly women with MDD.
Regarding antidepressant augmentation with ω3-FAs, the two trials by Jazayeri et al.9 and Carney et al.10 were excluded because pharmacological agents of known efficacy were started simultaneously with ω3-FA. Although simultaneous start and add-on to antidepressant failure trials address different clinical questions, they both provide valuable data on the adjunctive use of ω3-FA for depression. A sensitivity analysis comparing inclusion versus exclusion of these studies might have been more appropriate.
Selection of study subgroups
Regarding the Lucas et al.11 study that recruited women with psychological distress, we do not consider B&H's inclusion of the subgroup of 29 patients with a MDE at baseline to be justified because Lucas et al. state on page 642 of their paper that patients with severe MDE (scores of 26 on the 21-item Hamilton Depression Rating Scale (HDRS)) were excluded for ethical reasons. The mean baseline HDRS score for those with MDE was 16.7 in Lucas et al. versus a pooled mean HDRS baseline value of 22.0 (range 18.6–30.1) for the other MDD studies employing the HDRS in this meta-analysis. Thus, the small subgroup in Lucas et al. labelled as having ‘MDE’ was unlikely to be representative of patients with MDD as a whole, being biased towards patients with mild depressive symptoms.
Outcome measures used
B&H employed a hierarchical strategy for use of outcome measures, prioritizing HDRS, then Montgomery Åsberg Depression Rating Scale, then other measures.1 Although it is important to strive for consistency in outcomes entered into meta-analyses, in this instance, their strategy may not have been optimal for the following reasons.
First, it may be preferable to report the predefined primary outcome, or where this is not clearly stated, the outcome statistically powered to detect treatment changes, or the outcome used to select patients with cut-off values above a predefined threshold for study inclusion. Apart from a few studies employing a single outcome, most reported such a ‘primary’ outcome (Table 1). However, it should be noted that many earlier studies were not registered, thus limiting confidence in the descriptions of these outcomes as ‘primary’. A number of these ‘primary’ outcomes differed from those employed by B&H. For example, two out of the three studies of perinatal depression used the Edinburgh Postnatal Depression Scale, a standard measure in this clinical setting, as a ‘primary’ outcome;12, 13, 14 however, B&H used the HDRS for these studies. In addition, their strategy was applied inconsistently as Grenyer et al.15 included data on both Beck depression inventory and HDRS (time-by-treatment interaction F statistic that can be used to compute a SMD), whereas B&H used Beck depression inventory for this study.
Second, given it is possible to calculate pooled mean outcome values in meta-analyses of studies employing multiple outcome measures,3, 16 use of combined outcomes using all available published data on change in depressive symptoms might have reduced potential bias arising from meta-analytic investigator decisions as to which outcomes to select and/or uncertainty as to the robustness of reported ‘primary’ outcomes.
Inconsistencies in SMD estimates
The SMD for Su et al.17 in MDD was calculated in B&H as 1.04 (95% CI 0.13, 1.94),1 about half that calculated in Appleton et al.,2, 18 Lin and Su,19 Ross et al.4 and Martins.3 In addition, the calculated SMD 95% CIs for the Grenyer et al.15 and Mischoulon et al.20 studies did not cross zero in B&H's analysis,1 when neither paper reported significant differences.15, 20
B&H found no relationship to EPA dose on meta-regression analysis in contrast to other published meta-analyses,3, 4 which may have resulted from their much lower SMD estimate for the Su et al.17 MDD study. Moreover, they did not perform a subgroup analysis contrasting supplements containing predominantly EPA versus those containing predominantly docosahexaenoic acid, when other analyses have shown that only predominantly EPA supplements appear to be effective.3, 4 Recently, a further meta-analysis has demonstrated that only supplements containing 60% EPA appear to be efficacious, hypothesized to be related to the unopposed dose of EPA (i.e., after competition with inactive docosahexaenoic acid has been overcome).5
Given the differences in SMD estimates between B&H and other meta-analyses, we sought to replicate their analysis using the same studies, outcome measures and subgroups as B&H1—the Replication Model. In addition, we performed a number of sensitivity analyses on the Replication Model exploring: (i) the effect of restricting study selection to exclusive MDD by excluding Rogers et al.6 and the Lucas et al.11 subgroup, (ii) comparing B&H's outcomes strategy versus use of ‘primary’ outcomes versus combined outcomes in exclusive MDD and (iii) examining differential effectiveness of supplement regimes containing 60% EPA versus those containing <60% EPA in exclusive MDD.
We conducted updated analyses that included recently published studies of adult patients with a psychiatric diagnosis of MDD and using combined outcomes (Table 1). We also performed sensitivity analyses examining: (i) the effect of including versus excluding MDD studies that examined ω3-FA started simultaneously with pharmaceutical antidepressants and (ii) the differential effectiveness of 60% EPA versus <60% EPA. Finally, the effect of publication bias has been assessed using Duval and Tweedie's trim and fill method and Egger's test.
The program comprehensive meta-analysis3, 16 was employed using fixed effects for the Replication Model and associated sensitivity analyses for direct comparability with B&H's analysis1 and random effects for the updated analyses, which we considered the most appropriate approach, given the significant heterogeneity between studies.
Compared with placebo, overall SMD in response to ω3-FA was 0.172 (95% CI 0.018, 0.325; P=0.028) for the Replication Model. A one-study-removed analysis showed that the Su et al.17 MDD study accounted for the biggest shift from insignificant to significant overall SMD estimates, suggesting that the low SMD estimate quoted for this study in B&H may have contributed to their overall insignificant SMD estimate. Compared with the Replication Model, our sensitivity analyses showed a larger SMD estimate when only MDD studies were selected. Regarding choice of outcomes, SMD estimates were highest using B&H's outcomes, smaller using ‘primary’ outcomes and the most conservative using combined outcomes, suggesting this latter strategy may be optimal. MDD studies employing 60% EPA showed a highly significant SMD of 0.892 (95% CI 0.543, 1.241; P<0.001), whereas those employing <60% EPA showed no effect. Of note, heterogeneity was no longer significant when using combined outcomes and accounting for the EPA content of regimes (Table 2A).
Our updated analyses, including recently published MDD studies (Table 1), and which used combined endpoints, confirmed significant overall SMD estimates whether ω3-FA/antidepressant simultaneous-start MDD studies were included or not. This larger group of MDD studies employing 60% EPA still showed a highly significant SMD of 0.621 (95% CI 0.334, 0.909; P<0.001), whereas those employing <60% EPA showed no effect (Table 2B; Figure 1). In addition, in contrast to B&H, our meta-regression analysis of MDD studies showed a significant relationship between EPA dose and efficacy (unrestricted maximum likelihood estimation point estimate for slope 0.270; 95% CI 0.044, 0.499; P=0.019). However, no relationship with docosahexaenoic acid dose was demonstrated (point estimate for slope –0.078; 95% CI –0.376, 0.219; P=0.606).
As identified in B&H and other meta-analyses2, 3 we confirmed a relationship between baseline depression severity and efficacy on subgroup analysis, patients with moderate or severe depression at baseline showed a significant SMD of 0.429 (95% CI 0.255, 0.702; P=0.002) whereas patients with mild depression at baseline showed an SMD of 0.071 (95% CI –0.538, 0.689; P=0.819).
Given the apparent marked differential effectiveness of regimes containing 60% EPA versus those containing <60% EPA, we considered it uninformative to assess the impact of publication bias on the whole group of studies. Therefore, we analyzed publication bias separately for these two groups of studies. No evidence of publication bias was found for studies employing <60% EPA using Duval and Tweedie's trim and fill method and on Egger's test (point estimate –2.26; 95% CI –4.80, 0.28; P=0.069). However, publication bias was evident for studies employing 60% EPA; three imputed studies were required to render the funnel plot symmetric resulting in an adjusted SMD estimate of 0.468 (95% CI 0.142, 0.794) using Duval and Tweedie's trim and fill method and Egger's test was significant (point estimate –2.02; 95% CI –3.42, –0.62; P=0.009).
Our critique of B&H suggests that their meta-analysis included patients who did not fulfill criteria for MDD and that their insignificant overall effect was most likely because of errors in the calculation of one or more SMD estimates and the inclusion of a biased subgroup of patients from Lucas et al.11 that had excluded patients with severe depressive symptoms.
Our updated reanalyses suggest that ω3-FA supplementation is beneficial in adult patients with MDD but that this effect is strongly dependent upon the EPA content of the regimen. Thus, studies employing regimens containing 60% EPA showed a highly significant pooled SMD estimate, whereas those containing <60% EPA showed no effect. After taking publication bias into account, the SMD estimate for the 60% EPA group of studies compared favorably with the mean weighted effect size of published (0.37; 95% CI, 0.33–0.41) and unpublished (0.15; 95% CI, 0.08–0.22) pharmaceutical antidepressant trials,21 although it should be noted that the adjusted SMD estimate for the 60% EPA group of studies may have been affected by the fact that, with the exception of Lésperance et al.,22 these studies tended to have smaller sample sizes than the <60% EPA group of studies. These data would suggest that it is inappropriate to consider all ω3-FA interventions to be the same, and that EPA, given its low side-effect profile and tolerability, should be more extensively evaluated as an antidepressant treatment in clinical trials with sample sizes, study durations and resources normally afforded to pharmaceutical antidepressants.
However, the confidence of these findings is limited by the small sample sizes of many studies, the evidence of publication bias in the 60% EPA group of studies, the fact that very few early studies were registered, the tendency for registered studies to show smaller SMD estimates or no effect compared with unregistered studies and the trend for more recently published larger studies to show no effect. Regarding this latter point, it should be noted that the more recent studies employing EPA have tended to employ an EPA dose of around 1 g/day, often quoting the earlier dose-ranging study by Peet and Horrobin23 in support of this choice. Given that we have demonstrated here a dose-response relationship with EPA on meta-regression analysis, an insufficient dose of EPA could be another possible explanation as to why more recent studies employing EPA have not shown efficacy. Thus, if EPA is to be further evaluated as an antidepressant, then further studies examining doses of up to 4.4 g/day17 should be undertaken.
About this article
Journal of Neural Transmission (2017)