Summary of main results
We cannot make any comment on the comparative efficacy of individual antidepressant compounds. However, based on our meta-analyses, there was no evidence that the magnitude of intervention effects (compared with placebo) was modified by individual drug class. Pooled results showed some evidence that depression symptoms were reduced and that remission/response rates were higher in those taking a newer generation antidepressant compared to placebo. The clinical importance of these effects is unclear. Overall, the reduction in depression symptoms was relatively small at 3.51 (on a scale with a range from 17 to 113) (Summary of findings for the main comparison) and remission rates overall increased from 380 per 1000 (based on the median risk in the placebo group at post intervention) to 448 per 1000 (Summary of findings for the main comparison). Even for those drugs that had consistent evidence of efficacy across more than one outcome (fluoxetine and escitalopram), the reduction in average scores was 5.63 lower for those taking fluoxetine and 2.67 lower for those on escitalopram compared with those on placebo. Poznanski 1996 found a difference of 25 points on the CDRS-R scale between a clinically referred depressed group (n = 60) and a non clinical group (n = 223), and a difference of 19 points between clinically referred groups with (n = 60) and without depressive disorder (n = 18) (Poznanski 1996, pg 53). There was also a relatively small effect on functioning with those receiving fluoxetine and escitalopram having a score on average 3.08 higher and 2.28 points higher respectively than placebo on the CGAS (range 0 to 100). For fluoxetine, remission rates overall increased from 214 per 1000 (based on the median risk in the placebo groups at post intervention) to 315 per 1000 (a difference of 101 with 95% confidence interval (CI) from 6 more to 231 more). While significantly more young people recover if on fluoxetine, the remission rates are relatively low overall indicating many young people do not experience a clinically meaningful reduction in symptoms. Methodological issues including large attrition rates and inappropriate outcome measurement, with the associated potential for selective reporting, make it difficult to draw conclusions about the clinical benefit.
In addition, the efficacy results need to be balanced with evidence about adverse outcomes. Over all drugs, the risk of a suicide-related outcome for those taking antidepressants was 58% (95% CI 2% to 245%) higher compared to those taking placebo. This equates to an increased of risk in a group with a median baseline risk from 25 in 1000 to 40 in 1000 (Summary of findings for the main comparison). There was some variability in the risk of suicide-related outcomes across the drugs (I2 = 22%; 95% CI 0% to 58%), but this was not statistically significantly different. Suicidal ideation was only measured in two trials with no effect of antidepressant treatment evident. Adverse events were greater for those taking an antidepressant, and there was no effect of antidepressant on the rate of trial completion.
There were few data available to investigate whether age modified the intervention effect. There was no evidence from the available data that age modified the effects of antidepressant medication. The evidence of the effectiveness of antidepressants for children was inconsistent, with a small overall reduction in depression symptoms (3.18) and remission rates in one study (Emslie 2002) of 44% for those receiving fluoxetine compared to 18% for those receiving placebo and in the other (Emslie 2006) 69% for those receiving escitalopram and 52% for those receiving placebo. For adolescents, there was consistent evidence across both these outcomes of antidepressants compared with placebo. However, the same cautions apply in terms of methodological considerations and the size of the effects.
Overall completeness and applicability of evidence
In this review, to our knowledge, we have presented data on efficacy and adverse outcomes, including suicide-related outcomes, from all published and unpublished trials examining the use of newer antidepressants for child and adolescent depressive disorder. Despite attempts to contact all trial authors, as well as pharmaceutical companies responsible for funding the included trials, there are many instances of missing data in terms of effect estimates. In two cases there is very limited reporting of trials by the Medicines and Healthcare Products Regulatory Agency (MHRA), with publication in peer-reviewed journals not yet available (Mirtazapine Trial 1 & 2). Likewise the trial by Simeon 1990, which was stopped early, has never been published. We were unable to obtain any further report of the trial Glod 2004, the preliminary findings of which were published in a conference abstract.
The comparator group has an impact on the size of the effect and, in this case, the size of the effect in the placebo groups in these trials has been commented on (e.g. Jureidini 2004) and authors of the included trial reports themselves have commented, suggesting, for example, that this may be to do with the large amount of contact trial participants receive. Trials consistently include regular (often weekly) assessments and in a significant number of trials some sort of supportive contact or therapy was allowed. The interaction between participants and trial investigators was seldom standardised, as shown in a trial specifically investigating this issue, and while this interaction is typical of what takes place in real world clinical encounters, it impacts in an unknown way on the detection of differences between the active drug and placebo (Dunlop 2010).
The characteristics of the participants also impact on the intervention effect. It is possible, for example, that the effect shown in the fluoxetine trials may be due to exclusion of placebo responders in the lead-in time to the start of the trial. The placebo remission rate was higher in trials of other compounds and has previously been cited as a cause for concern regarding the efficacy of selective serotonin reuptake inhibitors (SSRIs) (Newman 2004). Again, it is notable that in the trials of fluoxetine the attrition rate was consistently higher in the placebo group compared with the fluoxetine group. There were generally very high attrition rates in all of the trials and it is unclear what effect this has on treatment estimates.
The young people in the trials are not likely to be representative of clinical populations typically seen in public child and adolescent mental health services as most trials included recruitment using advertisement, excluded many comorbid disorders, and in many trials excluded those who responded to placebo in the lead-in stage of the trial. Also excluded were those at risk of suicide despite suicide-related behaviours and comorbidity being significant features of clinically referred young people with depressive disorders (Birmaher 1996; Kovacs 1984; Marttunen 1998; Petersen 1993). The baseline severity of young people included was in the moderately severely ill range, with a large proportion of included participants experiencing their first episode of depression (although in many cases this first episode was of a long duration). The effectiveness of newer generation antidepressants in young people with more severe disorders and complex presentations, including comorbid conditions and suicide risk, is therefore unknown.
Quality of the evidence
There was limited information on the conduct of trials in relation to allocation concealment, blinding and compliance. Blinding is an issue when clinician-rated scales are the main outcome, particularly in the context of an inactive placebo where it may be possible to guess the assigned treatment group given side and other physiological effects likely in this group (Moncrieff 2004). In most cases the write-up simply indicated that the trial was 'double-blind', some stating that the placebo and medication capsules were identical and some stating that clinicians or trial personnel and participants were blinded. Only one trial specifically stated that outcome assessors were blinded.
The issue of reporting bias is important. Kirsch 2008 highlight in their meta-analysis of all trials of antidepressants submitted to the Food and Drug Administration (FDA) that effect sizes are smaller when unpublished studies are included and Turner 2008 showed that whether and how trials of antidepressants were published depended on the outcome of the trial. We located several trials that have not been fully reported or published at all. It is also worth noting that in many cases we were unable to obtain the required data from the published paper but had to contact authors or the pharmaceutical company. Reporting bias within the published reports of included trials was difficult to assess given the conduct of a trial can be obscured in the write-up for publication. Full and explicit reporting of changes in outcome definition was only undertaken by one investigator, however the primary outcome was reported and findings discussed (Emslie 1997; Jureidini 2004). The possibility of reporting bias was highlighted in a letter to the editor regarding post hoc alterations of response definitions in the trial by Keller 2001 (Jureidini 2003).
What level of improvement constitutes a meaningful clinical outcome is uncertain given response and remission were defined and reported variously both within and across trials, with the noted possibility of alteration of this definition, and the possibility of reporting bias as a result.
There was evidence of inappropriate methods of imputation with trialists often using last observation carried forward data (Sterne 2009). It was often the case that some randomised patients were not included in the final analysis.
The majority of trials were pharmaceutically funded. Two of the four fluoxetine trials were not pharmaceutically funded (Emslie 1997; TADS 2004) (the TADS trial had an unrestricted education grant from Eli Lily). Research has shown that across different health fields, pharmaceutically funded studies are more likely to have results favouring the pharmaceutical company's product (Lexchin 2003; Sismondo 2008).
A standard definition of remission or response would have been ideal; however, to calculate this individual patient data would have been necessary. Further, given a diagnosis of major depressive disorder on DSM criteria was an entry criterion for most of the trials, this may be considered the most desirable outcome measure.
Additionally, given symptom improvement or resolution does not necessarily correlate with improvements in functioning (see Winters 2005 for a review), the latter would seem a more clinically important patient outcome to collect. However, this was inconsistently measured and reported on scales that often did not have established psychometric properties. In trials that used the Children's Global Assessment Scale (CGAS) (an adaptation of the Global Assessment of Functioning (GAF) used in adults), there was some evidence that functioning was improved for those on an antidepressant compared with placebo. Self report data may also tap an outcome that is meaningful to the young person, however, few individual trials reported this.
The trials are designed only to examine the short-term effects of antidepressant medication, however, this does not preclude the possibility that the effectiveness of treatment is only apparent over a longer period of time. Long-term follow-up would be required to assess this.
Potential biases in the review process
It should be noted that the review process included collection of data from various sources. There were more complete data for the trials on paroxetine due to publication of trial reports by SmithKline Beecham on the internet. Details of aspects of trial methodology were relatively brief even in this case. Information and data from other trials were taken variously from scientific journal publications, from the MHRA data and, in some cases, obtained directly from trial authors and pharmaceutical companies.
We were unable to obtain a large number of potentially relevant publications to screen for inclusion. The majority of these were conference proceedings and there is some evidence that such data are often not published and that publication is associated with positive results (Sherer 2007). However, funnel plots for the outcomes remission and suicide-related behaviour were not suggestive of small-study effects and, moreover, the contour-enhanced funnel plots did not indicate that statistically significant results were more likely to be reported.
Agreements and disagreements with other studies or reviews
The issue of the efficacy of antidepressants and the associated suicide risk has been the topic of much research, review and debate in the literature.
Healy was one of the first to publish regarding the possibility of an increased risk of suicide associated with SSRIs (Healy 2003). Since then, observational studies have highlighted an association between antidepressant use and suicide risk in adolescents. For example, Olfson 2006 showed an association between use of antidepressants and suicide attempt in children and adolescents and Simon 2006b showed a greater risk of suicide attempt that resulted in hospitalisation for children and adolescents compared with adults.
Data on risk of suicide for children and adolescents have also come from meta-analyses based on trial data submitted to the FDA (Hammad 2006) and the Committee on Safety in Medicines in the UK (Dubicka 2006) that show similar results to our review. These findings have more recently been extended to young adults (18 to 25-year olds) (Stone 2009). Additionally, Dubicka 2006 analysed these data by type of suicide-related outcome showing suicidal ideation in 1.2% of those on SSRIs compared with 0.8% on placebo (odds ratio (OR) 1.45; 95% CI 0.54 to 3.88), self harm in 3.3% of those on SSRIs versus 2.6% on placebo (OR 1.44; 95% CI 0.70 to 2.97) and attempts in 1.9% of those on an SSRI versus 1.2% on placebo (OR 1.70; 95% CI 0.76 to 3.81) (Dubicka 2006). Another review (Bridge 2007) calculated the risk difference for suicide-related behaviours for antidepressant versus placebo for major depressive disorder (MDD) (13 trials), obsessive-compulsive disorder (OCD) (six trials) and non OCD anxiety disorders (six trials). The risk difference for suicide-related behaviours was not statistically significant for MDD alone, although the risk ratio was 1.6; 95% CI 1.0 to 2.7 (Bridge 2007). The risk difference calculated by Bridge 2007 was 1% compared with a risk difference of 1.5% based on risk ratio (assuming a median baseline risk) calculated in this review.
Counter to this evidence showing an increase in suicide-related behaviours, several arguments have been mounted.
Firstly, observational and ecological studies have shown no increased suicide risk related to antidepressant use. For example, observational studies by Valuck 2004 (a retrospective cohort study) and Sondergard 2006 (a register-linkage study) showed no statistically significant increases in suicide attempt (hazard ratio (HR) 1.43; 95% CI 0.89 to 2.82) and completed suicide (RR 4.47; 95% CI 0.95 to 20.96) (respectively) for children and adolescents taking antidepressants. Two reviews of studies investigating this link highlighted that observational studies have not shown a strong association between antidepressants and suicide risk (Hall 2006; Isacsson 2010). Hall 2006 highlighted that observational studies have tended to show that those at increased risk of suicide were more likely to be prescribed an antidepressant (Hall 2006) and therefore the association between antidepressant use and adolescent suicide-related behaviours may be more likely to be explained by severity of depression, and not caused by the antidepressant (Friedman 2007; Simon 2006). However, issues of potential confounding bias arising in observational studies are minimised in trials where, on average, those in the intervention and control groups are comparable in both measured and unmeasured confounders.
Hall 2006 and Isacsson 2010 have highlighted that the majority of ecological studies have shown that suicide mortality has decreased when SSRI use has increased in various counties or regions (Hall 2006). However, these studies either have not included children and adolescents or analysed this age group separately, with the exception of Olfson 2003 in the US. Olfson 2003 showed that increasing use of antidepressants was associated with a decreased suicide rate. A subsequent study using US and Dutch data showed a decrease in SSRI prescriptions associated with an increased suicide rate in both countries (Gibbons 2007).
There are shortcomings to these types of ecological studies, which have been discussed in the debate in this area. It is commonly acknowledged that ecological studies are not able to establish a causal link. For example, it could be that there are lower suicide rates in countries with less stigma associated with seeking intervention for mental health disorders, where antidepressant prescriptions rates are therefore higher (Gunnell 2004; Simon 2006). Ecological studies have focused on short time periods, for example, Healy 2009 and Reseland 2008 contend that ecological studies have omitted data from the 1960s and 1970s when suicide rates climbed despite the largest increase in antidepressant use seen in the last few decades; Safer 2007 similarly argues that time periods prior to the introduction of SSRIs have not been included when drops in suicide rates were observed. Another issue of ecological studies is that the relationships observed between variables at the aggregated level may not be the same as those observed at the individual level. An example of this is the possibility that increased use of antidepressants may be due to chronic users not new users, whose risk of treatment-induced suicide is less (Reseland 2008). There are issues of confounding, for example fluctuations in suicide rates may be due to cultural and religious factors (De Leo 2002). Further, there is the issue of redefinition of the outcome over time. For example, reanalysis of data on suicide rates, antidepressant sales and autopsy in Nordic countries highlighted the possibility that observed decreases in suicide rates may be due to falling autopsy rates, and an associated increased in the rate of deaths from unspecified causes (Reseland 2008).
Finally, the observation of increased suicide rates associated with a decrease in the rate of SSRI prescriptions may be because the recent censure of SSRIs has resulted in pessimism about being able to provide effective treatment. This may have caused both a decrease in help seeking and a disinclination to diagnose depressive disorders, supported by data showing decreased rates of diagnosis for child and adolescent depression in the USA (Libby 2007).
The fact that there were no reports of completed suicide in a total sample of 2240 young people has been highlighted. However, suicide is a rare event and much larger sample sizes with longer follow-up would be needed to assess the risk fully. Healy 2009 cautions that it is more accurate to state that no suicides were recorded and notes that large loss to follow-up makes it possible that suicides did occur. Research suggests that the best predictor of eventual suicide completion is previous suicidal ideation and suicidal behaviour so that an increase in rates of these outcomes may increase the risk of future suicide completion (Andrews 1992; Brent 1986; Brent 1993; Lewinsohn 1996). The trials included in our review and others do not provide long enough follow-up to determine whether short-term increased risk of suicide-related outcomes is subsequently followed by a longer-term risk reduction. An increase in risk followed by a steadily decreasing risk has been shown within one month of starting antidepressant use in one trial of adults (Jick 2004). Data on timing of suicidal risk has been analysed from the TADS 2004 trial and showed that there was no decrease in risk after one month as seen in adult data, again meaning it is unclear if there is a longer-term risk reduction.
In all reviews the rates of suicide-related behaviour are much lower than reports of suicidal ideation and behaviour in adolescents from the community. In some population-based studies suicidal ideation is reported by as many as 12% to 23% of young people and suicide attempts by 4% to 8% of adolescents aged 13 to 18 years (AHRG 2003; Grunbaum 2002; Lewinsohn 1996; Sawyer 2001). This could be explained by exclusion of those at risk of suicide from all but two trials (Emslie 2002; Von Knorring 2006). Most trials also excluded those with many co-morbid conditions. Co-morbidity is related to an increased risk of suicidal ideation and suicidal behaviours (Andrews 1992; Asarnow 1992; Brent 1986; Esposito 2002; Kovacs 1993; Shaffer 1996; Wetzler 1996).
The findings regarding an increased risk of suicide-related behaviours need to be balanced with concerns regarding the risk of untreated depression and the need to balance benefits and harms (AACAP 2004; Ebmeier 2006; Simon 2006).
Several reviews of the efficacy of SSRIs in children and adolescents with depression have been published. Earlier reviews based conclusions on narrative summaries of individual trial results (Brent 2004; Cheung 2005; Wagner 2005) not meta-analyses. Of these, the conclusions of Cheung 2005 are the most conservative, stating data are limited, commenting specifically on the inappropriateness of the outcomes used and cautioning clinicians to consider the use of SSRIs carefully.
A number of meta-analyses have been also published. In earlier meta-analyses Jureidini 2004 highlighted methodological problems, particularly regarding reporting, and concluded cautiously that there is only a small benefit of SSRIs that should be balanced with the risks. In contrast Cohen 2004 concluded SSRIs are an effective treatment for adolescent depression based on data published up to that point. In the same year, a review by Whittington 2004 highlighted the change in risk-benefit profile when unpublished data were included; their inclusion resulted in a favourable risk-benefit profile only for fluoxetine, with a number needed to treat to benefit (NNTB) of 6 (95% CI 4 to 15) for remission, and 5 (95% CI 4 to 13) for response. More recently, Tsapakis 2008 has published a systematic review including meta-analyses of any antidepressant (including tricyclic antidepressants (TCAs), SSRIs, monoamine oxidase inhibitors (MAOIs) and new antidepressants) for children and adolescents up to the age of 20 with a depressive disorder. The authors' conclusions were cautious, highlighting limited efficacy, particularly for children and in studies that combine adolescents and children. They found little appreciable difference between different types of antidepressants, however, suggesting that fluoxetine may be more effective, especially for adolescents. Similar to our findings, they questioned the appropriateness of outcome measures and noted that the young people included in trials may not be representative of more classic 'adult major depression' and may be less severe that young people typically seen in clinical practice.
In a meta-analysis of antidepressants for all age groups, Ioannidis 2008 questioned the appropriateness of outcome measures used and highlighted issues related to selective and distorted reporting and interpretation, calling into the question the effectiveness of antidepressants even in adult populations.
The results of our review present a dilemma for those treating young people with depressive disorders. While overall the reduction in depression symptoms is statistically significant, there are questions about the clinical effectiveness. The trials are of young people not representative of those typically presenting for treatment in clinics and furthermore they had some significant methodological shortcomings, making it difficult to draw firm conclusions. Potential benefit must be balanced with the finding that SSRIs are associated with a statistically significant increased risk of suicide-related behaviour (a combination of suicidal ideation and definitive suicidal behaviour). Much debate on this matter has taken place, with variable views.
While there is a lack of power to adequately examine rare events such as suicide-related outcomes in trials, meta-analysis offers a method for obtaining reliable evidence for these rare outcomes across trials. Given a lack of consistent measurement of various aspects of suicide risk, the clinical implications of the findings are still somewhat unclear.
It is unknown how children and adolescents with a depressive disorder and co-morbid conditions, who are at risk of suicide (i.e. those more typical of the young people who present at health services), would respond to SSRIs. Overall, the data regarding the benefits of SSRIs for child and adolescent depression are far from compelling while the information on the risks is limited.
Clinically, it must be kept in mind that there are significant risks in not treating depressive disorder, which has an increased risk of suicide completion, as well as impacts on academic and social functioning (Brent 2002; Ebmeier 2006; Fleming 1993; Lewinsohn 1998; NHMRC 1997; Rao 1995).