Search strategy

We searched MEDLINE (from 1950 to week 2 of January 2008), EMBASE (from 1980 to week 4 2008), PsycINFO (from 1967 to week 2 January 2008), CINAHL (from 1982 to week 2 January 2008), the Cochrane Central Register of Controlled Trials and the Cochrane Database of Systematic Reviews (1800–2008). We included randomised, double-blind controlled trials in any language covered by the above sources. The subject headings applicable to each of the above medical literature databases can vary. Thus, subject headings or keywords relevant to the disorder, treatments and methodological features of interest specific to each database were used and combined in consultation with a librarian experienced in conducting systematic reviews. These included the terms for bipolar disorder and depression that were combined with the Boolean operator OR, and then combined using the term AND with [lithium OR valproic acid OR carbamazepine OR lamotrigine OR topiramate OR gabapentin OR all benzodiazepines OR all typical antipsychotics OR aripiprazole OR clozapine OR olanzapine OR quetiapine OR risperidone OR ziprasidone]. The terms adult AND randomised clinical trial were then combined, and, using the operator AND, were added to the above disorder and treatment-specific search terms.

We also monitored the table of contents of the British Journal of Psychiatry, the Journal of Clinical Psychiatry, New England Journal of Medicine, The Lancet, the Archives of General Psychiatry, the American Journal of Psychiatry, Bipolar Disorders and the Journal of Affective Disorders up to 1 April 2008 for further reports. Moreover, we reviewed reference lists of acquired studies and searched the clinical trials database at www.clinicaltrials.gov.

Titles and abstracts of all studies were reviewed by both authors. Articles were selected for full-text review if inclusion criteria were met or if uncertainty regarding their eligibility existed. Disagreements at any stage of article selection were resolved by discussion between the two reviewers and if agreement could not be reached, a third reviewer was recruited to determine eligibility. Initial agreement between the initial two reviewers was calculated using the Φ statistic ^{Reference Meade, Cook, Guyatt, Groll, Kachura and Bedard6} as it is superior to weighted kappa in cases where agreement is particularly high or low. All disagreements were resolved by discussion; and intervention by the third reviewer was not required.

Study eligibility

We included all published, randomised, placebo-controlled and active comparator trials described as double-blind and that included a mood stabiliser treatment group in adults (aged 18–65) with bipolar disorder and acute major depression. Acute in this context refers to treatments administered to reduce symptoms of an existing episode of depression and that continued until the depressive episode was in remission or the study ended. Studies were eligible regardless of whether participants were medication free or not at baseline.

Studies also had to meet a methodological quality threshold, which we set at a score of three or more on the Jadad scale. ^{Reference Jadad, Moore, Carroll, Jenkinson, Reynolds and Gavaghan7} This instrument awards one point to a study if it is described as randomised, another if it is double-blind and an additional point if there is a description of withdrawals and individuals who drop out. Studies also receive one point if the method used to generate the randomisation sequence is described and appropriate and/or if the method of double-blinding is described and appropriate. One point is deducted for each of inappropriate application of randomisation or double-blinding. In keeping with suggestions regarding the assessment of articles about therapy, ^{Reference Guyatt8} an extra point was given if participants were analysed in the groups to which they were randomised. This occurred only after inclusion based on the original five-point scale was determined. Despite the fact that questions exist regarding the validity of utilising scale scores in systematic reviews, ^{Reference Juni, Witschi, Bloch and Egger9} the Jadad measure was utilised because it has been widely used and is well-recognised. Its utilisation reflects our desire to convey a concise but meaningful description of the methodological quality criteria applied to study selection for practising healthcare professionals. The overall quality of the evidence for each outcome was assessed according to grades of recommendation, assessment, development and evaluation (GRADE) working group criteria. ^{Reference Schunemann, Jaeschke, Cook, Bria, El-Solh and Ernst10}

If studies contained diagnostically heterogeneous populations, we required that data be reported separately for people with bipolar disorder. Studies examining lithium, valproic acid, carbamazepine, lamotrigine, topiramate, gabapentin, the benzodiazepines and the typical and atypical antipsychotics were eligible for consideration if the medications were started during the depressive phase of bipolar disorder. Medication dosages could be fixed or flexible. Eligible studies could contain participants who were not taking other medications at baseline as well as those in which the investigational medication was added to ongoing treatment.

We did not include studies that examined participants in mixed states. Most studies that have enrolled individuals in mixed states focus on the treatment of mania and fail to report data separately for mixed and manic groups. Crossover trials were excluded because mood stabilisers are likely to exert therapeutic effects beyond their cessation. Despite the fact that the mood stabilisers as a class represent a pharmacologically heterogeneous class of medication, results were pooled and then the effects of separate medications and medication classes were explored in subsequent sensitivity analyses. Data were extracted independently by both authors using a predetermined data extraction form and disagreements were discussed. If no consensus could be reached a third reviewer extracted data and those results were used. Double entry of the data by both authors prevented the need for involvement by the third reviewer.

Data analysis

Data were double entered by each author into RevMan 4.2 for Windows (Nordic Cochrane Centre, Copenhagen, www.cc-ims.net/) for analysis. Mood stabiliser monotherapy was compared with placebo, with combination therapy (mood stabiliser plus another pharmacological agent) initiated at the same time as mood stabiliser monotherapy and to antidepressant add-on or monotherapy in separate analyses for each outcome. These outcomes included rates of response, remission, all-cause discontinuation (a proxy for the acceptability of treatment to individuals), affective switching, suicidal behaviour and clinically significant weight gain. All outcomes were dichotomous and so pooled relative risks (RR) with 95% confidence intervals were calculated. For treatment response and remission, relative risks above one represent an increased likelihood of a positive outcome (response or remission) for mood stabiliser monotherapy relative to comparators. For adverse events (affective switching, suicidal behaviour and weight gain), relative risks above one indicated that mood stabilisers posed a greater risk than their respective comparators. All attempts to utilise intention-to-treat data were made for studies that contained information relevant to the outcomes of interest. We did not impute values for efficacy or safety outcomes and defined the total number of participants eligible for efficacy and safety analyses in each trial as those who had had at least one post-baseline assessment value. When data were not reported in studies, authors were contacted. If no reply was received, values were estimated based on figures in the manuscript.

Meta-analysis using the random-effects method of DerSimonian & Laird ^{Reference DerSimonian and Laird11} was undertaken. This approach assumes that included studies represent a random sample of a larger population of studies and takes into account the possibility that studies that could be included are underway or not yet published. ^{Reference Fleiss12} Indeed, our search of www.clinicaltrials.gov using the term bipolar disorder and then depression indicated that 14 registered trials that may meet inclusion criteria for this review are currently in progress.

The presence of heterogeneity between studies for each outcome was tested for by visual inspection of forest plots and supplemented with Cochran's Q statistic. ^{Reference Cochran13} The magnitude of heterogeneity for each outcome was determined using the I ² statistic. ^{Reference Higgins, Thompson, Deeks and Altman14} If statistical heterogeneity was present, reflected by a Q-test P<0.05 (two-tailed), additional post hoc subgroup analyses were undertaken to attempt to understand reasons for it. A priori sensitivity analyses were applied to placebo-controlled studies alone because active comparator trials used a wide variety of treatments and sample sizes for these comparisons were smaller, further reducing the power to detect meaningful differences. The sensitivity analyses conducted sought to examine whether treatment effect sizes differed between each of the individual medications, antipsychotic and anticonvulsant classes, bipolar type, gender or those with psychotic features or a rapid cycling course. Unfortunately, papers included in the meta-analysis only contained data that permitted comparisons between individual medications and between medication classes. For these comparisons, tests of interaction ^{Reference Altman and Bland15} were conducted to compare the levels of each characteristic for each outcome (e.g. lamotrigine v. olanzapine; antipsychotics v. anticonvulsants), as per the method of Altman & Bland. ^{Reference Altman and Bland15} Tests of interaction are used to compare two independent estimates of the same quantity (in this case, relative risks) from separate analyses. Significance levels for these comparisons were set at 0.01 (two-tailed) rather than the 0.005 generated by the Bonferroni correction, as it was felt to unduly reduce the power to detect effects. Number needed to treat (NNT) or number needed to harm (NNH) statistics were calculated using Visual Rx for Windows (Cates, www.nntonline.net) using relative risks and designating the median value of the control/comparator groups for each outcome as the control event rate. These were only calculated when relative risks were significant.

There is uncertainty regarding the best means of detecting publication bias and no current method is universally accepted because of problems with unreliability and low power. ^{Reference Gavaghan, Moore and McQuay16} Publication bias is, however, an area of increasing concern in clinical trials research, particularly in psychiatry. ^{Reference Turner, Matthews, Linardatos, Tell and Rosenthal17} Thus, despite its limitations, a decision was made to examine funnel plots visually as a means of assessing publication bias in cases where ten or more studies were present. If these plots were clearly asymmetric or analysis was not possible, they were classified as having possible publication bias; otherwise publication bias was deemed to be unlikely.

Search findings

In keeping with the quality of reporting meta-analyses statement, ^{Reference Moher, Cook, Eastwood, Olkin, Rennie and Stroup18} a trial flow summary is included (Fig. 1). Fifteen publications representing 18 unique studies, ^{Reference Calabrese, Bowden, Sachs, Ascher, Monahan and Rudd19–Reference Shelton and Stahl33} containing 4105 randomised participants were eligible. All studies used DSM criteria except Fieve and colleagues, ^{Reference Fieve, Platman and Plutchik30} which relied on the then current clinical criteria for diagnosis.

Fig. 1 Study flow diagram.

Twelve studies were multisite and six recruited participants from a single centre. The number of participants in these studies varied from 19 to 833. Of the 12 studies containing data eligible for placebo-controlled comparisons, 5 examined lamotrigine, 1 carbamazepine, 2 valproic acid, 2 olanzapine and 2 quetiapine. Studies followed patients from 3 to 26 weeks, although the majority were 6–8 weeks in duration. Fifteen contained participants with a diagnosis of bipolar disorder type I or II only and three also contained those with bipolar disorder NOS (not otherwise specified). None of the eligible studies reported results for all relevant outcomes separately for different bipolar types, rapid cycling or psychotic participants or by gender. Six studies were not directly sponsored by a pharmaceutical company. Only two were prospectively registered in a clinical trials registry. Jadad scale scores ranged from three to six. Overall quality of the evidence for all outcomes assessed according to GRADE working group criteria was moderate, mainly because of high rates of drop out in both treatment and comparator groups in the trials. Funnel plots could only be examined for clinical response in the mood stabilising v. placebo comparison and suggested that publication bias was unlikely. For the remainder we concluded that publication bias may exist. Remaining features of all studies contributing data to this review are contained in online Table DS1. Results of the meta-analysis and a summary of the quality of the evidence for each outcome are found in Table 1.

Comparison 1: mood stabiliser monotherapy v. placebo

Of the 12 studies containing data eligible for placebo-controlled comparisons, ^{Reference Calabrese, Bowden, Sachs, Ascher, Monahan and Rudd19–Reference Calabrese, Huffman, White, Edwards, Thompson and Ascher27} 5 examined lamotrigine, 1 carbamazepine, 2 valproic acid, 2 olanzapine and 2 quetiapine.

Comparison 2: mood stabiliser monotherapy v. combination therapy

Six studies involving 1322 participants ^{Reference Tohen, Vieta, Calabrese, Ketter, Sachs and Bowden20,Reference Amsterdam and Shults23,Reference Zhang, Kang, Tan, Li, Gao and Zhang26,Reference Brown, McElroy, Keck, Deldar, Adms and Togen28,Reference Sachs, Nierenberg, Calabrese, Marangell, Wisniewski and Gyulai29,Reference Shelton and Stahl33} compared mood stabiliser monotherapy plus or minus placebo to combination treatments involving a mood stabiliser and an augmenting agent. This was an antidepressant in all cases but one. ^{Reference Zhang, Kang, Tan, Li, Gao and Zhang26} Only one added treatment to ongoing therapy. ^{Reference Shelton and Stahl33} Sachs and colleagues' participants also received a psychosocial intervention, the effect of which did not differ between groups, ^{Reference Sachs, Nierenberg, Calabrese, Marangell, Wisniewski and Gyulai29} but 18% of participants included in the monotherapy group were on more than one mood stabilising medication.

Comparison 3: mood stabiliser monotherapy v. antidepressants

Five studies ^{Reference Amsterdam and Shults23,Reference Fieve, Platman and Plutchik30–Reference Shelton and Stahl33} involving only 106 participants compared mood stabilisers with antidepressants started during an acute major depressive episode and three involved participants already taking a mood stabiliser at baseline. ^{Reference Young, Joffe, Robb, MacQueen, Marriott and Patelis-Siotis31–Reference Shelton and Stahl33}

Main findings

When medications were examined in the aggregate, the efficacy of mood stabiliser monotherapy was superior to placebo in the treatment of acute episodes of bipolar depression. Furthermore, monotherapy with mood stabilising medications appears to have equivalent efficacy to combination or antidepressant therapy in the acute treatment of this phase of bipolar disorder. Mood stabilisers were not less acceptable to participants than placebo for the acute treatment phase. Although limited by a short duration of follow-up and small number of events overall, we also observed no differences in terms of affective switching or suicidal behaviour between mood stabilising medications and placebo or other active treatment groups. In the mood stabiliser v. placebo comparison, no individual agent or class of medication was superior to another for any outcome. It should be noted however that the tests of interaction underlying this conclusion may have been underpowered to detect differences. Few studies are available comparing the use of mood stabilisers or antidepressants as add-on treatments for bipolar depression. These small sample sizes also limit the conclusions that can be drawn from comparisons between these treatment options. The quality of the evidence that supports the findings for all outcomes in this systematic review is moderate only, largely because of high rates of loss to follow-up in the trials.

Clinical implications

Although our systematic review supports the recommendations of more recently published practice guidelines that suggest that lamotrigine or quetiapine be considered as first-line agents for the treatment of acute bipolar depression, ^{Reference Yatham, Kennedy, O'Donovan, Parikh, MacQueen and McIntyre34,35} randomised, double-blind, placebo-controlled trials suggest that olanzapine and carbamazepine monotherapy may also have utility for this indication. Caution should be exercised however in comparing the effect size estimates for the agents contained in this study with Gijsman and colleagues' systematic review of antidepressant treatment for bipolar depression ^{Reference Gijsman, Geddes, Rendell, Nolen and Goodwin4} since this would be an indirect comparison. Direct comparisons of these therapies such as those contained in Sachs and colleagues' work ^{Reference Sachs, Nierenberg, Calabrese, Marangell, Wisniewski and Gyulai29} provide a stronger basis for inference. However, available evidence suggests that combination therapy with antidepressants fails to outperform mood stabiliser monotherapy, which is relevant to decisions about the role of antidepressant medication in bipolar disorder. Overall, the included studies suggest that there are mood stabilising agents available to treat acute episodes of bipolar depression without increasing the risk of switching and without adding an unacceptable side-effect burden. Given these data and the uncertainty regarding the safety and utility of antidepressant medications in bipolar depression, at the present time there appears to be little support for the first-line use of antidepressant medications in the treatment of an acute episode of bipolar depression unless there are specific reasons why the mood stabilising medications examined here could not be used for individual patients. However, it is also important to acknowledge that head-to-head trials comparing mood stabilising medications against antidepressant medications in bipolar depression are few in number and somewhat contradictory in their findings, perhaps because of small samples sizes and different approaches. Indeed, more study is required to determine the role and optimal use of antidepressants in bipolar depression

The time-limited nature of the trials conducted to date makes it difficult to generalise to the long-term treatment of people with bipolar disorder, where maintenance treatment is a key component of the overall treatment approach. Another limitation of the data is that the majority of participants in the studies had bipolar disorder type I, and it is therefore impossible to know whether the effect of these treatments would be similar in people with type II bipolar disorder.

Unfortunately, the design and reporting of the original trials did not allow us to calculate separate effect sizes for the different genders, or for individuals with psychotic features or a rapid cycling course.

Whereas clinical practice guidelines also frequently list lithium as a first-line treatment for bipolar depression, only one parallel group randomised trial involving lithium was eligible for inclusion in the current review. Although a number of randomised, masked crossover studies that have examined the efficacy of lithium for bipolar depression exist, they were not included in our systematic review because of the difficulties associated with their interpretation owing to the probable carryover effects of lithium treatment into placebo periods. We were however able to locate eight published masked, placebo-controlled trials ^{Reference Zornberg and Pope36} and two active treatment ^{Reference Fieve, Platman and Plutchik30,Reference Mendels, Secunda and Dyson37} comparison studies containing outcome data for lithium treatment of bipolar depression for 174 participants. All but one has been interpreted as indicating a benefit for lithium over placebo.

Despite the established role of lithium as the gold standard mood stabilising medication, and its apparent anti-suicide properties, it is worth noting that high-quality trials confirming its efficacy in the short-term treatment of bipolar depression are limited. Furthermore, there is a significant absence of data informing clinicians how to treat participants who become depressed while on therapeutic doses of lithium.

Given that rates of clinical response are usually higher than remission, our finding that the NNT for response to mood stabiliser monotherapy v. placebo was greater than the NNT for remission seems counterintuitive. Closer examination of the data shows that in the studies that comprise the remission data, the percentage of participants remitting in the active treatment groups were lower than the rates of response. This is in keeping with what might be expected based on clinical experience and existing trial data. However, remission rates in the control groups of these studies were much lower than those for response, which led to the higher relative likelihood of remission compared with response (and subsequently, lower NNTs). The finding that fewer participants need treatment to achieve remission than response is likely because of the fact that the studies that reported remission data were among those with the highest remission and response rates. In other words, the studies that were included in the remission analysis were not a random sample of all studies. These studies actually had unusually high rates of treatment success (response and remission) and thus were not broadly representative of all studies that compared mood stabilisers with placebo.

One possible criticism of the current study is the appropriateness of synthesising data on a group of pharmacologically heterogeneous agents. Although this was considered, it was felt that classifying them as a single group was a reasonable strategy given that there is little evidence of striking differences in the success rates of treating bipolar depression with these agents. This decision was further supported by the absence of significant differences between individual drugs in our own sensitivity analyses.

Overall, the results of this systematic review and meta-analysis support the use of mood stabilising medications as first-line treatments for people with acute bipolar depression. Studies examining combination treatments are urgently required as there are few studies to guide treatment decisions when individuals become depressed while on mood stabilising medications. The modest effect sizes observed in studies of each class of medication examined, however, suggest that simply optimising or combining currently available medications is unlikely to eliminate the morbidity associated with this phase of illness. Novel treatments, both pharmacological and otherwise, are required for the depressed phase of bipolar disorder.