We read with interest Kessing et al’s timely and welcome paper ^{Reference Kessing, Hellmund, Geddes, Goodwin and Andersen1} supporting, by way of observational cohort study, the findings of BALANCE. ^{Reference Geddes, Goodwin, Rendell, Azorin, Cipriani and Ostacher2} Lithium again is shown to be superior to valproate for the management of bipolar disorder. The strength in this case comes from bridging the gap between the relatively brief follow-up in randomised control trials (RCTs) and the real-life situation faced by clinicians managing a lifelong illness of unpredictable course. Although the enriched study design in BALANCE aimed to maximise the generalisability of the findings to a clinical population, limitations inevitably remained in terms of including patients who had shown a differential previous response to either lithium or valproate, diagnostic heterogeneity within the sample population, and frequency of comorbidity compared with the general population. The limitations of observational cohort studies are multiple and well documented. One key concern is confounding by indication, but more general problems exist with group biases and masking of cause and effect relationships.

Kessing et al used ‘switch to’ and ‘add on’ as proxy outcomes for the efficacy of mood stabilisers. It would have been interesting, if possible, to separate the ‘switch to’ group from the ‘add on’ groups. The ‘add on’ outcome probably represents a treatment failure; however ‘switch to’ is likely to be a combination of lack of efficacy and poor tolerability. Indeed, their findings suggest that the initial, very rapid increase in incidence of switch/add on is related to tolerability rather than efficacy, whereas in BALANCE this finding would have been lost by drop-out during the run-in period. This is unlikely, however, to explain the superiority of lithium that is clearly present in both outcome measures.

It was previously argued that observational studies would overestimate treatment effects and that they hold little value in assessing therapies; however, comparative studies with RCTs, across various branches of medicine have now dismissed this. ^{Reference Black3} This sort of complementary approach, reconfirming findings from RCTs over long follow-up periods, is an important addition to the evidence base for treatment. This is especially true in areas where the disorder under investigation is chronic relapsing–remitting, and when the exclusion criteria of RCTs can often mean that external validity is low. If, as has been suggested, bipolar disorder is a heterogeneous condition with subtypes associated with preferential response to specific mood stabilisers ^{Reference Alda and O'Donovan4} (which can be identified by symptoms, clinical course and family history), then the observational study carries even more weight when compared with the RCT as it ‘allocates’ patients to treatments on the basis of predicted response, rather than randomisation. Bias can then be minimised by propensity score matching ^{Reference Rosenbaum and Rubin5} (controlling for unmeasured bias between study groups), although this method was not employed by Kessing et al.