However, the observational nonrandomized design that was used makes the study susceptible to so-called "confounding by indication." Assignment to the two treatment groups was likely dependent on characteristics associated with the outcome rather than simply the implied obverse, or in other words, it is not straightforward to understand how many of the patients did poorly because they were discontinued from antidepressants versus how many were discontinued from antidepressants because they were doing poorly. As a consequence, the results might not hold up after adjustment for confounding factors, usually handled in a stratified or multivariate analysis. For instance, a rapid-cycling course, which is a risk factor for poor outcome and a common reason why antidepressants are discontinued, was apparently not measured or controlled for. The findings of the study would have been more informative had they resulted from adjusted analysis in the Cox proportional hazards model used, after control for all relevant clinical variables, such as the prevalence of rapid-cyclers, bipolar I and II subtypes, comorbidities (especially substance abuse), enrollment centers with possibly different treatment strategies, and other treatments given, especially mood stabilizers. Additional relevant baseline variables, not reported or considered (purportedly because of lack of statistical significance), could have been incorporated in a multivariable analysis because univariate p value comparisons are not informative in assessing potential confounders (2).