This analysis supports the hypothesis that SSRIs begin to have observable beneficial effects in depression during the first week of treatment. The early treatment effect was seen on the primary outcome of differences in depressive symptom rating scale scores and on a secondary outcome of increased likelihood of achieving a 50% reduction in the HDRS score. The best-fitting model is described using a response variable in which the greatest absolute effect is observed in the first week and incremental responses by week diminish. This model was significantly better than all the other models, apparently excluding the possibility that treatment response from antidepressant drugs is subject to a period of delay.
The underlying placebo response was taken into account by analyzing the difference between the SSRI and placebo groups, and so the differences seen reflect drug-attributable effects. An early simple numerical difference in score on symptom rating scales may reflect an effect on particular symptoms rather than a true antidepressant drug effect. For example, an antidepressant agent with hypnotic effects could improve symptoms of insomnia before other effects were apparent. If this was the case, then it might be expected that a more global measure, such as the Clinical Global Impression ratings of improvement, would not show the same early benefits of treatment compared with placebo, and rating scale scores might show a stepped pattern of improvement after a period of delay. Neither of these patterns are apparent in this analysis, indicating that the benefits of SSRI treatment seen by the end of week 1 are true antidepressant drug effects.
A key question for physicians and patients is whether the early effects of SSRIs are clinically observable. This may indeed be the case. Of the improvement in symptom ratings attributable to treatment, that is, that seen in addition to placebo, approximately one third of the total effect after 6 weeks of treatment is seen in the first week (eg, HDRS: −1.07/−3.30=0.32). One week of treatment is also associated with an increased probability of achieving treatment response (RR, 1.64; 95% CI, 1.2 to 2.25). The absolute benefits of treatment for individuals depend on baseline risk; however, these benefits increase further across time, with, at a best estimate, a number needed to treat for 1 additional person to respond to SSRIs rather than placebo by week 6 of approximately one third of that by week 1. Therefore, individuals involved in treatment decisions will continue to need to wait several weeks for key treatment goals, such as remission,53
to be met.
Next, one has to consider the reliability of the estimates of effect calculated. Some sources of potential bias related to incomplete data may be of particular note in meta-analyses, including publication bias, failure to identify all available trials, and incomplete reporting of trial data.54
The search strategy used herein was comprehensive, identifying 6153 patients randomized to receive SSRIs and including data from 3618 patients compared with only 1549 in a recent analysis.55
However, there are likely to be eligible unpublished trials that could not be included in the analysis. Not all studies that took repeated outcome measurements presented the results for all time points. The effect of all these factors may be to overestimate the true treatment effect.56
However, there does not seem to be any compelling reason to expect the pattern of the response across time to differ in a systematic way between included and excluded trials.
A key issue is whether it is biologically plausible for SSRIs to produce early symptomatic response. Inhibition of serotonin transporters occurs rapidly in vitro and in vivo. A delayed response to antidepressant drug treatment is often linked with the time taken for a variety of adaptive neurobiologic changes to occur, for example, desensitization of serotonin 1A receptors and expression of neurotrophic factors such as brain-derived neurotrophic factor.57,58
However, there is evidence in human volunteers of downstream neurochemical59
effects of SSRI treatment in the first week of use that are associated in clinical populations with improvement.62,63
It is possible that the shape of the response curve for SSRIs across time is in some way determined by the design of the trials that contributed data. In particular, the use of last-observation-carried-forward analysis to account for missing data, which has become standard in trials of pharmaceutical treatments for depression and is used by approximately half of the trials analyzed herein, could make a constant effect appear log linear. This is because participants who leave a trial early would otherwise on average have been expected to improve regardless of their treatment allocation. Thus, last observation carried forward could add a negative bias to the results across time. An alternative and attractive hypothesis is that the effect of treatment may be constant on a relative scale, and as improvement occurs across time the absolute benefits achieved from each additional week decrease proportionately. Thus, the changes in rating scale scores () would be analogous to the results found for improvement in Clinical Global Impression score (), with the absolute benefit changing week to week while the relative benefit is constant across time. However, if either hypothesized effect applies to the present data, neither would indicate that the effects of treatment are delayed for a period, and, thus, they do not undermine the overall conclusion.
Why have many previous analyses not found evidence of antidepressant drug effects as early as the end of week 1? It is not surprising that statistically significant early differences are infrequently seen in individual RCTs powered to demonstrate treatment effects at a trial end point. There is an approximately inverse square law relationship between sample size and effect size such that to have equal power to find an effect of one third the magnitude requires a 9-fold increase in sample size.64,65
A variety of approaches can be taken in trial design to maximize sensitivity to early antidepressant drug effects, including more frequent assessments early in treatment and the use of pattern analysis or survival analytic approaches.8,18,66
The classic studies of Quitkin et al5,8
establishing the delayed-onset hypothesis differed in several respects from those analyzed herein. The participants included some with depressive illnesses of milder severity than those included herein; for example, in the 1987 study, the minimum score on the 21-item HDRS required for inclusion was 10. They also used a range of non-SSRI antidepressant agents, and as is often seen in clinical practice, these were titrated up to therapeutic doses during the first 2 weeks of treatment. This dose titration period, which is much less typical with SSRIs, may be the key difference, particularly because some other analyses using non-SSRIs with more rapid dose titration have found that effects begin to emerge more rapidly.9,12,45
One of these studies45
included a paroxetine treatment arm, finding a time of onset of its effects during the second week of treatment; one issue here may have been its relatively small size, with 82 patients randomized to 3 treatments.
The present analysis does not readily provide answers to some important related questions, such as whether there are particular patient characteristics, such as sex or disease severity at baseline, or features early in treatment that predict better eventual outcome. Studies66
specifically designed to address these questions can provide some answers; however, it may be that new, more sensitive measures of treatment response will have to be developed before we can reliably identify early response in smaller samples. It is intriguing to speculate that the types of investigation that have already revealed subtle early effects of SSRIs in the laboratory60
might be adapted in due course to the clinic. In summary, treatment with SSRIs is associated with symptomatic improvement in depression by the end of the first week of use. An early response is not necessarily a placebo response.