How to Read Antidepressant Effect Sizes Without Overclaiming

The short answer

When you read that antidepressants beat placebo with a standardized mean difference (SMD) of about 0.30, the honest interpretation is this: on average, the drugs move depression scores in the right direction, and that movement is real but small. The 2018 network meta-analysis led by Andrea Cipriani and colleagues in The Lancet pooled 522 trials and 116,477 participants across 21 drugs, and found every one of them more effective than placebo. Yet the average effect landed in the range statisticians call small. Both facts are true at once. Learning to hold them together, rather than picking the one that fits a prior belief, is the whole skill of reading this literature.

What the number actually is

A standardized mean difference expresses a treatment effect in units of standard deviation, which lets you compare results measured on different depression scales. By the conventional benchmarks that Jacob Cohen proposed, roughly 0.2 is small, 0.5 is medium, and 0.8 is large. The Cipriani analysis put the pooled antidepressant-versus-placebo SMD near 0.30. Even the highest-performing agent, amitriptyline, reached only about 0.48, which the accompanying commentary by Chittaranjan Andrade in the Journal of Clinical Psychiatry described as still sitting in the small range.

Those benchmarks are conventions, not laws of nature. Cohen offered them as rough guides and warned against treating them as hard cutoffs. A 0.30 in a condition where few things work well is not interpreted the same way as a 0.30 in a domain full of highly effective treatments. Context sets the meaning; the number alone does not.

The paradox worth sitting with

Here is what trips people up. The same dataset that yields a modest SMD also shows favorable response rates. In the Cipriani work, odds ratios for treatment response ran from about 1.37 for reboxetine to 2.13 for amitriptyline, meaning most drugs roughly doubled the odds of a meaningful response relative to placebo. Andrade framed the tension precisely: you can be pleased by the strong response and remission figures or disappointed by the low SMDs, and both reactions draw on the same numbers.

Why do continuous and categorical measures diverge? An SMD averages the entire distribution of symptom change, including the many people who improve only a little and the placebo group, which in depression trials often improves substantially on its own. A response rate, by contrast, counts how many people cross a threshold, such as a 50 percent symptom reduction. A small average shift can still push a nontrivial share of people over that line. Neither metric is lying. They answer different questions, and a careful reader asks which question matters for the decision at hand.

Why small can still matter at scale

A modest average effect for an individual can be a large effect for a population. If a treatment nudges the whole distribution of outcomes, the number of additional people who reach response, translated into a number needed to treat, can be clinically worth acting on. Depression is common, recurrent, and costly in years of life impaired. When a widely used intervention produces even a small reliable improvement across millions of exposed people, the aggregate burden lifted is not small. This is the same logic public health uses for interventions with small per-person effects and enormous reach.

The reverse caution holds too. Population-scale benefit does not license overclaiming about any single person. The average tells you what to expect across many people, not what will happen to one. Effect sizes describe distributions, not destinies.

Reading the meta-analysis without being fooled

A few habits protect against overreach. First, check the certainty of the evidence, not only the point estimate. The Cipriani authors rated the certainty of their findings from moderate to very low, and a majority of the included trials carried at least moderate risk of bias. A precise-looking summary number can rest on shaky underlying studies. Second, note the population the estimate generalizes to. These results describe acute treatment in adults with major depression who were generally nonpsychotic and medically stable, and Andrade cautioned that they do not automatically transfer to refractory or complicated cases. Third, resist the ranking trap. A network meta-analysis can order drugs, but small differences between adjacent agents often fall within overlapping uncertainty, and the head-to-head rankings were less certain than the simpler drug-versus-placebo contrasts.

Fourth, remember that the placebo arm is doing real work. Much of the improvement seen in trial participants occurs in the placebo group, which is why the drug-minus-placebo difference, rather than the raw improvement on drug, is the honest measure of pharmacological effect.

A working translation

If someone tells you antidepressants "barely beat placebo," they are leaning on the SMD and ignoring the response rates. If someone tells you they are "highly effective," they are leaning on the response rates and ignoring the SMD. The disciplined reading names both: reliably better than placebo, with an average effect that is small, and a benefit that becomes meaningful when spread across a large population of people who need help. That sentence is longer and less satisfying than a headline. It is also what the evidence actually supports.

This article is educational and not medical advice. Decisions about starting, changing, or stopping any treatment belong to a person and their own clinician, informed by their specific situation.