Exercise beats no treatment for back pain, but only if the small trials stay in

Exercise therapy for chronic low back pain has a Cochrane review behind it, 249 trials deep, finding a pain reduction that meets the threshold its authors set for clinical importance [s1]. It also has a reanalysis of that same evidence, published in BMJ Open in 2025, showing that the effect falls below that threshold once the smallest trials are removed [s2]. Both are defensible readings of the same data, and the difference between them is the honest answer to the question.

What Cochrane found

The 2021 review assessed exercise treatment against no treatment, usual care, placebo or other conservative treatments in adults with non-specific low back pain lasting more than 12 weeks [s1]. It included 249 trials, conducted in Europe (122 studies), Asia (38), North America (33) and the Middle East (24) [s1]. Participants had a mean age of 43.7 years, and on average 59% of study populations were female [s1]. Sixty-one per cent of studies examined two or more different exercise types, and 57% compared exercise with a non-exercise comparison [s1].

The reviewers pre-specified what would count as clinically important: a 15-point difference in pain and a 10-point difference in functional limitations, on scales rescaled to 0–100 [s1].

Against no treatment, usual care or placebo, moderate-certainty evidence showed a pain reduction of 15.2 points (95% CI −18.3 to −12.2) at earliest follow-up — meeting their threshold [s1]. For functional limitations the reduction was 6.8 points (−8.3 to −5.3), which did not meet the prespecified threshold [s1]. Most trials were judged at risk of bias, including 79% at risk of performance bias, because exercise treatments are difficult to blind [s1].

What happens when the small trials are removed

The 2025 BMJ Open analysis re-examined the trials in that Cochrane review using trial-sequential and cumulative meta-analysis, asking when the evidence base first became conclusive and what the later trials added [s2].

Its finding on timing is stark. The superiority boundaries for pain and disability were crossed in 2004, after four and five trials respectively (358 and 415 participants) [s2]. A further 43 trials with 2,626 participants were subsequently included in the Cochrane review [s2]. The 2004 mean effect sizes — pain −12.85 (95% CI −24.89 to −0.81) and disability −6.67 (95% CI −11.27 to 3.36) — were similar to those reported by Cochrane in 2021 [s2].

Its finding on trial size is the more consequential one. Thirty-three of the pain trials and 36 of the disability trials had fewer than 70 participants [s2]. When those are excluded, the limits of the 95% confidence intervals exclude the clinically important differences: pain −8.8 (95% CI −11.38 to −5.63) and disability −4.27 (−6.12 to −2.24) [s2]. The authors conclude that including small studies has produced biased results in previous meta-analyses, that further trials of exercise against usual care are difficult to justify, and that exercise treatments might not have a clinically important effect in people with chronic low back pain [s2].

An apparent sign error in the source. The disability confidence interval given for 2004 runs from −11.27 to 3.36 [s2] — an interval that crosses zero, which would contradict the same sentence's claim that the superiority boundary had been crossed. The most likely reading is that the upper bound should be −3.36. Neither figure can be confirmed from the abstract, so the 2004 disability estimate should be treated as unreliable as printed. The trial-size finding, which is the paper's substantive contribution, does not depend on it.

How much exercise, and which kind

A Bayesian network meta-analysis published in 2024 modelled the dose-response relationship rather than a single average effect, using 82 trials with 5,033 participants and searching six databases to June 2023 [s3].

It found a non-linear, U-shaped dose-response between total exercise and pain [s3]. The maximum significant response occurred at 920 MET-minutes (standardised mean difference −1.74, 95% credible interval −2.43 to −1.04), and the minimum needed to achieve a clinically important pain improvement was 520 MET-minutes per week [s3]. The dose required varied by exercise type, with Pilates the most effective, and the clinical effect most pronounced with Pilates exercise [s3]. Certainty of evidence was very low to moderate across outcomes [s3].

A U-shaped curve means more is not simply better — the modelled benefit falls away above the peak, a finding that is unusual in this literature and rests on modelling rather than on trials that randomised people to high doses.

What a reader can conclude

Exercise for chronic low back pain is safe, widely recommended and supported by an enormous number of trials. What is genuinely contested is whether the size of its benefit clears the bar that patients and clinicians would recognise as worthwhile. On the Cochrane reading, pain relief does and function does not [s1]. On the reanalysis, neither does once small-study bias is stripped out [s2].

The dose evidence adds a further caution: the effect appears to depend on how much exercise is actually done, on a curve that has a peak rather than a plateau [s3]. Averages pooled across trials using very different doses will understate the benefit for some participants and overstate it for others.

None of this argues against exercising with back pain. It argues against the confidence with which a single number is usually quoted.

This article is informational and is not medical advice. Back pain with fever, unexplained weight loss, or new bladder, bowel or leg-weakness symptoms requires prompt medical assessment.

Sources

• Exercise therapy for chronic low back pain — Cochrane Database of Systematic Reviews, 2021-09-28
• Research waste in randomised trials of exercise treatments for chronic low back pain: trial sequential and cumulative meta-analyses by publication date and size — BMJ Open, 2025-03-25
• The Best Exercise Modality and Dose for Reducing Pain in Adults With Low Back Pain: A Systematic Review With Model-Based Bayesian Network Meta-analysis — Journal of Orthopaedic & Sports Physical Therapy, 2024-03-08