In the mid-1990s, a team led by David Burr at Indiana University screwed strain gauges directly into the shinbones of two living volunteers and asked them to run. Uphill, downhill, zigzagging, sprinting — the conditions Israeli infantry recruits train under. The hardest efforts registered just under 2,000 microstrain, roughly three times what walking produced. Even at a downhill sprint, the tibia never went higher.

Five years later, at the same university, Yun-Feng Hsieh and colleagues loaded the forearms of adult rats and measured how much strain it took before new bone appeared on the surface. At one point along the bone the threshold was 1,343 microstrain. Three millimetres away it was 2,284. Three millimetres further, 3,074.

Put those findings side by side and you have the central problem in exercise prescription for bone. The load-threshold idea is sound, and it is the best available explanation for why walking programs do not build spine density. But the threshold is not a number anyone can hand you: it moves along a single bone, it was measured in a species whose skeleton is not yours, and the strains a human can generate voluntarily sit at or below it. What the trial evidence supports is looser and more useful than "cross 2,000 microstrain" — bone responds in proportion to how hard the load is, and most programs prescribed for osteoporosis are not hard.

Walking does not fail at everything — it fails at the spine

Two independent meta-analyses have asked what walking alone does to bone mineral density (BMD) in women after menopause, and they agree. Martyn-St James and Carroll pooled eight controlled trials in 2008 and found no significant change at the lumbar spine: a weighted mean difference of 0.007 g/cm² (95% CI −0.001 to 0.016, p = 0.09). At the femoral neck the pooled estimate was 0.014 g/cm² (95% CI 0.000 to 0.028, p = 0.05) — positive, but sitting exactly on the boundary, with substantial disagreement between trials (I² = 51.4%).

Ma, Wu and He repeated the exercise in 2013 with ten trials and reached the same place. Nothing at the lumbar spine, the radius, or the whole body. At the femoral neck, an effect appeared only in interventions running longer than six months. Their conclusion was blunt: walking as a singular exercise therapy does not preserve spine density.

The water evidence is thinner but points the same way. A 10-month randomised trial in Kosovo assigned 58 postmenopausal women with osteoporosis to land-based or aquatic exercise; lumbar spine density improved significantly more with land exercise (p < 0.001). One small trial without a non-exercising control is not a literature — but it is consistent with the mechanism, and no aquatic trial has shown the opposite.

None of this makes walking or water aerobics a waste of time. Both do things that matter a great deal. They are simply not the stimulus that builds vertebral bone, and they have been prescribed for decades as though they were.

The threshold is real, and it is not one number

Harold Frost's mechanostat — restated for clinicians in his 2003 update — treats bone as a negative feedback system with two switching points, like a thermostat with a heating and a cooling threshold. Strains above the upper point switch on the machinery that adds strength; strains below the lower one switch on the machinery that takes it away. It has organised the field for forty years.

Hsieh's rat experiment is what happens when you try to put a number on it. Loading the ulna daily for two weeks at five magnitudes produced a clean dose-response on the periosteal surface — more strain, more bone — but only once strain passed a threshold, and that threshold tracked how much load the site was already used to. The busiest region needed the most. On the endocortical surface, where strains were under 20% of periosteal values, there was no dose-response at all.

Charles Turner, a co-author on that paper, had already published a formal objection to the mechanostat. His 1999 argument was that bone cells react strongly to transients and then accommodate to steady signals — so a fixed strain threshold predicts the wrong thing over time, and the sequence of loading matters as much as its size. The threshold framing is contested inside the literature that produced it.

Site and condition	What was measured	Peak strain
Strain thresholds for new bone formation on the rat ulna (Hsieh et al., 2001) against peak strains measured in living human tibiae (Burr et al., 1996). Species and sites differ; the comparison is illustrative, not equivalent.
Rat ulna, proximal	Threshold for periosteal bone formation	1,343 µε
Rat ulna, mid-diaphysis	Threshold for periosteal bone formation	2,284 µε
Rat ulna, distal	Threshold for periosteal bone formation	3,074 µε
Human tibia, walking	Strain during activity, implanted gauges	About one-third of the running value
Human tibia, uphill and downhill zigzag running	Strain during activity, implanted gauges	Just under 2,000 µε

Rate may be doing work that magnitude cannot

Burr's group reported something in 1996 that gets less attention than the strain ceiling: strain rates during sprinting and downhill running reached 0.050 per second, higher than anything previously recorded in humans. Magnitude was capped. Speed of loading was not.

Milgrom and colleagues followed this up in 2001, inserting three strain-gauged staples into the mid-tibia of two volunteers and comparing calisthenics against jogging and walking. Zigzag hopping produced the highest compression, tension and shear strains and the highest compression strain rates; walking sat in the lowest or second-lowest group on every variable. Both studies are n = 2 — that is what invasive strain gauging in living humans permits — and both are the only direct evidence of what human bone actually experiences.

Sorting the trials by intensity roughly triples the effect at the spine

The most direct test of the load-threshold hypothesis in humans came in 2021, when Kistler-Fischbacher, Weeks and Beck re-analysed the exercise literature by intensity rather than pooling it. Their systematic review covered 100 trials and 120 interventions; the meta-analysis of the randomised subset covered 53 trials and 63 interventions. The distribution tells you most of the story: 19 low intensity, 40 moderate, and 4 high.

Intensity	Interventions	Lumbar spine (g/cm²)	Femoral neck (g/cm²)	Total hip (g/cm²)
Exercise effect on bone mineral density by intensity, pooled mean differences with 95% confidence intervals (Kistler-Fischbacher et al., 2021; 53 randomised trials).
Low	19	0.010 (0.005–0.015)	0.011 (0.006–0.016)	No significant effect
Moderate	40	0.012 (0.008–0.017)	0.011 (0.007–0.015)	0.008 (0.004–0.012)
High	4	0.031 (0.012–0.049)	No effect observed	Insufficient data

At the lumbar spine the high-intensity estimate is about three times the low- and moderate-intensity ones, which is the pattern the animal work predicts. At the femoral neck it vanishes — the authors attribute that to a lack of statistical power, since four interventions is not much to pool, but attribution is not evidence. Read the table honestly and it says: intensity clearly matters at the spine, and we do not yet know what it does at the hip.

The strongest trials are eight months long, closely supervised, and mostly from one group

LIFTMOR is the trial that changed the conversation. Watson and colleagues randomised 101 postmenopausal women with low bone mass (mean age 65 ± 5) to eight months of twice-weekly, 30-minute supervised high-intensity resistance and impact training — five sets of five repetitions above 85% of one-repetition maximum — or a low-intensity home program. Lumbar spine density rose 2.9 ± 2.8% in the training group and fell 1.2 ± 2.8% in the controls (p < 0.001). Femoral neck density held roughly steady in the training group (+0.3 ± 2.6%) while controls lost 1.9 ± 2.6% (p = 0.004). Compliance was 92%. There was one adverse event: a minor back spasm costing two missed sessions.

The safety question was examined separately. A 2019 analysis of the same cohort found no new vertebral fracture classifications after eight months of heavy lifting, and a measurable reduction in thoracic kyphosis in the training arm. A second trial, MEDEX-OP, randomised 115 women to the same protocol or a low-intensity Pilates-based comparator and reproduced the spine result: +1.9 ± 0.3% versus +0.1 ± 0.4%.

Three caveats belong in the same breath. These are eight-month trials with roughly 50 women per arm. Every session was supervised by trained staff, and the LIFTMOR authors are explicit that their safety finding applies under those conditions. And the senior author on LIFTMOR, MEDEX-OP and the intensity meta-analysis directs a clinic that sells the resulting program commercially — a conflict the papers disclose. The findings are consistent and were replicated; they have not yet been replicated by people with nothing riding on them.

A vibrating plate at 0.3 g complicates the whole story

If load magnitude were the only axis, whole-body vibration should do nothing. It applies accelerations around 0.3 g — a small fraction of what landing from a hop delivers.

A 2023 systematic review and meta-analysis of 23 studies (20 pooled) found otherwise. Across the parameters tested, the combination that carried high-quality GRADE evidence was high frequency (around 30 Hz), low magnitude (around 0.3 g), and a high cumulative dose (around 7,000 minutes) — and it improved lumbar spine density. Higher-magnitude vibration protocols produced results too, but at lower evidence quality.

That is difficult to reconcile with a simple magnitude threshold, and easy to reconcile with Turner's transient-response model. Bone appears to read several variables — how big the load is, how fast it arrives, how many times, and how unfamiliar it is. Magnitude is the one we can prescribe most easily, which is not the same as it being the only one that counts.

Every number in this piece is density, not a broken bone

This is the limitation that outranks the others. With one exception, everything above is a surrogate endpoint: grams of mineral per square centimetre, measured by dual-energy X-ray absorptiometry. Kelley and colleagues, pooling 28 effect sizes across roughly 1,600 participants in 2012, found small but significant improvements at both the femoral neck and lumbar spine and then estimated that these would cut 20-year relative fracture risk by about 11% and 10%. That is a modelled projection from density change, not a fracture count.

The exception runs through a different mechanism entirely. A 2019 meta-analysis of 25 randomised trials found that exercise reduced fall-related injuries in older adults (RR 0.879, 95% CI 0.832–0.928) and fractures specifically (RR 0.561, 95% CI 0.366–0.860). But the protocols that did it were balance training and combined programs — the fracture reduction came from not hitting the ground, not from denser vertebrae. Two distinct benefits, frequently sold as one.

Because the density changes at stake are small — one to three percent over eight months — the precision of whatever assessment is tracking them matters as much as the training does. That is a question worth asking of any bone assessment, whether it is DXA or radiation-free REMS (Radiofrequency Echographic Multi Spectrometry): what is the smallest change this method can actually detect?

What a load-first prescription looks like when it is written honestly

The 2017 Exercise and Sports Science Australia position statement is the clearest guideline document on this, and it does not offer a microstrain target either. It recommends impact activities and high-intensity progressive resistance training for bone, alongside balance and mobility work for falls, with adequate calcium and vitamin D, avoidance of loaded spine flexion, and modification of impact loading where osteoarthritis or frailty are present. Intensity is prescribed as a proportion of what the individual can lift, not as a property of the bone.

There is one more study worth holding onto, because it removes almost every confounder. Allison and colleagues asked 34 men, mean age 70, to hop briefly each day on one leg only, for twelve months, and then imaged both hips by CT. The exercised leg gained cortical mass surface density across the proximal femur (2.7%, p < 0.001), with local increases above 6% at the front and back of the femoral neck, and endocortical trabecular density rose 6.4%, with gains above 12% at specific sites. The other leg, attached to the same man with the same hormones, diet and medication, did not match it. And the gains were not spread evenly — they appeared where the load went.

That is what "bone responds to load, not to movement" actually looks like when you can see it. Nobody can tell you the number of microstrain your femoral neck requires. What the evidence supports is narrower and more actionable: the load has to be genuinely hard relative to your own current capacity, it has to be progressed, it goes where you put it, and — on the strength of the trials that exist — it should be supervised while you learn it.

This article is for education and is not a substitute for medical advice, diagnosis, or treatment. Talk to your physician about your individual situation.