Hormones interact non-linearly in osteoarthritis, study maps sex differences

Osteoarthritis has long been known to strike women far more often than men, a disparity so consistent that researchers have suspected sex hormones must be involved. Yet decades of studies examining individual hormones in isolation have produced contradictory and often frustrating results, leaving clinicians without a coherent explanation for why women bear the greater burden of this degenerative joint disease. A new study published in Biology of Sex Differences now argues that the problem lies in the way the question has been asked. Rather than acting alone, sex hormones appear to shape osteoarthritis risk through a dynamic, non-linear network of interactions whose effects depend critically on a person’s sex, menopausal status, and circulating estrogen levels.

The research, led by Bin Deng and Zugui Huang of the Department of Endocrinology at Nanping First Hospital Affiliated to Fujian Medical University in China, together with colleagues including corresponding author Xiaobin Ji, analyzed data from 13,848 adults drawn from the U.S. National Health and Nutrition Examination Survey (NHANES), covering the 2013–2016 and 2021–2023 survey cycles. NHANES, a nationally representative program run by the National Center for Health Statistics, provided the team with both hormone measurements and physician-diagnosed osteoarthritis status, along with the complex survey design information needed to generalize findings to the broader U.S. population. The work was supported by the Fujian Provincial Natural Science Foundation of China.

At the heart of the study is the free androgen index, or FAI, a calculated measure that reflects the biologically active fraction of testosterone available to tissues. Instead of treating FAI as a standalone predictor, the researchers positioned it on a two-dimensional plane alongside estradiol, the principal estrogen, and asked how osteoarthritis risk varies across the combined landscape of the two hormones. This approach embodies the team’s central hypothesis: androgen activity influences osteoarthritis not in isolation but within a hormonal ecosystem whose meaning changes with context.

To probe that ecosystem, the investigators deployed an unusually integrated analytical framework. They began with complex-survey regression models that respect the weighting and clustering of the NHANES design, estimating conventional odds ratios for the association between FAI and osteoarthritis within three strata: post-menopausal women, pre-menopausal women, and men. They then layered on two-way restricted cubic splines and generalized additive models, or GAMs, flexible statistical techniques that allow the risk surface to bend, curve, and tilt across the estradiol–FAI plane rather than forcing it into a single straight line. Finally, they turned to SHAP-based machine learning, an interpretability technique derived from cooperative game theory that quantifies how much each variable contributes to a model’s predictions and reveals where, along a variable’s range, its influence intensifies or reverses.

The regression results alone told a strikingly conditional story. Higher FAI was significantly associated with lower osteoarthritis risk only in post-menopausal women, with an odds ratio of 0.85 and a 95 percent confidence interval of 0.74 to 0.97, and a P value of 0.021. In pre-menopausal women the association pointed in the same protective direction, with an odds ratio of 0.75, but did not reach statistical significance (P = 0.145). In men, the association was essentially flat, with an odds ratio of 0.94 and a P value of 0.714. In other words, the same hormone told three different stories depending on who was carrying it and what else was circulating in their bloodstream.

The non-linear models went further, revealing distinct regional patterns that a simple regression could never capture. In post-menopausal women, the team identified what they describe as a “buffering” pattern. The GAM analysis was statistically significant across this group (P = 0.002), and a regional synergy measure averaging 0.077 highlighted one particular zone of the hormonal landscape: the quadrant where both estradiol and FAI are low. Post-menopausal women who had neither robust estrogen nor robust androgen activity occupied the highest-risk corner of the map. This finding suggests a possible biological redundancy, in which estrogen and androgen signaling can each, to some extent, compensate for the absence of the other in protecting joint tissue — and where losing both at once removes that safety net.

Men displayed a different geometry altogether. The GAM surface for males was significant at P = 0.042 and showed a “cross-over” pattern, meaning the direction of the androgen association flipped depending on estradiol levels. In one estrogenic context, higher androgen activity accompanied lower osteoarthritis risk; in another, the relationship inverted. Pre-menopausal women, by contrast, served almost as a natural negative control: their risk surface was essentially flat, showing no meaningful regional structure across the estradiol–FAI plane. This makes biological sense, the authors note, because pre-menopausal women maintain comparatively high and stable estrogen levels, which may mask or override any modulating influence of androgen activity.

A key element of the study’s statistical honesty lies in how it treats its own boundaries. Using SHAP dependence plots, the team identified candidate thresholds where the risk landscape appeared to change — for example, an estradiol level of approximately 2.63 and a log-transformed FAI of approximately −0.54 in post-menopausal women. But the researchers are explicit that these cut-points are hypothesis-generating indicators, not validated clinical thresholds, and they formally flag them as requiring independent confirmation before anyone should use them in practice. This restraint matters in a field where data-driven boundaries are easily overinterpreted.

Perhaps the most humbling quantitative result came from the machine learning stage. When the SHAP analysis tallied the standalone predictive contribution of each sex hormone, the values were remarkably small, with mean absolute SHAP values no greater than 0.0135. Sex hormones, in other words, are not powerful standalone predictors of who develops osteoarthritis. Their influence is real but contextual, operating as a modulator of risk within a larger hormonal and physiological background rather than as a primary driver. This may explain why single-hormone studies have been so inconsistent: a weak contextual signal will only emerge when the context itself is modeled.

The implications stretch beyond osteoarthritis. The pronounced female predominance in osteoarthritis incidence has long implicated the hormonal system, and this study provides a conceptual scaffold for understanding that disparity through steroid interaction rather than through any single hormone’s deficit or excess. It suggests that the menopausal transition is not merely a drop in estrogen but a reshuffling of an entire hormonal network, and that the joint consequences of that reshuffling depend on where a person lands in the multidimensional hormone space. The low-estradiol, low-FAI quadrant in post-menopausal women stands as the robust core of the findings — the one pattern the authors identify as most defensible.

The study does have limits that the authors acknowledge. NHANES data are cross-sectional, capturing a single moment in time, so the analysis cannot establish whether hormone levels precede the development of osteoarthritis or change as a consequence of the disease and its behavioral sequelae. Hormone measurements were taken at one point, yet the menstrual cycle, diurnal rhythms, and the years surrounding menopause all create substantial within-person fluctuation. Osteoarthritis status relied on self-reported physician diagnosis, which introduces the possibility of misclassification, and the candidate thresholds identified by SHAP remain unvalidated. As a secondary analysis of de-identified, publicly available data, the study required no new ethical approval, since the original NHANES protocols were approved by the NCHS Research Ethics Review Board and all participants provided written informed consent.

Even so, the methodological message may prove as influential as the biological one. By fusing survey-weighted regression, splines, GAMs, and interpretable machine learning, the researchers demonstrated that the tools of modern data science can map a hormonal ecosystem rather than merely test a coefficient. Their work suggests that future studies of sex hormones in chronic disease should routinely model interactions, test for non-linearity, and stratify by life stage — because the answer to “does testosterone protect against osteoarthritis?” turns out to be: it depends on estrogen, and on sex, and on menopausal status, and on where on the map you happen to be standing. For the millions of people, disproportionately women, facing the pain and disability of osteoarthritis, that nuanced map may one day help identify who is most at risk and when hormonal context matters most.