Reverse Causation in Ageing and Longevity Research

Who This Is Useful For

This page is useful for readers trying to understand why some observational findings in ageing research look biologically surprising, especially when a marker usually considered favorable appears associated with worse survival in older adults. It is particularly relevant when interpreting studies on physical activity, body weight, cholesterol, blood pressure, dementia risk, or late-life biomarkers. [3] [5] [6] [7] [9]

Reverse causation is one of the main reasons observational longevity findings can be misread. Instead of exposure causing outcome, the developing outcome can alter the measured exposure. In ageing research, this is common because disease, frailty, and the dying process often begin reshaping activity levels, appetite, body weight, blood pressure, lipids, and healthcare use years before a formal event is recorded. [1] [2] [5] [6] [7] [9]

That timing problem matters because many longevity papers are framed as if a baseline measure cleanly predicts future ageing. In reality, the measure may already sit inside an ongoing disease trajectory. When that happens, the association can exaggerate, weaken, or even invert the apparent effect of the exposure. [1] [2] [3] [4]

Concept at a Glance

1. What Reverse Causation Means

In ordinary causal language, reverse causation means that the arrow is running at least partly from the outcome process back to the exposure. A variable measured as an apparent predictor may actually be changing because the disease, disability, or mortality process has already begun. This is conceptually distinct from confounding, though both can appear in the same study. [1] [2]

The problem is especially important in cohort studies because a temporal ordering such as "exposure was measured before diagnosis" does not guarantee biological ordering. If the outcome has a long preclinical phase, then a baseline measurement can still be downstream of disease onset. [2] [5]

2. Why Ageing Research Is Especially Vulnerable

Many ageing-related outcomes emerge gradually. Dementia, frailty, disability, multimorbidity, and mortality often follow long trajectories during which appetite, mobility, exercise tolerance, inflammatory state, body composition, and treatment exposure are already changing. That makes it harder to identify a clean "before disease" measurement window than in acute conditions. [2] [5] [6] [7] [9]

The issue becomes stronger in late-life cohorts because participants have already been shaped by prior survival, comorbidity, frailty, and selective participation. A value measured at age 80 or 90 may not mean the same thing biologically that it meant at midlife. [6] [7] [9]

3. Physical Activity and Dementia

Physical activity is a useful example because the association looks plausible in both directions. Physical inactivity could contribute to dementia risk through vascular and metabolic pathways, but developing dementia can also reduce activity years before diagnosis. In a large individual-participant meta-analysis, inactivity measured within 10 years of diagnosis was associated with dementia, whereas inactivity measured 10 or more years before diagnosis was not. That pattern is consistent with reverse causation from the prodromal phase. [5]

Similar concerns appear in mortality studies. Analyses of physical activity that rely on recent measurements or short follow-up tend to show stronger protective associations than analyses using repeated measures and lagged exposure windows, suggesting that illness-related declines in activity can otherwise inflate apparent benefit. [3] [4]

4. BMI, Weight Loss, and Mortality

Late-life associations between low body mass index and higher mortality are often interpreted too literally. Longitudinal data show that BMI tends to decline in the years before death, with steeper declines in decedents than survivors. That means low BMI in older adults can be a marker of illness burden or approaching death rather than a straightforward cause of mortality. [7]

A related issue appears with weight loss. Observational studies often associate weight loss with higher mortality, but randomized evidence on intentional weight loss in older adults does not support a simple interpretation that weight loss itself is inherently harmful. The observational pattern can be partly driven by illness-related, unintentional weight loss. [8]

5. Cholesterol, Blood Pressure, and Terminal Decline

In older populations, low cholesterol and low blood pressure can look paradoxically associated with worse survival. Longitudinal evidence shows that total cholesterol declines in the last years of life, which can make low cholesterol appear harmful in non-randomized late-life analyses. [6]

Blood pressure shows a related interpretive problem. Observational work in frailer older adults often finds that lower systolic pressure is not clearly protective and may correlate with worse outcomes, but this can reflect reverse causation from frailty, functional limitation, and physiological decline rather than proof that lower pressure caused mortality. [9]

6. Protopathic Bias as a Special Case

A specific form of reverse causation occurs when a treatment is started for early symptoms of a disease that has not yet been diagnosed. In pharmacoepidemiology this is often called protopathic bias. It can make the treatment appear to raise risk when it was actually prescribed because the disease process had already started. [11]

Ageing research encounters this problem in areas such as sleep medications, anxiety treatments, or symptom-directed prescribing before dementia diagnosis. Analyses that apply longer lag periods often attenuate these associations, which is one reason lag choice matters so much in this literature. [11]

7. How Researchers Try to Reduce the Problem

There is no single fix, but several design choices help. Researchers may exclude people with baseline disease, remove events occurring in the first years of follow-up, use repeated measurements instead of a single baseline value, and separate exposure assessment from the risk period with lagged analyses. These steps do not eliminate bias automatically, but they test how much the association depends on likely outcome-driven changes in the exposure. [2] [3] [4]

Stronger causal designs can also help. Randomized trials reduce reverse causation because allocation is not determined by the participant's emerging disease trajectory, and Mendelian randomization can reduce some forms of reverse causation because genotype is fixed before disease onset. Even so, those designs have their own assumptions and limitations. [8] [10]

8. Warning Signs When Reading a Study

• The exposure is measured close to diagnosis or death: recent values are more likely to already reflect disease-related change. [3] [5] [6]
• The reported association weakens after excluding early follow-up: that often suggests at least part of the original finding reflected reverse causation. [3] [4] [5]
• The finding looks paradoxical in late life: examples include low cholesterol, low BMI, or low blood pressure linking with higher mortality. These may be real markers of vulnerability without being the original cause. [6] [7] [9]
• The exposure could plausibly change during a prodrome: this is especially relevant for behavior, cognition, body composition, and symptom-driven prescribing. [5] [11]

What This Does Not Mean

• It does not mean every late-life association is false; some exposures can both influence outcomes and be influenced by early disease. [1] [5]
• It does not mean lagging the data always solves the issue; an inappropriate lag can remove real causal exposure time or fail to cover a long prodromal window. [3] [5] [11]
• It does not mean observational ageing research is useless; it means causal interpretation depends heavily on timing, repeated measurement, and sensitivity analyses. [2] [4]
• It does not mean randomized or genetic designs are perfect; they reduce some reverse-causation problems but introduce other assumptions and practical limits. [8] [10]

Practical Interpretation Examples

• If a study finds that inactive older adults have much higher dementia risk: check how long before diagnosis activity was measured and whether long-lag analyses gave the same result. [5]
• If low cholesterol predicts higher mortality after age 80: consider whether the paper examined cholesterol trajectories and terminal decline rather than assuming low cholesterol was the original cause. [6]
• If low BMI appears harmful in older age: look for evidence on recent weight loss, multimorbidity, smoking, and years-to-death patterns. [7]
• If a medication seems linked to dementia: ask whether it might have been prescribed for early symptoms of the same disease process and whether longer exposure lags were tested. [11]

Related Reading

• Correlation, Confounding, and Causation in Longevity Research
• Selection Bias and Healthy User Bias in Longevity Studies
• Randomized Controlled Trials vs Observational Studies in Longevity Research
• How to Read the Methods and Limitations of a Longevity Study

Summary

Reverse causation is a central interpretive problem in ageing and longevity research because many exposures can be altered by disease long before diagnosis or death. The strongest readings of an observational finding therefore come from asking whether the exposure was measured early enough, whether repeated and lagged analyses were used, and whether the result survives designs less vulnerable to outcome-driven feedback. [2] [3] [5] [10]

References

1. Hernan, M. A., & Robins, J. M. (2006). Journal of Epidemiology and Community Health. https://pmc.ncbi.nlm.nih.gov/articles/PMC2652882/
2. Hernan, M. A., & Robins, J. M. (2016). American Journal of Epidemiology. https://pmc.ncbi.nlm.nih.gov/articles/PMC4832051/
3. Lee, D. H., et al. (2021). European Journal of Epidemiology. https://pmc.ncbi.nlm.nih.gov/articles/PMC8035269/
4. Mok, A., et al. (2019). International Journal of Epidemiology. https://pmc.ncbi.nlm.nih.gov/articles/PMC7124507/
5. Kivimaki, M., et al. (2019). BMJ. https://pmc.ncbi.nlm.nih.gov/articles/PMC6468884/
6. Charlton, J., et al. (2018). Journals of Gerontology Series A. https://pubmed.ncbi.nlm.nih.gov/29028914/
7. Pai, H., & Gulliford, M. C. (2022). BMJ Open. https://pmc.ncbi.nlm.nih.gov/articles/PMC9341213/
8. Shea, M. K., et al. (2011). American Journal of Clinical Nutrition. https://pmc.ncbi.nlm.nih.gov/articles/PMC3155925/
9. Supiano, M. A., Pajewski, N. M., & Williamson, J. D. (2018). Journal of the American Geriatrics Society. https://pmc.ncbi.nlm.nih.gov/articles/PMC5777872/
10. Davey Smith, G., & Hemani, G. (2014). Human Molecular Genetics. https://pmc.ncbi.nlm.nih.gov/articles/PMC4170722/
11. Penninkilampi, R., & Eslick, G. D. (2018). CNS Drugs. https://pubmed.ncbi.nlm.nih.gov/29926372/