The Biomarker Interaction Trap: A Causal Audit of Spurious Discoveries in NHANES

In the vast landscape of modern medicine, researchers often turn to massive digital archives of patient records to find hidden patterns that might predict disease. These databases, filled with thousands of measurements from blood tests and physical exams, act like a giant map where scientists hope to spot new connections between biological markers and health outcomes. The logic is straightforward: if two specific numbers in a blood test consistently rise and fall together in people with a certain illness, that relationship might reveal a new cause of the disease or a way to treat it. When a pattern appears in one group of people and then shows up again in a completely different group, scientists usually celebrate it as a confirmed discovery. However, this confidence can be misplaced. Just because a pattern repeats does not mean it is real; it might simply be a reflection of how the data was collected or who was included in the study. If the same hidden factor influences both the measurements and the disease in the same way across different groups, the pattern will look identical and convincing, even if it has nothing to do with biology.

This is the central puzzle James Brody set out to solve in a recent audit of the National Health and Nutrition Examination Survey, a massive, ongoing study of the American population. Brody examined three specific interactions between blood markers that had been identified as strong predictors of disease risk. These interactions had passed the test of replication, appearing in multiple independent groups of people, yet Brody demonstrated that each one was a statistical illusion rather than a genuine biological signal. He showed that these "discoveries" were actually the result of three distinct methodological traps: one caused by how the body's fluid levels change before a blood draw, another caused by the way doctors decide who gets medication, and a third caused by how population-wide averages fail to account for differences between men and women. By peeling back the layers of these artifacts, the study reveals that what looked like a complex biological dance was actually a simple mirage created by the data itself.

The first trap involves the way blood concentration changes when a person is dehydrated or has been standing for a long time. Blood is mostly water, and the proteins and fats floating in it are dissolved in that water. When the water volume in the blood drops slightly—perhaps because a person is thirsty or because a tourniquet was left on their arm too long during the blood draw—the remaining liquid becomes more crowded. This causes the measured levels of large molecules, such as albumin and cholesterol, to rise together, not because the body is producing more of them, but because there is less water to dilute them. In the study, researchers found a strong link between high albumin and high cholesterol that seemed to predict uncontrolled high blood pressure. However, when they adjusted the analysis to account for hydration status using simple markers like sodium levels, the link vanished completely. The apparent interaction was not a sign of disease; it was a measurement artifact caused by the physical state of the blood sample. Acting on this false signal could be dangerous, potentially leading doctors to prescribe diuretics to a patient who is already dehydrated, worsening their condition.

The second trap arises from the way doctors prescribe medicine, a phenomenon known as confounding by indication. In the real world, physicians do not give drugs to everyone; they give them to the sickest patients who need them most. This creates a situation where the group of people taking a specific drug looks very different from the group that is not, simply because the drug was chosen for their specific health problems. In the study, an analysis of patients taking statins (drugs for cholesterol) suggested that a combination of low hemoglobin and high creatinine predicted a failure to control blood sugar. This looked like a dangerous side effect of the drug. Yet, when the researchers looked at patients who were not taking statins but had the same low hemoglobin and high creatinine levels, they found the exact same pattern of poor blood sugar control. The link was not caused by the drug at all. Instead, these markers were signs of advanced kidney disease, a condition that is common in severe diabetes and often leads to the prescription of statins. The drug was merely a passenger on the journey of a severe disease, and the statistical pattern reflected the severity of the illness, not the effect of the treatment.

The third trap occurs when researchers apply a single set of rules to measure biological markers that naturally differ between men and women. Many substances in the blood, such as uric acid and hemoglobin, have different normal ranges for men and women due to hormonal differences. When a study uses a single cutoff point for the entire population, it inadvertently sorts people into groups based on their sex rather than their health. For instance, a "low" uric acid level might be normal for a woman but very low for a man, while a "high" hemoglobin level might be normal for a man but very high for a woman. In the study, an analysis suggested that having low uric acid or high hemoglobin was protective against diabetes, but having both anomalies at once was risky. This strange pattern disappeared when the researchers accounted for age and sex. The groups that appeared "protected" were simply younger men or younger women, who naturally have lower rates of diabetes. The interaction was not a biological discovery but a demographic sorting error, where the data was grouping people by their age and gender rather than by their metabolic health.

The implications of these findings extend far beyond these three specific examples. The study argues that the standard practice of finding a pattern in one group and confirming it in another is not enough to prove a biological truth if the underlying cause of the pattern is a systematic error that exists in both groups. Whether the error comes from how blood is drawn, how patients are selected for treatment, or how reference ranges are applied, the result is the same: a convincing but false story about disease. To avoid these pitfalls, the author proposes a simple checklist for researchers. Before declaring a new interaction, they should adjust for age and sex, check for signs of dehydration, and test whether the pattern holds up in people who are not taking the medication in question. If the pattern disappears after these checks, it was never a real biological signal to begin with. This approach does not dismiss the value of large data sets, but it insists that the search for truth requires more than just finding a repeating number; it requires understanding the context in which that number was measured.