The Increasing Predictive Validity of Self-Rated Health

Ask a stranger one simple question — "How is your health?" — and their answer can predict whether they'll be alive a decade later better than almost any single clinical test. Self-rated health, a one-question survey measure where people classify their own health on a scale from excellent to poor, has repeatedly been shown to be a remarkably strong predictor of mortality. A recent meta-analysis found that people who report poor health face about twice the risk of dying from any cause compared to those who report excellent health. This relationship holds even after adjusting for diagnosed diseases, biomarkers, and functional limitations. But here's the tension that motivated Jason Schnittker and Valerio Bacak's study: some scholars worried that this signal was degrading. Medicalization — the process by which ordinary sensations and life problems get recast as medical disorders — has expanded rapidly, creating new diagnostic categories faster than old ones disappear. Direct-to-consumer advertising has grown sharply since nineteen eighty-nine. Internet searches for health information rose from twenty-five percent of American adults in two thousand to fifty-nine percent by two thousand ten. In any given month, seventy-five to eighty percent of people experience at least one illness or injury, even though most never see a doctor. There was a fear that all this noise — more conditions labeled as diseases, more sensitivity to minor symptoms, and higher expectations for good health — would contaminate people's self-assessments, inflating reports of poor health among those who weren't actually at elevated mortality risk. To formalize this, Schnittker and Bacak used a two-by-two framework of concordant and discordant assessments. A concordant case is someone who rates their health as poor and does in fact die early, or who rates it as excellent and lives long. A discordant case is someone who calls themselves sick but lives a full life, or someone who says they're fine but dies unexpectedly. Medicalization, the argument suggests, should produce more of the latter — more noise and more off-diagonal cases — weakening the link between what people say and what their bodies do. The data used to test this comes from the nineteen eighty to two thousand two waves of the General Social Survey, a repeated cross-sectional study of American adults, linked to National Death Index records through two thousand eight. The analytic sample includes twenty-three thousand three hundred seven respondents. Schnittker and Bacak used Cox proportional hazards models — a standard survival analysis tool that estimates relative risk of dying over time — with age handled through entry age and censoring rather than as a covariate, and they controlled for sex, marital status, and race. What they found was the opposite of what the theoretical case predicted. The relationship between self-rated health and mortality strengthened substantially between nineteen eighty and two thousand two. People, on average, got better at knowing when they were truly sick. In the baseline model, the dose-response pattern is already clear: reporting good health carries about a fifteen percent higher hazard of death than reporting excellent health. Fair health carries about a thirty-seven percent higher hazard. Poor health carries about a seventy-eight percent higher hazard. But when the authors model how these associations changed over time, the shift is striking. For the poor health category, the hazard ratio relative to excellent health was one point one six four in nineteen eighty. By two thousand two, it had grown to three point four zero five. The annual interaction term is one point zero five zero — meaning the hazard ratio for poor health compounded at roughly five percent per year over those twenty-two years. This widening holds across all response categories, not just the extremes: the gap between excellent and good health grew, and so did the gap for fair health. Schnittker and Bacak plotted survival curves for the four combinations of health rating and era — poor and excellent in nineteen eighty, poor and excellent in two thousand two — and the two thousand two separation easily encompasses the nineteen eighty separation. The signal didn't erode. It sharpened. So the next question is: why? The authors systematically worked through the obvious candidates. First, education. More schooling does improve the accuracy of self-rated health — the interaction between self-rated health and years of schooling carries a hazard ratio of one point zero one zero, with a p-value below zero point zero one. Likewise, higher verbal cognitive ability, measured with the General Social Survey's ten-item wordsum vocabulary test scored zero to ten, also helps: its interaction with self-rated health has a hazard ratio of one point zero two nine, also with a p-value below zero point zero one, and wordsum itself is associated with lower mortality at zero point nine two zero. Both of these findings make intuitive sense — better-educated, more cognitively capable people may integrate health signals more accurately. But neither explains the time trend. When either variable is added to the model, the year-by-self-rated-health interaction remains statistically significant and nearly unchanged in magnitude. Second, changing causes of death. One plausible story was that accidental deaths — car crashes, falls, and other events that no one sees coming — were declining over this period, and their decline was mechanically improving the self-rated health signal because people can't anticipate accidents. Schnittker and Bacak re-estimated their models after removing deaths from external causes, cancer, and cardiovascular disease in turn. The trend persists. For the external-causes exclusion, the year-by-poor-health interaction is still one point zero five zero and highly significant. Accidental deaths aren't driving the result. Having closed those doors, the paper turns to health information exposure — and this is where the findings get both most compelling and most specific. In the final two General Social Survey waves, two thousand and two thousand two, the questionnaire included an information module asking respondents how they sought health information. The fraction reporting any frequent health-information seeking rose from zero point one eight two in two thousand to zero point two four one in two thousand two. Critically, those who sought information gave more valid self-reports: the self-rated health hazard ratio among information seekers was two point one six four, compared to one point three seven seven among non-seekers — a substantial difference. However, the source of that information turned out to matter as much as the fact of seeking it. For individuals who consulted a doctor for health information, the self-rated health hazard ratio was one point nine four eight — indicating a strong and statistically significant link between their self-assessment and actual mortality. For those who did not consult a doctor, it was one point four one nine. A similar pattern appeared for traditional media: those who sought health information from daily newspapers had a self-rated health hazard ratio of two point two seven eight, versus one point four nine zero among non-readers. The internet tells a different story. Among respondents who did not use the internet for health information, the hazard ratio was one point five five four, which is highly significant. Among those who did go online for health information, it dropped to one point two two three — and was no longer statistically significant. Internet health information appears to weaken rather than strengthen the link between self-assessment and survival. Schnittker and Bacak are careful about what to make of this. The information module only appears in two survey waves, with a sample of one thousand seven hundred twenty-nine respondents and two hundred three deaths, so these source-specific results are suggestive rather than definitive. However, the pattern is consistent with what we know about how online health information works: it's often organized to help users reach a self-diagnosis, it tends to surface rare conditions, it may increase sensitivity to symptoms without improving specificity, and it requires strong reading-comprehension skills to evaluate critically. The result is that internet users may be rating themselves as sicker for conditions that carry low mortality risk — exactly the discordant pattern the medicalization argument predicted, just concentrated in one specific channel. What the full picture suggests is that the quality of self-rated health as a mortality predictor tracks the quality of information people receive about their own health. Clinical contact and traditional media appear to give people a clear signal. Early internet use, at least in two thousand and two, may have given them noise. The broader implication is worth considering. Humans are, on average, getting better at knowing when they are truly sick. Self-rated health is not a sociological curiosity or a rough proxy for objective measurement. It reflects genuine self-knowledge that tracks survival, and that knowledge seems to be improving as health information becomes more available through reliable channels. For public health researchers, a single survey question remains a valid — and increasingly valid — tool for measuring population health. For the rest of us, it's a reminder that what we feel about our health carries real information, and that where we learn about health shapes whether we can interpret those feelings accurately. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.