Distortion in Risk Estimation Driven by Study Period Adjustment and Young-Age Event Sparsity: A Critical Evaluation of Secondary Analyses on the Nagoya City HPV Vaccination Survey

In the world of public health, researchers often look for patterns in large groups of people to understand whether a medical treatment causes harm or provides benefit. A central challenge in this work is separating the effects of a treatment from the natural changes that happen as people grow older. For instance, if a study tracks teenagers over several years, the types of health issues they experience will naturally shift as they move from childhood into adolescence. To get an accurate picture of a vaccine's safety, scientists must carefully adjust their calculations to account for these age-related changes, ensuring that a symptom appearing in a teenager is not mistakenly blamed on a shot given years earlier, nor is a real side effect hidden by the noise of normal growing pains. When these adjustments are done correctly, the data reveals the true relationship between a vaccine and a health outcome. When they are done incorrectly, however, the numbers can tell a completely different, and potentially misleading, story.

A recent evaluation by Sadao Suzuki, a professor emeritus at Nagoya City University, scrutinizes a specific statistical method used to analyze data from a major survey on the human papillomavirus (HPV) vaccine in Nagoya City. The original survey, conducted in 2015, tracked 24 different symptoms in young people to see if the vaccine increased the risk of developing them. While the original analysis found no evidence that the vaccine caused these symptoms, a later study by researchers Yaju and Tsubaki challenged those findings. They proposed a new way of looking at the data that involved a specific adjustment for the "study period," which is the length of time between when a person received their first vaccine dose and when the survey was conducted. Suzuki's paper argues that this new method contains a fundamental mathematical flaw that distorts the results, creating the illusion of danger where none exists.

The core of the problem lies in how the researchers handled the timing of symptoms. In the original survey, some young people reported symptoms that started before they ever received the vaccine. The later study decided to remove these early cases from the vaccinated group, reasoning that a symptom starting before the shot could not be caused by it. Suzuki points out that while this seems logical on the surface, the way it was combined with the "study period" adjustment created a hidden trap. The method treated the vaccinated and unvaccinated groups differently. For the vaccinated group, the observation window started only after the first shot. For the unvaccinated group, the window started at age twelve, regardless of when the survey took place. This created an uneven playing field where the unvaccinated group was observed for a longer stretch of their early youth than the vaccinated group.

This structural imbalance became critical because of a natural phenomenon known as "young-age event sparsity." Many of the symptoms tracked in the survey are rare in very young children but become more common as teenagers get older. Because the unvaccinated group was observed starting from age twelve, their data included a higher proportion of these rare, early-life events simply due to the longer observation window. The vaccinated group, whose clock started later, missed many of these early occurrences. When the statistical model tried to compare the two groups, it interpreted this difference in timing as a difference in risk. The result was a mathematical artifact that artificially inflated the risk numbers for symptoms that are rare in young children, making it look as though the vaccine was causing them, when in reality, the calculation was just counting the same events differently.

Suzuki's analysis breaks down the later study into three different models to show how the error compounds. The first model, which simply removed the early symptoms, already showed a bias, but the numbers still held some resemblance to the original findings. The second model introduced the "study period" adjustment, which caused the correlation with the original, reliable data to drop sharply. In this stage, the method began to systematically overestimate risks for the rare, early-life symptoms while underestimating risks for common ones. The final model added a complex interaction term that assumed the risk changed steadily over time. This step forced the mathematical results to peak at the very beginning of the study period, a point where the data was extremely thin and unreliable. In this final version, the risk for one specific symptom, which the original study found to be lower in vaccinated people, was calculated to be more than six times higher.

The paper concludes that these dramatic shifts are not discoveries of new risks but are the inevitable result of a broken mathematical tool. The author of the later study focused their discussion on the few symptoms where their new method showed a high risk, ignoring the fact that the same method produced impossible results for other symptoms, such as suggesting the vaccine might protect against a common ailment. By selecting only the results that supported their hypothesis and ignoring the contradictions, the later study presented a skewed view of the data. The evaluation confirms that the original survey's findings remain valid and that the alternative analysis relies on a structural bias that distorts the truth. The lesson here is that in epidemiology, the way you measure time and define your groups is just as important as the data itself; if the measurement tool is flawed, the picture it paints will be false, no matter how sophisticated the math appears.