📊 statistics

A Probabilistic Framework for Estimating the Modal Age at Death

This paper proposes a probabilistic framework that models death counts with Poisson distributions and employs either independent gamma priors or second-order random-walk priors to estimate the full posterior distribution and credible intervals of the modal age at death, demonstrating superior accuracy and robustness compared to existing single-number estimators and bootstrap methods.

Original authors: Silvio C. Patricio, Paola Vazquez-Castillo

Published 2026-11-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Silvio C. Patricio, Paola Vazquez-Castillo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

To understand how long people are living, demographers often look at the average age at death. But averages can be misleading, especially when a few very young deaths pull the number down, hiding the reality that most people are living much longer. A more revealing measure is the "modal age at death." This is simply the single age where the most people die. If you were to line up every person who died in a country over a year and count how many died at age 80, how many at 81, and so on, the modal age is the number on the peak of that hill. Unlike life expectancy, which changes if a few children die young, the modal age tells us specifically about the survival of older adults. It tracks whether the peak of the death curve is sliding toward older ages, signaling that people are not just living longer, but that the entire process of aging is being pushed back.

For decades, researchers have tried to pin down this number, but it is notoriously difficult to measure with precision. In any given year, the number of deaths at age 89 might be only slightly higher than at age 90. Because these numbers are based on real human lives, they fluctuate randomly. A small, random bump in the data can make it look like the peak has shifted by a year, even if the underlying reality hasn't changed. Traditional methods usually spit out a single number as the answer, perhaps with a range of uncertainty calculated separately. This approach struggles to capture the fact that, in many cases, the data simply cannot distinguish between two neighboring ages. It treats the peak as a fixed point rather than a fuzzy area of probability.

A new study by Silvio C. Patricio and Paola Vazquez-Castillo, researchers at the University of Southern Denmark, offers a different way to look at this problem. Instead of forcing the data into a single number, they built a framework that calculates the entire probability distribution of the modal age. Imagine the researchers asking, "Given the deaths we observed, what is the chance the peak is at 89? What is the chance it is at 90? What is the chance it is 91?" By running thousands of simulations that account for the natural randomness in death counts, they produced a full picture of uncertainty. Their method treats the death rates at each age as a puzzle where the pieces are connected. They used a statistical approach that allows the data from neighboring ages to inform one another, smoothing out the random noise while respecting the actual shape of the mortality curve.

The researchers tested their method using simulated data where they knew the true answer in advance. They created scenarios with different population sizes and different patterns of death, including some where the data was messy enough to have a secondary, smaller peak. In these tests, their new method proved to be significantly more accurate than existing techniques. When the data was clean and showed a single peak, their estimates were far more precise than those derived from older, independent calculations. In some cases, the error in their estimates was more than ten times smaller than the error in the standard approach. Even when the data was complicated by a secondary peak, their method remained robust, performing nearly as well as the best existing smoothing techniques. Crucially, their method provided a reliable measure of confidence. While other methods produced intervals that were either too narrow to be trustworthy or too wide to be useful, the new approach consistently captured the true answer within its calculated range, covering the correct age in nearly every simulation.

The team then applied their method to real-world data for women in Denmark, France, Japan, and Sweden, looking at records from 1960 to the present day. The results showed a clear trend: the modal age at death has been rising steadily in all four countries, confirming that the peak of human mortality is shifting to older ages. In Japan, for instance, the peak moved from roughly 79 to 92 years over this period. The study also demonstrated the stability of their approach. When the researchers changed the starting point of their analysis—asking the model to ignore deaths below age 40, or below age 50, or 60—their results barely changed. This stability is a major advantage over older models, which could shift their answers significantly depending on which ages were included in the calculation.

Perhaps the most powerful application of this new framework appeared when the researchers looked at the year 2020, a time of global disruption. In France, the data suggested a decline in the modal age that year. While the standard error bars for 2019 and 2020 individually overlapped, making it hard to say if a change had truly occurred, the new method calculated the probability of the change directly. It found a 99 percent probability that the modal age had indeed declined. This ability to quantify the likelihood of a shift, rather than just stating a range of possibilities, offers a clearer tool for understanding how specific events impact longevity. The study concludes that by embracing the full distribution of uncertainty, rather than hiding it behind a single number, we can get a much truer picture of how human life is changing.

Technical Summary: A Probabilistic Framework for Estimating the Modal Age at Death

Problem Statement

The modal age at death ( $M$ ), defined as the age at which the life-table distribution of deaths peaks, is a standard indicator of longevity that tracks the postponement of deaths into later life. Unlike life expectancy, it is independent of mortality at young ages. However, estimating $M$ is challenging because the distribution of deaths is often flat near its peak. Consequently, small random fluctuations in single-year death counts can shift the identified maximum by a year or more.

Existing methods typically report $M$ as a single point estimate (e.g., via parametric Gompertz models, P-spline smoothing, or discrete derivative tests) with uncertainty assessed separately via bootstrap or delta methods. These approaches yield confidence intervals describing hypothetical repetitions of data but fail to assign a probability to each candidate age or reveal when neighboring ages have nearly equal support in the data. This paper addresses the need for a method that estimates the full posterior distribution of the modal age, thereby quantifying uncertainty directly in terms of the probability that a specific age is the mode.

Methodology

The authors propose a Bayesian framework that propagates uncertainty from death rates through the life table to the modal age. The core innovation is treating the modal age not as a fixed parameter but as a random variable derived from the posterior distribution of death rates.

1. Likelihood and Independent Priors

The model assumes death counts ( $D_x$ ) at age $x$ follow a Poisson distribution given the death rate ( $m_x$ ) and exposure ( $E_x$ ):
$D_x | m_x \sim \text{Poisson}(m_x E_x)$
Initially, the authors consider independent Gamma priors for the rates ( $m_x \sim \text{Gamma}(\alpha, \beta)$ ). Due to conjugacy, the posterior for each rate is also Gamma:
$m_x | D \sim \text{Gamma}(\alpha + D_x, \beta + E_x)$
This setup allows for exact Monte Carlo integration. By drawing samples of rates from these posteriors, converting them to life-table deaths ( $d_x$ ), and identifying the maximum, the authors construct the posterior distribution $P(M=x | D)$ .

2. Random-Walk Prior (RW2)

Recognizing that single-year rates are noisy and that mortality rates (on a log scale) typically increase linearly with age (Gompertz law), the authors introduce a second-order random-walk (RW2) prior on the log rates ( $\eta_x = \log m_x$ ):
$\eta_x - 2\eta_{x-1} + \eta_{x-2} \sim N(0, \lambda^{-1})$
This prior encourages smoothness by penalizing large changes in the slope of mortality from one age to the next, effectively sharing information across neighboring ages. The smoothing parameter $\lambda$ is integrated out using a Laplace approximation, allowing the data to determine the degree of smoothing.

3. Estimation and Interpolation

Point Estimates: The posterior mean, median, and mode of the interpolated modal age ( $\tilde{M}$ ) are used as point estimates. The interpolation fits a parabola through the life-table deaths of the modal interval and its neighbors to estimate a continuous age.
Uncertainty: Credible intervals are derived from the quantiles of the posterior draws.
Computation: For the RW2 prior, where no closed-form posterior exists, the authors use an Integrated Nested Laplace Approximation (INLA) approach. They find the posterior mode of the log-rates via penalized Fisher scoring and approximate the posterior as a Gaussian, integrating over the precision parameter $\lambda$ .

Key Contributions

Probabilistic Formulation: The paper shifts the estimation of the modal age from a point estimate to a full posterior distribution, providing the probability that any specific age is the mode.
Exact and Approximate Inference: It provides an exact solution for independent priors and a computationally efficient Laplace approximation for the more robust RW2 prior.
Interpolation of Uncertainty: The method propagates uncertainty through the life table transformation and uses a consistent interpolation scheme (Kannisto, 2001) applied to every posterior draw to estimate continuous modal ages.
Local Mode Analysis: The framework allows for the calculation of the posterior probability of local modes, offering insight into the "roughness" of the mortality schedule.

Results

Simulation Study

The authors compared their method (RW2 posterior mean) against Gompertz, P-spline, and Discretized Derivative Tests (DDT) estimators using simulated data from mixtures of Gompertz distributions.

Accuracy: In unimodal scenarios, the RW2 posterior mean was 1.4 to 10.6 times more accurate (in terms of mean absolute error) than the posterior mean under independent priors.
Comparison to Other Estimators:
- In unimodal distributions, RW2, P-spline, and Gompertz estimates were all highly accurate, though RW2 had slightly higher error due to interpolation bias.
- In distributions with a secondary peak (representing subpopulations), the Gompertz model was heavily biased (misspecified), while the RW2 posterior mean remained relatively accurate. The P-spline estimator was the most accurate in these multimodal cases.
Interval Coverage: The RW2 credible intervals were conservative, covering the true mode in 98% to 100% of replications across all scenarios. In contrast, bootstrap intervals for other estimators often had lower coverage (e.g., Gompertz intervals never covered the true mode in secondary peak scenarios) or were significantly wider (DDT intervals were 4–8 times wider than RW2 intervals in unimodal cases).

Application to Real Data

The method was applied to female mortality data from Denmark, France, Japan, and Sweden (1960–2020).

Trends: The RW2 estimates showed consistent increases in modal age across all countries.
Stability: The RW2 posterior mean was highly stable with respect to the starting age of the estimation ( $x_0$ ). Changing $x_0$ from 40 to 70 altered the estimate by an average of only 0.03 years, compared to 0.42 years for the Gompertz model.
2020 Decline: The framework successfully quantified the probability of a decline in modal age in 2020. For France, despite overlapping marginal intervals for 2019 and 2020, the posterior probability of a decline was 0.99. In Sweden, the probability was 0.94.

Significance and Claims

The paper claims that its primary contribution is providing a probabilistic framework that explicitly quantifies the uncertainty of the modal age, which is particularly valuable when neighboring ages are difficult to distinguish due to data noise or flat mortality peaks.

Robustness: The RW2 prior effectively borrows strength across ages, reducing bias and variance compared to independent priors without the rigid structural assumptions of the Gompertz model.
Decision Support: The posterior distribution allows researchers to make probabilistic statements (e.g., "there is a 99% probability the modal age declined") that are not possible with standard confidence intervals.
Practical Utility: The method is stable across different age ranges and provides conservative, reliable credible intervals, making it suitable for routine estimation in both small and large populations.

The authors modestly note limitations: the single smoothing parameter in the RW2 model may not perfectly adapt to complex multimodal distributions (where P-splines performed slightly better in simulations), and the intervals are conditional on the Poisson model and known exposures, excluding errors in exposure data or age reporting. However, the framework is presented as a flexible tool that can be extended to other life-table quantities.