Social factors and lifespan inequality: a four-way factorial analysis of U.S. lifespan

This paper pioneers a four-way factorial analysis of U.S. lifespan data to demonstrate that while education and its interactions are the primary drivers of between-group lifespan inequality, such heterogeneity accounts for only a small fraction (7–10%) of total lifespan variance, with the majority arising from individual stochasticity.

The Gist

The Big Question: Why Do People Die at Different Ages?

Imagine you are looking at a giant jar of marbles. Some marbles are red, some blue, some green. Some are big, some are small. Some are smooth, some are rough. If you roll them down a ramp, they will all stop at different spots.

The big question this paper asks is: Why do they stop at different spots?

Is it because of their color (social factors like race or education)? Or is it just because of luck (random chance)?

For a long time, scientists looked at these factors one by one. They asked, "Does being male vs. female matter?" or "Does being rich vs. poor matter?" But in the real world, people are a mix of all these things at once. This paper is the first to look at four factors at the same time (Sex, Marital Status, Education, and Race) to see how much they actually explain the differences in how long people live.

The Experiment: The "Life Table" Recipe

The author used a massive dataset from the United States (2015–2019) that broke down the population into 54 different groups. Think of these groups as 54 different "flavors" of life:

• A married, college-educated, non-Hispanic White woman.
• A single, high-school-educated, Hispanic man.
• And so on...

For each of these 54 groups, the researchers calculated the average life expectancy and the "variance" (how much the ages of death varied within that group).

The Magic Math: Splitting the "Inequality Cake"

The core of the paper is a statistical trick called Variance Partitioning. Imagine the total difference in how long people live is a giant cake. The researchers wanted to slice this cake to see who gets the biggest piece.

They cut the cake into two main layers:

1. The "Group" Layer (Between-Group): This is the part of the cake caused by the differences between the 54 groups. (e.g., Do college grads live longer than non-grads?)
2. The "Luck" Layer (Within-Group): This is the part of the cake caused by pure chance inside each group. (e.g., Two people with the exact same background, same job, same health habits; one lives to 90, the other to 75. Why? Just bad luck or good luck.)

The Shocking Result:
Even when you account for four major social factors and all their complicated interactions, the "Group" layer is tiny.

• The "Luck" Layer accounts for 90% to 93% of the cake.
• The "Social Factors" Layer only accounts for 7% to 10% of the cake.

The Analogy: Imagine you are playing a video game. You can choose your character's class (Warrior, Mage, Rogue) and your starting gear (Gold, Silver, Bronze). These choices matter! But once the game starts, the biggest reason you win or lose isn't your class or your gear; it's the random number generator (RNG) of the game. You can be the best-equipped Mage, but if the dice roll against you, you lose. In human life, stochasticity (random chance) is the dice roll.

Who Gets the Biggest Slice of the "Social" Cake?

Even though social factors only explain a small slice of the total cake, some factors are bigger than others. The researchers used a method called Sobol' indices to measure importance.

• Education is the King: Education (and how it mixes with other factors) is the single biggest driver of social inequality in lifespan. It explains about half of the social difference.
• Marital Status and Sex: These are the next biggest players.
• Race: Surprisingly, in this specific model, race explained a smaller portion of the variance than education or marital status.
• The "Interaction" Effect: The paper also looked at how factors mix (e.g., Does being a woman and having a college degree change things differently than just being a woman or just having a degree?). These "interaction" slices of the cake were incredibly tiny—almost invisible.

The Two Ways to Slice the Cake (Mixing Distributions)

The author had to make a choice on how to weigh the groups. This is like deciding how to count the marbles in the jar.

1. The "Flat" Mix (The Experiment): Imagine you take exactly 1,000 people from every single one of the 54 groups, even if some groups are tiny in real life. This treats every "flavor" of life as equally important to understand the mechanics of the factors.
2. The "Population-Weighted" Mix (The Reality): Imagine you take a random sample of 1,000 people from the actual US population. In this scenario, the "Married, White, College-Educated" group is huge, so they dominate the results. The "Single, Hispanic, High-School-Educated" group is tiny, so they barely register.

The Result: Both methods gave the same big picture: Chance is the boss. But the "Population-Weighted" method showed slightly less inequality because the huge, dominant groups (who tend to live longer) smoothed out the extremes.

Why Does This Matter?

1. We Can't Eliminate Luck:
The paper argues that we should stop trying to "fix" lifespan inequality by only focusing on social factors. Even if we eliminated all poverty, racism, and lack of education tomorrow, 90% of the difference in how long people live would still exist because of individual randomness. We have to accept that "luck" is a fundamental part of life.

2. Education is the Lever:
If we do want to reduce the social gap, the paper suggests focusing on education. It has the biggest impact on the "social" slice of the cake.

3. The "Perfect Knowledge" Myth:
The author asks: "What if we knew everything about everyone? What if we tracked their DNA, their diet, their stress levels, and their bank accounts?"
He cites another study that tried to predict death using 32 different variables. Even with all that data, they could only explain 1.3% of the difference in lifespan. The rest was still just luck. Knowing more about people won't eliminate the randomness of life.

The Takeaway

Think of life as a long, winding road.

• Social factors (Education, Race, Sex) are like the type of car you are driving. A Ferrari (high education) generally goes faster and further than a beat-up sedan (low education).
• Individual Stochasticity is the weather, the potholes, and the random traffic jams.

This paper tells us that while having a Ferrari helps, the weather and the potholes are the main reason why two cars starting at the same time end up at different destinations. We can't control the weather, but we can try to make sure everyone has a decent car (education). However, we must accept that no matter how good the car is, the road is still full of surprises.

Technical Summary

1. Problem Statement

Lifespan inequality arises from two distinct sources:

1. Heterogeneity: Systematic differences between groups defined by observable characteristics (e.g., sex, race, education) or latent traits (e.g., frailty).
2. Individual Stochasticity: Random variation ("luck") inherent in survival processes, where individuals subject to identical risks experience different lifespans due to the probabilistic nature of mortality.

Previous research has largely focused on partitioning variance in lifespan by analyzing one factor at a time (e.g., education alone or race alone). This approach fails to capture the contributions of interactions between factors (e.g., how the effect of education on lifespan differs by race). Furthermore, existing multi-factor studies are rare due to data scarcity. The paper addresses the need to quantify the contributions of multiple social factors and their complex interactions to lifespan inequality simultaneously.

2. Methodology

The study employs a four-way factorial design to partition the variance in remaining longevity.

• Data Source: The analysis utilizes life tables from Bergeron-Boucher et al. (2024), covering the U.S. population from 2015–2019. The data includes all 54 combinations of four factors:
  • A: Sex (Female, Male)
  • B: Marital Status (Married, Previously Married, Never Married)
  • C: Education (High school or less, Some college, College degree)
  • D: Race (Black, Hispanic White, Non-Hispanic White)
• Variance Partitioning Framework:
  • The total variance ($V_{total}$) is decomposed into Within-group variance ($V_{within}$, due to stochasticity) and Between-group variance ($V_{between}$, due to heterogeneity).
  • $V_{between}$ is further decomposed into main effects (4 factors), two-way interactions (6 pairs), three-way interactions (4 triplets), and a four-way interaction.
  • Formulas for marginal means and mixing distributions are derived using Kronecker products and vectorization operators to handle the multi-dimensional arrays.
• Mixing Distributions: To calculate weighted variances, two mixing distributions ($\Pi$) are tested:
  1. Flat Mixing: Assigns equal weight ($1/54$) to every factor combination. This isolates the pure effects of the factors, treating them as an experimental design.
  2. Population-Weighted (Rank-One) Mixing: Weights combinations based on their relative abundance in the actual U.S. population. This reflects the real-world composition of the population.
• Sensitivity Analysis: The paper applies Sobol' global sensitivity indices to quantify the total contribution of each factor (including all its interactions) to the total variance.
• Age Conditioning: Calculations are performed for starting ages ranging from 30 to 85 years to observe how inequality evolves with age.

3. Key Contributions

• First Multi-Factor Variance Partitioning: This is the first study to apply variance partitioning to a four-factor factorial design in human demography, moving beyond single-factor analyses.
• Quantification of Interactions: It explicitly calculates the variance contributions of two-way, three-way, and four-way interactions, demonstrating that while they exist, their magnitude is negligible compared to main effects.
• Methodological Framework: The paper provides explicit mathematical formulas for decomposing variance in $K$-factor designs, offering a generalizable tool for demographic and biological studies.
• Distinction of Analytical Approaches: It clarifies the difference between "experimental" (flat mixing) and "survey" (population-weighted) approaches to weighting data, arguing that the choice depends on whether the goal is to understand factor effects or population composition.

4. Key Results

• Dominance of Stochasticity: Even when accounting for four major social factors and all their interactions, individual stochasticity accounts for approximately 90–93% of the total variance in remaining longevity.
  • Flat Mixing: Between-group heterogeneity explains 10.6% of the variance.
  • Population-Weighted Mixing: Between-group heterogeneity explains 7.9% of the variance.
• Role of Main Effects vs. Interactions:
  • Education is the single most significant factor, contributing roughly 46% of the between-group variance (flat mix) and 40% (weighted mix).
  • Interactions are negligible: Two-way interactions contribute an order of magnitude less than main effects. Three-way and four-way interactions are orders of magnitude smaller still.
  • Race contributes the least among the four factors (approx. 3.7% of between-group variance in flat mix).
• Sobol' Indices: When pooling main effects and all interactions involving a specific factor, Education remains the dominant predictor of lifespan inequality, followed by Marital Status, Sex, and Race.
• Age Trends: The ratio of between-group to total variance ($K$) remains relatively stable or declines slightly with age. The contributions of interactions converge at older ages (above 60), though the magnitude remains small.

5. Significance and Implications

• Reframing Inequality: The study confirms that while social factors create significant differences in average life expectancy (e.g., an 18-year gap between the best and worst groups), they explain very little of the variance (inequality) in individual lifespans. The vast majority of lifespan variation is due to unavoidable chance.
• Policy and Interpretation: The results suggest that policy interventions targeting social determinants (like education) can reduce the gap between groups but will have a limited impact on the overall inequality (variance) within the population because stochasticity is unavoidable.
• Future Research Directions:
  • The paper argues for evaluating factors based on their contribution to between-group variance rather than total variance, as this better isolates the effect of the factor itself.
  • It highlights that even with perfect knowledge of all individual characteristics (e.g., via machine learning on census data), the variance explained remains low (referencing Breen & Seltzer, 2024, who found only 1.3% variance explained by 32 variables), reinforcing the inevitability of stochasticity.
  • The methodology opens the door for analyzing other demographic outcomes (healthy longevity, lifetime reproduction) using multi-factor designs.

In conclusion, Caswell demonstrates that while social stratification creates distinct life expectancy profiles, the randomness of individual survival is the primary driver of lifespan inequality. The study provides a rigorous mathematical framework to disentangle these effects, showing that interaction terms are statistically present but practically insignificant compared to main effects like education.