Randomly plugging numbers. Say there are like 250,000 'undead' people. Say that on average they've been on dead for average of 50 years (crazy! I know!). This works out to an extra 12.5 million years of life. There are 127 million Japanese. If we assume average life expectancy of like 70 years, works out to 8.9 billion years of life. If we drop those 12.5 million or so years that don't exist... we're still at like 8.9 billion years.
It's true that the 'undead' wouldn't skew the mean life span much. But given the way life expectancy is calculated (http://en.wikipedia.org/wiki/Life_expectancy), it ends up being closer to median life span - so an overstated number of centenarians (people over 100) could significantly skew things.
And it's worth noting that while other developed nations are close to Japan in terms of life expectancy, they're WAY behind in terms of number of centenarians. Switzerland for example has an average life expectancy of 81.8 years, vs. Japan's 82.7. But Japan supposedly has 3.5x as many centenarians per capita (http://en.wikipedia.org/wiki/Centenarian#Centenarian_populat...).
Not always, no. They measure different things and fail in different ways.
Take the dataset [7, 8, 9], with a median and mean of 8. Adding a 100 to the set results in a median of 8.5 and a mean of 31, so the mean moves much farther. This is probably the effect you're thinking of: the mean can take extreme values into account "too much".
But I can also make the median move more. Take the dataset [0, 50, 100]. The median and mean are both 50. If I add [100, 100] to it so it becomes [0, 50, 100, 100, 100], the mean moves to only 70, but the median moves all the way up to 100! There was a "gap" in the numerical sequence that the median could jump over, but the mean couldn't.
Here's a different way of moving the median further. Take the dataset [1, 1, 1, 1, 2, 3, 4, 5, 5, 5, 5]. Bathtub-shaped data. As I add fives to the set, the mean goes 3, 3.17, 3.3. But the median goes 3, 3.5, 4! Medians move past thin spots in distributions very quickly.
Mean is sensitive to distant outliers; median is sensitive to unevenly distributed data and numerical gaps.
To come back on topic, while I don't have a reference for the age-at-death distribution, I think it's bathtub-shaped. Hence, the median might be more sensitive to extra values at the top than the mean would be.
While those are interesting counterexamples, they don't come close to modeling a realistic "age-at-death" distribution. There isn't any such distribution whose median would be significantly skewed by a tiny minority of centenarian outliers. These distributions are basically unimodal with the exception of a some degree of infant mortality, and the average (median or mean) is not located at a thin spot in the distribution (quite the opposite) [1].
In the case of life expectancy at age X, it is the mean, conditional on having already survived to age X. (In other words, life expectancy at age 20 excludes everyone who died between birth and age 20.)
Well to plug more numbers, those 250,000 'undead' people having been 'undead' for the last 20 years collecting pension checks amounting to say, $20k/year would mean:
5 million 'years' * $20k/year = $100 billion
Japan's pension fund is severely underfunded, so even this amount would help.
