"Akhenaten: Egypt's False Prophet", Nicholas Reeves, Thames and Hudson (2001).

An up-to-date biography with more recent studies. Quite well written and as usual for T&H books on Egypt, extremely well manufactured (even the paperbacks) with copious photographs and illustrations.

At the same time when you read the older and newer books about Akhenaten, you can be amazed at how little an extra half century of investigations has furthered our knowledge and how you really, as always, can't believe definitive conclusions. At this point the most recent opinion was the 1998 finding that the bones of KV55 were of a 35 year old Akhenaten. Nicholas does say, though, in what may have been a sardonic comment that if the medical experts don't change their minds again, then we can consider the matter closed. Of course, they did, and we can't.

But you wonder. Just how how much stock we can put in judgments based on methods that go from predicting an age of 18 to 20 years in 1966 to an age of 35 years in 1998 and back to 20 - 25 two years later?

One problem here is the uncertainty in the such predictions are often large , but not properly reported. For instance, one forensic anthropologist examined a skull (of a dead person, of course, but not Akhenaten) and came to the conclusion that the individual was between 26 and 35 years old. That really pins it down doesn't it?

But a more fundamental problem is partly due to the ubiquity of computers. Today it is way, way, way too easy to crunch data and come up with a set of equations that look like you can predict things. So you can take physical characteristics of a "data set" of a bunch of bones of people of certain age, sex, height, etc., and "regress" the data on your handy-dandy computer. Plug in information from some bones found in the ground, and hey presto! The computer spits out the age, sex, height, etc. of the person when he or she died. Not bad, huh?

What is not appreciated (at least not enough) is that a good data fit for an equation doesn't mean it predicts. This, of course, is the old and well known statement that correlation does not mean cause. Numerically what this means is you may have regression equations that produce a low root mean squared error and high correlation coefficient, but they fall flat when you plug in really new data.

And what is really not known is that failure of a mathematical regression to "generalize to new data" as it's called, isn't rare. Instead it's quite common, maybe even the norm. So unless it can be shown that a "modern, "up-to-date", and "state-of-the-art" method does meaningfully predict new data - not just fit the old - then you shouldn't accept the predictions.

And when you get nice equation that does handle new data, you'll be surprised at how huge the uncertainties can be. That's pretty heady stuff, but properly developed and validated equations and precisely stated error ranges can make fancy pants predictions look pretty silly.