Air safety update

Readers of Chapter 5 of Numbers Rule Your World already know that the U.S. has a strong air safety record. The MIT scholar, Arnold Barnett, has concluded that we have "nowhere to run for safety". Dying in a plane crash is in essence a random event.

Here are several recent articles updating the situation:

USA Today reports that the safety record has improved further in the last few years (link)

Wall Street Journal says the regulators are increasingly focused on "surface threats" rather than in-the-air incidents (link).

A researcher claims to have found a relationship between profitability and safety risk (link): I have yet to examine this study; based on this write-up, I don't think it's ready for prime time. He proposes that safety risk goes up whether an airline is above or below profitability target. It's unclear whether he looks into whether there is a third factor that affects both safety risk and risk of making a profitability target. Also, if Barnett is right that safety risk is essentially random, does it make sense to find factors that explain a random factor?

Comments

[Floormaster Squeeze]

I am not sure I understand your comment about the Madsen work. It is possible his work has problems (I have only looked at the article) but I am not sure how the Barnett work is related. Did Barnett specifically look at the specific statistical relation Madsen investigated and see no relationship? Did Barnett even look at profitability as a possible predictor (I do not know so he may have)?

It seems to me that Madsen looked at relationship that many investigators might miss (a non-linear relationship).

My point is that to investigate something and conclude that the effect is "random" does not preclude further (better?) specified hypotheses from being "non-random".

[Kaiser]

Floormaster: Good point. I'm speaking from the perspective of someone who is convinced by Barnett's studies (which has been ongoing for over 30 years). Given that "model", any discovery of an "effect" is likely to be spurious. If one doesn't believe that model, then it is possible to find effects. Maybe a Bayesian model with a noninformative prior can be used here. I'm also worried by trying to look for complex relationships when there is so little signal (fatalities/accidents).