There's this idea going around that we should spend more to save an ``identified'' life than a ``statistical'' life, on efficiency grounds. First let me summarize the argument that Robert Frank, Mark Kleiman and dozens of others have bought into; then let me explain why it's wrong. Their argument: Suppose in a world of a million people, each places a $1 value on avoiding a one in a million chance of death. Then efficiency requires us to spend up to a million dollars (but no more) preventing one statistical death. But now suppose one of these million is trapped in a mine, with a 1/2 chance of dying. And suppose he places a $4 million value on avoiding that chance of dying. Then efficiency requires us to spend up to $4 million to save him. Okay, now here's what's wrong: First, the argument, as I've paraphrased it (and, I claim, as presented by every single one of its proponents, at least as far as I can tell), plays fast and loose with the distinction between willingness to pay and willingness to accept. Does placing a $4 million value on avoiding a 1/2 chance of death mean that a) this is what the miner will pay to avoid that chance of death or b) this is what we'd have to pay him to accept it? It's important to be consistent about that, though I haven't found anyone in the argument-making camp who is. Let's consider the consequences of being consistent in either direction. First, let's use willingness to accept. What would I have to pay you to accept, say, a 100% chance of death? Plausibly, all of the world's wealth. But nobody believes we should spend all of the world's wealth to rescue a single miner. So willingness to accept can't be the right criterion here. That leaves willingness to pay (sometimes called equivalent variation. And here is my main claim: Using a willingness-to-pay criterion, the numbers in the above example are impossible. I'll illustrate with a population of four people (though the argument works equally well with a million). Let U(p,I) be the utility associated with death probability p and income I. Let x be the equivalent variation for a 1/4 chance of death and y the equivalent variation for a 1/2 chance of death. Then, since a 1/4 chance of death is the same as a 1/2 chance of a 1/2 chance of death, we have U(0,I-x) = U(1/4,I) = (1/2)U(1/2,I)   +   (1/2)U(0,I) = (1/2)U(0,I-y)   +   (1/2)U(0,I)   so if we assume that the function U(0,-) is convex (i.e. that people are not risk-preferring), we have 2x>y, so 4x>2y. But 4x is what we should pay to save a statistical life and 2y is what we should pay to save an identified life. In other words, the numbers in the example at the top are impossible. Statistical lives are worth more than identified lives, not less. If you want to get out of this by positing risk preference (a pretty desperate ploy to begin with), I respond that you don't have to wait for a mine accident to satisfy risk preference. People will satisfy it all on their own by making fair bets with each other until they're carrying all the risk they want, at which point utility in additional income has to be convex. Or, of course, you could argue on something other than efficiency grounds---but the statistical/identified life crowd seems loath to do that. On a separate note, Robert Frank has made the curious argument that we should value idenitified lives more highly because that's what people seem to want to do. Of course if your position is that people should always just follow their instincts wherever they lead, then there's no need for any kind of moral inquiry in the first place.