A comparison of procedures for estimating the parent probability distribution from a given set of fractiles

On one hand, eliciting subjective probabilities (fractiles) is a well-established procedure. On the other hand, knowledge of a subjective variable’s central moments or distribution function is often assumed. However, the problem of converting elicited fractiles into the required moments or distribution function has been largely ignored. We show that this conversion problem is far from trivial, and that the most commonly used conversion procedures often produce huge errors. Alternative procedures are proposed; the ‘Tocher's curve’ and ‘linear function of fractiles’ methods are shown to be much more accurate than the commonly used procedures.