Why is diagnosis using belief networks insensitive to imprecision in probabilities?