Josh from PMStudent.com has a post about estimating. There is a discussion of using 3 point estimates.

Some time ago, I came to the conclusion that making 3 point estimates is a bad idea. Instead use confidence intervals for risk ranges and drive the variance from those bands with a Monte Carlo Simulator.

This is a critical topic that goes to the credibility of any estimate no matter the domain.

The starting point is to understand the concept of anchoring and adjustments to the estimate and their impact on false optimism (or pessimism). The paper to start with is "Anchoring and Adjustment in Software Estimation," Aranda and Easterbrook.

When you read this, you may rethink the approach that asks the estimator the Most Likely estimate and Optimistic and Pessimistic estimates.

The original work in this area is from Tversky and Kahneman and their "Prospect Theory." A good place to look for some details (besides their work) is* Against The Gods*, Peter Bernstein, in Chapter 16, The Failure of Invariance."

It's time to move on from the classic "simple" approach found in places like PMBOK(r) to the processes used in other places that depend on probabilistic and statistical analysis - engineering, finance, medicine, physics.