Anchoring and Adjustment in Software Estimation

http://www.cs.toronto.edu/~jaranda/pubs/AnchoringAdjustment.pdf

This paper presented Jorge's results about anchoring and adjustment when estimating software project time consumption. Looking back at my notes on this paper, everything seems to be presented fairly well, all the numerical results seem sound. Couple of quick things:

The Anchoring and Adjustment phenomenon is clearly observable when the problem can be expressed as a number in a range. Can it be applied to other classes of estimations?

Why is the description of COCOMO so verbose? It don't see how it contributes to the paper, other than to demonstrate that current software estimation techniques don't work.

What is a null hypothesis?

Now, I want to take this opportunity to discuss an estimation technique that a former colleague of mine once described to me:

Estimating the time required to complete a software project is inherently a random activity, and it is generally accepted that estimating smaller, individual tasks within a project is more accurate than trying to estimate the project as a whole.

Lets divide our hypothetical project P into n subprojects, P0 ... Pn-1. For each of these, instead of guessing a completion time, we provide a low ball and high ball estimate (ie. P0 will definitely take more than a week, but less than 4).

We take the sum of all the low ball estimates, and the sum of all the high ball estimates, and we get two figures, one for the earliest possible completion time for P and one for the latest. While this may be sufficient for some, one further refinement is to apply a Gaussian distribution between these two figures, that way you could say with some degree of reasoning, that project P has an 80% chance of finishing in X weeks.

Ferenc, I apologize if I've gotten any of this wrong.