Page 161 -

132 PART TWO MANAGING SOFTWARE PROJECTS

Total estimated effort for the CAD software range from a low of 46 person-months
(derived using a process-based estimation approach) to a high of 58 person-months
Do not expect that all (derived using an FP estimation approach). The average estimate (using all three
estimates will agree approaches) is 53 person-months. The maximum variation from the average esti-
within a percent or mate is approximately 13 percent.
two. If the estimates
are within a 20 What happens when agreement between estimates is poor? The answer to this
percent band, they can question requires a re-evaluation of information used to make the estimates. Widely
be reconciled into a divergent estimates can often be traced to one of two causes:
single value.
1. The scope of the project is not adequately understood or has been misinter-
preted by the planner.
2. Productivity data used for problem-based estimation techniques is inappro-
priate for the application, obsolete (in that it no longer accurately reflects the
software engineering organization), or has been misapplied.

The planner must determine the cause of divergence and then reconcile the estimates.

5.7 EMPIRICAL ESTIMATION MODELS
An estimation model for computer software uses empirically derived formulas to pre-
dict effort as a function of LOC or FP. Values for LOC or FP are estimated using the
approach described in Sections 5.6.2 and 5.6.3. But instead of using the tables described
An estimation model in those sections, the resultant values for LOC or FP are plugged into the estimation
reflects the population model.
of projects from which The empirical data that support most estimation models are derived from a lim-
it has been derived.
Therefore, the model is ited sample of projects. For this reason, no estimation model is appropriate for all
domain sensitive. classes of software and in all development environments. Therefore, the results
obtained from such models must be used judiciously. 13

5.7.1 The Structure of Estimation Models
A typical estimation model is derived using regression analysis on data collected from
past software projects. The overall structure of such models takes the form [MAT94]

E = A + B x (ev) C (5-2)
where A, B, and C are empirically derived constants, E is effort in person-months, and
ev is the estimation variable (either LOC or FP). In addition to the relationship noted
in Equation (5-2), the majority of estimation models have some form of project adjust-

13 In general, an estimation model should be calibrated for local conditions. The model should be
run using the results of completed projects. Data predicted by the model should be compared to
actual results and the efficacy of the model (for local conditions) should be assessed. If agreement
is not good, model coefficients and exponents must be recomputed using local data.