The rationale for using linear regression and the r-squared (R^2) value
as a measure of fit is that the predictions of a model may be
SYSTEMATICALLY off target in a way well accounted for by a
linear model. Implicit in the use of linear regression between
the predicted or estimated development effort (independent variable)
and actual effort (dependent variable), is the definition
of a new composite model: the model formed by machine induction
plus the linear model formed through regression). Using such a
composite model, we would generate a prediction using the composite
model and then adjust it using a previously derived best fitting line.
This assessment measure may better match our intuitions and keeps
assessment focused on MRE values exclusively.

Rather than giving the R^2 values, we could evaluate an inductive
model's predictions by simply looking at the MRE's of predictions
stemming from the composite model. This is what we do in the
table showing results with Kemerer's data (corresponds to Table
1 of article):

               Cart' Est + regress eq.  
       Cart'   (102.5 + 0.075x, 
 Act    Est    Table 2 of article)
(truncated) 
               effort   error(%) <-- relative to Act
 287   1893    244.5    14.8
  82    162    114.65   38.97
1107  11400    957.5    13.53
  86    243    120.73   38.93
 336   6600    597.5    77.67
  84    129    112.19   33.56
  23    129    112.19  383.58 
 130    243    120.73    7.34
 116   1272    197.9    70.6 
  72    129    112.19   55.82
 258    243    120.73   53.33
 230    243    120.73   47.67
 157    243    120.73   23.1
 246    243    120.73   51.1
  69    129    112.17   60.47
 
MRE                     64.71 (compared to 364% MRE)

Using the best fitting line (Table 2) constructed from all of
Kemerer's data may lead to an overly optimistic MRE that
we would realize in practice when using the compound model. Rather,
we can do a take-one-out (N-fold cross validation) strategy
-- use 14 (of the 15) predicted effort values for Kemerer's projects.
In the table below, we summarize the results in the table above,
but additionally show the linear equation of best fit generated
over each set of 14 (of the 15) and used to adject the
effort prediction for the 15th model. A different equation
for estimation is generated for each project (using the remaining
14 projects).


               Compound
       Cart'   Model             Compound Model Ests. using N-fold CV 
 Act    Est    Est        ------------------------------------------------
(truncated)                linear eq.      R     Est. effort     error (%)
               
 287   1893    244.5      99.70 + .0750x .9132     241.72        15.78 
  82    162    114.65    105.41 + .0748x .9113     117.53        42.46 
1107  11400    957.5     125.99 + .0351x .6234     525.92        52.50
  86    243    120.73    105.46 + .0748x .9116     123.64        42.28
 336   6600    597.5     106.65 + .0874x .9621     683.30       103.18
  84    129    112.19    105.07 + .0748x .9112     114.73        36.58
  23    129    112.19    110.51 + .0746x .9136     120.08       417.59
 130    243    120.73    101.71 + .0752x .9119     119.99         7.91
 116   1272    197.9     108.70 + .0749x .9156     204.01        75.87
  72    129    112.19    106.14 + .0747x .9114     115.79        60.82
 258    243    120.73     82.52 + .0775x .9325     101.36        60.82
 230    243    120.73     92.90 + .0762x .9200     111.42        51.70
 157    243    120.73     98.16 + .0756x .9122     116.54        25.77
 246    243    120.73     91.48 + .0764x .9222     110.04        55.43
  69    129    112.17     73.07 + .0790x .9085      83.27        19.13
 
MRE                                                              71.19

In sum, the high R^2 (0.83) for the CART' estimates of Table 1 (of article)
translates into a relatively low MRE (64.71% or 71.19% depending on
methodology) compared to Cart' MRE of 364%, after using the line of 
best fit to adjust Cart' estimates -- apparently, Cart's errors are 
systematic and can be adjusted to good effect.


*The Act and Cart' Estimated values are truncated for presentation
purposes only -- see Table 1 of article for values used here.