re: "Cost estimating relations", I think.
On Wed, 13 Dec 2000 17:55:06 GMT, [EMAIL PROTECTED] wrote:
> I am in immediate assistance of convincing DOD people that a low R-
> squared is not necessarily saying that, the computed CER using the
> Linear Least Squares Method is not a good predictor of the data set.
> If anyone knows of papers, studies, theories, or any agencies that have
> documentation of using a CER with a low R-squared. The cut off for DOD
> is 0.64. We have many between 0.40 and 0.50. We feel that the
> regression equations are a very good indicator of the data set.
>
An R-squared depends on the sample (and its range) as well as the
model. So, R-squared has that as a *problem*. The raw residuals are
typically going to be invariant in cases where the R-squared is not.
A crudely modeled time series might give you a really high R-squared
while having VERY low ability to predict beyond its lag-equal-one.
Time series has that as a problem. You can't compare a times-series
R-squared to a cross-section result.
It is interesting to see a number "0.64" floated, as something
sufficient. I am only guessing at what is predicted, but does that
imply something about having 50% cost over-runs, while the CER of .40
can imply 100% cost over-runs?
- I used www.google.com to search for < regression "CER" > and hit
(among other places) the DOD site on Cost Estimating Relationships,
http://www.acq.osd.mil/dp/cpf/pgv1_0/pgv2/pgv2c5.html
The commentary includes these sensible words:
============ citation from the site
5.7 - Identifying Issues And Concerns
Questions to Consider in Analysis 1. As you perform price/cost
analysis, consider the issues and concerns identified in this section,
whenever you use regression analysis.
- Does the r2 value indicate a strong relationship between the
independent variable and the dependent variable?
The value of r2 indicates the percentage of variation in the dependent
variable that is explained by the independent variable. Obviously, you
would prefer an r2 of .96 over an r2 of .10, but there is no magic
cutoff for r2 that indicates that an equation is or is not acceptable
for estimating purposes. However, as the r2 becomes smaller, you
should consider your reliance on any prediction accordingly.
================= end of citation.
- Maybe you can refute them from their site.
--
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
=================================================================
Instructions for joining and leaving this list and remarks about
the problem of INAPPROPRIATE MESSAGES are available at
http://jse.stat.ncsu.edu/
=================================================================
