StRD: How to Use

##
How to Use Statistical Reference Datasets

The Statistical Reference Dataset pages are intended to complement the
testing of statistical software by providing datasets and corresponding
"certified values" of statistical parameters which are commonly estimated
by software packages. To obtain statistical reference datasets for this
purpose, read this entire list of instructions and then proceed.
1) Click on *Dataset Archives*

2) Click on the statistical method of interest:

*
*- balanced one-way ANOVA
- linear regression
- nonlinear regression
* univariate methods
*

Under each statistical method, you will see several datasets for testing
statistical software. These datasets are roughly graded according to
difficulty, so for a quick test, one may try the most difficult problems.
Alternatively, since a fairly wide variety of examples are provided, one
might choose examples which are closest to a particular application.

3) Click on a dataset of interest to view the following:

*
*- dataset information
- statistical model
- methodology
- graphical display of the data (where appropriate)
- certified results
* data file
*

The data files are in ASCII format with header information in lines
1-60, with the data beginning on line 61. For convenience, the header information and certified results are
included in the file so that only one file needs to be downloaded for
each dataset of interest.

4) If the dataset is of interest, then:

- download the data file to your computer
- using the model provided, analyze the data with your software
- compare your results with the "certified values"

The "certified values" represent the "best available" solution. More
discussion of certified values can be found on pages dealing with the
various statistical methods.
5) Repeat this procedure for several datasets that are in your domain of
interest.

The question of how close one should expect results using a particular
software package to be to the certified values cannot (at least at present)
be answered in general. Testing software with these datasets can provide
useful information for evaluation, but the decision as to whether there
is a potential problem with a particular computer algorithm requires
judgement which includes, but should not be limited to, comparsion
with certified values.