[R] Fractional Factorial Design

[paulandpen]

Excellent points Steve,

I am ever expanding my understanding in the area and power is an interesting 
one.  I do a lot of choice modelling myself, and I am confounded (grin) by 
the optimal way to develop designs with conditional levels (deliberate 
confounds) etc.

Thanks for that.

Regards P


----- Original Message ----- 
From: "S Ellison" <S.Ellison at lgc.co.uk>
To: "Caio Azevedo" <cnaberdl at gmail.com>; "paulandpen" 
<paulandpen at optusnet.com.au>; "R - discussion list" 
<r-help at stat.math.ethz.ch>
Sent: Tuesday, April 29, 2008 1:48 AM
Subject: Re: [R] Fractional Factorial Design


> Paul;
>
> You asked
>>using  .......
>> optFederov(~.,dat,6)
>>... does the job with good efficiency.
>>
>>I would be interested to know what your objection to this is S
>
> I have no issue with AlgDesign in principle, but the question was
> specifically about _fractional_ factorials, so I answered that.
>
> As to which is best - well, first pick your definition of 'best'. Both
> can improve drastically on full factorials. For me, he advantage of a
> fractional factorial is that it retains balance and, more importantly
> from a design perspective, I get to choose which effects are confounded
> and can arrange matters so that some effects are guaranteed
> unconfounded. The deterministic nature of the selection also makes it a
> bit easier to build power considerations into the process if you're so
> minded. The price of that is that the number of observations is
> typically larger than the smallest algorithmic design that might do a
> broadly similar job, though never as large as a full factorial.
> As I see it, the main advantage of algorithmic design is that you get
> to pick the size of the experiment. A second plus is that you can handle
> arbitrarily constrained designs much more easily, which is a feature
> I've sometimes found important. The disadvantage is that you may incur
> bias in some of the effect estimates, and because the selection process
> to fit an arbitrary experiment size typically involves some random
> selection from a candidate list, you don't necessarily get to choose
> which effects are biased. I guess you will also have a more interesting
> job deciding how many observations you need for a given power, if that's
> relevant.
>
>
> Steve E.
>
>
> *******************************************************************
> This email and any attachments are confidential. Any u...{{dropped:8}}