If I understand you correctly, the uncontrollable variable you are
concerned about is time: you want to control for the possibility that
your dependent variable(s?) may change detectably during the period of
the experiment. The method you propose -- taking a number of control
observations throughout the duration of the experiment -- will help you
discover whether there exist any trend(s) in time. (Of course, if the
variable fluctuates with time, at a frequency with period close to or
less than the interval between observations, you can't detect that; but
I gather this is viewed as highly unlikely.)
Another approach is to randomize thoroughly the order of treatments
(cells in the factorial design). Then, even if there is a time trend,
it will not have a systematic effect on the effects of the design
variables, but will contribute to the underlying noise. This will make
the experiment less sensitive to real design effects, but the decrease
in sensitivity will be small if the variation with time is small. And
it isn't really making the experiment any less sensitive than it would
otherwise be, it's rather making it less likely that any such temporal
effects contaminate your measures of the design effects.
If you also take control observations at regular intervals, as you
propose, you will be able to filter out some of that noise. ("Some" and
not "all" because wile the control observations will permit you to fit a
systematic model -- e.g., linear, quadratic, cubic components in time --
the model itself may be in error or incomplete.)
On 18 Aug 2003, Dave Jeffries wrote:
> I'm trying to set up the analysis for some water research, and I'm
> hoping to elicit some advice. So far the bulk is a 2 or 3 factorial
> design (to be decided later) with no interaction expected. However,
> there is one other variable that I can't control - as I said, it's
> water research, and each sample will be collected discretely, albeit
> from the same source, but on different days. There is NO WAY to take
> multiple samples at the same time and store some for later analysis,
> as each treatment/test will take over 48 hours to run, and the sample
> would change anyway. While I don't expect any significant changes in
> the samples collected before each test, I need to confirm it.
>
> So far, the best method I could come up with is to have a control
> treatment, one that is run every 3 or 4 treatments throughout the
> entire experiment, and then do a run-test on the response from it.
> Then, if the responses are close enough to be accepted as statis.
> insignificant, just assume that there was no change through the other
> tests as well.
>
> However, what I'd like to do is find a more robust method that works
> the last bit into the factorial design, perhaps through replication,
> but without letting the number of tests required get to large. If
> there's anyone out there who might be able to come up with some idea
> of how to check this uncontrollable variable, I'd greatly appreciate
> your thoughts on the matter.
-----------------------------------------------------------------------
Donald F. Burrill [EMAIL PROTECTED]
56 Sebbins Pond Drive, Bedford, NH 03110 (603) 626-0816
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
. http://jse.stat.ncsu.edu/ .
=================================================================
