Ah.  Here some advice acquired in my undergraduate engineering courses
may be pertinent.  When fitting an empirical curve whose shape one knows
approximately, it is efficient to choose design points more closely
spaced in the vicinity of critical values like maxima, minima, and
points of inflection;  and more widely spaced in regions where the
curve in question is linear (just to check that it IS at least roughly
linear in those regions, and to estimate the linear domain with some
precision);  when the precision of one's measurement is roughly the same
throughout the domain of the predictor(s), as you say is the case.
 (When precision varies along the domain, these desiderata are somewhat
modified by wanting to take more design points where precision is poor,
and fewer where the data have high precision.)
 (For your sigmoidal curves, I'd guess that there aren't any maxima or
minima to worry about, but that there are at least two points of
inflection to be modelled.)

When the form of the curve is known exactly, usually from underlying
mathematical theory, the same general considerations apply, but now the
question is, what design points (values of X, generally, in X-Y space)
would be most helpful in (a) estimating the parameters of the formal
model and (b) distinguishing between regimes for which the model has
different parameters.
 (If you think of a linear model for [over?]-simplicity, there are two
parameters, one of which locates the response curve (a straight line)
vertically, and the other reflects the slope of the line.  Design points
near the middle of the interesting/useful domain (of values of X) are
usually most efficient in estimating location;  points near the extremes
of the domain are most efficient in estimating slope.  If the sigmoidal
model for your curves is well-established in form, quite possibly much
the same considerations apply;  if the functional form, or some but not
all of the parameters of the model, are still more or less unknown, I'd
be inclined to concentrate my points more around the rapidly-changing
regimes, while including some others as a sort of check on the
less-rapidly-changing regions.  At some level, you're really asking
"Which points can I expect to show me the best separation between models
that I want to distinguish between, which points will give me the best
available precision in estimating the parameters I'm currently looking
at, and which points will help me diagnose a badly fitting model?"  We
may agree that today (or perhaps yesterday) it might have seemed
reasonable to try to fit sigmoidal data with a cubic, or perhaps
quintic, function in X;  tomorrow it may become obvious that "Oh!  That
ought to be a logistic function!", which might lead to other choices for
apparently-optimal design points.)

Good luck!   -- DFB.

On Wed, 26 May 2004, Xinmiao wrote in part:

> ... I'm trying to design an experiment in which behavioral/neuronal
> response v.s. stimulus strength curves are to be measured. We know
> that the tuning curves are sigmoidal/linear, and my question was how I
> should spread out the sampling points along the stimulus dimension,
> say 4 or 7 or even more? We've also known that the error-variance, at
> least for neuronal responses, should be approximately same as mean,
> the square root of which shouldn't differ much between different
> independent variable (stimulus strength).  ...

<snip, the rest>

 ------------------------------------------------------------
 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/                    .
=================================================================

Reply via email to