Dennis Roberts wrote:
> 1. you could take several methods AT random (after you list out all 50) ...
This is the classical position, I think. However, in practice we
never require random sampling in order to treat people as random.
Clark argues ISTR that we should treat factors as random if sampling
them doesn't deplete the population being sampled. So people, words
and so forth should be treated as random (if you wish to generalize
beyond your sample).
>From a pragmatic view Clark's position seems defensible - in that
(whether sampling is random, pseudo-random or arbitrary) you will
underestimate the error term if you treat factors as fixed in these circumstances.
> maybe 3 ... and try ... and, if you find some differences amongst these 3
> ... then you might be able to generalize back to the entire set of 50 ...
> ie, there are differences amongst the others too
This seems an odd case. If you really wanted to generalize to only
50 methods I think you'd sample a different teaching method for each
subject. This would allow you to generalize to people and methods simultaneously.
(Sort of analagous to Clark's suggestion that you sample separate
items for each participant to avoid needing to correct the F ratio).
Thom
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at
http://jse.stat.ncsu.edu/
=================================================================
