[From Bill Powers (971130.1041 MST)]
Bruce Abbott (971130.1210 EST) --
I had to laugh at this example. Administrators (at least the ones I've
dealt with) would not ask what is wrong with the students, they would ask
what is wrong with the teacher. Of course, this also begs the question.
I had already given an example in which the conclusion was about the
teacher, so I thought I'd throw in one on the other side. Note that if the
teacher were blamed, that would also be incorrect. The problem is with the
administrator who gives the football players a free pass. The general
principle is that administrators always use statistics to blame someone
else for effects caused by the administrators. See Dilbert.
If you assume that group characteristics apply to every individual in the
group, but are simply masked by random influences, you gain no
understanding of the individuals. On the other hand if you study
individuals, most of the variance in the group statistics will be
explained, although the correct explanation is likely to be different for
each individual.
Ah, now you're talkin' my lingo. Murray Sidman, a protege of B. F. Skinner,
said exactly this in _Tactics_ (1960).
Seems to me you have a rather hard conflict to resolve, if you believe this
and also teach conventional statistics courses.
Our dispute is not about the use of group statistics for legitimate
purposes, those related to performance of the group. It is about attempts
to apply the results of group studies to individuals rather than studying
the individuals one at a time.
From my perspective, our dispute is not about that either. My position was
not that one can with certainty apply the results of group studies to
individuals, but that group studies may identify variables whose effects
(under the conditions studied) apply with sufficient regularity to
individuals that differences appear in group means.
I tried to discuss that a while back and you didn't respond. If it is true
that the group effect arises because the treatment affects all the members
of the group similarly, then your conclusion is correct. But even if the
treatment affects all the members of the group differently, it is still
quite possible that there will be a group effect. Simply by doing the group
study, you can't determine which is the case: whether the individuals are
affected differently or in the same way. The only way to settle that
question is to investigate the individuals one at a time. If you do
investigate the individuals one at a time, the group study is superfluous.
But if you don't do the individual studies, the group study tells you
nothing about the individuals.
The logic is clear: group studies tell us nothing about individuals.
Variables so identified
warrant further study on a case-by-case basis. This is a weaker position
that is supported by research experience. I don't know that you would
disagree with this, but I did feel that in the rush to condemn group-based
designs, such proper uses were not being mentioned, and so I wanted to point
out (rather innocuously, I thought) that useful information can be gained
from them under proper conditions.
I am not condemning group-based designs. They are fine for studying group
behavior. Anything you think you learn about individuals from the group
study, however, you can learn faster from the individual studies, which you
will have to do in any case.
The only condition under which I could agree with you would be when the
group study is used as a quick way of eliminating treatments that have NO
effect. Unfortunately, all treatments have some effect. If the effect
doesn't reach significance, you just use a larger sample; eventually, if
you have the resources to test hundreds of thousands or millions of people,
you will get a significant effect no matter what you manipulate. And then,
to show that it is valid for predicting individual behavior, you will have
to start testing individuals. So you might as well just start with the
individuals, if individual behavior is what you're interested in.
Remember the invalid syllogism:
A manipulation that affects all individuals in a group the same way will
affect the group behavior in the same way as the individuals. (fact)
A manipulation affects group behavior. (observation)
Therefore the individual behaviors are affected the same way as the group
behavior. (conclusion)
The correct conclusion is that from the group behavior you can infer
nothing about the individual behaviors.
Succinctly:
A --> B.
B;
Therefore A.
... is wrong.
Best,
Bill P.
···
There _are_ cases in which group-based functions and individual functions
look much alike, by the way. Some functions are sufficiently general that
they do not vary radically from individual to individual; one task within
any science is to identify such general relationships. For example, we
might expect certain parameters such as gain to vary across individuals
controlling cursor position in a tracking study, but would not expect
radical differences from individual to individual in the nature of the
function. For example, I have seen group data in a study of body-weight
control in rats (not our study) that looks like just what one would expect
of an individual control system's behavior in response to disturbance.
Having said this, I should probably remind everyone that I am a strong
proponent of the single-subject approach, and am well aware of the
shortcomings of group-based designs when it comes to applying their results
to individuals. Nor am I claiming that everyone doing psychological
research using these methods is as aware of these limitations as I.
Regards,
Bruce