Educational attainment predictive models

[Clive Downs]

Hi Everyone,

Does any one have any experience of developing predictive models (e.g. with
multiple regression, hierarchical linear models etc.) of educational
attainment, specifically for children 'in care' ( ie children receiving
social care in the responsibility of a state body).

(or perhaps is working on this?)...

and would be interested in exchanging ideas?

Thanks,

Regards

Clive.

Clive Downs
Reading, England.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Hector Maletta]

Clive,
I do not work specifically in the field, but have been involved in such kind
of predictive models, even for children's life chances including school
achievement. My two cents contribution here is of a more general nature,
casting a shadow of caution over some uses of such models.
Prediction models are a matter of probability. For instance, a logistic
regression models predicts the odds (or the probability) or some event
happening; a regression finds the expected value of an interval variable,
surrounded by a margin of error governed by a (usually normal) probability
distribution.
Probability is essentially something predicated of POPULATIONS. If you
obtain, say, a probability of 70% of children achieving high grades when
they have such and such background values, what you are actually saying is
that "out of every 1000 children with these characteristics, about 700 will
achieve high grades". By the same token, when you throw a coin, you have a
50% probability of getting heads, which means that "for every 1000 throws,
about 500 will be heads". What is worthwhile emphasizing is that there is no
valid prediction for each specific throw, or (something easily forgotten) no
valid prediction for each specific child. Each specific throw may be tails
or head, and the outcome is strictly indeterminate. Each specific child with
this background may or may not achieve high grades, and the outcome is
strictly indeterminate. You only know that 70 out of every hundred will do,
but not which ones.
This does not mean that you cannot use probability to take decisions about
specific cases. It is only that these decisions lead to sound outcomes only
in the aggregate, never in the individual case. Betting on the next coin to
be heads because the latest 10 throws were tails has no grounds in
probability. Betting on this particular child to achieve cum laude is also
wrong. But if you are allocating fellowships, using probabilities will give
you the best chance to minimize "errors" in the aggregate (defining losses
as allocating fellowships to children with low achievement, or not
allocating them to bright kids). By the same token, if you play a game where
one outcome has a 60% probability, you better bet on that outcome, because
you will win 60% of the time, but not necessarily on one particular
occasion.
It is quite probable that you are quite aware of all this, but on this mail
list I have seen many times people preoccupied about the "accuracy of
prediction" and the accuracy of the "classification table" produced by SPSS
LOGREG procedure, as if the result of such statistical procedures were to
predict individual cases accurately. Even if no individual case is predicted
accurately to happen, which may happen under specific circumstances (e.g.
rare events where a 50% probability is rarely or never predicted) the
procedure may be your best bet on the aggregate.

Hector
-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Clive Downs
Sent: 23 April 2008 11:26
To: [hidden email]
Subject: Educational attainment predictive models

Hi Everyone,

Does any one have any experience of developing predictive models (e.g. with
multiple regression, hierarchical linear models etc.) of educational
attainment, specifically for children 'in care' ( ie children receiving
social care in the responsibility of a state body).

(or perhaps is working on this?)...

and would be interested in exchanging ideas?

Thanks,

Regards

Clive.

Clive Downs
Reading, England.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Steve Peck]

Clive et al,

I would add to Hector's comments:
 that (simple) general linear modeling approaches assume homogeneity
among sampled units
 and this assumption is rarely warranted where studying the kinds of
complex social/human dynamics
 you implicate with your question.

You might find helpful some of the papers in this recent Journal of
Social Issues ("Unexpected educational pathways")
   in which a variety of approaches to modeling "educational attainment"
are described:
http://www.blackwell-synergy.com/toc/josi/64/1

Steve

Clive Downs wrote:

> Hi Everyone,
>
> Does any one have any experience of developing predictive models (e.g. with
> multiple regression, hierarchical linear models etc.) of educational
> attainment, specifically for children 'in care' ( ie children receiving
> social care in the responsibility of a state body).
>
> (or perhaps is working on this?)...
>
> and would be interested in exchanging ideas?
>
> Thanks,
>
> Regards
>
> Clive.
>
> Clive Downs
> Reading, England.
>
> =====================
> To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD
>
>
>
>

--
Stephen C. Peck
Research Investigator
Achievement Research Lab
Research Center for Group Dynamics
Institute for Social Research
University of Michigan
426 Thompson Street, # 5136
Ann Arbor, MI  48106-1248
[phone]; fax [phone]
http://www.rcgd.isr.umich.edu/garp/
[hidden email]

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD