In talk.politics.drugs Radford Neal <[EMAIL PROTECTED]> wrote:
> My guess is that both sides in this "debate" are "arguing" for
> pre-conceived notions.  However, Mr. Sandle's contributions do not
> appear to me to be "bizarre", nor any more illogical than those of his
> opponents.  For example, his hypothesis that having a "crack baby (or
> 4 year old)" causes socioeconomic stresses, so that controlling for
> socioeconomic status is not a correct procedure, appears to me to be
> rather dubious (in that the effect seems unlikely to be large enough
> to be a problem), but it certainly isn't illogical.  In an actual
> debate, his oppenents would offer counter-arguments, or even decide
> to control for socioeconomic status BEFORE the child's birth.


repeating szasz:

 a partial rank correlation is
>> simply a more complicated version of the same technique, except that
>> simply you're proposing to add in more variables to the model (to be
>> "partialled out", hence the term "partial" correlation),

So now the study is exploring any causes in the following:

impaired infant <-> pregnancy cocaine use

impaired infant <-> social factors after birth

impaired infant <-> social factors before birth

pregnancy cocaine use <-> social factors before birth

pregnancy cocaine use <-> social factors after birth

social factors before birth <-> social factors after birth


A partial correlation is not going to tell the direction of cause. But  it 
may tell that there is some likelihood of a cause.

When you work with SPSS and enter your figures for the 4 series, you 
will get back 6 results for correlations, one result for the extent 
of the relationship between each pair of series. No cause so far.

Then if you get say SPSS to do partial correlations you get each of
those 6�pairs having extra figures depending on how many of the
others you partial for. 

There are 5 partials for each with single variable partialling, 
naking 30. 

There are 5 partials for each with dual variable partialling, 
bringing total to 60. 

There are 4 partials for each with 3 variable partialling. 
Now up to 84.

There are 2 partials for each with 4 variable partialling,


So it seems like SPSS will produce 92 figures, a few with asterisks 
indicating they have significance.

So then we take the asterisked figures and compare them with the 
appropriate non-partialled 6. They may be the same as or different 
from the non-partialled corresponding figure. They may be tending to 
zero.

By comparing the non-partial and partial results it is possible to get
some idea of possible cause, or to rule it out, as any likelihood.

Now I am waiting to see the study, and what was done to show no cause 
from pregnancy cocaine -> impaired infant.

Taking a further look at

>    Linkname: PA 765: Partial Correlation
>         URL: http://www2.chass.ncsu.edu/garson/pa765/partialr.htm
>        size: 290 lines

> gives a good look at the processes.

> *********
>                             Partial Correlation

> Overview

>    Partial correlation is the correlation of two variables while
>    controlling for a third or more other variables. The technique is
>    commonly used in "causal" modeling of small models (3 - 5 variables).
>    For instance, r[12.34] is the correlation of variables 1 and 2,
>    controlling for variables 3 and 4. The researcher compares the
>    controlled correlation (ex., r[12.34]) with the original correlation
>    (ex., r[12] and if there is no difference, the inference is that the
>    control variables have no effect.


So in the case under consideration that would mean, according to the
claim of no effect of pregnancy cocaine on impaired infant that the
correlation figure for

impaired infant  <-> social factors after birth 

is the same whether or not pregnancy cocaine use is partialled out.

 If the partial correlation
>    approaches 0, the inference is that the original correlation is
>    spurious -- there is no direct causal link between the two original
>    variables because the control variables are either (1) common
>    anteceding causes, or (2) intervening variables.

And that might show that any pregnancy cocaine use <-> impaired 
infant correlation appearing from the non-partial correlation is 
spurious if it's partial is low or zero when social factors are 
partialled out.

 Other patterns and
>    inferences discussed below have to do with partial control and
>    suppression effects.
>    Partial correlation still requires meeting all the usual assumptions
>    of Pearsonian correlation: linearity of relationships, the same level
>    of relationship throughout the range of the independent variable
>    ("homoscedasticity"), interval or near-interval data, and data whose
>    range is not truncated.

So if that does not apply then presumably rank correlation and 
partial rank correlation must be used. Though I do not know if SPSS has 
partial rank correlation with significance yet.
 
>    Partial correlation is common when there is only one control variable
>    but is sometimes used when there are two or three. For large models,
>    researchers use path analysis or structural equation modeling when
>    data are near or at interval level, or use log-linear modeling for
>    lower-level data. Newer versions of structural equation modeling
>    software allow variables of any type on either side of the equation.
> [...]
> ********

And for homework there is still the question of what the above might
imply for thinking about impaired infant (hyperactive) -> social
factors, with the arrow in that direction.


  However,
> it seems that that's not the way things work in talk.politics.drugs,
> at least with respect to one poster.

>    Radford Neal

Politics is about trying to create a perception and how to expose that in 
an opponent. Sometimes there is honest questioning, and honest mistakes 
are sometimes admitted but may be fudged over. I am not quite sure what 
was what in the following, an excerpt from which I started out with in 
this article:

******************
Brian:
> The Pearson product-moment correlation would be used for linear data,

sazsz:
   Yes indeed. There are also techniques related to the Pearson R that can
be used to analyze data with non-linear models. 

Brian:
> which are given scores and are found to be symmetrically spread above 
and 
> below an average value. So it very often should not be used.

   Your mistake, among many: and yet a partial rank correlation is simply
a more complicated version of the same technique, except that simply
you're proposing to add in more variables to the model (to be "partialled
out", hence the term "partial" correlation), and of course variables that
already are obviously irrelevant and contaminating variables for the
purposes of drawing valid conclusions from the crack
baby study.
**********************

Of course it was not I who introduced the `contaminating variables', I 
just question the direction of any causes between them.

And as an extra, use `social factors' rather than `socioeconomic status'
as the rich may not be immune. I suppose the status could be made another
variable in the analysis.


-----= Posted via Newsfeeds.Com, Uncensored Usenet News =-----
http://www.newsfeeds.com - The #1 Newsgroup Service in the World!
-----==  Over 80,000 Newsgroups - 16 Different Servers! =-----
.
.
=================================================================
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