Posted by
Mike on
Oct 20, 2016; 7:10pm
URL: http://spssx-discussion.165.s1.nabble.com/contrast-orthogonal-coding-with-unequal-cell-frequencies-tp5733307p5733346.html
I think that there are bigger problems here
than the distinction between
contrast coefficient and contrast
coding. Let me point out what I think
IMHO are some greater problems:
(1) The OP calls nonexperimental variables
"independent variables"
which the OP may do by convention (everyone in
the area calls them
that) but from an experimental design
perspective, they are not
independent variables -- the OP may want to
call them"causal variables"
and provide a path diagram that shows how the
causal and other
variables affect the *outcome*
variable.
(2) Given the info below, one would think that
one would look at the
correlation matrix for all of the variables
to determine how all of
the variables are interrelated. In all
likelihood, all of the variables --
"causal", outcome, "confounder"/3rd variables"
-- are correlated.
The concept of "confounder" is peculiar in
this situation because
it doesn't seem that nurses were randomly
assigned to nursing
specialty and one now wants to determine
whether random assignment
worked (i.e., the nursing specialty groups are
statistically equivalent
on background variables of age, years of
experience, etc.).
A path diagram explicitly identifying the
relationships that one expects
on a theoretical basis, would be very helpful
in clearing up what
is/isn't correlated -- and don't even get
started on mediation and
moderation effects.
(3) An alternative way of conceptualizing what
the OP want to do
is think in terms of Analysis of
Covariance, that is, does mean level
of perceived stress vary significantly as a
function of nursing specialty
AFTER removing the effects of other variables
(i.e., age, etc.).
IMHO, this puts the focus on the relationship
of greatest interest.
I know that the equivalent can be done in
multiple regression (indeed,
superfans of MR like Pedhazur and other prefer
MR to traditional
ANOVA analyses) but then we get the situation
that we're in right
now. I think that the original question
was perhaps misunderstood
because complete information was not provided
and the issue of
orthogonal coding for unequal sample sizes was
maybe a side issue
or even irrelevant.
(4) I could be wrong but it seems to me that
what the OP wants
to do is a MR that enters all of the
background variables first,
determine if the is a significant relationship
between perceived
stress and these variables (and which ones
significant), and then
enter the variable nursing specialty
(categories appropriately
coded) to determine if provides a significant increase in the
variance accounted for or R^2.
(5) I think it may become relevant to ask
whether orthogonal coding
should be used of nursing specialty
categories because I don't think
it likely that N for all specialties are equal. The situation
is complicated
by background variables since it is likely
that nursing specialty
will differ on some/all of the background
variables. Again, I think
this is made clearer from an ANCOVA
perspective but I'm
sure that folks who think in regression terms
will disagree.
(6) I could be wrong (probably am) but maybe
the following
analysis should be conducted: regress
perceived stress on all
of the background variables and if there is a
significant relationship,
save the residuals or studentized residuals,
transform them to
perceived stress scores by adding the original
mean and multiplying
by the original standard deviation, and then
regress these new
scores on an orthogonal contrast representing
nursing specialty.
The new stress scores should represent the
variance that remains
after the effects of background variables have
been removed
(explicitly) and one can ask if there is any
relationship between
them and the coding for nursing
specialty..
(7) Does anyone think that generating
propensity scores for the
background variables for the regression of
perceived stress on
nursing specialty categories might be an
alternative analysis to
consider?
(8) Does anyone wonder if a single nurse might
report having
multiple specialties? If so, how is this
represented in the data?
(9) My understanding of the OP's situation
could be completely
wrong, so feel free to ignore everything I
said above. But I do
think that maybe we have been focusing on the
wrong issues.
-Mike Palij
New York University
----- Original Message -----
Sent: Thursday, October 20, 2016 12:36
PM
Subject: Re: contrast (orthogonal) coding
with unequal cell frequencies
Are you speaking in your initial question about contrast
coefficients (coefficients in a contrast, they sum to zero) or contrast
coding (values of contrast variables)? See http://stats.stackexchange.com/a/221868/3277
20.10.2016 19:07, Sidra пишет:
I'm sorry fellows, I do not have a background of statistics ..that's why I'm
having a hard time understanding your suggestions here. Perhaps I need to be
a little bot more comprehensive.
I have IVs of nursing specialty, qualification, age, years of experience,
work shift, marital status and childberaing status; independent variable
being perceived stress by nurses. I need to see the effect of nursing
specialty (as main variable of interest) on perceived stress while
controlled for confounders.
I will identify confounders by noting crude coefficient of nursing specialty
and then noticing the change in its coefficient when each IV is placed in
the model with nursing specialty (one variable at a time). If adding a
variable in regression model brings a change of more than 10% in coefficient
of nursing specialty, I ll treat that variable as a confounder. Since to
find out if a variable is a confounder, I have to put that confounder alone
along with nursing specialty in regression model, I am not sure if I can
treat contrast 1 and contrast 2 (in place of marital status and childbearing
status) as individual variables to see how much change each brings about in
the coefficient of nursing specialty separately.
I hope I have made myself sufficiently clear. Please bear with me and offer
your kind insight on this problem.
=====================
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