Hi, all. I've worked with NSCAW data at some length independently and as a doctoral student at UNC-Chapel Hill. After some discussions amongst ourselves and consulting with RTI and others, we decided to avoid using the baseline wave infant data. Here's the rationale: 1. Assessing infants is a task that really should be done by clinicians with expert training. The data collectors had some basic training but not really the expertise required.
2. The NSCAW instruments used to assess infants are the screener versions, which means they have few items. Hence, they are less reliable since any error is magnified. In doing growth modeling (and this was part of my dissertation), you can use the raw scores but you'll need to include time as a predictor. Really, it would be better to use the adjusted score since we already know the scores should
rise as time passes. In some diagnostic analyses I ran a few years ago on the NSCAW developmental instruments, I found most of them are most strongly influenced by immediate temporal influences, rather than historical ones. That is, the scores at Wave 4 are most closed correlated with other scores at Wave 4, not scores at Wave 3 or Wave 1. While you can generate growth curves (I used Bollen and Curran's SEM-based methodology but it's been shown that SEM and HLM based models produced basically identical results), there's a lot of missing data to fill in and the imputation models are necessarily complex (I used Markov Chain Monte Carlo but there are other systems that would work).
Hope this helps. Feel free to contact me on or off list if it's not. My office email is below my signature.
Chris
E. Christopher Lloyd, PhD
Assistant Professor
School of Social Work
University of Arkansas at Little Rock
2801 South University Avenue
Little Rock, AR 72204
501.569.8486 eclloyd@xxxxxxxx
--- On Wed, 6/30/10, Chaffin, Mark J. (HSC) <Mark-Chaffin@xxxxxxxxx> wrote:
From: Chaffin, Mark J. (HSC) <Mark-Chaffin@xxxxxxxxx>
Subject: RE: Longitudinal Analysis using PLS3 and BDI scores for children as dependent vars
To: "'Child Maltreatment Researchers'" <child-maltreatment-research-l@xxxxxxxxxxxxxxxx>
Cc: "'lchen@xxxxxxxxxxxxxx'" <lchen@xxxxxxxxxxxxxx>
Date: Wednesday, June 30, 2010, 9:04 AM
A different analytic tack might revel some different
findings. Early childhood developmental data is vulnerable to a variety
of analytic problems. Given the pace and discontinuity of child development
during this period, test items that were previously impossible suddenly become
trivial only a few months later. These stage-sequential patterns of
development are precisely why items often change on measures across early
childhood stages. In the developmental literature, analysts have
sometimes cautioned against analyzing standardized scores to model these types
of staged phenomena that do not approximate a smooth underlying growth pattern.
   Latent transition models may offer a better fit with the underlying
phenomena. Depending upon the task, these can be fitted at the item or
task level for failure/success, assuming that after the item is passed it would
stay passed even after it is dropped from the measure, and assuming that a new
item that is failed would also have been failed earlier had it been included.Â
This would clearly involve use of raw scores in the analysis, not standardized.
There are a number of other possibilities here that seem less
likely. Measure reliability may not be the issue, given that it would tend
to increase intercept variability, but not necessarily intercept position.Â
You might check variability of baseline scores to see if you have a
non-constant variability issue. I doubt that regression to the mean or
ceiling effects at baseline are the problem, but are worth checking, especially
checking models with vs. without slope-intercept correlation pathways.
The other possibility, which I unfortunately suspect may be the
actual one, Â is that this is not an artifact at all, but rather reflects real
development of children in child welfare and the growing impact of their
environments over time. That is, you might assume that infant measures
reflect mostly biologic or basic health status, whereas later measures come to
reflect suboptimal environments more and more the longer the child is in that
environment, then you would expect to see children whose development started
off fairly normal, but had growth that didn’t keep pace.
Mark
Â
From: Lijun Chen
[mailto:lchen@xxxxxxxxxxxxxx]
Sent: Monday, June 28, 2010 1:30 PM
Subject: Longitudinal Analysis using PLS3 and BDI scores for children as
dependent vars
Â
I am using the NSCAW
data to examine the developmental indicators, especially BDI-Cognitive and
PLS3, for infants (0-12 months at wave 1) through the 4 waves of data
collections. One thing that baffles me is that the BDI and PLS3 standard scores
for most children at the 2nd wave have dropped precipitously from the
baseline wave when they were under 12 months. This may indicate the poor
performance of these infants relative to the national norm. I wonder
whether the (un)reliability of the baseline scores should also be a
contributing cause. I would appreciate your opinions / comments on this.
IÂ plan
to adopt growth curve modeling in analyzing the development
trajectories of these infants using BDI and PLS3 scores as dependent variable.
Is it preferable to use the Standard Scores or the raw scores as the
dependent variables? Is it valid to use the raw scores as the dependent
variable since items included in the instrument at different waves are not
the same?
Your advice is
appreciated. Â Â Â Â
Chapin Hall at the
University of Chicago
|
