Page 315 - Invited Paper Session (IPS) - Volume 2
P. 315
IPS254 Thaddeus Tarpey et al.
For the diagnosis of autism spectrum disorder (ASD), the debate has
centered on whether distinct disease categories exist or is there is an
underlying “spectrum” of disease severity that includes for example ADHD
(Grzadzinski et al., 2011). This issue was the focus of a recent article Kim et al.
(2018) that proposes that ASD consists of three spectrums instead of a single
spectrum. Their statistical analysis that led them to conclude ASD consists fo
three spectrums was based on Latent Class Factor Analysis (LCFA). The LCFA
incorporates both categorical features for presumed latent classes as well as
“dimensional” features corresponding to within group continuous latent
factors. Their statistical strategy in determining an appropriate model to use
was to build up the model starting with only two latent classes and one
continuous factor per group and then increase the number of classes and
factors per class.
Another avenue of discovery in the context of psychiatric nosology is to
incorporate information of defining diagnosis categories not only on
symptoms but also on treatment outcome. Of course, a problem with this
approach is that the type of treatment usually depends on diagnosis. However,
many psychiatric illnesses are treated with the same types of medications.
A major statistical challenge for this work is the issue of aliasing. For
psychiatric nosology, finite mixtures of normal distributions are an attractive
approach to the unsupervised learning problem of estimating parameters for
discrete disease populations. However, it has long been recognized that finite
mixture distributions are often indistinguishable from homogenous
continuous distributions (Pearson, 1895). If the components of the mixture
distribution are normal, then the model is identifiable (e.g., Teicher, 1961). If
distinct populations do not exist and disease severity varies continuously, then
one can use an infinite mixture of normals which is also identifiable under
certain conditions (Bruni and Koch, 1985). If it is believed there are distinct
disease categories, then from a statistical point of view, this problem can be
cast in the context of unsupervised learning and finite mixture model
approaches may be suitable. However, if distinct disease classes do not exist,
and disease severity varies continuously along a spectrum, then an infinite
mixture model may be a more appropriate statistical approach to the
psychiatric nosology problem. For example, Tarpey and Petkova (2010)
proposed an infinite mixture model in the context of a simple regression where
the predictor variable is continuous and latent. If the latent predictor is
Bernoulli, then the regression model becomes a 2-component finite mixture
model. Tarpey and Petkova (2010) considered a predictor variable that has a
beta distribution whereby in the limiting cases, the beta can converge to a
Bernoulli. In this way, the 2-component mixture and infinite mixture can be
described in a single model. Tarpey and Petkova (2010) introduced this model
in order to model placebo response when treating depression.
302 | I S I W S C 2 0 1 9