Figure
4.3: Classification Tree
|
|---|
|
Example
of a classification tree. Patients with an acute myocardial
infarction are classified according to age, Killip
class and number of leads with ECG elevation (STE). The probability of
30-day mortality is shown for each classification, e.g. 20% in those
older than 73.7 years. Data are from a sample (n=752) of the GUSTO-I
data (Lee
et al., 1995); http://www.eur.nl/fgg/mgz/software.html.
Note that a tree based on the full data set (n=40,830) would
become very complex. |
|
Classification
trees are very attractive in their presentation, but are
relatively "data hungry." This is because subgroups are created
within every branch of the tree, which leads to reduced sample
size in every branch further down the tree. Essentially, interaction
effects are assumed, while regression models generally assume
no interaction. Another disadvantage is that continuous variables
need to be categorized.
Neural
networks vary in structure and implementation, but generally
include one or more so-called "hidden layers" between predictors
and the outcome. Neural networks are very flexible, and naturally
allow for nonlinearity and interaction in predictor variables.
The linear
and nonlinear regression models that we consider all fall within
the class of generalized linear models. These are characterized
by a linear regression formula. The relationship between predictors
and outcome is non-linear because of a link function between
the linear predictor and the outcome, such as the log odds or
logit in logistic regression analysis.
Figure
4.1: Logistic Link Function
|
|---|
|
Illustration
of the logistic link function. The relationship between
the probability of an outcome and the logit of the probability
is a characteristic curve. The logit is calculated
as: ln(probability/(1-probability)). When the logit
is 0, the probability is 50%. |
|
Interestingly,
neural networks can be viewed as implementations of statistical
models that are either more complex nonlinear models, generalized
additive models, or generalized nonlinear models (Hastie
and Tibshirani, 1990).