See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/240719582

An Introduction to Classification and Regression Tree (CART) Analysis

Article · January 2000

CITATIONS READS
273 15,127

1 author:

Roger J Lewis
Harbor-UCLA Medical Center
262 PUBLICATIONS   11,220 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Roger J Lewis on 13 January 2014.

The user has requested enhancement of the downloaded file.

 An Introduction to Classification and Regression Tree (CART) Analysis

Roger J. Lewis, M.D., Ph.D.

Department of Emergency Medicine
Harbor-UCLA Medical Center
Torrance, California

Presented at the 2000 Annual Meeting of the Society for Academic Emergency Medicine in San
Francisco, California.

Contact Information:

Roger J. Lewis, MD, PhD

Department of Emergency Medicine
Harbor-UCLA Medical Center, Box 21
1000 West Carson Street
Torrance, California 90509
Tel: (310) 222-6741
Fax: (310) 782-1763
Email: [email protected]
c:\roger\saem2000\introduction to cart
Introduction to CART
Roger J. Lewis, M.D., Ph.D.

Introduction
A common goal of many clinical research studies is the development of a reliable clinical
decision rule, which can be used to classify new patients into clinically-important categories. Examples
of such clinical decision rules include triage rules, whether used in the out-of-hospital setting or in the
emergency department, and rules used to classify patients into various risk categories so that appropriate
decisions can be made regarding treatment or hospitalization.
Traditional statistical methods are cumbersome to use, or of limited utility, in addressing these
types of classification problems. There are a number of reasons for these difficulties. First, there are
generally many possible “predictor” variables which makes the task of variable selection difficult.
Traditional statistical methods are poorly suited for this sort of multiple comparison. Second, the
predictor variables are rarely nicely distributed. Many clinical variables are not normally distributed and
different groups of patients may have markedly different degrees of variation or variance. Third, complex
interactions or patterns may exist in the data. For example, the value of one variable (e.g., age) may
substantially affect the importance of another variable (e.g., weight). These types of interactions are
generally difficult to model, and virtually impossible to model when the number of interactions and
variables becomes substantial. Fourth, the results of traditional methods may be difficult to use. For
example, a multivariate logistic regression model yields a probability of disease, which can be calculated
using the regression coefficients and the characteristics of the patient, yet such models are rarely utilized
in clinical practice. Clinicians generally do not think in terms of probability but, rather in terms of
categories, such as “low risk” versus “high risk.”
Regardless of the statistical methodology being used, the creation of a clinical decision rule
requires a relatively large dataset. For each patient in the dataset, one variable (the dependent variable),
records whether or not that patient had the condition which we hope to predic t accurately in future
patients. Examples might include significant injury after trauma, myocardial infarction, or subarachnoid
hemorrhage in the setting of headache. In addition, other variables record the values of patient
characteristics which we believe might help us to predict the value of the dependent variable. For
example, if one hopes to predict the presence of subarachnoid hemorrhage, a possible predictor variable
might be whether or not the patient's headache was sudden in onset; another possible predictor would be
whether or not the patient has a history of similar headaches in the past. In many clinically-important
settings, the number of possible predictor variables is quite large.
Within the last 10 years, there has been increasing interest in the use of classification and
regression tree (CART) analysis. CART analysis is a tree-building technique which is unlike traditional
data analysis methods. It is ideally suited to the generation of clinical decision rules. Because CART
analysis is unlike other analysis methods it has been accepted relatively slowly. Furthermore, the vast
majority of statisticians have little or no experience with the technique. Other factors which limit CART's
general acceptability are the complexity of the analysis and, until recently, the software required to
perform CART analysis was difficult to use. Luckily, it is now possible to perform a CART analysis
without a deep understanding of each of the multiple steps being completed by the software. In a number
of studies, I have found CART to be quite effective for creating clinical decision rules which perform as
well or better than rules developed using more traditional methods. In addition, CART is often able to
uncover complex interactions between predictors which may be difficult or impossible to uncover using
traditional multivariate techniques.
The purpose of this lecture is to provide an overview of CART methodology, emphasizing
practical use rather than the underlying statistical theory.

Classification and Decision Problems

A classification problem consists of four main components. The first component is a categorical
outcome or “dependent” variable. This variable is the characteristic which we hope to predict, based on
the “predictor” or “independent” variables. Typical outcome variables are survival, need for surgery, and
presence of myocardial infarction. The second component of a classification problem are the “predictor”

Page 2
Introduction to CART
Roger J. Lewis, M.D., Ph.D.

or “independent” variables. These are the characteristics which are potentially related to the outcome
variable of interest. In general, there are many possible predictor variables. The third component of the
classification problem is the learning dataset. This is a dataset which includes values for both the
outcome and predictor variables, from a group of patients similar to those for whom we would like to be
able to predict outcomes in the future. The fourth component of the classification problem is the test or
future dataset, which consists of patients for whom we would like to be able to make accurate predictions.
This test dataset may or may not exist in practice. While it is commonly believed that a test or validation
dataset is required to validate a classification or decision rule, a separate test dataset is not always
required to determine the performance of a decision rule.
A decision problem includes two components in addition to those found in a classification
problem. These components are a “prior” probability for each outcome, which represents the probability
that a randomly-selected future patient will have a particular outcome, and a decision loss or cost matrix.
The decision cost matrix represents the inherent cost associated with misclassifying a future patient. For
example, it is a much more serious error to classify a patient with an emergent medical condition as non-
urgent, than to misclassify a patient with a non-urgent medical condition as urgent. A sample cost matrix
is shown below, for a triage problem in which patients are classified as emergent, urgent, and non-urgent.
The worst possible error, consisting of classifying a truly emergent patient as non-urgent (undertriage), is
fifteen times as serious as misclassifying an urgent patient as emergent (overtriage).
As the first
example, consider the Classified by Tree as
problem of selecting the Emergent Urgent Non-Urgent
best size and type of True Value of Emergent 0 5 15
laryngoscope blade for Outcome Urgent 1 0 5
pediatric patients Variable Non-Urgent 3 2 0
undergoing intubation.
The outcome variable, the Example Decision Cost Matrix.
best blade for each patient
(as determined by a consulting pediatric airway specialist), has three possible values: Miller 0, Wis-
Hipple 1.5, and Mac 2. The two predictor variables are measurements of neck length and oropharyngeal
height. The learning dataset is shown below. As can be
seen from the figure, the smallest patients are best intubated
Laryngoscope Blade versus Neck and OP Measurements
with the Miller 0, medium sized patients with the Wis-
Hipple 1.5, and the largest patients with the Mac 2. 8
One possible approach to analyzing these data 7
Mac 2
would be to use multivariate logistic regression, using neck
length and oropharyngeal height as the two independent 6

predictor variables. A multivariate logistic regression 5

Neck Length

model yields regression coefficients which, when used in

4
logit expressions, give the probability that each blade is the
best for that patient. Logistic regression equations are very 3 Wis-Hipple 1.5
Miller 0
difficult to use in clinical practice, especially in situations
2
such as this, in which the outcome variable has more than
two levels. Furthermore, it is difficult to incorporate 1
possible interactions in a multivariate logistic regression 0
model, and the model makes parametric assumptions which 0 1 2 3
may not be valid. OP Height
As shown on the next page, the logistic regression
model yields regression coefficients for the two
independent variables. Both of these regression coefficients are statistically significant, suggesting that
both neck length and oropharyngeal height are important predictors of the best laryngoscope blade.

Page 3
Introduction to CART
Roger J. Lewis, M.D., Ph.D.

Although the interpretation of this Logistic Regression Results

output is beyond the scope of the
current lecture, suffice it to say that Parameter Odds
this type of model is rarely clinically Variable Estimate ± SE P value Ratio
useful (especially when faced with a
small child who needs emergent INTERCP1 22.4 ± 13.5 0.0974 .
airway management). INTERCP2 45.5 ± 13.5 0.0007 .
NECK_LEN -5.8 ± 2.3 0.0114 0.003
OP_HEIGH -10.5 ± 3.7 0.0051 0.000
Disclosure
The Classification and
Regression Tree (CART) software to be illustrated in this lecture is a commercial product manufactured
and sold by Salford Systems (http://www.salford-systems.com). Salford Systems has donated CDs which
contain a trial version of their CART software, some additional modeling software not to be discussed in
this lecture, and copies of the datasets used in this lecture (provided by the lecturer). The lecturer does
not intend this presentation to be an endorsement of this particular software package. This is the only
CART software with which the lecturer has significant personal experience, making it impossible for him
to comment on the capabilities of other competitive products. The lecturer receives no financial support
from Salford Systems, and SAEM has received no support from Salford Systems beyond the donation of
the above-mentioned CDs for the use of meeting attendees.

Binary Recursive Partitioning

CART analysis is a form of binary recursive partitioning. The term “binary” implies that each
group of patients, represented by a “node” in a decision tree, can only be split into two groups. Thus,
each node can be split into two child nodes, in which case the original node is called a parent node. The
term “recursive” refers to the fact that the binary partitioning process can be applied over and over again.
Thus, each parent node can give rise to two child nodes and, in turn, each of these child nodes may
themselves be split, forming
additional children. The term
Node 1
“partitioning” refers to the fact that
Yes Class = Miller 0 No the dataset is split into sections or
Neck Length <
2.45 partitioned.
The figure to the left shows
Node -1 No
Node 2
Class = WH 1.5 Yes the classification and regression
Class = Miller 0 OP Height < tree which results from analysis of
1.75
the laryngoscope blade selection
Node -3 Node -2
data shown above. This tree
Class = Mac 2 Class = WH 1.5 consists of a root node (Node 1),
containing all patients. This node
is split based on the value of the neck length variable. If the neck length is < 2.45 centimeters, then those
patients are put in the first terminal node, denoted Node -1, and the best blade is predicted to be a Miller
0. All other patients are placed in Node 2. The group of patients in Node 2 is initially assigned a Wis-
Hipple 1.5 blade but they are also split based on their oropharyngeal height. Those patients with an
oropharyngeal height less than 1.75 are placed in terminal Node -2, and assigned a Wis-Hipple 1.5 blade,
while those with an oropharyngeal height ≥ 1.75 are placed in terminal Node –3 and assigned a Mac 2
blade.
Several things should be pointed out regarding this CART tree. First, it is much simpler to
interpret than the multivariate logistic regression model, making it more likely to be practical in a clinical
setting. Secondly, the inherent “logic” in the tree is easily apparent, and it makes clinical sense.
Interestingly, it has been shown that clinical decision rules which make sense to clinicians are more
likely to be followed in clinical practice than rules in which the reasoning is not apparent.

Page 4
Introduction to CART
Roger J. Lewis, M.D., Ph.D.

In the two figures below, a visual illustration of the CART approach is given. The root node,
which contains all patients, is split in two, analogous to a horizontal line being drawn at neck length =
2.45. All patients below the line, which are those found in the first terminal node, are assigned a
predicted class of Miller 0. The group of patients above the original line are then split by a second line
drawn at oropharyngeal height = 1.75. Those to the left of this line are assigned a class of Wis-Hipple
1.5, while those to the right of the line are assigned the class Mac 2. It is important to note that this
second line applies only to one of the regions, which corresponds to the second parent node (Node 2).
This process of partitioning is easy to visualize in two dimensions (i.e., when there are only two possible
predictor variables) but is difficult or impossible to picture when there are five, ten, or dozens of possible
predictors.

Laryngoscope Blade versus Neck and OP Measurements Laryngoscope Blade versus Neck and OP Measurements

8 8

7 Mac 2 7
Mac 2
6 6

5 5
Neck Length

Neck Length
Wis-Hipple 1.5 Wis-Hipple 1.5
4 4

3 3

2 2
Miller 0 Miller 0
1 1

0 0
0 1 2 3 0 1 2 3
OP Height OP Height

The text box to the right shows the set of commands, contained within a “command” file, which
were used to produce the analysis represented by the CART tree. While there are 16 lines of commands,
half are generic formatting commands or the USE 'c:\cart\data\airway.sys'
definition of the misclassification costs. This LOPTIONS MEANS = NO, PREDICTION = NO,
command file is included in the CD which has PLOTS = YES, TIMING = YES, PRINT
been distributed, as is the dataset “airway.sys.” FORMAT = 3
MODEL best_bla
Advantages and Disadvantages of CART KEEP neck_len op_heigh
CART analysis has a number of CATEGORY best_bla = 3 [min = 1]
advantages over other classification methods, PRIORS EQUAL
Misclass UNIT
including multivariate logistic regression. First, it
Misclasify Cost = 1 Classify 1 as 2
is inherently non-parametric. In other words, no Misclasify Cost = 1 Classify 1 as 3
assumptions are made regarding the underlying Misclasify Cost = 1 Classify 2 as 3
distribution of values of the predictor variables. Misclasify Cost = 1 Classify 3 as 1
Thus, CART can handle numerical data that are Misclasify Cost = 1 Classify 3 as 2
highly skewed or multi-modal, as well as Misclasify Cost = 1 Classify 2 as 1
categorical predictors with either ordinal or non- LIMIT DEPTH=16
ordinal structure. This is an important feature, as ERROR CROSS = 20
it eliminates analyst time which would otherwise
be spent determining whether variables are normally distributed, and making transformation if they are
not.
As discussed below, CART identifies “splitting” variables based on an exhaustive search of all
possibilities. Since efficient algorithms are used, CART is able to search all possible variables as

Page 5
Introduction to CART
Roger J. Lewis, M.D., Ph.D.

splitters, even in problems with many hundreds of possible predictors. [While some listeners may
shudder at possible problems with overfitting and data dredging, these issues are dealt with in depth later].
CART also has sophisticated methods for dealing with missing variables. Thus, useful CART
trees can be generated even when important predictor variables are not known for all patients. Patients
with missing predictor variables are not dropped from the analysis but, instead, “surrogate” variables
containing information similar to that contained in the primary splitter are used. When predictions are
made using a CART tree, predictions for patients with missing predictor variables are based on the values
of surrogate variables as well.
Another advantage of CART analysis is that it is a relatively automatic “machine learning”
method. In other words, compared to the complexity of the analysis, relatively little input is required
from the analyst. This is in marked contrast to other multivariate modeling methods, in which extensive
input from the analyst, analysis of interim results, and subsequent modification of the method are
required.
Finally, CART trees are relatively simple for nonstatisticians to interpret. As mentioned above,
clinical decision rules based on trees are more likely to be feasible and practical, since the structure of the
rule and its inherent logic are apparent to the clinician.
Despite its many advantages, there are a number of disadvantages of CART which should be kept
in mind. First, CART analysis is relatively new and somewhat unknown. Thus, there may be some
resistance to accept CART analysis by traditional statisticians (some of whom consult for prestigious
medical journals). In addition, there is some well-founded skepticism regarding tree methodologies in
general, based on unrealistic claims and poor performance of earlier techniques. Thus, some statisticians
have a generalized distrust of this approach. Because of its relative novelty, it is difficult to find
statisticians with significant expertise in CART. Thus, it may be difficult to find someone to help you use
CART analysis at your own institution. Because CART is not a standard analysis technique, it is not
included in many major statistical software packages (e.g., SAS).

Steps in Cart
CART analysis consists of four basic steps. The first step consists of tree building, during which
a tree is built using recursive splitting of nodes. Each resulting node is assigned a predicted class, based
on the distribution of classes in the learning dataset which would occur in that node and the decision cost
matrix. The assignment of a predicted class to each node occurs whether or not that node is subsequently
split into child nodes. The second step consists of stopping the tree building process. At this point a
“maximal” tree has been produced, which probably greatly overfits the information contained within the
learning dataset. The third step consists of tree “pruning,” which results in the creation of a sequence of
simpler and simpler trees, through the cutting off of increasingly important nodes. The fourth step
consists of optimal tree selection, during which the tree which fits the information in the learning dataset,
but does not overfit the information, is selected from among the sequence of pruned trees. Each of these
steps will be discussed in more detail below.

Tree Building
Tree building begins at the root node, which includes all patients in the learning dataset.
Beginning with this node, the CART software finds the best possible variable to split the node into two
child nodes. In order to find the best variable, the software checks all possible splitting variables (called
splitters), as well as all possible values of the variable to be used to split the node. A number of clever
programming tricks are used to reduce the time required to search through all possible splits. In the case
of a categorical variable, the number of possible splits increases quickly with the number of levels of the
categorical variable. Thus, it is useful to tell the software the maximum number of levels for each
categorical variable.
In choosing the best splitter, the program seeks to maximize the average “purity” of the two child
nodes. A number of different measures of purity can be selected, loosely called “splitting criteria” or

Page 6
Introduction to CART
Roger J. Lewis, M.D., Ph.D.

“splitting functions.” The most common splitting function is the “Gini”, followed by “Twoing.”
Although the CART software manual recommends experimenting with different splitting criteria, these
two methods will give identical results if the outcome variable is a binary categorical variable.
As discussed below, each node (even the root node) is assigned a predicted outcome class. The
process of node splitting, followed by the assignment of a predicted class to each node, is repeated for
each child node and continued recursively until it is impossible to continue.

Assignment of Node Classes

Each node, even the root node, is assigned a
predicted class. This is necessary, as there is no way to Criteria for Assigning Classes to Nodes:
know during the tree-building process which nodes will
end up being terminal nodes after pruning. The C( j| i) is cost of classifying i as j.
predicted class assigned to each node depends on three π (i) is prior probability of i.
factors: (1) the assumed prior probability of each class Ni is number of class i in dataset.
within future datasets; (2) the decision loss or cost
Ni (t ) is number of class i in node.
matrix; and (3) the fraction of subjects with each
outcome in the learning dataset that end up in each node. Node is class i, if
The function used to assign predicted classes to each C( j|i)π (i) Ni (t ) N i
>
node is shown at right. This method of node class C(i| j)π ( j) N j (t ) N j
assignment ensures that the tree has a minimal expected for all values of j.
average decision cost for future datasets similar to the
learning dataset in which the probability of each
outcome is equal to the assumed prior probabilities.

Missing Variables
For each node, the “primary splitter” is the variable that best splits the node, maximizing the
purity of the resulting child nodes. When the primary splitting variable is missing for an individual
observation, that observation is not discarded but, instead, a surrogate splitting variable is sought. A
surrogate splitter is a variable whose pattern within the dataset, relative to the outcome variable, is similar
to the primary splitter. Thus, the program uses the best available information in the face of missing
values. In datasets of reasonable quality this allows all observations to be used. This is a significant
advantage of this methodology over more traditional multivariate regression modeling, in which
observations which are missing any of the predictor variables usually are often discarded.

Stopping Tree Building

As mentioned above, the tree building process goes on until it is impossible to continue. The
process is stopped when: (1) there is only one observation in each of the child nodes; (2) all observations
within each child node have the identical distribution of predictor variables, making splitting impossible;
or (3) an external limit on the number of levels in the maximal tree has been set by the user (“depth”
option).
The “maximal” tree which is created is generally very overfit. In other words, the maximal tree
follows every idiosyncrasy in the learning dataset, many of which are unlikely to occur in a future
independent group of patients. The later splits in the tree are more likely to represent over fitting than the
earlier splits, although one part of the tree may need only one or two levels, while a different branch of
the tree may need many levels in order to fit the true information in the dataset. A major breakthrough of
the CART methodology was the realization that there is no way during the tree-building process to know
when to stop, and that different parts of the tree may require markedly different depths.

Page 7
Introduction to CART
Roger J. Lewis, M.D., Ph.D.

Tree Pruning
In order to generate a sequence of simpler and simpler trees, each of which is a candidate for the
appropriately-fit final tree, the method of “cost-complexity” pruning is used. This method relies on a
complexity parameter, denoted α, which is gradually increased during the pruning process. Beginning at
the last level (i.e., the terminal nodes) the child nodes are pruned away if the resulting change in the
predicted misclassif ication cost is less than α times the change in tree complexity. Thus, α is a measure
of how much additional accuracy a split must add to the entire tree to warrant the additional complexity.
As α is increased, more and more nodes (of increasing importance) are pruned away, resulting in simpler
and simpler trees.

Optimal Tree Selection

The maximal tree will always fit the learning dataset with higher accuracy than any other tree.
The performance of the maximal tree on the original learning dataset, termed the “resubstitution cost,”
generally greatly overestimates the performance of the tree on an independent set of data obtained from a
similar patient population. This occurs because the maximal tree fits idiosyncrasies and noise in the
learning dataset, which are unlikely to occur with the same pattern in a different set of data. The goal in
selecting the optimal tree, defined with respect to expected performance on an independent set of data, is
to find the correct complexity parameter α so that the information in the learning dataset is fit but not
overfit. In general, finding this value for α
would require an independent set of data, but
this requirement can be avoided using the
Decision Cost
technique of cross validation (see below). Optimal Tree
Average

Test
The figure to the right shows the Data
relationship between tree complexity, reflected
“Underfit” “Overfit” Learning
by the number of terminal nodes, and the
decision cost for an independent test dataset Data
and the original learning dataset. As the
number of nodes increases, the decision cost 0 2 4 6 8 10 12 14
decreases monotonically for the learning data. Large α Complexity Small α
This corresponds to the fact that the maximal
tree will always give the best fit to the learning
dataset. In contrast, the expected cost for an independent dataset reaches a minimum, and then increases
as the complexity increases. This reflects the fact that an overfitted and overly complex tree will not
perform well on a new set of data.

Cross Validation
Cross validation is a computationally-intensive method for validating a procedure for model
building, which avoids the requirement for a new or independent validation dataset. In cross validation,
the learning dataset is randomly split into N sections, stratified by the outcome variable of interest. This
assures that a similar distribution of outcomes is present in each of the N subsets of data. One of these
subsets of data is reserved for use as an independent test dataset, while the other N-1 subsets are
combined for use as the learning dataset in the model-building procedure (see the figure on the next
page). The entire model-building procedure is repeated N times, with a different subset of the data
reserved for use as the test dataset each time. Thus, N different models are produced, each one of which
can be tested against an independent subset of the data. The amazing fact on which cross validation is
based is that the average performance of these N models is an excellent estimate of the performance of
the original model (produced using the entire learning dataset) on a future independent set of patients.

Page 8
Introduction to CART
Roger J. Lewis, M.D., Ph.D.

When cross validation is used in CART, the entire tree

Reserved Test Data
building and pruning sequence is conducted N times. Thus,
there are N sequences of trees produced. Trees within the
sequences are matched up, based on their number of terminal
nodes, to produce an estimate of the performance of the tree in Learning Data
predicting outcomes for a new independent dataset, as a function
of the number of terminal nodes or complexity. This allows a
data-based estimate of the tree complexity which results in the
Learning Data
best performance with respect to an independent dataset. Using
this method, a minimum cost occurs when the tree is complex
enough to fit the information in the learning dataset, but not so
complex that “noise” in the data is fit. The figure below right Learning Data
shows a typical minimum which should occur in the cross- l
validation estimate of the misclassification cost, as a function of l
the number of terminal nodes or complexity. l

Example: HIV-Triage
As our next example, consider a dataset involving the Learning Data
triage of self-identified HIV-infected patients who present to the
emergency department (ED) for care. The outcome variable is the “urgency” of the visit, which has three
levels: emergent, urgent, and non-urgent. These
urgency levels are based on a retrospective evaluation of
the final diagnosis and clinical course. The patient's classification Cost
historical and presenting features, and the results of an Optimal Tree
CV Mis-

abbreviated and focussed review of systems are the

predictor variables. The dataset includes 389
observations and is included on the CD distributed
during the lecture.
0 2 4 6 8 10 12 14
Initial Screen
The initial screen which appears when the
Complexity
CART program is run includes the standard “File” and
“View” menus. The “File” menu includes an option which allows one to submit an entire command file
in order to conduct an analysis. In addition, one can use the “View” to open a “Notebook” window,
through which a command file may be viewed, edited, and submitted for processing.
The command files included on the CD can be opened in the Notepad window or submitted
directly from the File option. The use of the Notepad window allows editing of the command file (e.g., to
give the correct location for the data on your computer) followed by submission for processing.
Alternatively, the sample dataset may be opened, and a graphical user interface for model building may
be used in lieu of a command file. When the graphical model-building interface is used, the associated
command file may be saved, so that it can be viewed, edited, and reused later.
The text box on the next page shows the command file (“hivtree.cmd”) which can be used to
analyze the HIV triage dataset included on the CD. The MODEL statement shows the outcome variable,
in this case a numerical representation of the urgency of the ED visit. The KEEP command lists the
possible predictor variables to be considered in building the decision tree. [The version on the CD does
not use the KEEP command, as only the correct variables have been included in the dataset.] The
PRIORS statement shows that the actual distribution of outcomes in the learning dataset is to be used as
the prior probabilities for outcomes when assigning outcome classes to nodes. The MISCLASSIFY
statement defines the costs of various misclassification errors which may occur. In this case, the most
serious error is to misclassify a patient who is truly in class 0 (emergent) as class 2 (non-urgent). The

Page 9
Introduction to CART
Roger J. Lewis, M.D., Ph.D.

model is limited to a depth of 16, which USE 'C:\CART\data\HIVNEW.SYS'

limits the size of the maximal tree which LOPTIONS MEANS = NO, PREDICTION = NO,
may be constructed. In addition, cross PLOTS = YES, TIMING = YES, PRINT
validation is used, with the sample size FORMAT = 3
divided into 10 subsets. Each of the MODEL NVISITCA
commands used in the command file is KEEP NSOBE, NDIARR, NVOMIT, NDYSPH,
described in detail in the online NATAXI, NBEHAV, NWORSEHA, NHEADA,
documentation included with the software. NSWEAT, NCHILL, NSOBA, NDIZZY, NCOUGH,
SBP, RESP, PULSE, TEMP, NCCSOB,
The graphical user interface used for
NCCVOMDI, NCCHEADA, NCCFEVCH, NCCWEAKN,
model setup, as well as the screen which NCCOTHER, CD4
appears during processing of the command PRIORS DATA
file are shown (below right and top of next CATEGORY NVISITCA = 3 [min= 0]
page). In the latter case, 20-fold cross Misclass UNIT
validation is being used, which requires the Misclasify Cost = 4 Classify 0 as 1
creation of 21 tree sequences (one sequence Misclasify Cost = 8 Classify 0 as 2
using all learning data, and 20 tree sequences Misclasify Cost = 4 Classify 1 as 2
in which part of the data has been withheld Misclasify Cost = 6 Classify 2 as 0
Misclasify Cost = 4 Classify 2 as 1
for testing). Misclasify Cost = 1 Classify 1 as 0
LIMIT DEPTH=16
Results and Output ERROR CROSS = 10
The original CART software was BOPTIONS SURROGATES = 5 COMPETITORS = 3,
written for non-graphical computers, and all TREELIST = 10, BRIEF
output had to be printable on old style line BOPTIONS SERULE = 1 ,IMPORTANCE = 1
printers. The current CART software still BOPTIONS COMPLEXITY = 0.0, NCLASSES = 8
produces the old text-based reports, although TREE 'C:\CART\data\HIVTREE.TR1'
it provides a Window interface for
navigating the text reports. The
software also provides a fully
graphical navigator, which presents
the same information in a more user-
friendly graphical and tabular format.
The CART report consists of
7 sections. The first section consists
of the tree sequence, which includes
the primary sequence of trees based
on the entire learning dataset, as well
as cross validation estimates of tree
misclassification costs for an
independent set of data. The second
section of the report gives detailed
node information, including the
splitting criteria of each node,
surrogate variables to be used if the primary splitter is missing, and the distribution of outcomes for the
learning dataset at each node. The third section of the report gives information on the terminal nodes,
including the distribution of outcome classes.
The next three sections of the report all involve misclassification rates. The fourth section of the
report gives general misclassification information for the learning dataset (which overestimates the
performance of the tree), while the fifth section give cross validation estimates of misclassification rates
for an independent dataset. The sixth section give additional information on misclassif ication rates for the
learning dataset. The last section of the report gives information on the relative importance of different

Page 10
Introduction to CART
Roger J. Lewis, M.D., Ph.D.

variables in the dataset. These

importance measures incorporate
information both on the use of
variables as primary splitters, and
also their relative worth as surrogate
variables when primary splitters are
missing. A discussion of variable
importance is beyond the scope of
this lecture.
The accompanying figures
on pages 11-13 show the primary
Report Screen, the tree sequence
from a typical report (the HIV
Triage example), as well as the
graphical navigator which can be
used to examine any tree in the tree
sequence in detail. In addition, the
navigator can be used to view
detailed reports on misclassification
rates, both for the learning dataset
and for an independent dataset
(based on cross validation).
The CART software can
also produce a script for the program
allCLEAR to be used to generate
graphical representations of trees for
exporting to other programs (e.g.,
Microsoft Word or Microsoft
Powerpoint). The decision tree
resulting from an analysis of the
HIV triage data, represented using
the allCLEAR software program is
illustrated on page 13. The program
allCLEAR is no longer
automatically
included with the Dependent variable: NVISITCA
software, and the
CART program can Terminal Cross-Validated Resubstitution Complexity
also print nice looking Tree Nodes Relative Cost Relative Cost Parameter
trees by itself, as well -----------------------------------------------------
1 48 0.828 +/- 0.059 0.256 0.000
as export graphical
5 35 0.798 +/- 0.059 0.299 0.008
formats. 6 28 0.784 +/- 0.059 0.341 0.010
In the HIV 7 25 0.731 +/- 0.059 0.361 0.011
Triage example, the 8 19 0.722 +/- 0.058 0.433 0.021
root node is split on 9 9 0.678 +/- 0.057 0.558 0.022
the initial heart rate, 10 8 0.678 +/- 0.057 0.578 0.033
with patients with a 11** 7 0.663 +/- 0.057 0.603 0.044
heart rate of 104 or 12 5 0.762 +/- 0.056 0.692 0.076
less going to the left, 13 2 0.951 +/- 0.050 0.901 0.120
14 1 1.000 +/- 0.000 1.000 0.170
and those with a

Page 11
Introduction to CART
Roger J. Lewis, M.D., Ph.D.

higher heart rate going to the right.

As can be seen from the tree, the CD4
count is only important for patients
who are not tachycardic, as the CD4
count appears only in the left-hand
side of the tree. This is typical of
insights which can be obtained
through this type of analysis, namely,
the importance of a variable (e.g.,
CD4 count) can vary tremendously
based on the value of other variables.
This makes clinical sense--a patient
with markedly abnormal vital signs
probably has an emergent medical
condition regardless of the degree of
immunosuppression. In contrast, a
patient who is markedly
immunocompromised may have an
urgent or emergent medical condition
even with relatively normal vital signs
at triage.

Future Datasets
The purpose of a decision tree
is usually to allow the accurate
prediction of outcome for future
patients, based on the values of their
predictor variables. Similarly, the best
way to test a tree using an independent
dataset is to “drop” cases from a new
dataset through the tree in order to
determine the observed
misclassification rates and costs. The
CART software provides a command
(the “tree” command) which allows
the decision tree to be saved, so that it
can be used with a new set of data in
the future to predict outcome. This
allows the testing of the tree on a new
independent dataset.

Page 12
Introduction to CART
Roger J. Lewis, M.D., Ph.D.

Additional Topics
There are a number of
additional topics that, although of
practical importance to those using
CART analysis, are beyond the scope
of this lecture. These include the
choice and use of different splitting
rules and purity measures, the choice
of alternative prior probability
distributions, the interpretation of
information on surrogate and
competitive variables, methods used
to rate the importance of different
variables, and details of CART's
handling of missing variables.
Insights into these topics can best be
obtained by actual use of the software,
reference to the manual and online documentation, and comparison of results while varying user options.
The dataset and command files included in the distributed CD should allow the listener to begin using the
CART software and gain experience with its use.

Node 1
Yes Class = 1 No
PULSE <= 104.500

Node 2 Node 5
No Class = 1 Yes Yes Class = 0 No
CD4 <= 42.000 TEMP <= 101.650

Node 3 Node 6
Yes Class = 2 No Node -1 No Class = 1 Yes Node -7
TEMP <= 100.450 Class = 1 PULSE <= 137.500 Class = 0

Node 4
No Class = 2 Yes Node -4 Node -6 Node -5
PULSE <= Class = 1 Class = 0 Class = 1
99.500

Node -3 Node -2
Class = 1 Class = 2

Conclusions
Classification and Regression Tree (CART) analysis is a powerful technique with significant potential
and clinical utility. Nonetheless, a substantial investment in time and effort is required to use the
software, select the correct options, and interpret the results. Nonetheless, the use of CART has been
increasing and is likely to increase in the future, largely because of the substantial number of important
problems for which it is the best available solution.

Page 13
 Introduction to CART
Roger J. Lewis, M.D., Ph.D.

References

Primary Reference

1. Breiman L, Friedman JH, Olshen RA, Stone CJ. Classification and Regression Trees. Chapman &
Hall (Wadsworth, Inc.): New York, 1984.

Examples

2. Steadman HJ, Silver E, Monahan J, Apelbaum PS, Robbins PC, Mulvey EP, Grisso T, Roth LH,
Banks S. A classification tree approach to the development of actuarial violence risk assessment
tools. Law and Human Behavior 2000;24:83-100.
3. Hess KR, Abbruzzese MC, Lenzi R, Raber MN, Abbruzzese JL. Classification and regression tree
analysis of 1000 consecutive patients with unknown primary carcinoma. Clinical Cancer Research
1999;5:3403-3410.
4. Rainer TH, Lam PK, Wong EM, Cocks RA. Derivation of a prediction rule for post-traumatic acute
lung injury. Resuscitation 1999;42:187-196.
5. Dart RG, Kaplan B, Varaklis K. Predictive value of history and physical examination in patients with
suspected ectopic pregnancy. Annals of Emergency Medicine 1999;33:283-290.
6. Tsien CL, Fraser HS, Long WJ, Kennedy RL. Using classification tree and logistic regression
methods to diagnose myocardial infarction. Medinfo 1998;9:493-497.
7. Nelson LM, Bloch DA, Longstreth WT Jr., Shi H. Recursive partitioning for the identification of
disease risk subgroups: a case-control study of subarachnoid hemorrhage. Journal of Clinical
Epidemiology 1998;51:199-209.
8. Germanson TP, Lanzino G, Kongable GL, Torner JC,Kassell NJ. Risk classification after aneurysmal
subarachnoid hemorrhage. Surgical Neurology 1998;49:155-163.
9. Kastrati A, Schomig A, Elezi S, Schuhlen H, Dirschinger J, Hadamitzky M, Wehinger A, Hausleiter
J, Walter H, Neuman FJ. Predictive factors of restenosis after coronary stent placement. Journal of
the American College of Cardiology 1997;30:1428-1436.
10. Crichton NJ, Hinde JP, Marchini J. Models for diagnosing chest pain: Is CART helpful? Statistics
in Medicine 1997;16:717-727.
11. Hadzikadic M, Hakenewerth A, Bohren B, Norton J, Mehta B, Andrews C. Concept formation vs.
logistic regression: Predicting death in trauma patients. Artificial Intelligence in Medicine
1996;8:493-504.
12. Mair J, Smidt J, Lechleitner P, Dienstl F, Puschendorf B. A decision tree for the early diagnosis of
acute myocardial infarction in nontraumatic chest pain patients at hospital admission. Chest
1995;108:1502-1509.
13. Selker HP, Griffith JL, Patil S, Long WJ, D'Agostino RB. A comparison of performance of
mathematical predictive methods for medical diagnosis: Identifying acute cardiac ischemia among
emergency department patients. Journal of Investigative Medicine 1995;43:468-476.
14. Li D, German D, Lulla S, Thomas RG, Wilson SR. Prospective study of hospitalization for asthma.
A preliminary risk factor model. American Journal of Respiratory and Critical Care Medicine
1995;151:647-655.
15. Falconer JA, Naughton BJ, Dunlop DD, Roth EJ, Strasser DC, Sinacore JM. Predicting stroke
inpatient rehabilitation outcome using a classification tree approach. Archives of Physical Medic ine
and Rehabilitation 1994;75:619-625.
16. Hasford J, Ansari H, Lehmann K. CART and logistic regression analyses of risk factors for first dose
hypotension by an ACE-inhibitor. Therapie 1993;48:479-482.

Page 14

View publication stats