Journal of Pharmacokinetics and Pharmacodynamics, Vol. 33, No.

6, December 2006 (© 2006)

DOI: 10.1007/s10928-006-9033-1

Simulation of Correlated Continuous and Categorical

Variables using a Single Multivariate Distribution
Stacey J. Tannenbaum,1,5 Nicholas H. G. Holford,2,3 Howard Lee,2
Carl C. Peck,2 and Diane R. Mould4

Received 25 March 2006—Final 22 August 2006—Published online October 12, 2006

Clinical trial simulations make use of input/output models with covariate effects; the virtual
patient population generated for the simulation should therefore display physiologically rea-
sonable covariate distributions. Covariate distribution modeling is one method used to create
sets of covariate values (vectors) that characterize individual virtual patients, which should
be representative of real subjects participating in clinical trials. Covariates can be continuous
(e.g., body weight, age) or categorical (e.g., sex, race). A modeling method commonly used
for incorporating both continuous and categorical covariates, the Discrete method, requires the
patient population to be divided into subgroups for each unique combination of categorical
covariates, with separate multivariate functions for the continuous covariates in each subset.
However, when there are multiple categorical covariates this approach can result in subgroups
with very few representative patients, and thus, insufficient data to build a model that charac-
terizes these patient groups. To resolve this limitation, an application of a statistical method-
ology (Continuous method) was conceived to enable sampling of complete covariate vectors,
including both continuous and categorical covariates, from a single multivariate function. The
Discrete and Continuous methods were compared using both simulated and real data with
respect to their ability to generate virtual patient distributions that match a target popu-
lation. The simulated data sets consisted of one categorical and two correlated continuous
covariates. The proportion of patients in each subgroup, correlation between the continuous
covariates, and ratio of the means of the continuous covariates in the subgroups were varied.

1 Novartis Pharmaceuticals Corp., One Health Plaza 435/1125, East Hanover, NJ 07936,
USA.
2 Center for Drug Development Science, UC Washington Center, School of Pharmacy, Uni-
versity of California San Francisco, 1608 Rhode Islands Road, NW, Washington, DC
20036, USA.
3 Department of Pharmacology and Clinical Pharmacology, University of Auckland, 85
Park Rd, Private Bag 92019, Auckland, New Zealand.
4 Projections Research, Inc., 535 Springview Lane, Phoenixville, PA 19460, USA.
5 To whom correspondence should be addresed. E-mail: [email protected]

773

1567-567X/06/1200-0773/0 © 2006 Springer Science+Business Media, Inc.

774 Tannenbaum et al.

During evaluation, both methods accurately generated the summary statistics and proper pro-
portions of the target population. In general, the Continuous method performed as well as
the Discrete method, except when the subgroups, defined by categorical value, had markedly
different continuous covariate means, for which, in the authors’ experience, there are few clini-
cally relevant examples. The Continuous method allows analysis of the full population instead
of multiple subgroups, reducing the number of analyses that must be performed, and thereby
increasing efficiency. More importantly, analyzing a larger pool of data increases the precision
of the covariance estimates of the covariates, thus improving the accuracy of the description
of the covariate distribution in the simulated population.

KEY WORDS: covariate; covariate distribution modeling; continuous covariate; categorical

covariate; multivariate distribution; clinical trial simulation.

INTRODUCTION
Clinical trial simulation (CTS) can be a valuable tool to improve
drug development (1–3). By synthesizing the available knowledge about the
drug, patients, and clinical program (e.g., pharmacokinetics and pharma-
codynamics, disease progress, demographics) into a stochastic model, the
user can investigate, in silico, aspects of the clinical study plan (dosing reg-
imens, study designs, patient populations, formulations), allowing the clin-
ical team to make rational, informed decisions with regards to optimizing
the development plan of a new compound (4–8).
A clinical trial simulation model consists of three main components
(1): a clinical trial execution model, an input–output (IO) model, and a
covariate distribution model. The execution model describes aspects of the
study conduct such as compliance with dosing schedules, and subject drop-
outs. The IO model is a collection of models describing the disease progress
during the study period, and the pharmacokinetics and pharmacodynam-
ics of the drugs being tested. The covariate distribution model incorporates
patient-specific factors that may account for inter-individual differences in
observed pharmacokinetics and pharmacodynamics and contribute to vari-
ability in individual parameter values. Based on the established or hypoth-
esized impact of the covariates on the IO model, the simulated covariate
information is then used to predict IO model parameters for a virtual patient
with a particular combination of demographics and characteristics.
Covariate distribution modeling can be used to generate virtual patients
for clinical trial simulation (3, 9). Each patient is represented by a set of intrin-
sic or extrinsic factors (called a covariate vector) which collectively describe
the characteristics of the patient. Useful covariates typically include demo-
graphics (age, weight, sex, race), concomitant drug use (which may also include
abused drugs, tobacco and alcohol), and disease risk or health status biomar-
kers (e.g., blood pressure, cholesterol concentrations, creatinine clearance, liver
enzymes, disease severity). Note that patient covariates may be continuous
Simulation of Correlated Continuous and Categorical Variables 775

(such as age and weight) or categorical (such as sex, race, or smoking status).
These covariates are frequently correlated between individuals (e.g., women
are more likely to weigh less and have lower creatinine clearances than men).
Since the covariates are used to predict elements of the IO model that influence
a patient’s trial outcome, it is critical that the covariates associated with each
virtual patient be realistic and consistent with the projected patient popula-
tion. Therefore, some care should be given to the development of the covariate
distribution model.
There are a number of techniques that use the covariates of an exist-
ing patient population (e.g., a patient population with the same indica-
tion, or patient information from a previous study of the same drug)
to create new virtual patient populations for clinical trial simulation (9).
The simplest method is to sample complete patient covariate vectors from
observed values in the existing database (also called the empirical distribu-
tion), with or without replacement of that vector in subsequent sampling.
The benefit of sampling from an empirical distribution is that covariate
combinations are guaranteed to be realistic, as they are extracted directly
from real patient data. However, no new patient covariate vectors can be
created using this approach.
Rather than sampling complete vectors from single subjects, the
individual empirical distributions of each covariate can be used to create
vectors that do not exist in the empirical database. In this method, covari-
ates are sequentially sampled from their individual empirical distributions,
with each subsequent covariate chosen from a constrained set of values
based upon previously selected covariates (e.g., after randomly sampling
age from the observed age distribution, one would then randomly sam-
ple creatinine clearance from its observed distribution; however, the choice
of values would be limited to creatinine clearance measurements obtained
from patients with the age chosen in the first step). Such “conditional dis-
tributions” will preserve the correlation between the covariates. It should
be noted, however, that as each additional covariate is selected, subsequent
distributions become increasingly constrained, potentially limiting values
to a highly restricted subset; shuffling the selection order may partly alle-
viate this problem (9). In addition, the process of sequentially selecting
covariates can be computationally inefficient.
Random sampling of covariate vectors from a multivariate normal
distribution (MVND, Fig. 1) preserves the benefits of the previously
described covariate profile generation methods (generation of unique sub-
jects with realistic covariate vectors), while reducing their limitations (e.g.,
computational inefficiency and sampling from overly constrained distribu-
tions). A MVND is represented by two parameters: a vector of means
of the individual covariates, and matrix consisting of the variances of the
776 Tannenbaum et al.

Fig. 1. Illustration of multivariate normal distribution for two covariates (cov1 and cov2).
Covariate combinations that occur naturally in the target population have a high probability
of being selected by the covariate distribution model, whereas unrealistic or physiologically
impossible combinations are selected with lower frequency.

covariates along the main diagonal, and the covariances between each pair
of covariates in the other matrix positions (10).
Two important assumptions fixed in the definition of an MVND must
be considered. First, all covariates in the MVND are assumed to follow
the same known distribution (e.g., normal or log-normal). Second, while
covariance defines the basic association between two covariates, it is not
sufficient to fully define the shape of the relationship; sampling from a
MVND will always result in linearly related simulated covariates. Thus,
regardless of the observed distributions of the covariates, and the shapes
of the relationships between them in the empirical distribution, covari-
ates sampled from an MVND based upon this data will be normally (or
log-normally) distributed and linearly related; the simulated results may
therefore only approximate the original target population. It should be
noted, however, that most common covariates, such as age and weight, are
generally normally or log-normally distributed; in addition, within normal
ranges of covariate values, it is usual to see a linear relationship between
common covariates. Therefore, because the MVND defines the individ-
ual covariate distributions as well as maintains the systematic relationship
between the covariates, the generated covariate vectors should be physio-
logically realistic.
Software packages such as NONMEM (11) and Pharsight
Trial
Simulator (Version 2.1.2, Pharsight Inc, Mountain View, CA) allow sam-
ples to be obtained from multivariate normal distributions (MVND), but
this can only be accomplished when all covariates are continuous. Because
categorical covariates are not continuous they have not previously been
considered for inclusion in MVNDs. However, the method for covari-
ate distribution modeling that will be introduced creates a single MVND
Simulation of Correlated Continuous and Categorical Variables 777

which includes both continuous and categorical covariates. This is accom-

plished by deliberately ignoring the categorical nature of a covariate and
treating it as if it arose from a continuous distribution. Evaluation of the
properties of this method (the Continuous method) was undertaken by
comparing it with a standard method (the Discrete method) with respect
to their abilities to generate virtual patient distributions that match a real
or simulated target population.

METHODS
Discrete Method
A commonly used method for dealing with both continuous and cat-
egorical covariates is to use a separate MVND for each unique combina-
tion of categorical covariate values. This will henceforth be designated the
Discrete method. For example, if sex and smoking are the two categori-
cal covariates, the population is divided into four groups (female smokers,
female nonsmokers, male smokers, and male nonsmokers). Subgroup spe-
cific MVNDs are then created from which continuous covariates (e.g. age,
weight) are sampled. Although the Discrete method is frequently used,
there are significant limitations, which arise from subdividing the patient
population.
First, the Discrete method may be impractical to implement when
there are multiple categorical covariates (e.g. sex ×2, smoking ×3, race
×4, disease status ×5 would lead to 120 separate MVNDs). Even if it
were feasible to simulate this many MVNDs, estimation of their param-
eters may be impossible because of limited empirical observations of the
continuous covariates in each of the categorical subgroups. When there are
too few patients in a subgroup, there may be insufficient data to create
a reliable MVND; specifically, if there less than N + 1 subjects in a sub-
group (where N is the number of covariates in the MVND), the variance–
covariance matrix that is generated will be singular. A worst case scenario
is a subgroup in which there are no patients in the empirical distribution,
yet patients with this combination of categorical covariates could poten-
tially be enrolled in a future clinical trial. Because the relationship between
the continuous covariates in this subgroup is unknown, it is impossible to
determine if the simulated patient covariate vectors are appropriate for this
patient group.
Although there may be no data about the association between co-
variates (continuous or categorical) for a specific subgroup, it seems rea-
sonable to assume that the variance structure may be similar to those
778 Tannenbaum et al.

observed in subgroups occurring more frequently in the empirical dis-

tribution. Thus, we propose an approximate method to capture these
correlations in order to simulate correlated categorical and continuous co-
variates. The method can be applied using estimates from a sparse (for
some combinations) empirical distribution.

Continuous Method
In the Continuous method the parameters of a single MVND are esti-
mated by treating all categorical covariates as if they are continuous val-
ues, a procedure seen commonly in statistical simulation (12–15). In order
to constrain all biological covariates to be positive, we typically assume
a log-normal multivariate distribution. Thus, the MVND variance–covari-
ance matrix is defined in terms of the logarithms of the covariate values.
Likewise, categorical values must all be coded to possess positive values.
Complete patient covariate vectors (both continuous and categorical cova-
riates) are then sampled from a single MVND; because the sampled val-
ues are logarithmic, each component of the vector is then exponentiated to
obtain the true covariate values. Note that because categorical covariates
are sampled from a continuous MVND, sampled values for the virtual
patients for these covariates will therefore be nondiscrete (e.g., if nonsmok-
ers = 1 and smokers = 2 in the empirical distribution, a value such as 1.3
would be a possible value for smoking status in vectors sampled from a
MVND). These continuous values must then be mapped to discrete cate-
gorical values, based on a continuous critical value (CrV).
The CrV is determined from the inverse of the lognormal cumulative
distribution with a given mean, standard deviation, and cumulative prob-
ability (16), according to the following equation:

CrV(µ, σ, Pi ) = eµ+σ ·NORMINV(Pi ) . (1)

For a categorical covariate X with k discrete values, µ = mean(ln(X )),

σ = SD(ln(X )), Pi is the proportion of subjects in the empirical distri-
bution with categorical value X i (i ≤ k), and NORMINV is the inverse
of the standard normal distribution. Figure 2 illustrates the CrV calcula-
tion for an example of a categorical covariate, smoking status, with three
levels. The data set contained 18.6% nonsmokers, 49.3% former smokers,
and 32.1% smokers, coded 1, 2, and 3, respectively. The corresponding
cumulative probabilities (Pi ) were 0.186, 0.679 (0.186 + 0.493), and 1. The
original smoking status values were log-transformed and the mean and
standard deviation of the transformed data were determined (µ = 0.694,
σ = 0.378). These values were used (Eq. 1) to determine two CrV val-
ues (1.43 and 2.39). Thus, if the continuous value for smoking status was
Simulation of Correlated Continuous and Categorical Variables 779

Nonsmokers (18.6%)
Former smokers (49.3%)
Current smokers (32.1%)

0 1 2 3 4 5 6
Continuous value for smoking status
Fig. 2. Illustration of the calculation of discrete values for smoking status with three possible
values (nonsmoker, former smoker, current smoker). The mean and standard deviation define
the log-normal probability distribution of smoking status in the empirical distribution, as if it
was a continuous covariate. The histogram is derived from an empirical distribution used to
estimate the continuous distribution parameters. The cumulative probability (P) corresponds
to the area under the probability distribution curve (solid line). The CrVs are indicated by
arrows on the x-axis.

less than or equal to 1.43, the subject would be designated a nonsmoker;

a value between the two CrVs would designate a former smoker, and if
greater than 2.39, the subject would be designated a smoker.

Qualification of Methods
The performances of the Discrete and Continuous methods were eval-
uated based upon their abilities to reproduce the summary statistics of
target population covariate distributions, using both real and simulated
populations. All simulations were performed using Trial Simulator, and
statistics were determined using SPLUS
(Version 6.0, Insightful Corp.,
Seattle, WA).
For the Continuous method, the parameters of a single MVND were
estimated using all covariates (continuous and categorical) from the real
780 Tannenbaum et al.

or simulated data. The following steps were carried out in S-PLUS. First,
summary statistics (geometric mean, minimum, and maximum) were com-
puted for each covariate. All covariates were log-transformed (for categor-
ical covariates with a value of zero in the empirical distribution, 1 was
added to all values prior to log-transformation). The variance–covariance
matrix of the transformed values was then determined. The summary sta-
tistics and variance–covariance matrix were entered into Trial Simulator
to define the MVND. All covariates were classified as continuous, regard-
less of their original type (categorical, continuous) in the empirical distri-
bution. After the covariate vectors were sampled from the MVND, each
element was exponentiated. The “continuous” categorical covariate values
were then discretized based upon the appropriate CrV.
For the Discrete Method, the values from the real or simulated data
were subset into groups corresponding to each unique combination of cat-
egorical covariates. In Trial Simulator, each covariate was classified as cat-
egorical or continuous according to its type in the empirical distribution.
The method outlined in the previous paragraph was then applied for the
continuous covariates in each subgroup.
1000 subjects were simulated using both the Discrete and Continuous
methods. The population summary statistics and distributions of contin-
uous covariates and proportions of categorical covariate values generated
by both methods were compared to those of the corresponding “observed”
(real or simulated) data. In addition, a test of the method’s ability to
preserve the correlation coefficients between the covariates was examined.
A variance–covariance matrix was created from the simulated population
data, and compared to the variance–covariance matrix obtained from the
original data; the percent difference between the values in the same posi-
tions in the two matrices was calculated.
For each original data set (real, simulated), the Discrete and Contin-
uous methods were replicated 10 times. With 1000 subjects simulated per
replicate, good marginal statistics for the covariates, including 95% confi-
dence intervals, could be obtained. Since the results were fairly invariant
between the replicates (based on a small standard error of the mean), 10
replicates were judged to be sufficient to obtain precise estimates of the
covariate summaries.
The Continuous and Discrete methods were applied to both real data
and 27 simulated covariate data sets, as follows:

Empirical distribution of covariates (real data example)

The real data example was based on 467 subjects with seven continu-
ous covariates (age, weight, body mass index, diastolic and systolic blood
pressure, total cholesterol, fasting blood glucose) and three categorical
Simulation of Correlated Continuous and Categorical Variables 781

Table I. Parameter values for the simulated target population covar-

iate data sets

Parameter CONT1 CONT2

Mean (CAT = 1) 10,50,90* 90

Mean (CAT = 2) 100 100
CV (%) 30 30
Minimum 0 0
Maximum 1000 1000

Each subgroup (CAT = 1 and CAT = 2) was simulated with a sepa-

rate log-normal distribution based upon the mean, coefficient of vari-
ation, and range. All parameter values were fixed except for the mean
of CONT1 for subgroup CAT=1, which varied dependent upon the
mode ratio, MR* (see Table II).

covariates (sex, smoking status, and diagnosis), with 2, 3, and 4 catego-

ries, respectively. Subdividing the population by unique combinations of
categorical covariates created 2 · 3 · 4 = 24 subgroups, with between 1 and
83 subjects per subgroup (median = 9). In this example, due to the large
number of subgroups and the resulting small number of subjects in many
of these subgroups, the Discrete method proved to be impractical.

Simulated distributions of covariates (simulated data example)

The simulated covariate distribution consisted of one categorical
covariate with 2 levels (CAT = 1, CAT = 2), and two correlated continu-
ous covariates (CONT1 and CONT2). Each subpopulation (CAT = 1 and
CAT = 2) was simulated with a separate log-normal distribution for each
continuous covariate, according to the summary statistics in Table I. Both
continuous covariates had a bimodal distribution (the means of the two
subgroups differed). The summary statistics for CONT2 were fixed; for
CONT1, all parameters were fixed except for the mean of the CAT = 1
subgroup.
To test different simulation scenarios, three factors were varied: the
percentage of patients in the CAT = 1 subgroup, the correlation between
CONT1 and CONT2, and the extent of “bimodality” (overlap) of the dis-
tributions of CONT1 for the two subgroups, designated as the mode ratio,
MR. The lower the value of MR, the more distinct the modes of the two
subgroups were and the more the overall population distribution appeared
bimodal. Therefore a low MR value (0.1) resulted in a distribution that
appeared to be strongly bimodal whereas a high MR value such as 0.9
indicated a large degree of overlap between the two subgroups, leading to
the appearance that the population is in fact unimodal. Table II lists the
three values chosen for each factor.
782 Tannenbaum et al.

Table II. Variables defining the simulation scenarios (n = 27) for the
simulated target population covariate data sets

% (CAT = 1) Corr MR

10 0 0.1
25 0.45 0.5
50 0.9 0.9

The factors varied were the percentage of patients in subgroup

CAT = 1, the correlation between CONT1 and CONT2, and the
mode ratio, MR, which, when multiplied by 100, defined the mean
of CONT1 for subgroup CAT = 1, and indicated the degree of over-
lap between the two subgroups. The mean of CONT1 for subgroup
CAT = 2 was set to 100.

The MVND parameters for each of the possible scenarios were used
to simulate 27 covariate data sets with Trial Simulator.

Categorical Covariate Coding

An analysis was completed to assess whether the order in which a cat-
egorical covariate was coded had a significant impact on the results. For
the real data example, the original coding for smoking status was set at 1
for nonsmokers, 2 for former smokers, and 3 for current smokers (1/2/3).
Two alternate data sets were created with the codes shuffled (2/3/1 and
3/1/2). The Continuous method was performed for the two alternate data
sets and the summary statistics compared to the original data as well as
to the simulated data with the initial coding.
To determine if the results of the methodology were invariant to the
actual values of the codes (for example, 1/2/3 vs. 1/2/10 vs. 1/2/50), a sim-
ilar analysis to the previous was performed.

RESULTS
Empirical Distribution of Covariates (Real Data Example)
Table III and Figs. 3 and 4 display the covariate summary statistics of
the real data set and those generated using the Continuous Method. The
mean, standard deviation, and range of the continuous covariates in the
original population are maintained in the simulated population. In addi-
tion, the proportion of each value of the categorical covariates in the orig-
inal population is maintained in the simulated population, showing that
the mapping from continuous to discrete value calculation is appropriate.
Simulation of Correlated Continuous and Categorical Variables 783

Table III. Real Data Example: summary statistics of the categorical covariates of target pop-
ulation (n = 467) and the covariate data generated using the Discrete and Continuous meth-
ods (n = 1000 × 10 replicates)

Obs. Continuous. Discrete.

(%) % (SE) % (SE)

Sex 1 80.3 79.9 (0.94) 80.5 (1.51)

2 19.7 20.1 (0.94) 19.5 (1.41)
Smoking Status 1 18.6 19.0 (1.45) 18.9 (0.78)
2 49.3 50.2 (2.13) 48.6 (1.52)
3 32.1 30.8 (1.53) 32.5 (0.96)
Diagnosis 1 40.0 40.6 (1.43) 40.0 (0.24)
2 38.0 37.7 (1.50) 38.1 (1.41)
3 6.0 5.88 (0.63) 6.18 (0.45)
4 16.0 15.8 (0.71) 15.7 (1.38)

SEX SMOKING STATUS

1.0 1.0
Proportion

Proportion

0.5 0.5

0.0 0.0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Replicate Replicate

DIAGNOSIS

1.0
Proportion

0.5

0.0
0 1 2 3 4 5 6 7 8 9 10
Replicate

Fig. 3. Real Data Example: Proportion of patients in each level for the three categorical
covariates. The leftmost bar in each plot represents the target population data (n = 467).
Each additional bar represents one replicate of 1000 patients, generated using the Continuous
method. The varying bar colors represent the proportion of subjects in each category.

The standard errors of the mean (continuous) or proportion (categorical)

for each covariate (10 replicates) demonstrate high precision of the method
with negligible bias.
784 Tannenbaum et al.

AGE BODY WEIGHT BODY MASS INDEX

DIASTOLIC BP SYSTOLIC BP

CHOLESTEROL GLUCOSE

Fig. 4. Real Data Example: Box and whisker plot showing the distribution of values for the
continuous covariates. The leftmost box in each plot represents the target population data
(n = 467). Each additional box represents one replicate of 1000 patients, generated using the
Continuous method. The boxes contain the 25th to 75th percentiles, the horizontal bar repre-
sents the median value, the whiskers represent the 10th to 90th percentiles, and the lines out-
side the whiskers represent outliers.

Due to the underlying assumptions behind the MVND, the simulated

data will have a linear relationship between the covariates, which may not
necessarily reflect the true shape of the relationship in the observed data.
However, a test of the method’s ability to preserve the correlation coeffi-
cients was examined by comparing the variance–covariance matrix for the
observed and simulated population covariates. The maximum percent error
in the covariance terms for all scenarios was 10%, indicating that the Con-
tinuous method preserved the correlation between the covariates.
Table IV displays the covariate summary statistics of the real data
set and those generated using the Discrete Method. The accuracy, preci-
sion, and bias are similar to those of the Continuous method. For the real
data, however, the statistics from the Discrete method using the continu-
ous covariates are based on results from only 16 of the 24 subsets. The
MVNDs for the remaining 8 subsets, containing between 1 and 7 sub-
Simulation of Correlated Continuous and Categorical Variables 785

Table IV. Real Data Example: summary statistics of the continuous covariates in the tar-
get population (n = 467) and the covariate data generated using the Discrete and Continuous
methods (n = 1000 × 10 replicates)

Observed Continuous method Discrete method

Mean ± SD Mean (SE) ± SD (SE) Mean (SE) ± SD (SE)

Age 68.7 ± 8.2 69.7 (0.23) ± 7.20 (0.15) 69.9 (0.21) ± 4.98 (0.14)
Weight 72.6 ± 12.8 74.0 (0.27) ± 10.8 (0.19) 70.5 (0.44) ± 7.19 (0.34)
BMI 25.8 ± 3.5 26.1 (0.04) ± 2.92 (0.07) 25.8 (0.08) ± 2.41 (0.09)
Cholesterol 205.5 ± 44.0 214.0 (1.50) ± 43.7 (1.32) 212.2 (2.29) ± 30.95 (1.19)
Diastolic BP* 77.9 ± 11.0 79.4 (0.36) ± 10.2 (0.16) 79.4 (0.36) ± 10.2 (0.16)
Systolic BP* 146.1 ± 18.4 148.1 (0.36) ± 17.1 (0.42) 147.0 (1.06) ± 12.5 (0.45)
Glucose 5.97 ± 1.91 6.70 (0.07) ± 2.36 (0.09) 6.48 (0.11) ± 1.39 (0.07)

* BP: Blood Pressure

jects, failed to simulate reasonable values for the continuous covariates.

Additional analyses were performed to determine the causality of this fail-
ure; it was found that if there are data from less than N + 1 subjects in a
subgroup (where N is the number of covariates in the MVND), the vari-
ance–covariance matrix that is generated will be singular. This is clearly a
potential disadvantage of the Discrete method.

Simulated Distributions of Covariates (Simulated Data Example)

Figure 5 displays the proportion of patients in the (CAT = 1) sub-
group, and shows that both the Continuous and Discrete methods gener-
ate the expected proportion of patients in each subgroup.
As shown in Fig. 6a, for the situations in which MR equals 0.1, indi-
cating little overlap between the subgroups, the Continuous method fails
to describe the distribution of CONT1 in the two subgroups. However, as
MR increases, the subgroup distributions begin to overlap, masking the
bimodal characteristics of the distribution, and the Continuous method
begins to describe the original distributions more accurately. When the
overlap is substantial (MR = 0.9, Fig. 6c), the distribution of CONT1 gen-
erated by the Continuous method matches the simulated covariate data
distribution almost exactly. For all values of MR, the Discrete method
adequately distinguishes the distributions of the individual subgroups for
CONT1; this is expected since each subgroup is simulated separately.
The percent prediction errors (%PE) in the summary statistics of
CONT1 for the Continuous method are shown in Fig. 7. %PE is cal-
culated from 100· (predicted-true)/true, where true is the true mean of
CONT1, and “predicted” is the mean of CONT1 in the covariate data
786 Tannenbaum et al.

Fig. 5. Simulated Data: Bar chart showing the percentage of patients in the subgroup
(CAT = 1) for the target population covariate data and the covariate data generated by
the Continuous and Discrete methods. There should be 10, 25, and 50% in the (CAT = 1)
subgroup, respectively, for each set of 9 scenarios.

simulated by the Continuous method. For the whole population, the Con-
tinuous method reliably simulates covariates with mean and coefficient of
variation close to the true values. The %PE is negligible, and is relatively
independent of MR and number of patients in the subgroups, with only
a slight negative %PE for both the mean and CV at a MR value of
0.1. However, for the individual subgroup summary statistics, the %PE is
highly dependent upon both MR and the percentage of patients in that
subgroup. For MR = 0.1, the Continuous method results overestimate the
mean and CV of CAT = 1 (as shown by a large positive %PE); as the per-
centage of patients in this subgroup increases, however, the error decreases.
As MR increases, the errors approach zero for both mean and CV in the
subpopulations. For the Discrete method, there are negligible errors in the
mean and SD for the subgroups and for the whole population, which are
independent of the values of MR or the percentage of patients in each
subgroup (results not shown).
Figure 8 shows the correlation between CONT1 and CONT2 for
the true simulated covariate data, and for the Continuous and Discrete
method results. Comparing the plots of the simulated covariate data and
the Continuous method results indicates that for a MR value of 0.1, the
continuous method fails to capture the relationship between CONT1 and
CONT2, but adequately captures the correlation for larger ratios. For all
Simulation of Correlated Continuous and Categorical Variables 787

Fig. 6. (a) Simulated Data: Population distribution of CONT1 for MR = 0.1. The target pop-
ulation covariate data (gray bars) is overlaid with the Continuous method results (top) and
Discrete method results (bottom), shown as transparent bars. Only the scenarios for correla-
tion = 0 between CONT1 and CONT2 are shown, but the plots for correlations of 0.45 and
0.9 look similar. (b) Simulated Data: Population distribution of CONT1 for MR = 0.5. (c)
Simulated Data: Population distribution of CONT1 for MR = 0.9.
788 Tannenbaum et al.

Fig. 6. Continued.

ratios, the Discrete method accurately captures the correlation between

CONT1 and CONT2.
Figure 9 shows the %PE for the correlation between CONT1 and
CONT2. For MR = 0.1, the Continuous method underestimates the cor-
relation between CONT1 and CONT2 (large negative %PE) for both the
subpopulations and the whole population. As MR increases, the absolute
errors decrease in the subpopulations, and are negligible in the whole pop-
ulation. There does not appear to be a relationship between percent of
patients in each subgroup and the correlation. For the Discrete method,
there are negligible errors in correlation for the subgroups or for the whole
population for all ratios and proportion of patients in each subgroup
(results not shown).

Categorical Covariate Coding

Shuffling the order of the coding by definition altered Pi (see Eq. 1), but
also transformed the mean and standard deviation for smoking status, and
thus the CrVs. Paired with the new variance–covariance matrix (the row and
column for smoking status was altered), the new parameters for the MVND
compensated for the shuffled coding. Therefore, the order of categorical
Simulation of Correlated Continuous and Categorical Variables 789

ALL CAT=1 CAT=2

150
CAT=1: 10%
CAT=1: 25%
% PE (mean)

110 CAT=1: 50%

70

30

-10

-50
150

110
% PE (CV)

70

30

-10

-50
0.1 0.5 0.9 0.1 0.5 0.9 0.1 0.5 0.9

Fig. 7. Simulated Data: Percent error in mean (top row) and CV (bottom row) of CONT1
for the values generated by the Continuous method, as a function of mode ratio and %
(CAT = 1).

ALL CAT=1 CAT=2

35
%PE (Corr)

-35
CAT=1:10%
CAT=1:25%
CAT=1:50%
-70
0.1 0.5 0.9 0.1 0.5 0.9 0.1 0.5 0.9

Ri
Fig. 8. Simulated Data: Correlation between CONT1 and CONT2 for CAT = 1 (light gray)
and CAT = 2 (dark gray). Only the scenarios with 50% of patients in each subgroup are
shown, but the plots for 10% and 25% in the CAT = 1 subgroup look similar.
790 Tannenbaum et al.

Fig. 9. Simulated Data: Percent error in correlation between CONT1 and CONT2 for the
values generated by the Continuous method, as a function of mode ratio and % (CAT = 1).

covariate coding had no effect upon the simulated results. The summary
statistics of the continuous covariates, and the proportion of each value of
the categorical covariates in the original population, including smoking sta-
tus, were nearly identical to that of the original observed population.
While the value of the codes (1/2/3 vs. 1/2/50) did impact the simu-
lated results when the third value became large, this analysis was really not
necessary; should codes such as the latter appear in a data set, the third
value could be simply recoded (i.e., all values of 50 changed to 3) for the
creation of the MVND, and then transformed back (to 50) in the simu-
lated data set.

DISCUSSION
As demonstrated by the real data example, both the Continuous
method and Discrete methods generate accurate summary statistics for
the covariates of the target population. The mean, standard deviation,
and range of the continuous covariates in the target population, and the
Simulation of Correlated Continuous and Categorical Variables 791

12 17 22 27 32 37 40 60 80 100 0 5 10 15 20
90
80
AGE 70
60
50
37
32
27
BMI
22
17
12
350
300
250
CHOL 200
150
100

100

80
DBP
60

40
210

160
SBP
110

60
20
15
10 DIAB
5
0
100

80
WT
60

40
50 60 70 80 90 100 150 200 250 300 350 60 110 160 210 40 60 80 100

Fig. 10. Scatterplot matrix for observed continuous covariates in real data set. Lines repre-
sent Loess smooths of the individual plots, and indicate that the relationships between the
covariates are relatively linear; thus, it is appropriate to enter all covariates into the MVND.

proportion of patients with each of the categorical covariate values, are

maintained in the population simulated using either method. In addition,
the standard error of the 10 replicates shows high precision of both meth-
ods, while the prediction error indicates negligible bias.
However, despite the fact that the Discrete method appears to gener-
ate the proper values for the target population summary statistics, these
results are misleading because these statistics were calculated from only 2/3
(16 of 24) of the unique combinations of categorical covariate values. The
remaining subsets contained fewer than 8 subjects (one plus the number of
continuous covariates); the amount of data in each subset was thus inad-
equate to obtain a nonsingular variance–covariance matrix between the
covariates. As the number of patients in each subgroup increases, the
covariate variance–covariance matrix derived from these data becomes
more precisely estimated. Consequently the summary statistics of the sim-
ulated data more closely match the original statistics, indicating that the
model is representative of these patient subsets in the clinical trial. By
not dividing the patients into subsets, the Continuous method ensures that
there will be sufficient data to generate the MVND; in addition, since each
792 Tannenbaum et al.

subset requires calculation of summary statistics and creation of a separate

MVND, the number of analyses that must be performed is reduced.
In the simulated covariate data case, both the Continuous and
Discrete method results match the summary statistics of the whole popula-
tion, with low error in the means and CVs, as well as the expected propor-
tion of patients in each subgroup. However, whereas the Discrete method
adequately recreates the distribution of the CONT1 values for all values of
MR, the Continuous method does not. Because the Continuous method
assumes a unimodal distribution for the covariate in the whole population
but in fact a bimodal distribution for a covariate exists, the Continuous
method is not able to recreate this distribution when MR is low. However,
as the subgroups overlap (i.e., the value of MR increases), the bimodal
characteristics of the distribution of CONT1 become obscured, and the
overall population distribution appears unimodal. Our example suggests
that when MR is greater than or equal to 0.5, the performance of the
Continuous method is adequate.
The Continuous method has limited utility for populations when the
subgroup means are very different; however, to our knowledge, there are
few clinically relevant examples in which such a low value of MR might
be seen. Two possible examples are testosterone/estrogen concentration in
males versus females (17), or carboxyhemoglobin concentration in smokers
versus nonsmokers (18). For most commonly used covariates, such as weight
and age, the means of these covariates are usually quite similar between
subjects with different sexes, races, or smoking status, with an MR value of
perhaps 0.8 or higher. However, even when MR is low between subgroups,
the Continuous method may still be utilized when the Discrete method is
challenged by inadequate numbers of subjects in subgroups in empirically
derived datasets. Rather than completely subdividing the population into
each unique combination of categorical covariate value, the subgroups with
a low value of MR may be separated out and the Continuous method could
then be applied to describe the remaining covariates.
Because a MVND simulates covariates whose pairs are linearly
related, the results may not represent the true shape of the covariate rela-
tionships in the underlying empirical distribution. It is therefore recom-
mended that a scatterplot matrix be created for the covariates, as shown
in Fig. 10 for the real data example. This will allow a visual examination
of the relationship between the covariates, and the opportunity to identify
any covariate that shows a significantly nonlinear association with another.
For example, if covA and covB are related nonlinearly, there are meth-
ods that can be carried out prior to creating the MVND to correct this.
First, one might attempt to transform the covariates in some way to lin-
earize their relationship (e.g., for a relationship represented by a power
Simulation of Correlated Continuous and Categorical Variables 793

model, the log transformed covariates are linearly correlated). Alternately,

the relationship between covA and covB may be modeled with both fixed
and random effects to obtain a mathematical expression; covB could then
be excluded when estimating the parameters for the MVND. After sam-
pling values for the covariates from the MVND, for each value of covA
an estimate of covB can be obtained from the quantitative relationship
between the two covariates. Similarly, this method could be used to recon-
cile two covariates that are linearly dependent, which may cause the vari-
ance–covariance matrix to be singular if both are left in the analysis.
The Continuous method can be valuable when the original covar-
iate data from which a MVND is derived does not exactly match the
desired target patient population for the clinical trial simulation. Although
the individual covariates themselves may be the same, it may be desir-
able to simulate virtual patients in a different age group, or perhaps to
explore the outcome of the model given a larger percentage of smok-
ers than in the original data set. To do this, one simply has to adjust
the inclusion–exclusion criteria for the simulation study without changing
the MVND. For example, if it was desired to have 50% males and 50%
females in a 100 subject study, the criteria would be set such that once
50 males were selected, all subsequent covariate vectors from male sub-
jects would be excluded. Similarly, if older patients were desired, covari-
ate vectors with patients below a certain age would be excluded. Once all
subjects are selected, the demographics of this final population should still
match that of the original patient population, due to the fact that the sub-
jects were selected from the MVND created from these data. It is assumed
that the MVND created from the original population represents the inher-
ent interrelationships between the selected covariates for a typical subject
population. These relationships may be borrowed from other similar sub-
ject populations, even if the overall demographics (mean age, percentage
of smokers, etc.) were different.
We hope in the future to investigate this method for more elaborate
covariate distribution models (for example, time-varying covariates).
Compared to the Discrete method, the Continuous method has a
number of benefits that result from analyzing the whole population instead
of small subsets. Including a large amount of data in the creation of the
variance–covariance matrix enhances its stability and, as a consequence,
the reliability of the generated covariate combinations. In addition, by
allowing all covariates to be described by a single MVND (rather than
one for each unique combination of categorical covariates), the number
of analyses that must be performed is reduced, increasing efficiency. With
the exception of the rare instance of a low mode ratio between continu-
ous distributions, the Continuous method appears to efficiently generate
794 Tannenbaum et al.

unbiased, precise covariates for the purposes of simulating virtual patient

covariate vectors in a clinical trial simulation.

REFERENCES
1. N. H. G. Holford, M. Hale, H. C. Ko, J.-L. Steimer, and C. C. Peck (eds.). P.
Bonate, W. R. Gillespie, T. Ludden, D. B. Rubin, L. B. Sheiner, and D. Stanski
(contributors). Simulation in Drug Development: Good Practices. http://cdds.georgetown.
edu/research/sddgp723.html
2. P. L. Bonate. Clinical trial simulation in drug development. Pharm. Res. 17:252–256
(2000).
3. N. H. G. Holford, J. Monteleone, H. Kimko, and C. Peck. Simulation of clinical trials.
Annu. Rev. Pharmacol. Toxicol. 40:209–234 (2000).
4. S. Chabaud, P. Girard, P. Nony, and J. P. Boissel. Clinical trial simulation using ther-
apeutic effect modeling: Application to ivabradine efficacy in patients with angina pec-
toris. J. Pharmacokinet. Pharmacodyn. 29(4):339–363 (2002).
5. H. J. M. Lemmens, D. R. Wada, C. Munera, A. Eltahtawy, and D. R. Stanski.
Enriched analgesic efficacy studies: An assessment by clinical trial simulation. Contemp.
Clin. Trials 27(2):165–173 (2006).
6. C. Veyrat-Follet, R. Bruno, and R. Olivares. Clinical trial simulation of docetaxel in
patients with cancer as a tool for dosage optimization. Clin. Pharmacol. Ther. 68:677–
678 (2000).
7. H. Kastrissios, S. Rohatagi, J. Moberly, K. Truitt, Y. Gao, D. R. Wada, M. Takahashi,
K. Kawabata, and D. Salazar. Development of a predictive pharmacokinetic model for
a novel cyclooxygenase-2 inhibitor. J. Clin. Pharmacol. 46(5):537–548 (2006).
8. K. G. Kowalski and M. M. Hutmacher. Design evaluation for a population pharmaco-
kinetic study using clinical trial simulation: A case study. Stat. Med. 20:75–91 (2001).
9. D. R. Mould. Defining covariate distribution models for clinical trial simulation. In H.
C. Kimko and S. B. Duffull (eds.), Simulation for Designing Clinical Trials: A Phar-
macokinetic– Pharmacodynamic Modeling Perspective, Marcel Dekker, New York, 2003,
pp. 31–53.
10. M. Evans, N. Hastings, and B. Peacock. Statistical Distributions, John Wiley & Sons,
Inc., New York, 1993, pp. 102–105.
11. S. L. Beal and L. B. Sheiner (eds.). NONMEM Users Guides, Icon Development Solu-
tions, Ellicott City, MD (1989–98).
12. B. Schmeiser. Advanced input modeling for simulation experimentation. In P. A. Far-
rington, H. B. Nembhard, D. T. Sturrock, and G. W. Evans (eds.), Proceedings of the
1999 Winter Simulation Conference, 1999, pp. 110–115.
13. M. Kaut and S. W. Wallace. Evaluation of scenario-generation methods for stochastic
programming (2003). http://citeseer.ist.psu.edu/cache/papers/cs/29430/http:zSzzSzwww.iot.
ntnu.nozSz mkautzSzCV and studyzSzSG evaluation.pdf/kaut03evaluation.pdf
14. S. Ghosh and S. G. Henderson. Chessboard distributions and random vectors with
specified marginals and covariance matrix. Working paper, Department of Industrial
and Operations Engineering, University of Michigan, Ann Arbor (2000).
15. M. C. Cario and B. L. Nelson. Modeling and generating random vectors with arbi-
trary marginal distributions and correlation matrix. Technical Report, Department of
Industrial Engineering and 35 Management Sciences, Northwestern University, Evans-
ton, Illinois (1997).
16. L. Lapin. Probability and Statistics for Modern Engineering (2nd edn), PWS-KENT
Publishing Company, Boston, 1983, pp. 215–217.
17. R. H. Williams and P. R. Larson (eds.), Williams Textbook of Endocrinology (10th
edn), W.B. Saunders Co., Illinois, 2002, pp. 622, 726.
18. P. McDonough and R. Moffatt. Smoking-induced elevations in blood carboxyhaemo-
globin levels. Effect on maximal oxygen uptake. Sports Med. 27:275–283 (1999).