You are on page 1of 10

STATISTICS IN MEDICINE

Statist. Med. (in press)


Published online in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/sim.2108

Application of hidden Markov models to multiple sclerosis


lesion count data
Rachel MacKay Altman1;; and A. John Petkau2
1 Department

of Statistics and Actuarial Science; Simon Fraser University; Canada


of Statistics; University of British Columbia; U.S.A.

2 Department

SUMMARY
This paper is motivated by the work of Albert et al. who consider lesion count data observed on
multiple sclerosis patients, and develop models for each patients data individually. From a medical
perspective, adequate models for such data are important both for describing the behaviour of lesions
over time, and for designing ecient clinical trials. In this paper, we discuss some issues surrounding
the hidden Markov model proposed by these authors. We describe an ecient estimation method and
propose some extensions to the original model. Our examples illustrate the need for models which
describe all patients data simultaneously, while allowing for inter-patient heterogeneity. Copyright ?
2005 John Wiley & Sons, Ltd.
KEY WORDS:

Hidden Markov model; multiple sclerosis; time series; count data

1. INTRODUCTION
Multiple sclerosis (MS) is a debilitating disease of the central nervous system. Patients with
this disease may have problems with vision, co-ordination, sensation, gait, endurance, and
bowel, bladder, cognitive and sexual functions. It is now believed that such symptoms are
related to the development of lesions (areas of demyelination) in the brain and spinal cord.
Lesions may persist indenitely, or may disappear temporarily only to reappear at a later time,
or may disappear altogether. Relapsing-remitting MS is a particular type of this disease where
symptoms tend to worsen and then improve in alternating periods of relapse and remission.
It has been shown that relapse rate is positively associated with numbers of T2 lesions [1].
Magnetic resonance imaging (MRI) is one method for detecting and measuring MS lesions.
The development of a model for MS=MRI lesion count data may lead to medical insight into
Correspondence

to: R. M. Altman, Department of Statistics and Actuarial Science, Simon Fraser University, 8888
University Drive, Burnaby, BC V5A 1S6, Canada.
raltman@stat.sfu.ca

E-mail:

Contract=grant sponsor: Natural Sciences and Engineering Research Council of Canada

Copyright ? 2005 John Wiley & Sons, Ltd.

Received May 2004


Accepted September 2004

R. M. ALTMAN AND A. J. PETKAU

the behaviour of the disease. Furthermore, an adequate model is necessary for designing ecient clinical trials for new MS therapies. Albert et al. [2] (henceforth called AMSF) provide
a starting point for determining such a model. These authors discuss a study of relapsingremitting MS patients who received monthly MRI scans for a period of approximately 30
months. Each month, the numbers of lesions observed on these scans were recorded. AMSF
report the data for three particular patients (see Figure 1). The observed lesion counts range
from 0 to 19, with a mean of 4.5 and a median of four lesions per scan.
AMSF propose three dierent models for these data, and t each to the three patients
data individually. The rst model assumes that the monthly lesion counts are independent
and Poisson distributed with a common mean. The second assumes that, conditional on past
observations, each count is Poisson distributed with mean depending on these past observations. The third model is a hidden Markov model (HMM), where the lesion count at time t is
assumed to be Poisson distributed with mean depending on the patients unobserved disease
state at that time. As AMSF point out, MS is known to be a very heterogeneous disease.
Not surprisingly, their analysis suggests that none of the models is appropriate for all three
patients.
In this paper, we discuss some issues surrounding AMSFs HMM. Of the three models, we
focus on the HMM because we feel that it best reects the nature of the disease. Specically,
given the results of Reference [1], we expect the number of lesions to depend on the patients
underlying disease state (relapse or remission).
AMSFs HMM can be expressed as follows. Let Y t be the lesion count at month t. Let
{Z t } be an unobserved, stationary Markov chain taking on values in the set {1; 1} and with
transition probabilities given by P(Z t =j | Z t1 = i) = , i=j. The stationary distribution of {Z t }
is thus j P(Z t = j) = 12 , j =1; 1. Let Y t | Z1 ; : : : ; Z t be distributed as Poisson( t ), where


t =

 t1 ;

Zt = 1

(1=) t1 ; Z t = 1


For identiability, we assume that 1. Letting S t = tj = 1 Zj and letting  0 be the mean
count at baseline, we can rewrite this assumption as  t =  0  S t . It is also assumed that,
given Z1 ; : : : ; Z t , Y t is independent of Y1 ; : : : ; Yt1 ; Yt+1 ; : : : ; Yn . AMSF use the EM algorithm
to estimate the parameters ,  0 , and  for each patient separately.
The purpose of our manuscript is twofold. The rst is to discuss a more ecient means
of estimating the parameters of the HMM than the EM algorithm. The second is to provide
some insight into the underlying assumptions of AMSFs model, and to propose some possible
extensions. Our work will suggest that, in order to make substantive statements about the
course of lesion activity, we require a much larger collection of patients and a exible model
which describes all patients data simultaneously.
Our paper is organized as follows. In Section 2, we write AMSFs model in a more
standard form. This form allows us to evaluate the likelihoodand hence the maximum
likelihood estimates (MLEs)very eciently. We outline the details of our estimation method
in Section 3. In Section 4, we describe some features of the original HMM in more detail,
and suggest some extensions to the model. We conclude with Section 5, where we discuss
the need for a model which borrows strength across patients.
Copyright ? 2005 John Wiley & Sons, Ltd.

Statist. Med. (in press)

APPLICATION OF HIDDEN MARKOV MODELS

Patient 1

Lesion Count

15

10

0
0

10

20

30

Month
Patient 2

Lesion Count

15

10

0
0

10

20

30

Month
Patient 3

Lesion Count

15

10

0
0

10

20

30

Month

Figure 1. MS=MRI data reported by AMSF.

Copyright ? 2005 John Wiley & Sons, Ltd.

Statist. Med. (in press)

R. M. ALTMAN AND A. J. PETKAU

2. WRITING THE AMSF MODEL IN STANDARD FORM


AMSF do not express their HMM in the usual way. In particular, standard HMMs satisfy the
property
P(Y t | Z1 ; : : : ; Z t ) = P(Y t | Z t )
In AMSFs model,  t depends not just on Z t , but on Z1 ; : : : ; Z t1 as well.
However, it is possible to write AMSFs model as a standard HMM. Specically, dene
S0 = 0 and U t = (St1 ; S t ), t = 1; : : : ; n. We will then think of {U t }, rather than {Z t }, as the
(two-dimensional) hidden process. The process {U t } is a Markov chain, and, as in the standard model, P(Y t | U1 ; : : : ; U t ) = P(Y t | U t ). Let A t = {t; t + 2; : : : ; t 2; t }. The transition
probabilities for {U t } are given by
P(x2 ; x1 ); (x1 ; x0 ) (t) P {U t = (x1 ; x0 ) | Ut1 = (x2 ; x1 )}

1 ; x0 A t ; x0 = x1 + 1; x1 = x2 + 1

1 ; x0 A t ; x0 = x1 1; x1 = x2 1

x0 A t ; x0 = x1 + 1; x1 = x2 1
= ;

;
x0 A t ; x0 = x1 1; x1 = x2 + 1

0;
otherwise
This formulation of the model elucidates one of its key features: the state space of the hidden
processand hence the transition probabilitiesvary with time. In other words, this model
is a non-homogeneous HMM, and, although {Z t } is stationary, {Y t } is not.
In the case of a stationary HMM where the hidden process takes on only a nite number
of values, the MLEs are consistent and asymptotically normal under quite general conditions
[3, 4]. Similar results hold in the case when {Z t } belongs to a compact set and is possibly
non-stationary [5]. In these cases, the observed information converges in probability to the
Fisher information matrix [4, 5]. In addition, in the comparison of nested stationary HMMs
with a common, known number of hidden states, the likelihood ratio test (LRT) statistic has
the usual asymptotic 2 distribution [6]. However, we are not aware of any results in the
literature regarding the asymptotic properties of non-homogeneous HMMs. (AMSF did not
provide standard errors or carry out formal inference in their analysis.)
To make inferences about AMSFs model and its extensions, we will rely on the above
results for homogeneous HMMs. Because these results have not been shown to hold for nonhomogeneous HMMs, our conclusions should be considered only informal. If more formal
conclusions were desired, one could t a stationary Poisson HMM with an appropriate value
of K to the data. Specically, in AMSFs model, the mean lesion count at time t is restricted
to a discrete number of values (evenly spaced on the log scale). If we assume that the
observed process is stationary, it is reasonable to use a nite approximation to these mean
values, i.e. to assume that the mean at time t is one of K values, K. This new model
is simply a stationary Poisson HMM with K hidden states and with some restrictions on the
transition probabilities; hence, standard inference results would certainly apply.
Copyright ? 2005 John Wiley & Sons, Ltd.

Statist. Med. (in press)

APPLICATION OF HIDDEN MARKOV MODELS

3. COMPUTATIONAL ISSUES
AMSF use the EM algorithm to obtain the MLEs of the parameters in their model. This
algorithm can be a useful means of computing the MLEs when there are missing data (e.g.
the hidden states), and is very popular in the HMM literature.
However, as pointed out in Reference [7], HMM likelihoods can be computed very simply
as a product of matrices. Hence, direct maximum likelihood estimation is typically much
more ecient than the EM algorithm. In particular, using the formulation of Section 2 and
dening f(Y t | S t ) as the Poisson( t ) distribution, we can express the likelihood associated
with AMSFs model as
L=


S1 A 1

S1 A1


Sn An

S1 f(Y1 | S1 )

S1 f(Y1 | S1 )

S2 A2

n

t=2

P(St2 ; St1 ); (St1 ; S t ) f(Y t | S t )

P(S0 ; S 1 );( S 1 ; S2 ) f(Y2 | S2 )


Sn A n

P(Sn2 ; S n1 ); (Sn1 ; S n ) f(Yn | Sn )

By dening n appropriate matrices (one for each of the summations), we can compute the
likelihood as a product of these n matrices. This formulation allows the ecient evaluation
of the likelihood, even for relatively large numbers of hidden states. (See Reference [8] for
an example of tting a similar model with up to ve hidden states.)
After evaluating the likelihood, we can then obtain the MLEs numerically. We use a quasiNewton routine [9]. For the cases we have considered, maximizing the likelihood directly
produces the MLEs far more quickly than the EM algorithm. Finally, HMM likelihoods often
have many local maxima, so good starting values can be of critical importance. The speed of
the direct MLE computations allows us to try a variety of starting valuesor to do a grid
search over a set of reasonable valueswithin a reasonable time frame. The same search
using the EM algorithm would be computationally intensive. For these reasons, we much
prefer the use of direct likelihood maximization for tting HMMs.
4. HMMS FOR MS=MRI DATA
AMSFs model has two unusual features: the constraint that P(Z t = j | Zt1 = i) = , i=j, and
the structure of  t . The former implies that patients spend half the time in a state of deterioration and the other half in a state of improvement. The latter implies that the conditional
mean lesion count either increases or decreases each month; it cannot remain stable. These
assumptions may not be realistic for all patients.
In this section, we suggest some models with more exible structures. We t each model
to AMSFs data using the method of direct maximum likelihood estimation described in
Section 3. The quasi-Newton routine provides an estimate of the inverse Hessian of the loglikelihood, from which we obtain approximate standard errors.
4.1. Original model
We rst t AMSFs model. We can facilitate the maximization of the likelihood by transforming the parameters so that their range is the entire real line, i.e. 0 = log  0 ,  = log( 1),
and  = log(=(1 )).
Copyright ? 2005 John Wiley & Sons, Ltd.

Statist. Med. (in press)

R. M. ALTMAN AND A. J. PETKAU

Table I. Parameter estimates and standard errors for AMSFs model.


Patient 1
Parameter
0




Patient 2

Patient 3

Transformation

Estimate

SE

Estimate

SE

Estimate

SE

log 
log(
 1)


log 1

1.070
11.083
NA

0.091
NA
NA

1.362
0.282
1.241

0.178
0.245
0.575

2.223
1.128
1.336

0.244
0.560
0.847

log L

66.814

72.029

65.363

Table I gives parameter estimates and approximate standard errors. Our results, with the
addition of the standard errors, are similar to AMSFs. One dierence is that we have achieved
a higher value of the likelihood in the case of Patients 2 and 3. An explanation for this dierence might be that we considered a larger collection of starting values. A second dierence
is that we have omitted the estimate of  for Patient 1. For this patient,  is estimated as
1.000, and the model reduces to that for independent Poisson counts. So, the estimates for 
given by the quasi-Newton routine (and that given by AMSF) are, in fact, arbitrary. Since
 = 1 is on the boundary of the parameter space, there is no guarantee that the usual standard
error for the estimate of  is even approximately correct. Hence, we omit this value.
One question raised by AMSF is whether the complexity of the HMM is warranted, or
whether the independent Poisson count model is sucient to describe the variability in the
data. In principle, we should be cautious about making such inferences, since the test of  = 1
is a boundary problem. Informally, though, in the case of Patient 1, there is no evidence to
suggest that the simpler model is inadequate. In the case of Patients 2 and 3, the 95 per cent
condence intervals for  ([1.47, 2.22] and [1.11, 1.97], respectively) suggest evidence against
the hypothesis that  = 1. Thus, the HMM structure seems to be more appropriate for these
patients than the simpler model.
4.2. Generalization of the transition probabilities
The rst extension we consider is the use of general transition probabilities. In particular, we
assume that P(Z t = 1 | Zt1 =1) =  and P(Z t =1 | Zt1 = 1) = . This generalization allows
patients to spend diering portions of time in states of deterioration and improvement. Based
on the analysis in Section 4.1, we do not apply this new model to the data from Patient 1.
The parameter estimates for the other patients are given in Table II. In the case of Patient 3,
 is estimated as 1.000, which is on the boundary of the parameter space. Hence, we do not
include a standard error.
To test the validity of the assumption that  = , we note that AMSFs model is nested
within the more general model. We then use the LRT to compare the two models, assuming
that the LRT statistic has an asymptotic 12 distribution. Surprisingly, the more general model
does not t substantially better for either of the two patients (p-value = 0.64 and 0.13 for
Patients 2 and 3, respectively).
We have two possible explanations for these results. First, making inferences about the
hidden process is usually a dicult problem. The standard errors of the estimates of  given
in Table I are quite large relative to the estimates themselves, and relative to the standard
Copyright ? 2005 John Wiley & Sons, Ltd.

Statist. Med. (in press)

APPLICATION OF HIDDEN MARKOV MODELS

Table II. Parameter estimates and standard errors for the model with general transition probabilities.
Patient 2
Parameter
0




Patient 3

Transformation

Estimate

SE

Estimate

SE

log  0
log(
 1)


log 1



log 1

1.360
0.281
1.489

0.173
0.244
0.801

2.186
1.416
0.975

0.143
0.273
0.515

1.039

0.690

18.570

NA

log L

71.921

64.222

Table III. Parameter estimates and standard errors for the model with a general conditional
mean structure.
Patient 1
Parameter
0
0
1


Transformation
log  0
log 0
log
 1 

log 1

Estimate
1.168
0.006
0.006
NA

log L

66.652

Patient 2

Patient 3

SE

Estimate

SE

Estimate

SE

0.200
0.089
0.089
NA

0.987
0.484
0.621
0.974

0.256
0.621
0.120
0.708

2.446
0.466
0.398
2.384

0.247
0.166
0.158
0.813

70.401

63:904

errors of the estimates of 0 and  . The same is true of the standard errors of the estimates
of  and  in Table II.
A second explanation may lie in the structure of  t : the proportional increase in the mean
when Z t = 1 is assumed equal to the proportional decrease in the mean when Z t = 1. In the
case where there is no overall trend in the data (as is true for these particular patients, as
well as for relapsing-remitting patients in general when observed over a short time period),
the number of transitions from decreasing to increasing mean is forced to equal approximately
the number of transitions from increasing to decreasing mean. This statement is equivalent to
AMSFs assumption that 1 = 1 = 0:5. Since 1 = =( + ) and 1 = =( + ), we have
that  = .
4.3. Generalization of the conditional mean structure and of the transition probabilities
In light of the discussion in Section 4.2, we might consider modelling  t more generally while
leaving the transition probabilities as in Section 4.1, e.g.

0 t1 ;
if patient is deteriorating at time t
t =
(1=1 )t1 ; if patient is improving at time t
The parameter estimates and standard errors are given in Table III. When 0 = 1=1 ,
the model implies that {Y t } are independent with Y t distributed as Poisson( 0  t ), so that
 is arbitrary. Thus, for Patient 1, we omit the estimate of  . This modication does not
Copyright ? 2005 John Wiley & Sons, Ltd.

Statist. Med. (in press)

R. M. ALTMAN AND A. J. PETKAU

Table IV. Parameter estimates and standard errors for model with general transition probabilities
and conditional mean structure.
Patient 2
Parameter
0
0
1



Patient 3

Transformation

Estimate

SE

Estimate

SE

log  0
log 0
log 1



log 1



log 1

1.889
0.811
0.383

0.235
0.169
0.071

2.445
0.466
0.398

0.233
0.154
0.149

2.022

1.134

2.372

1.045

0.399

0.505

2.396

1.052

log L

68.695

63.904

signicantly improve the t for Patient 1 ( p-value = 0.57), but we observe some evidence of
an improved t for Patients 2 and 3 ( p-value = 0.071 and 0.088, respectively).
It was anticipated that, for Patients 2 and 3, the t might be further improved by using
general transition probabilities (as in Section 4.2). The estimates of the transformed parameters
and standard errors are given in Table IV. The LRTs comparing this model to the model with
 =  yield the p-values 0.065 and 1.00 for Patients 2 and 3, respectively. Thus, for Patient
2, there is some support for the expanded model.
4.4. Addition of a third hidden state
Our nal question of interest regarding AMSFs model involves the number of hidden states.
We consider the addition of a third hidden state, state 0, where the patients condition is
stable. This modication can be expressed as

t =

 t1 ;

if patient is deteriorating at time t

 t1 ;

if patient is stable at time t

(1=)t1 ;

if patient is improving at time t

= 0

St

We represent the transition probabilities


as Pi;j P(Z t = j | Zt1 = i), i; j {1; 0; 1}. Of

course, we have the constraint j Pi; j = 1 for all i.
One disadvantage of such an extension is the introduction of the problem of computing the
stationary probabilities, which are used as initial probabilities for the hidden Markov chain.
In the two-dimensional case, we have the simple closed form for the stationary distribution
given in Section 4.2. In order to compute the stationary distribution in the three-dimensional
case, however, a system of three linear equations must be solved at each iteration of the
quasi-Newton algorithm. Another disadvantage of this model is the large number of unknown
parameters. However, we could reduce this number if we were willing to place restrictions on
the transition probabilities (as in AMSFs model), for example by assuming that the transition
probability matrix is symmetric.
Copyright ? 2005 John Wiley & Sons, Ltd.

Statist. Med. (in press)

APPLICATION OF HIDDEN MARKOV MODELS

Table V. Parameter estimates and standard errors for 3-state model.


Patient 1

Patient 2

Patient 3

Parameter

Transformation

Estimate

SE

Estimate

SE

Estimate

SE

0

log  0
log ( 1)


log 1pp11p2


log 1pp12p2


log 1pp33p4


log 1pp34p4


log 1pp55p6


log 1pp56p6

1.061
0.003

0.139
0.028

1.372
0.671

0.155
0.130

2.119
0.395

0.110
0.068

0.417

0.476

3.139

0.229

0.184

0.167

0.536

0.816

4.316

0.315

0.126

0.195

0.173

0.537

0.142

0.643

6.825

6.846

0.370

0.250

6.376

1.247

2.210

0.349

0.547

0.856

1.757

1.002

7.741

6.168

1.525

0.136

7.079

1.051

1.170

1.401

P1;
1

P1;
0

P0; 1
P0; 0
P1; 1
P1; 0

log L

65.931

69.448

63.514

The parameter estimates and standard errors are given in Table V. In this case, the likelihood functions are quite at (likely due to the large number of parameters and relatively
small sample sizes) and hence dicult to maximize. The parameter estimates are not entirely
reliable, and may correspond to a local maximum.
When we compare the likelihoods in Table V to those in Table I, we see that substantial
decreases occur for Patient 2 in particular. Thus, we might surmise that this 3-state model
is more appropriate than AMSFs model. It would be a mistake, however, to use the 52
distribution to gauge the extremity of the LRT statistic. The test comparing models with
diering numbers of hidden states amounts to the hypothesis that some of the transition
probabilities are zero. Thus, this test is a boundary problem and does not satisfy the conditions
required for the usual results for LRTs. Moreover, we cannot assume that methods such as the
Akaike information criterion or the Bayesian information criterion provide consistent estimates
of the number of hidden states. The question of estimating the number of hidden states in
stationary HMMs is addressed in Reference [8], but the non-homogeneous case has not been
considered in the literature.
5. DISCUSSION
Our analyses in Section 4 illustrate a number of diculties in applying AMSFs model in
particular, and in modelling MS=MRI data in general.
Formal inference about non-homogeneous HMMs is challenging. It is dicult to have
condence in applying these models in the absence of tools for assessing their t. Further
research in this eld is warranted. In particular, additional theory is required in order to
formally examine the adequacy of AMSFs models and its extensions for MS=MRI data. A
technique for studying the t of a stationary HMM is proposed in Reference [10]; this work
may be a starting point for treating the non-homogeneous case.
The issue that is, perhaps, of greatest importance in the context of models for MS=MRI
data is the planning of clinical trials (see e.g. Reference [11]). The selection of a reasonable
Copyright ? 2005 John Wiley & Sons, Ltd.

Statist. Med. (in press)

R. M. ALTMAN AND A. J. PETKAU

model is key to achieving this goal. For example, in order to design clinical trials for new MS
therapies, we need a model which allows the incorporation of a treatment eect. Furthermore,
we require a model which describes all patients data simultaneously. In particular, in our
analyses, the standard errors associated with the estimates of the parameters of the hidden
process are relatively large. Assuming the same model for each patient is one means of
reducing this uncertainty. However, the behaviour of lesion counts in MS patients is often
highly variable. It thus seems more reasonable to allow at least some model parameters to vary
across patients. Random eects are a useful means of capturing between-patient dierences
while still borrowing strength across patients. Such models are discussed in detailin a very
general settingin Altman (manuscript under revision), where a new class of HMMs for
multiple time series, called mixed hidden Markov models (MHMMs), is developed. This
class is based on the generalized linear mixed model framework. For MS=MRI data, we might
assume, for example, that, conditional on a patient-specic random eect, u, and the hidden
disease state, Z, the lesion count at that time is Poisson distributed with mean depending on
u and Z. Including random eects in the transition probabilities is an option as well. In this
case, the model would allow the percentage of time spent in each disease state to vary among
patients. MHMMs also readily allow the incorporation of covariates, including a treatment
eect. The exibility of MHMMs makes them a promising possibility for modelling MS=MRI
data and planning clinical trials. Good choices of models from this class for such data are
currently under investigation.
ACKNOWLEDGEMENTS

This manuscript includes work from the rst authors Ph.D. thesis, and was partially supported by a
research grant and postdoctoral fellowship from the Natural Sciences and Engineering Research Council
of Canada. We would like to express our appreciation to Paul Albert, Henry McFarland, and the Joseph
Frank Experimental Neuroimaging Section, Laboratory of Diagnostic Radiology Research, Clinical Center, NIH, for providing the MS=MRI data.
REFERENCES
1. Sormani MP, Bruzzi P, Beckmann K, Wagner K, Miller DH, Kappos L, Filippi M. MRI metrics as surrogate
endpoints for EDSS progression in SPMS patients treated with IFN beta-1b. Neurology 2003; 60(9):14621466.
2. Albert PS, McFarland HF, Smith ME, Frank JA. Time series for modelling counts from a relapsing-remitting
disease: application to modelling disease activity in multiple sclerosis. Statistics in Medicine 1994; 13(57):
453 466.
3. Leroux BG. Maximum-likelihood estimation for hidden Markov models. Stochastic Processes and their
Applications 1992; 40(1):127143.
4. Bickel PJ, Ritov Y, Ryden T. Asymptotic normality of the maximum-likelihood estimator for general hidden
Markov models. Annals of Statistics 1998; 26(4):16141635.
5. Douc R, Matias C. Asymptotics of the maximum likelihood estimator for general hidden Markov models.
Bernoulli 2001; 7(3):381 420.
6. Giudici P, Ryden T, Vandekerkhove P. Likelihood-ratio tests for hidden Markov models. Biometrics 2000;
56(3):742747.
7. MacDonald IL, Zucchini W. Hidden Markov Models and Other Models for Discrete-Valued Time Series.
Chapman & Hall: London, 1997.
8. MacKay RJ. Estimating the order of a hidden Markov model. The Canadian Journal of Statistics 2002;
30(4):573589.
9. Nash JC. Compact Numerical Methods for Computers: Linear Algebra and Function Minimisation. Wiley:
New York, 1979.
10. Altman RM. Assessing the goodness-of-t of hidden Markov models. Biometrics 2004; 60(2):444 450.
11. McFarland HF, Frank JA, Albert PS, Smith ME, Martin R, Harris JO, Patronas N, Maloni H, McFarlin DE.
Using gadolinium-enhanced magnetic resonance imaging lesions to monitor disease activity in multiple sclerosis.
Annals of Neurology 1992; 32(6):758766.

Copyright ? 2005 John Wiley & Sons, Ltd.

Statist. Med. (in press)

You might also like