STATISTICS IN MEDICINE

Statist. Med. 2002; 21:1115–1127 (DOI: 10.1002/sim.1068)

Functional data analysis with application to periodically

stimulated foetal heart rate data. II: Functional logistic
regression

Sarah J. Ratcli
e 1; ∗; † , Gillian Z. Heller 1 and Leo R. Leader 2

1 Department of Statistics; Macquarie University; NSW 2109; Australia
2 School of Obstetrics and Gynaecology; Faculty of Medicine; University of New South Wales;
NSW 2052; Australia

SUMMARY
We present a basis solution for the modelling of a binary response with a functional covariate plus any
number of scalar covariates. This can be thought of as singular longitudinal data analysis as there are
more measurements on the functional covariate than subjects in the study. The maximum likelihood
parameter estimates are found using a basis expansion and a modied Fisher scoring algorithm. This
technique has been extended to model a functional covariate with a repeated stimulus. We used period-
ically stimulated foetal heart rate tracings to predict the probability of a high risk birth outcome. It was
found that these tracings could predict 94:1 per cent of the high risk pregnancies and without the stim-
ulus, the heart rates were no more predictive than chance. Copyright ? 2002 John Wiley & Sons, Ltd.
KEY WORDS: foetal habituation; functional data analysis; generalized linear models; maximum
likelihood

1. INTRODUCTION

Techniques for analysing functional data, where data on each subject can be thought to repre-
sent a continuous curve, have received much attention in recent years. Like longitudinal data,
the measurements are correlated over time. However, functional data contain many more mea-
surements per subject than subjects, resulting in singular equations if traditional longitudinal
methods are used. New methods, which have come to be known as functional data analy-
sis [1; 2] (FDA), have been developed for this type of data.
Many researchers have presented methods for FDA; they include functional principal com-
ponent analysis [3–5], functional ANOVA [6] and functional regression [1]. Existing regres-
sion techniques deal with only continuous [2; 7] and functional responses [8].
In this paper we present a method for FDA with a binary response, which we call functional
logistic regression. We then extend this method to the case of a functional covariate with a

∗ Correspondence to: Sarah J. Ratcli
e, Department of Biostatistics and Epidemiology, University of Pennsylvania,
School of Medicine, 831 Blockley Hall, 423 Guardian Drive, Philadelphia, PA 19104-6021, U.S.A.
† E-mail: [email protected]

Received February 2000

Copyright ? 2002 John Wiley & Sons, Ltd. Accepted May 2001
1116 S. J. RATCLIFFE, G. Z. HELLER AND L. R. LEADER

repeated stimulus. These techniques are used to determine if human foetal heart rate responses
to repeated vibroacoustic stimulation, as presented in Part I [9], are predictive of an infant’s
risk status at birth. We also investigate whether the same results can be found without the
stimuli.
As stated in Part I [9], this type of data has usually been analysed using the concept of
habituation, which is a decrease, leading to a cessation of the response that follows a repeated
stimulus. For example, links between impaired habituation and high risk [10] and traumatized
newborn infants [11] have been found. However, by using functional techniques, we can
model the data without incorporating the subjective notion of habituation.

2. METHODOLOGY

2.1. Functional logistic regression

We look at extending functional linear models, using standard generalized linear modelling
(GLM) techniques [12] to the case of a binary response variable, resulting in a functional
logistic regression with maximum likelihood parameter estimates.
For each of n subjects, the observed response, yi , is assumed to come from a Bernoulli
distribution with probability of success
i . As with standard GLM, the data is modelled using
the logistic link function, giving the functional logistic regression model

 

i T
log = zi  + xi (t)(t) dt; i = 1; : : : ; n (1)
1 −
i

where  = [ 0 1 : : :  r ]T is the vector of parameters for the covariates, zi = [1 z1 : : : zr ]T is

the vector of covariates for the ith subject, xi (t) is the functional covariate for the ith subject
and (t) is the functional parameter.
Solving this problem directly requires the inversion of a singular matrix. We follow instead
the functional regression techniques of Ramsay and Silverman [2] and model the functional
data and time parameter by linear combinations of known basis functions, (t). These basis
functions should be chosen to reect the characteristics of the data; Fourier basis functions
could be used to model periodic data, otherwise B-spline basis functions could be used as
they have the advantage of nite support. Thus, we have

(t) = T (t)b (2)

and

xi (t) = ci (t)
(3)
⇒ X (t) = C(t)

where b is a vector of basis coecients which need to be estimated. Estimates for C can be
found using standard least squares (as described in Part I).

Copyright ? 2002 John Wiley & Sons, Ltd. Statist. Med. 2002; 21:1115–1127
 FUNCTIONAL LOGISTIC REGRESSION WITH PERIODICALLY STIMULATED DATA 1117

Substituting for (t) and xi (t), the regression model (1) becomes

 

i
log = zi  + ci (t)T (t)b dt
T
1 −
i
= ziT  + ci Wb



(4)
⇒ log = Z + CWb
1−

 

= [Z CW ]
b

where W = (s)T (s) ds.
This model is now in a similar form to standard logistic regression models, and maximum
likelihood parameter estimates can be found via Fisher scoring [13]:
   −1  
ˆ Z T w∗ Z Z T w∗ CW ZT
= w∗  (5)
b̂ (CW )T w∗ Z (CW )T w∗ CW (CW )T

where w∗ is the diagonal weights matrix with
i (1 −
i ) as its diagonal elements and

i = ˆ i + (yi −
ˆ i )=wii∗
ˆ i = ziT ˆ + ci Wb

Since  depends on the parameter estimates and these estimates in turn depend on , an
iterative algorithm (as given in McCullagh and Nelder [12]) is used to estimate the parameters
in practice.
We have used cross-validated log-likelihoods to determine the number of basis functions,
M , to be used in the model. Thus, the optimal M value is the one which maximizes the
following equation:

CV(M ) = −2 (yi log
ˆ i;−i + (1 − yi ) log(1 −
ˆ i;−i )) (6)

i

where
ˆ i; −i is the predicted probability of a success for subject i when the parameter estimates
are found without subject i.
To compare functional logistic regression models, the residual deviance can be used. For
functional logistic regression, this will reduce to (see McCullagh and Nelder, reference [12],
p. 121 for details)

ˆ = −2ˆ T
ˆ − 2
D(y;
) log(1 −
ˆ i ) (7)
i

For standard logistic regression, the deviance function is not uniquely dened; it depends on
whether the data are grouped or ungrouped. However, with functional logistic regression we
will always have ungrouped data since each subject has a unique functional covariate.

Copyright ? 2002 John Wiley & Sons, Ltd. Statist. Med. 2002; 21:1115–1127
1118 S. J. RATCLIFFE, G. Z. HELLER AND L. R. LEADER

2.2. Functional logistic regression with a repeated stimulus

We extend the functional logistic regression model to the situation where there is a special
structure in the functional data, namely a repeated stimulus. The functional covariate now has
two parts: the curve measured within the time frame of each stimulus, and the timing of the
stimulus in relation to the other stimuli. This special structure is needed to model the foetal
heart rate data.
With the special structure, the regression model becomes

 q 

i
log = ziT  + s xi; s (t)(t) dt; i = 1; : : : ; n (8)
1 −
i s=1

where s is the parameter for the sth stimulus. s = 1; : : : ; q (for the foetal data q = 19), and
xi; s (t) is now the functional covariate for the ith subject, measured at time t within the sth
stimulus.
As for functional logistic regression, we have modelled both the functional covariate within
each stimulus minute and the functional parameter via M basis functions. (Note that M is again
chosen using the cross-validated log-likelihood.) Thus, the regression equation (8) becomes

 q 

i
log = ziT  + s Ci (t)T (t)b dt
1 −
i s=1

= ziT  + T Ci Wb (9)

As before, the maximum likelihood parameter estimates can be found via Fisher scoring
(see Appendix A). We have

   −1  
ˆ Z T w∗ Z Z T w∗ D ZT
= T ∗ w∗  (10)
b̂ D w Z DT w ∗ D DT
ˆ T w∗ E(E T w∗ E)−1
ˆT = ( − Z ) (11)

where
 
ˆT C1 W
 .. 
D = . 
T
ˆ Cn W (12)
T
E = [(C1 W b̂) : : : (Cn W b̂)]
i = ziT  + T Ci Wb

In practice, ;
ˆ b̂ and ˆ are found using a combination of the GLM algorithm [12] and the
weighted iterative least squares algorithm from Part I [9]. Details of this new algorithm are
given in Appendix B.

Copyright ? 2002 John Wiley & Sons, Ltd. Statist. Med. 2002; 21:1115–1127
 FUNCTIONAL LOGISTIC REGRESSION WITH PERIODICALLY STIMULATED DATA 1119

3. THE FOETAL HEART RATE DATA

We now return to the foetal heart rate data. The infant’s risk category at birth is a binary
response, with subjects divided into groups depending on their antenatal course and other birth
outcomes: normal and high risk.
We examine two groups of data: heart rate tracings from foetuses that had and had
not been stimulated. Since the foetal heart rates are assumed to return to their base rate
sometime after each stimulus, but before the next stimulus is applied, Fourier basis func-
tions have been used for the stimulated group. Consequently, Fourier basis functions were
also used for the unstimulated group. So, 1 (t) = 1; 2k (t) = cos 2
kt; 2k+1 (t) = sin 2
kt;
k = 1; : : : ; M=2.

3.1. Unstimulated heart rates

For the unstimulated foetuses, the heart rates were recorded every 0:2 seconds for 16–19
minutes. However, only the rst 5 minutes were used for this part of the analysis. Of the 10
subjects in this group, seven were normal and three were high risk.
Without the functional covariate, it was found that none of the scalar predictors could
be used to produce a good logistic regression model for the risk category. Similarly, using
functional logistic regression, none of the predictors was found to be important, including the
functional heart rate covariate (p¿0:10).
However, the best functional logistic model was

 

ˆ i ˆ dt
log = 10:905 + xi (t)(t)
1 −
ˆ i
ˆ
with (t) (Figure 1(a)) estimated using three basis functions. Under this model, only one of
the high risk subjects was correctly classied, while all remaining subjects were classied as
normal (Figure 1(b)), using a cut-o
probability of 0.55. The residual deviance of the model
was 10.514 on 6 degrees of freedom.

3.2. Stimulated heart rates

Of the subjects with stimulated heart rates, 46 of the 63 births (73.0 per cent) were consid-
ered to be normal, while 17 (27.0 per cent) were considered high risk. None of the scalar
covariates was signicant in a logistic regression model using only the scalar covariates. How-
ever, the best of these insignicant models is given below. Included in this model were the
mother’s age group and her parity. Mothers who have never borne a child are 2.48 times
more likely to have a high risk pregnancy than mothers who have borne previous children.
Also, mothers over the age of 35 years are 2.57 times more likely to have a high risk preg-
nancy than younger mothers (see Table I). Under this model, using a classication cut-o
of

ˆ = 0:5, only 3=17 (17.6 per cent) of the high risk pregnancies were correctly identied (see
Table II). The cut-o
of
ˆ = 0:5 gave the best result in terms of balancing the false positive
rate and the sensitivity of the model.
Using functional logistic regression with a repeated stimulus, the best model contained only
the functional pulse covariate; all other scalar covariates were not signicant, as expected.
Since the heart rates were assumed to have returned to a steady state within a minute of the

Copyright ? 2002 John Wiley & Sons, Ltd. Statist. Med. 2002; 21:1115–1127
1120 S. J. RATCLIFFE, G. Z. HELLER AND L. R. LEADER

Figure 1. Functional logistic regression results for unstimulated foetal heart rates.

Table I. Logistic regression summary for risk category.

Logistic regression analysis for risk category

MODEL: parity + age group
Covariate Coecient Standard deviation P Exp(coef)
Constant −0:875 0.373 0.019
Parity (1) −0:909 0.648 0.161 0.403
Agegp35 (1) 0.945 0.706 0.181 2.573
D = 70:314

Table II. Logistic regression – summary of classications for

risk category with cut-o
of
ˆ = 0:5.

Classication table for risk category

MODEL: parity + age group
Predicted Observed level
Normal High risk Total
Normal 44 14
High risk 2 3

% correct 95.7% 17.6% 74.6%

stimulus being applied, Fourier basis functions were used to model the data and the parameter
estimate.
The functional model was found to be

 19


ˆ i ˆ dt
log = −0:910 + ˆs xi; s (t)(t)
1 −
ˆ i s=1

Copyright ? 2002 John Wiley & Sons, Ltd. Statist. Med. 2002; 21:1115–1127
 FUNCTIONAL LOGISTIC REGRESSION WITH PERIODICALLY STIMULATED DATA 1121

Table III. Functional logistic regression – summary of

classications for risk category with cut-o
of
ˆ = 0:6.

Classication table for risk category

MODEL: pulse
Predicted Observed level
Normal High risk Total
Normal 45 1
High risk 1 16

% correct 97.8% 94.1% 96.8%

Figure 2. Parameter estimates from functional logistic regression for risk category.

with a residual deviance of 84.497 on 38 degrees of freedom. The functional time parameter,
(t), was modelled using M = 5 basis functions. The classication table for this model is given
in Table III. Overall, 96.8 per cent of all the subjects were correctly classsied. Using the
best classication cut-o
of
ˆ = 0:6, only one normal and one high risk birth were incorrectly
classied.
The estimated functional time parameter, (t),ˆ and minute parameters, ˆs , are shown
in Figure 2. The histogram of the predicted probabilities is given in Figure 3. This shows
clearly a di
erentiation in the predicted probabilities between the normal and high risk
groups.
Comparing the logistic and functional logistic models, we see that the functional model is
signicantly better than the simple logistic model. The functional model correctly detected
94.1 per cent of the high risk pregnancies, while the logistic model only correctly detected
17.6 per cent. The ROC curves [14] (Figure 4) also show that the functional logistic model
is superior to the logistic model.

Copyright ? 2002 John Wiley & Sons, Ltd. Statist. Med. 2002; 21:1115–1127
1122 S. J. RATCLIFFE, G. Z. HELLER AND L. R. LEADER

Figure 3. Histogram of probabilities of a high risk split by observed response.

Figure 4. ROC curves for the functional and simple logistic regression models for risk category.

3.3. Controls
The results of the stimulated heart rates were validated using the 10 unstimulated subjects as
controls. Since the models were based on 19 minutes worth of heart rate measurements, the
subjects with less than 19 minutes had the start of their measurements replicated at the end
to produce a full 19 minutes worth of data. The predictors and heart rate measurements were
then placed into the above model, found using the stimulated heart rates, in order to predict
the risk category for the controls.

Copyright ? 2002 John Wiley & Sons, Ltd. Statist. Med. 2002; 21:1115–1127
 FUNCTIONAL LOGISTIC REGRESSION WITH PERIODICALLY STIMULATED DATA 1123

Table IV. Summary of classications of risk category for the

controls, cut-o
probability = 0.6.

Classication table for risk category

MODEL: pulse
Predicted Observed level
Normal High risk Total
Normal 4 2
High risk 3 1

% correct 57.1% 33.3% 50.0%

The classication summary, using the functional logistic model found above, for the controls
is given in Table IV. It appears that there is no relationship between the true and predicted
outcomes. Thus, without the stimuli being applied, the heart rate does not give a useful
prediction of the infant’s risk category.

4. CONCLUSION

We have presented a method for logistic regression with a functional predictor. Both the
functional predictor and its corresponding functional parameter were modelled using basis ex-
pansions. The maximum likelihood parameter estimates were then calculated using the standard
logistic regression algorithm [12]. This basis solution has the advantage that it is relatively
simple and can be implemented with only minor changes to existing standard logistic regres-
sion computer algorithms.
Also presented is a modied form of functional logistic regression for use with a functional
predictor with a repeated stimulus. This method was used to model the foetal heart rate data.
Unlike current techniques, no denition of habituation was needed for the model.
This study found that the stimulated foetal heart rates were a good predictor of an infant’s
risk category at birth. The functional model correctly predicted 94.1 per cent of the high
risk pregnancies, compared to 17.6 per cent for the best standard logistic regression model.
However, it should be borne in mind that these percentages are optimistic, as the parameter
estimates were based on the same data on which predictions were made. Also, the functional
model used substantially more degrees of freedom than the standard logistic model.
It was found that the repeated stimulus was necessary for the heart rates to be predictive.
Without the stimulus, the heart rates were no more predictive of the infant’s birth outcome
than chance.
The other covariates in the study were found to have no e
ect on the infant’s birth out-
come. This included maternal age, smoking, drinking and ca
eine intake levels. However, the
smoking, drinking and ca
eine intake rates were low in the sample group.
The ndings from both parts of this study may have important clinical implications. Using
the functional model after stimulation as a clinical test may help alert the obstetrician to a high
risk birth, and to determine the optimal time to deliver a foetus in a complicated pregnancy.
It may also have a useful role in alerting parents about possible developmental delay, so that
early intervention could be undertaken.

Copyright ? 2002 John Wiley & Sons, Ltd. Statist. Med. 2002; 21:1115–1127
1124 S. J. RATCLIFFE, G. Z. HELLER AND L. R. LEADER

It is now planned to undertake a prospective randomized trial to see if using functional

analysis as described in this study is clinically useful.

APPENDIX A: DERIVATION OF PARAMETER ESTIMATES FOR

FUNCTIONAL LOGISTIC REGRESSION WITH A REPEATED
STIMULUS

As with standard GLMs, the actual responses, yi , are assumed to come from a Bernoulli(
i )
distribution. Thus, the log-likelihood function can be expressed as
n

 

i
l(
; y) = yi log + log(1 −
i ) (A1)
i=1 1 −
i
Parameter estimates can be found by maximizing equation (A1) with respect to ; b and .
This is done using the Fisher scoring method [13].
First, the derivative of the log-likelihood function with respect to
i is
9l yi −
i
=
9
i
i (1 −
i )
Using



i
log = i = ziT  + T Ci Wb
1 −
i
and the chain rule, the derivative of the log-likelihood with respect to  k is

9l n 9l d
9
i i
=
9 k i=1 9
i di 9 k

n yi −
i
=
i (1 −
i )zi;Tk
i=1
i (1 −
i )

n
= (yi −
i )zi;Tk (A2)
i=1

The Fisher information for  is given by

2 
9 l n 1 9
i 9
i
−E =
9j 9 k i=1
i (1 −
i ) 9 j 9  k
n 1
=
i (1 −
i )zi;Tj
i (1 −
i )zi;Tk
i=1
i (1 −
i )
n
=
i (1 −
i )zi;Tj zi;Tk
i=1

= {Z T w∗ Z }jk

Copyright ? 2002 John Wiley & Sons, Ltd. Statist. Med. 2002; 21:1115–1127
 FUNCTIONAL LOGISTIC REGRESSION WITH PERIODICALLY STIMULATED DATA 1125

where w∗ is the diagonal weights matrix, as given in the algorithm. By Fisher’s scoring
method, the new estimate of ; ˆ new , is given by

2 
2 
9 l 9 l 9l
−E 2 ˆ new = −E 2 ˆ +
9  9  9
Now


92l
−E 2 ˆ = Z T w∗ Z ˆ
9 
= Z T w∗ (ˆ − Db̂)

Thus, the new estimate is found by solving

Z T w∗ Z ˆ new = Z T w∗ (ˆ − Db̂) + Z T (y −
)

ˆ

= Z T w∗ ( − Db̂)

⇒ ˆ = (Z T w∗ Z)−1 Z T w∗ ( − Db̂)

Similarly

b̂ = (DT w∗ D)−1 DT w∗ ( − Z )
ˆ

ˆ T w∗ E(E T w∗ E)−1
ˆT = ( − Z )

Note that the equations for ˆ and b̂ can be combined and written as
   T ∗ −1  
ˆ Z w Z Z T w∗ D ZT ∗
= T ∗ w 
b̂ D w Z DT w ∗ D DT

APPENDIX B: ALGORITHM FOR ESTIMATING PARAMETERS

i
Let i = log 1−
i
and yi = the binary response for the ith subject. Then the iterative algorithm
proceeds as follows:
Step 0. Find an initial estimate for i ; ˜i , using the response for subject i; yi . We use
˜i = (yi + 0:5)=2
for an initial estimate. The 0.5 is necessary since yi is either 0 or 1 (see
McCullagh and Nelder [12]). Using this estimate, nd estimates for the pa-
rameters ;  and b. These estimates are found by minimizing the residual sum

Copyright ? 2002 John Wiley & Sons, Ltd. Statist. Med. 2002; 21:1115–1127
1126 S. J. RATCLIFFE, G. Z. HELLER AND L. R. LEADER

of squares, resulting in an iterative algorithm for functional regression, with ˜i

as the response:
   T −1  
ˆ Z Z Z TD ZT
= ˜
b̂ DT Z DT D DT

ˆ T E(E T E)−1
ˆT = (˜ − Z )

where ˜ = [˜1 : : : ˆ n ]T .
Step 1. Find the linear predictors ˆ i for the regression using the parameter estimates:

ˆ i = ziT ˆ + ˆT Ci W b̂

Step 2. Calculate the probability of a success for each subject:

exp(ˆ i )

ˆ i =
1 + exp(ˆ i )
Step 3. Calculate the adjusted dependent variable i and weights wi :

wi =
ˆ i (1 −
ˆ i )

yi −
ˆ i
i = ˆ i +
wi

Step 4. Re-estimate the parameter values via a weighted version of the functional re-
gression algorithm [9], with i as the response variable. This weighted version
nds the parameter estimates via a weighted least squares with weights wi :

i = ziT  + T Ci Wb

where the parameter estimates in the algorithm are found using the following
equations with w∗ as the diagonal weights matrix and  the vector of adjusted
dependent variables:
   T ∗ −1  
ˆ Z w Z Z T w∗ D ZT ∗
= w  (B1)
b̂ DT w ∗ Z DT w ∗ D DT

ˆ T w∗E(E T w∗ E)−1
ˆT = (v − Z ) (B2)

Step 5. Check for convergence. If the algorithm has converged, then stop; else repeat
from step 1.
Convergence is determined by looking at the change in the parameter estimates from one
iteration to the next. If the changes are suciently small, the algorithm is assumed to have
converged.

Copyright ? 2002 John Wiley & Sons, Ltd. Statist. Med. 2002; 21:1115–1127
 FUNCTIONAL LOGISTIC REGRESSION WITH PERIODICALLY STIMULATED DATA 1127

REFERENCES
1. Ramsay J, Dalzell C. Some tools for functional data analysis. Journal of the Royal Statistical Society, Series
B 1991; 53:539–572.
2. Ramsay J, Silverman B. Functional Data Analysis. Springer: New York, 1997.
3. Besse P, Ramsay J. Principal components analysis of sampled functions. Psychometrika 1986; 51:285 – 311.
4. Bouhaddou O, Obled C, Dinh T. Principal component analysis and interpolation of stochastic processes: methods
and simulation. Journal of Applied Statistics 1987; 14:251– 266.
5. Pezzulli S, Silverman B. Some properties of smoothed principal components analysis for functional data.
Computational Statistics 1993; 8:1–16.
6. Fan J, Zhang J. Functional linear models for longitudinal data. Technical report, Department of Statistics, UNC-
Chapel Hill, 1998.
7. Goutis C. Second-derivative functional regression with applications to near infra-red spectroscopy. Journal of
the Royal Statistical Society, Series B 1998; 60:103–114.
8. Faraway J. Regression analysis for a functional response. Technometrics 1997; 39:254–261.
9. Ratcli
e SJ, Leader LR, Heller GZ. Functional data analysis with application to periodically stimulated foetal
heart rate data. I: Functional regression. Statistics in Medicine 2002; 21:1103–1114.
10. Eisenberg R, Coursin D, Rupp N. Habituation to an acoustic pattern as an index of di
erences among human
neonates. Journal of Auditory Research 1966; 6:239–248.
11. Bronstein A, Itina N, Kamenetsaia A. The orienting reaction in newborn children. In Orienting Re
ex and
Exploratory Behavior, Varonin L, Leontiev A, Luris A, Sokolov E, Vinogradova O (eds). Moscow Academy
of Pedagogical Sciences of RSFSR: Moscow, 1968.
12. McCullagh P, Nelder J. Generalized Linear Models. 2nd edn. Chapman & Hall: London, 1989.
13. Fisher R. Theory of statistical estimation. Proceedings of the Cambridge Philosophical Society 1925; 22:
700 –725.
14. Hanley J. Receiver operating characteristics (ROC) methodology: The state of the art! Critical Reviews in
Diagnostic Imaging 1989; 29:307–335.

Copyright ? 2002 John Wiley & Sons, Ltd. Statist. Med. 2002; 21:1115–1127