Event-Related fMRI: Characterizing Differential Responses
Event-Related fMRI: Characterizing Differential Responses
Event-Related fMRI: Characterizing Differential Responses
1053-8119/98 $25.00 30
Copyright r 1998 by Academic Press
All rights of reproduction in any form reserved.
EVENT-RELATED fMRI 31
study of visually presented words. The stimuli em- signed to span the space of likely responses in a much
ployed were either novel (in the context of the experi- more compact way. In this paper we use a very simple
ment) or had been subject to encoding just prior to basis set: a synthetic hemodynamic response function
scanning. We will use these exemplar data to demon- and its derivative (see Fig. 1). The hemodynamic re-
strate differential responses expressed in terms of the sponse function comprised the sum of two gamma
magnitude of the response and in terms of the response functions and models a hemodynamic response with a
latency, in relation to stimulus onset. This paper is slight undershoot typical of our data. Gamma functions
concerned with the theoretical aspects of the analysis. provide reasonable and comprehensive models of the
A neurobiological interpretation of these data will be hemodynamic response function (Friston et al., 1994;
presented separately. Boynton et al., 1996). The inclusion of derivatives
allows for differential latencies, in the response, among
THEORETICAL BACKGROUND different brain areas and some latitude when (mis)speci-
fying the onset of the stimulus relative to when the data
The Statistical Model of Evoked Responses were actually acquired.
We now create a new set of explanatory variables
In Friston et al. (1994) we presented a model of vij(t ) which represent the original stimulus functions
observed hemodynamic responses, in fMRI time-series, ui(t ) convolved with the jth basis function
that obtains when the underlying neuronal activity
(inferred on the basis of changing task conditions) is
convolved, or smoothed with a hemodynamic response vi j(t ) 5 e b (t) · u (t 2 t) dt;
j i
function. This model was subsequently elaborated in
the context of the general linear model (Friston et al., substituting these expression into Eq. (1) gives
1995a, b; Worsley and Friston, 1995) and has been em-
ployed recently in the analysis of event-related responses
(Josephs et al., 1997; Friston et al., 1997). This model
y(t ) < g0 1 o o g v (t )
i j
ij ij
oe
T
y(t) < h0 1 hi(t) · ui(t 2 t) dt. (1) This is a general linear model with response variable y,
i
0 a column vector representing the observed time-series,
h0 5 g0,
hi(t) 5 o g b (t).
j
ij j (2)
a column vector of error terms e, and a design matrix the response itself can be associated with the synthetic
X 5 [X1 X2 . . . 1 G]. 1 corresponds to a column of ones hemodynamic response function and the effects mod-
and G is a collection of other uninteresting effects or eled by the derivative can be interpreted as a shift of
confounds such as low-frequency artifacts. The parti- this response in time. To find increases in the response
tions of the design matrix Xi contain the explanatory magnitude elicited by event type A, relative to B, we
variables vi j(t) for the ith event type, i.e., the convolved would specify a contrast that was 1 for A’s response
stimulus functions sampled at the times that the scans function, 21 for B’s, and zero elsewhere. The square of
were acquired. The unknown parameters gi j and g0 the resulting SPM5t6 would be exactly the same as the
constitute the initial elements of the column vector g. equivalent SPM5F 6, using Eq. (4), if the derivatives (i.e.,
timing effects) had been treated as confounds. We can
Making Inferences about Evoked Responses
test for differential latencies using a contrast of the
Having formulated the model in this way we can now derivatives in a similar way.
use standard procedures developed for serially corre-
lated fMRI time-series that employ the general linear
Estimating the Standard Error of Evoked Responses
model (Friston et al., 1995a; Worsley and Friston,
1995). These procedures provide parameter estimates A useful feature of using basis functions, when
(e.g., estimates of the basis function coefficients and, formulating the model as above, is that the estimated
implicitly, the responses themselves) and statistical responses are continuous functions of time. This means
parametric maps (SPMs) testing the significance of all we can estimate responses at latencies not actually
or some of the modeled response components. The measured. The estimated response to a particular
parameter estimates g of g are computed using stan- stimulus, at time t, is simply a linear compound of the
dard least-squares and are used to calculate the esti-
parameter estimates weighted by the values of the
mates hi(t) of the responses hi(t).
basis functions at t [from Eq. (2) hi(t) 5 Sg i j · bj(t)].
In the present context we are interested in making
statistical inferences about (i) responses to any stimu- This means that we can use standard results to obtain
lus type and (ii) differential responses. To test for any the standard error of the estimated response at t, again
response we can create an SPM5F 6 testing for all the as a continuous function of time SE5hi(t)6. The Appendix
effects in [X1 X2 . . .], treating the constant and other (A.1) describes how to do this. An alternative character-
specified terms as confounds. P values, corrected for the ization, of the reliability of the estimated response,
volume analyzed, are then assigned to maxima in the would be to use confidence intervals. The confidence
ensuing SPM5F 6 using the theory of Gaussian fields intervals follow simply from the standard error and
(Worsley, 1994). their derivation is presented in the Appendix (A.2). In
The final step is to assess differential responses. In this paper we use the standard error.
general this involves rearranging the design matrix so Finally we can estimate the standard error of the
that we can distinguish between the evoked hemody- timing of the response if we assume that the error in
namic response common to all types of stimulus and the hi(t) is attributable to a stochastic error in timing t*,
differences among these responses. The common re- where each estimate of hi(t) is a realization of hi(t 1 t*).
sponses are then treated as confounds and a new The Taylor expansion
SPM5F 6 is computed that tests for, and only for, the
differences. The common or main effect is simply Xc 5 hi(t) 5 hi(t 1 t*) 5 hi(t) 1 t* · dhi(t)/dt 1 · · ·
[X1 1 X2 1 . . .], the sum of all the stimulus-type-
specific effects. The differences Xd are given by orthogo-
allows one to posit
nalizing X with respect to Xc 5 X 2 Xc · pinv(Xc) · X . In
the case of just two stimulus types, Xd 5 X1 2 X2. This
‘‘partitioning’’ of the design matrix into commonalities SE5hi(t)6 < SE(t*) · dhi(t)/dt. (5)
and differences leads to an equivalent linear model
The standard error of these temporal perturbations is
y 5 [Xd Xc 1 G] · g 1 e (4) equivalent to the standard error of any estimated
latency li (i.e., SE5li 6 5 SE(t*)), giving
and the SPM5F 6 tests for the effects in Xd, treating the
remaining partitions as confounds. In the present SE5li 6 < SE5hi(t)6/(dhi(t)/dt). (6)
example we can use a simpler approach and assess the
differences directly using the original model [Eq. (3)] We will use this result in the next section to estimate
and a linear compound or contrasts of the parameter the temporal precision of event-related fMRI and to
estimates to give a SPM5t6. This is because we have make inferences about differences in response laten-
used a small and largely orthogonal basis set, where cies.
EVENT-RELATED fMRI 33
AN ILLUSTRATIVE APPLICATION The basis functions used in these analyses are shown
in Fig. 1. The stimulus functions u1(t) and u2(t) were set
Experimental Design and Data Acquisition to unity at the presentation of each stimulus type and
In this section we apply the theory presented above zero elsewhere. The SPM5F 6 reflecting the significance
to a fMRI time-series obtained from three normal of evoked responses modeled in X1 and X2 (or more
subjects during exposure to visually presented words formally, testing the null hypothesis that all h1(t) and
that were either (i) novel to the experiment or (ii) had h2 (t) were jointly zero) is shown in Fig. 2 with the
been studied in an encoding task prior to scanning. The design matrix. The left-hand side of the design matrix
data were acquired at 2 T using a Magnetom VISION comprises the explanatory variables X1 and X2. The
(Siemens, Erlangen) whole-body MRI system, equipped remaining columns contain the constant 1 used to
with a head volume coil. Contiguous multislice T2*- estimate g0 and other effects designated as confounds
weighted fMRI images were obtained with a gradient (low-frequency artifacts). The partitions were arranged
echo-planar sequence using an axial slice orientation to yield subject-specific estimates. The SPM5F 6 has been
(TE 5 40 ms, TR 5 3.22 s, 64 3 64 3 16 matrix size, thresholded (P , 0.05 corrected for multiple compari-
3 3 3 3 3 mm voxels). After discarding the first few sons) and shows widespread responses, notably in the
scans (to allow for magnetic saturation effects) the striate, extrastriate, medial parietal, anterior cingu-
time-series comprised 480 volume images, per subject, late, right dorsolateral prefrontal cortex, left lingual
with 3-mm isotropic voxels. Words were presented and parahippocampal gyri, left sensorimotor cortex
every 16 s in a pseudorandom order. The subject was (hand area), and left inferior postcentral gyrus (BA 40).
asked to discriminate novel from studied words by a It can be seen that the most intense responses were
keypress using the middle or index finger of the right elicited in the sensorimotor area associated with the
hand. executive components of the subjects’ finger move-
ments. Other extremely significant responses were
Data Preprocessing observed in bilateral basal posterior temporal cortices.
An example of an estimated evoked response, to old
The data were analyzed with SPM97 (Wellcome words, is shown in Fig. 3 for a voxel in the left
Department of Cognitive Neurology, http://www.fil- sensorimotor area, in terms of the adjusted data (dots)
.ion.ucl.ac.uk/spm). The time-series were realigned, and the estimated or fitted responses (solid lines). The
corrected for movement-related effects, and spatially standard error was computed as described in the
normalized into the standard space of Talairach and Appendix and the broken line represents plus and
Tournoux (1988) using the subject’s coregistered struc- minus one standard error. There are several important
tural T1-weighted MRI scan (Friston et al., 1995c, things to note about this characterization: (i) The
1996). The data were spatially smoothed with a 4-mm standard error seems smaller than one might have
isotropic Gaussian kernel and temporally smoothed predicted on the basis of the scatter of the adjusted
with a 4-s Gaussian kernel. data, because the estimated response at any single
point in time is based on data from all points in time.
Event-Related Responses This is a feature of using basis functions, as opposed to
The data were analyzed using the design matrix in simply taking the average in some limited time win-
Eq. (3) to assess the effect of either stimulus type and to dow. In other words by using a parsimonious model of
provide the stimulus-specific response estimates h1(t) expected responses, the responses can be estimated
and h2(t). The analysis of differential responses em- with a greater degree of precision than when using a
ployed the appropriate contrasts, testing for differences less specified model. (ii) It may be noted that the
in response magnitude and differences in response standard error of the response is zero when the esti-
latency. The first analysis gives a SPM5F 6 testing for the mated response is itself zero. Although this may seem
significance of an event-related response to one or both counterintuitive it is entirely sensible: Although the
stimulus types. The second analysis generates SPM5t6s response may actually differ from zero between stimuli,
reflecting the significance of any differential responses. these effects are modeled by other components of the
We modeled the response to stimuli separately for each design matrix (e.g., the low-frequency confounds). (iii)
subject. This means that the parameters that identify It appears as if the evoked response starts before the
the associated response functions were estimated inde- onset of the stimulus. There are two factors that relate
pendently for each subject, allowing us to assess the to this: The first is that the data (and the design matrix)
reproducibility of the form of differential responses were temporally smoothed in accord with the matched
from subject to subject. The data were normalized by filter theorem (Friston et al., 1995a). This convolution
dividing by the whole brain mean and multiplying by smears the initial responses into the prestimulus pe-
100. The resulting effects are then expressed as a riod. Second, the onset of the stimulus was defined in
percentage of whole brain mean. terms of the start of volume acquisition. In our mul-
34 FRISTON ET AL.
FIG. 2. (Left) SPM5F 6 testing for the significance of event-related responses to either word type. This is a maximum intensity projection of
a statistical image of the F ratio, following a multiple regression analysis at each voxel. The format is standard and provides three orthogonal
projections in the standard space of Talairach and Tournoux (1988). The grayscale is arbitrary and the SPM5F 6 has been thresholded at P ,
0.05 (corrected). (Right) The design matrix used in this analysis. The design matrix comprises the explanatory variables in the general linear
model. It has one row for each of the 3 3 480 scans and one column for each explanatory variable or effect modeled. The left-hand columns
contain the explanatory variables of interest vij(t) for each subject, where vij(t) is word stimulus function ui(t) convolved with the basis function
bj(t) in Fig. 1. The remaining columns contain covariates or effects of no interest designated as confounds. These include (left to right) a
constant term and periodic (discrete cosine set) functions of time to remove low-frequency artifacts with a high-pass cutoff period of 64 s.
Differential Responses
FIG. 3. Fitted response in the left sensorimotor area to previ- The subsequent analyses identified significant differ-
ously seen words. This voxel represents the maximum of the SPM5F 6 ences in several regions including the visual, inferotem-
of the previous figure. The fitted response (solid line) is shown with poral, and prefrontal cortices. We will concentrate here
the adjusted data (dots) plus and minus one standard error of the
fitted response (broken lines). The adjusted data are the original data
on two examples that demonstrate the various forms
with confounds (and the response to previously seen words) removed. that these differences can take. In the first we focus on
The data are plotted as a function of peri-stimulus time. a result that replicated not only over all three subjects
EVENT-RELATED fMRI 35
FIG. 4. (Left) SPM5F 6 thresholded at P , 0.05 (corrected) superimposed on a structural MRI testing for a significant response to either
word type. The posterior regions highlighted correspond to the lingual and parahippocampal gyri. (Right) Event-related responses at the voxel
marked by the cross-hairs on the SPM5F 6; (216, 244, 24 mm). The evoked responses are plotted in terms of estimated responses (solid lines)
and their standard errors (broken lines) as a function of time following stimulus onset.
but was expressed bilaterally in homologous ventrolat- the response to this stimulus, in this area, and com-
eral prefrontal regions. In these regions, again in all pares very favorably to the repeat time of 3.22 s. It is
subjects, novel words elicited strong activations whereas possible to make explicit inferences about differential
words that had been seen previously evoked negative onset latency using the standard errors of the timing as
responses or deactivations (although in the third sub- described in the Appendix (A.3). The examples given
ject the response to old words was more biphasic). The above demonstrate that the approach presented here is
results for each of the three subjects are shown sepa- sensitive to differences in timing as well as to differ-
rately in Figs. 5, 6, and 7. In each figure the responses ences in the magnitude of response components. It
are shown in terms of the fitted response (solid line) should be noted that the differential response latencies
and standard error (broken lines). The inset corre-
observed are far in excess of those that might be
sponds to the SPM5t6 (after transformation to a SPM5Z 6)
predicted by electrophysiological studies.
showing all voxels above a threshold of P 5 0.05
(uncorrected) on a standard structural MRI (coronal
section). The cross-hairs locate the voxel whose re- DISCUSSION
sponse is depicted in the graph. For example in Fig. 5
this voxel come from the inferior frontal gyrus (BA 37). We have presented an approach to characterizing the
This differential effect is reasonably significant Z 5 difference between hemodynamic responses in fMRI
4.18 (P , 0.001 uncorrected). The Z scores for the that are elicited by different sorts of stimuli. This
voxels in the remaining subjects were 3.10 (P 5 0.001 approach is based on a linear convolution model and
uncorrected, Fig. 6) and Z 5 3.59 (P , 0.001 uncor- standard inferential statistics. In particular we have
rected, Figure 7). modeled evoked responses and their differences in
Contrast the sort of differences in the previous sec-
terms of basis functions of peri-stimulus time. This
tion with those identified when testing for differential
facilitates a characterization of the response profiles
latencies. Figure 8 shows the responses of a voxel in the
that has a relatively high effective temporal resolution.
fusiform gyrus (238, 278, 210 mm, Z 5 4.14, P , 0.001
uncorrected). As above, this differential effect was We were able to demonstrate differences in both the
bilaterally represented and shows that novel words magnitude and in the latency of hemodynamic re-
evoked a much earlier response than previously seen sponses to visually presented words, depending on
words. In this example response latencies were as- whether or not the words had been processed prior to
sessed by the time to reach a third of the response scanning. As a general point it is worth emphasizing
maximum. The differential latency (about 3 s) is rela- that ‘‘response’’ in this paper refers to a stereotyped
tively large compared with the standard errors of the hemodynamic transient that is expressed relative to
timing of the responses. The black bars in Fig. 8 show the prestimulus baseline. Thus, as with event-related
these standard errors according to Eq. (6). The stan- potentials, the response is hard to interpret without
dard error for previously seen words was only 299 ms. reference to what the brain was doing during the
This can be taken as the effective temporal resolution of baseline.
36 FRISTON ET AL.
FIG. 5. (Inset) SPM5Z 6 thresholded at P , 0.05 (uncorrected) superimposed on a structural MRI testing for a significant differential
response in the first subject. The highlighted areas correspond to the inferior frontal sulci. (Graph) Event-related responses at the voxel
marked by the cross-hairs in the inset; (238, 22, 24 mm). The evoked responses are plotted in terms of estimated responses (solid lines) and
their standard error (broken lines) as a function of time following stimulus onset. Old denotes responses to previously seen words, and new
denotes responses to novels words.
The Basis Functions number of scans, not the repetition time. This is an
important consideration when choosing the basis func-
A key aspect of the approach presented here is the
tions because it is possible to properly model temporal
use of basis functions of peri-stimulus time. There are
frequencies that are higher than would normally be
clear advantages to using basis functions, compared to
simply averaging trial-based responses. Not least among allowed, given the limitations on temporal sampling
these is the facility to increase effective temporal imposed by the repetition time. Of course if one has
resolution well above the repetition time and indeed restricted sampling to a small number of discrete
estimate responses as a function of continuous time. A points in peri-stimulus time, then the basis set should
more fundamental advantage is that it does not require be chosen to reflect that fact.
the scans to be acquired with any fixed temporal The choice of basis functions is clearly a function of
relationship to the stimuli (this is because there is no the nature of the data acquired and the questions being
post hoc binning of the data for averaging). This may be asked of them. In general we have found that more
important in instances where the stimuli are not under robust results obtain when the basis set is small and
experimental control (e.g., perceptual reversal during temporally compact, especially when the interstimulus
presentation of ambiguous figures). It is worthwhile interval is short. In this paper we have used perhaps
considering that, using basis functions, the maximum the simplest set possible: a synthetic hemodynamic
effective temporal resolution is the interstimulus inter- response function and its derivative. This precludes
val divided by the total number of scans, although in response forms that are substantially more protracted
this instance there will be only one observation per than the normal response to a short-lived burst of
peri-stimulus time point. The key point though is that neuronal activity but was sufficient to give meaningful
the effective resolution is generally related to the results in the present study. There are other situations
EVENT-RELATED fMRI 37
FIG. 6. As for Fig. 5 but showing differential responses in the second subject.
(e.g., sentence processing) when the neuronal time that can arise is when one looks at the parameter
constants may call for more comprehensive sets of basis estimates in isolation (either in terms of estimated
functions that span longer periods of time (e.g., Vander- responses or inferences using the SPM5t6). In the pre-
berghe et al., 1997). In the context of characterizing sent paper we avoided linear dependence among the
phasic events we are currently exploring the general basis functions by using a sufficiently long interstimu-
approach of using the partial derivatives of a ‘‘canoni- lus interval and by using one function and its deriva-
cal’’ response function. In this framework an arbitrary tive for each stimulus type. In doing this we were able
hemodynamic response function f is selected with a to use the SPM5t6 with impunity.
small number of model parameters pi f (t, p1, p2, . . .).
The basis set would then comprise f, f/t, f/p1, Response Latencies
f/p2, . . . . The simplest version of this has been used
in this paper (i.e., f and f/t). The advantage of this One component of the work presented here is the
approach is that each modeled effect in the design characterization of differential responses in terms of
matrix has some physical meaning. Consequently con- response latency and the associated standard error.
trasts of the associated parameter estimates can be The standard errors of response onsets (as judged by
interpreted. We have used this above to characterize the time to reach a third of the maximal response)
differential responses in terms of differences in magni- suggest that in some instances fMRI can discriminate
tude (testing for differences among the f ) and differ- between dynamics on a 100-ms timescale despite rela-
ences in latency (differences among the f/t). We do, tively long repetition times. This is very encouraging
however, appreciate that differential responses may be but should not be confused with the notion that fMRI
very complicated and do not necessarily conform to this can be used to look at response latencies among differ-
simple dichotomous classification. ent areas (as used in electrophysiology to infer the
A final point, in relation to the basis set employed, is temporal sequence of areas that are recruited during
one of collinearity or correlations among the basis the promulgation of neuronal dynamics from one area
functions. Generally this is not problematic from the to the next). fMRI cannot be used to do this because of
point of view of statistical inference. Collinearity within the endogenous regional variability in the latency of
a design matrix partition (e.g., interesting effects or hemodynamic responses to underlying changes in neu-
confounds) has no effect on the ensuing statistical ronal activity (not to mention artifactual difference due
quotients when using the F statistic. The only problem to the timing of slice acquisitions) (DeYoe et al., 1992).
38 FRISTON ET AL.
FIG. 7. As for Fig. 6 but showing differential responses in the third subject.
However, it is in principle possible, using the methods another instance of the usefulness of the SPM5F 6. The
outlined above, to make inferences about the difference importance of the SPM5F 6, as opposed to SPM5t6 or t
in differential response latencies between areas (i.e., maps, is that it reflects the significance of a whole set of
stimulus 1 incurs a differential response latency of 650 parameter estimates, in this instance the collection of
ms, relative to stimulus 2, in area A but only 15 ms in coefficients that describe the hemodynamic responses
area B). This would constitute a region 3 stimulus- [hi(t)] or their differences (not reported in this paper).
specific response latency interaction that would be We envisage that the SPM5F 6 will find an increasing
independent of the (main) effect of region-specific differ- role in fMRI as the models of hemodynamic responses
ences. become more sophisticated and the number of param-
We have elected to characterize latency differences in eters increases. Some device is required to make an
terms of the standard error of the onset time as opposed inferences about these parameters en masse. The
to the standard error of the response itself. This is in SPM5F 6 is one such device.
contradistinction to some techniques in electrophysiol- It should be noted that the model employed to create
ogy where the differential latency is defined in terms of the SPMs was a linear convolution model relating a set
when the responses can be shown to be significantly of stimulus functions to the hemodynamic response. We
different (e.g., Thorpe et al., 1996). The fundamental have demonstrated previously (Friston et al., 1997)
distinction here is between defining an ‘‘onset time’’ that there are significant nonlinear components in
based on the parameter estimates of the response (our hemodynamic responses; however, these are expressed
approach) and defining it in terms of the standard error when stimuli are presented close together in time (such
of the parameter estimates (the alternative based on that the response to one stimulus is modulated by the
the demonstration of a significant difference). We have response to a preceding stimulus). In the context of
adopted our approach for the obvious reason that the event-related fMRI, with reasonable interstimulus in-
estimate of onset time should be a function of the tervals, these nonlinear effects can be discounted and
expected response, not the error of its measurement. one would normally be quite comfortable with a linear
model of the sort described here.
Statistical Considerations
Neurobiological Interpretation
One aspect of the techniques presented in this paper
is the use of the SPM5F 6 to make inferences about the Throughout this paper we have referred to the input
significance of responses and their differences. This is ui(t) as a stimulus function and y(t) as the hemody-
EVENT-RELATED fMRI 39
FIG. 8. As for the previous figures (Figs. 5–7) but in this instance showing an example where responses (of a single subject) differ in terms
of their latencies. The evoked responses are plotted in terms of (i) estimated responses (solid lines), (ii) their standard error (broken lines), and
(iii) the standard error of the time to third peak response (bars). This is for a voxel in the left fusiform gyrus (238, 278, 210 mm).
namic response. One can conceptualize the components responses observed can be assigned to differential
of the mapping from stimulus to response in terms of (i) processing at a neuronal level.
a transformation of the stimulus into evoked neuronal
dynamics and (ii) a translation of these neuronal Conclusion
changes into hemodynamic responses. If one demon-
strates a significant difference among hemodynamic We have presented a simple extension to previous
responses in the same area, then there are two explana- models of evoked responses in event-related fMRI that
tions: (i) The stimuli elicited different patterns of allows for the characterization of differences among
neuronal activity or (ii) the hemodynamic response to stimulus-specific responses using standard techniques
neuronal transients has itself changed. Clearly we developed for the analysis of functional neuroimaging
have assumed that the first is the proper explanation. data.
However, it is possible that the relationship between
neuronal dynamics and hemodynamic response may
APPENDIX
also have changed, for example, changes in cerebral
physiology associated with hyperventilation during A.1. The Standard Error of a Linear Compound of
anxiety-provoking stimuli relative to neutral stimuli Parameter Estimates for Serially Correlated Data
(here the resulting changes in global cerebral perfusion
and relative oxygenation may be expressed in terms of In this section we derive the standard error for the
a change in the hemodynamic responsiveness to under- estimated response hi(t) as a function of response
lying neuronal activation, i.e., a subtle change in the latency t. From Eq. (2) we have hi(t) 5 Sgi j · bj(t).
hemodynamic response function). On balance though, Because we know both the standard error of the
given that the stimuli are carefully chosen and well parameter estimates gi j and the values of the basis
matched, in terms of both their attributes and the functions at time t, bj(t), we can derive the standard
context in which they are presented, one might be quite error of the estimated response at t. In matrix notation
justified in assuming that the hemodynamic response hi(t) 5 bi(t) · g where, for example b1(t) 5 [b1(t), . . . ,
function in a given area will not change markedly. In bJ(t), 0, . . . , 0] and similarly for b2 , b3 and so on. The
such circumstances any difference in the hemodynamic standard error of the estimated responses is then given
40 FRISTON ET AL.