In this paper, we tested the time-invariance and linearity of SCRs as these constitute two central assumptions for SCR analysis both in classical methods (Boucsein, 1992
) and in model-based strategies, for example in a general linear convolution framework (Bach et al., 2009
). We conclude that the time-invariance assumption seems to be met. While there was trend-level evidence for repetition suppression even at ISIs of 9 s, we found no evidence that this suppression depended on ISI and therefore no convincing evidence for nonlinear interactions among overlapping SCRs, although CNS mechanisms might lead to repetition suppression independent of peripheral factors.
We previously reported residual variance between 20% and 40% on orienting responses to loud noise and aversive pictures, and could extend this now to defensive reactions and responses to events that are rendered salient by experimental instruction alone. In a visual detection task, we included a baseline period without target stimuli that was used to estimate baseline fluctuations. This exceeded the residual variance during evoked responses by almost a factor two. That is, it appears that spontaneous fluctuations are suppressed when SCRs are elicited by external events. This suggests that a major component of residual variance during evoked responses stems from spontaneous fluctuations (i.e. the components that contribute to baseline variance). In other words, spontaneous or non-specific fluctuations are more than a sufficient explanation for variability in evoked SCRs, and the time-invariance assumption can be regarded as met for practical purposes. This generalises to recordings from foot and middle phalanx of the fingers where very similar results were obtained.
Across participants, the estimated response functions looked comparable in different experimental paradigms. On this basis, we were able to estimate a canonical response function (CRF), derived from 1278 SCRs from 64 individuals. This was extended to form an informed basis set that can account for between-subject differences in response shape, and for between-condition differences; such as slightly later responses to aversive pictures (see ). In comparison with previously reported response functions (Lim et al., 1997; Alexander et al., 2005
) our CRF has a much longer recovery time (see ). It is however encouraging to note that our onset latency and peak are comparable with Lim et al. (1997)
. This is particularly important because their data were sampled at a fixed interval and adequately over-sampled, whereas we used a suboptimal (variable) sampling rate. The much shorter latency of the response function proposed by Alexander et al. (2005)
is due to the fact that their function does not include the time due to sensory processing and sudomotor transmission (which occurs prior to the measurable SCR response). It is worth noting that our CRF can also be used for foot and finger recordings without compromising model fit, indicating that the between-subject variance in response shape within one recording site dominates over the variance caused by latency differences between recording sites. Finger and foot recordings might be advantageous as thenar/hypothenar recordings are prone to movement artefacts. Recordings from distal phalanges of the fingers have been reported in the literature, and although we only acquired data from the middle phalanx, one might tentatively speculate that the time-invariance property and the canonical response function also apply at the distal phalanx.
Also, we found the linearity principle seems to hold, providing one allows for categorical differences in the underlying neuronal response to an initial stimulus, relative to subsequent stimuli. Between 3 s and 10 s after an event onset, responses to a second event of the same type were attenuated. Although this effect was not statistically significant, it speaks to a classical repetition suppression that may be useful to model in some designs (see below). Crucially, however, there was no effect of ISI, suggesting that the SCR to the second stimulus does not depend on the overlap with the SCR to the first. This is consistent with a linear model, provided the model incorporates simple repetition effects. It seems plausible that this kind of repetition effect reflects central adaption to the stimulus (for an overview see Boucsein, 1992
) but it could also be due to peripheral (e.g. sympathetic nerve) adaptation. Previous work (Bach et al., 2009
) suggested that the response shape might be different at different ISIs but we could not confirm this in the present study. This does not rule out non-linearities in the SCR response function, but they are probably difficult to detect without a clear hypothesis about their form.
Repetition effects could be an issue for any form of SCR analysis, be it semi-visual or mathematical. Here, we suggest how they can be dealt with in the framework of a linear convolution model (Bach et al., 2009
). If events engage a responsive system that has systematic variability, such that distinct levels of repetition are associated with different response amplitude, these levels can be encoded in different regressors (or, in classical analysis, in distinct levels of the experimental factor) (see Friston et al., 1998
for a discussion in the context of fMRI responses). Second, if there are systematic non-linearities in the responsiveness of the system, one can explicitly model these within the linear convolution framework using a Volterra series. Here, non-linearities are parameterised as coefficients of second and higher order Volterra kernels; together with the coefficients of the standard response function (first-order kernels). However, this is only necessary for more complex paradigms where there is no experimental control over repetition effects, or if we are specifically interested in such non-linearities, such as in studies of peripheral adaptation. Further work is needed to validate such models for the analysis of SCR and can be motivated by the fact that response habituation implies that the system we are modelling has memory.
An interesting possibility arises from the individual response functions depicted in , where aversive pictures evoke responses slightly later than other stimulus classes. There is no reason why the peripheral output system should exhibit a different response to one stimulus class than to any other. Given that images contain much more information than any other of the studied stimuli, this longer latency may reflect longer processing time in the central nervous system. The twist is that it might be possible to estimate characteristics of central nervous function by deconvolving the observed signal given a canonical response function. Recently proposed deconvolution/decomposition methods (Alexander et al., 2005; Benedek and Kaernbach, in press
) have addressed this issue by attempting to recover the sudomotor time series; however they did not provide a model for how this sudomotor activity relates to processes in the CNS. We will pursue this in future work.
We emphasise that the present approach not only assists in analysis of SCRs, but also connects to the broader literature on psychophysiological responses and sets out a framework of how analysis assumptions and research questions can be formulated mathematically. For example, it has been proposed that a higher rate of biphasic as opposed to monophasic SCRs might differentiate defensive from orienting responses (Uno and Grings, 1965
). In order to test this assumption, instead of visually scoring and counting responses (see Boucsein, 1992
), the present framework allows one to parameterise condition-specific response functions and test for their differences.
In conclusion, we have investigated two central assumptions of SCR analysis that are usually not made explicit, but which are entailed for example in a general linear convolution framework (Bach et al., 2009
). The time-invariance assumption could be supported, and there was no evidence for a violation of the linearity assumption, given repetition suppression is modelled. We suggest that repetition suppression does not preclude SCR analysis; indeed, non-linear methods can be incorporated into the linear framework, even if we do not have control over ISI. Thus, explicit modelling methods appear to be more powerful for SCR analysis than classical, semi-visual methods. Our findings pertain to many models that could be used for SCR. Indeed one might anticipate more refined models in the future, which might even relax the need for filtering the data prior to analysis. As noted by one of our reviewers, an interesting challenge here is to accommodate fluctuations in tonic skin conductance level, which can be substantially larger in amplitude than the SCR itself.