Roebroeck Seth Valdes-SosaTable 1: Types of Influence def<strong>in</strong>ed by absence of the correspond<strong>in</strong>g <strong>in</strong>dependence relations.See text for acronym def<strong>in</strong>itions.Global( All horizons)Local(Immediate future)ContemporaneousStrong(Probability Distribution)By absence ofstrong, conditional, global <strong>in</strong>dependence:X 2 (t)SCGi X 1 (t)||X 3 (t)By absence ofstrong, conditional, local <strong>in</strong>dependence:X 2 (t)SCLi X 1 (t)||X 3 (t)By absence ofstrong, conditional, contemporaneous<strong>in</strong>dependence:X 2 (t)SCCi X 1 (t)||X 3 (t)Weak(Expectation)By absence ofweak, conditional, global <strong>in</strong>dependence:X 2 (t)WCGi X 1 (t)||X 3 (t)By absence ofweak, conditional, local <strong>in</strong>dependence:X 2 (t)WCLi X 1 (t)||X 3 (t)By absence ofweak, conditional, contemporaneous<strong>in</strong>dependence:X 2 (t)WCCi X 1 (t)||X 3 (t)this def<strong>in</strong>ition is appropriate for po<strong>in</strong>t processes, discrete and cont<strong>in</strong>uous time series,even for categorical (qualitative valued) time series. The only problem with this formulationis that it calls on the whole probability distribution and therefore its practicalassessment requires the use of measures such as mutual <strong>in</strong>formation that estimate theprobability densities nonparametrically.As an alternative, weak concepts of <strong>in</strong>fluence can be def<strong>in</strong>ed based on expectations.Consider weak conditional local <strong>in</strong>dependence <strong>in</strong> discrete time, which is def<strong>in</strong>ed:E [ X 1 [t + ∆t]| X 1 [t,−∞], X 2 [t,−∞], X 3 [t,−∞]] = E [ X 1 [t + ∆t]| X 1 [t,−∞], X 3 [t,−∞]](7)When this condition does not hold we say X 2 weakly, conditionally and locally <strong>in</strong>fluences(WCLi) X 1 given X 3 . To make the implementation this def<strong>in</strong>ition <strong>in</strong>sightful,consider a discrete first-order vector auto-regressive (VAR) model for X = [X 1 X 2 X 3 ]:X [t + ∆t] = AX [t] + e[t + ∆t] (8)For this case E [ X[t + ∆t]| X[t,−∞]] = AX [t], and analyz<strong>in</strong>g <strong>in</strong>fluence reduces to f<strong>in</strong>d<strong>in</strong>gwhich of the autoregressive coefficients are zero. Thus, many proposed operationaltests of WAGS <strong>in</strong>fluence, particularly <strong>in</strong> fMRI analysis, have been formulated as testsof discrete autoregressive coefficients, although not always of order 1. With<strong>in</strong> the samemodel one can operationalize weak conditional <strong>in</strong>stantaneous <strong>in</strong>dependence <strong>in</strong> dis-84
Causal analysis of fMRIcrete time as zero off-diagonal entries <strong>in</strong> the co-variance matrix of the <strong>in</strong>novations e[t]:Σ e = cov[X [t + ∆t]|X [t,−∞]] = E [︀ X [t + ∆t] X ′ [t + ∆t]|X [t,−∞] ]︀In comparison weak conditional local <strong>in</strong>dependence <strong>in</strong> cont<strong>in</strong>uous time is def<strong>in</strong>ed:E [Y 1 [t]|Y 1 (t,−∞],Y 2 (t,−∞],Y 3 (t,−∞]] = E [Y 1 [t]|Y 1 (t,−∞],Y 3 (t,−∞]] (9)Now consider a first-order stochastic differential equation (SDE) model for Y = [Y 1 Y 2 Y 3 ]:dY = BYdt + dω (10)Then, s<strong>in</strong>ce ω is a Wiener process with zero-mean white Gaussian noise as a derivative,E [Y[t]|Y(t,−∞]] = BY (t)and analys<strong>in</strong>g <strong>in</strong>fluence amounts to estimat<strong>in</strong>g the parametersB of the SDE. However, if one were to observe a discretely sampled versionX[k] =Y (k∆t) at sampl<strong>in</strong>g <strong>in</strong>terval ∆tand model this with the discrete autoregressive modelabove, this would be <strong>in</strong>adequate to estimate the SDE parameters for large ∆t, s<strong>in</strong>ce theexact relations between cont<strong>in</strong>uous and discrete system matrices are known to be:A = e B∆t = I + ∞ ∑︀Σ e = ∫︀ t+∆tti=1∆t ii! Bie Bs∑︀ ω e Bs dsThe power series expansion of the matrix exponential <strong>in</strong> the first l<strong>in</strong>e shows A to bea weighted sum of successive matrix powers B i of the cont<strong>in</strong>uous time system matrix.Thus, the Awill conta<strong>in</strong> contributions from direct (<strong>in</strong> B) and <strong>in</strong>direct (<strong>in</strong> i steps <strong>in</strong>B i )causal l<strong>in</strong>ks between the modeled areas. The contribution of the more <strong>in</strong>direct l<strong>in</strong>ks isprogressively down-weighted with the number of causal steps from one area to anotherand is smaller when the sampl<strong>in</strong>g <strong>in</strong>terval ∆t is smaller. This makes clear that multivariatediscrete signal models have some undesirable properties for coarsely sampled signals(i.e. a large ∆t with respect to the system dynamics), such as fMRI data. Critically,entirely rul<strong>in</strong>g out <strong>in</strong>direct <strong>in</strong>fluences is not actually achieved merely by employ<strong>in</strong>g amultivariate discrete model. Furthermore, estimated WAGS <strong>in</strong>fluence (particularly therelative contribution of <strong>in</strong>direct l<strong>in</strong>ks) is dependent on the employed sampl<strong>in</strong>g <strong>in</strong>terval.However, the discrete system matrix still represents the presence and direction of<strong>in</strong>fluence, possibly mediated through other regions.When the goal is to estimate WAGS <strong>in</strong>fluence for discrete data start<strong>in</strong>g from a cont<strong>in</strong>uoustime model, one has to model explicitly the mapp<strong>in</strong>g to discrete time. Mapp<strong>in</strong>gcont<strong>in</strong>uous time predictions to discrete samples is a well known topic <strong>in</strong> eng<strong>in</strong>eer<strong>in</strong>gand can be solved by explicit <strong>in</strong>tegration over discrete time steps as performed <strong>in</strong> (11)above. Although this def<strong>in</strong>es the mapp<strong>in</strong>g from cont<strong>in</strong>uous to discrete parameters, itdoes not solve the reverse assignment of estimat<strong>in</strong>g cont<strong>in</strong>uous model parameters fromdiscrete data. Do<strong>in</strong>g so requires a solution to the alias<strong>in</strong>g problem (Mccrorie, 2003) <strong>in</strong>cont<strong>in</strong>uous stochastic system identification by sett<strong>in</strong>g sufficient conditions on the matrixlogarithm function to make Babove identifiable (uniquely def<strong>in</strong>ed) <strong>in</strong> terms of A.Interest<strong>in</strong>g <strong>in</strong> this regard is a l<strong>in</strong>e of work <strong>in</strong>itiated by Bergstrom (Bergstrom, 1966,(11)85