Granger causality

Let y and x be stationary time series. To test the null hypothesis that x does not Granger-cause y, one first finds the proper lagged values of y to include in a univariate autoregression of y:

${\displaystyle y_{t}=a_{0}+a_{1}y_{t-1}+a_{2}y_{t-2}+\cdots +a_{m}y_{t-m}+{\text{error}}_{t}.}$

Next, the autoregression is augmented by including lagged values of x:

${\displaystyle y_{t}=a_{0}+a_{1}y_{t-1}+a_{2}y_{t-2}+\cdots +a_{m}y_{t-m}+b_{1}x_{t-1}+\cdots +b_{q}x_{t-q}+{\text{error}}_{t}.}$

One retains in this regression all lagged values of x that are individually significant according to their t-statistics, provided that collectively they add explanatory power to the regression according to an F-test (whose null hypothesis is no explanatory power jointly added by the x's). In the notation of the above augmented regression, p is the shortest, and q is the longest, lag length for which the lagged value of x is significant.

The null hypothesis that x does not Granger-cause y is accepted if and only if no lagged values of x are retained in the regression.

Multivariate Granger causality analysis is usually performed by fitting a vector autoregressive model (VAR) to the time series. In particular, let ${\displaystyle X(t)\in \mathbb {R} ^{d\times 1}}$ for ${\displaystyle t=1,\ldots ,T}$ be a ${\displaystyle d}$-dimensional multivariate time series. Granger causality is performed by fitting a VAR model with ${\displaystyle L}$ time lags as follows:

${\displaystyle X(t)=\sum _{\tau =1}^{L}A_{\tau }X(t-\tau )+\varepsilon (t),}$

where ${\displaystyle \varepsilon (t)}$ is a white Gaussian random vector, and ${\displaystyle A_{\tau }}$ is a matrix for every ${\displaystyle \tau }$. A time series ${\displaystyle X_{i}}$ is called a Granger cause of another time series ${\displaystyle X_{j}}$, if at least one of the elements ${\displaystyle A_{\tau }(j,i)}$ for ${\displaystyle \tau =1,\ldots ,L}$ is significantly larger than zero (in absolute value).[10]

The above linear methods are appropriate for testing Granger causality in the mean. However they are not able to detect Granger causality in higher moments, e.g., in the variance. Non-parametric tests for Granger causality are designed to address this problem.[11] The definition of Granger causality in these tests is general and does not involve any modelling assumptions, such as a linear autoregressive model. The non-parametric tests for Granger causality can be used as diagnostic tools to build better parametric models including higher order moments and/or non-linearity.[12]

Previous Granger-causality methods could only operate on continuous-valued data so the analysis of neural spike train recordings involved transformations that ultimately altered the stochastic properties of the data, indirectly altering the validity of the conclusions that could be drawn from it. In 2011, however, a new general-purpose Granger-causality framework was proposed that could directly operate on any modality, including neural-spike trains.[23]

Neural spike train data can be modeled as a point-process. A temporal point process is a stochastic time-series of binary events that occurs in continuous time. It can only take on two values at each point in time, indicating whether or not an event has actually occurred. This type of binary-valued representation of information suits the activity of neural populations because a single neuron's action potential has a typical waveform. In this way, what carries the actual information being output from a neuron is the occurrence of a “spike”, as well as the time between successive spikes. Using this approach one could abstract the flow of information in a neural-network to be simply the spiking times for each neuron through an observation period. A point-process can be represented either by the timing of the spikes themselves, the waiting times between spikes, using a counting process, or, if time is discretized enough to ensure that in each window only one event has the possibility of occurring, that is to say one time bin can only contain one event, as a set of 1s and 0s, very similar to binary.

One of the simplest types of neural-spiking models is the Poisson process. This however, is limited in that it is memory-less. It does not account for any spiking history when calculating the current probability of firing. Neurons, however, exhibit a fundamental (biophysical) history dependence by way of its relative and absolute refractory periods. To address this, a conditional intensity function is used to represent the probability of a neuron spiking, conditioned on its own history. The conditional intensity function expresses the instantaneous firing probability and implicitly defines a complete probability model for the point process. It defines a probability per unit time. So if this unit time is taken small enough to ensure that only one spike could occur in that time window, then our conditional intensity function completely specifies the probability that a given neuron will fire in a certain time.