- Research article
- Open Access
Granger causality vs. dynamic Bayesian network inference: a comparative study
© Zou and Feng; licensee BioMed Central Ltd. 2009
Received: 17 December 2008
Accepted: 24 April 2009
Published: 24 April 2009
In computational biology, one often faces the problem of deriving the causal relationship among different elements such as genes, proteins, metabolites, neurons and so on, based upon multi-dimensional temporal data. Currently, there are two common approaches used to explore the network structure among elements. One is the Granger causality approach, and the other is the dynamic Bayesian network inference approach. Both have at least a few thousand publications reported in the literature. A key issue is to choose which approach is used to tackle the data, in particular when they give rise to contradictory results.
In this paper, we provide an answer by focusing on a systematic and computationally intensive comparison between the two approaches on both synthesized and experimental data. For synthesized data, a critical point of the data length is found: the dynamic Bayesian network outperforms the Granger causality approach when the data length is short, and vice versa. We then test our results in experimental data of short length which is a common scenario in current biological experiments: it is again confirmed that the dynamic Bayesian network works better.
When the data size is short, the dynamic Bayesian network inference performs better than the Granger causality approach; otherwise the Granger causality approach is better.
Based upon high throughput data, to reliably and accurately explore the network structure of elements (genes, proteins, metabolites, neurons etc.) is one of the most important issues in computational biology [1–6]. Currently, there are two main approaches which are often used to infer causal relationships  or interactions among a set of elements [8, 9]. One is the Granger causality approach [10, 11], and the other is the Bayesian network inference approach [12, 13]. The latter is often applied to static data. However, one can employ the dynamic Bayesian networks to deal with time series data for which the Granger causality has been solely developed. The Granger causality has the advantage of having a corresponding frequency domain decomposition so that one can clearly find at which frequencies two elements interact with each other.
Giving a multi-variable time series dataset, the Granger causality and dynamic Bayesian networks  can both be applied. The Granger causality notation, which was firstly introduced by Wiener and Granger [15, 16], proposed that we can determine a causal influence of one time series on another: the prediction of one time series can be improved by incorporating the knowledge of the second one. On the other hand, The Bayesian network  is a special case of a diagrammatic representation of probability distributions, called probabilistic graphical models [18–20]. The Bayesian network graph model comprises nodes (also called vertices) connected by directed links (also called edges or arcs) and there is no cycle in the graph. To learn the structure and the parameters for the Bayesian networks from a set of data, we should search the space(s) of all possible graph representations, and find out which structure is most likely to produce our data. If we have a scoring function (or likelihood function) which can determine the structure and parameter likelihood from the data, then the problem is to find the highest score (maximum likelihood) structure among all the possible representations.
The causal relationship derived from these two approaches could be different, in particular when we face the data obtained from experiments. Therefore it is of vital importance to compare these two causal inferring approaches before we could confidently apply them to biological data. By doing the comparison, one expects to find the advantages, performances and stabilities for each technique.
Adopting the most common existing methods to find the coefficients of the time series in both approaches in the literature, we compare the dynamic Bayesian network with the Granger causality both in the linear and nonlinear model. Interestingly, a critical point of the data length is found. When the data length is shorter than the critical point, the dynamic Bayesian network approach outperforms the Granger causality approach. But when the data length is longer, the Granger causality is more reliable. The conclusion is obtained via intensive computations (more than 100 computers over a few weeks). A biological data set of gene microarray is analyzed using both approaches, which indicates that for a data set with a short sampling length the dynamic Bayesian network produces more reliable results. In summary, we would argue that the dynamic Bayesian network is more suitable for dealing with experimental data.
To illustrate and compare the differences between the dynamic Bayesian network inference and the conditional Granger causality, a simple multivariate model with fixed coefficients, which has been discussed in many papers to test the Granger causality, is tested first. We then extend our comparisons to the more general case of the model with random coefficients, which requires considerable computational resources. More than 100 networked computers are used to perform the comparisons for more than a week. Both the Granger causality and the dynamic Bayesian network are applied to nonlinear models. Finally, we test our approach on a set of microarray data recently acquired from a comparison of mock and infected Arabidopsis leaf.
Synthesized data: linear case
where W1, W2, ⋯, W9 are uniformly distributed random variables in the interval [-1,1]. The randomly generated coefficients are also required to make the system stable. The stability can be tested by using the z-plane pole-zero method, which states if the outermost poles of the z-transfer function describing the time series are inside the unit circle on the z-plane pole-zero plot, then the system is stable.
The above toy model is then used to test the two different causality approaches: Bayesian network inference and Granger causality. They are applied with different sample sizes. For each sample size, we randomly generated 100 different coefficient vectors [W1, W2, ⋯, W9], which corresponds to100 different toy models in Example 1. For each different coefficient vectors model, we applied the same approach as in Example 1, using Monte Carlos method to construct 95% confidence interval for the Granger causality approach and chose high-confidence arcs (appearing in at least 95% of all samplings) for the Bayesian network inference approach. The total number of arcs (or causalities) is 500 (5 interactions for each realization) for each sample size. However we cannot expect to detect the maximum number of arcs in our system, since the coefficients are randomly generated, which could be significantly small.
Now we are in the position to find out why the dynamic Bayesian network is better than the Granger causality when the data length is short, and vise verse. In Figure 3B, we compare the performances on different coefficients (strength of interaction) for a fixed sample size of 900 (super-critical case). The x-axis shows the absolute value of coefficients, and y shows the corresponding causality (1 indicates positive causality and 0 indicates no causality). For visualization purposes, the figure for the Granger causality is shifted upward. From the five graphs, we can see that there is no difference between these two approaches if the coefficients are significant large (strong interactions with an absolute value of coefficients being greater than 0.15): both approaches can infer the correct connections. For most cases, the Granger causality approach performs with more stability when the coefficients are larger than 0.15, and the Bayesian network inference approach shows slightly more oscillations around this point. Hence we conclude that the Granger causality is less sensitive to the small value of the connection when the data length is large (see also the nonlinear case below).
We now compare the fitting accuracy of the two approaches, as shown in Figure 3C. We use the average mean-square error as a measurement of the fitting. Not surprisingly, the dynamic Bayesian network approach considerably outperforms the simple fitting algorithm in the Granger approach [15, 16], in particular when the data length is short.
In conclusion, when the data is a reasonable fit to the original model, the Granger causality works better. This is due to the fact that the Granger causality approach is more sensitive to a small value of the interactions. When the data length is short, the Bayesian approach can fit the data much more reliably and it outperforms the Granger approach.
Synthesized data: non-linear case
In real situations, all data should be nonlinear and a linear relationship as described above is only an approximation. To address the nonlinear issue, we turn our attention to kernel models. As we all know, any nonlinear relationship can be approximated by a series of kernel functions.
Therefore, all conclusions in the linear case are confirmed in the nonlinear model. In the literature , the result they obtained shows the same direction as we did here, which finds that the Granger causality performs better than the dynamic Bayesian network inference concerning a nonlinear kernel model of genetic regulatory pathways and for a sufficiently large sample size (2000 data points).
We first apply the dynamic Bayesian network inference approach on these two data sets. The two network structures for two cases are shown in Figure 7B. In the next step, the conditional Granger causality approach is applied. By using the bootstrapping method, we construct 95% confidence intervals as shown in Figure 7Cc. Finally, we can also obtain two network structures for two different cases shown in Figure 7Ca and Figure 7Cb. It is clearly seen that the globe patterns for the mock case and the infected case are different.
From the literature, there are three well known connections among the whole structure for the mock case. (1) It is known that GI alone is independent of the remaining six genes in the circuit. There should be no connection to and from the GI node (node 2 in the Figure 7) in our derived network. From Figure 7Ba and Figure 7Ca, we find that the dynamic Bayesian network inference method clearly picks this up, but the conditional Granger causality approach fails to detect this property. The Granger causality approach derived two false positive arcs which were connected to a GI node as shown in Figure 7Ca. (2) It is known that PRR7 and LHY share a feedback loop. In other words, there should be two directed arcs connected from node 1 (PRR7) to node 5 (LHY) and from node 5 to node 1. The network structures derived from both approaches are in agreement with this known relationship. (3) It is known that ELF4 has some interactions with both LHY and CCA1. There should be some connections between node 4 (ELF4) to node 5 (LHY), and between node 4 (ELF4) and node 6 (CCA1). From our derived structures, both approaches can detect these connections, which are in agreement with the known structure in the literature [24, 25].
According these three known relationships in the structure, we can find that the Bayesian network structure is in agreement with all three rules, but the network structure derived from the conditional Granger causality is not: two more false positive interactions are found. Again for a small sample size, the Bayesian network inference approach could be more reliable than the conditional Granger causality approach.
A fair comparison
In our results presented here, one of the key issues which is the cause of the critical point of the sampling size between the dynamic Bayesian approach and the Granger causality lies in the fact that a batch fitting approach is used in the Granger causality approach. One might argue that we could use the sequential fitting approach as in the Bayesian network to improve the performance of the Granger causality approach. This is certainly the case. However, due to the thousand publications in both topics , we simply adopted the most common approaches in the dynamic Bayesian network approach and the Granger causality. Developing one of the approaches, for example the Granger causality, so that it could always outperform the other is an interesting future research topic.
How long is long enough?
Although we have found the critical point of the two approaches, in practical applications, we have no certain idea where the critical point is. Hence, we still have to choose one of them to tackle the data. In molecular biology, we have to deal with a very limited data size; but in physiology, for example, neurophysiology, the data we record is usually very long. Hence one could argue that we use the dynamic Bayesian network in gene, protein or metabolite data, and apply the Granger causality to physiology data. The dynamic Bayesian network is more often reported in molecular biology, but the Granger causality has been very successfully applied in neurophysiological data  and fMRI. The result we chose to use was always chosen through experimental validation, as we did here for the plant data.
As we emphasized at the beginning, the advantage of the Granger causality over the dynamic Bayesian network is the frequency decomposition, which is usually informative when we deal with temporal data. For example, in neurophysiology data, we know the brain employs different frequency bands to communicate between neurons and brain areas [28, 29]. We would expect a similar situation to arise in genes, proteins and metabolites, although we lack a detailed analysis due to the limited data length. To this end, we have also presented frequency decomposition results in Appendix 1 (see Additional files 1 and 2) for the dynamic Bayesian network.
In our synthesized data, for both approaches, we did not find any false positive links in our experiments. However, there were a few false positive links found when we applied the conditional Granger causality and also partial Granger causality (data not shown, ) on the gene data. One might ask why this is the case; there are several different reasons. Firstly, the experimental data is not strictly stationary: it is a natural process and evolves with time. As a first approximation, we treat it as stationary. Of course, we could use ARIMA rather than ARMA model to fit the data in the Granger causality. Secondly, the seven gene network is only a small network embedded in complete and large network, so there are latent variables. Using the partial Granger causality  which was originally developed for eliminating latent variables, GI still has links with the other six genes. Whether the dynamic Bayesian network could do a better job in the presence of latent variables is another research topic.
The meaning of the found motifs
Two circuits are found: one with the mock plant and one with the infected plant. The plant rewires its circadian circuit after infection. Ignoring the issue of identifying the molecular mechanisms which control circuit rewiring, which is itself an interesting and challenging problem, we intend to discuss the functional meaning of the two circuits. To this end, we could assign a dynamics to the network and try to decipher the implications of the rewiring. Interestingly, we found that GI is recruited to save the network: if we leave GI as it is in the mock case, the whole network will rapidly converge to a fixed point state (a dead state). We will publish the results elsewhere.
Reasons for short size data
In our synthesized data, we test both short and long data samples and come to the conclusion that there is a critical size, at which the two approaches behave differently. In our experimental data, we only tested it for the short data set. Of course, as we mentioned above, in neurophysiological data, we have recordings of long time traces and the Granger causality is widely used there. However, we have to realize that all in vivo recordings are very dynamic and stationary of data will become a key issue once we apply both approaches to a long dataset. Furthermore, when the dataset is long, both approaches could do well and it is more difficult to find the difference between the two. Hence we have only compared the results for short data length in the experimental setup.
Reasons for small size of variables
In our synthesized data, we only used 5 variables to simulate a small interacting network; the number of variables could affect the result we derived. As expected, see also , the estimation of the Granger causality becomes unfeasible when the number of variables is large and the amount of the data sets is small. Hence, all results in the literature on estimating Granger causality are exclusive for small networks (around the order of 10), as we considered here. This is more or less true for dynamic Bayesian network inference as well. Extending the Granger causality and the dynamic Bayesian network inference to large networks is a challenging problem, even before we carry out the same comparison study on these two approaches as we did here.
In this paper, we carried out a systematic and computationally intensive comparison between the two network structures derived from two common approaches: the dynamic Bayesian network inference and the Granger causality. These two approaches are applied on both synthesized and experimental data. For synthesized data (both linear model and non-linear model), a critical point of the data length is found, and the result is further confirmed in experimental data. The dynamic Bayesian network inference performs better than the Granger causality approach, when the data length is short, and vice versa.
Causal influence measurement notation for time series was firstly proposed by Wiener-Granger. We can determine a causal influence of one time series on another, if the predication of one time series can be improved by incorporating the knowledge of the second one. Granger applied this notation by using the context of linear vector auto-regression (VAR) model of stochastic processes [30–33]. In the AR model, the variance of the prediction error is used to test the perdition improvement. For instance, assume two time series; if the variance of the autoregressive prediction error of the first time series at the present time is reduced by inclusion of past measurements from the second time series, then one can conclude that the second time series have a causal influence on the first one. Geweke [15, 16] decomposed the VAR process into the frequency domain, it converted the causality measurement into a spectral representation and made the interpretation more appealing.
The pairwise analysis introduced above can only be applied to bivairate time series. For more than two time series, a time series can have a direct or indirect causal influence to other time series. In this case, pairwise analysis is not sufficient or misleading for revealing whether the causal interaction between a pair is direct or indirect. In order to distinguish the direct and indirect causal affect, one introduces the conditional causality which takes account of the other time series' effect in a multivariate time series. In this paper, we used conditional causality to compare with the Bayesian network inference introduced before.
Linear conditional Granger causality
When the causal influence from Y to X is entirely mediated by Z, the coefficient b2iis uniformly zero, and the two autoregression models for two time series and three time series will be exactly same, thus we can get var(ε1t) = var(ε3t). We then can deduce F Y→X|Z = 0, which means Y can not futher improve the prediction of X including past measurements of Y conditional on Z. For var(ε1t) > var(ε3t) and F Y→X|Z = 0, we can say that there is still a direct influence from Y to X conditional on the past measurements of Z.
Non-linear conditional Granger causality
That is, the joint probability distribution is the product of the local distributions of each node and its parents. If node X i has no parents, its local probability distribution is said to be unconditional, otherwise it is conditional. This decomposition is useful for Bayesian networks inference algorithm to deal with the uncertain situation and incomplete data.
The detailed inductions of parameter estimations are given in the next chapter.
For learning the structure of the Bayesian network, one needs to search the space of all the possible structures and find out the best one which can be used to describe the input data, which is to maximum the conditional probability P(Data|θ, M) of data (Data) by give the parameters (θ) and the network structure (M). In order to balance the complex and concise of the structure, we can use BIC (Bayesian Information Criterion) as a scoring function, which includes an extra penalty term.
Parameter learning for linear Gaussian
The parameter learning part can be approached by fitting a linear-Gaussian model. The goal of learning parameter in Bayesian network is to estimate the mean and covariance of the conditional Gaussian distribution, thus we can deduce the parameters of μ, σ and W in the equation (19).
Thus the parameters in the Bayesian network can be learned from above two equations.
Parameter learning for non-linear Gaussian
We can also extend our linear model to a non-linear model like that for the Granger causality case. Suppose we have two variables which can be expressed as in equation (9). The kernel function is also chosen as described in equation (10) and equation (11).
where w is the connection weights between node X and it parents. It can be estimated by using the simple linear regression method.
There are two very different approaches to structure learning: one is constraint-based and the other is search and score algorithm. For the constraint-based algorithm, we start with a fully connected network and then remove the arcs, which are conditional independent. This has the disadvantage that repeated independence tests lose statistical power. For the latter algorithm, we perform a search on all possible graphs and select one graph which best describes the statistical dependence relationship in the observed data.
Unfortunately, the number of possible graphs increases super-exponentially with the number of nodes, so some search algorithms are required for overcoming this kind of complex problem rather than doing a exhaustive search in the space. There are several searching algorithms that can be applied; such as annealing search, genetic algorithm search and so on. The question could become easier if we know the total order of the nodes. The K2 algorithm allows us to find the best structure by selecting the best set of parents for each node independently. In the dynamic Bayesian networks, the order of nodes can be interpreted as the sequence of time lags represented for each node, so the K2 algorithm is applied for Bayesian network structure learning in this paper (see Appendix 2 in Additional file 1 for more details descriptions). The K2 algorithm tests parent insertion according to the order. The first node cannot have any parent, for other nodes, we can only choose the parent nodes which are behind it in this order. Then the scoring function can be applied to determine the best parent set, i.e. the one which gives the best score.
Where Data is the observed data, θ is the estimated value of the parameters, d is the number of parameters and N is the number of data cases. The term of is regarded as a penalty term in order to balance both simple and accurate structure representation.
Where j is the index of the nodes or variables in the Bayesian network, pa(j) is the set of parents of node j, and θ j are the parameters that define the conditional probability of Y j giving its parents.
In this paper, the Bayesian networks inference can then be approached by following procedure: initially, K2 algorithm is applied to search the space of possible graphs. For each possible structure, we can use the parameter learning algorithm to estimate the parameters of the networks. The BIC scoring function assigns a corresponding score through the estimated parameters and observed data set. The best network we can get is the highest score structures among all the possible graphs .
This work was supported by the CARMAN e-science project funded by the EPSRC (EP/E002331/1) and an EU grant (BION).
- Klipp E, Herwig R, Kowald A, Wierling C, Lehrach H: Systems Biology in Practice: Concepts, Implementation and Application. Weinheim: Wiley-VCH Press; 2005.View ArticleGoogle Scholar
- Feng J, Jost J, Qian M: Networks: From Biology to Theory. London: Springer Press; 2007.View ArticleGoogle Scholar
- Alon U: Network motifs: theory and experimental approaches. Nat Rev Genet 2007, 8(6):450–461. 10.1038/nrg2102View ArticlePubMedGoogle Scholar
- Tong AH, Lesage G, Bader GD, Ding H, Xu H, Xin X, Young J, Berriz GF, Brost RL, Chang M, Chen Y, Cheng X, Chua G, Friesen H, Goldberg DS, Haynes J, Humphries C, He G, Hussein S, Ke L, Krogan N, Li Z, Levinson JN, Lu H, Ménard P, Munyana C, Parsons AB, Ryan O, Tonikian R, Roberts T, Sdicu AM, Shapiro J, Sheikh B, Suter B, Wong SL, Zhang LV, Zhu H, Burd CG, Munro S, Sander C, Rine J, Greenblatt J, Peter M, Bretscher A, Bell G, Roth FP, Brown GW, Andrews B, Bussey H, Boone C: Global Mapping of the Yeast Genetic Interaction Network. Science 2004, 303: 808. 10.1126/science.1091317View ArticlePubMedGoogle Scholar
- Tsai TY, Choi YS, Ma W, Pomerening JR, Tang C, Ferrell JE Jr: Robust, Tunable Biological Oscillations from Interlinked Positive and Negative Feedback Loops. Science 2008, 321: 126. 10.1126/science.1156951PubMed CentralView ArticlePubMedGoogle Scholar
- Lee TI, Rinaldi NJ, Robert F, Odom DT, Bar-Joseph Z, Gerber GK, Hannett NM, Harbison CT, Thompson CM, Simon I, Zeitlinger J, Jennings EG, Murray HL, Gordon DB, Ren B, Wyrick JJ, Tagne JB, Volkert TL, Fraenkel E, Gifford DK, Young RA: Transcriptional Regulatory Networks in Saccharomyces cerevisiae. Science 2002, 298: 799. 10.1126/science.1075090View ArticlePubMedGoogle Scholar
- Pearl J: Causality: Models, Reasoning, and Inference. Cambridge: Cambridge Univ. Press; 2000.Google Scholar
- Albo Z, Di Prisco GV, Chen Y, Rangarajan G, Truccolo W, Feng J, Vertes RP, Ding M: Is partial coherence a viable technique for identifying generators of neural oscillations. Biological Cybernetics 2004, 90: 318. 10.1007/s00422-004-0475-5View ArticlePubMedGoogle Scholar
- Horton PM, Bonny L, Nicol AU, Kendrick KM, Feng JF: Applications of multi-variate analysis of variances (MANOVA) to multi-electrode array data. Journal of Neuroscience Methods 2005, 146: 22. 10.1016/j.jneumeth.2005.01.008View ArticlePubMedGoogle Scholar
- Guo S, Wu J, Ding M, Feng J: Uncovering interactions in the frequence domain. PLoS Computational Biology 2008, 4(5):e1000087. 10.1371/journal.pcbi.1000087PubMed CentralView ArticlePubMedGoogle Scholar
- Wu J, Liu X, Feng J: Detecting causality between different frequencies. Journal of Neuroscience Methods 2008, 167: 367. 10.1016/j.jneumeth.2007.08.022View ArticlePubMedGoogle Scholar
- Jansen R, Yu H, Greenbaum D, Kluger Y, Krogan NJ, Chung S, Emili A, Snyder M, Greenblatt JF, Gerstein M: A Bayesian Networks Approach for Predicting Protein-Protein Interactions from Genomic Data. Science 2003, 302: 449. 10.1126/science.1087361View ArticlePubMedGoogle Scholar
- Sachs K, Perez O, Pe'er D, Lauffenburger DA, Nolan GP: Causal Protein-Signaling Networks Derived from Multiparameter Single-Cell Data. Science 2005, 308: 523. 10.1126/science.1105809View ArticlePubMedGoogle Scholar
- Ghahramani Z: Learning Dynamic Bayesian Networks. Berlin: Springer Press; 2004.Google Scholar
- Geweke J: Measurement of Conditional Linear Dependence and Feedback Between Time Series. Journal of the American Statistical Association 1982, 79(388):907. 10.2307/2288723View ArticleGoogle Scholar
- Geweke J: Measurement of Linear Dependence and Feedback Between Multiple Time Series. Journal of the American Statistical Association 1982, 77(378):304. 10.2307/2287238View ArticleGoogle Scholar
- Jensen FV: An introduction to Bayesian networks. London: UCL Press; 1996.Google Scholar
- Bach FR, Jordan MI: Learning Graphical Models for Stationary Time Series. IEEE transactions on signal processing 2004, 52(8):2189. 10.1109/TSP.2004.831032View ArticleGoogle Scholar
- Buntine WL: Operations for Learning with Graphical Models. Journal of Artificial Intelligence Research 1994, 2: 159.Google Scholar
- Friedman N: Inferring Cellular Networks Using Probabilistic Graphical Models. Science 2004, 303: 799. 10.1126/science.1094068View ArticlePubMedGoogle Scholar
- Guo S, Seth AK, Kendrick KM, Zhou C, Feng J: Partial Granger Causality-Eliminating Exogenous Inputs and latent Variables. Journal of neuroscience methods 2008, 172(1):79. 10.1016/j.jneumeth.2008.04.011View ArticlePubMedGoogle Scholar
- Chen Y, Rangarajan G, Feng J, Ding M: Analyzing multiple nonlinear time series with extended Granger causality. Physics Letters A 2004, 324: 26. 10.1016/j.physleta.2004.02.032View ArticleGoogle Scholar
- Marinazzo D, Pellicoro M, Stramaglia S: Kernel-Granger causality and the analysis of dynamic networks. Physical review E 2008, 77: 056215. 10.1103/PhysRevE.77.056215View ArticleGoogle Scholar
- Locke JC, Kozma-Bognár L, Gould PD, Fehér B, Kevei E, Nagy F, Turner MS, Hall A, Millar AJ: Experimental Validation of a predicted feedback loop in the multi-oscillator clock of Arabidopsis thaliana. Molecular Systems Biology 2006, 2: 59. 10.1038/msb4100102PubMed CentralView ArticlePubMedGoogle Scholar
- Ueda HR: Systems biology flowering in the plant clock field. Molecular Systems Biology 2006, 2: 60. 10.1038/msb4100105PubMed CentralView ArticlePubMedGoogle Scholar
- ISI Web of knowledge:On 24th July, 2008, a search on ISI tells us that there are 3531 papers on Bayesian network in the area of Comp. Sci., Math., Math + Comp. Biol. and Business + Economics, and 1125 papers on Granger causality. [http://www.isiknowledge.com]
- Wang S, Chen Y, Ding M, Feng J, Stein JF, Aziz TZ, Liu X: Revealing the dynamic causal interdenpendence between neural and muscular signals in Parkinsonian tremor. Journal of Franklin Institute-Engineering and Applied Mathematics 2007, 344(3–4):180. 10.1016/j.jfranklin.2006.06.003View ArticleGoogle Scholar
- Wu J, Kendrick K, Feng J: Detecting Hot-Spots in Multivariates Biological Data. BMC Bioinformatics 2007, 8: 331. 10.1186/1471-2105-8-331PubMed CentralView ArticlePubMedGoogle Scholar
- Zhan Y, Halliday D, Jiang Ping, Liu X, Feng J: Detecting the time-dependent coherence between non-stationary electrophysiological signals-A combined statistical and time-frequency approach. Journal of Neuroscience Methods 2006, 156: 322. 10.1016/j.jneumeth.2006.02.013View ArticlePubMedGoogle Scholar
- Akaike H: Fitting Autoregressive Models for Regression. Annals of the Institute of Statistcal Mathmatics 1969, 21: 243. 10.1007/BF02532251View ArticleGoogle Scholar
- Beamish N, Priestley MB: A Study of Autoregressive and Window Spectral Estimation. Applied Statistics 1981, 30(1):41. 10.2307/2346656View ArticleGoogle Scholar
- Morettin PA: Levinson Algorithm and Its Applications in Time Series Analysis. International Statistical Review 1984, 52(1):83.View ArticleGoogle Scholar
- Morf M, Vieira A, Lee DTL, Kailath T: Recursive Multichannel Maximum Entropy Spectral Estimation. IEEE transactions on geosciences electronics 1978, GE-16(2):85. 10.1109/TGE.1978.294569View ArticleGoogle Scholar
- Ancona N, Marinazzo D, Stramaglia S: Radial Basis Function Approach to Nonlinear Granger Causality of Time Series. Physical Review E 2004, 70: 056221. 10.1103/PhysRevE.70.056221View ArticleGoogle Scholar
- Bishop CM: Neural Networks for Pattern Recognition. Oxford: Oxford Univ. Press; 1995.Google Scholar
- Bishop CM: Pattern Recognition and Machine Learning. New York: Springer Press; 2006.Google Scholar
- Murphy K: Bayes Net Toolbox for Matlab. It can be downloaded from the website.[http://www.cs.ubc.ca/~murphyk/Software/BNT/bnt.html]
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.