Approximate Bayesian computation schemes for parameter inference of discrete stochastic models using simulated likelihood density
© Wu et al.; licensee BioMed Central Ltd. 2014
Published: 6 November 2014
Mathematical modeling is an important tool in systems biology to study the dynamic property of complex biological systems. However, one of the major challenges in systems biology is how to infer unknown parameters in mathematical models based on the experimental data sets, in particular, when the data are sparse and the regulatory network is stochastic.
To address this issue, this work proposed a new algorithm to estimate parameters in stochastic models using simulated likelihood density in the framework of approximate Bayesian computation. Two stochastic models were used to demonstrate the efficiency and effectiveness of the proposed method. In addition, we designed another algorithm based on a novel objective function to measure the accuracy of stochastic simulations.
Simulation results suggest that the usage of simulated likelihood density improves the accuracy of estimates substantially. When the error is measured at each observation time point individually, the estimated parameters have better accuracy than those obtained by a published method in which the error is measured using simulations over the entire observation time period.
In recent years, quantitative methods have become increasingly important for studying complex biological systems. To build a mathematical model of a complex system, two main procedures are commonly conducted . The first step is to determine the elements of the network and regulatory relationships between the elements. In the second step, we need to infer the model parameters according to experimental data. Since biological experiments are time-consuming and expensive, normally experimental data are often scarce and incomplete compared with the number of unknown model parameters. In addition, the likelihood surfaces of large models are complex. The calibration of these unknown parameters within a model structure is one of the key issues in systems biology . The analysis of such dynamical systems therefore requires new, effective and sophisticated inference methods.
During the last decade, several approaches have been developed for estimating unknown parameters: namely, optimization methods and Bayesian inference methods. Aiming at minimizing an objective function, optimization methods start with an initial guess, and then search in a directed manner within the parameter space [3, 4]. The objective function is usually defined by the discrepancy between the simulated outputs of the model and sets of experimental data. Recently, the objective function has been extended to a continuous approach by considering simulation over the whole time period  and a multi-scale approach by including multiple types of experimental information . Several types of optimization methods can be found in the literature, among which two major types are called gradient-based optimization methods and evolutionary-based optimization methods. Based on these two basic approaches, various techniques such as simulated annealing . linear and non-linear least-squares fitting , genetic algorithms  and evolutionary computation [10, 11] have been attempted to build computational biology models. Using optimization methods, the inferred set of parameters produces the best fit between simulations and experimental data [12, 13]. which have been successfully applied for biological systems, however, there are still some limitations with these methods such as the problem of high computational cost when significant noise exists in the system. To address these issues, deterministic and stochastic global optimization methods have been explored .
When modeling biological systems where molecular species are present in low copy numbers, measurement noise and intrinsic noise play a substantial role , which is a major obstacle for modeling. Bayesian inference methods have been used to tackle such difficulties by extracting useful information from noise data . The main advantage of Bayesian inference is that it is able to infer the whole probability distributions of parameters by updating probability estimates using Bayes' Rule, rather than just a point estimate from optimization methods. Also. Bayesian methods are more robust than using other methods when they are applied to estimate stochastic systems, which is not that obvious for modeling of deterministic systems . Developments have taken place during the last 20 years and recent advances in Bayesian computation including Markov chain Monte Carlo (MCMC) techniques and sequential Monte Carlo (SMC) methods have been successfully applied to biological systems [18, 19].
For the case of parameter estimation when likelihoods are analytically or computationally intractable, approximate Bayesian computation (ABC) methods have been applied successfully [20, 21]. ABC algorithms provide stable parameter estimates and are also relatively computationally efficient, therefore, they have been treated as substantial techniques for solving inference problems of various types of models that were intractable only a few years ago . In ABC. the evaluation of the likelihood is replaced by a simulation-based procedure using the comparison between the observed data and simulated data . Recently, a semi-automatic method has been proposed to construct the summary statistics for ABC . These methods have been applied in a diverse range of fields such as molecular genetics, epidemiology and evolutionary biology etc. [24–26].
Despite substantial progress in the application of ABC to deterministic models, the development of inference methods for stochastic models is still at the very early stage. Compared with deterministic models, there are a number of open problems in the inference of stochastic models. For example, recent work proposed ABC to infer unknown parameters in stochastic differential equation models . Our recent computational tests  showed the advantages and disadvantages of a published ABC algorithm for stochastic chemical reaction systems in . In this work, we propose two novel algorithms to improve the performance of ABC algorithms using the simulated likelihood density.
Results and discussion
The first test system with four reactions
Comparison of averaged error and mean count number for estimated rate constants over five iterations using algorithms 1 and 2 with simulation number of 10 for system 1.
In these performances, we used ∈ k = (0.07, 0.06, 0.055, 0.05, 0.045) and (0.05, 0.045, 0.04, 0.035, 0.03) for algorithm 1 with step sizes Δt = 3 and 5, respectively. For algorithm 2, these values are ∈ k = (0.095, 0.08, 065, 0.05, 0.04) and (0.059, 0.055, 0.05, 0.045, 0.04). An interesting observation is that the values of mean count number are very large in the first iteration, then decrease sharply and stay within a value stably. We have a detailed test of using different values of the fitness tolerance ∈ k and found that when using step size of Δt = 3, mean count number stays at one if ∈ k ≥ 0.1; but it starts to increase sharply to a large number if ∈ k < 0.1. The observation numbers using a step size of Δt = 3 is 10 and the maximum error that can incur calculated from step 2.d is 0.1 with one hundred particles. Similarly, this critical ∈ k value is 0.06 for a step size of Δt = 5.
Meanwhile all averaged errors have a decreasing trend over iterations. Looking at different cases with various values of discrepancy tolerance α , it is also observed that using α = 0.1 results in more discrepancies of the estimated parameters on average than the other two cases, in particular, than the case α = 0.05. Thus in our following tests, we just concentrate on the cases of α = 0.05 and varying α. In addition, we observe that by taking α = 0.05 for the case with step size of Δt = 3, it leads to more accurate approximation since α = 0. 05 is less than most values of α in the case of varying values of α. It is consistent with the cases of a step size of Δt = 5 in which little differences can be found comparing strategies using α = 0.05 and α = ∈ k since the values of ∈ k are quite close to 0.05. In the case of varying values of α, a small value of ε5 leads to a small value of α5, which results in a substantial increase in mean count number. However, this large mean count number does not necessary bring more accurate estimated parameters. With these findings, we simulated results using α = 0.05 and α = ∈ only for algorithm 2. Consistent results are obtained using algorithm 2. Moreover, results obtained using algorithm 2 is more accurate than those from algorithm 1.
The second test system with eight reactions
This gene network includes five species, namely DNA, messenger RNA, protein product, dimeric protein, and the compound formed by dimeric protein binding to the DNA promoter site, which are denoted by DNA, mRNA, P, P2 and DNA · P2, respectively. In this network, the first two reactions R1 and R2 are reversible reactions for dimeric protein binding to the DNA promoter site. Reactions R3 and R7 are transcriptional and translation processes for producing mRNA and protein, respectively. Reactions R5 and R6 represent the interchange between protein P and dimeric protein P2. The system ends up with a degradation process of protein P .
The prior distribution of each parameter follows a uniform distribution π(θ) ~ U(0, B). For rate constants c1 ~ c8, the values of B are (0.5,2,1,0.1,0.5, 5,1,0.1). The proposed two algorithms were implemented over five iterations and each iteration contains 100 particles. We choose step sizes Δt = 2 or 5 and the number of stochastic simulation B k = 10.
Comparison of averaged error and mean count number for estimated rate constants of system 2 using algorithms 1 and 2.
Same ∈ k
Diff. ∈ k
Same ∈ k
Diff. ∈ k
Same ∈ k
Diff. ∈ k
Same ∈ k
Diff. ∈ k
For algorithm 2, we carried out similar tests. In the first case, we set ∈ k =(0.24,0.23,0.22,0.21,0.2), which is applied to the strategy of fixing α = 0.05 and varying α with α= ∈ k that is the case "Same ∈ k " in Table 2. Again, the averaged count numbers of varying α strategy are much smaller than those using a fixed α. Thus we decreased the value to (0.095,0.09,0.085,0.08, 0.075), which is the case "Diff. ∈ k " in Table 2; However, the averaged count numbers in the "Diff. ∈ k " case are similar to those of the previous two strategies, namely a fixed a and "Same ∈ k ". For algorithm 2, Table 2 suggests that the varying α strategy generates estimates that are more accurate than those obtained from the fixed α strategy. However, the best estimates in Table 2 are obtained using algorithm 1 and fixed α strategy.
To uncover the information of biological systems, we proposed two algorithms for the inference of unknown parameters in complex stochastic models for chemical reaction systems. Algorithm 1 is in the framework of ABC SMC and uses transitional density based on the simulations over two consecutive observation time points. Algorithm 2 generates simulations of the whole time interval but differs from the published method in the error finding steps by comparing errors of simulated data to experimental data at each time point. The proposed new algorithms impose stricter criteria to measure the simulation error. Using two chemical reaction systems as the test problems, we examined the accuracy and efficiency of proposed new algorithms. Based on the results of two algorithms for system 1, we discovered that taking smaller values of discrepancy tolerance α will result in more accurate estimates of unknown model parameters. This conclusion is confirmed by the second system that we tested under different conditions. Numerical results suggested that the proposed new algorithms are promising methods to infer parameters in high-dimensional and complex biological system models and have better accuracy compared with the results of the published method . The encouraging result is that new algorithms do not need more computing time to achieve such accuracy. Our computational tests showed that the selection for the value of fitness tolerance is a key step in the success of ABC algorithms. The advantage of the population Monte-Carlo methods is the ability to reduce the fitness tolerance gradually over populations. Generally, a smaller value of fitness tolerance will lead to a larger number of iteration count and consequently larger computing time. For deterministic inference problems, a smaller value of fitness tolerance normally will generate estimates with better accuracy. However, for stochastic models, this conclusion is not always true. In addition to the fitness tolerance, our numerical results suggested that other factors, such as the simulation algorithm for chemical reaction systems and the strategy of discrepancy tolerance, also have influences on the accuracy of estimates. Thus more skilled approaches, such as the adaptive selection process for the fitness tolerance, should be considered to improve the performance of ABC algorithms.
In this work, we used the SSA to simulate chemical reaction systems . This approach may be appropriate when the biological system is not large. In fact, for the two biological systems discussed in this work, the computing time of inference is still very large. To reduce the computing time, more effective methods should be used to simulate the biological systems, such as the τ-leap methods  and multi-scale simulation methods [36, 37]. Another alternative approach is to use parallel computing to reduce the heavy computing loads. All these issues are potential topics for future research work.
ABC SMC algorithm
ABC algorithms bypass the requirement for evaluating likelihood functions directly in order to obtain the posterior distributions of unknown parameters. Instead, ABC methods simulate the model with given parameters, compare the observed and simulated data, and then accept or reject the particular parameters based on the error of simulation data. Thus there are three key steps in the implementations of ABC algorithms. The first step is the generation of a sample of parameters θ* from the prior distribution of parameters or from other distributions that are determined in ABC algorithms. The second step is to define distance function d(X, Y) between the simulated data X and experimental observation data Y. Finally, a tolerance value is needed as a selection criterion to accept or reject the sampled parameter θ*. Based on the generic form of ABC algorithm , a number of methods have been developed including ABC rejection sampler and ABC MCMC [38, 39]. The ABC rejection algorithm is one of the basic ABC algorithm that may result in long computing time when a badly prior distribution that is far away from posterior distribution is chosen. ABC MCMC introduces a concept of acceptance probability during the decision making step which saves computing time. However, this may result in getting stuck in the regions of low probability for the chain and we may never be able to get a good approximation. To tackle these challenges, the idea of particle filtering has been introduced. Instead of having one parameter vector at a time, we sample from a pool of parameter sets simultaneously and treat each parameter vector as a particle. The algorithm starts from sampling a pool of N particles for parameter vector θ through prior distribution π(θ). The sampled particle candidates will be chosen randomly from the pool and we will assign each particle a corresponding weight ω to be considered as the sampling probability. A perturbation and filtering process following through a transition kernel q(·|θ*) finds the particles θ**. Similarly with θ**, data Y can be simulated and compared with experimental data X to further fulfil the requirements for estimating posterior distribution.
The basic form of algorithm described above is as follows :
Algorithm: ABC SMC
1. Define the threshold values , start with iteration k = 1.
2. Set the particle indicator i = 1.
and D0 is the experimental data set.
If k > 1, sample θ from the previous population with weights w k−1 and perturb the particle to obtain θ* using a kernel function .
If π(θ*) = 0 or b k (θ*) = 0, return to the beginning of step 3.
If i <N, update i = i + 1 and return to step 3.
5. Normalize the weights If k <K, update k = k +1 and go back to step 2.
A number of algorithms have been developed using the particle filtering technique, such as the partial rejection control, population Monte-Carlo and SMC. Each of them differs in the formation of weight w and the transition kernels.
ABC using simulated likelihood density
In the simulated likelihood density (SLD) methods, this transitional density is approximated by that obtained from a large number of simulations.
where d(x, y) is a distance measure between x and y.
Here we propose a new algorithm that uses the simulated transitional density function as the objective function. Unlike ABC SMC algorithm , the new method considers the transitional density function from t i −1 to t i only at each step. Based on the framework of ABC SMC, the new algorithm using transitional density is proposed as follows.
ABC SLD algorithm 1
1. Given data X and any assumed prior distribution π(θ), define a set of threshold values ∈1, · · ·, ∈ K .
2. For iteration k = 1,
(a) Set the particle indicator i = 1, sample θ* ~ π(θ).
(b) For time step l = 1,2, · · ·, n, use initial condition X l − 1 and parameter θ* to generate data Y at t l for B k times.
(c) For m = 1, ·· ·, B k , calculate the value of discrepancy and test for
where α is a defined constant.
If it is true, let β ml (θ*) = 0, otherwise it is one. Then determine
If ∈ <∈ k , update and move to the next particle i = i +1.
(e) Assign weight for each particle.
4. For iteration k = 2, · · ·, K
(a) Start with i = 1, Sample using the calculated weights .
(b) Perturb θ* through sampling θ** ~ q(θ|θ*>), where or q = U(a, b). Here values of a,b depend on θ* and .
(c) Generate simulations and calculate the error ∈ using the same steps as in 2(b) ~ (d).
(d) For each particle, assign weights
An alternative approach is to generate simulations over the observation time period but compare the error to experimental data at each time point. The approach locates somewhere between ABC SMC algorithm  and the proposed Algorithm 1. which is presented below. For simplicity we do not give a detailed algorithm, but just provide the key steps 2.b) ~ 2.d) that are different from those in Algorithm 1.
ABC SLD algorithm 2
2.b) Generate data Y B k times using θ*.
If it is true, let b ml (θ*) = 0, otherwise it is one.
The authors would like to thank the Australian Research Council for the Discovery Project (T.T. DP120104460). T.T. is also an ARC Future Fellow (FT100100748).
The publication costs for this article were funded by the corresponding author.
This article has been published as part of BMC Bioinformatics Volume 15 Supplement 12, 2014: Selected articles from the IEEE International Conference on Bioinformatics and Biomedicine (BIBM 2013): Bioinformatics. The full contents of the supplement are available online at http://www.biomedcentral.com/bmcbioinformatics/supplements/15/S12.
- Zhan C, Yeung LF: Parameter estimation in systems biology models using spline approximation. BMC systems biology. 2011, 5: 14-10.1186/1752-0509-5-14.PubMed CentralView ArticlePubMedGoogle Scholar
- Kikuchi S, Tominaga D, Arita M, Takahashi K, Tomita M: Dynamic modeling of genetic networks using genetic algorithm and S-system. Bioinformatics. 2003, 19 (5): 643-650. 10.1093/bioinformatics/btg027.View ArticlePubMedGoogle Scholar
- Gadkar KG, Gunawan R, Doyle FJ: Iterative approach to model identification of biological networks. BMC bioinformatics. 2005, 6: 155-10.1186/1471-2105-6-155.PubMed CentralView ArticlePubMedGoogle Scholar
- Gonzalez OR, Küper C, Jung K, Naval PC, Mendoza E: Parameter estimation using Simulated Annealing for S-system models of biochemical networks. Bioinformatics. 2007, 23 (4): 480-486. 10.1093/bioinformatics/btl522.View ArticlePubMedGoogle Scholar
- Deng Z, Tian T: A continuous approach for inferring parameters in mathematical models of regulatory networks. BMC bioinformatics. 2014, 15: 256-10.1186/1471-2105-15-256.PubMed CentralView ArticlePubMedGoogle Scholar
- Tian T, Smith-Miles K: Mathematical modelling of GATA-switching for regulating the differentiation of hematopoietic stem cell. BMC bioinformatics. 2014, 8 (S8): S8-Google Scholar
- Kirkpatrick S, Gelatt CD, Vecchi MP: Optimization by Simulated Annealing. Science. 1983, 220 (4598): 671-680. 10.1126/science.220.4598.671.View ArticlePubMedGoogle Scholar
- Mendes P, Kell D: Non-linear optimization of biochemical pathways: applications to metabolic engineering and parameter estimation. Bioinformatics. 1998, 14 (10): 869-883. 10.1093/bioinformatics/14.10.869.View ArticlePubMedGoogle Scholar
- Srinivas M, Patnaik LM: Genetic algorithms: A survey. Computer. 1994, 27 (6): 17-26.View ArticleGoogle Scholar
- Ashyraliyev M, Jaeger J, Blom J: Parameter estimation and determinability analysis applied to Drosophila gap gene circuits. BMC Systems Biology. 2008, 2: 83-10.1186/1752-0509-2-83.PubMed CentralView ArticlePubMedGoogle Scholar
- Moles CG, Mendes P, Banga JR: Parameter estimation in biochemical pathways: a comparison of global optimization methods. Genome research. 2003, 13 (11): 2467-2474. 10.1101/gr.1262503.PubMed CentralView ArticlePubMedGoogle Scholar
- Lall R, Voit EO: Parameter estimation in modulated, unbranched reaction chains within biochemical systems. Computational biology and chemistry. 2005, 29 (5): 309-318. 10.1016/j.compbiolchem.2005.08.001.View ArticlePubMedGoogle Scholar
- Lillacci G, Khammash M: Parameter estimation and model selection in computational biology. PLoS computational biology. 2010, 6 (3): el000696-View ArticleGoogle Scholar
- Goel G, Chou IC, Voit EO: System estimation from metabolic time-series data. Bioinformatics. 2008, 24 (21): 2505-2511. 10.1093/bioinformatics/btn470.PubMed CentralView ArticlePubMedGoogle Scholar
- Raj A, van Oudenaarden A: Nature, nurture, or chance: stochastic gene expression and its consequences. Cell. 2008, 135 (2): 216-226. 10.1016/j.cell.2008.09.050.PubMed CentralView ArticlePubMedGoogle Scholar
- Wilkinson DJ: Bayesian methods in bioinformatics and computational systems biology. Briefings in, bioinformatics. 2007, 8 (2): 109-116.View ArticleGoogle Scholar
- Toni T, Welch D, Strelkowa N, Ipsen A, Stumpf MP: Approximate Bayesian computation scheme for parameter inference and model selection in dynamical systems. Journal of the Royal Society Interface. 2009, 6 (31): 187-202. 10.1098/rsif.2008.0172.PubMed CentralView ArticleGoogle Scholar
- Battogtokh D, Asch DK, Case ME, Arnold J, Schüttler HB: An ensemble method for identifying regulatory circuits with special reference to the qa gene cluster of Neurospora crassa. Proceedings of the National Academy of Sciences. 2002, 99 (26): 16904-16909. 10.1073/pnas.262658899. [http://www.pnas.org/content/99/26/16904.abstract]View ArticleGoogle Scholar
- Sisson SA, Fan Y, Tanaka MM: Sequential monte carlo without likelihoods. Proceedings of the National Academy of Sciences. 2007, 104 (6): 1760-1765. 10.1073/pnas.0607208104.View ArticleGoogle Scholar
- Beaumont MA, Zhang W, Balding DJ: Approximate Bayesian Computation in Population Genetics. Genetics. 2002, 162 (4): 2025-2035.PubMed CentralPubMedGoogle Scholar
- Marjoram P, Molitor J, Plagnol V, Tavaré S: Markov chain Monte Carlo without likelihoods. Proceedings of the National Academy of Sciences. 2003, 100 (26): 15324-15328. 10.1073/pnas.0306899100.View ArticleGoogle Scholar
- Pritchard JK, Seielstad MT, Perez-Lezaun A, Feldman MW: Population growth of human Y chromosomes: a study of Y chromosome microsatellites. Molecular Biology and Evolution. 1999, 16 (12): 1791-1798. 10.1093/oxfordjournals.molbev.a026091.View ArticlePubMedGoogle Scholar
- Fearnhead P, Prangle D: Constructing summary statistics for approximate Bayesian computation: semi-automatic approximate Bayesian computation. Journal of the Royal Statistical Society: Series B (Statistical Methodology). 2012, 74 (3): 419-474. 10.1111/j.1467-9868.2011.01010.x.View ArticleGoogle Scholar
- Marjoram P, Tavaré S: Modern computational approaches for analysing molecular genetic variation data. Nature Reviews Genetics. 2006, 7 (10): 759-770. 10.1038/nrg1961.View ArticlePubMedGoogle Scholar
- Tanaka MM, Francis AR, Luciani F, Sisson S: Using approximate Bayesian computation to estimate tuberculosis transmission parameters from genotype data. Genetics. 2006, 173 (3): 1511-1520. 10.1534/genetics.106.055574.PubMed CentralView ArticlePubMedGoogle Scholar
- Thornton K, Andolfatto P: Approximate Bayesian inference reveals evidence for a recent, severe bottleneck in a Netherlands population of Drosophila melanogaster. Genetics. 2006, 172 (3): 1607-1619.PubMed CentralView ArticlePubMedGoogle Scholar
- Picchini UL: Inference for SDE models via Approximate Bayesian Computation. Journal of Computational and Graphical Statistics in press. 2014Google Scholar
- Wu Q, Smith-Miles K, Tian T: Approximate Bayesian computation for estimating rate constants in biochemical reaction systems. Bioinformatics and Biomedicine (BIBM). 2013, 416-421. IEEE International Conference on 2013Google Scholar
- Daigle BJ, Roh MK, Petzold LR, Niemi J: Accelerated maximum likelihood parameter estimation for stochastic biochemical systems. BMC bioinformatics. 2012, 13: 68-10.1186/1471-2105-13-68.PubMed CentralView ArticlePubMedGoogle Scholar
- Gillespie DT: Exact stochastic simulation of coupled chemical reactions. The journal of physical chemistry. 1977, 81 (25): 2340-2361. 10.1021/j100540a008.View ArticleGoogle Scholar
- Tian T, Xu S, Gao J, Burrage K: Simulated maximum likelihood method for estimating kinetic rates in gene expression. Bioinformatics. 2007, 23: 84-91. 10.1093/bioinformatics/btl552.View ArticlePubMedGoogle Scholar
- Wang Y, Christley S, Mjolsness E, Xie X: Parameter inference for discretely observed stochastic kinetic models using stochastic gradient descent. BMC systems biology. 2010, 4: 99-10.1186/1752-0509-4-99.PubMed CentralView ArticlePubMedGoogle Scholar
- Golightly A, Wilkinson DJ: Bayesian inference for stochastic kinetic models using a diffusion approximation. Biometrics. 2005, 61 (3): 781-788. 10.1111/j.1541-0420.2005.00345.x.View ArticlePubMedGoogle Scholar
- Reinker S, Altman R, Timmer J: Parameter estimation in stochastic biochemical reactions. IEE Proceedings-Systems Biology. 2006, 153 (4): 168-178. 10.1049/ip-syb:20050105.View ArticlePubMedGoogle Scholar
- Tian T, Burrage K: Binomial leap methods for simulating stochastic chemical kinetics. The Journal of chemical physics. 2004, 121 (21): 10356-10364. 10.1063/1.1810475.View ArticlePubMedGoogle Scholar
- Pahle J: Biochemical simulations: stochastic, approximate stochastic and hybrid approaches. Briefings in bioinformatics. 2009, 10: 53-64.PubMed CentralView ArticlePubMedGoogle Scholar
- Burrage K, Tian T, Burrage P: A multi-scaled approach for simulating chemical reaction systems. Progress in biophysics and molecular biology. 2004, 85 (2): 217-234.View ArticlePubMedGoogle Scholar
- Boys RJ, Wilkinson DJ, Kirkwood TB: Bayesian inference for a discretely observed stochastic kinetic model. Statistics and Computing. 2008, 18 (2): 125-135. 10.1007/s11222-007-9043-x.View ArticleGoogle Scholar
- Golightly A, Wilkinson DJ: Bayesian parameter inference for stochastic biochemical network models using particle Markov chain Monte Carlo. Interface Focus. 2011, 1 (6): 807-820. 10.1098/rsfs.2011.0047.PubMed CentralView ArticlePubMedGoogle Scholar
- Hurn AS, Jeisman J, Lindsay KA: Seeing the wood for the trees: A critical evaluation of methods to estimate the parameters of stochastic differential equations. Journal of Financial Econometrics. 2007, 5 (3): 390-455. 10.1093/jjfinec/nbm009.View ArticleGoogle Scholar
- Hurn A, Lindsay K: Estimating the parameters of stochastic differential equations. Mathematics and computers in simulation. 1999, 48 (4): 373-384.View ArticleGoogle Scholar
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/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.