## Time series analysis via mechanistic models. In review; pre-published at arxiv.org/abs/0802.0021 (2008)

Citations: | 13 - 5 self |

### BibTeX

@MISC{Bretó08timeseries,

author = {Carles Bretó and Daihai He and Edward L. Ionides and Aaron and A. King},

title = {Time series analysis via mechanistic models. In review; pre-published at arxiv.org/abs/0802.0021},

year = {2008}

}

### OpenURL

### Abstract

The purpose of time series analysis via mechanistic models is to reconcile the known or hypothesized structure of a dynamical system with observations collected over time. We develop a framework for constructing nonlinear mechanistic models and carrying out inference. Our framework permits the consideration of implicit dynamic models, meaning statistical models for stochastic dynamical systems which are specified by a simulation algorithm to generate sample paths. Inference procedures that operate on implicit models are said to have the plug-and-play property. Our work builds on recently developed plug-and-play inference methodology for partially observed Markov models. We introduce a class of implicitly specified Markov chains with stochastic transition rates, and we demonstrate its applicability to open problems in statistical inference for biological systems. As one example, these models are shown to give a fresh perspective on measles transmission dynamics. As a second example, we present a mechanistic analysis of cholera incidence data, involving interaction between two competing strains of the pathogen Vibrio cholerae. 1. Introduction. A

### Citations

1363 |
Generalized linear models
- McCullagh, Nelder
- 1990
(Show Context)
Citation Context ...tem. This relationship is discussed further in Section 5. For generalized linear models, over-dispersion is commonplace, and failure to account properly for it can give rise to misleading conclusions =-=[53]-=-. Phrased another way, including sufficient stochasticity in a model to match the unpredictability of the data is essential if the model is to be used for forecasting, or predicting a quantitative ran... |

1147 | A tutorial on particle filters for online nonlinear/non-gaussian bayesian tracking
- Arulampalam, Maskell, et al.
(Show Context)
Citation Context ...Indeed, it is too widespread to give a comprehensive review and we instead list some examples: molecular biochemistry [48]; wildlife ecology [55]; cell biology [37]; economics [25]; signal processing =-=[4]-=-; data assimilation for numerical models [34]. The study of infectious disease, however, has a long history of motivating new modeling and data analysis methodology [3, 6, 27, 36, 42]. The freedom to ... |

757 |
Exact stochastic simulation of coupled chemical reactions
- Gillespie
- 1977
(Show Context)
Citation Context ...en if the SDE is an acceptable approximation to the disease dynamics, there are technical reasons to prefer a discrete model. Standard methods allow exact simulation for continuous time Markov chains =-=[13, 28]-=-, whereas for a nonlinear SDE this is at best difficult [9]. In addition, if an approximate Euler solution for a compartment model is required, non-negativity constraints can more readily be accommoda... |

706 |
Optimal Filtering
- Anderson, Moore
(Show Context)
Citation Context ...omes nonlinear regression) or if the stochastic dynamical system is perfectly observed [7]. Here we address the general case with both forms of stochasticity. Despite considerable work on such models =-=[2, 15, 20, 50]-=-, statistical methodology which is readily applicable for a wide range of models has remained elusive. For example, Markov Chain Monte Carlo and Monte Carlo Expectation-Maximization algorithms [15] ha... |

674 | Lévy processes and infinitely divisible distributions (english ed.)., volume 68 of Cambridge studies in advanced mathematics - Sato - 2005 |

498 |
A second course in stochastic processes
- Karlin, Taylor
- 1981
(Show Context)
Citation Context ... variation in the rates (discussion of this decision is postponed to Sections 4 and 5). We refer to white noise as the derivative of an integrated noise process with stationary independent increments =-=[40]-=-. The integral of a white noise process over an interval is thus well defined, even when the sample paths of the integrated noise process are not formally differentiable. Specifically, we introduce a ... |

410 |
Statistical Inference
- Casella, Berger
- 2001
(Show Context)
Citation Context ...π(n, x, µ, σ) in (5) then follows by substituting in the appropriate definitions. To calculate the infinitesimal moments, we note that the moment generating function of a gamma random variable (e.g., =-=[16]-=-) gives E[e −µ∆Γ ] = (1 + µτ) −δ/τ , where τ = σ 2 . It follows that V ar[e −µ∆Γ ] = E[e −2µ∆Γ ] − � E[e −µ∆Γ ] � 2 = (1 + 2µτ) −δ/τ − (1 + µτ) −2δ/τ . A Taylor series expansion around δ = 0 gives E[e... |

399 |
Monte Carlo Strategies in Scientific Computing
- Liu
- 2001
(Show Context)
Citation Context ...omes nonlinear regression) or if the stochastic dynamical system is perfectly observed [7]. Here we address the general case with both forms of stochasticity. Despite considerable work on such models =-=[2, 15, 20, 50]-=-, statistical methodology which is readily applicable for a wide range of models has remained elusive. For example, Markov Chain Monte Carlo and Monte Carlo Expectation-Maximization algorithms [15] ha... |

318 |
Markov chains, Gibbs fields, Monte Carlo simulation, and queues
- Br'emaud
- 1999
(Show Context)
Citation Context ..., below, shows that (2) defines a continuous time Markov chain when the conditions (P1–P5) hold. A finite-state continuous time Markov chain is specified by its infinitesimal transition probabilities =-=[13]-=-, which are in turn specified by (2). Theorem 2 determines the infinitesimal transition probabilities resulting from (2), supposing the conditions (P1–P7). When the infinitesimal transition probabilit... |

318 |
Time series analysis by state space methods
- Durbin, Koopman
- 2002
(Show Context)
Citation Context ...is that the model structure should be chosen based on scientific considerations, rather than statistical convenience. Although linear Gaussian models give an adequate representation of some processes =-=[21]-=-, nonlinear behavior is an essential property of many systems. This leads to a need for statistical modeling and inference techniques applicable to rather general classes of processes. In the absence ... |

250 | Stochastic Differential Equations - Oksendal - 2000 |

225 |
Core Team (2006): R: A language and environment for statistical computing. R Foundation for Statistical Computing
- Development
(Show Context)
Citation Context ...n a χ 2 approximation [e.g., 5]. Each point corresponds to an optimization carried out as described in the caption of Table 2. Local quadratic regression implemented in R via loess with a span of 0.6 =-=[57]-=- was used to estimate the profile likelihood, following Ionides [35]. 16 γsB (p < 10 −6 , likelihood ratio test) for which the epidemiologically relevant cases are only the severe cases that are likel... |

211 | Data assimilation using an ensemble Kalman filter technique
- Houtekamer, Mitchell
- 1998
(Show Context)
Citation Context ...ehensive review and we instead list some examples: molecular biochemistry [48]; wildlife ecology [55]; cell biology [37]; economics [25]; signal processing [4]; data assimilation for numerical models =-=[34]-=-. The study of infectious disease, however, has a long history of motivating new modeling and data analysis methodology [3, 6, 27, 36, 42]. The freedom to carry out formal statistical analysis based o... |

191 |
Approximate accelerated stochastic simulation of chemically reacting systems
- Gillespie
- 2001
(Show Context)
Citation Context ...lation methods are available [28]. In practice numerical schemes based on Euler approximations may be preferable— Euler schemes for Markov chain compartment models have been proposed based on Poisson =-=[29]-=-, binomial [63] and multinomial [14] approximations. By choosing a model with a convenient numerical solution (such as the procedure in Figure 1), the issue of simulating sample paths becomes simpler ... |

174 |
Contributions to the mathematical theory of epidemics
- Kermack, McKendrick
- 1927
(Show Context)
Citation Context ...onomics [25]; signal processing [4]; data assimilation for numerical models [34]. The study of infectious disease, however, has a long history of motivating new modeling and data analysis methodology =-=[3, 6, 27, 36, 42]-=-. The freedom to carry out formal statistical analysis based on mechanistically motivated, non-linear, non-stationary, continuous time stochastic models is a new development which promises to be a use... |

139 |
Infectious Diseases of Humans
- Anderson, May
- 1991
(Show Context)
Citation Context ...onomics [25]; signal processing [4]; data assimilation for numerical models [34]. The study of infectious disease, however, has a long history of motivating new modeling and data analysis methodology =-=[3, 6, 27, 36, 42]-=-. The freedom to carry out formal statistical analysis based on mechanistically motivated, non-linear, non-stationary, continuous time stochastic models is a new development which promises to be a use... |

139 | Sequential Monte Carlo samplers - Moral, Doucet, et al. - 2006 |

131 |
Combined parameter and state estimation in simulationbased filtering
- Liu, West
- 2001
(Show Context)
Citation Context ...ed method [36] which provides a way to calculate a maximum likelihood estimate via sequential Monte Carlo, a plug and play filtering technique. Approximate Bayesian sequential Monte Carlo methodology =-=[49]-=- has also been proposed. In Section 2, we introduce a new and general class of implicitly specified models. Section 3 is concerned with inference methodology and includes a review of the iterated filt... |

121 | R: A language and environment for statistical computing. R Foundation for statistical computing, Version 2.3.0 - Team - 2006 |

119 | Generalized Linear Models. 2nd ed - McCullagh, Nelder - 1989 |

103 |
Inference and Asymptotics
- Barndorff-Nielsen, Cox
- 1994
(Show Context)
Citation Context ...xed i and n gives rise to estimates ˙ ℓni for the partial derivatives of the conditional log likelihoods. Standard errors of parameters are found from the estimated observed Fisher information matrix =-=[5]-=-, with entries given by Îik = � n ˙ ℓni ˙ ℓnk. We prefer profile likelihood calculations, such as Figure 7, to derive confidence intervals for quantities of particular interest. However, standard erro... |

94 | Statistical modelling: the two cultures
- Breiman
- 2001
(Show Context)
Citation Context ...for transition probabilities or sample paths are not required. The ability to analyze implicit models reduces the separation between algorithmic methods and model-based analyses identified by Breiman =-=[12]-=-. The goal of this paper is to develop plug-and-play inference for a general class of implicitly specified stochastic dynamic models, and to show how this capability enables new and improved statistic... |

60 | Exact and computationally efficient likelihood-based estimation for discretely observed diffusion processes
- Beskos, Papaspiliopoulos, et al.
- 2006
(Show Context)
Citation Context ...ange of models has remained elusive. For example, Markov Chain Monte Carlo and Monte Carlo Expectation-Maximization algorithms [15] have technical difficulties handling continuous time dynamic models =-=[9]-=-; these two approaches also lack the plug-and-play property. Several inference techniques have previously been proposed which are compatible with plug-andplay inference from partially observed Markov ... |

50 |
Statistical Inference for Stochastic Processes
- Basawa, Rao
- 1980
(Show Context)
Citation Context ...ical analysis is simpler if stochasticity can be confined to the observation process (the statistical problem becomes nonlinear regression) or if the stochastic dynamical system is perfectly observed =-=[7]-=-. Here we address the general case with both forms of stochasticity. Despite considerable work on such models [2, 15, 20, 50], statistical methodology which is readily applicable for a wide range of m... |

44 | Sequential Monte Carlo without likelihoods - Sisson, Fan, et al. - 2007 |

42 |
Binomial leap methods for simulating stochastic chemical kinetics
- Tian, Burrage
- 2004
(Show Context)
Citation Context ...are available [28]. In practice numerical schemes based on Euler approximations may be preferable— Euler schemes for Markov chain compartment models have been proposed based on Poisson [29], binomial =-=[63]-=- and multinomial [14] approximations. By choosing a model with a convenient numerical solution (such as the procedure in Figure 1), the issue of simulating sample paths becomes simpler at the expense ... |

39 |
A self-organizing state-space model
- Kitagawa
- 1998
(Show Context)
Citation Context ...by a random walk θ1:N, with E[θ0] = θ and E[θn|θn−1] = θn−1 for n > 1, the calculation of ˆ θn = E[θn|y1:n] and Vn = V ar(θn|y1:n−1) is a well-studied and computationally convenient filtering problem =-=[44, 49]-=-. Additional stochasticity of this kind is introduced in steps 4 and 12 of Figure 2. This leads to time-varying parameter estimates, so ¯ θi(tn) in Figure 2 is an 6sMODEL INPUT: f(·), g(·|·), y1, . . ... |

35 | Parameter estimation for differential equations: a generalized smoothing approach
- Ramsay, Hooker, et al.
- 2007
(Show Context)
Citation Context ...ne interpretation of a compartment model is to write the flows as coupled ordinary differential equations (ODEs), d dt Nij = µijXi(t). (9) Data analysis via ODE models has challenges in its own right =-=[58]-=-. One can include stochasticity in (9) by adding a slowly varying function to the derivative [62]. Alternatively, one can add Gaussian white noise to give a set of coupled stochastic differential equa... |

32 | Bayesian inference for a discretely observed stochastic kinetic model - Boys, Wilkinson, et al. - 2008 |

27 |
Noisy clockwork: Time series analysis of population fluctuations
- Bjørnstad, Grenfell
(Show Context)
Citation Context ...logical systems. Mathematical models for the temporal dynamics of biological populations have long played a role in understanding fluctuations in population abundance and interactions between species =-=[11, 52]-=-. When using models to examine the strength of evidence concerning rival hypotheses about a system, a model is typically required to capture not just the qualitative features of the dynamics but also ... |

26 |
Ecological and immunological determinants of influenza evolution
- Ferguson, Galvani, et al.
- 2003
(Show Context)
Citation Context ...effective vaccines and vaccination strategies [32]. Previous analyses relating mathematical consequences of strain structure to disease data include studies of malaria [33], dengue [23], in12sfluenza =-=[24, 45]-=- and cholera [47]. For measles, the strain structure is considered to have negligible importance for the transmission dynamics [18], another reason why measles epidemics form a relatively simple biolo... |

24 | Estimating Dynamic Equilibrium Economies: Linear versus Nonlinear Likelihood
- Fernández-Villaverde, Rubio-Ramírez
- 2005
(Show Context)
Citation Context ...in many other contexts. Indeed, it is too widespread to give a comprehensive review and we instead list some examples: molecular biochemistry [48]; wildlife ecology [55]; cell biology [37]; economics =-=[25]-=-; signal processing [4]; data assimilation for numerical models [34]. The study of infectious disease, however, has a long history of motivating new modeling and data analysis methodology [3, 6, 27, 3... |

24 | Approximate Bayesian computation scheme for parameter inference and model selection in dynamical systems - Toni, Welch, et al. |

23 |
The effect of antibodydependent enhancement on the transmission dynamics and persistence of multiple-strain pathogens
- Ferguson, Anderson, et al.
- 1999
(Show Context)
Citation Context ...on, and developing effective vaccines and vaccination strategies [32]. Previous analyses relating mathematical consequences of strain structure to disease data include studies of malaria [33], dengue =-=[23]-=-, in12sfluenza [24, 45] and cholera [47]. For measles, the strain structure is considered to have negligible importance for the transmission dynamics [18], another reason why measles epidemics form a ... |

22 |
Stochastic Population Models in Ecology and Epidemiology
- Bartlett
- 1960
(Show Context)
Citation Context ...onomics [25]; signal processing [4]; data assimilation for numerical models [34]. The study of infectious disease, however, has a long history of motivating new modeling and data analysis methodology =-=[3, 6, 27, 36, 42]-=-. The freedom to carry out formal statistical analysis based on mechanistically motivated, non-linear, non-stationary, continuous time stochastic models is a new development which promises to be a use... |

22 |
Unifying the epidemiological and evolutionary dynamics of pathogens
- Grenfell, Pybus, et al.
- 2004
(Show Context)
Citation Context ... of the strain structure can be key to understanding the epidemiology of the disease, understanding evolution of resistance to medication, and developing effective vaccines and vaccination strategies =-=[32]-=-. Previous analyses relating mathematical consequences of strain structure to disease data include studies of malaria [33], dengue [23], in12sfluenza [24, 45] and cholera [47]. For measles, the strain... |

22 |
Equation-free: The computer-assisted analysis of complex multiscale systems
- Kevrekidis, Gear, et al.
(Show Context)
Citation Context ...l techniques that require only simulation from the model (i.e., for which the model could be replaced by a black box which inputs parameters and outputs sample paths) have been called “equation free” =-=[43, 65]-=-. We will use the more descriptive expression “plug and play”. ∗ Supported by National Science Foundation grant EF 0430120. AMS 2000 subject classifications: Primary 62M10; secondary 62M05 Keywords an... |

21 |
Inference for nonlinear dynamical systems
- Ionides, Breto, et al.
- 2006
(Show Context)
Citation Context ...ay inference from partially observed Markov processes. Nonlinear forecasting [41] is a method of simulated moments which approximates the likelihood. Iterated filtering is a recently developed method =-=[36]-=- which provides a way to calculate a maximum likelihood estimate via sequential Monte Carlo, a plug and play filtering technique. Approximate Bayesian sequential Monte Carlo methodology [49] has also ... |

20 |
2000 Time series modelling of childhood diseases: a dynamical systems approach
- Finkenstädt, Grenfell
(Show Context)
Citation Context |

20 | Climate Change 2007: The Physical Science Basis. Contribution of Working Group I to the Fourth Assessment - Miller |

18 |
Why do populations cycle? A synthesis of statistical and mechanistic modeling approaches. Ecology 80
- Kendall, Briggs, et al.
- 1999
(Show Context)
Citation Context ...the plug-and-play property. Several inference techniques have previously been proposed which are compatible with plug-andplay inference from partially observed Markov processes. Nonlinear forecasting =-=[41]-=- is a method of simulated moments which approximates the likelihood. Iterated filtering is a recently developed method [36] which provides a way to calculate a maximum likelihood estimate via sequenti... |

16 |
Likelihood-based estimation of continuous-time epidemic models from time-series data: application to measles transmission in London
- Cauchemez, Ferguson
- 2008
(Show Context)
Citation Context ...mics, to statistical models, which both capture the mechanistic basis of the system and statistically describe the data, are well documented by a sequence of work on the dynamics of measles epidemics =-=[3, 6, 10, 17, 27, 54]-=-. Measles is no longer a major developed world health issue but still causes substantial morbidity and mortality, particularly in sub-Saharan Africa [18, 31]. The availability of excellent data before... |

14 |
Dynamics of measles epidemics: Estimating scaling of transmission rates using a time series SIR model, Ecol
- Bjornstad, Finkenstadt, et al.
(Show Context)
Citation Context ...mics, to statistical models, which both capture the mechanistic basis of the system and statistically describe the data, are well documented by a sequence of work on the dynamics of measles epidemics =-=[3, 6, 10, 17, 27, 54]-=-. Measles is no longer a major developed world health issue but still causes substantial morbidity and mortality, particularly in sub-Saharan Africa [18, 31]. The availability of excellent data before... |

14 |
On the asymptotic distribution of the size of a stochastic epidemic
- SELLKE
- 1983
(Show Context)
Citation Context ...ansition. The memoryless property of the exponential distribution makes this equivalent to the construction of Theorem 1, where clocks are restarted only for individuals who make a transition (Sellke =-=[61]-=-). Sellke’s construction is convenient for the proof of Theorem 1. Note 2: The trajectories of the individuals are coupled through the dependence of µ(t, X(t)) on X(t), and through the noise processes... |

13 |
Earn, Dynamical resonance can account for seasonality of influenza epidemics
- Dushoff, Poltkin, et al.
- 2004
(Show Context)
Citation Context ... review [17]. From another perspective, the properties of stochastic dynamic epidemic models have been studied extensively in the context of continuous time models with only demographic stochasticity =-=[8, 22, 64]-=-. We go beyond previous approaches, by demonstrating the possibility of carrying out modeling and data analysis via continuous time mechanistic models with both demographic and environmental stochasti... |

13 |
Antigenic diversity and the transmission dynamics of Plasmodium falciparum
- Gupta, Trenholme, et al.
- 1994
(Show Context)
Citation Context ...e to medication, and developing effective vaccines and vaccination strategies [32]. Previous analyses relating mathematical consequences of strain structure to disease data include studies of malaria =-=[33]-=-, dengue [23], in12sfluenza [24, 45] and cholera [47]. For measles, the strain structure is considered to have negligible importance for the transmission dynamics [18], another reason why measles epid... |

13 |
Uses and abuses of mathematics in biology
- May
(Show Context)
Citation Context ...logical systems. Mathematical models for the temporal dynamics of biological populations have long played a role in understanding fluctuations in population abundance and interactions between species =-=[11, 52]-=-. When using models to examine the strength of evidence concerning rival hypotheses about a system, a model is typically required to capture not just the qualitative features of the dynamics but also ... |

12 | Bayesian analysis of single-molecule experimental data - Kou, Xie, et al. - 2005 |

11 |
Appropriate models for the management of infectious diseases, PLoS Med
- Wearing, Rohani, et al.
(Show Context)
Citation Context ... review [17]. From another perspective, the properties of stochastic dynamic epidemic models have been studied extensively in the context of continuous time models with only demographic stochasticity =-=[8, 22, 64]-=-. We go beyond previous approaches, by demonstrating the possibility of carrying out modeling and data analysis via continuous time mechanistic models with both demographic and environmental stochasti... |

11 | Stochastic differential equations, 5th ed - Øksendal - 1998 |