Statistical Inference for Partially Observed Markov Processes via the R Package pomp

Partially observed Markov process (POMP) models, also known as hidden Markov models or state space models, are ubiquitous tools for time series analysis. The R package pomp provides a very ﬂexible framework for Monte Carlo statistical investigations using nonlinear, non-Gaussian POMP models. A range of modern statistical methods for POMP models have been implemented in this framework including sequential Monte Carlo, iterated ﬁltering, particle Markov chain Monte Carlo, approximate Bayesian computation, maximum synthetic likelihood estimation, nonlinear forecasting, and trajectory matching. In this paper, we demonstrate the application of these methodologies using some simple toy problems. We also illustrate the speciﬁcation of more complex POMP models, using a nonlinear epidemiological model with a discrete population, seasonality, and extra-demographic stochasticity. We discuss the speciﬁcation of user-deﬁned models and the development of additional methods within the programming environment provided by pomp .


Introduction
A partially observed Markov process (POMP) model consists of incomplete and noisy measurements of a latent, unobserved Markov process.The far-reaching applicability of this class of models has motivated much software development (Commandeur et al. 2011).It has been a challenge to provide a software environment that can effectively handle broad classes of POMP models and take advantage of the wide range of statistical methodologies that have been proposed for such models.The pomp software package (King et al. 2017) differs from previous approaches by providing a general and abstract representation of a POMP model.Therefore, algorithms implemented within pomp are necessarily applicable to arbitrary POMP models.Moreover, models formulated with pomp can be analyzed using multiple methodologies in search of the most effective method, or combination of methods, for the problem at hand.However, since pomp is designed for general POMP models, methods that exploit additional model structure have yet to be implemented.In particular, when linear, Gaussian approximations are adequate for one's purposes, or when the latent process takes values in a small, discrete set, methods that exploit these additional assumptions to advantage, such as the extended and ensemble Kalman filter methods or exact hidden Markov model methods, are available, but not yet as part of pomp.It is the class of nonlinear, non-Gaussian POMP models with large state spaces upon which pomp is focused.
A POMP model may be characterized by the transition density for the Markov process and the measurement density1 .However, some methods require only simulation from the transition density whereas others require evaluation of this density.Still other methods may not work with the model itself but with an approximation, such as a linearization.Algorithms for which the dynamic model is specified only via a simulator are said to be plug-and-play (Bretó et al. 2009;He et al. 2010).Plug-and-play methods can be employed once one has "plugged" a model simulator into the inference machinery.Since many POMP models of scientific interest are relatively easy to simulate, the plug-and-play property facilitates data analysis.Even if one candidate model has tractable transition probabilities, a scientist will frequently wish to consider alternative models for which these probabilities are intractable.In a plug-and-play methodological environment, analysis of variations in the model can often be achieved by changing a few lines of the model simulator codes.The price one pays for the flexibility of plug-and-play methodology is primarily additional computational effort, which can be substantial.Nevertheless, plug-and-play methods implemented using pomp have proved capable for state-of-the-art inference problems (e.g., King et al. 2008;Bhadra et al. 2011;Shrestha et al. 2011Shrestha et al. , 2013;;Earn et al. 2012;Roy et al. 2012;Blackwood et al. 2013a,b;He et al. 2013;Bretó 2014;Blake et al. 2014).The recent surge of interest in plug-and-play methodology for POMP models includes the development of nonlinear forecasting (Ellner et al. 1998), iterated filtering (Ionides et al. 2006(Ionides et al. , 2015)), ensemble Kalman filtering (Shaman and Karspeck 2012), approximate Bayesian computation (ABC; Sisson et al. 2007), particle Markov chain Monte Carlo (PMCMC; Andrieu et al. 2010), probe matching (Kendall et al. 1999), and synthetic likelihood (Wood 2010).Although the pomp package provides a general environment for methods with and without the plug-and-play property, development of the package to date has emphasized plug-and-play methods.
The pomp package is philosophically neutral as to the merits of Bayesian inference.It enables a POMP model to be supplemented with prior distributions on parameters, and several Bayesian methods are implemented within the package.Thus pomp is a convenient environment for those who wish to explore both Bayesian and non-Bayesian data analyses.
The remainder of this paper is organized as follows.Section 2 defines mathematical notation for POMP models and relates this to their representation as objects of class 'pomp' in the pomp package.Section 3 introduces several of the statistical methods currently implemented in pomp.Section 4 constructs and explores a simple POMP model, demonstrating the use of the available statistical methods.Section 5 illustrates the implementation of more complex POMPs, using a model of infectious disease transmission as an example.Finally, Section 6 discusses extensions and applications of pomp.1: Constituent methods for class 'pomp' objects and their translation into mathematical notation for POMP models.For example, the rprocess method is set using the rprocess argument to the pomp constructor function.

POMP models and their representation in pomp
Let θ be a p-dimensional real-valued parameter, θ ∈ R p .For each value of θ, let {X(t ; θ), t ∈ T } be a Markov process, with X(t ; θ) taking values in R q .The time index set T ⊂ R may be an interval or a discrete set.Let {t i ∈ T, i = 1, . . ., N }, be the times at which X(t ; θ) is observed, and t 0 ∈ T be an initial time.Assume t 0 ≤ t 1 < t 2 < • • • < t N .We write X i = X(t i ; θ) and X i:j = (X i , X i+1 , . . ., X j ).The process X 0:N is only observed by way of another process Y 1:N = (Y 1 , . . ., Y N ) with Y n taking values in R r .The observable random variables Y 1:N are assumed to be conditionally independent given X 0:N .The data, y * 1:N = (y * 1 , . . ., y * N ), are modeled as a realization of this observation process and are considered fixed.We suppose that X 0:N and Y 1:N have a joint density f X 0:N ,Y 1:N (x 0:N , y 1:N ; θ).The POMP structure implies that this joint density is determined by the initial density, f X 0 (x 0 ; θ), together with the conditional transition probability density, f Xn|X n−1 (x n | x n−1 ; θ), and the measurement density, f Yn|Xn (y n | x n ; θ), for 1 ≤ n ≤ N .In particular, we have (1) Note that this formalism allows the transition density, f Xn|X n−1 , and measurement density, f Yn|Xn , to depend explicitly on n.

Implementation of POMP models
pomp is fully object-oriented: in the package, a POMP model is represented by an S4 object (Chambers 1998;Genolini 2008) of class 'pomp'.Slots in this object encode the components of the POMP model, and can be filled or changed using the constructor function pomp and various other convenience functions.Methods for the class 'pomp' class use these components to carry out computations on the model.Table 1 gives the mathematical notation corresponding to the elementary methods that can be executed on a class 'pomp' object.
The rprocess, dprocess, rmeasure, and dmeasure arguments specify the transition probabilities f Xn|X n−1 (x n | x n−1 ; θ) and measurement densities f Yn|Xn (y n | x n ; θ).Not all of these arguments must be supplied for any specific computation.In particular, plug-and-play methodology by definition never uses dprocess.An empty dprocess slot in a class 'pomp' object is therefore acceptable unless a non-plug-and-play algorithm is attempted.In the package, the data and corresponding measurement times are considered necessary parts of a class 'pomp' object whilst specific values of the parameters and latent states are not.Applying the simulate function to an object of class 'pomp' returns another object of class 'pomp', within which the data y * 1:N have been replaced by a stochastic realization of Y 1:N , the corresponding realization of X 0:N is accessible via the states method, and the params slot has been filled with the value of θ used in the simulation.
To illustrate the specification of models in pomp and the use of the package's inference algorithms, we will use a simple example.The Gompertz (1825) model can be constructed via R> library("pomp") R> pompExample(gompertz) which results in the creation of an object of class 'pomp', named gompertz, in the workspace.The structure of this model and its implementation in pomp is described below, in Section 4. One can view the components of gompertz listed in

Initial conditions
In some experimental situations, f X 0 (x 0 ; θ) corresponds to a known experimental initialization, but in general the initial state of the latent process will have to be inferred.If the transition density for the dynamic model, f Xn|X n−1 (x n | x n−1 ; θ), does not depend on time and possesses a unique stationary distribution, it may be natural to set f X 0 (x 0 ; θ) to be this stationary distribution.Otherwise, and more commonly in the authors' experience, no clear scientifically motivated choice of f X 0 (x 0 ; θ) exists and one can proceed by treating the value of X 0 as a parameter to be estimated.In this case, f X 0 (x 0 ; θ) concentrates at a point, the location of which depends on θ.

Covariates
Scientifically, one may be interested in the role of a vector-valued covariate process {Z(t)} in explaining the data.Modeling and inference conditional on {Z(t)} can be carried out within the general framework for nonhomogeneous POMP models, since the arbitrary densities f Xn|X n−1 , f X 0 and f Yn|Xn can depend on the observed process {Z(t)}.For example, it may be the case that The covar argument in the pomp constructor allows for time-varying covariates measured at times specified in the tcovar argument.An example using covariates is given in Section 5.

Methodology for POMP models
Data analysis typically involves identifying regions of parameter space within which a postulated model is statistically consistent with the data.Additionally, one frequently desires to assess the relative merits of alternative models as explanations of the data.Once the user has encoded one or more POMP models as objects of class 'pomp', the package provides a variety of algorithms to assist with these data analysis goals.Table 2 provides an overview of several inference methodologies for POMP models.Each method may be categorized as full-information or feature-based, Bayesian or frequentist, and plug-and-play or not plug-andplay.
Approaches that work with the full likelihood function, whether in a Bayesian or frequentist context, can be called full-information methods.Since low-dimensional sufficient statistics are not generally available for POMP models, methods which take advantage of favorable low-dimensional representations of the data typically lose some statistical efficiency.We use the term "feature-based" to describe all methods not based on the full likelihood, since such methods statistically emphasize some features of the data over others.
Many Monte Carlo methods of inference can be viewed as algorithms for the exploration of high-dimensional surfaces.This view obtains whether the surface in question is the likelihood surface or that of some other objective function.The premise behind many recent methodological developments in Monte Carlo methods for POMP models is that generic stochastic numerical analysis tools, such as standard Markov chain Monte Carlo and Robbins-Monro type methods, are effective only on the simplest models.For many models of scientific interest, therefore, methods that leverage the POMP structure are needed.Though pomp has sufficient flexibility to encode arbitrary POMP models and methods and therefore also provides a platform for the development of novel POMP inference methodology, pomp's development to date has focused on plug-and-play methods.However, the package developers welcome contributions and collaborations to further expand pomp's functionality in non-plug-and-play directions also.In the remainder of this section, we describe and discuss several inference methods, all currently implemented in the package.

The likelihood function and sequential Monte Carlo
The log likelihood for a POMP model is (θ) = log f Y 1:N (y * 1:N ; θ), which can be written as a sum of conditional log likelihoods, where and we use the convention that y * 1:0 is an empty vector.The structure of a POMP model implies the representation 1).Although (θ) typically has no closed form, it can frequently be computed by Monte Carlo methods.Sequential Monte Carlo (SMC) builds up a representation of f Xn|Y 1:n−1 (x n | y * 1:n−1 ; θ) that can be used to obtain an estimate, ˆ n|1:n−1 (θ), of n|1:n−1 (θ) and hence an approximation, ˆ (θ), to (θ).SMC (a basic version of which is presented as Algorithm 1), is also known as the particle filter, since it is conventional to describe the Monte Carlo sample, {X F n,j , j in 1 : J} as a swarm of particles representing f Xn|Y 1:n (x n | y * 1:n ; θ).The swarm is propagated forward according to the dynamic model and then assimilated to the next data point.Using an evolutionary analogy, the prediction step (line 3) mutates the particles in the swarm and the filtering step (line 7) corresponds to selection.SMC is implemented in pomp in the pfilter function.The basic particle filter in Algorithm 1 possesses the plugand-play property.Many variations and elaborations to SMC have been proposed; these may improve numerical performance in appropriate situations (Cappé et al. 2007) but typically Algorithm 1: Sequential Monte Carlo (SMC, or particle filter): pfilter( P, Np = J), using notation from Table 1 where P is a class 'pomp' object with definitions for rprocess, dmeasure, init.state,coef, and obs.input: Simulator for f Xn|X n−1 (x n | x n−1 ; θ); evaluator for f Yn|Xn (y n | x n ; θ); simulator for f X 0 (x 0 ; θ); parameter, θ; data, y * 1:N ; number of particles, J. Initialize filter particles: simulate X F 0,j ∼ f X 0 ( • ; θ) for j in 1 : J. for n in 1 : N do Simulate for prediction: Evaluate weights: w(n, j) = f Yn|Xn (y * n |X P n,j ; θ) for j in 1 : J.
Apply Algorithm 2 to select indices k 1:J with P [k j = m] = w(n, m).
Resample: set X F n,j = X P n,k j for j in 1 : J.
Construct cumulative sum: c j = j m=1 wm , for j in 1 : J. Draw a uniform initial sampling point: U 1 ∼ Uniform(0, J −1 ).Construct evenly spaced sampling points: U j = U 1 + (j − 1)J −1 , for j in 2 : J. Initialize: set p = 1.for j in 1 : J do while U j > c p do Step to the next resampling index: set p = p + 1. end Assign resampling index: set k j = p.end output: Resampling indices, k 1:J .complexity: O(J) lose the plug-and-play property.Arulampalam et al. (2002), Doucet and Johansen (2009), and Kantas et al. (2015) have written excellent introductory tutorials on the particle filter and particle methods more generally.
Basic SMC methods fail when an observation is extremely unlikely given the model.This leads to the situation that at most a few particles are consistent with the observation, in which case the effective sample size (Liu 2001) of the Monte Carlo sample is small and the particle filter is said to suffer from particle depletion.Many elaborations of the basic SMC algorithm have been proposed to ameliorate this problem.However, it is often preferable to remedy the situation by seeking a better model.The plug-and-play property assists in this process by facilitating investigation of alternative models.
for n in 1 : N do Perturb parameters: Θ P n,j i ∼ Normal Θ F n−1,j i , (a m−1 σ i ) 2 for i ∈ I, j in 1 : J.
Simulate prediction particles: Evaluate weights: w(n, j) = f Yn|Xn (y * n |X P n,j ; Θ P n,j ) for j in 1 : J.
Apply Algorithm 2 to select indices k 1:J with P [k u = j] = w (n, j).
Resample particles: X F n,j = X P n,k j and Θ F n,j = Θ P n,k j for j in 1 : J.
Filter mean: θn i = J j=1 w(n, j) Θ P n,j i for i ∈ I.
In line 6 of Algorithm 1, systematic resampling (Algorithm 2) is used in preference to multinomial resampling.Algorithm 2 reduces Monte Carlo variability while resampling with the proper marginal probability.In particular, if all the particle weights are equal then Algorithm 2 has the appropriate behavior of leaving the particles unchanged.As pointed out by Douc et al. (2005), stratified resampling performs better than multinomial sampling and Algorithm 2 is in practice comparable in performance to stratified resampling and somewhat faster.

Iterated filtering
Iterated filtering techniques maximize the likelihood obtained by SMC (Ionides et al. 2006(Ionides et al. , 2011(Ionides et al. , 2015)).The key idea of iterated filtering is to replace the model we are interested in fitting -which has time-invariant parameters -with a model that is just the same except that its parameters take a random walk in time.Over multiple repetitions of the filtering procedure, the intensity of this random walk approaches zero and the modified model approaches the original one.Adding additional variability in this way has four positive effects: A1.It smooths the likelihood surface, which facilitates optimization.A2.It combats particle depletion by adding diversity to the population of particles.
A3.The additional variability can be exploited to explore the likelihood surface and estimate the gradient of the (smoothed) likelihood, based on the same filtering procedure that is required to estimate the value of the likelihood (Ionides et al. 2006(Ionides et al. , 2011)).
A4.It preserves the plug-and-play property, inherited from the particle filter.
Iterated filtering is implemented in the mif function.By default, mif carries out the procedure of Ionides et al. (2006).The improved iterated filtering algorithm (IF2) of Ionides et al. (2015) has shown promise.A limited version of IF2 is available via the method = "mif2" option; a full version of this algorithm will be released soon.In all iterated filtering methods, by analogy with annealing, the random walk intensity can be called a temperature, which is decreased according to a prescribed cooling schedule.One strives to ensure that the algorithm will freeze at the maximum of the likelihood as the temperature approaches zero.
The perturbations on the parameters in lines 2 and 7 of Algorithm 3 follow a normal distribution, with each component, [θ] i , of the parameter vector perturbed independently.Neither normality nor independence are necessary for iterated filtering, but, rather than varying the perturbation distribution, one can transform the parameters to make these simple algorithmic choices reasonable.
Algorithm 3 gives special treatment to a subset of the components of the parameter vector termed initial value parameters (IVPs), which arise when unknown initial conditions are modeled as parameters.These IVPs will typically be inconsistently estimable as the length of the time series increases, since for a stochastic process one expects only early time points to contain information about the initial state.Searching the parameter space using time-varying parameters is inefficient in this situation, and so Algorithm 3 perturbs these parameters only at time zero.
Lines 7-12 of Algorithm 3 are exactly an application of SMC (Algorithm 1) to a modified POMP model in which the parameters are added to the state space.This approach has been used in a variety of previously proposed POMP methodologies (Kitagawa 1998;Liu and West 2001;Wan and Merwe 2000) but iterated filtering is distinguished by having theoretical justification for convergence to the maximum likelihood estimate (Ionides et al. 2011).
Initialization: compute ˆ (θ 0 ) using Algorithm 1 with J particles.for m in 1 : M do Draw a parameter proposal, θ P m , from the proposal distribution: Compute ˆ (θ P m ) using Algorithm 1 with J particles.

Particle Markov chain Monte Carlo
Full information plug-and-play Bayesian inference for POMP models is enabled by particle Markov chain Monte Carlo (PMCMC) algorithms (Andrieu et al. 2010).PMCMC methods combine likelihood evaluation via SMC with MCMC moves in the parameter space.The simplest and most widely used PMCMC algorithm, termed particle marginal Metropolis-Hastings (PMMH), is based on the observation that the unbiased likelihood estimate provided by SMC can be plugged into the Metropolis-Hastings update procedure to give an algorithm targeting the desired posterior distribution for the parameters (Andrieu and Roberts 2009).PMMH is implemented in pmcmc, as described in Algorithm 4. In part because it gains only a single likelihood evaluation from each particle filtering operation, PMCMC can be computationally relatively inefficient (Bhadra 2010;Ionides et al. 2015).Nevertheless, its invention introduced the possibility of full-information plug-and-play Bayesian inferences in some situations where they had been unavailable.

Synthetic likelihood of summary statistics
Some motivations to estimate parameters based on features rather than the full likelihood include: B1. Reducing the data to sensibly selected and informative low-dimensional summary statistics may have computational advantages (Wood 2010).
B2.The scientific goal may be to match some chosen characteristics of the data rather than all aspects of it.Acknowledging the limitations of all models, this limited aspiration may Algorithm 5: Synthetic likelihood evaluation: probe(P, nsim = J, probes = S), using notation from Table 1 where P is a class 'pomp' object with defined methods for rprocess, rmeasure, init.state, and obs.input: Simulator for f Xn|X n−1 (x n | x n−1 ; θ); simulator for f X 0 (x 0 ; θ); simulator for f Yn|Xn (y n | x n ; θ); parameter, θ; data, y * 1:N ; number of simulations, J; vector of summary statistics or probes, S = (S 1 , . . ., S d ).
B3.In conjunction with full-information methodology, consideration of individual features has diagnostic value to determine which aspects of the data are driving the full-information inferences (Reuman et al. 2006).
B4. Feature-based methods for dynamic models typically do not require the POMP model structure.However, that benefit is outside the scope of the pomp package.
B5. Feature-based methods are typically doubly plug-and-play, meaning that they require simulation, but not evaluation, for both the latent process transition density and the measurement model.
When pursuing goal B1, one aims to find summary statistics which are as close as possible to sufficient statistics for the unknown parameters.Goals B2 and B3 deliberately look for features which discard information from the data; in this context the features have been called probes (Kendall et al. 1999).The features are denoted by a collection of functions, S = (S 1 , . . ., S d ), where each S i maps an observed time series to a real number.We write S = (S 1 , . . ., S d ) for the vector-valued random variable with S = S(Y 1:N ), with f S (s ; θ) being the corresponding joint density.The observed feature vector is s * where s * i = S i (y * 1:N ), and for any parameter set one can look for parameter values for which typical features for simulated data match the observed features.One can define a likelihood function, S (θ) = f S (s * ; θ).Arguing that S should be approximately multivariate normal, for suitable choices of the features, Wood (2010) proposed using simulations to construct a multivariate normal approximation to S (θ), and called this a synthetic likelihood.

Compute observed probes: s
Compute simulated probes: , but performance will depend on the choice of , τ i , and σ i , as well as on the choice of probes S.
Simulation-based evaluation of a feature matching criterion is implemented by probe (Algorithm 5).The feature matching criterion requires a scale, and a natural scale to use is the empirical covariance of the simulations.Working on this scale, as implemented by probe, there is no substantial difference between the probe approaches of Kendall et al. (1999) and Wood (2010).Numerical optimization of the synthetic likelihood is implemented by probe.match,which offers the choice of either the subplex method (Rowan 1990;King 2015) or any method provided by optim or the nloptr package (Johnson 2014;Ypma 2014).

Approximate Bayesian computation (ABC)
ABC algorithms are Bayesian feature-matching techniques, comparable to the frequentist generalized method of moments (Marin et al. 2012).The vector of summary statistics S, the corresponding random variable S, and the value s * = S(y * 1:N ), are defined as in Section 3.4.The goal of ABC is to approximate the posterior distribution of the unknown parameters given S = s * .ABC has typically been motivated by computational considerations, as in point B1 of Section 3.4 (Sisson et al. 2007;Toni et al. 2009;Beaumont 2010).Points B2 and B3 also apply (Ratmann et al. 2009).
The key theoretical insight behind ABC algorithms is that an unbiased estimate of the likelihood can be substituted into a Markov chain Monte Carlo algorithm to target the required posterior, the same result that justifies PMCMC (Andrieu and Roberts 2009).However, ABC takes a different approach to approximating the likelihood.The likelihood of the observed features, S (θ) = f S (s * ; θ), has an approximately unbiased estimate based on a single

Monte Carlo realization Y
where B d is the volume of the d-dimensional unit ball and τ i is a scaling chosen for the ith feature.The likelihood approximation in Equation 6 differs from the synthetic likelihood in Algorithm 5 in that only a single simulation is required.As becomes small, the bias in Equation 6decreases but the Monte Carlo variability increases.The ABC implementation abc (presented in Algorithm 6) is a random walk Metropolis implementation of ABC-MCMC (Algorithm 3 of Marin et al. 2012).In the same style as iterated filtering and PMCMC, we assume a Gaussian random walk in parameter space; the package supports alternative choices of the proposal distribution.

Nonlinear forecasting
Nonlinear forecasting (NLF) uses simulations to build up an approximation to the one-step prediction distribution that is then evaluated on the data.We saw in Section 3.1 that SMC evaluates the prediction density for the observation, f Yn|Y , by first building an approximation to the prediction density of the latent process, f Xn|Y for some choice of positive lags c 1:L .The prediction errors when this model is applied to the data give rise to a quantity called the quasi-likelihood, which behaves for many purposes like a likelihood (Smith 1993).The implementation in nlf maximizes the quasi-likelihood computed in Algorithm 7, using the subplex method (Rowan 1990;King 2015) or any other optimizer offerered by optim.The construction of the quasi-likelihood in nlf follows the specific recommendations of Kendall et al. (2005).In particular, the choice of radial basis functions, f k , in line 5 and the specification of m k and s in lines 3 and 4 were proposed by Kendall et al. (2005) based on trial and error.The quasi-likelihood is mathematically most similar to a likelihood when min(c 1:L ) = 1, so that Q (θ) approximates the factorization of the likelihood in Equation 2. With this in mind, it is natural to choose contiguous lags c 1:L = 1 : L. However, Kendall et al. (2005) found that a two-step prediction criterion, with min(c 1:L ) = 2, led to improved numerical performance.It is natural to ask when one might choose to use quasi-likelihood estimation in place of full likelihood estimation implemented by SMC.Some considerations follow, closely related to the considerations for the synthetic likelihood and ABC (Sections 3.4 and 3.5): C1.NLF benefits from stationarity since (unlike SMC) it uses all time points in the simulation to build a prediction rule valid at all time points.Indeed, NLF has not been considered applicable for non-stationary models and, on account of this, nlf is not appropriate if the model includes time-varying covariates.An intermediate scenario between stationarity and full non-stationarity is seasonality, where the dynamic model is forced by cyclical covariates, and this is supported by nlf (cf.B1 in Section 3.4).
C2. Potentially, quasi-likelihood could be preferable to full likelihood in some situations.It has been argued that a two-step prediction criterion may sometimes be more robust than the likelihood to model misspecification (Xia and Tong 2011) (cf.B2).
C4. NLF does not require that the model be Markovian (cf.B4), although the pomp implementation, nlf, does.
C6.The regression surface reconstruction carried out by NLF does not scale well with the dimension of the observed data.NLF is recommended only for low-dimensional time series observations.NLF can be viewed as an estimating equation method, and so standard errors can be computed by standard sandwich estimator or bootstrap techniques (Kendall et al. 2005).The optimization in NLF is typically carried out with a fixed seed for the random number generator, so the simulated quasi-likelihood is a deterministic function.If rprocess depends smoothly on the random number sequence and on the parameters, and the number of calls to the random number generator does not depend on the parameters, then fixing the seed results in a smooth objective function.However, some common components to model simulators, such as rnbinom, make different numbers of calls to the random number generator depending on the arguments, which introduces nonsmoothness into the objective function.

A first example: The Gompertz model
The plug-and-play methods in pomp were designed to facilitate data analysis based on complicated models, but we will first demonstrate the basics of pomp using simple discrete-time models, the Gompertz and Ricker models for population growth (Reddingius 1971;Ricker 1954).The Ricker model will be introduced in Section 4.5 and used in Section 4.6; the remainder of Section 4 will use the Gompertz model.The Gompertz model postulates that the density, X t+∆t , of a population of organisms at time t + ∆t depends on the density, X t , at time t according to In Equation 9, K is the carrying capacity of the population, r is a positive parameter, and the ε t are independent and identically-distributed lognormal random variables with log ε t ∼ Normal(0, σ 2 ).Additionally, we will assume that the population density is observed with errors in measurement that are lognormally distributed: Taking a logarithmic transform of Equation 9gives On this transformed scale, the model is linear and Gaussian and so we can obtain exact values of the likelihood function by applying the Kalman filter.Plug-and-play methods are not strictly needed; this example therefore allows us to compare the results of generally applicable plug-and-play methods with exact results from the Kalman filter.Later we will look at the Ricker model and a continuous-time model for which no such special tricks are available.
The first step in implementing this model in pomp is to construct an R (R Core Team 2017) object of class 'pomp' that encodes the model and the data.This involves the specification of functions to do some or all of rprocess, rmeasure, and dmeasure, along with data and (optionally) other information.The documentation (?pomp) spells out the usage of the pomp constructor, including detailed specifications for all its arguments and links to several examples.
To begin, we will write a function that implements the process model simulator.This is a function that will simulate a single step (t → t + ∆t) of the unobserved process (Equation 9).
R> gompertz.proc.sim<-function(x, t, params, delta.t,...) { + eps <-exp(rnorm(n = 1, mean = 0, sd = params["sigma"])) + S <-exp(-params["r"] * delta.t)+ setNames(params["K"]^(1 -S) * x["X"]^S * eps, "X") + } The translation from the mathematical description (Equation 9) to the simulator is straightforward.When this function is called, the argument x contains the state at time t.The parameters (including K, r, and σ) are passed in the argument params.Notice that x and params are named numeric vectors and that the output must likewise be a named numeric vector, with names that match those of x.The argument delta.tspecifies the time-step size.In this case, the time-step will be 1 unit; we will see below how this is specified.
Next, we will implement a simulator for the observation process, Equation 10.
R> gompertz.meas.sim<-function(x, t, params, ...) { + setNames(rlnorm(n = 1, meanlog = log(x["X"]), sd = params["tau"]), "Y") + } Again the translation from the measurement model Equation 10 is straightforward.When the function gompertz.meas.sim is called, the named numeric vector x will contain the unobserved states at time t; params will contain the parameters as before.This return value will be a named numeric vector containing a single draw from the observation process (Equation 10).
Complementing the measurement model simulator is the corresponding measurement model density, which we implement as follows: R> gompertz.meas.dens<-function(y, x, t, params, log, ...) { + dlnorm(x = y["Y"], meanlog = log(x["X"]), sdlog = params["tau"], + log = log) + } We will need this later on for inference using pfilter, mif and pmcmc.When the function gompertz.meas.dens is called, y will contain the observation at time t, x and params will be as before, and the parameter log will indicate whether the likelihood (log = FALSE) or the log likelihood (log = TRUE) is required.
With the above in place, we build an object of class 'pomp' via a call to pomp: R> gompertz <-pomp(data = data.frame(time= 1:100, Y = NA), times = "time", + rprocess = discrete.time.sim(step.fun= gompertz.proc.sim,delta.t= 1), + rmeasure = gompertz.meas.sim,t0 = 0) The first argument (data) specifies a data frame that holds the data and the times at which the data were observed.Since this is a toy problem, we have as yet no data; in a moment, we will generate some simulated data.The second argument (times) specifies which of the columns of data is the time variable.The rprocess argument specifies that the process model simulator will be in discrete time, with each step of duration delta.ttaken by the function given in the step.funargument.The rmeasure argument specifies the measurement model simulator function.t0 fixes t 0 for this model; here we have chosen this to be one time unit prior to the first observation.
It is worth noting that implementing the rprocess, rmeasure, and dmeasure components as R functions, as we have done above, leads to needlessly slow computation.As we will see below, pomp provides facilities for specifying the model in C, which can accelerate computations manyfold.
Before we can simulate from the model, we need to specify some parameter values.The parameters must be a named numeric vector containing at least all the parameters referenced by the functions gompertz.proc.simand gompertz.meas.sim.The parameter vector needs to determine the initial condition X(t 0 ) as well.Let us take our parameter vector to be R> theta <-c(r = 0.1, K = 1, sigma = 0.1, tau = 0.1, X.0 = 1) The parameters r, K, σ, and τ appear in gompertz.proc.simand gompertz.meas.sim.The initial condition X 0 is also given in theta.The fact that the initial condition parameter's name ends in .0 is significant: it tells pomp that this is the initial condition of the state variable X.This use of the .0suffix is the default behavior of pomp: one can however parameterize the initial condition distribution arbitrarily using pomp's optional initializer argument.
We can now simulate the model at these parameters: R> gompertz <-simulate(gompertz, params = theta) Now gompertz is identical to what it was before, except that the missing data have been replaced by simulated data.The parameter vector (theta) at which the simulations were performed has also been saved internally to gompertz.We can plot the simulated data via R> plot(gompertz, variables = "Y") Figure 1 shows the results of this operation.

Computing the likelihood using SMC
As discussed in Section 3, some parameter estimation algorithms in the pomp package are doubly plug-and-play in that they require only rprocess and rmeasure.These include the nonlinear forecasting algorithm nlf, the probe-matching algorithm probe.match,and approximate Bayesian computation via abc.The plug-and-play full-information methods in pomp, however, require dmeasure, i.e., the ability to evaluate the likelihood of the data given the unobserved state.The gompertz.meas.densabove does this, but we must fold it into the class 'pomp' object in order to use it.We can do this with another call to pomp: R> gompertz <-pomp(gompertz, dmeasure = gompertz.meas.dens) The result of the above is a new class 'pomp' object gompertz in every way identical to the one we had before, but with the measurement-model density function dmeasure now specified.
To estimate the likelihood of the data, we can use the function pfilter, an implementation of Algorithm 1.We must decide how many concurrent realizations (particles) to use: the larger the number of particles, the smaller the Monte Carlo error but the greater the computational burden.Here, we run pfilter with 1000 particles to estimate the likelihood at the true parameters: Now let us compute the log likelihood at a different point in parameter space, one for which r, K, and σ are each 50% higher than their true values.
R> theta.guess<-theta.true<-coef(gompertz) R> theta.guess[c("r","K", "sigma")] <-1.5 * theta.true[c("r","K", "sigma")] R> pf <-pfilter(gompertz, params = theta.guess,Np = 1000) R> loglik.guess<-logLik(pf) R> loglik.guess[1] 24.89781 In this case, the Kalman filter computes the exact log likelihood at the true parameters to be 36.01,while the particle filter with 1000 particles gives 35.75.Since the particle filter gives an unbiased estimate of the likelihood, the difference is due to Monte Carlo error in the particle filter.One can reduce this error by using a larger number of particles and/or by re-running pfilter multiple times and averaging the resulting estimated likelihoods.The latter approach has the advantage of allowing one to estimate the Monte Carlo error itself; we will demonstrate this in Section 4.3.

Maximum likelihood estimation via iterated filtering
Let us use the iterated filtering approach described in Section 3.2 to obtain an approximate maximum likelihood estimate for the data in gompertz.Since the parameters of Equations 9 and 10 are constrained to be positive, when estimating, we transform them to a scale on which they are unconstrained.The following encodes such a transformation and its inverse: R> gompertz.log.tf<-function(params, ...) log(params) R> gompertz.exp.tf<-function(params, ...) exp(params) We add these to the existing class 'pomp' object via: R> gompertz <-pomp(gompertz, toEstimationScale = gompertz.log.tf,+ fromEstimationScale = gompertz.exp.tf) The following code initializes the iterated filtering algorithm at several starting points around theta.true and estimates the parameters r, τ , and σ.
R> estpars <-c("r", "sigma", "tau") R> library("foreach") R> mif1 <-foreach(i = 1:10, .combine= c) %dopar% { + theta.guess<-theta.true+ theta.guessTable 3: Results of estimating parameters r, σ, and τ of the Gompertz model (Equations 9 and 10) by maximum likelihood using iterated filtering (Algorithm 3), compared with the exact MLE and with the true value of the parameter.The first three columns show the estimated values of the three parameters.The next two columns show the log likelihood, ˆ , estimated by SMC (Algorithm 1) and its standard error, respectively.The exact log likelihood, , is shown in the rightmost column.An ideal likelihood-ratio 95% confidence set, not usually computationally available, includes all parameters having likelihood within qchisq(0.95,df = 3)/2 = 3.91 of the exact MLE.We see that both the mif MLE and the truth are in this set.In this example, the mif MLE is close to the exact MLE, so it is reasonable to expect that profile likelihood confidence intervals and likelihood ratio tests constructed using the mif MLE have statistical properties similar to those based on the exact MLE.
also that we have used the foreach package (Kane et al. 2013;Revolution Analytics and Weston 2014) to parallelize the computations.
Each of the 10 mif runs ends up with a different point estimate (Figure 2).We focus on that with the highest estimated likelihood, having evaluated the likelihood several times to reduce the Monte Carlo error in the likelihood evaluation.The particle filter produces an unbiased estimate of the likelihood; therefore, we will average the likelihoods, not the log likelihoods.
R> mf1 <-mif1 [[which.max(pf1)]]R> theta.mif<-coef(mf1) R> loglik.mif<-replicate(n = 10, logLik(pfilter(mf1, Np = 10000))) R> loglik.mif<-logmeanexp(loglik.mif, se = TRUE) R> theta.true<-coef(gompertz) R> loglik.true<-replicate(n = 10, logLik(pfilter(gompertz, Np = 20000))) R> loglik.true<-logmeanexp(loglik.true, se = TRUE) For the calculation above, we have replicated the iterated filtering search, made a careful estimation of the log likelihood, ˆ , and its standard error using pfilter at each of the resulting point estimates, and then chosen the parameter corresponding to the highest likelihood as our numerical approximation to the maximum likelihood estimate (MLE).Taking advantage of the Gompertz model's tractability, we also use the Kalman filter to maximize the exact log likelihood, , and evaluate it at the estimated MLE obtained by mif.The resulting estimates are shown in Table 3. Usually, the last row and column of Table 3 would not be available even for a simulation study validating the inference methodology for a known POMP model.
In this case, we see that the mif procedure is successfully maximizing the likelihood up to an error of about 0.1 log units.To carry out Bayesian inference we need to specify a prior distribution on unknown parameters.The pomp constructor function provides the rprior and dprior arguments, which can be filled with functions that simulate from and evaluate the prior density, respectively.Methods based on random walk Metropolis-Hastings require evaluation of the prior density (dprior), but not simulation (rprior), so we specify dprior for the Gompertz model as follows.

Full information Bayesian inference via PMCMC
R> hyperparams <-list(min = coef(gompertz)/10, max = coef(gompertz) * 10) R> gompertz.dprior<-function (params, ..., log) { + f <-sum(dunif(params, min = hyperparams$min, max = hyperparams$max, The PMCMC algorithm described in Section 3.3 can then be applied to draw a sample from the posterior.Recall that, for each parameter proposal, PMCMC pays the full price of a particle filtering operation in order to obtain the Metropolis-Hastings acceptance probability.For the same price, iterated filtering obtains, in addition, an estimate of the derivative and a probable improvement of the parameters.For this reason, PMCMC is relatively inefficient at traversing parameter space.When Bayesian inference is the goal, it is therefore advisable to first locate a neighborhood of the MLE using, for example, iterated filtering.PMCMC can then be initialized in this neighborhood to sample from the posterior distribution.The following adopts this approach, running 5 independent PMCMC chains using a multivariate normal random walk proposal (with diagonal variance-covariance matrix, see ?mvn.diag.rw).Comparison with the analysis of Section 4.3 reinforces the observation of Bhadra (2010) that PMCMC can require orders of magnitude more computation than iterated filtering.Iterated filtering may have to be repeated multiple times while computing profile likelihood plots, whereas one successful run of PMCMC is sufficient to obtain all required posterior inferences.However, in practice, multiple runs from a range of starting points is always good practice since convergence cannot be reliably assessed on the basis of a single chain.To verify the convergence of the approach or to compare the performance with other approaches, we can use diagnostic plots produced by the plot methods (see Figure 3).

A second example: The Ricker model
In Section 4.6, we will illustrate probe matching (see Section 3.4) using a stochastic version of the Ricker map (Ricker 1954).We switch models to allow direct comparison with Wood (2010), whose synthetic likelihood computations are reproduced below.In particular, the results of Section 4.6 demonstrate frequentist inference using synthetic likelihood and also show that the full likelihood is both numerically tractable and reasonably well behaved, contrary to the claim of Wood (2010).We will also take the opportunity to demonstrate features of pomp that allow acceleration of model codes through the use of R's facilities for compiling and dynamically linking C code.
The Ricker model is another discrete-time model for the size of a population.The population size, N t , at time t is postulated to obey In addition, we assume that measurements, Y t , of N t are themselves noisy, according to where φ is a scaling parameter.As before, we will need to implement the model's stateprocess simulator (rprocess).We have the option of writing these functions in R, as we did with the Gompertz model.However, we can realize manyfold speed-ups by writing these in C. In particular, pomp allows us to write snippets of C code that it assembles, compiles, and dynamically links into a running R session.To begin the process, we will write snippets for the rprocess, rmeasure, and dmeasure components.
Each element of plist is a function of a single argument.Each of these functions can be applied to the data in ricker and to simulated data sets.Calling pomp's function probe results in the application of these functions to the data, and to each of some large number, nsim, of simulated data sets, and finally to a comparison of the two.[Note that probe functions may be vector-valued, so a single probe taking values in R k formally corresponds to a collection of k probe functions in the terminology of Section 3.4.]Here, we will apply probe to the Ricker model at the true parameters and at a wild guess.
R> pb.truth <-probe(ricker, probes = plist, nsim = 1000, seed = 1066L) R> guess <-c(r = 20, sigma = 1, phi = 20, N.0 = 7, e.0 = 0) R> pb.guess <-probe(ricker, params = guess, probes = plist, nsim = 1000, + seed = 1066L) Results summaries and diagnostic plots showing the model-data agreement and correlations among the probes can be obtained by R> summary(pb.truth)R> summary(pb.guess)R> plot(pb.truth)R> plot(pb.guess) An example of a diagnostic plot (using a smaller set of probes) is shown in Figure 4.Among the quantities returned by summary is the synthetic likelihood (Algorithm 5).One can attempt to identify parameters that maximize this quantity; this procedure is referred to in pomp as "probe matching".Let us now attempt to fit the Ricker model to the data using probematching.
R> pm <-probe.match(pb.guess,est = c("r", "sigma", "phi"), transform = TRUE, + method = "Nelder-Mead", maxit = 2000, seed = 1066L, reltol = 1e-08) This code runs optim's Nelder-Mead optimizer from the starting parameters guess in an attempt to maximize the synthetic likelihood based on the probes in plist.Both the starting parameters and the list of probes are stored internally in pb.guess, which is why we need not specify them explicitly here.While probe.matchprovides substantial flexibility in the choice of the optimization algorithm, for situations requiring greater flexibility, pomp provides the function probe.match.objfun,which constructs an objective function suitable for use with arbitrary optimization routines.
To put the synthetic likelihood approach into context, let us compare the results of estimating the Ricker model parameters using probe-matching and using iterated filtering (IF), which is based on the likelihood.The following code runs 600 IF iterations starting at guess: R> mf <-mif(ricker, start = guess, Nmif = 100, Np = 1000, transform = TRUE, + cooling.fraction= 0.95^50, var.factor = 2, ic.lag = 3, + rw.sd = c(r = 0.1, sigma = 0.1, phi = 0.1), max.fail = 50) R> mf <-continue(mf, Nmif = 500, max.fail = 20) Table 4 compares parameters, Monte Carlo likelihoods ( ˆ ), and synthetic likelihoods ( ˆ S , based on the probes in plist) at each of (a) the guess, (b) the truth, (c) the MLE from mif, and (d) the maximum synthetic likelihood estimate (MSLE) from probe.match.These results demonstrate that it is possible, and indeed not difficult, to maximize the likelihood for this model, contrary to the claim of Wood (2010).Since synthetic likelihood discards some of the by means of maximum likelihood via iterated filtering (Algorithm 3).The row labeled "guess" contains the point at which both algorithms were initialized.That labeled "truth" contains the true parameter value, i.e., that at which the data were generated.The rows labeled "MLE" and "MSLE" show the estimates obtained using iterated filtering and maximum synthetic likelihood, respectively.Parameters r, σ, and τ were estimated; all others were held at their true values.The columns labeled ˆ and ˆ S are the Monte Carlo estimates of the log likelihood and the log synthetic likelihood, respectively; their Monte Carlo standard errors are also shown.While likelihood maximization results in an estimate for which both ˆ and ˆ S exceed their values at the truth, the value of ˆ at the MSLE is smaller than at the truth, an indication of the relative statistical inefficiency of maximum synthetic likelihood.information in the data, it is not surprising that Table 4 also shows the statistical inefficiency of the maximum synthetic likelihood relative to that of the likelihood.

Bayesian feature matching via ABC
Whereas the synthetic likelihood approach carries out many simulations for each likelihood estimation, ABC (as described in Section 3.5) uses only one.Each iteration of ABC is therefore much quicker, essentially corresponding to the cost of SMC with a single particle or the synthetic likelihood approach with a single simulation.A consequence of this is that ABC cannot determine a good relative scaling of the features within each likelihood evaluation and this must be supplied in advance.One can imagine an adaptive version of ABC which modifies the scaling during the course of the algorithm, but here we do a preliminary calculation to accomplish this.We return to the Gompertz model to faciliate comparison between ABC and PMCMC.
R> plist <-list(probe.mean(var= "Y", transform = sqrt), + probe.acf("Y",lags = c(0, 5, 10, 20)), + probe.marginal("Y",ref = obs(gompertz))) + psim <-probe(gompertz, probes = plist, nsim = 500) + scale.dat<-apply(psim$simvals, 2, sd) R> abc1 <-foreach(i = 1:5, .combine= c) %dopar% { + abc(pomp(gompertz, dprior = gompertz.dprior),Nabc = 4e6, + probes = plist, epsilon = 2, scale = scale.dat,+ proposal = mvn.diag.rw(c(r= 0.01, sigma = 0.01, tau = 0.01))) + } The effective sample size of the ABC chains is lowest for the r parameter (as was the case for PMCMC) and is 430, as compared to 180 for pmcmc in Section 4.4.The total computa- tional effort allocated to abc here matches that for pmcmc since pmcmc used 100 particles for each likelihood evaluation but is awarded 100 times fewer Metropolis-Hastings steps.In this example, we conclude that abc mixes somewhat more rapidly (as measured by total computational effort) than pmcmc.Figure 5 investigates the statistical efficiency of abc on this example.We see that abc gives rise to somewhat broader posterior distributions than the full-information posteriors from pmcmc.As in all numerical studies of this kind, one cannot readily generalize from one particular example: even for this specific model and dataset, the conclusions might be sensitive to the algorithmic settings.However, one should be aware of the possibility of losing substantial amounts of information even when the features are based on reasoned scientific argument (Shrestha et al. 2011;Ionides 2011).Despite this loss of statistical efficiency, points B2-B5 of Section 3.4 identify situations in which ABC may be the only practical method available for Bayesian inference.

Parameter estimation by simulated quasi-likelihood
With the pomp package, it is fairly easy to try a quick comparison to see how nlf (Section 3.6) compares with mif (Section 3.2) on the Gompertz model.Carrying out a simulation study with a correctly specified POMP model is appropriate for assessing computational and statistical efficiency, but does not contribute to the debate on the role of two-step prediction criteria to fit misspecified models (Xia and Tong 2011;Ionides 2011).The nlf implementation we will use to compare to the mif call from Section 4.3 is R> nlf1 <-nlf(gompertz, nasymp = 1000, nconverge = 1000, lags = c(2, 3), + start = c(r = 1, K = 2, sigma = 0.5, tau = 0.5, X.0 = 1), + est = c("r", "sigma", "tau"), transform = TRUE) Figure 6: Comparison of mif and nlf for 10 simulated datasets using two criteria.In both plots, the maximum likelihood estimate (MLE), θ, obtained using iterated filtering is compared with the maximum simulated quasi-likelihood (MSQL) estimate, θ, obtained using nonlinear forecasting.(A) Improvement in estimated log likelihood, ˆ , at the point estimate over that at the true parameter value, θ. (B) Improvement in simulated log quasi-likelihood ˆ Q , at the point estimate over that at the true parameter value, θ.In both panels, the diagonal line is the 1-1 line.
where the first argument is the class 'pomp' object, start is a vector containing model parameters at which nlf's search will begin, est contains the names of parameters nlf will estimate, and lags specifies which past values are to be used in the autoregressive model.The transform = TRUE setting causes the optimization to be performed on the transformed scale, as in Section 4.3.In the call above lags = c(2, 3) specifies that the autoregressive model predicts each observation, y t using y t−2 and y t−3 , as recommended by Kendall et al. (2005).The quasi-likelihood is optimized numerically, so the reliability of the optimization should be assessed by doing multiple fits with different starting parameter values: the results of a small experiment (not shown) indicate that, on these simulated data, repeated optimization is not needed.nlf defaults to optimization by the subplex method (Rowan 1990;King 2015), though all optimization methods provided by optim are available as well.nasymp sets the length of the simulation on which the quasi-likelihood is based; larger values will give less variable parameter estimates, but will slow down the fitting process.The computational demand of nlf is dominated by the time required to generate the model simulations, so efficient coding of rprocess is worthwhile.
Figure 6 compares the true parameter, θ, with the maximum likelihood estimate (MLE), θ, from mif and the maximized simulated quasi-likelihood (MSQL), θ, from nlf.  showing that likelihood-based inference is almost as good as nlf at optimizing the simulated quasi-likelihood criterion which nlf targets.Figure 6 suggests that the MSQL approach may be inefficient, since it can give estimates with poor behavior according to the statistically efficient criterion of likelihood.Another possibility is that this particular implementation of nlf was unfortunate.Each mif optimization took 22.5 sec to run, compared to 3.3 sec for nlf, and it is possible that extra computer time or other algorithmic adjustments could substantially improve either or both estimators.It is hard to ensure a fair comparison between methods, and in practice there is little substitute for some experimentation with different methods and algorithmic settings on a problem of interest.If the motivation for using NLF is preference for 2-step prediction in place of the likelihood, a comparison with SMC-based likelihood evaluation and maximization is useful to inform the user of the consequences of that preference.

A stochastic, seasonal SIR model
A mainstay of theoretical epidemiology, the SIR model describes the progress of a contagious, immunizing infection through a population of hosts (Kermack and McKendrick 1927;Anderson and May 1991;Keeling and Rohani 2008).The hosts are divided into three classes, according to their status vis-à-vis the infection (Figure 7).The susceptible class (S) contains those that have not yet been infected and are thereby still susceptible to it; the infected class (I) comprises those who are currently infected and, by assumption, infectious; the removed class (R) includes those who are recovered or quarantined as a result of the infection.Individuals in R are assumed to be immune against reinfection.We let S(t), I(t), and R(t) represent the numbers of individuals within the respective classes at time t.
It is natural to formulate this model as a continuous-time Markov process.In this process, the numbers of individuals within each class change through time in whole-number increments as discrete births, deaths, and passages between compartments occur.Let N AB be the stochastic counting process whose value at time t is the number of individuals that have passed from compartment A to compartment B during the interval [t 0 , t), where t 0 is an arbitrary starting point not later than the first observation.We use the notation N • S to refer to births and N A • to refer to deaths from compartment A. Let us assume that the per capita birth and death rates, and the rate of transition, γ, from I to R are constants.The S to I transition rate, the so-called force of infection, λ(t), however, should be an increasing function of I(t).For many infections, it is reasonable to assume that the λ(t) is jointly proportional to the fraction of the population infected and the rate at which an individual comes into contact with others.
Here, we will make these assumptions, writing λ(t) = β I(t)/P , where β is the transmission rate and P = S + I + R is the population size.We will go further and assume that birth and death rates are equal and independent of infection status; we will let µ denote the common rate.A consequence is that the expected population size remains constant.
The continuous-time Markov process is fully specified by the infinitesimal increment probabilities.Specifically, writing ∆N (t) = N (t + h) − N (t), we have together with statement that all events of the form Taking expectations of Equations 14 and 15, dividing through by h, and taking the limit as h ↓ 0, one obtains a system of nonlinear ordinary differential equations which is known as the deterministic skeleton of the model (Coulson et al. 2004).Specifically, the SIR deterministic skeleton is The fact that the observed data are linked to an accumulation, as opposed to an instantaneous value, introduces a slight complication, which we discuss below.

Implementing the SIR model in pomp
As before, we will need to write functions to implement some or all of the SIR model's rprocess, rmeasure, and dmeasure components.As in Section 4.5, we will write these components using pomp's Csnippets.Recall that these are snippets of C code that pomp automatically assembles, compiles, and dynamically loads into the running R session.
To start with, we will write snippets that specify the measurement model (rmeasure and dmeasure): R> rmeas <-" + cases = rnbinom_mu(theta, rho * H); + " R> dmeas <-" + lik = dnbinom_mu(cases, theta, rho * H, give_log); + " Here, we are using cases to refer to the data (number of reported cases) and H to refer to the true incidence over the reporting interval.The negative binomial simulator rnbinom_mu and density function dnbinom_mu are provided by R. The logical flag give_log requests the likelihood when FALSE, the log likelihood when TRUE.Notice that, in these snippets, we never declare the variables; pomp will ensure that the state variable (H), observable (cases), parameters (theta, rho), and likelihood (lik) are defined in the contexts within which these snippets are executed.
For the rprocess portion, we could simulate from the continuous-time Markov process exactly (Gillespie 1977); the pomp function gillespie.simimplements this algorithm.However, for practical purposes, the exact algorithm is often prohibitively slow.If we are willing to live with an approximate simulation scheme, we can use the so-called "tau-leap" algorithm, one version of which is implemented in pomp via the euler.simplug-in.This algorithm holds the transition rates λ, µ, γ constant over a small interval of time ∆t and simulates the numbers of births, deaths, and transitions that occur over that interval.It then updates the state variables S, I, R accordingly, increments the time variable by ∆t, recomputes the transition rates, and repeats.Naturally, as ∆t → 0, this approximation to the true continuous-time process becomes better and better.The critical feature from the inference point of view, however, is that no relationship needs to be assumed between the Euler simulation interval ∆t and the reporting interval, which itself does not even need to be the same from one observation to the next.
Under the above assumptions, the number of individuals leaving any of the classes by all available routes over a particular time interval is a multinomial process.For example, if ∆N SI and ∆N S are the numbers of S individuals acquiring infection and dying, respectively, over the Euler simulation interval [t, t + ∆t), then where 1 − e −(λ(t)+µ) ∆t , p S→ = µ λ(t) + µ 1 − e −(λ(t)+µ) ∆t . ( By way of shorthand, we say that the random variable (∆N SI , ∆N S ) in Equation 18has an Euler-multinomial distribution.pomp provides convenience functions for such distributions, which arise with some frequency in compartmental models.Specifically, the functions reulermultinom and deulermultinom respectively draw random deviates from, and evaluate the probability mass function of, such distributions.As the help pages relate, reulermultinom and deulermultinom parameterize the Euler-multinomial distributions by the size (S(t) in Equation 18), rates (λ(t) and µ), and time interval ∆t.Obviously, the Euler-multinomial distributions generalize to an arbitrary number of exit routes.
The help page (?euler.sim)informs us that to use euler.sim,we need to specify a function that advances the states from t to t + ∆t.Again, we write this in C to realize faster run-times: R> sir.As before, pomp will ensure that the undeclared state variables and parameters are defined in the context within which the snippet is executed.Note, however, that in the above we do declare certain local variables.In particular, the rate and dN arrays hold the rates and 1.For many infections, the contact rate is seasonal: β = β(t) varies in more or less periodic fashion with time.
2. The host population may not be truly closed: imported infections arise when infected individuals visit the host population and transmit.
3. The host population does not need to be constant in size.If we have data, for example, on the numbers of births occurring in the population, we can incorporate this directly into the model.
4. Stochastic fluctuation in the rates λ, µ, and γ can give rise to extrademographic stochasticity, i.e., random process variability beyond the purely demographic stochasticity we have included so far.
To incorporate seasonality, we would like to assume a flexible functional form for β(t).Here, we will use a three-coefficient Fourier series: There are a variety of ways to account for imported infections.Here, we will simply assume that there is some constant number, ι, of infected hosts visiting the population.Putting this together with the seasonal contact rate results in a force of infection λ(t) = β(t) (I(t) + ι) /N .

Conclusion
The pomp package is designed to be both a tool for data analysis based on POMP models and a sound platform for the development of inference algorithms.The model specification language provided by pomp is very general.Implementing a POMP model in pomp makes a wide range of inference algorithms available.Moreover, the separation of model specification from the inference algorithm facilitates objective comparison of alternative models and methods.The examples demonstrated in this paper are relatively simple, but the package has been instrumental in a number of scientific studies (e.g., King et al. 2008;Bhadra et al. 2011;Shrestha et al. 2011;Earn et al. 2012;Roy et al. 2012;Shrestha et al. 2013;Blackwood et al. 2013a,b;Lavine et al. 2013;He et al. 2013;Bretó 2014;King et al. 2015).
As a development platform, pomp is particularly convenient for implementing algorithms with the plug-and-play property, since models will typically be defined by their rprocess simulator, together with rmeasure and often dmeasure, but can accommodate inference methods based on other model components such as dprocess and skeleton (the deterministic skeleton of the latent process).As an open-source project, the package readily supports expansion, and the authors invite community participation in the pomp project in the form of additional inference algorithms, improvements and extensions of existing algorithms, additional model/data examples, documentation contributions and improvements, bug reports, and feature requests.
Complex models and large datasets can challenge computational resources.With this in mind, key components of the pomp package are written in C, and pomp provides facilities for users to write models either in R or, for the acceleration that typically proves necessary in The package contains more examples (via pompExample), which can be used as templates for implementation of new models; the R and C codes underlying these examples are provided with the package.In addition, pomp provides a number of interactive demos (via demo).Further documentation and an introductory tutorial are provided with the package and on the pomp website, http://kingaa.github.io/pomp.

Figure 2 :
Figure 2: Convergence plots can be used to help diagnose convergence of the iterated filtering (IF) algorithm.These and additional diagnostic plots are produced when plot is applied to a class 'mif' or class 'mifList' object.

Figure 3 :
Figure 3: Diagnostic plots for the PMCMC algorithm.The trace plots in the left column show the evolution of 5 independent MCMC chains after a burn-in period of length 20000.Kernel density estimates of the marginal posterior distributions are shown in the right column.The effective sample size of the 5 MCMC chains combined is lowest for the r variable, being equal to 180: the use of 40000 proposal steps in this case is a modest number.The density plots at right show the estimated marginal posterior distributions.The vertical line corresponds to the true value of each parameter.

Figure 4 :
Figure 4: Results of plot on a class 'probed.pomp'object.Above the diagonal, the pairwise scatterplots show the values of the probes on each of the 1000 data sets.The red lines show the values of each of the probes on the data.The panels along the diagonal show the distributions of the probes on the simulated data, together with their values on the data and two-sided p values.The numbers below the diagonal are the Pearson correlations between the corresponding pairs of probes.

Figure 5 :
Figure 5: Marginal posterior distributions using full information via pmcmc (solid line) and partial information via abc (dashed line).Kernel density estimates are shown for the posterior marginal densities of log 10 (r) (left panel), log 10 (σ) (middle panel), and log 10 (τ ) (right panel).The vertical lines indicate the true values of each parameter.

Figure 9 :
Figure 9: One realization of the SIR model with seasonal contact rate, imported infections, and extrademographic stochasticity in the force of infection.

Table 1
Executing pompExample() lists other examples provided with the package.

Table 2 :
Inference methods for POMP models.For those currently implemented in pomp, the function name and a reference for description are provided in parentheses.Standard expectation-maximization (EM) and Markov chain Monte Carlo (MCMC) algorithms are not plug-and-play since they require evaluation of f Xn|X n−1 (x n | x n−1 ; θ).The Kalman filter and extended Kalman filter are not plug-and-play since they cannot be implemented based on a model simulator.The Kalman filter provides the likelihood for a linear, Gaussian model.The extended Kalman filter employs a local linear Gaussian approximation

Table 4 :
Parameter estimation by means of maximum synthetic likelihood (Algorithm 5) vs.
Diagram of the SIR epidemic model.The host population is divided into three classes according to infection status: S, susceptible hosts; I, infected (and infectious) hosts; R, recovered and immune hosts.Births result in new susceptibles and all individuals have a common death rate µ.Since the birth rate equals the death rate, the expected population size, P = S + I + R, remains constant.The S→I rate, λ, called the force of infection, depends on the number of infectious individuals according to λ(t) = β I/N .The I→R, or recovery, rate is γ.The case reports, C, result from a process by which new infections are recorded with probability ρ.Since diagnosed cases are treated with bed-rest and hence removed, infections are counted upon transition to R.
C(t 1 , t 2 ), recorded within a given reporting interval [t 1 , t 2 ) might perhaps be modeled by a negative binomial processC(t 1 , t 2 ) ∼ NegBin(ρ ∆N SI (t 1 , t 2 ), θ), (17)where ∆N SI (t 1 , t 2 ) is the true incidence (the accumulated number of new infections that have occured over the [t 1 , t 2 ) interval), ρ is the reporting rate, (the probability that an infection is observed and recorded), θ is the negative binomial "size" parameter, and the notation is meant to indicate that EC(t 1 , t 2 ) | ∆N SI (t 1 , t 2 )It is typically impossible to monitor S, I, and R, directly.It sometimes happens, however, that public health authorities keep records of cases, i.e., individual infections.The number of cases,