# Leading indicators of mosquito-borne disease elimination

- 2.4k Downloads
- 11 Citations

## Abstract

Mosquito-borne diseases contribute significantly to the global disease burden. High-profile elimination campaigns are currently underway for many parasites, e.g., *Plasmodium* spp., the causal agent of malaria. Sustaining momentum near the end of elimination programs is often difficult to achieve and consequently quantitative tools that enable monitoring the effectiveness of elimination activities after the initial reduction of cases has occurred are needed. Documenting progress in vector-borne disease elimination is a potentially important application for the theory of critical transitions. Non-parametric approaches that are independent of model-fitting would advance infectious disease forecasting significantly. In this paper, we consider compartmental Ross-McDonald models that are slowly forced through a critical transition through gradually deployed control measures. We derive expressions for the behavior of candidate indicators, including the autocorrelation coefficient, variance, and coefficient of variation in the number of human cases during the approach to elimination. We conducted a simulation study to test the performance of each summary statistic as an early warning system of mosquito-borne disease elimination. Variance and coefficient of variation were highly predictive of elimination but autocorrelation performed poorly as an indicator in some control contexts. Our results suggest that tipping points (bifurcations) in mosquito-borne infectious disease systems may be foreshadowed by characteristic temporal patterns of disease prevalence.

## Keywords

Transcritical bifurcation Critical transition Vector-borne disease Disease elimination Critical slowing down Vector-borne disease model Malaria## Introduction

Vector-borne diseases constitute approximately 17 % of the estimated worldwide infectious disease burden (World Health Organization 2014). Mosquito-borne infections, in particular, contribute substantially to this burden, causing millions of deaths and hundreds of millions of cases annually (World Health Organization 2013). Mosquito-borne pathogen elimination remains a high priority on the global public health agenda, with elimination campaigns underway for malaria, onchocerciasis, and lymphatic filariasis (Hopkins 2013). The introduction of insecticides such as DDT in the mid twentieth century led to the implementation of global eradication campaigns for many of the most deleterious mosquito-borne diseases, including malaria, dengue virus, and yellow fever virus. These campaigns were partially successful in that diseases were eliminated in some countries and effectively controlled in others. For example, malaria was eliminated in 79 countries between 1945 and 2010 (Chiyaka et al. 2013) and the reduction of *Aedes aegypti* led to the near elimination of dengue and yellow fever in the Americas (Gratz 1999; Gubler 1998). However, the incidence of mosquito-borne infections has resurged since the 1970s due to a nexus of factors including globalization, increased global movement of people, development of vector resistance to insecticides, changes in land use, and loss of financial support and public health infrastructure for control and elimination (Institute of Medicine (US) Forum on Microbial Threats 2008; Gubler 2010; Mackenzie et al. 2004). Sustaining political will near the end of elimination campaigns is often difficult to achieve since there may be marginal return on investment (in terms of case reductions) (Cohen et al. 2012). Government and public health agencies require markers of progress to justify the billions of dollars that are spent on interventions (Cohen et al. 2010). Quantitative evaluation tools that will encourage governments and philanthropic organizations to choose the optimal level of investment in control and elimination activities after the initial reduction of cases has occurred are needed.

Mosquito-borne disease elimination, which involves halting transmission through interventions until no parasites remain (Breman et al. 2011; Klepac et al. 2013), e.g., through vector control strategies and/or treatment of infected individuals, corresponds to a critical transition in the disease transmission system. Criticality occurs at the point where the basic reproduction number, *R* _{0}, the average number of secondary infected cases arising from a single infected case in an entirely susceptible population, is equal to one. Tipping points in complex systems may be described mathematically as bifurcations if the change in the external driver variable is slow relative to the characteristic speed of the internal variables. If control efforts are implemented sufficiently slowly over time such that elimination eventually occurs, the elimination strategy causes the system to cross a bifurcation point. The critical transition may be anticipated because prior to reaching the dynamical threshold, the system gradually loses stability (“critical slowing down” Strogatz 1994; Scheffer et al. 2009). In continuous-time systems, the loss of stability is measured by the dominant eigenvalue of the linearized system. Signatures of the slow return rate to the underlying equilibrium may be detectable through summary statistics. Statistical patterns of critical slowing down may be a diagnostic of the proximity to disease elimination (e.g., O’Regan and Drake 2013). To establish how close a disease is to elimination or emergence, typically, a stochastic epidemic model (e.g., a compartmental model or a branching process model) is fitted to time series data to estimate the basic reproduction number of a disease. The disadvantage of this approach is having to specify parametric models that make strong assumptions about the form of transmission. Additionally, the fit of stochastic epidemic models to data is usually evaluated using a goodness-of-fit criterion such as statistical likelihood, but many models may yield near-identical fits. Detection of critical slowing down would circumvent the problem of model structural uncertainty and improve efficiency by making use of higher order information in time series that is not used by other approaches that focus on trends in the mean. Methods that detect the presence of critical slowing down would therefore capture key structural features of transmission that are exhibited by a large family of mosquito-borne disease models. Non-parametric approaches for emergence and elimination prediction that are independent of model-fitting and evaluation would advance infectious disease forecasting significantly.

In this paper, we use stochastic differential equations to model the gradual implementation of four control activities on mosquito-borne disease elimination: the use of bed nets that hamper the per-capita human biting rate, e.g., (Nyarango et al. 2006), indoor residual insectide spraying procedures that shorten adult mosquito lifespan (Giardina et al. 2014) or late-acting insecticides (Read et al. 2009), the administration of drugs that reduce the human infectious period (Lawpoolsri et al. 2009), and the reduction of mosquito abundance due to administration of insecticides or elimination of breeding sites through larviciding or water drainage (Goodman et al. 1999; Tusting et al. 2013). We use the models to develop and validate leading indicator summary statistics for documenting disease elimination in mosquito-borne disease systems. We focus our analysis on malaria but our findings are relevant to other mosquito-borne infections such as yellow fever. Our results indicate that critical slowing down is detectable prior to mosquito-borne disease elimination.

## Theory for elimination of vector-borne diseases

### Mean field theory of Ross-Macdonald model

*H*(

*t*) and

*M*(

*t*) respectively, the model is

*N*

_{ h }is assumed to be constant and consequently, the number of susceptible human hosts

*S*

_{ h }is equal to (

*N*

_{ h }−

*H*), i.e.,

*N*

_{ m }is assumed constant and the number of susceptible mosquitoes

*S*

_{ m }is (

*N*

_{ m }−

*M*), i.e.,

*R*

_{0}are listed in Table 1. The model has two equilibria, the disease-free equilibrium (0, 0) and the endemic equilibrium

*R*

_{0}> 1, the endemic equilibrium is stable and the disease-free equilibrium is unstable. When

*R*

_{0}<1, the stabilities of the equilibria are switched. Mathematically, a transcritical bifurcation, whereby the disease-free equilibrium and endemic equilibrium meet and exchange stability, occurs at the point where the basic reproduction number is equal to one (Keeling and Rohani 2008). Transcritical bifurcation diagrams corresponding to various control measures are shown in Fig. 1. Interventions due to the biting rate and mosquito abundance modify the endemic equilbrium differently than those that affect the per-capita mosquito mortality rate and per-capita human host recovery rate. The concave down trend in Figs. 1a, b suggests that successful pathogen extinction may only occur at the very end of elimination campaigns that focus on either of reduction of mosquito abundance or mosquito biting rates.

Variables and parameters of the time-varying Ross-Macdonald models

Variable | Expression | Value |
---|---|---|

Number of infectious humans | | |

Number of infectious mosquitoes | | |

Per-capita human recovery rate | | 0.01 day |

Per-capita mosquito mortality rate | | 0.1 day |

Human population size | | 1000 |

Per-capita mosquito biting rate | | 0.3 day |

Transmission efficiency from mosquitoes to humans | | 0.5 (Smith and McKenzie 2004) |

Transmission efficiency from humans to mosquitoes | | 0.5 (Smith and McKenzie 2004) |

Mosquito population size | | 10000 |

Basic reproduction number | \(R_{0} = \frac {k^{2}p q N_{m}}{\delta \mu N_{h}}\) | |

Time critical point is reached | | |

Critical biting rate | \(k^{\ast } = \sqrt {(N_{h}\delta \mu )/(pqN_{m})} \) | |

Critical mosquito population size | \(N_{m}^{\ast } =(N_{h}\delta \mu )/(pqk^{2}) \) | |

Critical recovery rate | | |

Critical mortality rate | | |

Value of biting rate prior to application of control measures | | 0.3 day |

Rate of change in biting rate | | 0.0001 day |

Mosquito population size prior to application of control measures | | 10000 |

Rate of change of mosquito population size | | 1 day |

Mosquito mortality rate prior to application of control measures | | 0.1 day |

Rate of change of mosquito mortality rate | | 0.0025 day |

Value of human recovery rate prior to application of control measures | | 0.01 day |

Rate of change of recovery rate | | 0.001 day |

Time-varying biting rate | \(k(t) = \left \{\begin {array}{ll} k_{0} - k_{1} t, &t < t^{\ast } \\ k^{\ast } &t \geq t^{\ast } \end {array}\right .\) | |

Time-varying mosquito population size | \(N_{m}(t) = \left \{\begin {array}{ll} N_{m0} - N_{m1} t, &t < t^{\ast } \\ N_{m}^{\ast } &t \geq t^{\ast } \end {array}\right .\) | |

Time-varying mortality rate | \(\delta (t) = \left \{\begin {array}{ll} \delta _{0} + \delta _{1} t, &t < t^{\ast } \\ \delta ^{\ast } &t \geq t^{\ast } \end {array}\right .\) | |

Time-varying recovery rate | \(\mu (t) = \left \{\begin {array}{ll} \mu _{0} + \mu _{1} t, &t < t^{\ast } \\ \mu ^{\ast } &t \geq t^{\ast } \end {array}\right .\) | |

Environmental noise strength | | 0.05 |

*k*of mosquitoes decreases at rate

*k*

_{1}due to gradual increases in bed net usage, we write down the Ross-Macdonald equations as a fast-slow system (e.g., Kuehn 2011),

*𝜖*is a parameter that denotes the speed of evolution of the biting rate (assumed to be gradual, i.e., 0<

*𝜖*<<1) and the function

*f*describes the change in biting rate. The biting rate evolves over time to its critical value

*k*

^{∗}(achieved at

*R*

_{0}=1 at time

*t*=

*t*

^{∗}) according to

*μ*(

*t*)), the reduction of mosquito population numbers (

*N*

_{ m }(

*t*)), and the increase of per-capita mosquito mortality rate (

*δ*(

*t*)) over time

*t*through control applications (Table 1).

#### Stochastic description of Ross-Macdonald model

*N*

_{ h }(constant and equal to

*N*

_{ m }for the mosquito population). Individuals may transition from being susceptible to being infectious. Table 2 shows the transition probability fluxes into and out of the infectious populations (

*H*,

*M*). We use the diffusion approximation outlined in Allen (2003) to derive stochastic differential equations that incorporate demographic stochasticity and time-varying control measures. In Appendix 1, we provide the derivation of the equations.

Transition probability fluxes for Ross-Macdonald model. Numbers of infectious humans and populations are denoted by *X*=(*H*, *M*)^{ T }. The vector Δ*X*=(Δ*X* ^{ i })=(*H*(*t*+Δ*t*)−*H*(*t*), *M*(*t*+Δ*t*)−*M*(*t*))^{ T } denotes change in state, *i*=1, 2, …, 5

Event | Change in population sizes | (Δ | Transition probability |
---|---|---|---|

Infection of susceptible humans | \((H, M)\rightarrow (H+1, M)\) | \( \left (\begin {array}{l} 1 \\ 0 \end {array}\right ) \) | \(p_{1} = \frac {k(t) p M(N_{h}-H)}{N_{h}} {\Delta } t\) |

Recovery of infectious humans | \((H, M) \rightarrow (H-1, M)\) | \(\left (\begin {array}{l} -1 \\ 0 \end {array}\right )\) | |

Infection of susceptible mosquitoes | \((H, M) \rightarrow (H, M+1)\) | \(\left (\begin {array}{l} 0 \\ 1 \end {array}\right )\) | \( p_{3} = \frac {k(t) q H(N_{m}(t)-M)}{N_{h}} {\Delta } t\) |

Death of infectious mosquitoes | \((H, M) \rightarrow (H, M-1)\) | \(\left (\begin {array}{l} 0 \\ -1 \end {array}\right )\) | |

No change | \((H, M) \rightarrow (H, M)\) | \(\left (\begin {array}{l} 0 \\ 0 \end {array}\right )\) | \( 1 - \displaystyle \sum \limits _{i=1}^{4} p_{i} \) |

*t*. Specifically, in a small time interval Δ

*t*, we assume each time-varying control measure

*g*(

*t*) is subject to environmental stochasticity as follows,

*η*is a normal random variate with mean zero and unit variance and

*σ*denotes the strength of the environmental noise. Equation 5 replaces the appropriate parameter in the drift term obtained through the diffusion approximation (A.1 in Appendix 1), to represent an environmental perturbation to each process. For example, assuming per-capita biting rate is gradually reduced through control efforts, the change in the human host and mosquito populations in a small time interval [

*t*,

*t*+Δ

*t*) is

*η*

_{ i }are normally distributed random variables with mean zero and variance of unity. Letting Δ

*t*→0, and assuming existence and uniqueness of the stochastic integral (Allen 2003), then \(\eta _{i} \sqrt {\Delta t} \rightarrow 0\) and the system of equations converges in the mean square sense to a system of Ito stochastic differential equations,

Time-varying Ross-Macdonald equations with demographic and environmental stochasticity

Biting rate | \(\begin {array}{llll} dH = (\frac {k(t) p}{N_{h}} M(N_{h} - H) - \mu H) dt +\sqrt {\frac {k (t)p}{N_{h}} M(N_{h} - H) + \mu H} dW_{1} + \sigma \frac {M p (N_{h} - H)}{N_{h}} dW_{3} \end {array}\) |

\(\begin {array}{llll} dM = (\frac {k(t) q}{N_{h}} H(N_{m} - M) - \delta M) dt+\sqrt {\frac {k(t) q}{N_{h}} H(N_{m} - M) + \delta M}dW_{2} + \sigma \frac {H q (N_{m}-M)} {N_{h}} dW_{3} \end {array}\) | |

Mosquito population size | \(\begin {array}{llll} dH = (\frac {k p}{N_{h}} M(N_{h} - H) - \mu H) dt +\sqrt {\frac {k p}{N_{h}} M(N_{h} - H) + \mu H} dW_{1} \end {array}\) |

\(\begin {array}{lll} dM = (\frac {k q}{N_{h}} H(N_{m}(t) - M) - \delta M) dt+\sqrt {\frac {k q}{N_{h}} H(N_{m} - M) + \delta M}dW_{2} + \sigma \frac {k H q} {N_{h}} dW_{3} \end {array}\) | |

Recovery rate | \(\begin {array}{llll} dH = (\frac {k p}{N_{h}} M(N_{h} - H) - \mu (t) H) dt +\sqrt {\frac {k p}{N_{h}} M(N_{h} - H) + \mu (t) H} dW_{1} + \sigma H dW_{3} \end {array}\) |

\(\begin {array}{lll} dM = (\frac {k q}{N_{h}} H(N_{m} - M) - \delta M) dt+\sqrt {\frac {k q}{N_{h}} H(N_{m} - M) + \delta M}dW_{2} \end {array}\) | |

Mortality rate | \(\begin {array}{llll} dH = (\frac {k p}{N_{h}} M(N_{h} - H) - \mu H) dt +\sqrt {\frac {k p}{N_{h}} M(N_{h} - H) + \mu H} dW_{1} \end {array}\) |

\(\begin {array}{llll} dM = (\frac {k q}{N_{h}} H(N_{m} - M) - \delta (t) M) dt+\sqrt {\frac {k q}{N_{h}} H(N_{m} - M) + \delta (t) M}dW_{2} + \sigma M dW_{3} \end {array}\) |

### Leading indicators of elimination

To calculate leading indicators of disease elimination, we need to analyze the properties of the time-varying Ross-Macdonald system, in the neighborhood of the equilibrium point that would be present in the *𝜖*→0 limit (e.g., O’Regan and Drake 2013; Kuehn 2011). In the stochastic system, the equilibrium is a quasi-stationary state. Quantifying the behavior of the deviations from the equilibrium point can be achieved by deriving a locally linear approximate description of the probability distribution that is the solution of the forward Kolmogorov equation (A.6). Since the Ross-Macdonald system does not exhibit long transients and the time-varying functions represent non-stationary processes that are changing gradually through time, linearization of the drift and diffusion terms of the equations in Table 3 gives reasonable information for the qualitative behavior of fluctuations about the quasi-stationary state.

*λ*

_{1}and

*λ*

_{2}, respectively. Changes in these summary statistics as the bifurcation point is approached are indicators of critical slowing down.

Analytical expressions for quasi-stationary statistics about the endemic infectious human quasi-steady state *H* ^{∗} expressed in terms of the eigenvalues

Power spectrum | \(\frac {2(\alpha _{H}+D_{11} \omega ^{2})}{(\omega ^{2}-\lambda _{1}\lambda _{2})^{2}+(\lambda _{1}+ \lambda _{2})^{2} \omega ^{2}}\) |

Autocorrelation | \(\displaystyle \frac {1}{\pi }{\int }_{0}^{\infty } S_{I}(\omega ) \cos (\omega \tau ) d\omega \) |

Variance | \(\frac {\alpha _{H} +\lambda _{1}\lambda _{2} D_{11}}{2 \lambda _{1}\lambda _{2} (-\lambda _{1}- \lambda _{2})}\) |

Coefficient of variation | \(\frac {(\alpha _{H} +\lambda _{1}\lambda _{2} D_{11})^{\frac {1}{2}}}{\sqrt {2 \lambda _{1}\lambda _{2} (-\lambda _{1}- \lambda _{2})} H^{\ast }}\) |

Variable | Expression |
---|---|

| \(\frac {N_{h} (k(t)^{2} N_{m}(t) p q - N_{h} \delta (t)\mu (t))}{k(t) q (k(t) N_{m}(t) p + N_{h} \mu (t))}\) |

| \(\frac {k(t)^{2} N_{m}(t) p q - N_{h} \delta (t) \mu (t)}{k(t) p(k(t) q +\delta (t))}\) |

| − |

| |

| |

| − |

| |

| − |

| \(a_{22}^{2} D_{11}+a_{12}^{2} D_{22}-2D_{12}a_{12}a_{22}\) |

Variables substitutions for terms of the variance-covariance matrix of time-varying Ross-Macdonald models with demographic and environmental noise

Coefficient | Biting rate | Recovery rate |
---|---|---|

| \(\frac {k(t) p}{N_{h}} M^{\ast }(N_{h} - H^{\ast }) + \mu (t) H^{\ast } + (\sigma /Nh)^{2} p^{2} (M^{\ast })^{2} (N_{h} - H^{\ast })^{2} \) | \( \frac {k p}{N_{h}} M^{\ast }(N_{h} - H^{\ast }) + \mu (t)H^{\ast } + (\sigma H^{\ast })^{2}\) |

| ( | 0 |

| ( | 0 |

| \(\frac {k(t) q}{N_{h}}H^{\ast }(N_{m} - M^{\ast }) + \delta M^{\ast }+(\sigma /Nh)^{2}q^{2}(H^{\ast })^{2}(N_{m} - M^{\ast })^{2}\) | \(\frac {k q}{N_{h}}H^{\ast }(N_{m} - M^{\ast }) + \delta M^{\ast }\) |

Coefficient | Mosquito population size | Mortality rate |

| \(\frac {k p}{N_{h}}M^{\ast }(N_{h} - H^{\ast }) + \mu H^{\ast }\) | \(\frac {k p}{N_{h}} M^{\ast }(N_{h} - H^{\ast }) + \mu H^{\ast }\) |

| 0 | 0 |

| 0 | 0 |

| \(\frac {k q}{N_{h}}H^{\ast }(N_{m}(t) - M^{\ast }) + \delta M^{\ast } +\sigma ^{2} \frac {k q}{N_{h}}(H^{\ast })^{2}\) | \(\frac {k q}{N_{h}}H^{\ast }(N_{m} - M^{\ast }) + \delta (t) M^{\ast }+(\sigma M^{\ast })^{2}\) |

### Simulations

The preceding sections present an analytical theory of leading indicators of elimination for stochastic time-varying Ross-Macdonald models. To investigate the results of this theory for a particular parameter set, we calculated leading indicators of disease elimination in infectious human hosts using the models in Table 3, assuming that (a) the mean number of infectious individuals in the *𝜖*→0 limit is given by the deterministic endemic equilibrium (*H* ^{∗}, *M* ^{∗}) (2) or (b) by assuming it is given by the current state (*H*(*t*), *M*(*t*) of the system approaching elimination. We selected parameters consistent with malaria (Smith and McKenzie 2004) (Table 1).

To test the robustness of the theoretical predictions in Table 4, we simulated the approach to elimination 500 times for each of the models in Table 3. For each control scenario, the infectious time series approaching elimination were sampled at weekly intervals in each simulation, assuming perfect detection (no underreporting). The transcritical bifurcation in these scenarios was approached over a long time frame (e.g., 2800 days (400 weeks) when *k* _{1}=1/10, 000 day^{−1}) in the fast-slow Ross-Macdonald model approaching elimination. The length of the time series in each simulation set depended on how long it took to reach the critical threshold *R* _{0}=1 in each model (400 weeks in the changing biting rate model, 1280 weeks in the changing mosquito mortality model, 320 weeks in the changing recovery rate model, and 1424 weeks in the changing mosquito abundance model, respectively).

### Analysis over a moving window

Our simulation study evaluates the performance of summary statistics (lag-1 autocorrelation, variance, and coefficient of variation) as an early warning system. Here, we describe an algorithm to process input time series data that is independent of a specific model. To investigate the robustness of the changes in the early warning theoretical predictions over a moving window, i.e., as they would be used in online analysis of surveillance data, we used Gaussian filtering to remove the influence of the slowly varying trend (Dakos et al. 2008). We fitted a Gaussian kernel smoothing function with a fixed bandwidth across the infectious human host time series up to the time that the transcritical bifurcation was predicted (*t* ^{∗}). We obtained the residuals by subtracting the fit from each time series. We calculated the lag-1 autocorrelation, variance, and coefficient of variation of the residuals over a moving window half the length of each time series. We calculated the lag-1 autocorrelation coefficient of each replicate using the acf function in *R*. The coefficient of variation was found by calculating the mean and standard deviation of each infectious replicate. The median and 95 % prediction intervals for each of the statistics were calculated over the 500 replicates of each model. The prediction intervals were calculated using the quantile function in *R*. To quantify the association between time and the statistic for each replicate, we used Kendall’s correlation coefficient *τ*, a non-parametric statistic of association between two quantities that has values between 1 (positive correlation) and -1 (negative correlation). By repeating the calculation for each realization of leading indicators, we generated distributions of the temporal correlation *τ* for all of the simulation sets approaching criticality.

To assess leading indicator performance, we adapted the method described in Boettiger and Hastings (2012) to distinguish between statistics obtained from systems approaching elimination from those calculated from quasi-stationary systems. Distributions of Kendall’s *τ* were additionally generated using the statistics obtained from realizations of quasi-stationary epidemic systems (null models) that were processed using the procedure outlined above. That is, the null models have the same environmental noise structure as the test models (the systems of stochastic differential equations in Table 3) but assume *k*(*t*) = *k* _{0}, *N* _{ m }(*t*) = *N* _{ m0}, *δ*(*t*) = *δ* _{0} and *μ*(*t*) = *μ* _{0} in each model respectively. Null models were initialized from the deterministic equilibrium calculated from the values of *k* _{0}, *N* _{ m0}, *μ* _{0}, and *δ* _{0} in Table 1 and were simulated for the same length of time as required for the transition to be approached in the test models (systems approaching elimination).

From the distributions of Kendall’s *τ* obtained from null and test model realizations, we calculated receiver-operating characteristic (ROC) curves. The receiver-operating curve plots diagnostic sensitivity (true positive rate, y-axis) as a function of one minus diagnostic specificity (false positive rate). The false positive rate is the integral of the distribution of the correlation coefficent under the quasi-stationary system, to the right of a threshold line, and the true positive rate is the integral of the distribution of the correlation coefficent under the system approaching elimination. These rates make up the curve. If curves lie on or near a line with a constant slope of unity, this would imply the distributions overlap completely, and they cannot be used to distinguish between quasi-stationary systems from those approaching criticality (Boettiger and Hastings 2012; O’Regan and Drake 2013). The magnitude of departure from this line is summarized by the area under the curve, which reaches its maximum at one. If the area under the curve is close to one, then there is near perfect detection of sensitivity.

### Underreporting

To assess the effects of underreporting on leading indicator performance, we binomially sampled each time series in each set of stationary and supercritical intervention simulations with rates of detection ranging from 20 to 90 %. We calculated the ROC curves and reported the area under the curve for each simulation set.

## Results

### Predictions for mean behavior of leading indicators

### Simulation study results

### Imperfect detection results

## Discussion

Detecting the onset of critical transitions in infectious disease systems such as the emergence of novel pathogens and the elimination of disease would be of tremendous value for public health. The goal of our study was to develop a theory of leading indicators for vector-borne infectious disease transmission systems that follow the assumptions of the Ross-Macdonald model and that are forced through a critical transition through gradual increases in control efforts. Our main results include analytical expressions for summary statistics for a family of time-varying Ross-Macdonald models approaching elimination (Table 4). Numerical calculations assuming time-varying control activities indicate that trends in the observable statistics are discernible. Testing the robustness of these predictions via a simulation study suggests that critical slowing down is detectable in human-host fluctuations of Ross-Macdonald type mosquito-borne disease systems. These results are in broad agreement with the findings that variance, autocorrelation, and coefficient of variation were predictive of disease elimination, in supercritical SIR and SIS compartmental models with time-varying vaccination uptake and transmission rates respectively (O’Regan and Drake 2013).

The ROC procedure we describe here is not appropriate for measuring leading indicator statistics in time series of case reports. What is observed in practice is a single realization (case report time series) that may lie far from the mean model prediction due to stochastic fluctuations. Rather, we use the theory to calculate the theoretical values of the mean leading indicator summary statistics and compare changes in them as the critical point is approached with changes in median statistics from 500 stochastic simulations over a moving window (“Simulations” and “Analysis over a moving window” sections; and “Results” section). Changes in summary statistics as the critical point is approached are signatures of critical slowing down. Our simulation study tests the performance of each summary statistic as an early warning system, which we define as an information processing system that returns a binary signal at each time in the simulation.

Our online processing algorithm assumes that the input data is the number of infectious individuals at time *t*, but in practice, data would consist of incidence time series (the number of new cases at time *t*). Incidence data will closely agree with the number of currently infected individuals if the incidence time series has the same resolution as the duration of the infectious period (Ferrari et al. 2005). If the system is quasi-stationary or moving towards elimination sufficiently slowly, then the number of infectious individuals at a time equal to the infectious period should equal the number of new cases. However, this issue becomes complicated for diseases such as malaria where case reports in elimination settings may only include very sick individuals that seek treatment and not those who are infected but are not sick enough to seek treatment. Moreover, our results in Fig. 9 suggest that the effects of imperfect detection are not uniform across interventions. Development of theory for detection of critical slowing down accounting for underreporting and the presence of unobserved processes is needed.

The Ross-MacDonald model has long been used as a strategic model for understanding mosquito-borne infectious disease (Smith et al. 2012), but has not been studied under the realistic scenario of a slow environmental forcing. Actual elimination requires the growth of institutions (like hospitals, central government, land conversion like ditching, and mosquito control networks) that are slow to develop and need to be maintained until elimination is reached. While many control campaigns could be considered “pulse” interventions, in that the intervention is deployed as quickly as possible, the models considered here encompass “press” strategies (interventions that are maintained until elimination is reached). Our approach ignores integrated control measures that impact multiple components of vectorial capacity simultaneously, e.g., insecticides targeted on adult mosquitoes that affect mosquito abundance and adult longevity. However, it would be straightforward to derive the theoretical predictions for any such model, simply by appropriate modification of the formulas for the eigenvalues. A limitation of our time-varying models is that they do not include the extrinsic incubation period (EIP), to which transmission can be sensitive to variation. Including a delay in the form of an exposed mosquito class could be important for anticipating elimination if evolution-proof control activities that affect mosquito lifespan are implemented since sustained transmission between hosts and vectors requires the EIP to be less than adult lifespan. Additionally, our simple model of mosquito-borne disease elimination does not include many realistic aspects of mosquito-borne dynamics including waiting time distributions (Smith and McKenzie 2004), temperature dependence in the gonotrophic cycle and mosquito development time (Paaijmans et al. 2013; Read et al. 2009), spatial location of hosts and vectors, and spatial heterogeneity in adult and larval mosquito habitats (Hollingsworth et al. 2014; Perkins et al. 2013; Reiner et al. 2013). Future tactical models examining elimination dynamics should include these complexities but since most models of vector-borne diseases are of the Ross-Macdonald type (Reiner et al. 2013), it is reasonable to develop theoretical predictions for summary statistics using the existing Ross-Macdonald theory, combined with simple assumptions of control activities applied to malaria.

In conclusion, to our knowledge, this study constitutes the first theory for non-parametrically anticipating elimination of mosquito-borne diseases. Our analysis of Ross-Macdonald models parameterized for malaria shows that critical slowing down is detectable in fluctuations in human cases for mosquito-borne parasites approaching elimination, suggesting that disease elimination may be anticipated even in the absence of a detailed understanding of underlying mechanisms. Our work suggests that online algorithms for detecting changes in leading indicators may be achievable, possibly aiding sustainment of the gains made by elimination programs.

## Notes

### Acknowledgments

The authors would like to thank E. O’Dea, S. Bowden, and C. Dibble for helpful comments on the manuscript. This research was partially funded by the National Institutes of Health U01GM110744. The content is solely the responsibility of the authors and does not necessarily reflect the official views of the National Institutes of Health. JWL acknowledges funding support from the Georgia GIFT program. JMD acknowledges funding support from the RAPIDD program of the Science & Technology Directorate, Department of Homeland Security, and the Fogarty International Center, National Institutes of Health. SMO completed this work while a Postdoctoral Fellow at the National Institute for Mathematical and Biological Synthesis, an Institute sponsored by the National Science Foundation through NSF Award #DBI-1300426, with additional support from The University of Tennessee, Knoxville.

### Open Access

This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.

## References

- Allen EJ, Allen LJS, Arciniega A, Greenwood PE (2008) Construction of equivalent stochastic differential equation models. Stochastic Analysis AppliedGoogle Scholar
- Allen LJS (2003) An introduction to stochastic processes with applications to biology. Prentice Hall, Upper Saddle River, NJGoogle Scholar
- Boettiger C, Hastings A (2012) Quantifying limits to detection of early warning for critical transitions. J R Soc Interface 9(75):2527–2539. doi: 10.1098/rsif.2012.0125
- Breman JG, De Quadros CA, Dowdle WR, Foege WH, Henderson DA, John TJ, Levine MM (2011) The role of research in viral disease eradication and elimination programs: lessons for malaria eradication. PLoS Med 8(1):e1000,405. doi: 10.1371/journal.pmed.1000405 CrossRefGoogle Scholar
- Chiyaka C, Tatem AJ, Cohen JM, Gething PW, Johnston G, Gosling R, Laxminarayan R, Hay SI, Smith DL (2013) Infectious disease. The stability of malaria elimination. Science 339(6122):909–10. doi: 10.1126/science.1229509 CrossRefPubMedGoogle Scholar
- Cohen JM, Moonen B, Snow RW, Smith DL (2010) How absolute is zero? An evaluation of historical and current definitions of malaria elimination. Malar J 9(1):213. doi: 10.1186/1475-2875-9-213 CrossRefPubMedPubMedCentralGoogle Scholar
- Cohen JM, Smith DL, Cotter C, Ward A, Yamey G, Sabot OJ, Moonen B (2012) Malaria resurgence: a systematic review and assessment of its causes. Malar J 11(1):122. doi: 10.1186/1475-2875-11-122 CrossRefPubMedPubMedCentralGoogle Scholar
- Dakos V, Scheffer M, Van Nes EH, Brovkin V, Petoukhov V, Held H (2008) Slowing down as an early warning signal for abrupt climate change. Proc Natl Acad Sci USA 105(38):14,308–14,312. doi: 10.1073/pnas.0802430105 CrossRefGoogle Scholar
- Dakos V, van Nes EH, D’Odorico P, Scheffer M (2012) Robustness of variance and autocorrelation as indicators of critical slowing down. Ecology 93(2):264–271. doi: 10.1890/11-0889.1 CrossRefPubMedGoogle Scholar
- Ferrari MJ, Bjørnstad ON, Dobson AP (2005) Estimation and inference of R0 of an infectious pathogen by a removal method. Math Biosci 198(1):14–26. doi: 10.1016/j.mbs.2005.08.002 CrossRefPubMedGoogle Scholar
- Giardina F, Kasasa S, Sié A, Utzinger J, Tanner M, Vounatsou P (2014) Effects of vector-control interventions on changes in risk of malaria parasitaemia in sub-Saharan Africa: a spatial and temporal analysis. Lancet Glob Health 2(10):e601–15. doi: 10.1016/S2214-109X(14)70300-6 CrossRefPubMedGoogle Scholar
- Goodman CA, Coleman PG, Mills AJ (1999) Cost-effectiveness of malaria control in sub-Saharan Africa. Lancet 354(9176):378–85CrossRefPubMedGoogle Scholar
- Gore-Langton GR, Alenwi N, Mungai J, Erupe NI, Eves K, Kimwana FN, Soti D, Akhwale W, Hassan FA, Juma E, Allan R (2015) Patient adherence to prescribed artemisinin-based combination therapy in Garissa County, Kenya, after three years of health care in a conflict setting. Malar J 14 (1):125. doi: 10.1186/s12936-015-0645-z CrossRefPubMedPubMedCentralGoogle Scholar
- Gratz NG (1999) Emerging and Resurging Vector-Borne Diseases. Ann Rev Entomol 44(1):51CrossRefGoogle Scholar
- Gubler DJ (1998) Resurgent vector-borne diseases as a global health problem. Emerg Infect Dis 4(3):442–50. doi: 10.3201/eid0403.980326
- Gubler DJ (2010) The global threat of emergent/re-emergent vector-borne diseases. In: Atkinson PW (ed) Vector Biology, Ecology and Control, pp 39–63Google Scholar
- Hollingsworth TD, Pulliam JR, Funk S, Truscott JE, Isham V, Lloyd AL (2014) Seven challenges for modelling indirect transmission: Vector-borne diseases, macroparasites and neglected tropical diseases. Epidemics doi: 10.1016/j.epidem.2014.08.007
- Hopkins DR (2013) Disease Eradication. N Engl J Med 368:54–63CrossRefPubMedGoogle Scholar
- Institute of Medicine (US) Forum on Microbial Threats (2008) Vector-Borne Disease Emergence and ResurgenceGoogle Scholar
- Keeling MJ, Rohani P (2008) Modeling infectious diseases in humans and animals. Princeton University Press, PrincetonGoogle Scholar
- Klepac P, Metcalf CJE, McLean AR, Hampson K (2013) Towards the endgame and beyond: complexities and challenges for the elimination of infectious diseases. Phil Trans R Soc B 368(1623):20120,137. doi: 10.1098/rstb.2012.0137 CrossRefGoogle Scholar
- Kuehn C (2011) A mathematical framework for critical transitions: bifurcations, fast-slow systems and stochastic dynamics. Phys D 240:1020–1035CrossRefGoogle Scholar
- Lawpoolsri S, Klein EY, Singhasivanon P, Yimsamran S, Thanyavanich N, Maneeboonyang W, Hungerford LL, Maguire JH, Smith DL (2009) Optimally timing primaquine treatment to reduce Plasmodium falciparum transmission in low endemicity Thai-Myanmar border populations. Malar J 8(1):159. doi: 10.1186/1475-2875-8-159 CrossRefPubMedPubMedCentralGoogle Scholar
- Mackenzie JS, Gubler DJ, Petersen LR (2004) Emerging flaviviruses: the spread and resurgence of Japanese encephalitis, West Nile and dengue viruses. Nat Med 10(12 Suppl):S98–109. doi: 10.1038/nm1144 CrossRefPubMedGoogle Scholar
- Nisbet RM, Gurney WSC (1982) Modelling fluctuating populations. Wiley, New YorkGoogle Scholar
- Noor AM, Amin AA, Akhwale WS, Snow RW (2007) Increasing coverage and decreasing inequity in insecticide-treated bed net use among rural Kenyan children. PLoS Med 4(8):e255. doi: 10.1371/journal.pmed.0040255 CrossRefPubMedPubMedCentralGoogle Scholar
- Nyarango PM, Gebremeskel T, Mebrahtu G, Mufunda J, Abdulmumini U, Ogbamariam A, Kosia A, Gebremichael A, Gunawardena D, Ghebrat Y, Okbaldet Y (2006) A steep decline of malaria morbidity and mortality trends in Eritrea between 2000 and 2004: the effect of combination of control methods. Malar J 1:33. doi: 10.1186/1475-2875-5-33 CrossRefGoogle Scholar
- O’Regan SM, Drake JM (2013) Theory of early warning signals of disease emergence and leading indicators of elimination. Theor Ecol 6(3):333–357. doi: 10.1007/s12080-013-0185-5 CrossRefGoogle Scholar
- Paaijmans KP, Read AF, Thomas MB (2009) Understanding the link between malaria risk and climate. Proceedings National Academic Science USA pp 13,844–13,849 doi: 10.1073/pnas.0903423106
- Paaijmans KP, Cator LJ, Thomas MB (2013) Temperature-dependent pre-bloodmeal period and temperature-driven asynchrony between parasite development and mosquito biting rate reduce malaria transmission intensity. PloS One 8(1):e55,777 . doi: 10.1371/journal.pone.0055777 CrossRefGoogle Scholar
- Perkins TA, Scott TW, Le Menach A, Smith DL (2013) Heterogeneity, mixing, and the spatial scales of mosquito-borne pathogen transmission. PLoS Comput Biol 9(12):e1003,327. doi: 10.1371/journal.pcbi.1003327 CrossRefGoogle Scholar
- Read AF, Lynch PA, Thomas MB (2009) How to make evolution-proof insecticides for malaria control. PLoS Biol 7(4):e1000, 058. doi: 10.1371/journal.pbio.1000058 CrossRefGoogle Scholar
- Reiner RC, Perkins TA, Barker CM, Niu T, Chaves LF, Ellis AM, George DB, Menach AL, Pulliam JRC, Bisanzio D, Buckee C, Chiyaka C, Cummings DAT, Garcia AJ, Gatton ML, Gething PW, Hartley DM, Johnston G, Klein EY, Michael E, Lindsay SW, Lloyd AL, Pigott DM, Reisen WK, Ruktanonchai N, Singh BK, Tatem AJ, Kitron U, Hay SI, Scott TW, Smith DL (2013) A systematic review of mathematical models of mosquito-borne pathogen transmission: 1970–2010. J R Soc Interface 10(81):20120,921. doi: 10.1098/rsif.2012.0921 CrossRefGoogle Scholar
- Rogers DJ, Randolph SE, Snow RW, Hay SI (2002) Satellite imagery in the study and forecast of malaria. Nature 415(6872):710–5. doi: 10.1038/415710a CrossRefPubMedPubMedCentralGoogle Scholar
- Scheffer M, Bascompte J, Brock WA, Brovkin V, Carpenter SR, Dakos V, Held H, Van Nes EH, Rietkerk M, Sugihara G (2009) Early-warning signals for critical transitions. Nature 461(7260):53–59. doi: 10.1038/nature08227 CrossRefPubMedGoogle Scholar
- Smith DL, McKenzie FE (2004) Statics and dynamics of malaria infection in Anopheles mosquitoes. Malar J 3(1):13. doi: 10.1186/1475-2875-3-13 CrossRefPubMedPubMedCentralGoogle Scholar
- Smith DL, Battle KE, Hay SI, Barker CM, Scott TW, McKenzie FE (2012) Ross, Macdonald, and a theory for the dynamics and control of mosquito-transmitted pathogens. PLoS Pathog 8(4):e1002,588. doi: 10.1371/journal.ppat.1002588 CrossRefGoogle Scholar
- Strogatz SH (1994) Nonlinear dynamics and chaos with applications to physics, biology, chemistry and engineering. Addison-Wesley, Reading, MAGoogle Scholar
- Tusting LS, Thwing J, Sinclair D, Fillinger U, Gimnig J, Bonner KE, Bottomley C, Lindsay SW (2013) Mosquito larval source management for controlling malaria. Cochrane Database Syst Rev 8:CD008,923. doi: 10.1002/14651858.CD008923.pub2 Google Scholar
- World Health Organization (2013) World health report: research for universal health coverage. Technical ReportGoogle Scholar
- World Health Organization (2014) Vector-borne diseases. fact sheet No. 387Google Scholar

## Copyright information

**Open Access**This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.