Outline

Population heterogeneity, structure, Introduction: Heterogeneity and population structure and mixing Models for population structure Structure example: rabies in space Models for heterogeneity Jamie Lloyd-Smith • individual heterogeneity and superspreaders Center for Infectious Disease Dynamics • group-level heterogeneity Pennsylvania State University Population structure and mixing mechanisms Pair formation and STD

Heterogeneity and structure – what’s the difference? Modelling heterogeneity

Group-level heterogeneity and multi-group models Tough to define, but roughly… Heterogeneity describes differences among individuals or groups in Break population into sub-groups, each a population. of which is homogeneous. (often assume that all groups mix randomly) Population structure describes deviations from random mixing in a population, due to spatial or social factors. However, epidemiological traits of each individual are due to a complex blend of host, pathogen, and environmental factors, The language gets confusing: and often can’t be neatly divided into groups - models that include heterogeneity in host age are called (or predicted in advance). “age-structured”. Individual-level heterogeneity - models that include spatial structure where model parameters differ through space are called “spatially heterogeneous”. Allow continuous variation among individuals.

Models for population structure Models for population structure

Random mixing Multi-group Spatial mixing Random mixing or mean-field

• Every individual in population has equal probability of contacting any Network Individual-based model other individual. • Mathematically simple – “mass action” formulations borrowed from chemistry – but often biologically unrealistic. • Sometimes basic βSI form is modified to power law βSaIb as a phenomenological representation of non-random mixing.

1 Models for population structure Models for population structure

Multi-group Spatial mixing or metapopulation

• Divides population into multiple discrete groupings, based on spatial or social differences. • Used when individuals are distributed (roughly) evenly in space. • To model transmission, need contact matrix or Who Acquires • Can model many ways: From Whom (WAIFW) matrix: • continuous space models (e.g. reaction-diffusion or contact kernel)

βij = transmission rate from infectious individual in group i to • individuals as points on a lattice susceptible in group j • patch models (metapopulation with spatial mixing) • Or use only within-group transmission (so βij =0 when i≠j), and explicitly model movement among groups. • Used to study travelling waves, spatial control programs, influences of restricted mixing on disease invasion and persistence

Models for population structure Models for population structure

Social network Individual-based model (IBM) or microsimulation model

• Precise representation of contact structure within a population • The most flexible framework. • “Nodes” are individuals and “edges” are contacts • Every individual in the model carries its own attributes (age, sex, • Important decisions: Binary vs weighted? Undirected vs directed? location, contact behaviour, etc etc) Static vs dynamic? • Can represent arbitrarily complex systems ( = realistic?) and ask detailed • Basic network statistics include degree distribution (number of edges per questions, but difficult to estimate parameters and to analyze node) and clustering coefficient (How many of my friends are model output; also difficult for others to replicate the model. friends with each other?) • STDSIM is a famous example, used to study transmission and control of • Powerful tools of discrete mathematics can be applied. sexually transmitted diseases including HIV in East Africa.

Rabies in space Major Terrestrial Reservoirs of Rabies in the United States Rabies is an acute viral disease of mammals, that causes cerebral dysfunction, anxiety, confusion, agitation, progressing to delirium, abnormal behavior, hallucinations, and insomnia. Raccoon

Transmitted by infected saliva, most commonly through biting.

Latent period = 3 – 12 weeks (in raccoons) = 1 week (ends in death)

Pre-exposure offers effective protection. • Until mid 1970s, raccoon rabies was restricted to FL and GA. Post-exposure vaccination possible during latent period. • Then rabid raccoons were translocated to the WV-VA border, and a major began in the NE states.

2 Models of the spatial spread of rabies Spatial invasion of Rabies across the Northeastern U.S.

23 years (1977-1999)

Rabies in wildlife in US

Simple patch model (+ small long-range transmission term) was able to fit data well.

Smith et al (2002) PNAS 99: 3668-3672 Russell et al (2004) Proc Roy Soc B 271: 21-25.

The vital few and insignificant many – Individual heterogeneity the 20/80 rule: )

Macroparasitic diseases: 0 Heterogeneity in the worm burdens in individuals are overdispersed, and well- population 20-80 described by a negative binomial distribution.

Every host is equal STDs and -borne diseases: 20-20 Woolhouse et al (PNAS, 1998) analyzed contact rate data and

proposed a general 20/80 rule: Transmission potential 20% of hosts are responsible for 80% of transmission 20% of hosts account for 80% of pathogen transmission % Transmission potential (R potential Transmission % But how to approach other directly-transmitted diseases, 0 20406080100 for which contacts are hard to define? 0 20406080100 PercentagePercentage of of hosthost population population

Slide borrowed from Sarah Perkins

A model for individual heterogeneity Individual reproductive number, ν Expected number of cases caused by a particular infectious Basic reproductive number, R0 individual in a susceptible population. Expected number of cases caused by a typical infectious individual in a susceptible population. Z = actual number of cases caused by a particular infectious individual. Individual reproductive number, ν Expected number of cases caused by a particular infectious Stochasticity in transmission Æ Z ~ Poisson(ν ) individual in a susceptible population.

The offspring distribution defines Pr (Z=j ) for all j. ν varies continuously among individuals, with population mean R0.

R0

ν

ν Z = 0 Z = 1 Z = 2 Z = 3 …

3 Branching process: a stochastic model for disease invasion Contact tracing for SARS Observed offspring distribution into a large population.

Number of secondary cases, Z

For any offspring distribution, it tells you: • Pr(extinction) Estimated distribution of • Expected time of extinction and number of cases individual reproductive number, ν • Growth rate of major outbreak ν

Singapore SARS outbreak, 2003

Monkeypox SARS, Beijing Smallpox

What about other emerging diseases?

Pneumonic plague Hantavirus* Variola minor

Dynamic effects: stochastic extinction of disease

k = 0.1 rae heterogeneity greater

k = 1

Probability of extinction of Probability k→∞

Basic reproductive number, R0

Read more about individual heterogeneity and superspreading in Lloyd-Smith et al (2005) Nature 438: 355-359.

4 Transmission: mechanisms matter The R-matrix or next-generation matrix

Generalized R0 for a multi-group population Transmission dynamics are the core of epidemic models Rij = E(# cases caused in group j|infected in group i)

Usual approach considers group membership as static.

Æ Take time to think about the mechanisms underlying Di = expected infectious period, spent entirely in group i transmission, and to find the best tradeoff between model βij = transmission rate from group i to group j simplicity and biological realism. The expected number of cases in group j caused by an individual infected in group i is then: R = D β e.g. Between-group transmission in metapopulations ij i ij But what if the host moves and transmission is strictly local? Frequency-dependent transmission vs pair-formation models Dij = expected time spent in group j by individual infected in group i, while still infectious

βj = transmission rate within group j

Now Rij = Dijβj

Transmission in a metapopulation Analytic approach to R

If movement rules are Markovian, so

pij = Pr(move from group i to group j):

mj = Pr(recover or die while in group j) x The process can be described by an absorbing Markov chain, with overall transition matrix: x x x x

⎡Pn×n m⎤ ⎢ ⎥ x x ⎣ 0 1 ⎦

The expected residence times Dij are then given by the fundamental matrix: Simulate: D = (I − P)−1 • range of multi-group population structures • acute and chronic diseases R-matrix: Rij = Dijβj

Acute disease Chronic disease 1 R0= 20 R0= 10

0.8 R0= 5

0.6

0.4 R0= 2 0.2

1 group of 1000 25 groups of 40 100 groups of 10 Total proportion infected 0 −3 −2 −1 0 1 10 10 10 10 10 Acute and chronic diseases movement rate/recovery rate

with same R0 behave very differently when R0 does not predict invasion for this system! invading a metapopulation.

5 Units of analysis: R0 versus R* Units of analysis: R0 versus R*

• R* = the expected number of groups infected by the first infected group (a group-level R0). (Ball et al. 1997 Annals of Appl. Prob.)

• Analytical expressions for R0 or R* are hard to find for systems with mechanistic movement, finite group sizes, and finite numbers of groups. R R0* • Use “empirical” values: mean values from simulations where we track who infects whom. ˆ ˆ R0 R*

Predictors of disease invasion Approaching R*

1.0 1.0 μ/γ = expected number of movements between groups by an individual during its infectious period 0.8 0.8 pI = expected proportion of initial group infected following 0.6 0.6 β β the initial outbreak. If R0 is large, then pI ~1. 0.1 0.1 0.4 0.4 0.5 0.5 pInμ/γ = expected number of infectious individuals that will 1.0 1.0 0.2 4.3 0.2 4.3 disperse from the initial group

9.0 mean proportion infected 9.0 mean proportion infected mean proportion 15 0.0 15 0.0 R0 ≈ β /γ 024681012024681012 Rˆ Rˆ * 0 So: for a , we require R0 >1 and pI nμ/γ >1.

crudely, R* will increase with pI n μ/γ and with β /γ. R* is a much better predictor of disease invasion in a structured population than R 0 A proper mathematical treatment of this problem is needed! Cross et al. 2005 Eco. Letters

Summary on mechanisms in multi-group models A mechanistic model for STD transmission

• Need to consider timescales of relevant processes: rate = f (S,I) mixing, recovery, transmission, (susc. replenishment) S I • In some limits, simpler models do OK. STDs are often modelled using frequency-dependent • In general, and especially when different processes incidence: occur on similar timescales, ⎛ S ⎞ f (S, I) = cFD pFD ⎜ ⎟ I mechanistic models are needed to capture dynamics. ⎝ N ⎠

• Appropriate “units” aid prediction. cFD = rate of acquiring new partners

pFD = prob. of transmission in S-I partnership S/N = prob. that partner is susceptible

Read more about disease invasion in structured populations in I = density of infectives Cross et al (2005) Ecol Lett 8: 587-595 Read more about pair-formation models for STDs in Cross et al (2007) JRS Interface 4:315-324.. Lloyd-Smith et al (2004) Proc Roy Soc B 271: 625-634

6 Pair dynamics Pair dynamics

km km km km X Singles SS XS SI IS XI II Singles

kl

l lll P Pairs PSS PSI PII Pairs

X = single individual Xy = single individual of type y (where y = S or I) P = pair Pyz = pair of types y and z (where y,z = S or I) k = pairing rate (per capita) k = pairing rate (per capita) l = pair dissolution rate l = pair dissolution rate

myz = “mixing matrix”

Pair dynamics Pair-formation epidemic

λ σ k m k m k m k m S SS XS S SI I IS XI I II Singles μ μ μ μ

XS XI P l l P l l P Pairs SS SS SI SI SI II II μ μ

PSS PSI PII

βpairPSI Xy = single individual of type y (where y = S or I) μ σ μ σ μ Pyz = pair of types y and z (where y,z = S or I) Consider populations where pairing dynamics are much ky = pairing rate (per capita) faster than disease dynamics. lyz = pair dissolution rate myz = “mixing matrix” Timescale approximation: pairing dynamics are at

Transmission occurs only in S-I pairs (PSI), at rate βpair quasi-steady-state relative to disease dynamics (c.f. Heesterbeek & Metz (1993) J. Math. Biol. 31: 529-539.)

Timescale approximation for pairing ⎧dX S ⎪ = −kS X S + 2lSSPSS + lSI PSI k m k m k m k m dt S SS X S SI I IS X I II Singles ⎪ S I dX Fast ⎪ I = −k X + 2l P + l P ⎪ dt I I II II SI SI pairing Fast pairing ⎪ timescale l l l l dP PSS SS SI PSI SI II PII Pairs ⎪ SS 1 ⎨ = 2 kSmSS X S − lSSPSS dynamics ⎪ dt Timescale approximation ⎪dP SI = 1 k m X + 1 k m X − l P ⎪ dt 2 S SI S 2 I IS I SI SI Slow l βpair P*SI ⎪ S I Entire dP epidemic population ⎪ II = 1 k m X − l P sI ⎪ 2 I II I II II timescale mS mI ⎩ dt

S= XS + 2P SS + PSI I = XI + 2PII + PSI ⎧dS * ⎪ = λ − βpair PSI +σ I − μS Slow ⎪ dt disease ⎨ Challenge: find P* in terms of S, I, and pairing ⎪dI * SI = βpair PSI − ()σ + μ I * ⎩⎪ dt parameters. Then incidence rate = βpair P SI dynamics

7 * Finding P SI from fast equations Simplest case: uniform behaviour

Substitute: Disease status has no effect on pairing behaviour. k = pairing rate for all individuals S = X S + 2PSS + PSI ⎫ Total all susceptibles and ⎬ l = break-up rate for all partnerships I = X I + 2PII + PSI ⎭ infectives, single and paired kS X S ⎫ mSS = mIS = ⎪ kS X S + kI X I ⎪ * ⎛ k ⎞ SI ⎬ Assume random Incidence rate =β P = β ⎜ ⎟ k X pair SI pair m = m = I I ⎪ mixing in pair ⎝ k + l ⎠ N SI II ⎪ kS X S + kI X I ⎭ formation ⎛ S ⎞ Recall the FD incidence: cFD pFD ⎜ ⎟ I ⎝ N ⎠ * Set dPyz/dt’s = 0, solve for P SI

Pair-based transmission and frequency dependence Pair-based transmission and frequency dependence

c = rate of acquiring partners FD Frequency dependence can represent pair-based transmission 1 kl but timescale approximation is required. = = 1 1 l + k k + l Conversely: p = probability of transmission in S-I partnership FD We know STD dynamics are driven by pair-based transmission. FD models implicitly make timescale approximation. = 1−exp(−βpair ×1/l) ≈ βpair l (since βpair << l)

SI ⎛ k ⎞ SI cFD pFD ≈ βpair ⎜ ⎟ Use mechanistic derivation of FD to assess this assumption. N ⎝ k + l ⎠ N

Application to STD models When does FD adequately represent pair-based transmission?

Transient, highly-transmissible Chronic, less-transmissible Compare simulations: STDs STDs frequency-dependent incidence vs. • High chance of infection per • Low chance of infection per exposure exposure full simulation of pair dynamics and disease • Most individuals recover •No recovery! within a month • e.g. HIV, HSV-2 • e.g. gonorrhoea, chlamydia for different timescales of: • disease – bacterial and viral STDs Many bacterial STDs Many viral STDs • pairing dynamics – define average pair lifetime, D 1 1 D = = l k

8 Transient, Chronic, Modelling disease-induced behaviour changes highly-transmissible STD less-transmissible STD

k m k m k m k m S SS XS S SI I IS XI I II Singles

l l l l PSS SS SI PSI SI II PII Pairs

Four cases: 1. No effect on behaviour 2. Disease alters pair-formation rate, k ≠ k Frequency dependence is a good depiction of pair-based S I 3. Disease alters break-up rate, l ≠ l ≠ l transmission only when mixing occurs fast compared to SS SI II 4. Disease alters both k and l (y, z = S or I) disease timescales. y yz Timescale approximation breaks down badly for D ~ 3 days

Modelling disease-induced behaviour changes Case Rates, k φk(s,i)

kS=kI=k k For all four cases, the incidence rate takes a generalized 1 lSS=lSI=lII=l k + l frequency-dependent form: k ∫ k π π 2 S I S I l =l =l =l SI SS SI II π Ss +π Ii β pairφκ (s,i) N kS=kI=k π 3 where φ (s,i) is a function of s=S/N, i=I/N and lSS ∫ lSI ∫ lII 1 1 2 κ 2 + 2 1− 4aπ si k ∫ k the pairing parameters. S I π Sπ I 4 2 2 lSS ∫ lSI ∫ lII 1 ⎜⎛π s +π i + ()()π s +π i − 4a π π si ⎟⎞ 2 ⎝ S I S I S I ⎠

l ⎛ l ⎞ l ⎛ l ⎞ ⎛ l 2 ⎞ SI ⎜ SI ⎟ SI ⎜ SI ⎟ ⎜ SI ⎟ where s=S/N, i=I/N, πy=ky/(ky + lSI) and a = ⎜1− ⎟ + ⎜1− ⎟ + ⎜1− ⎟ kI ⎝ lSS ⎠ kS ⎝ lII ⎠ ⎝ lSSlII ⎠ If kS=kI, then πS= πI ªπ.

Calculation of R0 Calculation of stability threshold

R0 = lim (transmission rate per I individual × duration of infectiousness) S → N Consider stability threshold of the no-infection equilibrium, when ⎛ S 1 ⎞ population is wholly susceptible: = lim ⎜ β pair φ κ (s, i) × ⎟ S → N ⎝ N σ + μ ⎠

R0 > 1 ↔ no-infection equilibrium is unstable to perturbations in I β pair = lim ()φ κ (s, i) σ + μ S → N R > 1 ↔ ⎡∂f ⎤ dI 0 I > 0, where = f (S, I) ⎢ ∂I ⎥ dt I β pair ⎛ k I ⎞ ⎣ ⎦ S→N = ⎜ ⎟ in all four cases. σ + μ ⎝ k I + lSI ⎠ Yields the same result as R0 calculation in all four cases – though note that just because a quantity is an epidemic threshold • No dependence on k S parameter does not mean that it equals R0!! • No dependence on l or l SS II c p • Identical to standard FD result, R = FD FD e.g. (R )k for any k>0 also has an epidemic threshold at 1. 0 σ + μ 0

if cFD = contact rate of infected individuals

9