Vol 438|17 November 2005|doi:10.

1038/nature04153

LETTERS
Superspreading and the effect of individual
variation on disease emergence
J. O. Lloyd-Smith1,2, S. J. Schreiber3, P. E. Kopp4 & W. M. Getz1

Population-level analyses often use average quantities to describe data to investigate specific questions in light of observed variation8,16,
heterogeneous systems, particularly when variation does not arise but a broad understanding of the role of individual variation in
from identifiable groups1,2. A prominent example, central to our outbreak dynamics is lacking.
current understanding of epidemic spread, is the basic repro- As a theoretical basis, we introduce the ‘individual reproductive
ductive number, R 0, which is defined as the mean number of number’, n, as a random variable representing the expected number
infections caused by an infected individual in a susceptible of secondary cases caused by a particular infected individual.
population3,4. Population estimates of R 0 can obscure consider- Values for n are drawn from a continuous probability distribution
able individual variation in infectiousness, as highlighted during with population mean R 0 that encodes all variation in infectious
the global emergence of severe acute respiratory syndrome (SARS) histories of individuals, including properties of the host and patho-
by numerous ‘superspreading events’ in which certain individuals gen and environmental circumstances. In this framework, super-
infected unusually large numbers of secondary cases5–10. For spreading events (SSEs) are not exceptional events9, but important
diseases transmitted by non-sexual direct contacts, such as SARS realizations from the right-hand tail of a distribution of n (refs 7, 15).
or smallpox, individual variation is difficult to measure empiri- Stochastic effects in transmission are modelled using a Poisson
cally, and thus its importance for outbreak dynamics has process4, so that the number of secondary infections caused by
been unclear2,10,11. Here we present an integrated theoretical and each case, Z, is described by an ‘offspring distribution’ Pr(Z ¼ k)
statistical analysis of the influence of individual variation in where Z , Poisson(n).
infectiousness on disease emergence. Using contact tracing data By considering three possible distributions of n, we generate three
from eight directly transmitted diseases, we show that the distri- candidate models for the offspring distribution: (1) in generation-
bution of individual infectiousness around R 0 is often highly based models neglecting individual variation, n ¼ R 0 for all cases,
skewed. Model predictions accounting for this variation differ yielding Z , Poisson(R 0); (2) in differential-equation models with
sharply from average-based approaches, with disease extinction homogeneous transmission and constant recovery rates, n is expo-
more likely and outbreaks rarer but more explosive. Using these nentially distributed, yielding Z , geometric(R 0); (3) in a more
models, we explore implications for outbreak control, showing general formulation, we let n be gamma-distributed with mean R 0
that individual-specific control measures outperform population- and dispersion parameter k, yielding Z , negative binomial(R 0,k)
wide measures. Moreover, the dramatic improvements achieved (ref. 23). The negative binomial model includes the conventional
through targeted control policies emphasize the need to identify Poisson (k ! 1) and geometric (k ¼ 1) models as special cases. It
predictive correlates of higher infectiousness. Our findings indi- has variance R 0(1 þ R 0/k), so smaller values of k indicate greater
cate that superspreading is a normal feature of disease spread, and heterogeneity.
to frame ongoing discussion we propose a rigorous definition for We gathered empirical offspring distributions from detailed con-
superspreading events and a method to predict their frequency. tact tracing or surveillance data sets, and challenged the candidate
For sexually transmitted and vector-borne diseases, host contact models using model selection techniques24 (see Supplementary
rates have long served as surrogates for individual infectious- Notes). For SARS outbreaks in Singapore and Beijing, the negative
ness3,12–14, leading to the assertion of a general ‘20/80 rule’ (whereby binomial model is unequivocally favoured (Fig. 1a and Supplemen-
20% of cases cause 80% of transmission13) and to the influential tary Table 1). Conventional models assuming homogeneity cannot
concept of high-risk ‘core groups’3,12,13. For directly transmitted reproduce the observed transmission patterns. For the Singapore
infections, however, the overall infectiousness of each case—that is, outbreak, the maximum-likelihood estimate k^ is 0.16 (90% confi-
the number of other individuals infected during the infectious dence interval 0.11–0.64), indicating an underlying distribution of n
lifetime of a single individual—arises from a complex mixture of that is highly overdispersed (Fig. 1a, inset). According to this analysis,
host, pathogen and environmental factors (see Supplementary the great majority of SARS cases in Singapore were barely infectious
Notes). Consequently, the degree of infectiousness is distributed (73% had n , 1) but a small proportion were highly infectious (6%
continuously in any population4,7,11,15,16 and, crucially, distinct risk had n . 8). This is consistent with field reports from SARS-afflicted
groups often cannot be defined a priori2,11. This impedes the regions5,6 but contrasts with published SARS models9,10,25,26.
conventional approach to adding heterogeneity to epidemic models, Comparing results for eight directly transmitted infections reveals
in which populations are divided into homogeneous subgroups2–4,17. the differing degree of individual variation among diseases and
Research on continuous individual variation in infectiousness for outbreak settings (Fig. 1b, c and Supplementary Tables 1, 2). The
directly transmitted infections has been largely restricted to within- Poisson offspring distribution is almost always strongly rejected. The
household transmission18,19, or to variation in infectious period20,21 geometric model has considerable support for several data sets,
or social network22. Some recent studies have used contact tracing which indicates significant individual variation in transmission
1
Department of Environmental Science, Policy and Management, 140 Mulford Hall, University of California, Berkeley, California 94720-3114, USA. 2Biophysics Graduate Group,
University of California, Berkeley, California 94720-3200, USA. 3Department of Mathematics, The College of William and Mary, Williamsburg, Virginia 23187-8975, USA.
4
Centre for Mathematics, University of Hull, Hull HU6 7RX, UK.

355
© 2005 Nature Publishing Group
LETTERS NATURE|Vol 438|17 November 2005

rates because real infectious periods are less dispersed than the arbitrarily defined as Z $ 8 (ref. 6), Z $ 10 (ref. 5), Z . 10 (ref. 26)
exponential distribution20,21. The negative binomial model is selected or ‘many more than the average number’9, and different thresholds
decisively for several data sets, and enables comparative study of are surely needed for measles (R 0 , 11–18; ref. 3) or monkeypox
diseases through the dispersion parameter. Like SARS, measles (R 0 , 1).
in highly vaccinated populations shows high variation in two We propose this general protocol for defining a superspreading
surveillance data sets, with narrow confidence intervals excluding event: (1) estimate the effective reproductive number, R, for the
the conventional models (note that heterogeneous vaccination cover- disease and population in question; (2) construct a Poisson distri-
age is an important environmental factor contributing to bution with mean R, representing the expected range of Z due to
this pattern). Monkeypox and smallpox viruses show intermediate stochasticity without individual variation; (3) define an SSE as any
variation, consistent across multiple data sets, and pneumonic infected individual who infects more than Z (n) others, where Z (n) is
plague transmission is slightly less variable. Data limitations prevent the nth percentile of the Poisson(R) distribution. A 99th-percentile
definitive conclusions for other diseases. Comparing our findings to SSE is then any case causing more infections than would occur in
the 20/80 rule proposed for sexually transmitted and vector-borne 99% of infectious histories in a homogeneous population (Fig. 1d).
diseases13, no general rule emerges but the core principle of hetero- This approach complements a priori identification of potential
geneous transmission is certainly supported (Fig. 1c). superspreaders when that is feasible, as for sexually transmitted
Numerous reports of superspreading events provide further evi- diseases (where promiscuity drives risk)3,12. In addition, the defi-
dence for variation in n. We reviewed 37 published accounts of SSEs nition enables prediction of the frequency of SSEs once R 0 and k have
for 11 directly transmitted infections (Fig. 1d; see Supplementary been estimated (Supplementary Fig. 1)—an outstanding challenge in
Notes). Unrecognized or misdiagnosed illness is the most common emerging disease epidemiology8,9.
cause of these SSEs, followed by alternative modes of spread To assess the effect of individual variation on disease outbreaks, we
(especially airborne), high contact rates, and co-infections that aid analyse a branching process model with negative binomial offspring
transmission. High pathogen load or shedding rates are occasionally distribution, corresponding to gamma-distributed n (Fig. 2a; see
implicated, but are rarely measured. A consistent and general Supplementary Notes). Of primary interest is the probability of
definition of SSEs is currently lacking—for SARS, an SSE has been stochastic extinction, q, after the introduction of a single infected

Figure 1 | Evidence for variation in individual reproductive number n. in n. c, Proportion of transmission expected from the most infectious 20% of
a, Transmission data from the SARS outbreak in Singapore in 2003 (ref. 5). cases, for 10 outbreak or surveillance data sets (triangles). Dashed lines show
Bars show observed frequency of Z, the number of individuals infected by proportions expected under the 20/80 rule (top) and homogeneity
each case. Lines show maximum-likelihood fits for Z , Poisson (squares), (bottom). Superscript ‘v’ indicates a partially vaccinated population.
Z , geometric (triangles), and Z , negative binomial (circles). Inset, d, Reported superspreading events (SSEs; diamonds) relative to estimated
probability density function (solid) and cumulative distribution function reproductive number R (squares) for twelve directly transmitted infections.
(dashed) for gamma-distributed n (corresponding to Z , negative Lines show 5–95 percentile range of Z , Poisson(R), and crosses show the
binomial) estimated from Singapore SARS data. b, Expected proportion of 99th-percentile proposed as threshold for SSEs. Stars represent SSEs caused
all transmission due to a given proportion of infectious cases, where cases by more than one source case. ‘Other’ diseases are: 1, Streptococcus group A;
are ranked by infectiousness. For a homogeneous population (all n ¼ R 0), 2, Lassa fever; 3, Mycoplasma pneumonia; 4, pneumonic plague; 5,
this relation is linear. For five directly transmitted infections (based on tuberculosis. R is not shown for ‘other’ diseases, and is off-scale for
k^ values in Supplementary Table 1), the line is concave owing to variation monkeypox. See Supplementary Notes for details.
356
© 2005 Nature Publishing Group
 NATURE|Vol 438|17 November 2005 LETTERS
at low k=0.01, just one superspreader, hence
less % of outbreaks reaching 100 cases as the
the outbreaks can be easily controlled

Figure 2 | Outbreak dynamics with different degrees of individual variation top to bottom (values and colours as in a). c, Growth of simulated outbreaks
in infectiousness. a, The individual reproductive number n is drawn from a with R 0 ¼ 1.5 and one initial case, conditional on non-extinction. Boxes
gamma distribution with mean R 0 and dispersion parameter k. Probability show median and interquartile range (IQR) of the first disease generation
density functions are shown for six gamma distributions with R 0 ¼ 1.5 with 100 cases; whiskers show most extreme values within 1.5 £ IQR of the
(‘k ¼ Inf ’ indicates k ! 1). b, Probability of stochastic extinction of an boxes, and crosses show outliers. Percentages show the proportion of 10,000
outbreak, q, versus population-average reproductive number, R 0, following simulated outbreaks that reached the 100-case threshold (roughly 1 2 q).
introduction of a single infected individual. The value of k increases from

individual (Fig. 2b). For R 0 , 1, all invasions die out, as in standard rates strongly depend on variation in n (Fig. 2c and Supplementary
models. For R 0 . 1, increased variation strongly favours extinction8. Fig. 2e, f). Diseases with high individual variation show infrequent
For example, if R 0 ¼ 3 then q ¼ 0.06 under the assumption of but explosive epidemics after introduction of a single case. This
homogeneous n (k ! 1), or q ¼ 0.33 if k ¼ 1, but if k ¼ 0.16 (as pattern recalls SARS in 2003, for which many settings experienced
estimated for SARS) then q ¼ 0.76. Extinction risk rises owing to a no epidemic despite unprotected exposure to SARS cases27,28,
higher proportion of non-transmitting cases when n is overdispersed whereas a few cities suffered explosive outbreaks8–10,15,26. Our results,
(Figs 1a, 2a and Supplementary Fig. 2a). This effect thwarts invasion using k^ ¼ 0:16 for SARS, explain this simply by the presence or
by diseases that are very potent spreaders on average: for arbitrarily absence of high-n individuals in the early generations of each out-
high R 0, q ! 1 as k ! 0 (Supplementary Fig. 2b). The expected break6. In contrast, conventional models (with k ¼ 1 or k ! 1)
number of cases before extinction is hardly affected by k (Sup- cannot simultaneously generate frequent failed invasions and rapid
plementary Fig. 2c), because low-k outbreaks that fail probably growth rates without additional, subjective model structure.
y= difference
lacked SSEs and between
thus resemble homogeneous outbreaks with lower Disease control interventions could increase or decrease individual
R 0extiction , higher when
. Accordingly, value individual variation is large, extinction occurs variation in infectiousness. Infected individuals might reduce their
increaes
rapidly icnrease
or not at allin(Supplementary Fig. 2d). number of non-essential contacts, or governments might impose
extinction probability absed
For outbreaks avoiding stochastic extinction, epidemic growth quarantine or isolation on particular individuals. Here we explore
on individual specific control
as K increases, the more
homogenous the population
is hence higher extinctin
under individual specific
control
c= 0.4, ifnecitous of disease after control/durign control, the dispersion K
is reduced by 0.4 decreases and reproducitve number also
decreases, maintained less superspreader and
as control effeir increase, the successful control . while for SARS in beijign,
probability of extinction in the populaiton wide contol does not affect R0
individual but reduced dispersion ,individuals specieis
conrtol will be more effecitve thant populatino
probability of containment wide control
based on the types o f
control , solid= population
wide policy, dot= individual
specific , as K is low , when 3 types of control, 0.2= 20% population
have taregted individuals controlled , , targeting on the individual speicies
specific control, even with control show, as
icnreases k , the
contianment probaiblity is
hgiher 3
Figure , | Implications for control measures. a, Increase in extinction never reaching the 100-case threshold) for four diseases with R 0 ¼ 3 and
probability (q ind 2 q pop) under individual-specific control compared to k ¼ 0.1 (blue), k ¼ 0.5 (green), k ¼ 1 (black) or k ! 1 (purple). Control
population-wide control, for diseases with R 0 ¼ 3 and different degrees of policies are population-wide (solid lines), random individual-specific
individual variation, k, subject to control effort c. With population-wide (dotted lines), or targeted individual-specific (dashed lines, where half of all
control, the infectiousness of all individuals is reduced by a factor c. With control effort is focused on the most infectious 20% of cases). For k ! 1, all
individual-specific control, a proportion c of infectious individuals (selected individuals are identical, so targeting has no effect and dotted and dashed
at random) have their infectiousness reduced to zero. The outbreak is lines overlay one another. d, The factor by which targeting increases the
assumed to begin with one case, with control present from the outset. effect of control on preventing a major outbreak, relative to random
b, Estimates of R^ and k^ from outbreak data sets before and after control individual-specific control (see Supplementary Notes), when 20%, 40% or
measures were initiated (joined by solid lines; Supplementary Table 2), and 60% of the total population is controlled. Results in c and d are the mean of
post-control values of k c estimated from theoretical models of control as 10,000 simulations, with control beginning in the second generation of
described in the Supplementary Notes. c, Effect of random versus targeted cases.
control measures. The probability of outbreak containment (defined as
357
© 2005 Nature Publishing Group
LETTERS NATURE|Vol 438|17 November 2005

several idealized cases theoretically, for an outbreak with offspring epidemiology, and open challenges for further work. Explosive
distribution Z , negative binomial(R 0,k) before control (see Sup- epidemics demand rapid action by authorities and can strain health
plementary Notes). Consider the effect of control effort c, where infrastructures. High extinction probabilities indicate that disease
c ¼ 0 reflects no control and c ¼ 1 reflects complete blockage of introductions or host species jumps may be more frequent than
transmission. Under population-wide control, the infectiousness of currently suspected. Cluster-size surveillance for pathogen adap-
every individual in the population is reduced by a factor c (that is, tation29 or dwindling population immunity30 should be tuned to
npop
c ¼ ð1 2 cÞn for all individuals). Under random, individual- observed levels of variation. Realization of targeted control measures
specific control, a proportion c of infected individuals (chosen at requires a better understanding of factors determining individual
random) is traced and isolated completely such that they cause zero infectiousness. This work must be integrated with established theory
infections (that is nind
c ¼ 0 for a proportion c of infected individuals, of sexually transmitted diseases and social networks, where high-risk
and nind
c ¼ n for the rest). Individual-specific control raises the degree groups exert nonlinear influence on R 0 because contact rates
of heterogeneity in the outbreak as measured by the variance-to- affect infectiousness and susceptibility equally3,4,12,13,22. All diseases
mean ratio of Z, whereas population-wide control reduces hetero- probably show intermediate degrees of covariation between infec-
geneity. Both approaches yield effective reproductive number tiousness and susceptibility, a topic demanding empirical and theo-
R ¼ (1 2 c)R 0, so the threshold control effort for guaranteed disease retical study17. The central role of R 0 in epidemic analysis is
extinction is c $ 1 2 1/R 0 as in conventional models. For inter- unassailable, but our findings show that emerging disease outbreaks
mediate values of c, however, the individual-specific approach always cannot be fully understood if individual variation in infectiousness is
works better (Fig. 3a and Supplementary Fig. 3a, b), consistent with neglected. Examination of other population processes dependent on
our finding that higher variation favours disease extinction (Fig. 2b). small numbers of individuals may yield similar insights.
Branching process theory confirms that q ind . q pop whenever
c [ (0,1 2 1/R 0) (see Supplementary Notes). METHODS
To assess the realism of these idealized control scenarios, we Analysis of disease data. For data sets including the full distribution of Z, we
analysed contact tracing data from four outbreaks before and after estimated R^ 0 and k^ using maximum-likelihood methods. The candidate models
were compared using Akaike’s information criterion (AICc) modified for small
imposition of control measures. Control always lowered the esti-
sample size. Confidence intervals for k^ were estimated by bias-corrected non-
mated dispersion parameter (that is k^ c , k) ^ as predicted by the
parametric bootstrapping and corroborated by four other methods. For data sets
individual-specific model (Fig. 3b), although small sample sizes often including only estimates of R^ 0 and the proportion of cases not transmitting (p^ 0 ),
led to overlapping confidence intervals (Supplementary Table 2). we estimated k^ by solving p^ 0 ¼ ð1 þ R^ 0 =kÞ2k numerically, and evaluated the
This increased skew in transmission arose chiefly from undiagnosed candidate models using confidence intervals calculated by two methods.
or misdiagnosed individuals, who continued to infect others (and Expected proportions of transmission due to particular groups of infectious
even cause SSEs), whereas controlled individuals infected very few. individuals (Fig. 1b, c) were calculated using the gamma distribution of n with
estimated values of R^ 0 and k. ^ See Supplementary Notes for details, and for
To further examine our control theories, we calculated kpop c and kind
c
^ ind
for each data set; kc was always closer to kc , although twice it fell descriptions of data sets.
Branching process analysis. Analysis of branching process models centres on
between the two predictions, indicating a possible combination of
the probability
P generating function (pgf) of the offspring distribution,
control mechanisms (Fig. 3b). Real-world control thus seems to gðsÞ ¼ 1 j
j¼0 Pr ðZ ¼ jÞ s , defined for jsj # 1. When R 0 . 1, the long-
increase individual variation, favouring extinction but risking term probability of disease extinction after introduction of a single infected
ongoing SSEs. Larger data sets are needed to establish this pattern individual is the unique solution of q ¼ g(q) on the interval (0,1). For a
definitively. negative binomial offspring distribution Z < NegB(R 0 ,k), the pgf is
If highly infectious individuals can be identified predictively (see gðsÞ ¼ ð1 þ Rk0 ð1 2 sÞÞ2k . Under population-wide control, Z pop c < NegBðð1 2
Supplementary Notes) then the efficiency of control could be greatly cÞR0 ; kÞ and therefore g pop ðsÞ ¼ ð1 þ ð1 2 cÞ Rk0 ð1 2 sÞÞ2k ; and the variance-to-
increased (Fig. 3c, d). Focusing half of all control effort on the most mean ratio is 1 þ (1 2 c)R 0/k. Under
random individual-specific control, the
2k
infectious 20% of cases is up to threefold more effective than random exact pgf is g ind ðsÞ ¼ c þ ð1 2 cÞ 1 þ Rk0 ð1 2 sÞ with variance-to-mean ratio
control (Fig. 3d). When k ¼ 0.1 or 0.5, outbreak containment is 1 þ R0 =k þ cR0 . This scenario can be approximated by Z ind;NB < NegBðð1 2

c ind ind 2kind
assured for targeted control levels at roughly half the threshold level cÞR0 ; kind
c ), where k ind
c is the solution to p0 þ cð1 2 p0 Þ ¼ 1 þ R c =k c
c
and
of c ¼ 1 2 1/R 0 for random control. Gains in efficiency increase with decreases monotonically as c increases. Further details, descriptions of outbreak
more intense targeting of high-n cases, but saturate as overall cover- simulations and formal analysis of control measures are found in the Sup-
age c increases (Supplementary Fig. 3c, d). Again, branching process plementary Notes.
theory generalizes these findings: for a given proportion c of
Received 1 March; accepted 22 August 2005.
individuals controlled, greater targeting of higher-n individuals
leads to lower effective reproductive number R and higher extinction 1. Levin, S. A., Grenfell, B., Hastings, A. & Perelson, A. S. Mathematical and
probability q (see Supplementary Notes). computational challenges in population biology and ecosystems science.
The data sets analysed here were collected from published litera- Science 275, 334–-343 (1997).
2. Becker, N. G. & Britton, T. Statistical studies of infectious disease incidence.
ture, and may be subject to selection bias for successful invasions and J. R. Stat. Soc. B 61, 287–-307 (1999).
SSEs rather than typical disease behaviour. Surveillance data sets are 3. Anderson, R. M. & May, R. M. Infectious Diseases of Humans: Dynamics and
less vulnerable to this bias, but may under-report isolated cases. We Control (Oxford Univ. Press, 1991).
urge that detailed transmission tracing data be collected and made 4. Diekmann, O. & Heesterbeek, J. A. P. Mathematical Epidemiology of Infectious
Diseases: Model Building, Analysis, and Interpretation (Wiley, Chichester, 2000).
public whenever possible, even if unexceptional. At a minimum, we 5. Leo, Y. S. et al. Severe acute respiratory syndrome—Singapore, 2003. Morbid.
propose a new measure for inclusion in outbreak reports: the Mortal. Wkly Rep. 52, 405–-411 (2003).
proportion of cases not transmitting (p 0), which, together with R 0 6. Shen, Z. et al. Superspreading SARS events, Beijing, 2003. Emerg. Infect. Dis. 10,
is sufficient to estimate the degree of variation in n (Supplementary 256–-260 (2004).
7. Dye, C. & Gay, N. Modeling the SARS epidemic. Science 300, 1884–-1885
Fig. 4). As more data become available, trends may emerge in the
(2003).
degree of variation present, for example, for different modes of 8. Lipsitch, M. et al. Transmission dynamics and control of severe acute
spread or levels of virulence. Richer data sets may also enable testing respiratory syndrome. Science 300, 1966–-1970 (2003).
of the branching process assumption that case outcomes are inde- 9. Riley, S. et al. Transmission dynamics of the etiological agent of SARS in
pendent and identically distributed, by detecting possible corre- Hong Kong: Impact of public health interventions. Science 300, 1961–-1966
(2003).
lations in n values within transmission lineages or systematic 10. Bauch, C. T., Lloyd-Smith, J. O., Coffee, M. & Galvani, A. P. Dynamically
changes as outbreaks progress. modeling SARS and other newly-emerging respiratory illnesses: past, present,
Our results have broad implications for emerging disease future. Epidemiology 16, 791–-801 (2005).

358
© 2005 Nature Publishing Group
NATURE|Vol 438|17 November 2005 LETTERS

11. Koopman, J. Modeling infection transmission. Annu. Rev. Public Health 25, syndrome reveal similar impacts of control measures. Am. J. Epidemiol. 160,
303–-326 (2004). 509–-516 (2004).
12. May, R. M. & Anderson, R. M. Transmission dynamics of HIV infection. Nature 27. Ha, L. D. et al. Lack of SARS transmission among public hospital workers,
326, 137–-142 (1987). Vietnam. Emerg. Infect. Dis. 10, 265–-268 (2004).
13. Woolhouse, M. E. J. et al. Heterogeneities in the transmission of infectious 28. Park, B. J. et al. Lack of SARS transmission among healthcare workers, United
agents: Implications for the design of control programs. Proc. Natl Acad. Sci. States. Emerg. Infect. Dis. 10, 244–-248 (2004).
USA 94, 338–-342 (1997). 29. Ferguson, N. M., Fraser, C., Donnelly, C. A., Ghani, A. C. & Anderson, R. M.
14. Lloyd-Smith, J. O., Getz, W. M. & Westerhoff, H. V. Frequency-dependent Public health risk from the avian H5N1 influenza epidemic. Science 304,
incidence in models of sexually transmitted diseases: portrayal of pair-based
968–-969 (2004).
transmission and effects of illness on contact behaviour. Proc. R. Soc. Lond. B
30. Farrington, C. P., Kanaan, M. N. & Gay, N. J. Branching process models for
271, 625–-634 (2004).
surveillance of infectious diseases controlled by mass vaccination. Biostatistics
15. Anderson, R. M. et al. Epidemiology, transmission dynamics and control of
SARS: the 2002–-2003 epidemic. Phil. Trans. R. Soc. Lond. B 359, 1091–-1105 4, 279–-295 (2003).
(2004).
16. Gani, R. & Leach, S. Epidemiologic determinants for modeling pneumonic Supplementary Information is linked to the online version of the paper at
plague outbreaks. Emerg. Infect. Dis. 10, 608–-614 (2004). www.nature.com/nature.
17. Becker, N. & Marschner, I. in Stochastic Processes in Epidemic Theory
Acknowledgements We are grateful for comments and data suggestions from
(eds Picard, P., Gabriel, J. P. & Lefevre, C.) 90–-103 (Springer-Verlag, New
York, 1990). B. Bolker, J. Edmunds, N. Ferguson, A. Galvani, R. Gani, N. Gay, J. Gog,
18. Becker, N. Analysis of Infectious Disease Data 48–-59 (Chapman & Hall, London, B. Grenfell, H. Hethcote, D. Heymann, A. Hubbard, N. Jewell, J. Lauer, R. May,
1989). T. Porco, C. Roth, D. Smith and B. Williams. We thank R. Gani for providing
19. Bailey, N. T. J. The Mathematical Theory of Infectious Diseases 2nd edn, 263–-265 unpublished data from a previous publication, and L. Matthews for sharing work
(Griffin, London, 1975). ahead of print. Our research was supported by the NSF, NIH-NIDA, the James
20. Keeling, M. J. & Grenfell, B. T. Disease extinction and community size: S. McDonnell Foundation, the NSF/NIH Ecology of Infectious Disease Program,
Modeling the persistence of measles. Science 275, 65–-67 (1997). and the South African Centre for Epidemiological Modelling and Analysis
21. Lloyd, A. L. Destabilization of epidemic models with the inclusion of realistic (SACEMA).
distributions of infectious periods. Proc. R. Soc. Lond. B 268, 985–-993 (2001).
22. Meyers, L. A., Pourbohloul, B., Newman, M. E. J., Skowronski, D. M. & Author Contributions J.O.L.-S. and W.M.G. conceived the study. J.O.L.-S.
Brunham, R. C. Network theory and SARS: predicting outbreak diversity. collected and analysed outbreak data, conducted dynamic modelling, and
J. Theor. Biol. 232, 71–-81 (2005). drafted and revised the text. S.J.S. conducted formal analysis of branching
23. Boswell, M. T. & Patil, G. P. Random Counts in Models and Structures (ed. Patil, processes and control measures. W.M.G. provided technical input on
G. P.) 7–-8 (Pennsylvania State Univ. Press, University Park, 1970). superspreading and control analyses. All authors contributed conceptually, and
24. Burnham, K. P. & Anderson, D. R. Model Selection and Multimodel Inference: A edited or commented on the text.
Practical Information-Theoretic Approach (Springer, New York, 2002).
25. Lloyd-Smith, J. O., Galvani, A. P. & Getz, W. M. Curtailing transmission of Author Information Reprints and permissions information is available at
severe acute respiratory syndrome within a community and its hospital. Proc. npg.nature.com/reprintsandpermissions. The authors declare no competing
R. Soc. Lond. B 270, 1979–-1989 (2003). financial interests. Correspondence and requests for materials should be
26. Wallinga, J. & Teunis, P. Different epidemic curves for severe acute respiratory addressed to to J.O.L.-S. ([email protected]).

359
© 2005 Nature Publishing Group