Woodruff, Rosalie E.1; Guest, Charles S.1; Garner, Michael G.2; Becker, Niels1; Lindesay, Janette3; Carvan, Terence4; Ebi, Kristie5
Ross River virus is an arbovirus widely distributed throughout Australia. It has also been reported in many Pacific island countries, including Papua New Guinea, the Solomon Islands, American Samoa, Fiji, New Caledonia, and the Cook Islands. 1,2,3 The virus causes epidemic polyarthritis, with rash and fever in some patients, and joint symptoms in 95% of cases. 4 The arthritic symptoms may persist for months and can be severe and debilitating; the disease has been reported to linger in some patients for years. 5 The disease is an important public health issue in Australia, with an average of 5000 cases per year recorded since 1991. There is no treatment for the disease and, in the absence of a vaccine, prevention remains the sole public health strategy.
The primary enzootic cycle is between reservoir hosts (typically marsupials) and the mosquito vector. Given low immunity in the host population and suitable climatic conditions, massive virus amplification occurs, resulting in a spillover of infection into human populations. Weather conditions directly affect the breeding, survival, and abundance of arthropods and their extrinsic incubation period. 6,7 Elsewhere, 8 we undertook a national review of the epidemiology of Ross River virus disease (RRV disease) during the period 1991–1999, and observed the influence of climate, environmental modifications, and vector-host dynamics on disease patterns.
In the current study we test the hypothesis that weather variables have the potential to predict epidemics of RRV disease with sufficient accuracy to be of benefit to health authorities. Our aim was to develop simple models for the probability of epidemics. These could enable early warning of climate conditions conducive to outbreaks of disease, and could be useful as a decision tool for health authorities in risk-management planning.
In this paper, we present the results of predictive models for two regions. We also discuss the benefits and limitations of spatial analysis within bioclimatic regions as a method for arboviral prediction.
The study involved two regions surrounding the Murray River, the largest water source in southeastern Australia (Figures 1 and 2). The Murray area has a long record of human epidemics (documented since 1928), as well as a (comparative) wealth of information available on virus, host, and vector populations. The principal transmission vector in the area is Culex annulirostris, a freshwater-breeding mosquito. The primary amplification vector during this period is likely to have been C australicus. 9Aedes vigilax and A camptorynchus, which inhabit salt pans and brackish water, 7,10 and A sagax and A vittiger have also been implicated. 11
Region 1 is on the edge of the large internal landmass of Australia, and experiences hot dry summers and cold winters, whereas Region 2 (further south and closer to the moderating influence of the coast) has a temperate climate pattern (Figure 2). Average yearly rainfall is higher in Region 2 (563 mm) than Region 1 (343 mm). The majority of rainfall occurs in winter and early spring in both regions (Figure 3). The number and intensity of rainfall events over southeastern Australia has been related to the El Niño-Southern Oscillation (ENSO) cycle. 12
The base unit of observation is the statistical local area (SLA), an administrative division used by the Australian Bureau of Statistics. Table 1 shows the characteristics of the regions.
We obtained positive notifications 13 of Ross River virus disease for the period from July 1991 to June 1999 (8 years) for the study regions from the Commonwealth Disease Network of Australia New Zealand, National Notifiable Diseases Surveillance Scheme. Individual case data included the date when symptoms were estimated to have commenced, and the date when the relevant health authority was notified about the case.
Because notification data may be biased toward cases with typical clinical symptoms, 14 and because people with less severe illness may not seek medical help or may be misdiagnosed, the notification data are undoubtedly understated. In the two study regions, 2098 cases were available for analysis (Table 1).
We aggregated case data from postcode to SLA level using geographic information system software. 15 We obtained estimated annual resident population by sex and 5-year age category for each SLA for 1996 (mid-interval) from the Australian Bureau of Statistics, 16 and we calculated average annual standardized rates. We grouped SLAs with similar long-term rainfall, temperature and humidity conditions, and vegetation type, into two “bioclimatic” regions, using methods described elsewhere. 17 SLAs with 10 or fewer cases were excluded.
The Data Drill 19 provided monthly weather data relevant to the biology of the vector and reservoir hosts and the replication of the virus. Data initially were in the form of interpolated daily surfaces at 0.05-degree resolution (ie, approximately 5 km2 grid), derived from point observations recorded at weather stations across the country. 18 The interpolation method adjusted for elevation as well as latitude and longitude, and took into account the long-term variability of each variable at each weather station. 19 We calculated a monthly spatial summary value for the weather variables in each SLA (as listed in Table 2). Long-term average estimates for all the variables were also obtained, based on the period 1955 to 1999. Figure 3 shows the long-term average monthly total rainfall, average minimum, and average maximum temperatures for both regions.
Irrigation and mosquito control measures were potential confounders. The predominant irrigation method used by cropping industries along the Murray River is water release from the Hume Reservoir throughout the growing season (typically September through April). The Murray-Darling Basin Commission provided data on flow from the Reservoir (megaliters per day). No mosquito control activities were conducted in the study period on the northern border of the Murray (part of Region 1). In some SLAs to the south, limited interventions, in the form of sporadic larviciding, were conducted in all years.
Analyses were undertaken using the statistical package Stata 6.0. 20 Figure 4 summarizes the analytic methods used for this study. We conducted analyses using the “RRV disease year,” which runs from July (month 1) to the following June (month 12), to contain a typical annual epidemic curve. The dichotomous outcome variable was defined as whether or not an epidemic occurred in a SLA in one year. We defined an epidemic as the number of cases in any one year (July–June) that exceeded the mean plus one standard deviation of all cases recorded in that SLA during the study period (July 1991–June 1999). The standard deviation was based on the assumption that incidence has a Poisson distribution.
We expressed the probability of an epidemic year in terms of weather variables by a logistic regression model. The base unit for the weather variables was a month; averages of more than 1 month were derived for some variables (Table 2). Variables were first considered one at a time with the outcome (epidemic or nonepidemic year in a SLA). Provided they had a P < 0.05, they were included in a multivariable logistic regression model. The form of the equation was:
log (P/[1−P]) = α + β1x1 + β2x2 +. . .+ βnxn
where P is the probability of an epidemic, the xi are explanatory variables specific to a SLA, and α, β1. . .βn are constants that are estimated from the data. Backward stepwise regression was used initially to select an adequate model; variables with a P-value of > 0.2 were removed.
Our main criteria for judging a satisfactory model were that it have more than 90% accuracy in predicting the outcome values, and that it have significance on a Hosmer-Lemeshow goodness-of-fit test. 21 We considered a value of P to be an accurate reflection of an epidemic if its value lay above 0.75, and an accurate reflection of a nonepidemic if it fell below 0.25. In recognition of the fact that P is estimated, we considered the precision of the estimated regression coefficients.
We developed two models for each region. The “early warning” model includes weather conditions for the months July–November (late southern hemisphere winter to end of spring). Cases typically commence around November, and early warning of the likelihood of an epidemic season in November would be timely. Control measures at this point would be effective in reducing mosquito breeding. We also developed a second, or “late warning” model, which includes variables for December–February (the southern hemisphere summer). Initial cases would have appeared by this time, but in many years the bulk of cases are not apparent until March or April. As the interval between onset and symptoms is often as low as 7–9 days, 22 it would still be useful to check the probability of an epidemic at this time, and to issue public alerts if the results were conclusive.
We validated models within the dataset in a rotating fashion; each combination of 7 years in the series was used to predict the eighth. Public health practitioners in the regions suggested that a model would be satisfactory for early warning purposes if it correctly predicted more than 70% of the epidemics in validation. For this we used a cutoff probability of 0.5.
Widespread epidemics occurred in 1992–1993 (Region 1: 12 of 14 SLAs; Region 2: 24 of 24 SLAs) and 1996–1997 (Region 1: 11 of 14 SLAs; Region 2: 19 of 24 SLAs). Few epidemics were recorded in SLAs in other years. The 8-year average annual incidence rate and the number of cases and epidemics in each region are shown in Table 1.
Early Warning Models (Variable Month 1–5)
Table 3 shows the predictor variables, their coefficient estimates, standard errors, and 95% confidence intervals (CI). In general, epidemics were best predicted by early spring rainfall variables and late spring temperature variables in both regions.
Early winter (June–July) rainfall patterns differed between the two epidemic years (Figures 5 and 6). Nonetheless, number of rain days (RNDAY) in July (midwinter) was a predictor variable for both regions. There were minor differences between regions in terms of particular variables that best predicted precipitation in the late winter through early spring period: average August/September monthly rainfall totals (RNTOT) in Region 1, and August sea surface temperature (SST) and September RNDAY in Region 2. Sea surface temperature in the month of August was a positive predictor in the temperate Region 2. RNDAY456L, the average of RNDAY for the spring months of the year preceding the epidemic year, was negatively associated with epidemics in both regions.
November temperature variables were negatively associated with epidemics in both regions: average maximum temperature (TMAX) was a predictor in the Region 1 model (OR = 0.01; 95% CI = 0.00–0.37), and absolute maximum temperature (ABSTMAX) was a predictor in Region 2 (OR = 0.37; 95% CI = 0.24–0.57). This is consistent with the spring breeding habits of C australicus, the principal amplification vector, which shows greatest population growth at lower temperatures (range, 6–20°C). 9 Only 9% of epidemics occurred when November TMAX exceeded 22°C in Region 2, compared with 69% of nonepidemics. The pattern was the same for Region 1, although the TMAX level was 4°C higher. Relative humidity in late spring was a negative predictor of epidemics in Region 1 (the drier region), although the biological reason for this is unclear.
Late Warning Models (Variable Month 1–8)
Including absolute minimum temperature (ABSTMIN) and relative humidity at maximum temperature (RHTMAX) for February (late summer in the Southern Hemisphere) made a striking improvement to the accuracy of epidemic prediction in the Region 1 model (Table 4). In Region 2, February TMAX was positively associated with epidemics (and a better predictor in this region than ABSTMIN). Other model variables were absolute maximum temperature (ABSTMAX) in early summer (December) and vapor pressure in mid- to late summer (January–February). A different summer rainfall pattern between the two epidemic years prevented any of the rainfall variables for those months from adequately describing these conditions (compare December–February months in Figures 5 and 6).
In the 1996–1997 epidemic year, when lower than average summer rainfall was recorded, average temperatures for February were 27°C in Region 1 and 25°C in Region 2. Several studies have identified the increased survival of infected female mosquitoes as the single most important factor in contributing to infections. 23,24 The life span of C annulirostris is longest when mean ambient temperatures persist at around 25–27°C. 9,25 A predictor in the Region 2 model was February TMAX (the mean value for 1996–1997 was 32°C, 4°C higher than the long-term mean). At higher temperatures vectorial capacity is increased, owing to a reduction in the extrinsic incubation period. 26
Mosquito reproduction can be interrupted or terminated by even several days of cold weather in the summer months. 27 In Region 1, which abuts the Australian inland desert and is affected by temperature extremes, February ABSTMIN was a substantial predictor of epidemics. A study in the RRV disease endemic region of temperate northern Australia has also found increases in minimum temperature to be positively associated with case incidence. 28
Epidemics were predicted well for the two stages in both regions. Table 5 reports the accuracy of the models for each year in the period, and the sensitivity and specificity values for the models over all years is reported in Table 6. All models predicted nonepidemic years better than epidemic years.
In Region 1, the early warning model had a sensitivity of 62% (16 of 26 epidemics). Prediction improved dramatically for the late warning model, which had a sensitivity of 96% (25 of 26 epidemics). In Region 2, the early warning model had a sensitivity of 73% (32 of 44 epidemics). The late warning model, however, had a lower sensitivity of 66%.
These findings indicate that weather conditions at relatively coarse temporal and spatial resolutions can be used to predict RRV disease epidemics with sufficient accuracy and lag time for public health planning. Figure 7 summarizes the main results. Both regions had models with more than 70% sensitivity, and in Region 2 epidemics were predicted with 73% sensitivity as early as November (when cases typically commence).
The Influence of Rainfall Pattern
The timing, duration, and extent of rainfall events in epizootic regions are critical indicators of outbreaks. 29 Two different rainfall patterns were evident in this study: excess “winter” rainfall and excess “summer” rainfall. Breeding of the main transmission vector, C annulirostris, occurs from mid-spring to late autumn, 10 and populations typically peak in January or February. 30 Reports from previous epidemics in these regions have linked heavy rainfall or extensive flooding in the December to February period with intense breeding of Culex mosquitoes, and hence to a spillover of infection to human populations. Large epidemics in the years 1928, 1956, 1984, and 1971 demonstrate this pattern. 31,27,32,33 In the first epidemic year of the study period, 1992–1993, heavy rainfall (nearly twice the long-term mean) commenced in August, and above average rainfall was sustained throughout the spring and up until late January (see Figure 5).
Nicholls 34 has noted that severe outbreaks of RRV disease in temperate Australia appear to follow heavy summer rainfall, suggesting an association with La Niña, the cool phase of the ENSO cycle. Harley and Weinstein 35 found no association between the southern oscillation index (SOI) and RRV disease outbreak years for Australia as a whole, although some RRV disease outbreaks in southeastern Australia have been positively correlated with September values of the SOI. 36 We found August SST to be a better predictor of rainfall excess, and hence epidemics, than the SOI. This was not surprising, as calculation of the SOI incorporates air pressure values from both sides of the Pacific basin (Darwin and Tahiti), whereas SST records the rise and fall of temperatures directly in the El Niño region of the eastern Pacific (Table 2, note ∥).
Not all outbreaks in the Murray area have followed heavy summer rainfall. 4 In 1980–1981, abnormally high spring rainfalls were recorded before outbreaks. In the 1996–1997 epidemic year, total rainfall over summer was below the period average in both regions. Instead, higher than average rainfall commenced as early as June, and continued until late August in Region 1 and September in Region 2 (Figure 6). Although high rainfalls were recorded in some of these months in other years (1991–1992 and 1995–1996), they were interspersed with lower than average values in either July or August. These lower values may have been a factor in aborting a buildup in these years.
We speculate that sustained winter and spring rainfall, even in the absence of excess summer rainfall, could enhance transmission by two mechanisms. First, it would allow a longer time period for amplification of virus levels between the initiating vector (Aedes species) and host populations. Floodwater Aedes maintain the virus by transovarial transmission of infection through embryonated eggs. 37 These mosquitoes overwinter as drought-resistant eggs in mud flats and creek beds. 38 After heavy winter rainfall, the females emerge and infect the vertebrate hosts. 37 In 1996–1997, minimum temperatures during July and August were higher than average in both regions. The combination of these two factors is likely to have provided for early and prolific breeding of Aedes populations and an extended period of virus buildup, thus increasing transmission potential. Second, the prolonged heavy rainfall (which resulted in widespread flooding into October) acted to raise the water table in the regions, reducing absorption and runoff. As a consequence, ground pools were able to remain into summer, despite low summer rainfalls, providing sites for Culex and Aedes breeding. These species extended the infection to humans, and epidemics ensued.
Irrigation practices, with a special release from the Hume Reservoir, appear likely to have contributed to the maintenance of wet conditions throughout the 1996–1997 summer. The Hume modulates water levels for the Murray River, influencing flood patterns downstream. A “special release” of an excessive volume of water from the Hume was authorized in October 1996 (owing to fears of a crack in the reservoir wall), which artificially maintained the Murray River flow at flood levels for more than 30 days. Under normal circumstances, a maximum irrigation release of 25,000 ml per day is permitted, and negligible flooding occurs. Flow levels in October of 1996 were persistently above 25,000 ml per day (for 2 weeks they were more than 90,000 ml per day), and did not return to normal for 6 weeks. Under this volume of water the area inundated increased rapidly, 39 and billabongs and low depressions around the river filled up.
Extrinsic (ie, climatic) factors need to be combined with host-virus population dynamics in epidemic prediction. 40 Below-average rainfall in the spring of the preceding year was a necessary, but not sufficient, explanatory variable in all the models. We attribute this to the influence of vector activity on vertebrate host immunity. Higher than normal rainfalls during an October–December period would initiate and support mosquito virus activity and result in the infection of a proportion of the vertebrate population. As the breeding cycle of the primary vertebrate host (macropods) takes more than 1 year to complete, 41 the pool of susceptible vertebrates in the following year would be reduced, virus amplification would be minimal, and the probability of human cases would also be small. Conversely, in the year after an epidemic year, the high level of vertebrate population immunity would be sufficient to reduce the probability of successive epidemic years to very low levels, even if other climatic conditions were suitable (as occurred in 1993–1994 in the study regions). This finding has also been demonstrated by studies on RRV disease epidemic activity in arid northwestern Australia. 42
Limitations of the Study
The principal limitation of this study lies with the notification data, which were reported by place of residence rather than place of suspected infection. We made the assumption that in rural areas with relatively large SLAs, work and recreation (and hence transmission) generally occur within the same SLA. However, it is likely that a small portion of cases were misclassified, which would weaken the association. Furthermore, because of a lack of data, we were not able to control for the possible confounding effect of rice irrigation practices, which may vary by SLA. Although we understood mosquito control interventions to have been minimal throughout the period, estimates of the timing and effectiveness of sprays would have strengthened the findings.
The short span of data was a problem for some climate variables. Given the unusual number of El Niño events that occurred during the study period (1990–1994 and 1997–1998), we cannot be certain about the strength of the predictive power of sea surface temperature in the Region 2 early warning model. One La Niña (cool phase) occurred in 1998–1999, but was only of weak intensity during the critical late winter to spring months.
The cross-validation method assessed how well the variables in each of the models were able to predict epidemics, rather than the predictive performance of the estimated coefficients for the 8-year model. Epidemics occurred infrequently in this dataset (in general, in only 2 of the 8 years). In the rotating validation process, when 1 year of epidemic information was removed, only half of the data on epidemics remained to derive the parameter estimates of the predictive model for that year. Given that there are several variables in each model, it may be that some of these predicted the outcome by chance.
Climate, Climate Change, and Health
Beyond the seasonal effects studied here, our changing climate—including interannual cycles and longer-term natural and human-induced cycles—will continue to influence the extrinsic and intrinsic factors that drive vector-borne diseases. Climate projections for 2030 indicate that temperatures may rise by 0.4–2°C over most of Australia. By 2070 they may rise by 1–6°C. 43 Winter rainfall in the southeast of the country may decrease by up to 10% by 2030 (35% by 2070), and summer rainfall in these regions may increase by 10% to 20% by 2030 (35% to 60% by 2070) depending on whether the results of slab or coupled models are consulted. 43
The effect of these changes on the breeding and survival of arthropods and vertebrate host populations, even in a region as “localized” as southeastern Australia, is difficult to anticipate. Russell speculates that Aedes populations in dry areas (such as Region 1) may be adversely affected by decreased winter rainfall, possibly resulting in delayed or precluded virus activity. 6 We would expect this to result in an overall reduction in epidemics in years when the “winter pattern” might have occurred. Conversely, the predicted increase in summer rainfall may increase the availability of mosquito habitat that, combined with higher average temperatures, may lead to higher humidity, a lengthened season of abundance, and greater transmission levels. 6 Understanding how ENSO may change with global climate change is also essential in anticipating the impact of future climate on vector-borne diseases. Currently, climate models have mixed success in estimating interannual influence. 44
Weather forecasts can be used in conjunction with other surveillance techniques to identify conditions suitable for an epidemic of RRV disease. Seasonal arbovirus activity is already monitored in parts of southeastern Australia, with weekly trapping of mosquitoes throughout the Ross River season used to record population profiles and virus isolates. The multistaged approach (early and late warning models) developed in this study enables response plans to be adjusted as forecast certainty increases, allowing health authorities to make the most of limited resources. Although both weather and climate are pivotal in generating the conditions needed for epidemics of arboviral disease, human influence (through environmental modifications such as irrigation) can also contribute to events.
We acknowledge Keith Moodie and others at SILO, Queensland Department of Natural Resources and Mines, for their assistance with compiling the climate data. We acknowledge the Communicable Diseases Network of Australia and New Zealand for the disease data, and the Murray-Darling Basin Commission for providing data on releases from the Hume Reservoir. We thank staff at the Animal Health Science Unit, Department of Agriculture Fisheries and Forestry Australia, for the use of computing resources.
1. Rosen L, Gubler DG, Bennett PH. Epidemic polyarthritis (Ross River) virus infection in the Cook Islands. Am J Trop Med Hyg 1981; 30: 1294–1302.
2. Aaskov JG, Mataika JU, Lawrence GW, Rabukawaqa W, Tucker MJAR, Dalglish DA. An epidemic of Ross River virus infection in Fiji. Am J Trop Med Hyg 1981; 30: 1053.
3. Scrimgeour EM, Aaskoz JG, Matz LR. Ross River virus arthritis in Papua New Guinea. Trans R Soc Trop Med Hyg 1987; 81: 833–834.
4. Mudge PR, Aaskov JG. Epidemic polyarthritis in Australia, 1980–1981. Med J Aust 1983: 2: 269–273.
5. Westley-Wise VJ, Beard JR, Sladden TJ, Dunn TM, Simpson J. Ross River virus infection on the north coast of New South Wales. Aust N Z J Public Health 1996; 20: 87–92.
6. Russell RC. Mosquito-borne arboviruses in Australia: the current scene and implications of climate change for human health. Int J Parasitol 1998; 28: 955–969.
7. Kay BH, Aaskov JG. Ross River virus (epidemic polyarthritis). In: Monath TP, ed. The Arboviruses: Epidemiology and Ecology.
Volume IV. Boca Raton, Florida: CRC Press, 1989:93–112.
8. Woodruff RE, Garner MG, Ebi K. Ross River Virus disease: the influence of climate on distribution and trends. Submitted for Publication.
9. Dhileepan K. Mosquito seasonality and arboviral incidence in Murray Valley, southeast Australia. Med Vet Entomol 1996; 10: 375–384.
10. Dobrotworsky NV. The Mosquitoes of Victoria: (Diptera, Culicidae). London: Melbourne University Press, 1965.
11. Russell RC. Ross River virus: disease trends and vector ecology in Australia. Bull Soc Vect Ecol 1994; 19: 73–81.
12. Nicholls N, Kariko A. East Australian rainfall events: interannual variation, trends, and relationships with the Southern Oscillation. J Clim 1993; 6: 1141–1152.
13. Mackenzie JS, Broom AK, Calisher CH, et al
. Diagnosis and reporting of arbovirus infections in Australia. Commun Dis Intell 1993; 17: 202–207.
14. Tai K, Whelan PI, Patel MS, Currie B. An outbreak of epidemic polyarthritis (Ross River virus disease) in the Northern Territory during the 1990–1991 wet season. Med J Aust 1993: 158: 522–525.
15. MapInfo Professional
[computer program]. Version 5.0. New York: MapInfo Corporation, 1998.
16. Australian Bureau of Statistics. Integrated Regional Database (IRDB)
. Version 1998. Canberra: Commonwealth of Australia, 1998.
17. Woodruff RE, Garner MG, Guest CS. Using GIS to study the epidemiology of Ross River virus, an important zoonosis in Australia. Paper presented at GIS Veterinary Conference, Lancaster, UK, September, 2001.
18. Jeffrey SJ, Carter JO, Moodie KB, Beswick AR. Using spatial interpolation to construct a comprehensive archive of Australian climate data. Environ Model Softw 2001; 16: 309–330.
19. Hutchinson MF. Interpolating mean rainfall using thin plate smoothing splines. Int J Geog Info Sys 1995; 9: 385–403.
20. Stata Statistical Software
. Edition 6.0. College Station: Stata Press, 1999.
21. Hosmer DW, Lemeshow S. Goodness-of-fit tests for the multiple logistic regression model. Commun Stat 1980; A9: 1043–1069.
22. Fraser JRE, Cunningham AL. Incubation time of epidemic polyarthritis. Med J Aust 1980; 1: 550–551.
23. Scott TW, McLean RG, Francy DB, Card CS. A simulation model for the vector-host transmission system of a mosquito-borne avian virus, Turlock (bunyaviridae). J Med Entomol 1983; 20: 625–640.
24. de Moor PP, Steffens FE. A computer-simulation model of an arthropod-borne virus transmission cycle, with special reference to chickungunya virus. Trans R Soc Trop Med Hyg 1970; 64: 927–934.
25. Mottram P, Kay BH, Kettle DS. The effect of temperature on eggs and immature stages of Culex annulirostris
Skuse (Diptera: Culicidae). J Aust Entom Soc 1986; 25: 131–135.
26. Weinstein P. An ecological approach to public health intervention: Ross River virus in Australia. Environ Health Perspect 1997; 105: 364–366.
27. Hawkes RA, Boughton CR, Naim HM, Stallman ND. A major outbreak of epidemic polyarthritis in New South Wales during the summer of 1983/1984. Med J Aust 1985; 143: 330–333.
28. Tong S, Bi P, Parton K, Hobbs J, McMichael AJ. Climate variability and transmission of epidemic polyarthritis. Lancet 1998; 351: 1100.
29. Davies FG, Linthicum KJ, James AD. Rainfall and epizootic Rift Valley Fever. Bull World Health Organ 1985; 63: 941–943.
30. Russell RC. Seasonal activity and abundance of the arbovirus vector Culex annulirostris
skuse near Echuca, Victoria, in the Murray Valley of Southeastern Australia 1979–1985. Aust J Explor Biol Med Sci 1986; 64: 97–103.
31. Nimmo JR. An unusual epidemic. Med J Aust 1928; 1: 549–550.
32. Anderson SG, French EL. An epidemic exanthem associated with polyarthritis in the Murray Valley. Med J Aust
33. Seglenieks V, Moore BW. Epidemic polyarthritis in South Australia: report of an outbreak in 1971. Med J Aust 1974; 2: 552–556.
34. Nicholls N. El Niño-Southern. Oscillation and vector-borne disease. Lancet 1993; 342: 1284–1285.
35. Harley DO, Weinstein P. The Southern Oscillation Index and Ross River virus outbreaks. Med J Aust 1996; 165: 531–532.
36. Maelzer D, Hales S, Weinstein P, Zalucki M, Woodward A. El Niño and arboviral disease prediction. Environ Health Perspect 1999; 107: 817–818.
37. Lindsay MDA, Broom AK, Wright AE, Johansen CA, Mackenzie JS. Ross River virus isolations from mosquitoes in arid regions of Western Australia: implications of vertical transmission as a means of persistence of the virus. Am J Trop Med Hyg 1993; 49: 686–696.
38. Marshall ID, ed. The Epidemiology of Murray Valley Encephalitis in Eastern Australia–Patterns of Arbovirus Activity and Strategies of Arbovirus Survival. Arbovirus Research in Australia. Vol. 2. Brisbane: CSIRO Division of Animal Health and Queensland Institute of Medical Research, 1979.
39. River Murray Commission. Report on Flood Plain Management of Murray River Between Hume Dam and Yarrawonga
. Albury: Murray-Darling Basin Commission, September 1976.
40. Hay SI, Myers MF, Burke DS, et al
. Etiology of interepidemic periods of mosquito-borne disease. Proc Natl Acad Sci U S A 2000; 97: 9335–9339.
41. Caughley G, Shepherd N, Short J. Kangaroos. Their Ecology and Management in the Sheep Rangelands of Australia. Cambridge: Cambridge University Press, 1987.
42. Mackenzie JS, Lindsay MD, Broom AK. Effect of climate and weather on the transmission of Ross River and Murray Valley encephalitis viruses. Microbiol Aust 2000; 21: 20–24.
43. Climate Impacts Group. Climate Change Scenarios for the Australian Region. Melbourne: CSIRO Division of Atmospheric Research, 1996.
44. Hennessy KJ, Suppiah R, Page CM. Australian rainfall changes, 1910–1995. Aust Meteorol Mag 1999: 48: 1–13.
© 2002 Lippincott Williams & Wilkins, Inc.