VECTOR-BORNE AND ZOONOTIC DISEASES Volume 14, Number 6, 2014 ª Mary Ann Liebert, Inc. DOI: 10.1089/vbz.2012.1284
Analysis and Prediction of Ross River Virus Transmission in New South Wales, Australia Victoria Ng, Keith Dear, David Harley, and Anthony McMichael
Abstract
Background: Ross River virus (RRV) disease is the most widespread mosquito-borne disease in Australia. The disease is maintained in enzootic cycles between mosquitoes and reservoir hosts. During outbreaks and in endemic regions, RRV transmission can be sustained between vectors and reservoir hosts in zoonotic cycles with spillover to humans. Symptoms include arthritis, rash, fever and fatigue and can persist for several months. The prevalence and associated morbidity make this disease a medically and economically important mosquitoborne disease in Australia. Methods: Climate, environment, and RRV vector and reservoir host information were used to develop predictive models in four regions in NSW over a 13-year period (1991–2004). Polynomial distributed lag (PDL) models were used to explore long-term influences of up to 2 years ago that could be related to RRV activity. Results: Each regional model consisted of a unique combination of predictors for RRV disease highlighting the differences in the disease ecology and epidemiology in New South Wales (NSW). Events up to 2 years before were found to influence RRV activity. The shorter-term associations may reflect conditions that promote virus amplification in RRV vectors whereas long-term associations may reflect RRV reservoir host breeding and herd immunity. The models indicate an association between host populations and RRV disease, lagged by 24 months, suggesting two or more generations of susceptible juveniles may be necessary for an outbreak. Model sensitivities ranged from 60.4% to 73.1%, and model specificities ranged from 57.9% to 90.7%. This was the first study to include reservoir host data into statistical RRV models; the inclusion of host parameters was found to improve model fit significantly. Conclusion: The research presents the novel use of a combination of climate, environment, and RRV vector and reservoir host information in statistical predictive models. The models have potential for public health decision-making. Ross River virus disease—Predictive model—New South Wales—Australia—Vector-borne diseases—Mosquito vectors—Reservoir hosts.
Key Words:
Introduction
R
oss River virus (RRV) disease is the most widespread mosquito-borne disease in Australia. It is endemic and largely restricted to Australia, Papua New Guinea, and the Solomon Islands (Harley et al. 2001, Russell 2002). RRV infection may result in arthritis, rash, fever, and fatigue that can persist for several months (Harley et al. 2002, Mylonas et al. 2002). Since 1991, approximately 4300 notifications have been reported annually in Australia (National Notifiable Disease Surveillance System 2013). No vaccine exists and prevention relies on mosquito control (Carvan et al. 1986,
Tomerini et al. 2011) and personal protective measures (Westley-Wise et al. 1996, Harley et al. 2005, Ritchie et al. 2006). Increasingly, RRV predictive models are used (Williams et al. 2007, Jacups et al. 2008a, Jacups et al. 2011, Williams and Williams 2011). RRV is maintained in natural enzootic cycles between mosquitoes and reservoir hosts. Primary vectors include Aedes camptorhynchus, Ae. vigilax, and Culex annulirostris (Mackenzie et al. 1994, Russell 1994, Mackenzie et al. 1998, Russell 1998, Harley et al. 2001, Russell 2002). The apparent natural reservoir hosts are nonmigratory native macropods, primarily kangaroos and wallabies (Kay and Aaskov 1989,
National Centre for Epidemiology and Population Health, The Australian National University, Canberra, Australia.
422
ROSS RIVER VIRUS TRANSMISSION IN NEW SOUTH WALES, AUSTRALIA
Mackenzie and Smith 1996, Harley et al. 2001, Russell 2002). During outbreaks and in endemic regions, RRV transmission can be sustained between vectors and reservoir hosts in zoonotic cycles with spillover to humans (Hawkes et al. 1985, Aaskov and Doherty 1994, Harley et al. 2001, Smith et al. 2009). Climate and environmental factors influence breeding, survival, and abundance of RRV vectors and reservoir hosts. Many studies have demonstrated the link between climate, environment and RRV disease in humans (Tong et al. 1998, Tong and Hu 2001, Done et al. 2002, Tong and Hu 2002, Tong et al. 2002, Whelan et al. 2003, Jardine et al. 2004, Kelly-Hope et al. 2004a, Kelly-Hope et al. 2004b, KellyHope et al. 2004c, Tong et al. 2004, Tong et al. 2005, Biggs and Mottram 2008, Hu et al. 2010a, Vally et al. 2012). RRV predictive models have been developed using climate and environment data (Maelzer et al. 1999, Woodruff et al. 2002, Hu et al. 2004, Gatton et al. 2005, Bi et al. 2009, McIver et al. 2010, Jacups et al. 2011, Werner et al. 2011). Information on RRV vectors has also been used (Hu et al. 2006a, Hu et al. 2006b, Woodruff et al. 2006, Jacups et al. 2008b, Williams et al. 2009, Hu et al. 2010b). Reservoir host data have yet to be used in statistical models, but dynamic mathematical models suggest host data could improve model performance (Glass 2005, Carver et al. 2009b). Predictive models have been developed in Queensland, the Northern Territory, and Western Australia (Hu et al. 2004, Gatton et al. 2005, Hu et al. 2006a, Hu et al. 2006b, Woodruff et al. 2006, Jacups et al. 2008b, Hu et al. 2010a, McIver et al. 2010, Jacups et al. 2011), where notification rates are highest (Miller et al. 2005, Yohannes et al. 2006). In New South Wales (NSW), approximately 650 cases are reported annually, but there has been little research exploring climatic and environ-
423
mental associations. Our study objectives were to develop early warning predictive models in NSW using climate, environment, and RRV vector and reservoir host data, to investigate whether the inclusion of host parameters improves model fit, and to explore whether longer-term associations related to RRV activity can be incorporated into RRV models. Methods and Materials Study area
The study included four distinct bioclimatic regions in NSW (Fig. 1). Long-term climate classifications were obtained from the Australian Bureau of Meteorology (BOM) to define regions (Bureau of Meterology 2005a, Bureau of Meterology 2005b). Climate classifications for vegetation, seasonal rainfall, and temperature and humidity conditions were merged with Australian postcode boundaries in ArcGIS (Environmental Systems Research Institute [ERSI], Inc.), using a method described elsewhere (Woodruff et al. 2002; Woodruff 2003, Woodruff et al. 2006), to assign climate classifications to postcodes. Postcodes were grouped into four bioclimatic regions. Regions were selected for large numbers of cases and wide variation within years. The regions encompassed a wide range of climatic and environmental conditions (Table 1). Unit of analysis
The spatial unit of analysis was an Australian postcode because notifications are reported by postcode and postcodes can be grouped into homogeneous bioclimatic regions. The temporal units of observation were months, analyzed within a 24-month lag period.
FIG. 1. Four bioclimatic regions identified in New South Wales (NSW) as regional study areas for the regional Ross River virus (RRV) predictive models. Color images available online at www.liebertpub.com/vbz
424 Data
RRV disease notifications were obtained from the National Notifiable Diseases Surveillance System. A total of 8397 notifications were recorded in NSW between July, 1991, and June, 2004. The NSW population by Census Collection District (CCD) for 1991, 1996, and 2001 were obtained from the Australian Bureau of Statistics (ABS), and ArcGIS was used to redistribute CCD population to postcodes. Data for intercensus years were interpolated linearly (1991–2001) and extrapolated (2002–2004) using Stata (Release 10.0. Stata Corporation, College Station, TX). Daily BOM weather station data were obtained for evaporation, relative humidity, temperature, and rain days (Table 2), and monthly means were calculated. Postcodes that contained more than one weather station were assigned the average from those stations. Postcodes with no weather stations were assigned the inverse distance-weighted average of the closest stations up to a 75-km radius cutoff. Interpolated monthly climate surfaces were obtained for additional rainfall and temperature variables (Table 2). ArcGIS was used to assign climate surfaces data to postcodes. Other monthly data included sea surface temperature (SST), tidal height, Normalized Difference Vegetation Index (NDVI), water sources, distance from the coast, elevation, and Accessibility/Remoteness Index of Australia (ARIA) (Table 2). Climate and environment data were selected for their potential to influence vector and reservoir host ecology, or to act as an ecological proxy for conditions that support RRV activity. Weekly mosquito data were aggregated into monthly averages, and only trap sites that contained 35 or more months of data over the study period (156 months in total) were included (Table 2). Postcodes that contained more than one trap site were assigned the average trap catch. Postcodes that did not contain a trap site were assigned data from the nearest postcode neighbour with data. Yearly macropod population counts were collected by Kangaroo Management Zones (KMZ), of which there were 15 (Table 2). ArcGIS was used to redistribute kangaroo and wallaroo populations into postcodes in proportion to the percentage of area overlap. Populations were representative of the population in July (average month of surveying). Stata was used to interpolate monthly values from yearly population. Although kangaroo and wallaroo populations fluctuate over time, due to the relatively long breeding interval between young (8–12 months) (Shepherd 1987, TyndaleBiscoe and Renfree 1987, Dawson 1995, Pople and Grigg 1999), it was reasonable to interpolate monthly values from yearly populations. Statistical analyses
Statistical analyses were performed using Stata. The outcome variable for the models was the number of RRV disease over 12 months from July 1 to June 30, inclusive, because over 90% of cases are reported between November and June. Cases for July to June were predicted using monthly data, lagged to 24 months before January (Fig. 2). Thus, information from the first half of each year ( July to December) was used to predict cases for July to June. Only 15% of cases have been reported by December. This modeling strategy provides an average 2-month lead time for implementing control measures before the disease peak in March.
NG ET AL.
The outcome variable was regressed against each exposure variable one at a time using either a Poisson or a negative binomial regression model. For time-series exposure variables (Table 2), each monthly lag was tested individually in 24 separate univariate models. Person-days (population multiplied by the number of days in a month) were used as an offset variable to account for differences in the population density. Time-invariant and nonlagged variables associated with RRV disease cases at p < 0.10 were deemed candidate variables for multiple regression models. Timeseries monthly lags were analyzed separately, but their results were examined together. If one or more monthly lags were associated with RRV disease cases at p < 0.10, the time-series variable was deemed a candidate variable. This led to highly correlated co-variates in multiple regression models. To avoid fitting highly correlated co-variates, and to build a parsimonious model, a 4th degree polynomial distributed lag (PDL) approach was used (Supplementary Appendix A; Supplementary Data are available at www.liebertonline/ vbz/). This reduced the number of co-variates from 24 monthly lags to five polynomial terms for each candidate time-series variable. A large number of co-variates remained to be explored in multiple regression models, and a backward stepwise regression model with a p < 0.05 cutoff was used to remove candidate variables and their interaction terms from the final model. Variables that remained in the model were kept for the final model. Variables that were dropped were re-entered into the final model, one by one. Dropped variables were tested for their contribution to model fit individually (see next section). If improvement was made, the variable was set aside to be entered together with other dropped variables that improved model fit in a second backward stepwise regression with the initial group of retained variables. This process was repeated until no dropped variables improved model fit or until all variables were included. Model fit and cross-validation
As this was a count model, the predicted RRV count values could take any nonnegative value. For observed counts = 0, deviance was calculated as 2*predicted. For observed counts > 0, deviance was 2*(observed*ln(observed/predicted) (observed - predicted)). The deviance cutoff for a correct prediction was selected to reflect a range of values that could be considered a correct prediction (Table 3). An initial criterion for accepting a validated model was at least 75% prediction accuracy. A postestimation goodness-of-fit chisquared test for Poisson regression models was used to test the fit of the data to their respective distributions. An analogous postestimation test for negative binomial models was performed using the dispersion parameter alpha. Residual plots were also used to check that data followed model distribution. Finally, the Akaike information criterion (AIC) was used to compare the additional value of adding exposure variables into a series of nested models. Models were ranked according to their AIC statistic, with the lowest AIC indicating the best model. A jackknife approach was used to validate the final models; one postcode in each region was left out at a time with the remaining dataset used to fit the final model. The parameter
425
41 69,365 km2 344.3 meters 147,659 556 29.0 February (22.7%) 74.8% 3.19 31.1C 16.3C 15.8C 3.7C 1661.7 mm 232.3 mm 54.0 mm 585.3 mm 169.0 mm 141.3 mm
24 76,460 km2 87.1 meters 67,601 859 97.8 February (32.4%) 80.7% 3.41 31.5C 16.1C 15.8C 4.4C 1847.6 mm 269.9 mm 54.1 mm 375.1 mm 92.1 mm 103.9 mm
Region 2
27.4C 15.8C 15.8C 4.6C 1462.2 mm 173.8 mm 67.8 mm 876.7 mm 317.5 mm 142.6 mm
62 68,589 km2 388.3 meters 397,869 627 12.1 April (23.9%) 86.8% 1.79
Region 3
ARIA scores range from 0 to 1.84 (highly accessible), 1.85 to 3.51 (accessible), 3.52 to 5.80 (moderately accessible), 5.81 to 9.08 (remote) to 9.09 to 12 (very remote). RRV, Ross River virus.
a
Geography, demographics, and RRV disease Number of postcodes Total area Mean elevation Mean population (1991–2004) RRV disease notifications ( July 1991 to June 2004) Mean RRV disease incidence rate per 100,000 persons Peak notification month (% of all cases) Notifications occurring between January and June ARIA (Accessibility/Remoteness Index of Australia) scorea Climate and environment Mean maximum temperature (summer) Mean minimum temperature (summer) Mean maximum temperature (winter) Mean minimum temperature (winter) Mean annual evaporation Mean evaporation (summer) Mean evaporation (winter) Mean annual rainfall Mean rainfall (summer) Mean rainfall (winter)
Region 1
Table 1. Characteristics of Bioclimatic Regions
(continued)
27.6C 18.6C 19.7C 8.5C 1455.5 mm 165.6 mm 72.4 mm 1288.9 mm 447.5 mm 193.6 mm
44 30,245 km2 78.8 meters 461,789 2,288 38.1 April (24.0%) 89.1% 1.89
Region 4
426
2
Common wallaroos (2.2%) 29,973
November/February
November/January
Common wallaroos (not surveyed) 35,565
68.3%
78.9%
Eastern and western grey kangaroos (88.8%) Red kangaroos (9.0%)
A. camptorhychus (0.0%)
A. camptorhychus (0.06%)
Eastern and western grey kangaroos (75.1%) Red kangaroos (24.9%)
A. vigilax (1.5%)
A. vigilax (0.01%)
C. annulirostris (98.5%)
180.3 NDVI units (Sep) 65.0 NDVI units (May)
175.8 NDVI units (Sep) 53.5 NDVI units (May) C. annulirostris (99.93%)
31.3 days 4.8 days 64.8% Winter (6am) - 91.1% Summer (3pm) - 35.7% Temperate 119.5 NDVI units
21.2 days 1.9 days 60.8% Winter (6am) - 89.3% Summer (3pm) - 31.3% Dry grassland 107.7 NDVI units
Region 2
NDVI ranged from 0 being the lowest end of the spectrum (no vegetation) to 250 (dense vegetation).
RRV reservoir hosts Most common reservoir specie (% of host population) Second most common reservoir specie (% of host population) Third most common reservoir specie (% of host population) Mean macropod population per postcode
RRV vectors Most common vector specie (% of RRV vectors) Second most common vector specie (% of RRV vectors) Third most common vector specie (% of RRV vectors) RRV vectors as a proportion of trap total count Mosquito emergence month/Peak mosquito month
Climate and environment (continue) Number of rain days > 5mm per year Number of rain days > 25mm per year Mean annual humidity Mean maximum humidity Mean minimum humidity Dominant vegetation type Mean Normalized Differential Vegetation Index (NDVI)2 Mean maximum NDVI2 (month) Mean minimum NDVI2 (month)
Region 1
Table 1. (Continued)
Not surveyed
14,888
Eastern Grey kangaroos (not surveyed) Not found in region
Common wallaroos (not surveyed)
November/February
41.5%
A. camptorhychus (0.3%)
C. annulirostris (37.2%)
A. vigilax (62.5%)
172.2 NDVI units (Mar) 61.0 NDVI units (Nov)
58.5 days 13.4 days 70.0% Summer (6am) - 89.2% Winter (3pm) - 54.8% Subtropical 142.8 NDVI units
Region 4
Eastern and western grey kangaroos (62.36%) Common wallaroos (37.55%) Red kangaroos (0.09%)
November/February
46.2%
A. camptorhychus (0.2%)
C. annulirostris (41.7%)
A. vigilax (58.1%)
166.3 NDVI units (Mar) 92.3 NDVI units ( June)
45.5 days 7.9 days 70.1% Summer (6am) - 88.0% Winter (3pm) - 53.6% Temperate 138.5 NDVI units
Region 3
427
RRV disease notifications in NSW (outcome variable) NSW population by postcode (derived from CCDs, used as an offset variable) Mean total evaporation Mean relative humidity at 6 am Mean relative humidity at 3 pm Mean lowest daily maximum temperature Mean highest daily minimum temperature Mean number of rain days in a month – > 1 mm, 2 mm, 3 mm, 5 mm, 10 mm, 25 mm, and 50 mm Mean total rainfall Mean absolute maximum rainfall Mean temperature Mean maximum temperature Mean minimum temperature Mean absolute maximum temperature Mean absolute minimum temperature Mean sea surface temperature Mean, maximum and minimum tidal height above sea level
Description
July 1991 to June 2004 July 1991 to June 2004
C cm NOAA MHL
2004 2004 2004 2004 2004 2004 July 1991 to June 2004
June June June June June June C
to to to to to to
BOM
1991 1991 1991 1991 1991 1991
July July July July July July
mm mm C C C C
BOM BOM BOM BOM BOM BOM
July 1991 to June 2004
Count
BOM
2004 2004 2004 2004
July 1991 to June 2004
June June June June
C
to to to to
BOM
1991 1991 1991 1991
July July July July
July 1991 to June 2004
July1991 to June 2004
Time period
mm % % C
Count
Count
Unit
BOM BOM BOM BOM
ABS
NNDSS
Sourcea
(time-series) (time-series) (time-series) (time-series)
(time-series) (time-series) (time-series) (time-series) (time-series) (time-series)
Monthly (time-series) Monthly (time-series)
Monthly (time-series)
Monthly Monthly Monthly Monthly Monthly Monthly
Monthly (time-series)
Monthly (time-series)
Monthly Monthly Monthly Monthly
Annual
Annual
Temporal unit (variable type)b
Postcode Postcode
Postcode
Postcode Postcode Postcode Postcode Postcode Postcode
Postcode
Postcode
Postcode Postcode Postcode Postcode
Postcode
Postcode
Spatial unit
(continued)
a NNDSS, National Notifiable Disease Surveillance System; ABS, Australian Bureau of Statistics; BOM, Bureau of Meteorology; NOAA, Climate Diagnostic Centre, National Oceanic and Atmospheric Association; MHL, Manly Hydraulics Laboratory. b Time-series exposure variables, exposure variables that varied spatially and temporally for which monthly lags could be generated.
Sea surface temperature Tidal height
Climate data (surface)
Climate data (station)
Demographic data
Disease data
Category
Table 2. Summary of Variables Used for RRV Predictive Models
428
Mean, maximum and minimum vegetation (Normalized Differential Vegetation Index (NDVI)) Closest water source Number of water sources within 10 km Number of water sources within 20 km Number of water sources within 50 km Distance from the coast Mean postcode elevation Accessibility/Remoteness Index of Australia (ARIA) Mean total trap count Mean Ae. vigilax trap count Mean Ae. camptorhynchus trap count Mean Cx. annulirostris trap count Mean eastern and western grey kangaroo numbers (reporting of species was pooled) Mean red kangaroo numbers Mean common wallaroo numbers
Description
Count categorye Count categorye Count categorye CountCategorye Count Count Count
DECC DECC
km Count Count Count km Meters ARIA scored
GA GA GA GA GA GA GISCA ICPMR ICPMR ICPMR ICPMR DECC
NDVI unitc
Unit
ERIN
Sourcea
over over over over over over over
study study study study study study study
period period period period period period period
Jan 1992 to June 2004 Jan 1992 to June 2004
July 1991 to June 2004 July 1991 to June 2004 July 1991 to June 2004 July 1991 to June 2004 Jan 1992 to June 2004
Constant Constant Constant Constant Constant Constant Constant
Jan 1992 to June 2004
Time period
(nonlagged) (nonlagged) (nonlagged) (nonlagged) (time-series) Monthly (time-series) Monthly (time-series)
Monthly Monthly Monthly Monthly Monthly
(time-invariant) (time-invariant) (time-invariant) (time-invariant) (time-invariant) (time-invariant) (time-invariant)
Monthly (time-series)
Temporal unit (variable type)b
Postcode Postcode
Postcode Postcode Postcode Postcode Postcode
Postcode Postcode Postcode Postcode Postcode Postcode Postcode
Postcode
Spatial unit
a ERIN, Environmental Resources Information Network (ERIN) Department of the Environment, Water, Heritage and the Arts; GA, Geoscience Australia; GISCA, National Key Centre for the Social Applications of Geographic Information Systems; ICPMR, NSW Arbovirus Surveillance and Mosquito Monitoring Program, Institute of Clinical Pathology and Medical Research; DECC, Kangaroo Management Program, Department of Environment and Climate Change. b Time-series exposure variables, variables that varied over space and time for which monthly lags could be generated. Time-invariant exposure variables, variables that only varied over space (postcodes) but did not varied over time (remained constant over the study period), therefore monthly lags could not be generated. Nonlagged exposure variables, variables that did vary over space and time but due to missing data over non-monitoring months, polynomial terms could not be generated. c NDVI ranged from 0 being the lowest end of the spectrum (no vegetation) to 250 (dense vegetation). d ARIA scores range from 0–1.84 (highly accessible), 1.85–3.51 (accessible), 3.52–5.80 (moderately accessible), 5.81–9.08 (remote) to 9.09–12 (very remote). e Mosquito trap numbers were categorized as: 1 (0–99 mosquitoes), 2 (100–199) mosquitoes, 3 (200–299) mosquitoes, 4 (300–499 mosquitoes) or 5 ( ‡ 500 mosquitoes).
Reservoir host data
Socioeconomic/ rural indicator Mosquito data
Environment
Category
Table 2. (Continued)
ROSS RIVER VIRUS TRANSMISSION IN NEW SOUTH WALES, AUSTRALIA
429
FIG. 2. Monthly lagged variables tested for each time-series exposure variable and their corresponding months and year. RRV, Ross River virus.
estimates for the newly fitted model were then used to calculate the predicted counts in the excluded postcode. This was repeated until predicted count for each postcode was calculated. Overall accuracy, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) were calculated. Results
Region 1 is characterized by hot dry summers, cold wet winters, and dry grassland vegetation. The principal land use is agriculture and much of the region is irrigated. The main RRV vector is Cx. annulirostris, and the most common macropods are grey and red kangaroos, respectively (Table 1). There were 859 notifications; the 13-year mean notification rate was 97.8 per 100,000 persons. Three high notification years (years in which the number of disease cases exceeded the mean plus one standard deviation) were observed (1992/93, 1996/97, and 1998/99). The peak notification month was in February. The best-fitted Region 1 model (AIC 621.1502, goodness-of-fit chi-squared test p = 0.461) contained eight time-series exposure variables ( p < 0.05 for all) (Table 4 and Supplementary Appendix B). No timeinvariant or nonlagged exposure variables were retained in the final model. Region 2 is characterized by hot dry summers, cold winters, uniform rainfall throughout the year, and temperate vegetation. The principal land use is agriculture with rivers in the region diverted for irrigation. The dominant RRV vector is Cx. annulirostris and the most common macropods are grey kangaroos (Table 1). There were 556 notifications; the 13-year mean notification rate was 29.0 per 100,000 persons. Two high notification years were observed in 1995/96 and 1998/99 and the peak notification month was in February. The best Region 2 model (AIC 906.8076, goodness-of-fit chi-squared test p = 0.586) comprised eight time-series exposure variables ( p < 0.05 for all) (Table 4 and Supplemen-
tary Appendix C). No time-invariant or nonlagged exposure variables remained in the final model. Region 3 is characterized by warm wet summers, cool dry winters, and temperate vegetation. The principal land uses are agriculture, industry, and low-density urban settlement. The dominant RRV vectors are Ae. vigilax and Cx. annulirostris. The most common macropods are grey kangaroos and common wallaroos, respectively (Table 1). There were 627 notifications; the 13-year mean notification rate was 12.1 per 100,000 persons. High notification years occurred in 1996/7 and 2000/01; peak notifications were observed in April. The best-fitted Region 3 model (AIC 1165.910, goodness-of-fit chi-squared test p = 0.988) included seven time-series exposure variables ( p < 0.05 for all) (Table 4 and Supplementary Appendix D). Two time-invariant exposure variables were also retained; ARIA score (interrater reliability ratio [IRR] 1.59 per one score increase, p = 0.002) and number of water sources within 50 km (IRR 0.86 per increase by one in the number of water sources, p < 0.001). Region 4 is characterized by warm humid summers, mild winters, summer rainfall, and subtropical vegetation. The region is primarily urban and semiurban with little agriculture. The main RRV vectors are Ae. vigilax and Cx. annulirostris (Table 1). The main reservoir host species are the common wallaroo and the eastern grey kangaroo (although reservoir host data was not available for this region). There were 2288 RRV disease notifications reported; the 13-year mean notification rate was 38.1 per 100,000 persons. Three high notification years were observed (1995/96, 2002/03 and 2003/04) and peak notifications occurred in April. The best-fitted Region 4 model (AIC 1026.711, dispersion parameter, alpha = 0.11) retained six time-series exposure variables ( p < 0.05 for all) (Table 4 and Supplementary Appendix E). One nonlagged exposure variable was retained (Cx. annulirostris trap count in November, IRR 1.42 per categorical increase in mosquito trap numbers, p = 0.049). Interaction terms did not produce a lower AIC
430
statistic nor improve the model’s predictive accuracy for four models. Models were used to predict RRV cases from July to June each year. Because we used lags to 24 months, 11 years ( June 1 to July 30) were predicted over the 13-year period. Model accuracy ranged from 68.7% to 84.7% (Table 5). The mean accuracy by year across all regions was 78.8% (Table 5). The PPV ranged from 78.0% to 89.3% whereas the NPV ranged from 28.6% to 87.3% (Tables B1, C1, D1, E1 in Supplementary Appendices B–E). RRV disease cases in high notification years were predicted below the overall accuracy for all years in respective regions with the exception of one high notification year in Region 4 (Table 5). Sensitivity was between 60.4% and 73.1% whereas specificity was between 57.9% and 90.7% (Fig. 3). Region 3 and Region 4 models had higher sensitivities of 73.1% and 70.9%, respectively. Regions 1, 2, and 3 models gave specificity over 90% (90.2%, 90.7%, and 90.3%, respectively). Each regional model consisted of a unique combination of climate, environment, RRV vector, and reservoir host variables (Table 4). The effect of time-series exposure variables on RRV disease was distributed over the 2-year lag period with variations within and between regions (Supplementary Appendices B–E). All models retained at least one rainfall, one temperature, and one NDVI variable despite heterogeneity across regions. Grey kangaroo count remained in the three models where data were available. The inclusion of grey kangaroo data in the Regions 1 and 3 models improved model sensitivity by 2.0% and 2.8%, respectively, but reduced model sensitivity by 1.3% in the Region 2 model. Nonetheless, the delta AIC were - 7.489, - 13.395, and - 11.732 for the Regions 1, 2, and 3 models, respectively, compared to the models without host parameters indicating significant improvement on model fit with the inclusion of host parameters (Burnham and Anderson 2002). Discussion
This is the first study to use climate, environment, and RRV vector and reservoir host data to develop early warning predictive models for RRV disease in NSW, Australia. This is also the first study to include reservoir host data in statistical RRV models and show that the inclusion of host parameters improves model fit. Four regional models for bioclimatic areas were developed, each including different climate, environment, and RRV vector and reservoir host co-variates. Our study highlights the differences in RRV ecology and epidemiology between neighboring regions in NSW. This has also been observed in other multiregion studies (Tong and Hu 2002, Woodruff et al. 2002, Gatton et al. 2005, Williams et al. 2009). Polynomial distributed lag models were used to identify longer-term influences related to RRV activity. The effects of time-series exposure variables on RRV disease were distributed over the 2-year lag period, suggesting long-term influences from 2 years ago and from the previous year impact on RRV activity in the current year. Numerous RRV studies using moving average models for disease modeling have shown that the effects of climate and environment on RRV disease vary over time (Tong and Hu 2001, Tong et al. 2002, Hu et al. 2004, Tong et al. 2004, Hu et al. 2006a, McIver et al. 2010). Associations over 1–5 months lag are most commonly identified, as have associations in the pre-
NG ET AL.
vious year or season (Tong and Hu 2001, Tong and Hu 2002, Woodruff et al. 2002, Hu et al. 2004, Kelly-Hope et al. 2004c, Tong et al. 2004, Gatton et al. 2005, Tong et al. 2005, Hu et al. 2006a, Woodruff et al. 2006, Jacups et al. 2008b, Bi et al. 2009, Williams et al. 2009, McIver et al. 2010, Jacups et al. 2011, Werner et al. 2011), this is the first study to show that long-term effects of up to 2 years ago can influence RRV activity. Shorter-term associations (months in the current year; Fig. 2) are thought to be drivers of RRV transmission due to their influence on RRV vectors. The lag is attributed to the time required for optimal conditions to initiate widespread breeding in the mosquito population to reach a viable number for virus amplification. Our study identified positive shorterterm associations between rainfall, temperature, relative humidity, evaporation, tidal height, vegetation, and RRV disease that are indicative of conditions that promote mosquito breeding; other studies have found similar shorter-term associations (Tong et al. 1998, Tong and Hu 2001, Tong et al. 2002, Tong and Hu 2002, Woodruff et al. 2002, Hu et al. 2004, Kelly-Hope et al. 2004c, Tong et al. 2004, Gatton et al. 2005, Tong et al. 2005, Hu et al. 2006a, Woodruff et al. 2006, Jacups et al. 2008b, Bi et al. 2009, Williams et al. 2009, Hu et al. 2010a, McIver et al. 2010, Jacups et al. 2011, Werner et al. 2011). Conversely, we also observed negative shorter-term associations between rainfall, temperature, relative humidity, vegetation, and RRV disease that may reflect the threshold for optimal conditions; for example, an increase in relative humidity in dry and arid Region 1 in the current year was associated with an increase in RRV activity, whereas an increase in relative humidity in wet and humid Region 4 in the current year was associated with a decrease in RRV activity. Similar negative shorter-term associations have also been reported in other studies (Tong and Hu 2001, Tong et al. 2002, Tong and Hu 2002, Woodruff et al. 2002, Hu et al. 2004, Bi et al. 2009, Williams et al. 2009, Jacups et al. 2011). Long-term associations (months in the previous year and 2 years ago; Fig. 2) possibly reflect influences on RRV reservoir host breeding and distribution. Herd immunity in the reservoir host population is an important determinant of RRV transmission in humans (Carver et al. 2009a, Carver et al. 2010). Without sufficient nonimmune (susceptible) macropod juveniles, virus amplification would not occur, despite favorable conditions and an abundant vector population (Lindsay et al. 1996, Mackenzie 2001, Woodruff et al. 2002, Jacups et al. 2008b, Jacups et al. 2011). Consistent with this, seroprevalence among western grey kangaroos increased from < 25% before to 75% during an outbreak (C.J. Gordon, unpublished data, reported in Carver et al. 2009a, Carver et al. 2010). Outbreaks generally do not occur in successive years, possibly due to increased seroprevalence in reservoir hosts (Lindsay et al. 1996, Woodruff et al. 2002). NDVI and kangaroo populations in Regions 2 and 3, lagged 24 months, were correlated with RRV disease. Permanent pouch exit occurs 8–12 months after birth (Tyndale-Biscoe and Renfree 1987, Dawson 1995). A 2-year lag corresponds to approximately two generations of juveniles entering the population. Two or more generations of susceptible juveniles may therefore be necessary for an outbreak, although evidence from macropod serosurveys is needed to support this. Our models performed better than some other models but worse than others (Supplementary Appendix F). In part,
431
Table 3. Range of Predicted RRV Counts Considered Correct for Observed RRV Counts at a Residual Deviance Cutoff of Two
(continued)
432
Boxes marked with an ‘‘x’’ were considered incorrect predictions, while boxes marked with a ‘‘.’’ were considered correct predictions. Boxes marked with a ‘‘.’’ and in white indicate exact predictions where predicted counts were identical to observed counts.
Table 3. (Continued)
433
Goodness-of-fit chi-squared test p = 0.586
Akaike’s information criterion 1.49 (AIC) = 906.8076
Region 2
Goodness-of-fit chi-squared test p = 0.461
Akaike’s information criterion (AIC) = 621.1502
Region 1
Total rainfall (L2_Jan to L2_Mar [ - ]) SST (Nov to Dec [ + ]) Maximum temperature (L2_Mar to L1_Jul [ - ]) Total evaporation (L2_Apr to L2_Jun [ + ], L1_Jan to L1_Apr [ - ], Aug to Dec [ + ]) Relative humidity at 3pm (L1_Jul to L1_Nov [ - ], L1_Apr to Sep [ + ]) Mean NDVI ( Jul to Oct [ + ]) Maximum NDVI (L1_Jul to L1_Jan [ + ], L1_Jun to Nov [ - ]) Grey kangaroo count (Aug to Sep [ - ])
Refer to Appendix C for details on IRR and 95% confidence interval levels.
1. Total rainfall (Nov to Dec [ - ]) 2. Number of rain days > 1mm (L1_July to L1_Dec [ - ], Aug to Oct [ + ]) 3. Number of rain days > 5mm (L2_Jan [ + ], L2_Apr to L1_Jul [ - ], L1_Nov to L1_Feb [ + ], Jul to Oct [ - ], Dec [ + ]) 4. SST (L2_Jan to L2_Mar [ - ], L1_Aug to Sep [ + ], Dec [ - ]) 5. Absolute minimum temperature (L2_Jan to L2_Feb [ + ]) 6. Mean NDVI (L2_Jan to L2_Ap [ - ], L1_Jul to L1_Nov [ + ], Aug to Dec [ + ]) 7. Maximum NDVI (L2_Jan to L2_Mar [ + ], L2_Jun to L1_Dec [ - ], Nov to Dec [ - ]) 8. Grey kangaroo count (L2_Feb to L2_Apr [ + ], L1_Sep to L1_Dec [ - ])
Refer to Appendix B for details on IRR and 95% confidence interval levels.
1. 2. 3. 4. 5. 6. 7. 8.
Significant time-series exposure variables (p < 0.05) (Significant months* are in parentheses, direction of associations are in square brackets)
Table 4. Summary of Exposure Variables for the Four Regional Model
(continued)
No significant time-invariant or nonlagged exposure variables
No significant time-invariant or nonlagged exposure variables
Significant time-invariant or non-lagged exposure variables (p < 0.05)
434
Total rainfall (L2_Jan to L2_Feb [ - ], L2_Apr to L1_Nov [ + ], L1_Jun to Dec [ + ]) Number of rain days > 10mm (L2_Mar to L1_Sep [ - ], L1_May to Nov [ - ]) SST (L2_Jan to L2_Apr [ - ], L1_Sep to L1_Apr [ + ], Oct to Dec [ - ]) Absolute maximum temperature (L2_Jan to L2_May [ - ], L1_Sep to Jul [ + ]) Mean tidal height (L2_Jan to L2_Mar [ - ], L1_Jun to Nov [ + ]) Maximum NDVI (L2_Jan to L2_May [ + ], L1_May to Nov [ - ]) Grey kangaroo count (L2_Jan to L2_Mar [ + ], L2_May to L1_Oct [ - ], L1_May to Aug [ + ])
Refer to Appendix D for details on IRR and 95% confidence interval levels.
1. 2. 3. 4. 5.
Number of rain days > 10mm (L2_Jan to L2_Apr [ - ]) Maximum temperature (L2_Jan to L2_Mar [ - ], L1_Apr to Oct [ + ], Dec [ - ]) Minimum temperature (L2_Jan to L2_Mar [ - ], L2_Jun to L1_Dec [ - ]) Relative humidity at 6am (L2_Mar to L2_Jun [ - ], L1_Nov to L1_Feb [ + ], Jul to Oct [ - ]) Mean tidal height (L2_Jan to L2_Mar [ - ], L2_May to L1_Sep [ + ], L1_Dec to L1_May [ - ], Aug to Nov [ + ]) 6. Mean NDVI (Oct to Nov [ - ]).
Refer to Appendix D for details on IRR and 95% Confidence Interval levels
1. 2. 3. 4. 5. 6. 7.
1. Culex annulirostris trap count in November, IRR 1.42 per categorical increase in mosquito trap numbers, CI 1.00–2.00, p = 0.049)
1. ARIA score (IRR 1.59 per one score increase, CI 1.18–2.15, p = 0.002) 2. Number of water sources within 50km (IRR 0.86 per increase by one water source, CI 0.81–0.90, p < 0.001)
Significant time-invariant or non-lagged exposure variables (p < 0.05)
L2_Jan to L2_June: months from two years ago, L1_July to L1_June: months in the previous year, July to December: months in the current year (refer to Fig. 2 and corresponding Appendices for detail). IRR, interrater reliability ratio; SST, sea surface temperature; NDVI, Normalized Difference Vegetation Index; ARIA, Accesibility/Remoteness Index of Australia.
*
Goodness-of-fit dispersion parameter alpha = 0.11
Akaike’s information criterion (AIC) = 1026.711
Region 4
Goodness-of-fit chi-squared test p = 0.988
Akaike’s information criterion (AIC) = 1165.910
Region 3
Significant time-series exposure variables (p < 0.05) (Significant months* are in parentheses, direction of associations are in square brackets)
Table 4. (Continued)
ROSS RIVER VIRUS TRANSMISSION IN NEW SOUTH WALES, AUSTRALIA
Table 5. Overall Accuracy of the Four Regional Models by Year Yeara
Region 1
Region 2
Region 3
Region 4
1993/94 1994/95 1995/96 1996/97 1997/98 1998/99 1999/00 2000/01 2001/02 2002/03 2003/04 All years
91.7% 91.7% 62.5% 45.8%* 95.8% 70.8%* 79.2% 70.8% 87.5% 95.8% 100.0% 81.1%
82.9% 90.2% 63.4%* 78.0% 100.0% 65.9%* 73.2% 73.2% 92.7% 87.5% 73.2% 80.0%
93.2% 88.1% 78.0% 79.7%* 80.0% 91.9% 83.6% 75.0%* 88.5% 88.5% 84.2% 84.7%
57.1% 100.0% 54.2%* 84.6% 84.6% 58.6% 65.4% 72.7% 55.6% 64.0%* 68.8%* 68.7%
a 1 July to 30 June. *Denotes a high notification year, defined as a year in which the number of Ross River virus (RRV) disease cases exceeded the mean plus one standard deviation of all cases recorded in the region during the study period.
these findings may relate to variation in study area, covariates, outcome variables, methodologies, and spatial/ temporal unit of analysis between models. A validated model capable of predicting 70% or more of outbreaks is considered of value to public health professionals for early
435
warning purposes (Woodruff et al. 2002). Two of our models performed ‘‘satisfactorily’’ (Regions 3 and 4; sensitivity > 70%). A forecast of at least 2 months in advance is optimal in epidemic preparedness in public health (Myers et al. 2000), whereas the Regions 1 and 2 models only provided a 1-month lead time. The Regions 3 and 4 models provided a 3-month lead time, suggesting a 2-month lead time would have been possible for all regions had the data input cutoff been altered to reflect historical disease peaks. The models presented here demonstrate the potential for accurate and timely forecasting of RRV disease in NSW. Further research is needed to improve on predicting RRV disease in high notification years and in simplifying the modeling process for health authorities for models to be informative and effective as an operational tool. Notification data are prone to misclassification errors. However, these errors are likely random, not systematic, and the former are a lesser threat to validity than the latter. We used data collected for purposes other than ours and fitted these to our spatial and temporal unit of analysis. In assigning local data to large postcodes (weather stations and mosquito trap sites), assumptions were made that neighboring postcodes share similar climate and mosquito fauna, which may not be appropriate (Ryan et al. 2004). However, the best available data were used, and co-variates in the final models were correlated with RRV disease both in univariate and multivariate models. Confounders such as mosquito control,
FIG. 3. Sensitivity and specificity of the four regional models for all years. The ‘‘1’’ represents Ross River virus (RRV)count postcodes and ‘‘0’’ represents zero-count postcodes.
436
public awareness campaigns, irrigation practices, and other factors were not considered in the models. The impact of these in our study is unmeasured, but warrant further investigation. We investigated the use of signal predictors for RRV disease in four distinct bioclimatic regions in NSW, Australia. Each regional model consisted of a unique combination of predictors for RRV disease, highlighting the differences in the disease ecology and epidemiology in NSW. Vegetation and kangaroo populations lagged 24 months were correlated with RRV activity, and the inclusion of host parameters improved model fit significantly. Our findings suggest reservoir host data should be included in models. The models presented here demonstrate potential as a functional decision-making tool for public health practitioners. Acknowledgments
Funding was provided from an Australian National University PhD scholarship and a supplementary National Centre for Epidemiology and Population Health scholarship. We thank Dr. Rosalie Woodruff for her early contributions to this study. Author Disclosure Statement
No competing financial interests exist. References
Aaskov J, Doherty R. Arboviral zoonoses of Australasia. In: Beran G, Steele J, eds. Handbook of Zoonoses: Section B. Viral. Boca Raton, Florida: CRC Press, 1994:289–306. Bi P, Hiller J, Cameron A, Zhang Y, et al. Climate variability and Ross River virus infections in Riverland, South Australia, 1992–2004. Epidemiol Infect 2009; 19:1–8. Biggs A, Mottram P. Links between dryland salinity, mosquito vectors, and Ross River virus disease in southern inland Queensland–an example and potential implications. Australian J Soil Res 2008; 46:62–66. Bureau of Meterology. Climate classifications of Australia. Melbourne, Australia. 2005a. Bureau of Meterology. Thermal classification zone map construction. 2005b. Burnham K, Anderson D. Model Selection and Multimodel Inference: A Practical Information-Theoretic Approach. New York: Springer-Verlag, 2002. Carvan C, Russell R, Bryan J. Report of attempts to control the population of Culex annulirostris at Griffith, N.S.W during an outbreak of epidemic polyarthritis in the summer of 1983/84. Arbovir Res Aust 1986; 4:168–171. Carver S, Bestall A, Jardine A, Ostfeld R. Influence of hosts on the ecology of arboviral transmission: Potential mechanisms influencing dengue, Murray Valley encephalitis, and Ross River virus in Australia. Vector Borne Zoonotic Dis 2009a; 9:51–64. Carver S, Spafford H, Storey A, Weinstein P. Dryland salinity and the ecology of Ross River Virus: The ecological underpinnings of the potential for transmission. Vector Borne Zoonotic Dis 2009b; 9:611–622. Carver S, Kilpatrick A, Keunzi A, Douglass R, et al. Environmental monitoring to enhance comprehension and control of infectious diseases. J Environ Monitoring 2010; 12:2048– 2055.
NG ET AL.
Dawson T. Kangaroos—Biology of the Largest Marsupials. Sydney: University of New South Wales Press, 1995. Done S, Holbrook N, Beggs P. The quasi-biennial oscillation and Ross River virus incidence in Queensland, Australia. Int J Biometeorol 2002; 46:202–207. Gatton M, Kay B, Ryan P. Environmental predictors of Ross River virus disease outbreaks in Queensland, Australia. Am J Trop Med Hyg 2005; 72:792–799. Glass K. Ecological mechanisms that promote arbovirus survival: A mathematical model of Ross River virus transmission. Trans R Soc Trop Med Hyg 2005; 99:252–260. Harley D, Sleigh A, Ritchie S. Ross River virus transmission, infection, and disease: A cross-disciplinary review. Clin Microbiol Rev 2001; 14:909–932. Harley D, Bossingham D, Purdie D, Pandeya N, et al. Ross River virus disease in tropical Queensland: Evolution of rheumatic manifestations in an inception cohort followed for six months. Med J Aust 2002; 177:352–355. Harley D, Ritchie S, Bain C, Sleigh A. Risks for Ross River virus disease in tropical Australia. Int J Epidemiol 2005; 34:548–555. Hawkes R, Boughton C, Naim H, Stallman N. A major outbreak of epidemic polyarthritis in New South Wales during the summer of 1983–1984. Med J Aust 1985; 143: 330–333. Hu W, Nicholls N, Lindsay M, Dale P, et al. Development of a predictive model for Ross River virus disease in Brisbane, Australia. Am J Trop Med Hyg 2004; 71:129–137. Hu W, Tong S, Mengersen K, Oldenburg B. Rainfall, mosquito density and the transmission of Ross River virus: A time-series forecasting model. Ecolog Model 2006a; 196:505–514. Hu W, Tong S, Mengersen K, Oldenburg B, et al. Mosquito species (Diptera: Culicidae) and the transmission of Ross River virus in Brisbane, Australia. J Med Entomol 2006b; 42:375–381. Hu W, Clements A, Williams G, Tong S, et al. Bayesian spatiotemporal analysis of socio-ecologic drivers of Ross River virus transmission in Queensland, Australia. Am J Trop Med Hyg 2010a; 83:722–728. Hu W, Mengersen K, Dale P, Tong S. Difference in mosquito species (Diptera: Culicidae) and the transmission of Ross River virus between coastline and inland areas in Brisbane, Australia. Environ Entomol 2010b; 39:88–97. Jacups S, Whelan P, Currie B. Ross River virus and Barmah Forest virus infections: A review of history, ecology, and predictive models, with implications for tropical northern Australia. Vector Borne Zoonotic Dis 2008a; 8: 283–297. Jacups S, Whelan P, Markey P, Cleland S, et al. Predictive indicators for Ross River virus infection in the Darwin area of tropical northern Australia, using long-term mosquito trapping data. Trop Med Int Health 2008b; 13:943–952. Jacups S, Whelan P, Harley D. Arbovirus models to provide practical management tools for mosquito control and disease prevention in the Northern Territory, Australia. J Med Entomol 2011; 48:453–460. Jardine A, Lindsay M, Heyworth J, Weinstein P. Dry-season mosquito breeding associated with irrigation in the northeast Kimberley region of Western Australia: Potential impact on mosquito-borne disease transmission. EcoHealth 2004; 1:387–398. Kay B, Aaskov J. Ross River virus disease (epidemic polyarthritis). In: Monath T, ed. The Arboviruses: Epidemiology
ROSS RIVER VIRUS TRANSMISSION IN NEW SOUTH WALES, AUSTRALIA
and Ecology, volume IV. Boca Raton, Florida: CRC Press, 1989:93–112. Kelly-Hope L, Purdie D, Kay B. Differences in climatic factors between Ross River virus disease outbreak and non-outbreak years. J Med Entomol 2004a; 41:1116–1122. Kelly-Hope L, Purdie D, Kay B. El Nin˜o Southern Oscillation and Ross River virus outbreaks in Australia. Vector Borne Zoonotic Dis 2004b; 4:210–213. Kelly-Hope L, Purdie D, Kay B. Ross River virus disease in Australia, 1886–1998, with analysis of risk factors associated with outbreaks. J Med Entomol 2004c; 41:133–150. Lindsay M, Oliveira N, Jasinska E, Johansen C, et al. An outbreak of Ross River virus disease in southwestern Australia. Emerg Infect Dis 1996; 2:117–120. Mackenzie J. Ross River virus and Ross River virus disease. In: Service M, Ashford R, eds. Encyclopedia of ArthropodTransmitted Infections of Man and Domesticated Animals. Wallingford, UK: CABI Printing Series, 2001:443–447. Mackenzie J, Smith D. Mosquito-borne viruses and epidemic polyarthritis. Med J Aust 1996; 164:90–93. Mackenzie J, Lindsay M, Coelen R, Broom A, et al. Arboviruses causing human disease in the Australasian zoogeographic region. Arch Virol 1994; 136:447–467. Mackenzie J, Broom A, Hall R, Johansen C, et al. Arboviruses in the Australian region, 1990 to 1998. Commun Dis Intell 1998; 22:93–100. Maelzer D, Hales S, Weinstein P, Zalucki M, et al. El Nin˜o and arboviral disease prediction. Environ Health Perspect 1999; 107:817–818. McIver L, Xiao J, Lindsay M, Rowe T, et al. A climate-based early warning system to predict outbreaks of Ross River virus disease in the Broome region of Western Australia. Aust N Z J Public Health 2010; 34:89–90. Miller M, Roche P, Yohannes K, Spencer J, et al. Australia’s Notifiable Diseases Status, 2003. Annual Report of the National Notifiable Diseases Surveillance System. Commun Dis Intell 2005; 29:1–61. Myers, M., D. Rogers, J. Cox, A. Flahault, et al. Forecasting disease risk for increased epidemic preparedness in public health. Adv Parasitol 2000; 47:309–330. Mylonas A, Brown A, Carthew T, McGrath B, et al. Natural history of Ross River virus-induced epidemic polyarthritis. Med J Aust 2002; 177:356–361. National Notifiable Disease Surveillance System. National Notifiable Diseases Surveillance System, Communicable Disease Australia. 2013. Available at www9.health.gov.au/ cda/source/rpt_3.cfm Accessed September 1, 2013. Pople T, Grigg G. Biology of harvested species. In: Pople T, Grigg G, eds. Commercial Harvesting of Kangaroos in Australia. Canberra: Environment Australia, 1999. Ritchie S, Williams C, Montgomery B. Field evaluation of new mountain sandalwood mosquito sticks and new mountain sandalwood botanical repellent against mosquitoes in north Queensland, Australia. J Am Mosquito Control Assoc 2006; 22:158–160. Russell R. Ross River virus: Disease trends and vector ecology in Australia. Bull Soc Vector Ecol 1994; 19:73–81. Russell R. Vectors vs. humans in Australia—who is on top down under? An update on vector-borne diseases and research on vectors in Australia. J Vector Ecol 1998; 23:1–46. Russell R. Ross River virus: Ecology and distribution. Ann Rev Entomol 2002; 47:1–31. Ryan P, Lyons S, Alsemgeest D, Thomas P, et al. Spatial statistical analysis of adult mosquito (Diptera: Culicidae) counts:
437
An example using light trap data, in Redland Shire, Southeastern Queensland, Australia. J Med Entomol 2004; 41:1143–1156. Shepherd N. Condition and recruitment of kangaroos. In: Caughley G, Shepherd N, Short J, eds. Kangaroos: Their Ecology and Management in the Sheep Rangelands of Australia. Cambridge: Cambridge University Press, 1987:135– 158. Smith D, Hall R, Johansen C, Broom A, et al. Arbovirus Infections. In: Cook G, Manson P, Zumla A, eds. Manson’s Tropical Diseases. Elsevier Health Sciences, 2009:715–751. Tomerini D, Dale P, Sipe N. Does mosquito control have an effect on mosquito-borne disease? The case of Ross River virus disease and mosquito management in Queensland, Australia. J Am Mosq Control Assoc 2011; 27:39–44. Tong S, Hu W. Climate variation and incidence of Ross River virus in Cairns, Australia: A time-series analysis. Environ Health Persp 2001; 109:1271–1273. Tong S, Hu W. Different responses of Ross River virus to climate variability between coastline and inland cities in Queensland, Australia. Occup Environ Med 2002; 59:739–744. Tong S, Bi P, Parton K, Hobbs J, et al. Climate variability and transmission of epidemic polyarthritis. Lancet 1998; 351:1100. Tong S, Bi P, Donald K, McMichael A. Climate variability and Ross River virus transmission. J Epidemiol Community Health 2002; 56:617–621. Tong S, Hu W, McMichael A. Climate variability and Ross River virus transmission in Townsville Region, Australia, 1985–1996. Trop Med Int Health 2004; 9:298–304. Tong S, Hu W, Nicholls N, Dale P, et al. Climatic, high tide and vector variables and the transmission of Ross River virus. Int Med J 2005; 35:677–680. Tyndale-Biscoe C, Renfree M. Reproductive Physiology of Marsupials. Cambridge: Cambridge University Press, 1987. Vally J, Peel M, Dowse G, Cameron S, et al. Geogrphic Information Systems used to describe the link between the risk of Ross River virus infection and proximity to the Leschenault estuary, WA. Aust N Z J Public Health 2012; 36:229– 235. Werner A, Goater S, Carver S, Robertson G, et al. Environmental drivers of Ross River virus in southeastern Tasmania, Australia: Towards strengthening public health interventions. Epidemiol Infect 2011; 140:359–371. Westley-Wise V, Beard J, Sladden T, Dunn T, et al. Ross River virus infection on the north coast of New South Wales. Aust N Z J Public Health 1996; 20:87–92. Whelan P, Jacups S, Melville L, Groom A, et al. Rainfall and vector mosquito numbers as risk indicators for mosquitoborne disease in Central Australia. Commun Dis Intell 2003; 27:110–116. Williams C, Fricker S, Kokkinn M, Wakefield J. RR Forecaster: Tools for predicting mosquito and arbovirus activity Version 1.1. Adelaide, University of South Australia and South Australian Department of Health, 2007. Williams C, Fricker S, Kokkinn M. Environmental and entomological factors determining Ross River virus activity in the River Murray Valley of South Australia. Aust N Z J Public Health 2009; 33:284–288. Williams S, Williams C. South Australian Mosquito Intelligence Report No. 11. Adelaide, Mosquito and Public Health Research Group, University of South Australia, 2011. Woodruff R. Climate and Environment As Signal Predictors of Ross River VIrus Disease. Doctor of Philosophy, The Australian National University, 2003.
438
Woodruff R, Guest C, Garner M, Becker N, et al. Predicting Ross River virus epidemics from regional weather data. Epidemiology 2002; 13:384–393. Woodruff R, Guest C, Garner M, Becker N, et al. Early warning of Ross River virus epidemics: combining surveillance data on climate and mosquitoes. Epidemiology 2006; 17:569–575. Yohannes K, Roche P, Roberts A, Liu C, et al. Australia’s Notifiable Diseases Status, 2004, Annual Report of the National Notifiable Diseases Surveillance System. Commun Dis Intell 2006; 30:1–79.
NG ET AL.
Address correspondence to: Victoria Ng Centre for Public Health and Zoonoses (CPHAZ) Department of Population Medicine Ontario Veterinary College University of Guelph, Guelph Ontario, N1G 2W1 Canada E-mail:
[email protected]