Statistical Applications in Tropical Disease ResearchClimate-sensitive diseases and Early Warning Systems
Climate-sensitive diseases
Infectious diseases generally transmitted by insects(vector-borne), but can be food/water/air-borne
Particularly prevalent in developing countries
Climate directly impacts
Human behaviourDisease pathogenDisease vector
Michelle Stanton (Lancaster University) Statistical Applications in Tropical Disease Research November 2nd 2011 1 / 15
Statistical Applications in Tropical Disease ResearchClimate-sensitive diseases and Early Warning Systems
Climate-sensitive diseases
Infectious diseases generally transmitted by insects(vector-borne), but can be food/water/air-borne
Particularly prevalent in developing countries
Climate directly impacts
Human behaviourDisease pathogenDisease vector
Michelle Stanton (Lancaster University) Statistical Applications in Tropical Disease Research November 2nd 2011 2 / 15
Statistical Applications in Tropical Disease ResearchClimate-sensitive diseases
World Health Organization has identified 14 diseases for which climate can beused to inform disease predictions
Disease Transmission Climate-epidemic link
Influenza Air-bourne Decrease in temperatureCholera* Food and water-borne Increase in sea and air temperatureMalaria* Bite of female
Anopheles mosquito Changes in temp. and rainfallMeningococcal Air-borne Increases in temperature and decreasesMeningitis in humidityDengue Bite of female
Aedes mosquito High temp., humidity and rainfall
Using climate to predict infectious disease epidemics. World Health Organization (2005)
* Indicates that the relationship between climate and epidemics has beenquantified statistically.
Michelle Stanton (Lancaster University) Statistical Applications in Tropical Disease Research November 2nd 2011 3 / 15
Statistical Applications in Tropical Disease ResearchClimate-sensitive diseases
World Health Organization has identified 14 diseases for which climate can beused to inform disease predictions
Disease Transmission Climate-epidemic link
Influenza Air-bourne Decrease in temperatureCholera* Food and water-borne Increase in sea and air temperatureMalaria* Bite of female
Anopheles mosquito Changes in temp. and rainfallMeningococcal Air-borne Increases in temperature and decreasesMeningitis in humidityDengue Bite of female
Aedes mosquito High temp., humidity and rainfall
Using climate to predict infectious disease epidemics. World Health Organization (2005)
* Indicates that the relationship between climate and epidemics has beenquantified statistically.
Michelle Stanton (Lancaster University) Statistical Applications in Tropical Disease Research November 2nd 2011 4 / 15
Statistical Applications in Tropical Disease ResearchClimate-informed Early Warning Systems
In developing countries, the usual practice is to wait until an epidemic is underwaybefore implementing control measures.
EWS are intended to provide early identification of an epidemic.
Few operational EWS are in place in the health sector. However, due to:
advances in data availability (disease surveillance, GIS, remote sensing)
success of EWS outside of health sector
advances in statistical and epidemiological modelling
increasing awareness of climate change
the focus on EWS for epidemic diseases has increasedMichelle Stanton (Lancaster University) Statistical Applications in Tropical Disease Research November 2nd 2011 5 / 15
Statistical Applications in Tropical Disease ResearchClimate-informed Early Warning Systems
In developing countries, the usual practice is to wait until an epidemic is underwaybefore implementing control measures.
EWS are intended to provide early identification of an epidemic.
Few operational EWS are in place in the health sector. However, due to:
advances in data availability (disease surveillance, GIS, remote sensing)
success of EWS outside of health sector
advances in statistical and epidemiological modelling
increasing awareness of climate change
the focus on EWS for epidemic diseases has increasedMichelle Stanton (Lancaster University) Statistical Applications in Tropical Disease Research November 2nd 2011 6 / 15
Statistical Applications in Tropical Disease ResearchEWS considerations
EWS tend to be empirical rather than mechanisticImportant to consider the spatial and temporal scale of the EWS
Spatial: Often determined by surveillance data available. Generally spatiallyaggregated surveillance data, and station or gridded climate dataTemporal: Determined by the data and by the question we’re trying to answer(weekly, monthly?), but are generally short-term
Proportional influence of climate will differ at different spatial and temporalscales
Michelle Stanton (Lancaster University) Statistical Applications in Tropical Disease Research November 2nd 2011 7 / 15
Statistical Applications in Tropical Disease ResearchEWS considerations
EWS tend to be empirical rather than mechanisticImportant to consider the spatial and temporal scale of the EWS
Spatial: Often determined by surveillance data available. Generally spatiallyaggregated surveillance data, and station or gridded climate dataTemporal: Determined by the data and by the question we’re trying to answer(weekly, monthly?), but are generally short-term
Proportional influence of climate will differ at different spatial and temporalscales
Michelle Stanton (Lancaster University) Statistical Applications in Tropical Disease Research November 2nd 2011 8 / 15
Statistical Applications in Tropical Disease ResearchMeningitis Environmental Risk Information Technologies Project
Aim:
1 Improve the understanding of therelationship between meningitis andenvironmental/climate parameters
2 Use this knowledge to improve theefficacy of meningitis prevention andcontrol strategies
Michelle Stanton (Lancaster University) Statistical Applications in Tropical Disease Research November 2nd 2011 9 / 15
Statistical Applications in Tropical Disease ResearchMeningitis Environmental Risk Information Technologies Project
Case study: Niger
[0.548,0.897](0.897,1.33](1.33,1.97](1.97,2.65](2.65,6.55]
Average Incidence
010
2030
4050
Niger
Week
Inci
denc
e (p
er 1
00,0
00)
1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008
Climate Risk Factors
Temperature
Wind (U, V, speed)
Humidity
Dust Concentration
Other Risk Factors
December Incidence (inc.neighbours)
Population Density
Latitude
Michelle Stanton (Lancaster University) Statistical Applications in Tropical Disease Research November 2nd 2011 10 / 15
Statistical Applications in Tropical Disease ResearchMeningitis Environmental Risk Information Technologies Project
Case study: Niger
[0.548,0.897](0.897,1.33](1.33,1.97](1.97,2.65](2.65,6.55]
Average Incidence
010
2030
4050
Niger
Week
Inci
denc
e (p
er 1
00,0
00)
1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008
Climate Risk Factors
Temperature
Wind (U, V, speed)
Humidity
Dust Concentration
Other Risk Factors
December Incidence (inc.neighbours)
Population Density
Latitude
Michelle Stanton (Lancaster University) Statistical Applications in Tropical Disease Research November 2nd 2011 11 / 15
Statistical Applications in Tropical Disease ResearchEvaluating predictions
Considered fitting a model to Jan-May count data (dry season)
Used averaged climate variables prior to January as predictors
Fitted a negative binomial model to the data
No generally agreed criteria for assessing the accuracy of EWS.Aims to consider:
1 Predict the number of casesRMSE, R2
2 Evaluate whether or not a particular threshold will be exceededCalculate p = P(Incidence > 100 cases per 100, 000|Risk Factors)For each 0 < p < 1, calculate Sensitivity, Specificity, PPV, NPV
Observed EpidemicYes No
Predicted EpidemicYes TP FP PPV = TP
TP+FP
No FN TN NPV = TNFN+TN
Sensitivity Specificity= TP
TP+FN = TNFP+TN
Michelle Stanton (Lancaster University) Statistical Applications in Tropical Disease Research November 2nd 2011 12 / 15
Statistical Applications in Tropical Disease ResearchEvaluating predictions
Considered fitting a model to Jan-May count data (dry season)
Used averaged climate variables prior to January as predictors
Fitted a negative binomial model to the data
No generally agreed criteria for assessing the accuracy of EWS.Aims to consider:
1 Predict the number of casesRMSE, R2
2 Evaluate whether or not a particular threshold will be exceededCalculate p = P(Incidence > 100 cases per 100, 000|Risk Factors)For each 0 < p < 1, calculate Sensitivity, Specificity, PPV, NPV
Observed EpidemicYes No
Predicted EpidemicYes TP FP PPV = TP
TP+FP
No FN TN NPV = TNFN+TN
Sensitivity Specificity= TP
TP+FN = TNFP+TN
Michelle Stanton (Lancaster University) Statistical Applications in Tropical Disease Research November 2nd 2011 13 / 15
Statistical Applications in Tropical Disease ResearchSummary
At district-level, identified predictors were
Meridional (N-S) Wind, Wind Speed, Dust ConcentrationDecember Incidence (both in district, and average of neighbours)Population densityLatitude
Not all of the between-district variability explained by these variables
Model was better than a baseline model (persistence)
Improvements predominantly in identifying districts which exceeded theepidemic threshold (sensitivity)
Closing Remarks
Climate is unlikely to explain all of the spatio-temporal variability in a disease
The success of an operational EWS is not only the predictive skill of thesystem, but relies on engaging with decision-makers, and the efficiency ofcontrol measures
Michelle Stanton (Lancaster University) Statistical Applications in Tropical Disease Research November 2nd 2011 14 / 15
Statistical Applications in Tropical Disease ResearchSummary
At district-level, identified predictors were
Meridional (N-S) Wind, Wind Speed, Dust ConcentrationDecember Incidence (both in district, and average of neighbours)Population densityLatitude
Not all of the between-district variability explained by these variables
Model was better than a baseline model (persistence)
Improvements predominantly in identifying districts which exceeded theepidemic threshold (sensitivity)
Closing Remarks
Climate is unlikely to explain all of the spatio-temporal variability in a disease
The success of an operational EWS is not only the predictive skill of thesystem, but relies on engaging with decision-makers, and the efficiency ofcontrol measures
Michelle Stanton (Lancaster University) Statistical Applications in Tropical Disease Research November 2nd 2011 15 / 15