Upload
thomas-flynn
View
214
Download
2
Tags:
Embed Size (px)
Citation preview
1
Modelling malaria in Africa driven by Modelling malaria in Africa driven by DEMETER forecastsDEMETER forecasts
Anne Jones
Department of Geography
University of Liverpool
Liverpool
UK
NERC e-science studentship
Supervisor: Dr Andy Morse
ECMWF Forecast User Group Meeting, June 2006 [email protected]
2
OverviewOverview
• Climate and malaria
• Liverpool malaria model
• Predicting malaria in Botswana using DEMETER forecasts
• Discussion of results
• Model spin up
• Bias correction
• Timing issues
• Conclusions
ECMWF Forecast User Group Meeting, June 2006 [email protected]
3
Image from The Wellcome Trust
Anopheles Life Cycle and ClimateAnopheles Life Cycle and Climate
Anopheles breeding sites include cattle footprints, water tanks and rice fields
ECMWF Forecast User Group Meeting, June 2006 [email protected]
4
ECMWF Forecast User Group Meeting, June 2006 [email protected]
Image from The Wellcome Trust
Anopheles Life Cycle and ClimateAnopheles Life Cycle and Climate
The adult biting/laying cycle and survivorship depend on temperature
Gonotrophic cycle length according to Detinova (1962)
Mosquito survival according to Craig et al (1999)
5
Parasite Life Cycle and ClimateParasite Life Cycle and Climate
Temperature
18 °C
ECMWF Forecast User Group Meeting, June 2006 [email protected]
6
Structure of the Liverpool Malaria Model (LMM) Structure of the Liverpool Malaria Model (LMM)
Mosquito population
Malaria transmission -
mosquito
Malaria transmission -
human
Daily Malaria incidence (number of new cases) and prevalence (proportion of population infected)
10 day rainfall
daily temperature
daily temperature
humidity (10 day rainfall)
daily temperature
Input data: station or gridded datasets (ERA-40, DEMETER)
ECMWF Forecast User Group Meeting, June 2006 [email protected]
7
Botswana MalariaBotswana Malaria
MARA map
(Craig et al, 1999)
Tmin
Rain
Tmean
MARA run with ERA-40 1 deg (1958-2001)
MARA limiting variable (Tmin is min of 4 daily values)
ECMWF Forecast User Group Meeting, June 2006 [email protected]
8
Malaria Anomalies Malaria Anomalies
-2.0
-1.0
0.0
1.0
2.0
1982 1984 1986 1988 1990 1992 1994 1996 1998 2000
Year
Mal
aria
An
om
aly
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
Inci
den
ce A
no
mal
y
Malaria Index LMM incidence
Malaria index of Thomson et al. (2005) and ERA-40-driven LMM yearly total incidence.
Thomson et al. found a quadratic relationship between malaria and rainfall
9
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
1 31 61 91 121 151
Forecast Day
Ma
lari
a P
rev
ale
nc
eSeasonal ForecastsSeasonal Forecasts
• Assessed performance of forecasts made using LMM driven by DEMETER forecasts.
• Compared to ERA-40 driven forecasts (tier-2 validation) and yearly malaria anomalies (tier-3 validation)
• DEMETER daily rainfall and temperature series were corrected for model biases then used as input to the malaria model.
Botswana malaria forecast for February 1989, LMM driven by DEMETER multi-model
(ERA-driven model shown in red)
95
85
65
35
15
5ERA
ECMWF Forecast User Group Meeting, June 2006 [email protected]
10
Tier-3 ROC AreasTier-3 ROC Areasfor November malaria forecastfor November malaria forecast
ROC Area (<0.5 = no skill), Upper Tercile event, forecast 6 month totals for Botswana grid average, () 95% confidence intervals calculated from 1000 bootstrap samples
Validated against Thomson et al (2005) Malaria Index
LMM Input Data ROC Area
ERA-40 0.78 (0.56-0.97)
Raw DEMETER 0.31(0.11-0.56)
Bias-corrected DEMETER
T correction only
0.67 (0.29-0.80)
0.70 (0.35-0.90)
ERA-40 "persistence"
(wrong year ensemble)0.32
(0.07-0.63)
11
15
17
19
21
23
25
27
29
1 31 61 91 121 151
Forecast Day
Me
an
T (
de
gre
es
C)
Effect of temperature bias correctionEffect of temperature bias correction
• Temperature variability not a strong driver of malaria variability in this region
• However malaria model requires realistic temperatures
• DEMETER temperatures need to be bias corrected to achieve this, because models is sensitive to biases in uncorrected data of ~ 2 degrees
DEMETER temperature forecasts for Botswana, November 1997
95
85
65
35
15
5ERA
Uncorrected Temperature Corrected Temperature
15
17
19
21
23
25
27
29
1 31 61 91 121 151
Forecast Day
Me
an
T (
de
gre
es
C)
ECMWF Forecast User Group Meeting, June 2006 [email protected]
12
ECMWF Forecast User Group Meeting, June 2006 [email protected]
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
180 30 60 90 120 150
Forecast Day
Ma
lari
a p
rev
ale
nc
e
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
180 30 60 90 120 150
Forecast Day
Ma
lari
a p
rev
ale
nc
e
• Improvement in skill due to temperature correction
• If temperatures too low, delay in model is increased
Uncorrected input
Corrected temperature
DEMETER-driven malaria forecasts for November 1997
95
85
65
35
15
5ERA
Effect of temperature bias correction contd.Effect of temperature bias correction contd.
13
Tier-1 ROC Areas for November rainfall forecast
Event Uncorrected Corrected
Lower Tercile 0.88(0.67-1.0)
0.75(0.45-1.0)
Above the median
0.65(0.37-0.89)
0.68(0.39-0.92)
Upper Tercile 0.72(0.46-0.93)
0.63(0.37-0.86)
ROC Area (<0.5 = no skill),forecast 6 month totals for Botswana grid average, () 95% confidence intervals calculated from 1000 bootstrap samples
Validated against ERA-40 rainfall totals
Effect of rainfall bias correction Effect of rainfall bias correction
•Bias correction of rainfall causes decrease in skill
ECMWF Forecast User Group Meeting, June 2006 [email protected]
14
• Correct for frequency and intensity separately
• If distributions are very different, large numbers of rainfall days are removed
• Correction to mean climate but reduction in skill - need alternative method
Effect of rainfall bias correction contd.Effect of rainfall bias correction contd.
model crfc, daily rainfall over 20 years for Feb at 25E, 22.5S
= ERA-40, = uncorr. DEM, = corr. DEM
ERA-40 Crfc model
November rainfall totals in mm for all grid points in Botswana over 20 years
Daily rainfall (mm)
Cum
ulat
ive
freq
uenc
y
ECMWF Forecast User Group Meeting, June 2006 [email protected]
15
Timing IssuesTiming Issues
Botswana grid-averaged values for November 1988 forecast
Bias corrected data
Rainfall LMM Mosquitoes LMM Malaria
• Bad/lack of bias correction means rainfall can be too low
• Model peak can be outside the forecast window due to lag between rainfall and malaria cases (made worse when rainfall too low)
• Skill not improved by extending forecast window with previous year of ERA data
16
Tier-3 ROC AreasTier-3 ROC Areas--alternative model outputsalternative model outputs
LMM Input Data LMM Malaria Anomalies
LMM Mosquito Anomalies
ERA-40 0.78 (0.56-0.97)
0.89 (0.71-1.0)
Raw DEMETER 0.31(0.11-0.56)
0.73 (0.42-0.97)
Bias corrected DEMETER
0.67 (0.29-0.80)
0.56
(0.42-0.91)
ROC Area (<0.5 = no skill), Upper Tercile event, November forecast 6 month totals for Botswana grid average, () 95% confidence intervals calculated from 1000 bootstrap samples
Validated against Thomson et al (2005) Malaria Index
• Skill improved by using model mosquito numbers
• Bias correction decreases skill due to strong rainfall driver
• Cannot use in other areas where temperature a stronger driver
17
ConclusionsConclusions
• DEMETER-driven forecasts were skilful, better than climatology and persistence forecasts
• Bias correction of temperature is important even if variability in temperature not important - temperatures must be "realistic" for the application model
• Bias correction of rainfall is unsatisfactory - use of daily rainfall output is problematic and need to consider other methods using monthly anomalies instead (e.g. weather generator)
• Lag in model mean malaria cases may occur outside forecast window - can be solved for Botswana using mosquito model but not applicable to other areas
ECMWF Forecast User Group Meeting, June 2006 [email protected]
18
Thankyou
© L Anderson-Ptito, RBM Partnership Secretariat
ECMWF Forecast User Group Meeting, June 2006 [email protected]