Probabilistic prediction of climatesensitive diseases using multimodel seasonal forecasts

1

Probabilistic prediction of climate-sensitive
diseases using multi-model seasonal
forecasts Francisco J. Doblas-Reyes f.doblas-reye
s_at_ecmwf.int European Centre for Medium-Range
Weather Forecasts With significant contributions
from Renate Hagedorn (ECMWF), Simon Mason (IRI),
Andy Morse (Univ. of Liverpool), Tim Palmer
(ECMWF), Madeleine Thomson (IRI), Antje
Weisheimer (ECMWF)
2
Seasonal forecast objective
To utilize the ability to predict climate
variability on the scale of months to a year to
improve management and decision making in respect
to several socio-economic applications at local,
regional, and national scales.

• Requirements
• predict climate variability deal with
  uncertainties in climate prediction and assess
  skill
• seasonal-to-interannual time scales use coupled
  ocean-atmosphere general circulation models and
  empirical/statistical methods
• variable spatial scale downscaling

3
A user strategy the end-to-end approach

• A broad range of forecast products might be
  offered, but a specific analysis of the user
  requirements is necessary.
• End-to-end is based on collaboration and
  continuous feedback.
• Users develop their models taking into account
  climate prediction limitations.
• Users employ objective records of performance.
• The final level of forecast quality that provides
  added value is defined by the application -gt
  user-oriented verification. Users assess the
  final value of the predictions.
• Forecast reliability becomes a major issue.

4
What do we try to predict
Other phenomena yielding seasonal and interannual
predictability are tropical ocean variability,
soil moisture processes or snow and sea/ice
variability
5
How do we predict Dynamical prediction
Primitive equations (7 equations with 7
variables)
6
Sources of forecast uncertainty

• Initial conditions Lack of perfect knowledge of
  the initial conditions (for both the atmosphere
  and the ocean) due to inadequacy of the
  observational network, limitations of satellite
  data and errors in assimilation methods.
• Model error Forecast models are imperfect, have
  a coarse resolution and make strong use of
  parametrizations.
• Boundary forcing uncertainty The detailed
  composition of the atmosphere during the forecast
  is unknown -gt link to human-induced global change.

7
Observations for initial conditions
In situ observations
synop-ship
14956 buoy

3499 temp

580 pilot

949 plane
33634
Polar orbit satellites
Geostationary satellites
ATOVS, AIRS, SCAT, SSMI, Ozone
408751
14/03/2004 00UTC gt 1M obs
8
Sensitivity to initial conditions Ensembles
9
Ensemble prediction system

• Example Nine 6-month long simulations with
  slightly different initial conditions

Feb 87 May 87 Aug 87
Nov 87 Feb 88 ...
10
Model error

• Model error is a consequence of the
  approximations used in a forecast system, it is
  measured as the difference in the statistical
  properties of the simulations and an
  observational reference and has as consequences
• Systematic error in mean and climate variability
• Lack of forecast reliability
• Model equations are an average function of the
  actual situation of the atmosphere
• It is impossible to increase arbitrarily model
  resolution, so parameterizations will always have
  drawbacks

11
Model error sources

• Unfortunately, climate models are far from
  perfect because of the numerical approximations
  and the parameterizations

Satellite image of shallow convection
12
Model error and forecasting

• Predictability is a measure of the maximum
  prediction accuracy achievable for a physical
  system, while actual accuracy is an attribute of
  the forecast quality of a forecast system.
  Accuracy suffers from model error so that it
  underestimates predictability.
• Three ad-hoc methods have been devised to address
  the problem of model error in medium-range, s2d
  and climate change forecasting
• Multi-model ensemble
• Perturbed parameters
• Stochastic physics

13
System addressing model error

• Multi-model example Nine 6-month long
  simulations with different forecast systems

Feb 87 May 87 Aug 87
Nov 87 Feb 88 ...
14
End-to-end DEMETER
http//www.ecmwf.int/research/demeter/

• Research project funded by the Vth FP of the EC,
  with 11 partners.
• Integrated multi-model ensemble prediction system
  for seasonal time scales.
• More than a multi-model exercise seasonal
  hindcasts used to assess the skill, reliability
  and value of end-user predictions.
• Applications in crop yield and tropical
  infectious disease forecasting.
• Officially finished in September 2003, but with
  an operational follow up.

15
Multi-model ensemble approach
16
Multi-model ensemble system

• DEMETER system
• 7 coupled global circulation models

9 member ensembles ERA-40 initial conditions
SST and wind perturbations 4 start dates per
year 6 months hindcasts

• Hindcast production for 1980-2001 (1958-2001)

17
Multi-model ensemble system

• DEMETER system 7 coupled global circulation
  models

CNRM (FR) ECMWF (INT) INGV (IT) LODYC (FR) MPI
(DE) UKMO (UK) CERFACS (FR)
7 models x 9 ensemble members 63 member
multi-model ensemble
Feb 87 May 87 Aug 87
Nov 87 Feb 88 ...
18
Multi-model ensemble system

• DEMETER system 7 coupled global circulation
  models

CNRM (FR) ECMWF (INT) INGV (IT) LODYC (FR) MPI
(DE) UKMO (UK) CERFACS (FR)
Feb 87 May 87 Aug 87
Nov 87 Feb 88 ...
19
Multi-model ensemble system

• DEMETER system 7 coupled global circulation
  models

CNRM (FR) ECMWF (INT) INGV (IT) LODYC (FR) MPI
(DE) UKMO (UK) CERFACS (FR)
Feb 87 May 87 Aug 87
Nov 87 Feb 88 ...
20
Multi-model ensemble system

• DEMETER system 7 coupled global circulation
  models

CNRM (FR) ECMWF (INT) INGV (IT) LODYC (FR) MPI
(DE) UKMO (UK) CERFACS (FR)
63 member multi-model ensemble 1 hindcast
Feb 87 May 87 Aug 87
Nov 87 Feb 88 ...
21
ENSO predictions
Multi-model seasonal (MAM) predictions, Niño3.4
SSTs
22
Forecast quality assessment
Forecast quality assessment is a basic component
of the prediction process
Information about the quality and the uncertainty
of the predictions is as important as the
prediction itself
23
Verification of probabilistic forecasts
24
Predicting for users end-to-end
Climate forecast

62
4
3
2
1
63
25
Downscaling for s2d predictions
http//www.ecmwf.int/research/EU-projects/ENSEMBLE
S/news/index.html
26
Climate and infectious diseases
From Using Climate to Predict Infectious Disease
Outbreaks A Review, WHO 2004
27
Malaria in Botswana

• Increases in rainfall increase breeding site
  availability and, therefore, malaria vector
  populations. Rainfall is also associated with
  humidity increases, resulting in higher vector
  survivorship and a greater probability of
  transmission.
• Constraints to study impact on malaria incidence
• Deficient surveillance (but a notifiable disease
  in Botswana)
• Lack of confirmed case data (laboratory-confirmed
  cases are recorded in Botswana)
• Short time series for analysis (Botswana has
  annual records from 1982)
• Many confounding factors (dates of changes in
  drug policies in Botswana are known)

28
Malaria in Botswana
Standardized log malaria incidence and monthly
total precipitation. The disease follows the
maximum of the rainy season with a delay of a few
weeks
From S. Mason and M. Thomson (IRI)
29
Malaria early warning systems
gathering evidence for early and focused response
geographic/community focus
case surveillance alone late warning
From M. Thomson (IRI)
30
Seasonal forecasts for malaria warning
Precipitation composites for the five years with
the highest (top row) and lowest (bottom row)
standardised malaria incidence for DJF DEMETER
(left) and CMAP (right)
Quartiles define extreme events (outbreaks) for
malaria prediction
31
Seasonal prediction of malaria risk
DJF CMAP precipitation vs Botswana standardised
log malaria incidence for 1982-2002
32
Malaria warning with climate information
Probabilistic predictions of standardised malaria
incidence quartile categories in Botswana with
five months lead time
33
Operational prediction for malaria
Operational seasonal predictions of OND 2006
precipitation (August start date, 2-month lead
time) presented at the latest Southern Africa
Climate Outlook Forum and used at the Malaria
Outlook Forum (September). Skill is measured as
Spearmans rank order correlation of the ensemble
mean
From S. Mason (IRI)
34
Prediction for other diseases cholera
Seasonal prediction of precipitation over the
Indian subcontinent might be used to predict
cholera outbreaks
All Indian Rainfall Corr 0.39
V-shear index Corr 0.54
U-shear index Corr 0.74
35
Seamless prediction systems an example

• Users might optimally benefit from forecasts in
  different time scales -gt seamless system
• ECMWF probabilistic seamless forecast system
• 1-10 days medium range EPS
• 10 days-1 month monthly forecast system
• 1 month-12 months seasonal forecast system

12mth
1mth
10d
01/01
01/02
01/03
15/01
29/01
12/02
26/02
36
Seamless prediction seasonal forecasts and
climate change

• Users adapt to climate change by reacting to
  climate variability, predicted with s2d forecast
  systems.
• Changes in precipitation are presumably due to
• Increased atmospheric water vapor content
• Changes in circulation
• Can seasonal forecasts be used to estimate how
  reliable climate change estimates are in a
  seamless framework? Use two multi-model systems
• DEMETER seasonal forecasts
• IPCC AR4 climate change simulations

37
Summary

• Probabilistic seasonal forecasts based on
  multi-model ensembles have shown useful skill for
  several variables and regions, especially in the
  tropics.
• Seasonal forecasts are a useful contribution for
  the control of malaria outbreaks several months
  in advance in Botswana.
• The methodology can be extended to other
  infectious diseases.
• Seasonal forecasts are a useful tool for the
  analysis of climate change predictions and for
  the timely adaptation of vulnerable groups.

38
Questions?