Toward the Development of a
Drought Hazard Index:
Methods and Initial Results
Emily K. Grover-Kopec
International Research Institute for Climate Prediction
Maxx Dilley, UNDP
Bradfield Lyon, IRI
Régina Below, CRED
5th EM-DAT Technical Advisory Group Meeting
August 18-19, 2005
Background
 Initial analysis of relationship between
hydro-meteorological drought hazards
and drought disasters highlighted need
to review EM-DAT documentation
methods
 CRED and IRI develop joint project to:
1. Improve documentation of drought disasters
in EM-DAT
2. Advance the understanding of how drought
hazards associated with drought disasters
Characterizing the Hazard
Temporal and spatial nature of the hazard
make it difficult to define
Use drought impacts as ground-truth for
definition
Develop hazard index for characterizing
magnitude, duration, timing and location
Relating Hazards to Disasters
Drought Disasters
Societal
Vulnerability
PROXY
EM-DAT
Drought
Hazards
PROXY
Meteorological
Agricultural
Hydrological
Hazard Indices
Drought Disasters in EM-DAT
 360 disaster events
40
30
20
10
1998
1991
1984
1977
1970
1963
1956
1949
1942
1935
1928
1921
1914
1907
0
1900
Number of events
Number of Drought Disaster
Events in EM-DAT
Year
~1979 - 2004
Hazard data
availability
Consistent
disaster data
Africa
47%
Asia
22%
Europe
9%
South America
8%
Central America
8%
Australia/Oceania
3%
North America
3%
Drought Hazard Indicators
Meteorological
– SPI (Standardized Pcpn Index)
– WASP (Weighted Anomaly Standardized Pcpn)
Agricultural
– NDVI (Normalized Difference Vegetation Index)
– Soil Moisture
– PDSI (Palmer Drought Severity Index)
– WRSI (Water Requirement Satisfaction Index)
Drought Hazard Indicators
NDVI
SPI (3-month)
PDSI
GMSM
WRSI
WASP (3-month)
Drought Hazard Indicators
SPI
WASP
NDVI
WRSI
GMSM
PDSI
Indicators are a function of:
1. Time Scale
2. Time Lag
3. Threshold
Example: 3-Month SPI < -1.0 ; 0-4 months before disaster event
Converting Spatially-Continuous
Data to Country-Level Data
-------------------------------------------------------------------------------------
Hazard data = F(X,Y,T)
Disaster data = F(Country,T)
Problematic issues with taking a simple
average of data for each country
1. Average of large country generally not
representative of disaster event in EM-DAT
2. Relatively wet and dry regions in same
country can mute drought hazard signal
Problem 1: Average of Large Countries
Not Representative of Hazard
Apply land classification mask to remove
areas neither inhabited or used for
agriculture
Application of Land Use Mask
Problem 2: Simultaneous Wet and Dry
Areas Within a Country
Apply dry mask to remove all anomalously
wet areas
Spatially-Continuous Data
Converted to Country-Level Data
Applying land classification and dry masks to the data
and then averaging the result over national boundaries
generates hazard data that can be compared to the
point disaster data
-------------------------------------------------------------------------------------
Hazard
Hazarddata
data==F(Country,T)
F(X,Y,T)
-------------------------------------------------------------------------------------
Disaster data = F(Country,T)
Analysis Options: Not Regression
 Hazard indicators highly
correlated
 Autocorrelation present in
indicators with time scale
greater than 1 month
Regression is not an
appropriate analysis
technique
Hazard Indicators for Botswana (1979-2004)
JFMAMJJASOND
SPI
WASP
SM
PDSI
2
1.5
3
2
JFMAMJJASOND
1
1
0.5
0
-1
-1
-1.5
-2
Indicator time series with
3-Month time scale
Fe
b00
Fe
b02
Fe
b98
Fe
b94
Fe
b96
Fe
b92
Fe
b88
Fe
b90
Fe
b86
-0.5
Fe
b84
Fe
b80
Fe
b82
JFMAMJJASOND
Fe
b04
0
-2
-3
-4
-5
Analysis Options
 Condense hazard and disaster data to
binary, country-level indicators and then
use:
1. Contingency table statistics and skill scores
 Ongoing
2. Principle Component Analysis
 Planned
Creating the Contingency Tables
DISASTER OCCURS
YES
HAZARD
INDEX
DEFINITION
MET
NO
YES
a
b
NO
c
d
Creating the Contingency Tables
START
Country-level average
of masked data H(Ti)
Is H ≤ Thd?
Repeat
for Ti+1, n
and all countries
NO
YES
HB=0
Does disaster occur
in same country within
L months of Ti?
YES
HB=1
NO
c
d
Does disaster occur
in same country within
L months of Ti?
YES
NO
a
b
Result: Table for each combination of hazard, time scale, threshold and lag
Creating the Contingency Tables
Example: 6-Month WASP, Threshold=-1.25,
Lag=4 months
6-Month WASP Data
EM-DAT
Disaster
Afghanistan X
x x x x x…
Albania Jun x1979
x x x x…
.
.
Is value less than or
.
equal to -1.25?
.
.
.
Zimbabwe x x x x x …
Does a disaster start in
Check EM-DAT
for
Afghanistan
during
corresponding
disaster
Jun-Oct 1979?
YES
NO
Y
N
Hazard
Index Y
a
bb
N
c
d
Creating the Contingency Tables
Example: 6-Month WASP, Threshold=-1.25,
Lag=0-4 months
6-Month WASP Data
Afghanistan x X
x x x x…
Albania
x 1979
x x x x…
Jul
.
.
Is value less than or
.
equal to -1.25?
.
.
.
X
Zimbabwe x x x x x …
YES
EM-DAT
Disaster
Does a disaster start in
Check EM-DAT
for
Afghanistan
during
corresponding
disaster
Jul-Nov 1979?
YES
Y
N
Hazard
Index Y
aa
b
N
c
d
Contingency table for
DHI = [WASP6, Thd=-1.25,
Lag=0-4 Months]
Making Sense of It All
Statistics can be used to characterize
each hazard indicator’s table in terms of
how well it “predicts” disasters
Let these statistics tell us which is/are the
best indicator(s)
Contingency Table Statistics
3-Month WASP Skills at Multiple Thresholds
1
0.9
0.8
0.7
0.6
0.5
0.4
HR
TS
POD
FAR
HSS
0.3
0.2
0.1
0
T=-0.75
T=-1
T=-1.25
T=-1.5
Threshold
T=-1.75
T=-2
SPI and WASP
Skill Score SPI vs. WASP at Multiple Time Scales
WASP1
WASP3
WASP6
WASP9
WASP12
SPI1
SPI3
SPI6
SPI9
SPI12
0.05
0.04
HSS
0.03
0.02
0.01
0
T=-0.75
T=-1
T=-1.25
T=-1.5
Threshold
T=-1.75
T=-2
Initial Results
 WASP appears to have closer relationship with
disasters at all but shortest time scales
– Seasonality important
 For these meteorological indices:
– Time scale ~ 3-6 months
– Country-wide threshold ~ -1.0 (moderate drought)
 Contingency tables/stats
– Will be able to say more about contingency table
results after significance testing
– Additional motivation for using additional statistical
methods
Next Steps
Continue contingency table analysis for
remaining hazard indicators
Perform additional statistical methods
– PCA  Provide a series of independent,
weighted sums of the indicators which
maximize the amount of explained variance in
the disaster data
Next Steps
Apply above information to formulations of
single Drought Hazard Index (DHI)
– Most likely a weighted combination of
indicators, but may be a single indicator
Make DHI available via the IRI Data
Library Maproom
Potential applications of DHI in EWS and
methodology in regional/country-level case
studies
Principle Component Analysis
Basics
 Standarizing indicators gives equal weight to all.
Otherwise indicators with higher variance have more
weight.
 Combine indicators so those that are describing similar
aspects are described in a single metric
 Each combination (principal component):
– Measures different aspect of disaster behavior and is completely
uncorrelated with the others
– Has high variance (i.e., summarizes as much information as
possible)
– Are weighted sums of original indicators
Contingency Table Statistics
•HR = (a+d)/n
•TS = a/(a+b+c)
•POD = a/(a+c)
•FAR = b/(a+b)
•HSS = (ad-bc)/(a+c)(b+d)