I n s t r u m e n t s

a n d

M e t h o d s

taken twice a day (9:00 am and 1:00 pm) as well as various
othermeasurements.
So we can perform quite significant testsfrom thisin-
formation.
Oursample consists of123days with 140avalanches
for1111releaseattempts.Obviously,weselecteddays
without significant snowfall to eliminate statistical noise.

2.2.PROCEDURE

Wecalculated, for eachday from the sample, an avalanche
activity index as following:

Acti = nAval / nEvent
where
nAvalis the number of observed avalanches(arti-
ficially released ornot),
nEvent is the sum: (number of release attempts) +
(numberof natural avalanches).
For example, duringMarch7th 1995, 7 avalanches were
observed for9 release attempts and no natural avalanche
occurred: theavalancheactivity index was 0.78; the snow-
drift index was 30 g, the wind speed was nil, andthe wind
speed the day before was 15 knots.
We completed charts to visualise the statistical distri-
butions of avalanche activity accordingto wind speed and
snowdrift (fig.6).
Thechartsconfirmsthatavalancheactivitystrongly
depends onsnowdriftwhich seemstobemore informa-
tive than wind.Now let us quantify this intuitive deduc-
tion and let us turn to an index of association in the pre-
dictive sense.
Let us assume that oursample islarge enough to con-
sider that frequenciesand probabilities of events are equal
and let us describe the distributions bythe mean of joint
probabilitydistribution tables (fig.7).
Let us now characterise the relationship betweensnow-
drift and avalancheactivity.The index of predictive asso-
ciation developed by Goodman and Kruskal is:
lAS = [P(e) - P(e|S)] /P(e)[1]
where
.S is the snowdrift index given by the driftometer,
.P(e) is the lowest probability of an errorin avalanche
activity classification when S is unknown;
P(e)=1-max [!] P
(Pij: probability^ij
from cell [Si,Aj])[2]

.P(e| S)isthelowestprobabilityofanerrorinava-
lanche activity classification when S is known;

P(e|S)=[P(e|So).P(So)+P(e|S1).P(S1)]/[P(So)+P(S1 )]

Thequestionisnow:howmuchdoesknowing snow-
drift improve our ability to predict avalanches?In order to
give an answer we studied therelationship betweensnow-
drift and avalanche activity.

2.1.TEST SAMPLE

Weuseddatacollectedduringawhole winter season(1995-
96) at Alpe d'Huez ski resort. This ski resort is our 'outdoor
laboratory'forthe development of the forecasting system
NxLogandprovidesveryaccurate and reliable data:ski
patrolmendailyattempttotriggeravalanchesinsome
defined paths whatever the snowpack conditions are. For
ourexperiments,thesameslopesaretestedeveryday,
irrespectively of their expected stability state. This means
that we actually know, day by day, whether the snowpack
was stable ornot in these paths.
At the same time weather data are collected at a meas-
urementstationwhich isvery closetothepaths.Wind-
(instantaneousvalues)andsnowdriftmeasurementsare

P(e|S) =1-[!]
i^max j ^Pij

[3]

TheindexlASshowstheproportionalreduction inthe
probabilityof errorafforded by specifying snowdrift data:
if theknowledge of snowdrift doesnot reduce the prob-
abilityoferror,the index is0.On theotherhand, ifthe
index is 1, no error is made given the snowdrift classifica-
tion:there is complete predictive association.
Now let us calculate lAS.It comes from [1],[2] and [3]:
ijjiji

131