WMOD00478

• Title: Field Experimentation in Weather Modification
• Date: 3/1/1979
• Weather Modification - Doc Type: Report

--

1
,

i
\

----~,.

~

~,-.7",-""_''''''''''.",,",,,.7--i.ii"'''''_~ .....-

Gabriel: Field Experimentation in Weather Modification

K. RUBEN GABRIEL *

81

Comment

Dr. Braham has raised the question of how deeply
statisticians should be involved in weather modification
work. This is an interesting question which is presumably
equally relevant, mutatis mutandis, to other areas in
.which statistics is used.
To deal with this qu.estion, it is useful to distinguish
four stages of statistical activity in the experimental con-
text: (1) the experiment is designed i (2) it is performed i
(3) the principal hypothesis of the experiment is con-
firmed or refuted and the corresponding parameter(s)
estimated; and finally, (4) other aspects of the results
are explored. .The confirmatory stage (3) of analysis con-
centrates on the principal purpose of the experiment, that
hypothesis and/or those parameters for whose study the
experiment was designed. Thus, the principal motivation
of weather modification experiments usually is to deter-
mine whether the hypothesis, that seeding augments pre-
cipitation (the parameter), is sufficiently true to justify
operational cloud seeding-the proposed action. In ad-
dition, the data from such experiments are scanned for
effects on distributions of precipitation, and for relation
of seeding effects to season, to time, to extent of seeding,
and to a host of concomitant meteorological variables.
These are the exploratory analyses, which form stage (4)
. of statistical activity.
Few experiments are analyzed only for the single
principal hypothesis, and only rarely will an "ex peri-
'ment" be run just to allow undirected exploration of its
results. In general, experiments are aimed at a combina-
tion of confirmatory and exploratory analyses, the former
relating to the principal question posed by the experimen-
ter, the latter allowing him to fan out in the search for
novel ideals. It is not generally useful to label an entire
experiment as either confirmatory or exploratory, since
most experiments allow BOrne analyses of each kind.
Analysisand experimentation are a sequential or cumula-
tive process. The exploration of one experiment's out-
comes usually results in the formulation of new hypo-
theses which may then be subject to confirmation by the
next experiment.
Now consider the statistical methodology appropriate
to stages (3) and (4). (Planning these analyses also
determines the proper design and performance of the
experiment.) The confirmatory analysis stage (3) fits
very well into the framework of two-decision significance
testing theory. A hypothesis is formulated, variables
chosen, distributions postulated, parameters defined, and

· K. Ruben Gabriel is Professor of Statistics and Biostatistics at
the University of Rochester~ Rochester, NY 14627.

'--~~-_."'---.........:.._~ :.-.-

a statistic derived so as to maximize test power and/or
estimation efficiency. The experiment is supposed to have
m~en performed strictly according to the design, and to
culminate in the computation of a statistic which leads
to a decision-all this exactly following the protocol set up
in the design. The end product is a statement of signifi-
es.nce, accompanied by an estimate of the relevant pa-
rameter with error intervals.
The exploratory stage (4) consists of all conceivable
analyses of the experimental results classified in any
reasonable way, with a view to verifying hunches about
the special characteristics of seeding effects (e.g., rare
extreme "blockbuster" effects vs. common average
effects), about their relation to time (e.g., hours after
seeding, month of year), to place (.e.g., within-plume or
outside it, on-target or downwind), to type of storm
(e:.g., wind direction), to cloud type (e.g., tower height,
Wi!l.ter content, seeding nucleus prevalence), and about
the relevance of any other information. Some of the con-
comitants will have been suggested a priori by earlier
e:xperiments, others will occur as a result of inspecting
the data.
The statistical methodology required for such in-
vestigati~n is mainly descriptive and exploratory data
analysis. Methods ~f sifting through multiple cross-
cl:ilSSifications come into their own here, as do techniques
for tracing trends, locating differences, and identifying
J>f~tterns. Classieal methods of hypothesis testing are quite
marginal to this activity. The criteria a.re not maximal
power or efficiency but rather resolution of data into
interpretable regularities. In exploring many facets of the
<u~ta, one is very unlikely to know enough about each
facet to allow selection of "optimal" techniques. Instead,
one does better with simple techniques which are suffi-
cil~ntly easy to understand to allow the statistician to
work hand in hand with the meteorologist. They should
explore experimental results together, suggesting patterns
and ideas to each other and progressively sifting the data
and interpreting them in physical terms.
This stage of analysis is classified as exploratory be-
cause that is the main thrust of any freewheeling study
of an experiment's results. But exploration is always ac-
companied by some confirmation (Tukey 1977, p. vii):
any striking finding of the exploratory analysis will also
be judged in terms of how likely it is to have arisen by'
pure chance. The investigator will require something like
st:!l.ndard errors or P-values to be attached to expl9ratory
.findings so that he has some idea of how "nonrandom"
they are. One problem with such confirmatory calcula-