WMOD00478

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

10

::

treatments. If a randomization is thought to be inap-
propriate at the outset, then it should be set aside and a
new randomization obtained. Perhaps this could increase
the sensitivity of the experiment without seriously affect-
ing the properties of the significance test based on the un-
constrained randomization distribution. On the other
hand, to maintain the integrity of the randomization
distribution, the complete set of bad randomizations
should, as far as practical, be eliminated before a design
is obtained. For example, the randomized block design
may be visualized as one way of eliminating bad ran-
domizations. When constraints on the randomization do
not yield a known design, it may not be possible to rely
on the usual normal-theory approximations to the signifi-
cance tests. In this case, tests must be based on the
. randomization distribution directly.
The randomization oontroversy in Whitetop is more
difficult since a bad randomization was suspected only
after completion of the experiment. The experiment was
randomized and, from all evidence available to us, the
randomization appears to have been rigorously imple-
mented. The suspicion of a ba.d randomization was based
on the meteorological phenomenon of persistence; that
is, the rainfall in the target area during the time interval
of the experimental unit is positively correlated with the
upwind rainfall for the ten hours preceding the unit. The
problem is how to allow for this additional information
in the interpretation of the results. The answer should
hinge on the strength of the inference made under ran-
domization theory versus the support for the foundations
of the model that connects the additional information
with the experiment. Apparently Professor Braham is
more uncertain of the results of the additional analysis,
because of nonindependence, partitioning, and multiple
testing, than of the correlation model. In any case, if the
information is believed to be at all relevant, it should of
course be revealed. In this regard the cautionary note
of the Berkeley group is quite appropriate, although it
may have been worded too strongly since it has evidently
caused many to dismiss the results of the experiment
without weighing the issues mentioned.
Finally, we comment briefly on some of the specific

SHYRL M. DAWKINS and ELIZABETH L. SCOTT*

-

Journal Df the American Statistical Association, March 1979

issues raised by Professor Braham in the last section of
his article.
St8.tistics by its nature tends to be interdisciplinary.
Many statisticians have developed considerable expertise
in substantive areas without formal training. Surely, this
could be done in meteorology. Lack of training might
,obstruct communication between meteorologists and sta-
tisti{:ians in the initial contacts, but it would be very
unfortunate if this also obstructed mutual respect. Cer-
tainly, every meteorologist need not be a stand-aloile
statistician, but at the same time it seems to us that
resolution of many of the statistical issues involved in
weather modification experimentation does require con-
siderable statistical expertise.
W,e somewhat agree with the view that weather modifi-
cation research may not have reached the stage of con-
firmll~tory experimentation. In the presence of treatment-
unit nonadditivity, it is not clear which hypotheses should
be tested or which facts confirmed.
Combining data over experiments must always be done
with care. However, it may not be possible to combine
experiments when treatment-unit nonadditivity is pres-
ent. The problems here are great. At the very least, the
experiments to be combined should have the treatments
and units be the same qualitatively. Past weather modifi-
cation experiments have employed a variety of definitions
for experimental units; observations have been taken for
a fixed period of time or for periods of stormy weather,
and the target areas have been fixed or floating. There
seems little reason to suspect that even the qualitative
natw'e of the treatment-unit nonadditivity will remain
unchl!l.nged whEm the treatment and/or unit are varied.

REFERENCES

Cook, R. Dennis, and Holschuh, Norton (1978), "Statistical Design
for the Evaluation of Cloud Seeding in Minnesota," Technical
Report No. 309, School of Statistics, University of Minnesota.
Fisher, R.A. (1926), "The Arrangement of Field Experiments,"
JOUlrnal oJ the Ministry oj AgricuUure oJ Great Britain, 33, 503-513.
[paper 17 in Fisher, R.A., Contributi0n8 to Mathematical St.atisticB
(19fiO), New York: John Wiley &: Sons.]
Woodley, William L., Simpson, Joanne, Biondini, Ronald, and
Berkeley, Joyce (1977), "RainfaIl Results, 1970-1975: Florida
Areu Cumulus Experiment," Science, 195, 735-742.

Comment

1. INTRODUCTION

We agree with Professor Braham that weather modifi-
cation is an important societal problem; for more than

· Shyrl M. Dawkins is a Programmer and Elizabeth L. Scott is
Professor, both at the Statistical Laboratory, University of Cali-
fornia, Berkeley, CA 94720. This work was supported by the Office
of Naval Research, Contract No. ONR-NOOOI4-75-~159.

25 years, the research in the Statistical Laboratory at
Berkeley has included studies of weather modification,
especially the design and analysis of weather modification
experiiments. Whitetop is indeed an important experi-
ment "combining a randomized seeding experiment with
basic studies of cloud physics."