American Meteorological Society

Covariance Analysis Technique Based on Bivariate Log-Normal Distribution with Weather

Modification Applications
Author(s): Paul W. Mielke, Jr., James S. Williams and Sing-chou Wu
Source: Journal of Applied Meteorology (1962-1982), Vol. 16, No. 2 (February 1977), pp.
183-187
Published by: American Meteorological Society
Stable URL: https://www.jstor.org/stable/26178150
Accessed: 13-07-2023 15:15 +00:00

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

American Meteorological Society is collaborating with JSTOR to digitize, preserve and

extend access to Journal of Applied Meteorology (1962-1982)

This content downloaded from 206.84.140.5 on Thu, 13 Jul 2023 15:15:29 +00:00
All use subject to https://about.jstor.org/terms
 February 1977 P. W. MIELKE, J. S. WILLIAMS AND S-C. WU 183

Covariance Analysis Technique Based on Bivariate

Distribution with Weather Modification Applicat

Paul W. Mielke, Jr., and James S. Williams

Department of Statistics, Colorado State University, Fort Collins 80523

SlNG-CHOU WU

Department of Computer Science and Statistics, California Polytechnic State Univer

(Manuscript received 29 July 1976, in revised form 3 January 1977)

ABSTRACT

A statistical technique based on the bivariate log-normal distribution is presented for analyses of tr
ment (cloud seeding) effects on measurements (precipitation amounts) when appropriate covariates (c
lated control area measurements) are available. An example is given which evaluates effects of seeding
specific 500 mb temperature partitions of 24 h precipitation amount data from the Ï964—70 Wolf Creek
wintertime orographic cloud seeding experiment. In addition, an appendix includes an analogous analy
technique based on the bivariate log-normal distribution for cross-over designs.

1. Introduction
the bivariate log-normal distribution. The analysis
technique based on this distribution is presented in
The purpose of this article is to present a useful
Section 3. A detailed numerical example is given in
covariance analysis technique based on the bivariate
Section 4 which involves an evaluation of cloud seeding
log-normal distribution. A specific application of this
effects on specific 500 mb temperature partitions of
technique involves analyses of possible effects by cloud
24 h precipitation amount data from the 1964-70
seeding on target area precipitation data when ap
Wolf Creek Pass wintertime orographic cloud seeding
propriate (viz., non-contaminated, well correlated, etc.)
experiment sponsored by the State of Colorado (Grant
control area precipitation data are available.
and Elliott, 1974). Finally, an analytic technique
The primary considerations affecting the choice of
based on the bivariate log-normal distribution for a
a bivariate distribution involving asymmetric mar
cross-over design is contained in the Appendix. Al
ginals for analyses of paired measurements such as
daily precipitation amounts at two locations are noted. though the cross-over design is often limited in ap
plication because of contamination problems, it has
The first consideration is the ability of the distribu
on occasion been effectively utilized in weather modi
tion to describe the measurements in question. In
particular, the log-normal distribution is a reasonable fication studies including the outstanding Israeli
candidate for fitting both precipitation and relatedartificial rainfall stimulation experiment (Gabriel,
1967; Mielke, 1974).
measurements in comparison to alternative candidates
such as the gamma distribution and closed form beta
2. Bivariate log-normal distribution
distributions of the second kind (Mielke and Johnson,
1974). Other arguments in support of the log-normal Let X and Y denote bivariate log-normal random
distribution for describing precipitation have also beenvariables with marginal distribution shape parameters
proposed (Biondini, 1976). The second considerationa and ß, and scale parameters A and B, respectively.
involves the tractability of applying bivariate dis Then the bivariate log-normal density function of X
tributions for this purpose. In this respect, the biand V is
variate log-normal distribution is relatively simple to
employ. In contrast, a bivariate gamma distributionfx,y{x,y) = [_2irxyaß(l — p2)*]-1
with similar properties requires cumbersome computa
tions to implement such a covariance analysis tech
nique (Moran, 1969). Also, other contenders such as
x^(__L_lfîî^T
V 2(1—p2) IL a J
the bivariate beta distribution of the second kind are
similarly plagued with severe implementation problems.
Section 2 includes a description and properties of ^ |-ln(VA)jln(y/5)j^|-lnO;/£)j2J^

This content downloaded from 206.84.140.5 on Thu, 13 Jul 2023 15:15:29 +00:00
All use subject to https://about.jstor.org/terms
184 JOURNAL OF APPLIED METEOROLOGY Volume 16

where 0<min(z,y,a,/3,/l ,B) and p, the dependence

the continued-covariate design described by Wu et al.
parameter, satisfies |p|<l. If p= 0, then X and F The continued-covariate design is commonly
(1972).
are independent random variables. a necessity in weather modification experiments in
The moments of X and F about zero are givenvolving
by target and control areas because of anticipated
downwind seeding contamination in the target area
E{X>Y') = A>B' exp[(sV+ Pß?+2pstaß)/2~].
if the control area is seeded (and not vice versa). The
In particular, the mean and variance of X are example in Section 4 is typical of this situation. Thus,
while the optimal design for this example prescribed
E(X)=Ae°2» and Var(X) = A2ea2(e"2-l),
by Wu et al. (1972) is not feasible due to contamina
the mean and variance of F are tion (i.e., the prescribed optimal design would have
been a balanced cross-over design had contamination
E(Y)=BeP'2 and Var(F)= BV'(e^-l), not existed), the analysis technique associated with
and the covariance and correlation of X and Y are the continued-covariate design used is appropriate.
The following covariate analysis technique based on
Coy(X,Y)=AB(e^-\)e^^12, the bivariate log-normal distribution for testing the
Cor (X,Y) = (e'°0-l)/[(e°2-l) (e^-1 )]». null hypothesis (Ho: A n= A s) versus the alternative
hypothesis (H\\ An^A s) is the same analytic tech
3. Covariance analysis technique nique, following a transformation to a bivariate normal
distribution, which was described by Wu et al. (1972)
Let (xiji), i— 1, ..m+n, denote m+n indefor continued-covariate designs. They showed that
pendent pairs of precipitation amount measurements these simple design and analysis variations of the
from target and control areas, respectively, in either
classical covariance technique have optimal properties
a weather modification experiment or project. In this
for testing the hypotheses in question.
instance x and y denote the variates designated for The description of this covariance analysis tech
treatment and covariate, respectively. For conven nique is simplified by using the following set of suf
ience, we assume the first m and last n pairs of mea ficient statistics :
surements designate the non-seeded and seeded events
m m+n
in the target area, respectively. (Actually the arrange
UiN = m~lY,\nxi uis = n~l ]T lnxi
ment of pairs is inconsequential for this model with 1=1 i—m+l

time-independent responses. Any pattern, predeter m m+n

mined or resulting from randomization, is analyzed UïN = m

in the same manner.) This design is a continued tmI i=m+1

covariate design (also known as a target-control m m+n

ViN
design). If seeding has an effect, this effect is assumed = nrlYL^yi v\s=n~l £ hvyt
i'=I i—m+1
to appear only in scale parameter differences of the
log-normal responses associated with non-seeded and m+n

seeded target area measurements. Thus we let AN ViN 2 v2S=n~1

= XnriY.
0n:y;)2 (
i=m+l
and A s denote scale parameters associated with non
m+n
seeded target area measurements, respectively, in the
(uv)N = m~1 X! (lnx<)(lny,-) (uv)s = n~1 £ (hu.Xlnyi)
bivariate log-normal distribution. It is assumed that 1=1 i=m+1
a, ß, B and p are unaffected by seeding. The null
hypothesis that seeding does not change target area Under H0: AN
measurements in any way is therefore quantified by
An=As
The covariance analysis technique presented here
(1+H)s*
mn J [m+n "T*
is distributed as Student's t w
is directly related to an optimal small-sample analysis
technique (which is also asymptotically optimal)freedom
for where

T = Uis—Uin—G(Vis — Vin),

^[(wtOjv—Miiv!>iiv]+w[(tta)s—Missis]
G=
«t(tf2W — ®12v)+»(®2S —®îs)

mn (vis—vin)2
E=
m+n w(t>2jv_®ijv)+w(f2s_fis)
m (m2 w—m iv)+w («2 s ~ w I s) — G2[w (i'2 ,v—ïï.v)+» (i>2 s—"is) ]

m+n—3

This content downloaded from 206.84.140.5 on Thu, 13 Jul 2023 15:15:29 +00:00
All use subject to https://about.jstor.org/terms
February 1977 P. W. MIELKE, J. S. WILLIAMS AND S-C. WU 185

The point estimate of A S/A N is given by ticular, 100[(yl,s/^jv)

percentage increase d
Âs/Ân=eT denotes a decrease).
Specific point estimat
and the 1 —y confidence limits for A s/A m are given by
and B are given by
f rm-\-n "]5 I J".v=exp (mlv)
expj r± (1 -\-H)s2 tm+n_i(y/2) |,
I L mn J ) Âs=ÂNeT
Ê=exp[(mvw+nvis)/(m+n)]
where tm+n-z(y/2) is that point of the t d
with OT+»—3 degrees
Point estimates
of of the freedom
shape and dependence paramewhich
by chance alone y/2 proportion
ters a, ß and p are given by of the

a= wijv)+»(«2S—wis)]/ (m-\-n)}*,
ß = {[m (v2N - vIn)+n (v2 s~ vfs) ]/ (m+n)} *,

A
m[_(uv)N—Miat»ijv]+w[(m!))s—Mis»is]
P =
{\jti (u2 n—Mi n)+w {u2 s—wis) ~]\jn (i>2 n—v\n)+n (v2 s ~ vîs) ]} '

The above point estimates of a, ß and p are alsoment (cf. Mielke et al., 1971) was designed to compare
appropriate as given above (i.e., with no modificationnon-seeded and seeded daily precipitation amounts.
of the sufficient statistics) for the same parametersAs a consequence, a 24 h period was the experimental
of the bivariate log-normal distribution associated withunit of the Climax experiment. Since non-seeded and
the analytic technique given in the Appendix for the seeded 24 h periods of the WCP experiment were
cross-over design. intentionally not randomized, comparisons of non
seeded and seeded 24 h precipitation amounts from
4. Numerical example the three WCP precipitation sensors per se are highly
questionable because of climatological differences be
The data used to illustrate this covariance analysis
tween winter seasons. However, the use of a covariate
technique were obtained during the 1964-70 Wolf
based on appropriate control data provides a vehicle
Creek Pass (WCP) wintertime orographic cloud
to adjust comparisons of non-seeded and seeded 24 h
seeding experiment. This experiment was sponsored
period precipitation amounts for the climatological
by the State of Colorado and was both designed anddifferences between winter seasons. The control area
operated by Colorado State University under the
(an area 50-100 mi west of WCP) data consisted of
direction of Lewis 0. Grant. A description of this
six National Weather Service (NWS) precipitation
experiment has recently been given by Grant and
sensors located at Durango (NWS #05-2432, 1996 m),
Elliot (1974).
Mesa Verde (NWS #05-5531, 2155 m), Ouray (NWS
In particular, the target area of this experiment
#05-6203, 2359 m), Pleasant View (NWS #05-6591,
was the WCP vicinity. Three precipitation sensors
2091 m), Silverton (NWS #05-7656, 2841 m) and
spaced at roughly 5 mi intervals were used during
Telluride (NWS #05-8204, 2669 m). Since each of
each of the six winter seasons of this experiment.
these control area precipitation sensors is situated
These three precipitation sensors were designated
WCP West (2899 m), WCP Summit (3249 m) and50-100 mi west of WCP, contamination of any control
WCP East (2819 m). The target area was seededarea precipitation sensor by seeding of the WCP ex
periment is exceedingly remote.
during three winter seasons of this experiment
In this example x is the mean of precipitation
(1964-65, 1966-67 and 1968-69). The target area
amount measurements (inches of water) from the
was non-seeded during the other three winter seasons
three WCP precipitation sensors for a 24 h period
of this experiment (1965-66, 1967-68 and 1969-70).
providing (i) no more than one of the three measure
Contrary to a statement by Grant and Elliott (1974),
ments is missing and (ii) the mean precipitation
the WCP experiment was initiated during the 1964-65
winter season. amount is at least 0.015. Also y is 0.01 plus the mean
The winter season was the experimental unit of of precipitation measurements (inches of water) from
the WCP experiment since a primary goal of this the six control area NWS precipitation sensors for a
effort was to compare non-seeded and seeded winter24 h period providing at least three of the six mea
season snowpack runoff amounts. The concurrent surements are non-missing (the purpose of adding 0.01
Climax wintertime orographic cloud seeding experito each value is simply to insure that all covariates

This content downloaded from 206.84.140.5 on Thu, 13 Jul 2023 15:15:29 +00:00
All use subject to https://about.jstor.org/terms
186 JOURNAL OF APPLIED METEOROLOGY Volume 16

are positive prior to taking logarithms). Among ERT-71-01885-A03,

all and a cooperative agreeme
the Rocky Mountain Forest and Range Expe
the 357 pairs of target and control area measurements
Station and the Northern Forest Fire Labora
(Xi,yi) obtained during the WCP experiment, 185 (172)
pairs involved a non-seeded (seeded) target area.
USD A Forest Service. The authors appreciat
A further point worth noting is that lag correlation
ments by a referee.
estimates associated with these data were negligible
(near zero). APPENDIX

Using the previously described data of the WCP

Analytic Technique Based on Bivariate Log-No
experiment, we will illustrate the covariate analysis
Distribution for Cross-Over Designs
technique by evaluating two disjoint and exhaustive
500 mb temperature partitions of this data (viz., Suppose the m non-seeded events of the conti
<—23°C and ^ —23°C). As in previous investiga covariate design described in Section 3 are rep
with modified events such that the target are
tions, the 500 mb temperature is merely an estimate
of cloud top temperature. non-seeded and the control area is seeded. If now the
target area is termed target area I and the control
500 mb temperature < —23°C (m=83, n=72)
area is termed target area II and the two target areas
Muv=-2.0073 Mis=-1.9620 G=0.55332 are alternately seeded, then this modified design is a
uÏN= 5.2744 m2S= 4.8009 //=0.000001 cross-over design.
V\x— —2.8964 vxs= -2.8987 *2=0.76566 For a cross-over design to be effective, precipitation
amount measurements from the two target areas must
z<2Ar= 9.6943 ViS= 9.4144 r= 0.04657
be well correlated and neither target area can be
(■uv)N= 6.5825 Ms= 6.1937 t1M (0.025)= 1.975
seriously contaminated by seeding of the other target
T= 0.33 [2-sided P-value is 0.74] area. As in Section 3, let AN and A s denote scale
parameters associated with non-seeded and seeded
As/An= 1.05 [estimated 5% increase (non-significant)
target area I measurements, respectively. However,
due to seeding]
now let Bn and Bs denote scale parameters associated
(0.79,1.38) are 0.95 (95%) confidence limits for A S/A N
with seeded and non-seeded target area II measure
i*=0.134 is=0.140 £=0.055 ments, respectively. Thus either scale parameters AN
and Bn or scale parameters A ,s and Bs are associated
<$=1.053 j3=1.081 p=0.568
with a specific event. It is assumed that As/AN
500 mb temperature ^ —23°C (m= 102, n= 100)
= Btf/Bs (i.e., the effect of seeding on either target is
«1^=-2.0481 mis= -1.6170 G=0.68861 the same). Thus AS=AN implies BS=BN and visa
u2N= 5.8693 m2S= 3.9933 ff=0.000034 versa.

v1N= -2.6571 z>ls= -2.6420 s2=0.73743 Suppose Xi and y, are replaced with w,-= %,/)>,
V2N— 8.6894 v2S= 8.7331 r= 0.42070 Zi~ Xiyi, respectively, in each sufficient statistic
in Section 3. Then test statistic T in Section 3 is
(uv)N= 6.6600 (uv)s= 5.3812 /199(0.025)= 1.972
again distributed as Student's t with m+n—3 degrees
T = 3.48 [2-sided P-value is 0.0006] of freedom under Il<>: AN=A ,s and where each of the
statistics r, G, H and s2 now involve the modified
Âs/Ân= 1-52 [estimated 52% increase due to seeding]
sufficient statistics described in the previous sentence.
(1.20,1.93) are 0.95 (95%) confidence limits for A s/A N
However, the point estimate of either A s/An or Bn/Bs
iJV=0.129 is=0.196 £=0.071 is now given by
<î= 1.236 ß= 1.300 p=0.724 Âs/Ân— Ên/Ê s— eT,ï
Not only do the above results illustrate
and thethis
1—7 co
confidence limits for either As/AN or
variance analysis technique based on Bn/Bs
the bivariate
are given by
log-normal distribution, they also yield additional
support for previous findings pertaining to the cloudt /m+n V 1
exp -±
top temperature association with wintertime orographic
2 \ 4mn / J
{\+H)sAtm+n
cloud seeding (Grant and Elliott, 1974).
Also specific point estimate
Acknowledgments. This research was performed
As, Bs as and Bn are given by
part of the National Hail Research Experiment (Sub
^ljv=exp(MiAr),
contract S5013), managed by the National Center for 5s=
Âs=ÂNeTl2, ÊN=BseTi\
Atmospheric Research and sponsored by the Weather
Modification Program, Research Application Direc
torate, National Science Foundation, The
andpoint
alsoestimates
was of the shape and dependence
parameters
jointly supported by NSF, RANN, under Grant a, ß and p are identical (i.e., do not in

This content downloaded from 206.84.140.5 on Thu, 13 Jul 2023 15:15:29 +00:00
All use subject to https://about.jstor.org/terms
February 1977 P. W. MIELKE, J. S. WILLIAMS AND S-C. WU 187

and Probability,
volve modification of sufficient statistics Vol. 5, University of California Pre
given above)
91-113.
to those point estimates of a, ß and p given in
Section 3. Grant, L. 0., and R. E. Elliott, 1974: The cloud seeding tem
perature window. J. Appl. Meteor., 13, 355-363.
If the choice of a design is optional (i.e.,
Mielke, no
P. W., problem
1974: Squared rank test appropriate to weather
such as contamination exists), the optimal design
modification both
cross-over design. Technometrics, 16, 13-16.
can and should be calculated with the ,algorithm
L. O. Grant and given
C. F. Chappell, 1971: An independent
replication of the Climax wintertime orographic cloud
by Wu et al. (1972). In particular, the balanced cross
seeding experiment. J. Appl. Meteor., 10, 1198-1212; Cor
over design would be a most informative
rigendum, 15,design
801. for
the example in Section 4 (as a consequence of ä^ß).
, and E. S. Johnson, 1974: Some generalized beta distribu
tions of the second kind having desirable application features
in hydrology and meteorology. Water Resources Res., 10,
REFERENCES 223-226.
Moran, P. A. P., 1969: Statistical inference with bivariate
Biondini, R., 1976: Cloud motion and rainfall statistics. gamma
J. AppI.
distributions. Biometrika, 56, 627-634.
Meteor., 15, 205-224. Wu, S., J. S. Williams and P. W. Mielke, 1972: Some designs
Gabriel, K. R., 1967 : The Israeli artificial rainfall stimulation
and analyses for temporally independent experiments in
experiment. Statistical evaluation for the period 1961-65.
volving correlated bivariate responses. Biometrics. 28,
Proc. Fifth Berkeley Symposium on Mathematical Statistics
1043-1061.

This content downloaded from 206.84.140.5 on Thu, 13 Jul 2023 15:15:29 +00:00
All use subject to https://about.jstor.org/terms