, MULTI-STAGE SAMPLING ON SUCCESSIVE OCCASIONS WHERE FIRST- STAGE UNITS .ARli; DRAWN WITH UNEQUAL PROBABILITIES AND HITH REPLACEMEET by Niyom Purakam and. John C. Koop Institute of Statistics M:l.meograph Series No. 472 April, 1966
,
MULTI-STAGE SAMPLING ON SUCCESSIVE OCCASIONS
WHERE FIRST-STAGE UNITS .ARli; DRAWN
WITH UNEQUAL PROBABILITIES AND HITH REPLACEMEET
by
Niyom Purakam and. John C. Koop
Institute of StatisticsM:l.meograph Series No. 472April, 1966
•
,
ABSTRACT
PURAKAM, NIYOM. Multi-Stage Sampling on SUccessive Occasions where
First-Stage Units are Drawn with Unequal Probabilities and with
Replacement. (Under the direction of JOHN CLD:lENT KOOP).
A multi-stage sa~ling design, particularly intended for large
scale sample surveys on successive (or repeated) occasions is developed.
The sampling design is general in the sense that the probabilities of
selecting units (for the preliminary first-stage sa~le) are arbitrary.
Each of these first-stage units is drawn with replacement. The
technique of partial replacement of first-stage sa~ling units is based
on the order of occurrence of these units. The partial replacement
technique is developed to meet two basic objectives:
(i) To spread the burden of reporting among respondents which may
be expected to help in maintaining a high rate of response.
(ii) To enable the sampler to take advantage of the saItij?ling design
in the reduction of sampling variance of several estimators proposed.
Several ways of utilizing the past as well as the present informa
tion from the sampling design to estimate the total, and the change in
total of a population characteristic of interest, are presented. The
nature of the gain in efficiency from using the four different forms of
estimators in estimati~g the total, and the change in total, is explored.
The comparisons of efficiency among the estimators wherever possible,
are given under certain assumptions simiiar to the assumption of Second
Order or Weak Sense Stationarity· usedin conventional time series
analysis.
•
,
t,
The estimation theory is covered in detail for two-stage sampling
on two successive occasions. The extension to higher stage sampling on
more than two successive occasions is sufficiently indicated, In all,
the reduction in the variance of an estimator whenever achieved, is in
the total variance namely, the between first-stage units variance plus
the within first-stage units variance, and so on if there are more than
two stages of sampling.
..
iii
ACKNOWLEDGEMENTS
To Dr. J. C. Koop, the chairman of my advisory connnittee, I wish
to express my sincere appreciation and thanks for his guidance during
my graduate work here. It was mainly through his suggestion and
direction that the work reported in this thesis was completed.
My thanks are extended to other members of my advisory connnittee:
Dr. R. G. D. Steel, Dr. J. Levine, Dr. N. L. Johnson and Dr. B. B.
Bhattacharyya for their useful suggestions and criticisms of th~ draft.
I am also very grateful to The Agency for International Development
for the financial support which enabled me to pursue graduate study
in this country.
For their respective shares in the work of typing the draft and
the final version of this thesis, I wish to express my heartfelt thanks
to Mrs. Ann Bellflower and Mrs. Selma McEntire.
Finally, to my father who had to struggle all his life to support
my early education, but who never saw the result at this stage, this
thesis is humbly dedicated •
.. iv
TABLE OF CONTENTS
LIST OF TABLES • . . . • 0 • 0 0 • • • • • vii
Page
19 ~ • •. . .. .INTRODUCTION •1.
o • • • •1.1. Basis for the Present Investigation1.2. Nature of the Problem • • • • • • • 0 • •
• • 12
13
6
· .
. .. . .000000 •• 02 • REVIEW OF LITERATURE • • • •
3. A PROPOSED SAMPLING DESIGN
3.1. Description of a Method of Partial Replacement ofFirst-Stage Units ••••••• • • • • • • • • •• 12
3 .2 • Advantages of the Proposed Scheme of PartialReplacement of First-Stage Units • • • • • • • • • • 14
3.3. Specification of Probability System and the Methodof Selecting Sub-Units • • • • 16
4. ESTD1ATION THEORY • 0 • 0 • • 21
4.3.1.4.
4.3.1.5.
4.3.1.1.
4.3.1.2.
4.3.1.3.
4.1.4.2.4.3.
Introductory Remarks • • • • • • • • 21Estimation of Total for the First Occasion • • • 23Estimation of Total for the Second Occasion • • 27
4.3.1. A Linear Composite Estimator • ••• 28....
Expected Valu; of 2T •••• ,. " 29
Variance of 2~ • • • • • • • • • 30
Note on the Estimation of""-
Var (2~) • • • • • • • •• ,.. 42~
Efficiency of 2T • • • • • • • • 43
Choice of Q in the LinearConu;>osite Estimator ••• • • 47
4., .1.6. Choice of Ii ...•...•.• 524.,.1.7. Simultaneous Optimum Values of
Q and IJ. in the LinearComposite Estimator • • • • • 54
40,.1.8. Expression for the MinimumAttainable Variance of the ""Linear Conu;>oSite Estimator 2t!'. 56
4.3.2. A Modified Linear Composite Estimator '. '. '. 594.3.3. A General Linear Estimator • • • • • • • 66
4 .3 ., .1. Determination of Constants • 674.,.,.2. A Comment on the Form of th~
Estimator 2T* • • • • • • • • 73
v
TABLE OF CONTENTS (continued)
Page
99
76
79
7982
91
93
94949496. . .
4.3.3.3.
4·3.3·5.
Efficiency of the Estimator 2T*. 73Optimum Value of IJ. when the,
Estimator /i* Is Used. ,. .' ,.
4.3.4. A Ratio-Type Composite Estimator ••• • " "~*4.3.4.1. Expected Value of 2T •••. " "
4 4 A*·3. .2. Variance of 2T .•••• ,. " "
4 . l\i\*4.3 •.•3. EffJ.ciency of 2T .,. ••
Estimation of Change in Total between the First and'Second Occasion • • • • • • • • • • . • • • • • • •4.4.1. An Estimator Based on the Linear Composite
Estimator •• • • • • • • • • • • • •4.4.1.1. Estimator •• • ••4.4.1.2. Variance ••••••••.' .' .' .'
~
4.4.1.3. Efficiency of D1 •
4.4.1.4. A Comment about the Gain inEfficiency • • • • • • • •
4.4.1.5. A Remark about the Optimum .Value of Q • • • • • • • • • • 102
4.4.2. An Estimator Based on the Modified LinearComposite Estimator •• • • • • • •• 103
~t4.4.2.1. Variance of D1 •••••••• ,. 103
At4.4.2.2. Efficiencyof P1' P ,••••• " ,.104
4.4.3. A General Linear Estimator .•••••••• 1084.4.3.1. Determination' of Constant· a','
b " C I, d I • • • • • • • • • • 108
4.4.
..
4.4.3.2 • A Comment about the Form of. the,
""*Estimator D1 • • • • .,. ," 111A*4.4.3.3. Variance of D1 ••••••••• 111
4.4.3.4. Efficiency of the Estimator Dr 112
4.4.4. An Estimator Based on the Ratio-TypeComposite Estimator • I' • • • • •• •• ,. 114
A*4.4 .4 .1. Bias of D1
• • • • •• " " 114s*4.4.4.2. Variance of JJ1 • • • ••••• " 115
4.4.4.3. A Remark about Var (~~) •••• 117
vi
TABLE OF CONTENTS (continued)
Page
. . .5. MORE THAN TWO STAGES OF SAMPLING AND MORE THAN TWOSUCCESSIVE OCCASIONS • • • • •• • • . 119
· • 138• 146
• . 119• . 119• • 127• • 128• • 128
General Remarks • • • • • • • • • • • • • • • •More Than Two Stages of Sampling • • • • • • •More Than Two Occasions • • • • • • • • • •
5.3.1. Estimation of Total •••••••••••5.3.1.1. A Linear Composite Estimator5.3.1.2. A Modified Linear Composite
Estimator •••••••••5.3.1,.3. A General Linear Estimator5.3.1.4. A Ratio-Type Composite
Estimator • • • • • • • • • • • 152Estimation of Change in Total • • • • • • • • 156
5.3.2.1. The Estimator Based on theLinear Composite Estimator 156
5.3.2.2. The Estimator Based on theModified Linear CompositeEstimator • • • • • • • • • • • 156
5.3.2.3. A General Linear Estimator • • • 1575.3.2.4. The Estimator Based on the
Ratio-Type Estimator • • • • • 159
Note on the Estimation of Sampling VariancesNote on the Efficiency of' the Estimators when
1Vy :f 2Vy • • • • •••••• • • • • • • •
160
• 163
· .• ••• 160• ••• 161
. . .
. . . . .• • 0 • • • • 165
· . 165
••.•• 167
. .. . .
• • 0 • • • •
o 0 0 0 • • 0
. .• • 0 • 0 0'
. .• • • I)
. .• • • 0. .
8.1.8.2.
6. SUMMARY AND CONCLUSIONS •
6.1. Summary. • •6.2. Conclusions.
7. LIST OF REFERENCES • • • • • • • •
8. APPENDICES • • • • • • • • • • • •
J.
t.
vii
LIST OF TABLES
Page
;.1. Scheme of sample selection, partial replacement andconstituent samples ••.•••••• • • • • ~ • • • 15
4.1
4.2.
111 A
Percent gain in efficiency of 2T over 2T
1Optimum value of Q for ~ ~ 4 and percent
of 2~ over 2fE • • • • • • • • •
. . .gain in effici!3n~y
48
50
4.4.
4.6.
1Optimum value of Q for ~ ~ ~ and percent gain in efficiency,A A ./
of 2T over 2T • • • • • • • • • • • • • • • • • • • • • • 51
1Optimum value of Q for ~ ~ 2 and percent gain in eff1~iency,A A
of 2T over 2T • • • • • • • • • • • • • • • • • • • • • • 52
Optimum value of ~ for Q = ~ and the percent gain in~ A
efficiency of 2T over 2T • • • • • • • • • • • • • 54
* APercent gain in efficien~ of 2T over 2T for simultaneous
optimum values of Q and ~ • • • • • • • • • • • • • • • • 58
. .
•
r
4.7·
4.8.
4.10.
4.11
4.12 .
4.13.
4.14.
4.15·
~I AllPercent gain in efficiency of 2T over 2T for ~ = 2 ' Q = 2.
~
Optimum value of Q to be used in the estimator 2T for1 A
~ = 4 and the percent gain in efficiency over 2T ••••AI
Optimum value of Q to be used in the estimator 2T when1 A
~ = 3' and the percent gain in efficiency over 2T ••~I
Optimum value of Q to be used in the estimator 2T when1 A
~ = 2' and the percent gain in efficiency over 2T
A* A 1Percent gain in efficiency of 2T over 2T for ~ =4 . .Percent gain in efficiency of /r* over /i for ~ = ~
Percent gain in efficiency of 2fE* over /r for ~ = ~ ••••
Optimum value of ~ when the estimator 2T*iS used andA
the percent gain in efficiency over 2T • • • • • • • • •
* AllPercent gain in efficiency of Dl over Dl when ~ = 4 ' Q = 2'
63
64
65
66
74
75
76
78
97
viii
LIST OF TABLES (continued)
Page
4.16.
4.17.
4.18.
~ ~ 1 1Percent gain in efficiency of D1 over D1 for ~ =3' Q =2 98
~ ~ 1 1Percent gain in efficiency of D1 over D1 for ~ = 2' Q =2 99
Optimum value of Q to be used in the estimator ~1. and,~ 1
percent gain in efficiency over D1 for ~ = 4 . . .. . 100
• 101
• • • • 106• • • • It •• • • • • It • • •
~
Optimum value of Q to be used in the estimator D1 and~ 1
percent gain in efficiency over D1 for ~ =3' . • . .Optimum value of Q to be used in the estimator ~1 and
~ 1percent gain in efficiency over D1 for ~ = 2 . . . . . . . 102
At ~ 1Percent gain in efficiency of D1 over D1 for ~ = 2 '
1Q=2
4.20.
4.21.
4.19.
4.22.~,
Optimum values of Q to be used in the estimator D1 and~
percent gain in efficiency over D1 • 0 • • • 107
~Percent gain in efficiency of ;T
1 1~ =2' Q =2 . . . . . . . .
over
. . .";T for. . • 0 • 1;6
• • 146
151
• '1;8
• 14;
• 168
• •
. . .
over
. . . .. .
• .. • 0
The percent gain inefficiency of the estimator~ 1 1 2 (
2T, when Q = 2" ' ~ = 2" and 1VY = '4 2Vy)
~ "Percent gain in efficiency of ;T over ;T for some1 1
assumed values of P1 and P2, for ~ = 3' ' Q = 2 . . . 0 •
~ ~
Tt over T fo; ;. rPercent gain in efficiency of1 1
~=2,Q=2······· •••~ "Percent gain in efficiency of ; T' over ;T for
~ = ~ , Q = t for some values of PI and P2 •••
~* " 1Percent gain in efficiency of ;T over;T for ~ = 2~
2T
5.5·
8.1.
8.2.~ ~
The percent gain in efficiency of 2T over 2T when
Q =~ , ~ =~ and 1Vy =t (2Vy) ••• • ••• • • • •• 169
LIST OF TABLES (continued)
ix
8.4.
~
Percent gain in efficiency of 2T' over
Q=l,ll=l and V =2.(V)2 2 1y Ij: 2y
Percent gain in efficiency of 2~' over
Q=l,Il=1. and V =24
(V)2 2 1y 2y
. . .'"2T when
Page
••••• 170
• • • •• 171
,-L· INTRODUCTION
I:~l. Basis for the Pres'ent Investigation
In continuing sample surveys conducted at regular intervals (e.g.
quarterly) for investigating the time-dependent chaXacteristic of
certain dYnamic populations, it is frequently advantageous to use the
- 1/so-called rotation sampling- technique, Whereby a scheme of partial
replacement of sampling units i~ developed in such a way that the
sampling units to be used will be in the sample consecutively for some
fimte number of occasions, then they Will be replaced by newly selected
units • The replacement is done only to a portion of the sample while
the other portion is retained for the next occasion. To such plans of
sampling were attached different names by various authors, such as
I'sampling on succession occasions with partial replacement of units II
[12], "rotation sampling" [4], "sampling for a time-series" [7],
"successive Sampling" [17].
The main advantages of a technique of sampling where· partial
replacement of units is part of the overall sampling design over one
where there is no partial replacement of units (1:..!., taking a new set
of units, or using the same set of units every time) are as follows.
1. Partial replacement of units in the sample spreads the burden
of reporting among more respondents and hence results in better co-
operation from respondents. This is very iniportant from the standpoint
1/ II II- The name rotation sampling refers to the process of eliminat-ing some of the old elements from the sample and adding new elementsto the sample each time a new sample is drawn. (See Eckler [4]).
:2
of maintaining the rate of response when a human population is studied.
Experiences from many census or survey studies (by complete enumera-
tion or ·sSmpling methods) seem to indicate that the respondents tend
to become uncooperative during the third or fourth visit if the survey
is carried out repeatedly and the same sampling units are used. Even
with full cooperation, the respondents may be unwilling to give the
same type of information time a:f'ter time, or they may be inf'luenced by
the inf'ormation which they give and receive at earlier interviews, and
this may make them progressively less representative as time proceeds.
On the other hand, taking a new set of units every time is obviously
more expensive. The compromise is partial replacement of units.
:2. Partial replacement of' units in the sample permi-ts the use of
data from past samples to improve the current estimate of population
characteristics of interest. This can be accomplished by some appro-
priate methods of estimation which takes advantage of past as well as
present information to provide an estimate for the present occasion.
This theoretical advantage is perhaps the most important reason for
using partial replacement of units technique when we have to deal with
time-series characteristics.
A large scale sample survey such as a national sample survey, is
usually carried out in more- than one-stage to reduce time, labor and
costs. The new approach here is that the technique of partial replace-
ment of units can be applied to the units at any stage in the sam,pling
process. However, the U. S. Bureau of Census, in applying the technique
to various periodic sam,ple surveys such as the Current Population
•
3
Survey, the Monthly Retail Trade Survey, the Monthly Accounts Receivable
Survey, etc., partially replaces the last-stage units.
It is of theoretical and of practical interest to consider a some
what different scheme of partial replacement of units, where, in the
multi-stage sampling process, the partial replacement of units is
carried out at the first-stage. At the second-stage (or succeeding
stages, if more than two stages of sampling is used) the sampling may
be carried out in any appropriate fashion as the practical situation
may demand. The above problem seems to have received no attention so
far except by Tik.kiwal [11] in a different context.
1.2. Nature of the Problem
For the development of the theory underlying the proposed tech
nique of partial replacement of units in multi-stage sampling, we
w.i.ll fornmlate the problem as follows.
Consider a population whose characteristics of interest change
With timeJ for exampleJ the number of persons employed or unemployed
in the labor force of a country. These numbers are known to vary from
season to season. It is desired to conduct a large scale sample survey
to estimate such time-dependent characteristics of this population
periodically (say quarterly for a total period of 2 years) using a
multi-stage sampling plan. Also, in order to minimize some undesir
able effects resulting from interviewing the responsent repeatedly
over a long period of time, we seek some replacement techniques in
such a way that our sampling operation will utilize units to be in
Cluded for interviewing only for appropriate number of occasions.
•
4
Then replace them after some occasions by units not selected for inter-
viewing before. More specifically, suppose we are going to use a two
stage Samplini/ design, where, in our population, there are N definable
first-stage units:
t~, ~, , uJEach first-stage unit ui ' (i=1,2, ••• ,N) contains Ni second
stage units. For example, if the first-stage units are taken to be
villages or towns, the second-stage units might be households in those
villages or towns. In case of two-stage sampling, these second-stage
units if they are selected, 'Will actually be visited an.d interviewed
by the enumerators.
In the usual non-rotation sampling method, a sample of n units
out of the N first-stage units 'Would be selected from the population;
after that the procedure is to select ni second-stage units out of the
Ni·found in each selected ni • Hence, the total sample size under thisn
sampling scheme is E ni • If the sampling is done quarterly for 2i=l
years and those units are used every time, we will have to repeatn
visiting those E ni units 8 times. On the other hand, to selectn i=l
a new set of E ni units every quarter may be too costly. If thei=l
partial replacement techI)ique such as the one used by the U. S. Bureau
of the Census, is incorporated in the above two-stage sampling plan,
g/This case is chosen for illustration purposes. Later, we willconsider more tha.n two 3tages of sampling. Also, we 'Will not consider stratification because it does not make any difference in thedevelopment of our scheme of partial replacement of units in thesample •
a-
5
the selected second-stage units will be partially replaced on every
occasion. However, the same first-stage units are still being used on
every occasion. In sequel it will appear that such a sampling design,
using any appropriate estimator, brings about a reduction in variance
only in the within first-stage units part of the variance, while the
major contribution to the total variance of the estimator (!.~., the
between first-stage units part of variance) is not reduced.
Considering the fact that it is the between first~'stage units
part of the variance that contributes most heavily to the total vari
ance in many sub~sampling situations, a sampling design and estimation
procedure which reduces this variance would be desirable. The problem
then to be taken up in this thesis is "How would a sampling design
which reduces the total variance of an estimator be obtained? Further,
what is the appropriate estimation theory?" It is the purpose of this
thesis to investigate the above problem.
2. REVIEW OF LITERATURE
The first attempt to utilize the information obtained from
previous. samples to· improve the estimate for the current occasion
seems to have been made by Jessen [8]. To estimate the population
mean of a characteristic of interest, he conducted a survey on two
successive occasions. On the first occasion, a simple random sample
was drawn (in one stage). On the second occasion, he replaced a part
of the sample drawn on the first occasion, while the remaining portion
was retained for matching. 0 n th es econ d occasion, he obtained
two estimators which were correlated. One was the sample mean based
on the new units only, and the other was a regression estimate based
on the units which were kept for matching, and the overall sample mean
obtained from the units observed on the first occasion. He thus
obtained a new estimator by a linear combination of the two estimates
which he claimed to be a minimum variance linear unbiased estimator
of the population mean on the second occasion.
The form of the estimator used by Jessen is
I
Y. =Q y,' + (l-Q) y~2 2u c:;m
where
Y~u =Y2U =the mean of unmatched portion on the
second occasion, and
6
(2.2 )
In (2.2)
Y2m = the mean of the matched portion on the second
occasion.
Ylm =the mean of the matched portion on the fi~st
occasion.
Yl = the mean of the whole sanr,ple on the first
occasion.
" II dThe regression coefficient b is assume to be known.
1
7
Q =1 + 1
1
1 + 1
The extension of the theory to more than two occasions, also
confined to unistage simple random sampling, was made by Yates [20].
Yates presented the estimation theory which may be viewed as a
generalization of Jessen's theory. In addition, in the development
of the estimation theory for more than two occasions, Yates assumed
that the correlation between the same sampling units for the observed
characteristic on two different occasions is of an exponentially
decreasing type, Le., the correlation between observations made on
.,.
82
one occasion apart is p, two occasions apart is p , three occasions
a.part .is p3 and so on. Yates further assumed that the variances and
covariances did not change with time, i. e.
Cov (Yj'Yh ) = Cov (Yj+k' Yh+k )
where j, h, j+k and h+k denote the jth, hth, (j+k)th and (h+k)th
occasion respectively.
Patterson [12] further extended the theory given by Yates and
derives a necessary and sUfficient condition for a linear unbiased
estimator to be a minimum variance estimator 0 Patterson considered
also types of correlation patterns other than that given by Yates.
Similar theory for unistage simple random sampling is also discussed
in [1] and [15] with some slightly different approaches. other
contributors to the theory of sampling on-successive occasions were
Tikkiwal [16], [17] and Eckler [4] who developed specific schemes of
partial replacement of units and presented·the theory relevant to
their proposed sampling plans. Eckler was instrumental in introducing
the term "rotation sam;plingll, actually suggested to him by S. S. Wilks 0
In 1954, Hansen ~ a1. [6] redesigned the Current Population
Survey (c oP 0 So), from which information on employment, unemployment
and other related socio-economic data are compiled monthly. One
feature of sub~samplingwhich has an important bearing on the estima-
tion theory introduced in the new sample involves a scheme of partial
replacement or rotation of sampling units at the last sta.ge~ This
• 9
sampling technique was primarily intended to avoid a decline in
respondent cooperation (which may happen when the same unit is repeat-
ed1y interviewed ) and to reduce the variances of estimates.
For any given month, the C.P.S. sample is composed of eight sub-
samples or rotation groups. All of the units composing a·particular
rotation group enter and drop out of the sample at the same time. A
given rotation group stays in the sample for four consecutive months,
leaves the sample during the eight succeeding months, and then returns
for another four consecutive months. It is then dropped from the
sample completely.
It was in this kind of continuous sample survey that the so-called
"composite estimation procedure ll was first introduced. The composite
estimator used in this survey is of the form:
I' _ I"I( 11 +' ') + (1 1"1).'Ya - "'(, Ya-l Ya,a-l - Ya - lja · -"'(, Ya
where
O<Q<l
y~ is the composite estimate for month a,
Y,' is the regUlar ratio estimate based on the entirea
(2.6 )
\.
sample for month aj
Y~,a-l is the regular ratio estimate for month a but based
on the returns from the segments which are included in
the sample for both months a ~d a-I,
Y~-l,a is the regular ratio estimate for month (a-I) but
made from the returns from the segments which are in-
eluded in the sample in both month a and a-I.
10
The composite estimate takes advantage of accumulated information
from earlier samples as well as the information from the current one
and results in smaller variances of the current estimate and the esti
mate of the change of most of the characteristics of interest. But
the larger gains through the reduction of variances of the estimates
are usually realized for the estimate of the change. However, under
such a sampling scheme, only the within first-stage.unit component of
variance of the estimates is improved while the between first-stage
unit variance still remains the same as in the regular estimate.
This is because the same sample of first-stage units is used in every
month.
Onate [11], in developing multistage sampling designs for the
Philippine Statistical Survey of Households, adopted the ~ame princi
ple. He proposed the division of each sample barrio (a second-stage
unit that corresponds somewhat to a township in the united States)
into a small number of segments (less than 10), and imposed a
specific rotation scheme or a scheme of partial replacement to these
segments. This sampling technique was mainly intended to reduce the
response resistance of panel households. Moreover, Onate developed a
finite population theory for the composite estimator defined in (2.6)
for his specific sampling design. Rao and Graham (14] further extended
Onate's finite population theory for the composite estimator to a more
general pattern. The results presented by them were for uni-stage
sampling.
11
Other than those mentioned above, there were a few other a.uthors
who applied the partial replacement of sampling units technique to
their sampling work, for example, Ware and Gunia [18] presented a
theory which is applicable to continuous forest inventory sampling.
They also considered the problem of optimum replacement when differen
tial costs are taken into account. The method of sampling used by
them, however, was confined to uni-sta.ge simple random sampling for
two successive occasions only.
As mentioned earlier, Tikkiwal [17] in an unpublished (but
abstracted paper) proposed the technique of partial replacement of
first-stage units, but in a theoretical context different from that
treated in this thesis, as will appear in sequel.
Des Raj [3] recently proposed the selection of clusters with prob
abilities proportional to size for sampling over two successive occa
sions and indicated the application of the theory to double sampling.
From the review of literature, it is seen that the investigation
of a theory of multi-stage sampling with unequal selection probabilities
for first-stage units, and incorporating also the partial replacement
of a subset of the first-stage units on each occasion, remains an open
problem.
.. I
12
3 • A PROPOSED SAMPLING DESIGN
3.1. Description of a Method of Partial Replacementof First-Stage Units
Consider the situation when we plan to conduct successive sample
surveys for p occasions. The population U consists of N definable
first-stage units:
t ~, u2
, ~, •••••••••••••••••••••••••••••••• , ~ }
In each ~J there are Ni second-stage units (i=1,2, ••• ,N). Our objec
tive is to use n first-stage units on each occasion and selecting m
second-stage un1ts for interview. If the technique of partial replace:
ment of units is to be incorporated into the above sampling design on
the philosophy that on any two successive occasions, the first-stage
samples each of size n will contain a certain number of common units
which will be available for measuring the change over time (if ariy)
of the population characteristic of interest, the following scheme
of partial replacement of first-stage units can be used:
1. Assuming that the desired proportion of first-stage units to
be replaced after each occasion is IJ., (0 < IJ. < 1), draw a preliminary
first-stage sample of size n + (p-l)lJ.n with replacement from U. (The
probability system for selecting the first-stage units in this process
Will be defined later in this chapter.) Record the order or occurrence
of each unit. Conceivably, we would expect some identical :un:tts in the
preliminary sample since they are drawn with replacement.
2. The first-stage units which occurred from order 1 to nconsti-
tute the sample for the first occasion. Reject the first IJ.n units and
_i
13
retain the next (l-~)n units (as determined by the order of occurrence
of these units) for the second occasion, supplementing those retained
units by the next set of IJ.n units which occurred from order (n+l) to
(n+!J.n). Thus, the required s~le size of n first-stage units is
maintained on the second occasion With the assurance of having (l-~)n
units matched with those of the first occasion. On the third occasion,
reject the next IJ.n units which occurred from or~er (IJ.n+l) to 2~n while
the other (l-lJ.)n units which occurred from order 2~n+l to n+!J.n are
retained. Supplement those retained units by ~n units which occurred
from order n+J,.Ln+l to n+'2lJ.n. Do in a similar fashion for the succeeding
occasions. On the pth occasion, there will be (l-~)n units which are
matched with the (p_l)th occasion plus IJ.n unmatched units which
occurred from order n+(p-2)~n+l to n+(p-l)~n.
Example:
u =t~, ~, ~, 00. 00 • 00 •• , u60 ~
1N = 60, n = 9, p = 4 and IJ. = 3 ·
We have IJ.n = ~ =3, and n + (p-l)lJ.n = 9 + (4-1)3 = 18.
We need to draw a preliminary sample of size 18 from the above
population with replacement and record the order of their appearances.
There will be 6018 possible samples, and, of course, not all of them
are distinct. More generally, there are ~ possible samples. In our
particular example, a possible preliminary sample might be:
14
1 2 U§, 4 u5 6ui4U,2' ~7' ~4' 9
, u42 '~- 8 9 10 11 12 13 14
~3 ' U28 ' u4 ' ~5' u48 ' u19 ,~O
15 16 17 18u4 ' u4l ' u9 ' ~
,
where the superscripts indicate the order of appearance of the unit.
The structure of the preliminary sample and its four const;l.tuent
samples is sun:rma.r.ized in Table 3.1- which speaks for itself.
;5.2. Advantages of the Proposed Scheme of Partial Replacementof First-Stage units
This proposed method of partial replacement of the first-stage
sampling units when the sampling is done in two or more stages, can be
expected to minimize response resistance and other undesirable features
resulting from interviewing the same respondents over and over again.
Although the drawing of the preliminary sample of n + (p-1)lJ.n first-
stage units with replacement does permit the same first-stage units to
appear more than once, the second..stage units (and units at other
succeeding stages if more than two-stage sam.pling is used) can be
expected to be different if they are drawn independently in each re
peated first-stage unit. In sequel we shall discuss three different
methods of drawing second-stage units. Hence, the attempt to spread
the burden of reporting among more.-1!espondents is taken care of.
Another advantage of the proposed partial replacement of first-
stage units is that when these units are certain kinds of big-administra-
tive units such as villages or towns, the problem of encountering
missing units when successive sampling is done, is minimized.
.. r ,. t ... A
Table 3.1. Scheme of sample selection, partial replacement and constituent samples
Order of 1 2 3 4 5 6 7 8 9 10 11 12 13 14. 15 16 17 18appearance
~t drawn ~2 '\7 ~ '\4 u9 ~ '\4 u23 u28 ~ l)5 ~8 '\9 ~O ~ ~l u9 ~
~ample forOccasion --------------r---------------r--------------I .
1 I : r
l)2 '\7 u,: '\4 ~ ~ : '\4 ~3 ~8 \
2--------------r---------------~--------------t--------------:
: u14 ~ ~ l '\4 ~3 '\?8 ~'\ ~5 u48 IL_______________~--------------+--------------:.,..- - - - - - - --r
3 I I I
I ~4 ~3 '\?8 Iu4 ~5 '\8: ~9 '\?O u4 :L--------------t--------------i-·--------------r -------------_.
, f I4 I f1'\ ~5 '\8J ~9 ~O u4 I u4l ~ ~____________________________ L______________
t:
16
3.3. Specification of Probab1.lity System and the Methodof Selecting Sub-Units
To develop a complete multi-stage sampling design which can be
put into operation, the specification of the probability system to be
used in selecting units at each stage of sampling must be made. So
far, we have only outlined the method of partial replacement of first-
stage sampling units without specifying the probability system to be
used in selecting such units.
It is well known that in multi-stage sampling d~signs, the use of
unequal probabilities in selecting first-stage units often leads to
more efficient estimates than the use of equal probabilities.
The first-stage units may be selected with or without replacement
after each draw. However, as already stated, in this thesis the first-
stage units are selected with replacement after each draw because
selection of first-stage units with replacement confers statistical
independence between the units involved, resulting in extensive
simplifications in the estimation of variances and covariances in-
volved in the total variance of the estimator used, the expressions for
which are, to say the least, very long.
In actual applications, these selection probabilities may be
assigned proportional to the sizes of the first~stage units, ~.a., as
measured by the number of second-stage units in each first-stage unit.
(If the first-stage units are villages or towns, the size of these
villages or towns may be measured by the number of households in those
villages or towns.)
17
Next, we will consider appropriate methods of selecting second-
stage units in each selected first-stage unit. For this, a brief re-
view of three well-known methods for selecting second-stage units,
when the first-stage units are selected with replacement, will be
given.
Method I is generally attributed to Sukhatme [15]. In this method,th .
if the i first-stage unit is selected A.i
times, then mA.i
second-stage
units are selected with equal probabilities and without replacement
from that first-stage unit.
Method II is due to Cochran [1]. thIn this method, if the i
I
first-stage unit is selected A.i times, A.i
sUb-samples of size m are
independently drawn with equal probabilities and without replacement
thfrom the i first-stage unit, each sub-sample being replaced after it
is drawn.
Metbod III is due to Hartley and Des Raj [2] as pointed out by
Rao [13]. thIn this method, when the i first-stage unit is selected
A.i
times, a fixed size of thm second-stage units are drawn from the i
first-stage unit with equal probabilities and without replacement, and
ththe estimate from the i first-stage unit is weighted by A.i
•
Comments on the three methods:
Method I: In this method, it is assumed that
(i=1,2, •.• ,N).
,....
That is, the total size of each first-stage unit in the population is
relatively large as compared to the sample size at the second stage
of sampling so that the unfavorable case of drawing the same
\
18
first-stage unit up to A.i
times where mA.i > Ni has an extremely small
chance of occurring.l
In practice, especially in a large-scale sample survey, the above
problem may not arise for two reasons namely:
(i) The number of first-stage units in the population is usually
large so that the chance of drawing the same first-stage unit A.i times
where A.i is curiously large is very small.
(ii) The size of the first-stage units (namely the number of
second-stage units in each first-stage unit in the population) are
usually (or can be made) sufficiently large.
Hence, for the purpose of spreading the burden of reporting among
respondents, this method of selecting second-stage units should fit in
With our sampling design. However, when the above method was intro..
dueed, it was intended to be used only in a survey which was conducted
on one particular occasion. Hence, to apply this method is a successive
sampling scheme such as proposed earlier, some adjustment needs to be
made. In our sampling scheme, we have selected the n + (p-l)~.m first-
stage units with unequal probabilities and with replacement and the
scheme of partial replacement of units is based on these units as
described in detail previously. In that replacement scheme, it is
evident that the same first-stage unit may be selected more than once.
Hence, if we modify the assUIlI.Ption given originally to be such that the
sizes of the first-stage units in the population are sufficiently
large so that
(t = 1,2, ••• ,p)
\
t
19thwhere Ait = the number of times that the i first-stage unit is
th Iselected on the t occasion, then Sukhatme s method of selecting
second-stage units can be used. In practice, the above assumption
should not be unrealistic especially in most large scale sample
surveys.
Method II: This method is apparently free from any assumption
about the size of the first-stage units in the population since the
procedure is to select independent sets of m second-stage units with
equal probabilities and without replacement -every time from the same
first-stage unitwhich occurs more than once. We recall that in this
method, each sub-sample is replaced after it is drawn to cOIqp1ete one
particular set of m second-stage units. However, we may expect the
same second-stage units to occur more than once on a particular occa-
sion. This seemingly unfavorable event should be accepted by the
sampler since it will not increase any operational problem. All we
need to do is to go out and interview that unit once and record the
information, keeping .. the number of repetitions of occurrence as the
frequency or weight to be used in estimation procedure for that partic-
ular occasion. Should the same second-stage units be selected on the
next or succeeding occasions due to the nature of our sampling design,
the sampler also should accept these (unfavorable) events and go ahead
to interview those units according to the number of occasions in which
they are included in the sampling plan. These events should not be
many, and in View of maintaining the rate of response, we may instruct
the interviewers to try their best to explain to such would-be
respondents as to why they are interviewed for several occasions. In
20
conclusion, this method of selecting second-stage units will be adopted
in our partial replacement scheme in preference to Method I where the
assumption
PNi 2: m ( I: ;\'it)
t=l
may not be satisfied.
Method III: This method does not seem to fit into our sampling
design by its very nature, !.~., it will not help in spreading the
burden of reporting among respondents. Hence, we will not consider it.
.'
. 21
4. ESTDvlATION THEORY
4.1. Introductory Remarks
As stated earlier the sampling theory shows not o~ythe sampling
design, but also shows the estimation theory which follows "from it.
We will not consider the estimation procedures which can be made under
the sampling theory to be proposed. It should also be emphasized that
the estimation theory in statistical sampling generally does not de-
pend upon the concept of distribution of random variables. However,
the estimation theory in Statistical Sampling partially relies on basic
criteria in the classical estimation theory such as unbiasedness,
minim.um varianc,)../ or m.inimum mean squar~/, which are known to char
acterize a good estimator. When the two properties are about the, same,
. IIwe may have to add another criteria which by common sense is ease of
computation" so that only one estimator stands out as the best choice.
However, the added criteria has become less significant in the coun-
tries where modern electronic computers are extensively employed.
To develop an estimation theory which will be applicable to the
proposed sampling design, some of the criteria mentioned will be
utilized. The essential notation to be used is as follows:
2./However, Godambe [5] and more recently Kool> [9] have demonstratedthe non-existence of minimum variance unbiased estimators when unitsare drawn with unequal probabilities •
.!t/In the case of an unbiased estimator, these two are the same
since Mean Square Error = Variance + (Bias)2.
•
22
th th= the variate value of the j second-stage unit in the i
thfirst-stage unit on the t occasion.
tT = the total of population characteristic of interest on
ththe t occasion (t = 1,2, ••• ,p).
N = the number of first-stage units in the population.
n = the number of first-stage units in the sample (Le., the
first-stage sample size on each occasion).
Ni the number of second-stage units in the i th first-st~e
unit (i =1,2, ••• ,N).
m = the number of the second-stage units in the sample
(chosen to be equal for every ~irst-stage unit, to fit
with our sampling design) •
Ni . th= ~ t Yi ., the population total for the i first-stage
j=l J thunit on the t occasion.
Pi > 0, (i =1,2, ••• ,N), the probability of selecting the i th
Nfirst-stage unit such that, ~ Pi =1.
i=lS is a set of first-stage units selected in a specified order
and is a subset of the preliminary sample of n + (p-l )lJ.n
first-stage units. The size of this set (sample) will be
clear from the context in which it is used.
Other symbols will be defined where they are used.
In this thesis, we will be concerned in estimating the total of
the population characteristic and the change or difference between the
totals of population characteristics on two successive occasions. If
interest lies in the~ of the population characteristic and the
..
•
change or difference between the means of population characteristic,
the theory can also be used 'With a little change entirely in the
multiplying constant{s) that enter into the expressions for the esti-
mators and their variances.
For the development of the estimation theory, we will adopt the
method of selecting second-stage units as suggested by Cochran [1].
Although Sukhatme's method [15] of selecting second-stage units is not
so complex in principle and can be adopted, it is felt that such a
method requires a restrictive assumption about the size of the first-
stage units and leads to a complicated expression for the variance.
Thus, in the development of the estimation theory that 'Will follow,
whenever a first-stage unit is selected, m second-stage units are
independently selected from that first-stage unit 'With equal ~rob-
abilities and without replacement.
4.2. Estimation of Total for the First Occasion
To estimate the total of the population characteristic for the
first occasion under the proposed sampling design, is a straight-
forward procedure. As a background for comparison with the estimation
procedure for the second occasion, and as other derivations of vari-
ances and covariances have a bearing on these results, it is worthwhile
to include the relevant theory here, despite the fact that it is
already well known.
N NiTo estimate 1T = ~ ~ 1Yij' !.~., the total of the char
i=l j=l
acteristic of interest of the population on the first occasion.
24
An unbiased estimator of IT
\"- 1 n NiT = - .E -
1 n i=l Pi
m.E
j=l(4.1)
We show that IT as given in (4.1) is unbiased. Since two~stage
"-sampling is used, the expected value of IT is easily obtained by
applying the well-known theorem on conditional expectation; n~ely
(4.2)
•
..
where in our context E(·I S) refers to the conditional expectation
given the first-stage sample.
We now apply (4.2) to find
E( T) = E[E ! ~ Ni ~ lYij Is]1 n i=l Pi j=l m
Now
so that after some simplification
whereNi
lYi = j~ lYij •
25
FuI;.ther
\ N= . L: 1Yi = 1T
J.=l(4.4)
So that
,..Hence, IT is an unbiased estimator of IT as claimed.
,..The variance of 1T can be derived by several methods. One m..ethod
which seems to be most convenient for multi-stage sampling is t9 apply
an important theorem on conditional variance formalized by Madow (1949);
namely
We apply (4.5) to find the expression for Var (IT) .
Consider first, Var ~E(lTIS)}. We have
E( Tis) =1 ~ E(Ni ~ lYij ·Is) 0
1 n i=l Pi j=l m
(4.6)
..1= -n2
n
since
= 1 ~ P (1Yi _. T)2n i=l i Pi 1
E (lYi ) T for all i as demonstrated at (4.4).Pi = 1· ,
\
26
Now, con,si<ier the second part of the variance.in (4.5); nameiy,
E Var (lT1s) • Returning to (4.1), we have
(4.8)
It is well known that
where
(Ni
-m)
(Ni-1)
and
1~ ,
so that
and hence
Var ...L2n
\
N Ni 22 (N -m)
E {var (lTla)1 n 10"i i::; T .E Pi (-) - (Ni -1)Pi mn i=l
1 N ~ 1~ (Ni-m)= .En i=l Pi m N -1i
'"Therefore, the expression for Var (IT) is
(4.10 )
..::: 1. [~ p
n i:::1 i
27
(Ni-m) ](N.-1)
J.
(4.11)
::: ~ [the between first-stage units variance + the
average within first-stage units variance].
4.3. Estimation of Total for the Second Occasion
On the second occasion, we wish to estimate
NL:
i=l
We recall that the sample for this occasion is made of the first-stage
units appearing from the (IJ.n+1)th to (n+!J.n)th draw. Of this set of n
first· stage units, those units appearing from the (IJ.n+1)th to the nth
draw constitute the matched part. Those units -appearing from the
)th th(n+1 to the .(n+!J.n) draw constitute the part which replaced the IJ.n
units dropped after the first occasion.
If there is 'a change over time of the characteristics of the
population, then the (l-lJ.)n first-stage units which are kept in common
between the two occasions should serve as a natural measure of such
change. We will consider several possible estimators which can be
used to estimate the total for the second occasion. These estimators
are linear combinations of estimates based on different parts of the
sample and take advantage of past as well as present information from
the sample to provide an estimate for the present occasion (in this
case the second occasion).
28
4.3.1. A Linear Composite Estimator
We represent the structure of our sampling procedure covering the
first two occasions diagrammatically as follows:
1st occasion
IJ.n units (l-lJ.)n units
I t 1\
2nd occasion ! I(l-lJ.)n units IJ.n units
The set of units connected by the two way arrow represents the
matched portion. The linear co~osite estimator for the second occa-
sion is
• (4.12 )
where *2T = the linear composite estimator of the total for the
second Hoccasion:
'"1T = the unbiased estimator of the total for the first
occasion as defined in (4.1)
= 1 ~ ~ ~ 2Yij E
(l-lJ.)n i=fJ.n+1 Pi j=l m
(4.14)
'2./ 29A 1
n+J,J.n Ni m 2Yij2T = - Z Zn i=j..l.n+l Pi j=l m
!.!:.., the unbiased estimator of 2T, based 0 n all of the n first
stage units for the second occasion.
Remarks:
(i) The subscript i in the summations of (4.13), (4.14), (4.15)
refers to the order of occurrence of the relevant first-stage units
which are recorded in the preliminary drawing procedure.
(ii) Although lYij and 2Yij, are not the same, the associated
probabilities Pi are conservative.
1'". , /,'to.
Expected Value Of 2T. It is easy to verify that' 2T'. is
• also unbiased. We have
It can be shown along the same lines as in (4.4) that
Hence
'2./This estimator namely /r would be used if the sampling was
carried out only on~ occasion, or if past information from thefirst occasion is totally ignored.
4.3.1.2.~' ~
Variance of 2T. We recall that
30
...The expression for Vax (IT) has been established in (4.11),
namely:
(4.16)
The expressions for other relevant variances in (4.16) can be...
derived in the same fashion as for Vax (1T) • They are:
(Ni-m) ](Ni-1)
(4.17)
... 1 [. N 2Yi 2 N ~ 20";Vax ( T) = - Z P (- - T) + Z - (-)
2 n i=l i Pi 2 i=l Pi m
,
[
N lYi 2 N ~ 10"; (N1-m)]. Z Pi (7 - 1T) + Z P (m) (Il -1)1=1 1 1=1 1 1
( 4.18)
(Ni-m) ](Ni -1) • .
(4.19)
,31
where in (4.17) and (4.19)
and
= ~i 2Yijj=l Ni
The expressions for the relevant covariances in (4.16) will nowA A
be derived. Consider first, Cov (2,lT, 1,2T).
From the structure of the two estimators we have explicitly:
A
A (1 n Ni m 2YijCov (2,lT, 1,2T) = Cov (l-~)n L: L:
Pi m ,i=~n+l j=l..
1n Ni m
l~iJ J(l-~)nL:
PiL:
i=~n+l j=l
To obtain the expreSsion for the above covariance, we apply the
theorem on conditional covariances formalized by Madow [1 0] which is
as follows.
If U and V are random variables and A is a random event, then
Cov (U, V) = E[Cov (UIA, viA)] + Cov [E(uIA), E(V!A)]
A A-
To apply the above theorem for finding Cov (2,lT, 1,2T), letA A
2,lT play the role of U, 1,2T play the role of V, and S, here the set
of first-stage units selected from (~n+l)th to nth draw, play the role
of A. Then, it is not difficult to show that
~ 2Y
ii=jJ.n+l Pi
Ni, where 2Y' =!: 2Yi'
J. j=l J
and where
t1 n
COY !:. ( I-jJ. )n i=fJn+l
..
This leads to
( )
_ 1 N 2Yi 1YiCOy E(2,lTIS), E(1,2T1s) - (l-fJ)n !: P. (- - T)(- - T).
1=1 J. Pi 2 Pi 1
~ (Ni )2 ~i (2Yij - 2Yi)(lYi {lYi)
i=jJ.n+l Pi j=l Ni
1 (Ni-m)iii (N
i-1)
since second-stage units are selected without replacement in each
selected first-stage unit.
33
Hence
(4.21)
Corribining (4.20) and (4.21), we thus obtain
+ ~ ~ ~i (2Yij - 2Yi)( lYij - lYi)
i=l Pi j=l Ni(4.22 )
We may interpret the result given in (4.22) as follows.
The total covariance is made up of two parts name:Ly, the between
first-stage units covariance and the average within first-stage units
covariance.
We also notice that in our context, the above covariance virtually
measures the auto-regressive·· nature over time of the characteristic
under- .study.
Next, we consider cov-(lT, 1,2T). To establish the expression for
this covariance, we proceed as follows.
•
J
A A
From the structure of IT and 1,2T, we have
This can be rewritten as
A 1 IJ.n Ni mCOy (IT, 1,2T) = COY - (E -- E
n i=l Pi j=1
1 ~ Ni ~(l-';)n P
r- i=J.l.n+l i j=l
= 1 [ IJ.n(1-) -COY ( En IJ. n i=1
By virtue of mutual statistical independence between the first-
stage units u1 ' u2 ' ••• , uJ.l.n and ulJ.n+l' uJ.l.n-t2' •••••••• un' it can be
shown that
IJ.n N mCOY ( E pi E
i=l i j=l, (4.24 )
35
The second expression in the square bracket, namely
~, n Nim lYij n Ni m lYij
Cov( .E .E .E .E )Pi
,Pii=lln+l j=l m i=lln+l j=l m
is readily recognized as
n Ni m lYijVax ( .E .E )
i=jJ.n+l Pi j=l m
"which can bederived in the same way as Vax (1T), and we obtain
n Ni
mCOy (.E -.E
i=J.1n+l Pi j=l
lYijm '
Combining these results, we thus obtain
(Ni-m)}(Ni-l) •
(4.25 )
(4.26)
f
,.The very nature of COy (l~\ 1,2T) as revealed by (4.26) shows
that it is quite different from COy (2,lT, 1,2T).'-'
The expressions for other covariances involved in Vax (2~) will
now be derived along the same lines. By definition,
.,
tl {lJ.n Ni m lYij n Ni m lYij }= Cov.- ~ -- ~ + ~ -- ~ ,n i=l Pi j=l m i=lJ.n+l Pi j=l m
Using the same procedure as in the development of the expression
'" ....for COy (1,2T, 2,lT), it will be found that the second part, namely
~ nNi m lYij n Ni m
2:1JJCOy ~
Pil: , ~
Pi~
i=j..m+l j=l m i=lJ.n+l j=l
(4.28)
37
Now consider
r~nN
im lYij m N
im
2:iJ JCov L:Pi
L: , L:Pi
L:i=l j=l m i9J,n+l j=l
Iln n iNi m lYij N.e ~ 2mY.eJ] 6/
= L: L: Cov pL:. , P "i=l .e=lln+l i j=l m .e j=l
Again applying Madow's theorem, we find
·\Ni m lYij N.e ~ 2Ym.eJ}Cov P L:, , P "
i J=l m .e j=l
= Cov rE (:i ~ lYij I 8) ,L i j=l m
..'
( Ni
mlYiJ 18 N.e m 2Y.ej
I SJ .+ E Cov (Pi L: P.eL:
j=l m ' j=l m
The first part is
t Ni m lYij N m
2Y.ej I 5)Cov E(Pi
L: I 8 ), E(..1. .Ej=l m P.e j=l m
tlYi 2YJ} N N lY' 2T,e
= Cov Pi ' = .E L: Pi P (~ - IT)(p- - 2T)P.e i=l .e=l .e Pi ,e
2/It may be noted that the derivation here is diff~rent from thatn Ni m lY1J n N1 m 2Yij
of COV[ L: p ~ m ' L: P I: m ]; this is due to thei=lln+l 1 j=l 19J,n+l i j=l.
fact that the latter is the planned matched portion. The relevant probabilities are quite different; this should be apparent when one recallsthat the index i in the summation signs refers to the order of theoccurrence of the first-stage units. -----
38
For the second part, consider the covariance expression under the
expectation sign, we find, as in the derivation leading to (4.25) that
[Ni m lYij N,e m 2Y
m,ej I S)~E Cov (- !: . Is, -!: .
Pi j=l m P,e j=l
= 0
and hence
Cov {~n Ni ~i=l Pi j=l
~ 2~ijJ =j=l
o •
Similarly,
A A [1 n Ni m 1Yij 1 nT)J.n Ni ~ 2Yij]Cov (1T, T) = Cov - I: - I: , - I: - L..
2 n i=l Pi j=l m n i9.ln+1 Pi j=l m .
39
1Yijm ),
n Ni m 2Yij n+i-!n Ni ~ 2Ymij)J(I: pI: m + I: P L..
i9.ln+1 i j=l i=n+1 i j=l
_ .1:. [- tJ..Ln Ni m 1Yij- 2 -Cov I: P I: m '
n 1=1 i j=l
t n Ni m 1Yij n Ni m 2YiJ... ]+ Cov . I: - I: , I: - I:
i=J..Ln+1 Pi j=l m 1=J..Ln+1 Pi j=l m -
\ ~nNi m 1Yij n+i-!n Ni
m2~iJJ+ Cov I:
~I: , I:
PiI:
i=l j=l m i=n+1 j=l
{n N m 1Yij n+i-!n Ni m27r J]+ Cov I: pi I: I: I: J •
m,
Pi j=l mi=J..Ln+1 i j=l i=n+1
The ~, third a.nd fourth covaria.nces can be shown to be zero,
and we are left with
1 [ (n Ni= - Cov I: -n2 i=J..Ln+1 Pi
mI:
j=l
lYijm '
..'
~ 2:ij J]j=l
~J
40
Apart from constants, we find as in the derivation leading to (4.29)
that
(Ni-m)l
(Ni-l)}
(4.30)
'" '"The same procedure when applied to Cov (2,lT, 2T) and
'" '"Cov (1,2T, 2T), yields
and
2 N lYi 2By writing leTby for E Pi (p - 1T) ,
i=l i
N ~2
(Ni-m)2 l eTileTwy for E
Pi (Ni-1) ,
i=lm
(4.32)
.' 41
(Ni-m)(N
i-1) ,
and
we arrive at more concise expressions involved in the variance function
given by (4.16), viz.
COy (T T) = (l-~)n1 ' 2 2n
In terms of the notation given by (4.33) we find from (4.16)
Var (2~) - Q2 [1 (0:2 + 0"2 ) + 1 (0:2 +0"2 )- n 1 by 1 wy (1-1l )n 2 by 2 wy
42
1 2 2 2+ "'(l:---=Il~)-n (lO"by + 10"WY) + Ii (1.20"byy + 1.20"wyy)
_ g (0:2 + 0"2 )... 2 . ( 0: + 0" ~2 1 by 1 wy (l-ll)n 1.2 byy 1.2 WYYJ
+ (1_Q)2 [1 (~ + 0"2 )] + 2Q(1-Q) [(l-ll)n ( 0:n 2 by 2 wy n2 1.2 byy
+ 0" ) + 1 (0:2 + 0"2 ) _ 1. ( 0: + 0" )].1.2 wyy n 2 by 2 wy n ,1.2 byy 1.2 wyy
(4.34 )~
4.3.1.3. Note on the Estimation of Var (2~)' The expression forA.
Var (2T) involves the constants Q, IJ. and n and the population
2222+values of 10"by + 10"WY' 20"by + 20"WY and 1.20"byy 1.20"wyy· An un-
Abiased estimate of Var (2T) can be made by substii:;uting unbiased
. 222 2estl.mates of 10"by + 10"WY' 20"by + 20"WY and 1.20"byy + 1.20"wyy
respectively in (4.34). By virtue of our sampling design, it can be
shown that
2 2 n 1Y~ A. 2(i) 10"by + 10"WY is unbiasedly estimated by ~ (p - 1T) / (n-1),
i=l i
where
A _
2T and 2Yi are as defined previously.
...
The verification is given in the Appendix •
4.3.1.4. *Efficiency of 2T.
43
To see the usefulness of the linear
and
1Vy for 2 210"by + 10"WY ,
2Vy for 2 220"by + 20"WY ,
V for ~ + 0" ,1.2 Y 1.2 byy 1.2 wyy
to obtain more compact expressions for variances.A
The form of 2T has been given previously at (4.15), namely
A = 1 n~n Ni~T ~ Pc;. n i=lln+1 i
mr:
j=l
with variance given in (4.33), namely
= 1. (0".2 + ,; ) =n 2 by 2 wy
Using the above symbols in (4.34), and after some algebraic
manipulation, we find
1/This estimator would be used if no attempt is made to utilizethe prior information from the first occasion or if the sampling isconducted only for one occasion.
44
tVar
...
Var (2~) = Var (2T) + l~) [1} + 2:1- 2Q;t ~ + (l~J l.~Vy(4.35)
To make a meaningful comparison between the efficiency of 2~ and
"2T, we will make the following assumptions:
(i) V = V = V ~/ (say)1 y 2 Y Y
V(ii) p = 1.2 y
-V;V;~
lYi and 2Y
ip is the correlation betweenPi Pi
Under the above assumptions, (4.35) yields the relation
~ A 2Q2 V [Q 1 VVar (2T) = Var (2T) + (l~~) -! -2Q~ .1 + (l~~)J p -!' .
~/ThiS assumption is referred to, in time series analysis, assecond order stationarity or weak sense stationarity.
.,
tJ
~When p = 0, the expression for the variance of 2T is
This implies that in such cases (p=O), the linear composite
estimator will be less efficient than the regular unbiased estimatorA
2T. This should be no surprise to the sampler because in such cases
(i.e., no correlation over. time), we would not spend any effort using
the prior information from the first occasion in estimating the
characteristic of interest for the second occasion.
In actual application, we would expect that ~he correlation p is
sufficiently high so that estimation by a composite estimator is worth-
while. For this, we will now examine the nature of the gain in effi
ciency of 2~ over /r for some realistic values of p}.1 Since the~
nature of the gain in efficiency using 2T, also depends on the wei¢ht
Q and the proportion of first-stage units replaced namely IJ., we
consider such gain for some sets of values of Q and IJ. which ~ be of
. . 101practical interest.---
21The estimate of p can be made from the samples. The formula isI I
n {'i A- 2Yi A-
I: (p- - 1 2T) (p- - 2 IT)= i=pn+l i' i'
p II ~ (lY~ - T)2 ~ (2Yi _ T)2]i=jJ.n+l Pi 1,2 i=pn+l Pi 2,1
~IThe optimum value of both Q and IJ. will be discussed later.
Case I:
Here
~ AHence, it will be seen that 2T will be more efficient than 2T
if [~ - if. p] < 0 ,
4!.~., if p > "7 <= ·57 .
The per cent gain in efficiency of 2~ over 2T as measured by
46
is
The gain as a function of p is tabulated for some values of p
in Table 4.1.
Case II:
Here
Var <2~) =Var <i) + [it -~ p]:t ·We see again that 2~ will be more efficient than 2T if
!.~., if p > ~ = .60.
j
...
47A A
The percent gain in efficiency of 2T over 2T in this case is
( ) _ (5p-3)G2 P - l5-5p x 100.
The gain is also tabulated in Table 4.1.
Case III:
Here
(2~)A [~- t p] V
Var = Var (2T) ... .Ln ,
~ A
and thus 2T is more efficient than 2T if
!.~., if P > .67 .~ A
The percent gain in efficiency of 2T over 2T in this case is
The nature of the gain for some values of p is also tabulated in
Table 4.1.
4.3.1.5. Choice of Q in the Linear Com,posite Estimator. The
lit "gains in efficiency of 2T over 2T considered previously are for
1some specific sets of values of Q and IJ.. The weight Q =2' is
selected because it appears to be the most natural choice to begin
with. The replacement rates of' IJ. = t ' ~, and ~ are the most
natural to choose and perhaps will be of' practical interest.
,. ,.Table 4.1. Percent gain in efficieney of 2~ over 2T
.- ;
Case I: p~/=.6 p=·7 p=.8 p=·9 p=·95
1Q=-2< 1.00 3·90 7.14 10.60 12.41
1f.L = '4
Case 2: p=.6 p=.7 p=.8 p=·9 p=·95
1Q=-20.00 4.35 9·09 14.29 17·07
. 1f.L =-
3
Case:5 : p=.6 p=.7 p=.8 p=·9 1'=.95
1Q=-2negative 2.56 11.11 21.21 26.98
1f.L=-2
!:/Note: the number under each value of p is thecorresponding percent gain in efficiency.
48
~
From the expression for Var (2T),we will see that apart from the* ,.
value of p, the nature of the gain in efficiency of 2T over 2.T,
generally depends on both the values of Q and fJ.. In practice, the
sampler may not have enough freedom to choose the value of fJ. and may
have to choose the one that fits best with practical conditions of the
problem such as the costs ,of the survey, etc. Under suchconditions,
it becomes necessary to determine the best choice of the weight Q to
be used in the composi"tie estimator (4.12.) so that maximum gain in
efficiency (for a fixed fJ.) is realized.
We will now proceed to determine the optimum value of Q for some
specific replacement rates (!.!.,
practical interest, and the gains
realized.
Case I:
1fJ. = 4'
for some values of fJ.), 'Which are ·of* ,.
in efficiency of 2.T over 2.T therein
ItThe variance of 2.T given in (4.36) can be rewritten as
Var (2~J • [1 + 2Q2~ + (2Q1..e -(i~)- 2QVJ p] 1When fJ. =~, the above expression for Vax (2') simplifies to
[ 1 +~ - (~ +~ Jp] 1·For fixed n and V I the optimum va1ue of Q is given by
[
2 ,Y 2]V~ :1 + 2§ - (~ + f)p -f = 0
50
!.~.,
vJ. = 0n '
and we find
Q - -.2E- 8-2p
Table 4.2 shows the optimum values of Q and the percent gain in
efficiency in this case for some values of p , which are likely to
occur in practice.
Table 4.2.
p
.6
·7
.8
·9
·95
1Optimum value of Q for ~ =4 and percent gain in efficiency~ ,..
of 2T over 2T
Optimum value of Q!! %Gain in efficiency
.26 4.14
.32 5·90
037 8.11
.44 10.85
.47 12.45
!/Rounded to the nearest two digits.
Case II:
In this case, (4.37) simplifies to
[1 + '12- (23'1 +~) pJ ~ .
~
Minimizing Va:r (2~)' we find that the optimum value of Q = (3-P)
51
The optimum values of Q as function of p and the percent gain
in efficiency for this case are tabulated in Table 4.;.
Table 4.;. 1Optimum value o~ Q for ~ =3 and percent gain in
efficiency of 2T over 2T
p Optimum value of Q %Gain in efficiency
.6 .25 5.26
·7 .;0 7.6;
.8 .;6 10.74
.9 .4; 14·75
.95 .46 17·14
Case III:
1IJ. =-:2
In this case, (4.;7) reduces to
[1 + 2Q2 _ (Q + Q2 lP] ~ ,*Minimizing Var (:2T) results in the optimum value of Q = 4-~p •
The optimum values of Q as a function p and the percent gain in
efficiency for this case, are tabulated in Table 4.4 •
..2e.... 1We see that the optimum values of Q which are 8-2p for IJ. =4 '
;~p for IJ. =~ , and 4-~p for IJ. =~ approach ~ as p tends to
unity.
J..
Table 4.4. 1optimum value of Q for ~ =2 and percent gain inA A
efficiency of 2T over 2T
52
p optimum value of Q %Gain in efficiency
.6 .21 6.87
·7 .27 10.40
.8 ·33 15·33
·9 .41 22·56
·95 .45 27·23
1Hence, in practice, the choice of Q =2 should be satisfactory
when the correlation p is high, meaning any value between .9 and 1.
The deviation from maximum precision of the linear composite estimator
2~ will be small but considerable reduction in computational work is
achieved. However, for a low value of p, the sampler should try to
use the exact optimum value of the weight Q so that the entire
complex sampling design will be worthwhile.
4.3.1.6. Choice of ~. In the previous section, we have
considered the optimum value of Q for some specific values of ~,
1and found that for a high value of p, we may use Q =2 as the weight
in the linear composite estimator without sacrificing much efficiency
and also avoiding tedious computations. Another problem of interest
is to determine what is the best choice of ~ (the replacement rate)
when Q = ~ is used. Further, how would the gain in efficiency change
from the use of optimum value of ~ as against natural values of ~
such as ~ = ~ , ~ , ~ which have been considered previously.
.J
53
We proceed to determine the optimum value of I.l. as follows:
1Substituting Q =2' in (4.37) and after some simplification, we
have the following variance function for 2~
~2-~li:~r~ ~
Setting ~~ [(2-~Jfi:nl) ~] = 0
leads to the following quadratic equation in I.l.:
PI.l.2 - 2pI.l. + 2p - 1 = 0 •
Solving we find
= 1 + ""p(l-p)I.l. p
(4.38)
"
Since 0 < I.l. < 1, we see that the only admissible root of (4.38)
is
which is the required optimum value of I.l..
Table 4.5 shows the optimum value of I.l. as a function of p and. ~ A
the percent gain in efficiency of 2T over 2T •
.We see from Table 4.5 that when p is high, there is an
appreciable increase in the gain in efficiency. However, the optimum
value of I.l. may not be feasible when p is as high as .95 simply
because it involves replacing over three-fourths of the first-stage
1units. In such case, the choice I.l. =2' might be used without
sacrificing much efficiency.
Table 4.5. 1Optimum value of lJ. for Q =2" and the percent gain in
* "efficiency of 2T over 2T
54
p Optimum value of lJ. %Gain in efficiency
.6 .18 1.02
·7 .34 4.30
.8 ·50 11.11
·9 .66 25·00
·95 ·77 39·39
Such choice (i.e., 111 = 2" ) should serve well both from the point
'(
of view of the sampling operation and ease of computation.
4.3.1.7. Simultaneous Optimum. Values of Q and 11 in the Linear
Composite Estimator. The optimization so far considered is for Q
and 11 separately, !.!:.., the optimum. value of Q was determined by
holding 11 fixed and vice versa. In practice, one might be interested
in searching for the best combination of Q and 11 so that the linear
~composite estimator 2T as defined in (4.12) will have the smallest ..
variance or equivalently, the gain in efficiency using the linear
*composite estimator 2T will be fully realized. We will now consider
the simultaneous optimum. values of Q and 11.
We recall that
" [1 + 2Q2( 2Q2 2QIl - 2Q21l2) p] V
Var (2T) 11 + ..Jl.- (1-11 ) (1-11) n
2e C1 (Q"g.) . n ' say.
J
55
For fixed values of n and V, the simultaneous optimum valuesy
of Q and ~ maybe defined as the pair of values of Q and ~
which minimize Cl(Q,~) defined in the above identity.
Setting
we have
and
2Q2 [(l-"ll:"~~-l~ + [(1-" lt4Q)! - 2Q - ¥<2" ]
- (2QJ12
- 2QJ1 - 2Q2"2) { -1 )] (1_~)2 = 0
and noting that 0 < ~ < 1 we find, from (4.39 )
and from (4.40)
Q = (p _ 2p~ + p~2)
(1-2p~ + p~2)
From (4.41) and (4.42), we have
(4.39 )
(4.40)
(4.41)
(4.42)
... 56
This leads to a cubic equation in ~, namely
(1l-1) (P1l2
- 21l + 1) = o.
Since ~ f 1 , we have
2pll - 21l + 1 = 0,
yielding the roots
Il = (4.44)
.....
The admissible value of ~ must be
1 - .y 1-pP
(4.45 )
Substituting the value of ~ from (4.45) in (4.41) we obtain
(4.46)
The form of Q is quite suggestive in that it approaches ~ when p
approaches 1.
It can be verified that
~ = 1 - -v;;:;;p
and
,
4.3.1.8. Expression for the Minimum Attainable Variance of the~
Linear Composite Estimator 2T. If the simultaneous optimum values of
Q and ~ are actually used, from (4.37) we have
57
t·= [1 + 2a(1- -{l:;)} 2. [1 - ..-P;}
1 ... (1 ... ~),p
p
1- (1- -y:;:p)p
upon simplification, the expression is
~ [1 "'~J2J VVar ( T) = 1... - {1 - V1-p ..:Y-2 opt. 2 n
Or equivalently,
v..:Y-n
(4.48)
.... ....The percent gain in efficiency of 2'2 over 2T for the
simultaneous values of Q and IJ. as given by (4.46) and (4.45) 1s
58
The simultaneous optimum. values of Q and fJ and the percent~
gain in efficiency of 2T over 2'£' when such values are used, are
tabulated in Table 4.6.
Table 4.6.~ A
Percent gain in efficiency of 2T over 2T for
simultaneous optimum values of Q and fJ
p Optimum fJ Optimum. Q %Gain in efficiency
.6 .61 018 7·25
07 064 023 11.43
.8 069 028 18.05
.9 076 .34 30·55
095 082 .39 ·43.67
Comparing the nature of the gain in efficiency when the
simultaneous optimum. values of Q and fJ are used, With the previous
cases, we will see that there is an overall increase in the gain from
every case considered previously. When the exact optimum. values of Q
and fJ are not too convenient to be used, the sampler may adopt the
1 1choice of fJ:= 2 and Q:= 2 with a slight loss in efficiency. But
this is compensated by a simpler replacement procedure. A better
1strategy is to use fJ:= 2 (1.!.0' replace half of the first-stage
units on the second occasion) .and use the optimum. weight Q corespond-
ing to the p-value applicable, given in Table 404. This may be done
when it is evident that the correlation p is between 0.6 and 0.8.
1 1When p > 0.8, the choice of fJ =2 and Q = 2 still produces a
substantial gain in efficiency and may be adopted for simplicity.
59
4.3.2. A Modified Linear Composite Estimator
From the proposed sampling design, another estimator that bears a~
close resemblance to the linear composite estimator 2T may be
constructed. The structure of such an estimator which we Will refer
to as a modified linear composite estimator is similar to the limiting
form of Jessen's estimator [8] and to the estimator recently considered
by Des Raj [2]. It is
~
Q[lTA
1,2T] + [1 - Q]A
T' = + 2,lT - 2,2T2
where
A 1 nTJ.,Ln Ni m 2Yi2,2T =-- L;
PiL;
lln i=n+l j=l m (4.50 )
A
1,2T and Q are as~
2T' is unbiased.
!.!., the unbiased estimator of 2T, but based on those first-stage
units which are selected to replace the lln first-stage units asA A
described in the sampling design. IT, 2,lT,
defined preViously 0 It is easy to verify that~,
Intuitively, the estimator 2T will have a less complicated
expression for its variance than 2~. This is by virtue of ourA A ,A
sampling design. 2,2T is statistically independent of IT, 2,lT,A A
1,2T and hence there will be no covariance between 2,2T and each of
the three just mentioned. However, we should expect that the ~ppro-
*priate weight Q used in the construction of 2T may not be appro-~,
priate for 2T. We will consider the nature of the gain in efficiency
~Iof 2T for some specific values of II which might be of practical
interest.
• 60A
For this, we first derive the expression for the variance of 2~'.
From (4.49), we have
A
Var (2T') = Q2[var (IT) + Var (2,lT) + Var (1,2T)
A A
- Cov (1 2T, 2 2T)] •, ,
It can be shown by a method similar to and leading to (4.25) that. - ........ A ....
the last three covariances namely Cov (IT, 2,2T), Cov (2,lT, 2,2T) ,.... A
Cov (1,2T, 2,2T) are each zero, so that we find
(4·51)
Using previous notation, it can be shown that
(4.52 )
(4.52) together with (4.33) leads to the expression
~, 2rl ) 1 () 1 1Vax (2T ) = Q In (lVy + (l-~)n 2Vy + (l-~)n (lVyj
+ 2Q2[~ P-V<lVY)(2Vy) - ~ <IVy)
- (l:~)n P-V< IVy) (2Vy) J+ (1_Q)2 ~~ (2Vy)'
Under the assumption that lV = 2V = V ,(4.53) reduces toy y Y
V (~i) = [Q2~2<l-2j?) + ~(2Q-l) + (Q-lll '!.z.ax 2 ~(l-~) J n
61
~We will first consider the efficiency of 2T' as compared to the
original lineax composite estimator 2~ when the weight Q = ~ is
111used for the cases: ~ = 4" 3' and 2' respecti,vely.
Case I:
In this case, (4.54) leads to a relation
Var <2~') =[1 +~ - %pJ t= Vax ( ~) + [5 - 2pl ~
2 . 12 J n
> Vax (2T) for all 0 < P <1 •
p >
A ::=:It has already been demonstrated that Vax (2T) > Vax (2T) for
~ ~
.57. We see that in this case, 2T' is less efficient than 2TA
and also less efficient than the simple estimator 2T.
62
Case II:
In this case, (4.54) leads tO,a relation
'!.zn
\
Since we have shown that Var (2T) > Var (2~) for p >.6, we see
again that in this case 2~' is less efficient than 2~ and also lessA
efficient than the simple estimator 2T.
Case III:
In this case, (4.54) leads to a relation
Var ( ~i ) = [1 + (1 - 2p ~ ~2 4 . J n
= V ( ., TA ) + (1 - 2p) '!.zar 2 4 n •
* AHence, it is clear that Var (2T') will be less than Var (2T)
1when p > 2' .~
To see how the nature of the gain in efficiency using 2T' changes
*when it is used instead of the original linear composite estimator 2T,
* Awe compute the percent gain in efficiency of 2T' over 2T which for
.....:.,
this case is
G ( ) = (2p-l) x 1005 P 5-2p
The gain for this case is tabulated in Table 4.7 for some values of p.
I. ~, " 1Table 't.7. Percent gain in efficiency of 2'.1.' over 2T for fJ. =2' '1Q=-2
p 10 Gain in efficiency
.6 5.26
.7 11.11
.8 17.64
.9 25·00
·95 29·03
*Comparing with the gain in efficiency using 2T as given in
Table 1, we see that for each corresponding value of
higher gain in efficiency.
*T'p, 2 shows a
'"To get the idea about the full potential of 2~' when we are
free to choose the weight Q so that maximum gain in efficiency for
fJ. =~ , ~ and ~ is realized, we will now consider using the optimum Q
for each case and compare the nature of the gain in efficiency to thatA
of 2T.
Case I:
64
From (4.54)
Minimizing Var
= [(17 - 2p) Q2 _ 8Q + 41.::z .l 3 J n
~
(2T' ), with respect to Q, results in the optimum
value of Q given by
The optimum values as a function of p and the percent gain in
*efficiency of 2T' over 2T when optimum Qr S are used are- tabulated
in Table 4.8.
Table 4.8.
Case II:
~Optimum value of Q to be used in the estimator 2T for
1 A
J..l. = 4 and the percent gain in efficiency over 2TI
P Optimum Q %Gain in efficieacy
.6 .76 3·94
·7 ·77 8.33
.8 .78 13·23
·9 ·79 18.75
·95 .80 21.95
1J..l.=,
In this case~(from (4.54)),
Var (2~') = [~(5-P) - (6Q-3~ ;:
The optimum value of Q is found to be
Opt. Q := ~ •
Table 4.9 shows the nature of the optimum values of Q and the
65
AAt
gain in efficiency of 2TA
over 2T •
Table 4.9.
Case III:
*Optimum value of Q to be used in the estimator 2Tf
when ~:= ~ and the percent gain in efficiency over 2T
p Optimum Q %Gain in efficiency
.6 .68 4.76
·7 ·70 10.25
.8 ·71 16.67
.9 ·73 24.24
·95 .74 26.58
1f.l.=
2
In this case, (4.54) gives
Vex (2~') = [,,2(5 - 21» - IjQ. + ~ ~The optimum value of Q is
2Q := (5 - 2p)
We again tabulate the optimum values of Q and the percent gain
in efficiency in Table 4.10.
. Table 4.10.~
Optimum value of Q to be used in the estimator 2T'1 A
when ~ =2 and the percent gain in efficiency over 2T
p Optimum Q %Gain in efficiency
.6 ·53 5·55
·7 ·55 12·50
.8 .59 21.43
·9 .62 33·33
·95 .65 42.85
66
~Comparing the nature of the gain in efficiency to that of 2T in
Tables 4.2, 4.3, and 4.4, we see that except for p = 0.6, 2~' willlit
produce higher gains in efficiency than 2T (when the corresponding
optimum value of Q is used in each case.)
If the simultaneous optimum values of ~ and Q are required,
~they can be determined by minimizing Var (2T') with respect to Q
and ~ in (4.54) !.~. minimizing
[Q2~2(1_2e) + ~(2Q-l) + (Q-l)21C2(Q,~) = ~(l-~) ] .
Except for some tedious algebra, the procedure is straightforward
and will be omitted.
4.3.3. A General Linear Estimator
Another type of linear estimator which is relatively more general,
can also be constructed to estimate 2T, the total of population char
acteristic of interest for the second occasion. This type of estimator
", 67
was introduced by Hansen et al. [7) in their book and will be referred
to as a general linear estimator. It is given by
(4.55 )
where
-!...Iln N. m lYi,,\ J. (4.56 )T = L: L:1,1 Iln i=l Pi j=l m
is the unbiased estimator for IT based on the first-stage units"A A
appearing from order 1 to Iln, 1,2T, 2,lT and 2,2T are as defined in
(4.14), (4.13) and (4.50) respectively, and a, b, c, d are appropriate
constants to be determined so that 2T* is an unbiased estimator with
the least possible variance. The following diagram is helpful in
.understanding the structure of this estimator:
A A
1, IT 1,2T
,. - ... ~
1st occasion I: jJ.n (l-Il )nt
2nd occasion (l-Il )n Iln
..'" ~
--- -A A
2,lT
2,2T
4.3.3.1. Determination of Constants. As will be evident from the
structure of the above diagram, this estimator 2T* utilizes in a
different way the past and present information available from the
proposed s~ling design.
68
We have
"'* [ 1 f..Ln Ni m lYi oJT =a - L: - L: £.bl2 f..Ln i=l Pi j=l m
and, as noted earlier, the index i in the sunnnation signs refers to
the order of occurrence of each first-stage unit in the sample.
"'*First, we require that 2T shall be unbiased. For this we must
have
i.e.
so that we must have
(a + b) = 0
and
(c + d) = 1
giving
b = -a and d = (1 - c).
Under this restriction, we can now write (4.55) as
(4.58)
(4.60)
69
Next we search for the best combination of values of a and c- -"*in (4.58) so that the estimator 2T will have the least possible
variance. Bearing this in mind, we proceed to determine the best
values of a and e.
From (4.58) we have
Var (2T*) = a2
[var (l,lT) + Var (1,2T) - 2 COY (l,lT, 1,2T)]
2 " 2 "+ C Var (2,lT) + (l-c) Var (2,2T)
+ 2 ac [COY (l,lT, 2,lT) - COY (1,2T, 2,lTj
+ 2a(1-c) [COY (1,1T, 2,2T) - COY (1,2T, 2'2T~
+ 2c (l-c) COY (2,lT, 2,2T) •
In (4.59) it can be shown that
" AThe following five covariances in (4.59), viz. COy (l,lT, 1,2T),
COy (l,lT, 2,lT), COy (l,lT, 2,2T), COy (1,2T, 2,2T) andA ...
Cov (2,lT, 2,2T) can be shown to be zero.
Using some of the previous results from (4.33) we find
Var (2T*) = a2 [~~ (19 + (l-~)n (lVy~ + c
2 (l-~)n (2Vy)
+ (1_c)2 ~ (2vy) - 2ac (l:~)n (1.2Vy)
70
By collecting terms and si~lifying, the expression is
2 V [2 ]v (V)Var ( T*) = . a (~) + (c-l) + 2c~-~ ~ _ 2ac. 1.2 y •2 ~(l-~) n ~(l-~) n (l-~) n.
(4.61)
Setting the f'irst partial derivatives of' Var (2T*) with respect
to a and c to zero, we f'ind after si~lif'ication
~ (IVy) - c (1.2VY) = 0
(C-l)~+/-l) IVy - a (1.2Vy) = 0 •
Solving (4.62) and (4.63) f'or a and c, we obtain:
(4.62 )
where
a =
c =
2~/-lp - /-l P ~2 2 V(l-/-l p) 1 y
(l-/-l)2 2)(l-~ p
(4.64)
as previously def'ined.
To justify that the above values of' a and c minimize Var (/i*),we consider
02 t ~) -2 (1.2Vy)
B = aadc Var (2T*) = (l-/-l ) n ,
02 l ~ J 2 V.- A = oa2 Var (2T ) =
/-l(l-~ )(U)
n,
02 lVar (2T*)}2 V
C = = (~)oc
2 /-l(l-~ ) n
~ 71
Clearly A > 0, C > 0 for all IJ. > 0 and < L
We see that
4 [( V ~2_ (1V;)(2VYJ= 2 2 1.2 Y. 1J.2(l-IJ.) n
= 4[ ~p2 - , } (lVY)(2Vy)J
(l-IJ. )2n2
Since B2
- AC < 0 will imply that a and c given by (4.64)
and (4.65) minimize Var (2T*), we see that the above condition is
satisfied when
2 1 < 0P - 1J.2
or when
or
p < 1IJ.
..-Since 0 < IJ. < 1, it is evident that the above condition is
always satisfied.
72
Having obtained the optimum values of the constants a andc,
we can now write the exact form of what may be called a best unbiased
estimator T* as2
[ A AJ (l-Il)1,lT - 1,2T + 2 2
(l-Il P )
In terms of the observations, the form of this estimator is
(l-Il ) [1 nNi m 2:iJJ+ r:2 2 (l-Il)n i';n+l Pi(l-Il P ) j=l
2 [1 ni;Ln Ni m 2:q ]+ Il(l-p.p ) - r: r:2 2 Iln i=n+l Pi(l-Il p ) j=l
(4.66 )
73
4.3.3.2. A Comment on.the Form of the Estimator 2T*. The
estimator /i* requires knowledge of the values of 2Vy and lVy
and p • This might not be so handy in practice especially when the
values of 2VY and 1VY are not known and must be estimated from the
sample itself. (Usually in large scale sample surveys, the computation
of the sampling variances, if ever they are made, are made long after
the computation of the estimates of the population characteristics of
interest.) However, under the assumption that 2V = lV =V , they y y
estimator will involve only the value of p whose value may be
determined by past experiences or by judicious guessing.
4.3.3.3. Efficiency of the Estimator 2T*. Using the weights a
and c given by (4.64) and (4.65) in (4.61), we have
Var
V1.2 Y
After some simplification, the expression for the variance of
T* is2
Or in a more suggestive form:
A
We will now compare the efficiency of 2T* with 2T for some
values of 1..1. which might be adopted in practice.
Case I:
11..I.=lj:
In this case, Var (2T*) given by (4.61) is
[2 ] V'
Var ( T*) = .1 M 3p ~2 16 ,2 n.p -
= Var ( T) M .2L. (2Vy1 .
2 16Mp2 t n j
A* A
The percent gain in efficiency of 2T over 2T is
2G (p)= (3p ) x 100
6 16;'4p2
and is tabulated in Table 4.11.
A* A 1Table 4.11. Percent gain in efficiency of 2T over 2T for 1..1.. = lj:
p %Gain in efficiency
.6 1.42
·1 10·55
.8 14.28
·9 19.04
·95 21.81
14
75
Case II:
1IJ.=-
3
In this case,
Var ( T*) = [1 _ 2P2J. 2V
y2 9 2 n-P
.....* .....and the percent gain in efficiency of 2T over 2T is
2G (p) = (~) x 100 •7 9-3p2
These values are tabulated in Table 4.12 below •
Table 4.12.
Case III:
.....* ..... 1Percent gain in efficiency of 2T over 2T for IJ. = 3
p %Gain in efficiency
.6 9·09
·7 13·01
.8 18.08
·9 24.66
·95 28.62
1IJ.=
2
In this case,
A.. [ 2l 2V.. yVar (2T'ft") = 1 - 4:p2j n
76A* A
The percent gain in efficiency of 2T over 2T for this case
is
We again tabulate these values in Table 4.13.
Table 4.13.A* -, A 1
Percent gain in efficiency of 2T over 2T for ~ =2
p %Gain in efficiency
.6 10·98
·7 16.22
.8 23·53
·9 34.03
·95 41.11
,
A*4.3.3.5. Optimum Value of ~ when the Estimator 2T Is ~sed. It
is of some interest to determine the optimum value of ~ !.~., the
proportion of the first-stage units to be replaced after the first
occasion, and "to see how the gain in efficiency changes when the optimum
~ is used against natural values such as ~, ~ , ~. In (4.67),
setting the first partial derivative of Var (2T*), with respect to ~,
equal to zero we have
[(1_~2p2)(_p2) _ (1_~p2)( _2~p2 )] ~ = 0,(l_~2p2)2 n
or
2 2p ~ - 2~ + 1 = 0
j.l =
yielding the solution
1 !.V1 _ p2
2p
Since the acceptable value of j.lmust lie between 0 and 1, the
appropriate root which gives the optimum value of j.l is
77
= 1 - -V 1 - p2j.l 2
p(4.68)
We compare the gain in efficiency when the optimum value of j.l
is used to the gain in efficiency when ~ = ~;! and ~.
Substituting the optimum value j.l = 1 - 2 1-02 in (4.67) andp
after some simplification, we obtain:
Var ( T*) = [. p2 j 2Vy
2 opt. y 2 n2 - 2 1-p
11/
The percent gain in efficiency of 2T* over 2T when the optimum
j.l is used is
11/It is interesting to note that. there is
Var ( T*).. = ! t. p2 J 2V
y2 opt. 2 .... r--'2 n
- V 1-p'"
where j.l t = the optimum value given by (4.68).op •
a relation
V= -l (~)
2j.lopt. n
,)
..
78
The optimum values of ~ as a function of p and the gain in
efficiency are tabulated in Table 4.14.
Table 4.14. Optimum value of ~ when the estimator 2T* is"used and the percent gain in efficiency over 2T
p Optimum ~ %Gain in efficiency
.6 ·55 11.11
·7 .58 16.73
.8 .62 25.00
·9 .69 39·50
·95 ·77 53.33
Comparing the gain in efficiency 'using the optimum value of ~with
111the gain in efficiency using natural values of ~ = 4' ' 3' ' 2' (Tables
4.11, 4.12, 4.13), we see that the gain does not increase much from
1 1using ~ = 2' .,.. and in view of·· simplicity the sampler may use ~ = 2'
"*without sacrificing much efficiency of the estimator 2T •
Moreover, comparing the extent of the gain in efficiency using
"* ~ ~tthe estimator 2T with that of 2T and 2T for the three cases,
111 tnamely ~ = 4' ' 3' ' 2' ' and when the optimum Q s are used in both
estimators 2~ and 2T' (Tables 4.2, 4.3, 4.4, 4.8, 4.9, 4.10, 4.11,
) A*4.12, 4.13' , we see that the use of the estimator 2T will produce
the largest gain in efficiency.
"*It should also be noted that in comparing the efficiency of 2TA
with that of the simple estimator 2T, no assumption regarding the
79
stability of th~ variance from occasion to occasion is made as in the
former set of comparisons.
4.3.4. A Ratio-Type Composite Estimator
Another type of estimator that might be used to estimate 2T is
given by
where
.A
2~* = 'l[(2,1~){lT)] + (l-'l) [2~1,2T
A A. A A
2,lT, 1,2T, IT, 2T and Q are as defined previously.
A*4.3.4.1. Expected Value of 2T. From (4.69), we have
(4.70)
A
We have shown that E(2T) = 2T. To obtain an expression for
E[:2'1~)(lT)1, we refer to an unpublished lemma given by Koop.
1,2T JLemma I: If X, Y, Z are three random variables and Z:f 0 , then
the following relation holds:
E[X£] = ~[1 + COvx(~,y) COVx(~·Z) -
_ COv)~,:,Z)] + ir- (z : z)21XYZ ~ Z J
where X =E(X), Y = E(Y), Z=E(Z);
and Cov (X,Y,Z) = E[ (X - X)(Y - Y)(Z - Z)]
Cov (Y,Z)
YZ
(4.71 )
80
The proof of this lemma follows from the identity
~ == ~ [1 + 6X + b.Y +~Y - t:fZ, - t:.Xt::Z - b.YhZ - A'X't:.YhZ]z
+ [1m (t:fZ, )2] (4.72)zwhere
!:iZ = (z: Z)Z
We now apply the above lemma to find E[(2'1~) (IT) which is
1,2T Jone component of E(2~)' Recall that E(2,lT) =2T, E(1,2T) = IT
A A
= E(lT) and that 1,2T rO. Hence by the lemma, we have
[
A j [ A ATAT Cov ( T, T)E (2,1) (T) = (~)( T) :1 + 2,1 1
T 1 IT 1 2T1T1,2
81
Or equivalently,
From (4.70) and (4.73), we obtain
.2T A A Cov (211~::' 1,2T, IT)
- 1..fCov (IT, 1,2T) -
1..f
A
IT)2]] •+...1.. E[ :2!1~)(lT) k2T-l..f 1,2T
(4.74 )
~
Hence, 2T"', unlike the other estimators considered previously,
is not an unbiased estimator of 2T. The amount of bias is Q times
the function involving the four covariances, the triple covariance and
the remaining involved product-moment given in the square bracket.
However, upon examining the nature of the covariances (see (4.33)) and
the product-moments involved which are of different signs and are
scaled by the population totals and the square of the totals
respectively, we will see that the amount of bias is likely to be small,
especially when multiplied by Q which is less than one.
l'.:::4.3.4.2. Variance of 2T*. From (4.69), we have
A 2[Var (2T*) = Q :Var
A
To obtain the _essions for Var t(2.1;)(lT)} and.
A 1,2
Cov {(2'1.~)(lT), 2T 1, we will first establish two lemmas which W1ll
1,2T )
be applicable to our problem.
Lemma II: If X, ~ Z are three random variables which are not neces
sarily statistically independent, and Z r 0 , then
._w
Var -- 2 t(XY)'= [XY] Var (X) + Var(Y) + Var (Z) +2·Cov (X,Y)Z Z . X2 ;p. ~ Xy
~-- 2[_ 2 CO~£X,Z) _ 2 CO~ &Y,Z) + [~] Var (8XAY)
XZ YZ Z
+ Var (1SI.l:Il) + Var (6.YM.) :.. Var (8XAYM.)
+ 2 COY {b.X, 6.X6.YJ - 2 COY {b.X, 6.Xb.Z}
- 2 COY tb.X, 6.YM.J - 2 COY (b.X, 8XAY6.ZJ+ 2 COY tAY, 8XAYJ - 2 COy t6.Y, 1SI.l:Il) - 2 COY { bY., 6.YM. }
- 2 COY {AY, ~J - 2 COv( &, 6X6.YJ+2 COY{&,~) +2 COV{&, 6.Y6Z} +2 COV{ &, 8XAY6.ZJ
(continued on next page)
i 83
- 2 COY { AXAY, ~} - 2 COY {AXAY, f:::,.":f.8l, 1
- 2 COY [AXAY, AXA":f.8l,) + 2 COY \ l:!:I.b2., tJf:::,.":f.8l,J
+ 2 COy t=, =)] - Var ~~ (llZ)'
+ 2 COy tr, r (t:::..Z)2]
where
=t:::..X = (X:X), f:::,.Y = (Y:Y), 6Z = (Z:Z) •
X Y Z
Proof: The identity (4.72) shows that
n ~[ ]Z == -; 1 + t:::..X + f:::,.Y + t:::£L:::i - 6Z -~ - &M., - AXA":f.8l,Z
+ [~ (6Z)2)
(4.76)
Transposing the last term. on the right hand side to the left hand
side, we have
n n( )2 XY[- - - til =- 1 + t:::..X + t::J - til +~ - !SXAZ - f:::,.":f.8l,Z Z =Z
84
• • Vax [~] + Vax [~ (AZ)2] - 2 Cov [~ , ~ (AZ)2]
= [~t [vax (LIX) + Vax (£:;Y) + V';" (lIZ) + Vax (IIUIY) + Vax (LIXAZ)
+ Vax (6.Yb.Z) + Vax (1Sf.6.Yb.Z) + 2 Cov t6.X, 6.Y) - 2 Cov \ 6.X, t1l.J
+ 2 Cov t6.X, ISf.6.Y) - 2 Cov {6.X, 6.XbZ) - 2 Cov t6.X, 6.YDZ)
- 2 Cov {6.X, 6.XAYAZ) - 2 Cov t6.Y, AZ} + 2 Cov ( 6.Y, ISf.6.YJ
- 2 Cov t6.Y, 6.X!:fl..) - 2 Cov {6.Y, 6.Yl:1l} - 2 Cov { boY, ISf.6.YDZ)
- 2 Cov tAZ, ISf.6.Y) + 2 Cov {AZ, 6.XbZJ+ 2 Cov { AZ, 6.Yb.ZJ
- 2 Cov t1:::.X6.Y, 6.X!:::..Yb.Z) + 2 Cov t6.XbZ, 6.YDZ}
+ 2 Cov t6.Xtfl, ~)' + 2 Cov t1Sf.t:jZ, D.XAYDZJ].Now Var (6.X) = Var (X:X) = ~i Var (X)
X XSimilarly,
Var (6.Y) 1= =2 Vax (Y) ,
Y
Var (t1l.) = ~ Var (Z) •
Z
Cov t6.X, 6.Y) = Cov {(X:X), (y:y)} = =: Cov {X,Y),X Y- XY
85
and
... =: Cov tX,Z ~ ,xz
=: COv { Y,Z) .yz
. . from (4.77), we will have
+ [~l2 [var (=) + Var (=) + Var (l>YllZ)
+ Var (A':l.AY6Z) + 2 Cov {8X, A':l.AY) .. 2 Cov {8X, b:.XAZ }
.. 2 Cov (8X, l:::.Y6ZJ .. 2 Cov {8X, A':l.AY6Z}
+ 2 Cov tl:::.Y, A':l.AY) .. 2 Cov {l:::.Y, b:.XAZ )
+ 2 Cov { 6Z, A':l.AY6Z} .. 2 Cov lA':l.AY, t:::.Xt:::.ZJ
.. 2 Cov tA':l.AY, l:::.Y6Z} .. 2 Cov tA':l.AY; t:::.Xt:::.Y6ZJ+2 Cov(~,ISl6Z)
+ 2 cov {b:.XAZ, t:::.XAY6Z) + 2 Cov tb.Y6Z, A':l.AY6ZJ]
- Var [~ (t,Z)2] + 2 Cov [~ , l>f (t,Z)2J •
86
Lemma III: If X, Y, Z and W are random variables which are not
necessarily statistically independent and Z r 0 , then
(4.78}
COY [~ ,wl = X! [COy)X,W) + COY)Y,W) _ cov)z,w)lZ . X Y Z J
+ ~ fCoV (=, II) - Cov (=, II) - COv (EfAZ, II)
- Cov (=t1Z, II)] + Cov [~ (l>Z)2, II]
Proof: Using identity (4.72), we have
Cov [(~), III = COv[(~ (1 + 1St. + l!>f += -l>Z -=- IYUSYN,) + ~ (l>Zf} , II] • (4.79)
Hence, the R.H.S. is
[~l (COy (~,W) + COy (~Y,W) + COy (~Y,W) - Cov (~,W)
- COV (=,11) - Cov (6YllZ,II) - COV (=t1Z,tn]
+ COy t~ (~)2, W1 0
Since COy (~,w) = COy ((X~X) , W}
1 {- = J= X l E[(x-x)wl - E(X-X) E(W)
1:; E (xw) - X E(W)X
= : E (xw) - E(X)E(W)X
= ~ COy (x,w) 0
X
f 87
Similarly,
.... COY (AY,W) 1= -;- Cav (Y, W) ,Y
1COY (~,W) =-=- COY (Z,W) 0
Z
Hence, from (4.79)
COY [(~) , W] = X! [cov_(X,Wl + cov)Y,W) .. cov).z,W)]z X Y Z
+ X! [COY (AXAY,W) .. COY (AXAZ,W) .. COY (AYM.,W)Z .
.. COY (Az/W/\z,W)] +COV t~ (~)2., W) '0
In applying the two lemmas to our problem where the various
estimates under study play the role of random variables, we will use
only the leading terms, !o!o, the terms of second order or less in
(4076) and (4078) by assuming that the magnitudes of the remainder
terms in the two formulae are relatively small as compared to the
leading terms 0
Thus, the approximate expressions for
"'-
and Cov ((2., 1~ )(1T), 2.TJ1,2
T.
are:
88
and
+ Var (1,2T) + 2 Cov (2.z'lT, IT)
I T2 2TIT
2 OOv (:~ l}l] (4.'80 )
(4.81)
A
Hence, from (4.75), the approximate expression for Var (21!*) is
A 1Using previous results such as Var (2,lT) = (l-lJ.)n (2Vy)'
" " 1Cov (2,lT, 1,2T) = (l-lJ.)n (1.2Vy) etc., we can write the approximate
89A
expression for Var (2T*) as
~ 2[ 1 2T 2 1V 2T 2 1Var (2T*) ~ Q(l-~)n (2Vy) + (~) ~ + (~) (l-~)n (lVy)
1 1
22T 12T 1 2T 1 ]
+ 2 (IT) n (1.2VY) - 2 (IT) (l-~)n (1.2VY) - 2 (iT) n (lVY~
+ (1_Q)2 ! ( V ) + 2Q(1-Q) r! ( V ) + (2T
) (l-~)n (' V )n 2 y In 2 y 1T n2 1.2 y
Or, by collecting some terms, we have
where
R =
Remarks:
[2 2 2]+ 2R .Q~ - Q~ - Q ~
. (l-~)
V1.2 yn (4.82)
(i) Comparing (4.8,2) with (4.35), we see that the approximate~
formula for Var (2T*), is very similar to the exact formula forA
Var (2T) !.~., the variance of the linear composite estimator, except2T A A A .
for the ratio R = T· When E(2· IT) = E(lT) = E(l 2T) !.~.,1 ' ,
when 2T = IT the two formulae are identical.
(ii) The approximate formulae (4.80) and (4.81) ca,n also be
derived by the use of the Taylor approximation technique, ! ..!., by
.. 90
expanding
'"= (2,1T) ( T)
'" I1,2
T
into a power series about the point (2T, IT, IT). Neglecting the
terms of powers higher than one, we get
A
(2,1~)(lT) ~1,2T
(.2T
T)( IT ) + (. T- T} 0
1 2,1 2 0 T2,1
A
{ (2,1~)(lT)J1,2T
+ {T- T} l1 1 ~ A
°IT
which gives
A
IT':"IT
or
A
2 IT A {A} f A, } 2T
{ A J(~)(IT) s. 2T + 2,lT-2T + IT-IT -T - 1 2T-IT1 '1,2T
(4.83)
91
andfrom which the approximate formulae for Var {(2'1~)(lT)}1,2T
Cov { (2'1~)(lT), 2TJ can be arrived. However, the latter technique
1,2T
must rely on the assumption of differentiability near (2T'lT, IT)
and the existence of partial derivatives Which cannot be guaranteed
simply because the estimates involved are not continuous.
(iii) From (4.83), we also note that when 2T = 1T , we will have
'" '" '"IT + 2,lT - 1,2T
which in turn, implies that in such case
!.~., the ratio-type composite estimator will give approximately the
same estimate as does the linear composite estimator.
~*4.3.4.3. Efficiency of 2T. From (4.82), we see that the
A 2Ta.pproximate variance of 2T* involves the ratio R =T which is
1
unknown. However, in practice, we may use the sample ratio
as an estimate of R.
By rewriting (4.8,2), we have
•
Under the assumption that V = V = V and writing p V forly 2y Y Y
l.2Vy , we have
Var (2~*) ~ Var (2T) + ri (1~~) [1 + ff] ~
[2 2 2] V+ .. QJl - QJl - Q Jl .1l.
2 P R . (l-Jl) n •
From (4.84), we see that when p = 0 , we will have
(4.84)
which implied that in this case (!..!., no correlation over time), the
ratio-type composite estimator to the order of approximation involved,"'-
is less efficient than the simple estimator. 2T. In practice, we
would expect the correlation p to be sufficiently high so that the
use of the estimator 2~* will result in a gain in efficiency over theA
estimator 2T.A
Since the expression for Var (2T*) given by (4.84) does involve
2Tthe value of the ratio R =~ , we cannot make any specific comparison
~ 1about the efficiency of 2T*. However, we will examine the nature of
the interval value of R, which is usually unknown, for some specific~
cases of interest in which the use of the estimator 2T* results in aA
gain in efficiency over the simple estimator 2T •
We consider (4.8'4). To the order of approximation involved,
* "'-Var (2T*) will be less than Var (2T) if
Q2fJ. f- l + R21 ~+ 2 p R [QJl2
- QJl- Q2Jl21 ~ < 0(l-Jl) t J n . (1-fJ.) J n .•
93
For 0 < Q < 1, 0 < IJ. < 1, the above condition is equivalent to:
Specific cases for
* AThe estimator 2T* will be more efficient than 2T if
(ii)
A~ ~The estimator 2T will be more efficient than 2'1' if
~ [5p - -J25p2 - 9 ] < R <.~ [5p +-J 25p2 - 9 ]
1 1( iii) For Q =2" ' IJ. =2" •
~ A
The estimator 2T* will be more effiOient than 2T it
4.4. Estimation of Change in Total between the First andSecond Occasion
To estimate the change in the total of the population characteristic
=
94
several estimators can be constructed. To this end, we will consider
the estimators which bear close resemblance to the estimators used for
estimating the total. Also, attempts will be made wherever possible
to compare the relative efficiency of these estimators with respect to
the simple estimator
(4.86)
4.4.1. An Estimator Based on the Linear Composite Estimator
4.4.1.1. Estimator. One possible estimator that can be used to
'"estimate Dl is Dl whose structure is:
'"~ = T-1T1 2
'"'"where 2T is the linear composite estimator as defined in (4.12).~ A
Since we have seen that E(2T) = 2T and E(lT) = IT , this
implies that
= T - T.2 1
(4.88)
95
Using the results obtained previously for the relevant
variances and covariances in (4089), we Will have
Under the assumption that
and by writing p VY for 102VY , the expression for the variance of~Dl becomes
;... 2[ Vl 2[ ]VVar (Bl ) = Q 2(1-p) (l-~)rij + 2(1-Q) 1 - (l-~)p ++ 2Q(1-Q) [2(1-P) V~] ,
which can be rewritten in a more suggestive form as
V(.L..)
n
(4.91)
..
96~4.4.1.3. Efficiency of Dl • To see how the use of the estimator
AA ADl will result in a gain in efficiency over the simple estimator D
l,
we recall that
with
=! ( V ) +! ( V ) _ 2(1-~)n2y nly n (4.92)
we will haveUnder the assumption that IVy = 2VY =Vy
Vax (~) =2 [1 - (l-~)pJ V~ •
Hence, from (4.91) and (4.93), we have the relation
~We see again that when p =0 , the estimator Dl is less
A
efficient than the simple estimator Dl , for in such a case
V
(l~~) (+)
,.,.
since the second term is always a positive quantity.
However, when the correlation p is high, we would expect that the~
use of the estimator Dl will give a more precise result. For this,~ A
we will now examine the gain in efficiency of Dl over Dl for some
values of ~ and Q Which might be adopted in practice.
97
Case I:
Here we have
Var (~ ) =Var (D ) + (4 - 13p)1 1 24
v...L
n
~
The percent gain in efficiency of D1 over
( ) _ ( 13p - 4)gl P - 52 _ 49p x 100.
*The percent gain in efficiency of D1
shown in Table 4.15.
....over D1 for this case is
...
-
A ....Table 4.15. Percent gain in efficiency of D1 over D1 when
1 11-l=4' Q=2'
p . %Gain in efficiency
.6 12.39
·7 28.81
.8 50.00
.9 97.47
·95 153·21
Case II:
In this case, we will have
.. v...L
n
..
A ~
The percent gain in efficiency of D1 over D1 is
( ) _ ( 9p-3 )B2 P - 27-25p x 100 •
Table 4.16 shows the nature of the gain for some values of P
which may arise in practice.
98
Table 4.16.
Case III:
~ "-Percent gain in efficiency of D1 over D1 for
1 11-1=:3' Q=2'
P %Gain in efficiency
.6 20.00
·7 34.73
.8 60.00
·9 113·33
·95 170·77
\
1 11-1=-, Q=-2 2
In this case
90 A
The percent gain in efficiency of D1 over D1 is
( ) _ (3) -2 )g3 P - 10-9p x 100.
Table 4.17 shows the nature of the gain in efficiency for this
case.
Table 4.17.~ A
Percent gain in efficiency of D1 over D1 for1 1
1-1=2"' Q=2"
P %Gain in efficiency
.6 21.74
·7 40.34
.8 71.43
·9 131.58
·95 189.65
99
4.4.1.4. A Comment about the Gain in Efficiency. Comparing the~ A ~
nature of the gain in efficiency using D1 over D1 to that of 2TA
over 2T for the three cases considered, we see that sUbstantial gain
will be realized in estimating the change. In practice, we may have
to estimate both the change and the current tot.a1. This should present~
no problem for once 2T is computed,~ .
D1 is automatically obtainedA ~
by simply subtracting 1T from 2T. To see how the gain in effi-
ciency changes when the optimum value of Q is used in (4.88), we
will now consider the gain in efficiency for the three cases
considered previously.
Case I:
1 ~When 1-1 = 4' Var (D1 ) as given by (4.91) is
~ [ 2fL 2 2 (Q2 4tlQ , 1 VvVar (D1 ) = 2 (1 - T) + 3' Q - - 6 ~ Pj ~
~MinimiZing Var (D1 ) with respect to Q, results in the optimum value
Q =..2L-(4-p)
..
100
When this optimum Q is used we will have
A
The percent gain in efficiency over D1 is
3 2'::7..,....lL.... x 100.10 (l-p)
The percent gain in efficiency is tabulated in Table 4.18.
Table 4.18.
Case II:
A
'"Optimum value of Q to be used in the estimator D1
andA 1
percent gain in efficiency over D1 for ~ =4
p Optimum Q %Gain in efficiency
.6 ·53 16.87
.7 .64 30.62
.8 .75 60.00
.9 .87 151.87
·95 ·93 338.43
When 1~ = 3 ' (4.91) leads to
~ [ 2 2 Q2 + 4Q JVVar (Dl ) =2(1 - 3 p) + Q - 3 p ~.
~Minimizing Var (D
1) with respect to Q, results in the optimum
value of Q = (3:P) ·The percent gain in efficiency over
101
The optimum values of Q and the percent gain in efficiency over....D1 for this case, are tabulated in Table 4.19.
....Table 4.19. Optimum value of Q to be used in the estimator 1)1 and
percent gain in efficiency over 1)1 for ~ =~
p Optimum Q %Gain in efficiency
.6 ·50 20.00
·7 .61 36.29
.8 .73 71.11
·9 .85 180.00
·95 ·93 401.01
Case III:
1When ~ =2:' (4·91) gives
VVar (D ) = [2(1 - 2.) + 2Q2 .. (Q2 + 2Q)p] J:.
1 2 n
and the optimum value of Q is ~ •
'"The percent gain in efficiency over D1 is
x 100.
The optimum values of Q and the percent gain in efficiency over
'"D1 for this case, are tabulated in Table 4.20.
102AA
Table 4.20. Optimum value of Q to be used in the estimator Dl andA 1
percent gain in efficiency over D1 for ~ =2
p Optimum Q %Gain in efficiency
.6 .43 22·50
·7 .54 40.83
.8 .67 80.00
.9 .82 202·50
·95 ·90 451.25
4.4.1.:2. A Remark about the Optimum Value of Q. It should be
:::noted that although the estimator Dl is based on the linear composite
~ ~ A
Dl = 2T- 1T , the optimum values to be used in
1 ~and 2' the optimum value of Q for Dl
AA
optimum value for 2T •
1= 2 ' the optimum value of Q to be used in
A
the optimum value to be used in ~l is
~
estimator 2T, !.~.,
~ ~
2T and Dl are~ the
1 1namely, when ~ = '4 ' 3'turns out to be twice the
For example, when ~
~ whereas2(2-p) ,
same. In the three cases which we considered
,
In practice, the sampler must have the primary objective in mind
whether he wants to get an estimate of the total or the change in total
and choose the best weight which will result in the best gain. How-
ever, when the sampler wants to get good estimates of both the current
total and the change in total between the two occasions, the appro-
priate weight Q for estimating total may be used in both cases without
sacrificing much efficiency.
..103
4.4.2. An Estimator Based on the Modified Linear COmposite Estimator
Another possible estimator which can be used in estimating Dl
is'"
B~ = 2~' - IT (4.94 )
~,
where 2T is the modified linear composite estimator as defined in
(4.49) .
We have seen that
evident that
'"""!.~., D is also an unbiased estimator of Dl •
..By writing (4.94) as
or
we find
,)
(4.96)
104
Remembering that
A A
COV (2,2T, IT) = o ,
A A
COV (2,lT, 2,2T) = o ,
A A
COV (1,2T, 2,2T) = o ,
and using the expressions obtained earlier for the relevant variances
and covariances, we will have
+ [l_Q]2[..1..( V ) +! ( V )1j..I.n 2 y n 1 y J
+ ~ (1Vy )] •
+ 2Q( l-Q) [- 1 ( V )n 2 y
Under the assumption that 2Vy = lVy = Vy and L2Vy = pVy' the
expression now becomes
'" [ 2 2 2 2] VVar (n') = (l-Q) -l-Q j..I. + 2Qj..I. - j..I. ..JL +1 j..I.(l-j..I.) n
..
~I4.4.2.2. Efficiency of Dl • We will now consider the efficiency
of the estimator ~:i. relative to the siIll.Ple estimator £1 for some
specific cases which might be adopted in practice.
Case I:
105
In this case
Var (E~) = [~ - t p] v~
= Vax (D ) + [....2 + ! pJ :.z..1 J2 3 n'
where
var(l\)= [2 -i p] V~ , as given by (4.93).
A A ~,
Hence Vax (j)i) > Var (Dl ) for all °< P < 1 !o!!.. Dl is~
efficient than the simple estimator j)lo
Case II:
1IJ. =3' '
In this case
Var (~~) = [t -t p] +
> Var (Dl ), for °< P < 1 °
~
Hence, we see again that in this case, Di is also~ effi-
'"cient than the simple estimator Dl °
Case III:
In this case,
Var (~i) = Vax (Dl ) +~ V~
106
~IA
The percent gain in efficiency of Dl over Dl is given 'by
g7(P) := (2p;S1) x 100 .9- P
A AAI
The percent gain in efficiency of Dl over Dl for this case
is tabulated in Table 4.21.
Table 4.21. Percent gain in efficiency of ~~1 1
~=2"' Q=2"
A
oyer Dl for
..
P %Ga~n in efficiency
.6 3·70
.7 8.33
.8 14.30
·9 22.22
.95 27·27
We see from the above table that the amount of gain in efficiency
is not substantial.~t
To improve the efficiency of Dl for the three
cases considered above, the optimum weight Q for each case may be
used. Since the procedure is analogous to that used in the estimator~
Dl , we willmt present the details of the derivation, but only the results
will be given. These are found in Table 4.22.
A comment:At
Comparing the nature of the gain in efficiency of Dl to
.'
----~
~that of Dl when the optimum values of Q are used in the three
111cases namely ~ = 4 ' :3 ' 2' ' we see that the gains from using the
~t Aestimator Dl are slightly~ than the gain from using Dl • This
is due to the structures of the two estimators themselves. (See (4.88))
107~
Table 4.22. Optimum values of Q to be used in the estimator D''"
1and percent gain in efficiency over D1
Case I: 1 Optimum Q = (3p + 12)1l=1j: 17 - 2p
P Optimum Q %Gain in efficiency
.6 .87 11.98
·7 .90 26.34
.8 ·93 56.34
·9 .97 148.86
·95 .98 336.15
Case II: 1 Optimum Q = t; ~ pjIl=-3
.6 .81 13.79
·7 .86 30.67
.. .8 ·90 66.10
·9 ·95 173·10
·95 ·97 393.33
Case III: 1 Optimum . - f2 + p)Il=-2 Q -5 - 2p)
.6 .68 14.65•
·7 .75 33·33
.8 .82 72.88
·9 ·90 195·80
·95 .95 425.14
108
A "and (4.95». The difference 2,2T - IT in (4.95) provides l~ss
"information about 2T - IT than the difference 2T - IT in (4.88).
4.4.3. A General Linear Estimator
Another type of estimator Which can be used to estimate
A ,,~ A
where 1,lT, 1,2T, 2,lT and 2,2T are as defined previously, and,A
a', b', e', d' are constants to be determined so that Dr is an
unbiased estimator of Dl having least possible variance.
Determination of Constant a', b', c'; d'.
to be unbiased we must have E(Dr) = 2T - IT •
E(Dr) = (a' + b' )IT + (c' + d' )2T •
Imposing the condition of unbiasedness we have
A
For nr
(a' + b' ) = -1
(c'(4.100)
+ d') = 1
giving
b' = -(l+a' ) ,d' = (l-c') ,
so that
Dr = a' \l,lT} - (l+a') {1,2T} + c' t2,lT} + (l-c') [2,2T}.
(4.101)A*We now choose a' and c' in (4.1-01) so that Var (Dl ) is a minimum.
, 109
First we have
(4.102 )
Under the proposed sampling design, the only non-zero covarianceA A
is Cov (1,2T, 2,lT) •
Using previous results for the relevant variances and covariances
in (4.102), we will have
Var (n*) = (a l )2 (lVY
) + (1+a,)2 (A-) + (c,)2 (A)1 Iln· (l-ll)n (l-ll)n
v V+ (1_c,)2 (~) - 2(1+a') c' (~) •
Iln (l-ll)n
Taking partial derivatives of Var (D"~) with respect to a' and c'
and equating the results to zero, we have
(4.104)~ai Var (D~)
~Ci Var (Dt)
= [2a' + «1)' )1 V 2c' ( v) - 0~ 1-1l n] 1 y - (l-ll)n 1.2 y -
[2c' 2(1-c' )1 «1)') ( )
= (l-ll)n - Iln j 2Vy - l-~ n 1.2Vy =0 •
(4.105 )
..111
Using these weights in (4.101), we can now write the estimator as
D* =1
+
+
~l-~) + P~(l-~~fi J[ (1-11 p) Jp~(l-~) -Ji§;. + (l-~)l
(1_112p
2 ) ]
~_~2p2 _ p~(l-~) -J..1Vy
2V
y
(4.110)
"'*4.4.3.2. A Comment about the Form of the Estimator D1 • We see
again that the estimator D~ like the estimator 2T*, requires the
values of 1Vy and 2Vy and p. Under the assum,ption that
1VY = 2VY , the ratios under the radical signs in (4.110) become unity,
so that only the value of p is required in the weight function. It
may also be noted that these variances and p can be estimated from
the samples for two successive occasions as pointed out in Section
4.4.3.3. Variance of D~. Using the value of a', 1 + a', c'
and 1 - c' given in (4.106), (4.108), (4.107), (4.109) in (4.103),
we have
112
-J§. 22 2pl-l(l-l-l) =..Jl. + l-l P - l-l
lVyJ, A lVyVar (D~) = 2 2 l-ln(l-l-l p)
~2
(l-l-l) + Pl-l( l-l-l) =-.ylVlV
+ Y(l-l-l)n2 2
(l-l-l P )
l~(l_~)-J11 + (1-~) J2
V+ 2 Y
(l-l-l)n(1_l-l2p2 )
22 ~2
l-l - l-l P - pl-l(l-l-l) ~2Vy2V
+ y2 2 l-ln(l-l-l P )
2 2(l-l-l P )
After some algebraic manipulations, the exPression is found to be
A [2~V [2~VVar (D*) = l-p l-l !...y +l-p l-l U - 21 122 n 122 n-P l-l -P l-l
ii
......*4.4.;.40 Efficiency of the Estimator Dlo We will compare the~ ......
efficiency of Dl with that of the simple unbiased estimator Dlo
Under the assumption that 2Vy = lVy = Vy , we will have from
(4.111)
113
(4.113 )
We recall from (4.93) that
Hence we have the relation:
Var (Di) := Var (Dl
) _ 2 fp2J.1 - P:J.1
2]. V~ •
l 1 - P J.1
The percent gain in efficiency of Di over Dl as measured by
is
Var "'- "*(Dl ) - Var (Dl )
Var (D~)x 100
(4.114)
"* "'-The nature of the gain in efficiency of Dl over Dl is~ A
identical to that of Dl over D1 for the corresponding cases
111J.1 := 4' ' 3" and 2' when the respective optimum values of Q are used.
1We verify and find that when J.1:= '4
1 _ 3p2g(p, 4') - 16(1-p)
showing the same gains in efficiency as in Table 4.18. The cases when
J.1 := ~ and ~ can also be verified and we will find that
11)g(p, 3") = g5(P), g(p, 2')= g6(p •
"
114
The explanation: for this state of affairs is because when in Dr(as given by (4.110)) we set ,.,V = lV = V , we find the resulting
~ y y y~
form identical to D1 , as obtained when optimum values of Q are
used as weights i~ (4.88), and writing also
4.4.4. An Estimator Based on the Ratio-Type Composite Estimator
Another type of estimator that can be used to estimate Dl =2T - IT
A*is Dl given by
~~ = 2~* - lT
{2.1~)(lT)] + (l-Q) [2T] - " (4.115 )= IT
1,2T
4.4.4.1- A* With the result at (4.74) we findBias of Dlo
so that the amount of bias of
same as the amount of bias of
~*D1
as an estimator of 2T - IT is the
"'*2T as an estimator of 2T °
115
~* ~* ~* A4.4.4.20 Variance of Ill' We have from Dl = 2T - IT ,
A*The approximate expression for Var (2T ) has· been established in
(4.83 ) • The expression for Var (1T) is well known. To find the
A* Aapproximate expression for Cov (2T , 1T), we first recall that
2~ ~ Q r(2.1~) (1;) + (l-Q) [2T] •L1,2T J
Hence
(40116)
Now, to find the approximate expression for
we apply Lemma III. Using only the leading terms,
mate expression to be
A
2 IT A A
Cov [(~)(lT), lTl,1,2T .
we find the approxi-
(4.117)
116
Using previous results for the relevant expressions in (4.117),
we have
l~(4.118)
..
2 1TA A
Substituting the approximate expression for Cov [(~)(lT), 1T]
1,2T.A A
and the expression for Cov (2T, 1T) obtained earlier in (4.116) we
have
A
And hence With (4.83) and recalling that Var (1~)
~*the approximate expression for Var (D1 ) as
[2 2 2] V+ 2R .QIJ. - QIJ. - Q IJ. ~ +!. (V)
(1-1J.) n n 1 y
_ 2Q ( V) _ 2(1-Q) [(1-1J.) ( V)] ..n 1.2 y n 1.2 y
(4.120 )
Under the assumption that 1V = 2V = V , we obtain, after somey y y
simplification:
.....- where
~* V 2 2 VVar (D1 ) = 2[1 - (l-lJ,)p] ~ + Q (1~1J.) [1 + R ] ~
- 2 [f~~~) R + (l+R)~] P ~ (4.121)
117
**4.4.4.3. A Remark about Var (Dl
). Comparing the approximate
~ ~variance of lJl given by (4.121) to the exa.ct variance of Dl given
2Tby (4.91) we see that when R = T' the two expressions are the same.
A 1 VRecalling that Var (Dl ) = 2 [1 - (l-~)p] .:;, we have the relation
Var (~~) ~ Var (D1 ) + Q2 (l~~) [1 + ~] ~
•
~* AIn (4.122), it is evident that when p = 0, Var (Dl ) > Var (Dl ).
When the correlation p is sufficiently high, we would expect that
A*the estimator Dl when used, will result in some gain in efficiency
A **over the simple estimator Dl • However, since Var (D1 ) involves the
2Tratio R = T which is usually unknown, we cannot make any specific
1comparisons regarding the efficiency of this estimator.
We will examine the nature of the interval values of R where
**the use of Dl will result in a gain in efficiency.
:::::* "Now, to the order of approximation involved, Var (Dl ) < Var (Dl )
if
V VQ2 (l~~) [1 +~] 1- (i~) [Q~R + (l+R)(l-~)] P ~ < 0 ,
or if Q[l + R2] - 2[Q~R + (1 + R)(l - ~)] p < 0 •
Case I:
~* AHere the estimator lJl is more efficient than Dl if
~ [7P -1/49p2 + 48p - 16] < R < ~ [7P +-J49p2 + 48p - 16 ].
•
Case II:
118
~* AHere the estimator D1 is more efficient than D1 if
Case III:
Here the estimator ~~ is more efficient than £1 if
•
•
5. MORE THAN 'IWO STAGES OF SAMPLING AND MORE THAN 'lWOSUCCESSIVE OCCASIONS
5.L General Remarks
The estimation theory covered so far has been for two-stage
sampling for two successive occasions. When the number of stages of
sampling and the number of successive occasions are more than two, the
algebra becomes more involved. We will now consider the extension of
the theory to such cases. For this!J we will consider the two exten-
sions of the theory separately, first the theory for more than two
stages of sampling on two successive occasions, and then the theory for
more than two successive occasions where the sampling may be done in
two or more stages. .
5.2. More Than Two Stages of Sampling
When the practical situation demands that the sampling must be in
three or more stages and the sampler still wants to incorporate the
technique of partial replacement of first-stage units to such multi-
stage sampling design for the reasons described in Chapter 1, the basic
scheme as described in Chapter 3 is still applicable. The only addi-
tional thing to be considered is the appropriate method of selecting
units in the succeeding stages. Since the first-stage units are drawn
with unequal probabilities and with replacement and the second-stage
units are drawn with equal probabilities and without replacement, to
accomplish the purpose of spreading the burden of reporting among
respondents, the selection of units in the third and other succeeding
stages may be done as in the second stage, 1.~., selecting the units in
119
120
the third and other succeeding stages with equal probabilities and
\l'
without replacement. As an illustration of the general problem, the
occasion.
extension of the theory to four-stage sampling will be considered.
ththe variate value of the £ fourth-stage unit of
unit of the jth second-stage unit of the i th first-
Let tYijk£ beththe k third-stage
stage unit on the tth
We are interested in estimating
N Ni ~j NijkT = ~ ~ ~ ~ lYo °k O
1 i=l j=l k=l £=1 J.J ~
(i=1,2, ••• N; j=1,2, .•• ,Ni ; k=1,2, ... ,Nij , £=1,2, ... ,Nijk ).
The procedure for selecting units is as follows:
(i) The n + (p-l)~n first-stage units for the preliminary sampleN
are selected with probabilities Pi > 0 (~ Pi=l), and with replacei=l
ment after each draw and the order of appearance of the units noted,
as described in Chapter 3.
(ii) m second-stage units are independently selected with equal
probabilities and without replacement after each draw in each of the
first-stage units selected.
(iii) r third-stage units are selected with equal probabilities
and without replacement after each draw in each of the second-stage
units selected.
(iv) q fourth-stage units are selected with equal probabilities
and without replacement after each draw in each of the third-stage
units selected.
An unbiased estimator for lT based on the first n first-stage
units of the set of n + (p-l)~n units is
121
(5.2 )
where the index i again refers to the order of occurrence of the
first-stage units in the preliminary sample.
'"We show that E(lT) = lT •
By a well-known theorem on conditional expectation
E(lT) = E [ES f ES \ ES (IT)} ;]lt 1,2 1,2,3
where in (5.3) ES ( . ) is the conditional expectation of the1,2,3
function represented by ( • ) given the relevant third-stage, second-
stage and first-stage unit, and so on.
Now
'" 1 n N. m N.. r Ni'k (] Nijk· 1E ( T) = - 2: ~ 2: ...hl. 2: ~.. 2: - y81,2,3 1 n i=l Pi j=l m k=l r q £=1 Nijk 1 ijk£
m Not;2: ....bL
j=l m
where
Further,
1n N, m
=- 2: 2:, 2:n i=l Pi j=l
Ni.J. Ni , 1r. 2: J - lYiJ'km r k=l Nij
where
Again further
122
where
And hence,
,
+s {~ (ES (IT)n]N lYinl: Pi=
1 1,2 1,2,3 n i=l Pi
N N~
Ni Nio k= l: 1Yi = l: l: l: JlYijk.e = lT •
i=l i=l j=l k=l ,£=1
To find the expression for Var (IT) the theorem on conditional
variance is applied. In our case we find
,
where VarS (.) is the conditional variance of the function1,2,3
represented by the dot given the relevant first-stage, second-stage
and third-stage unit, and so on.
We proceed to evaluate each part of the variance.
123
Next
~1 n Ni m ]
= Var - L: - L: Y81 n i=l Pi j=l 1 ij
1 n N. 2= 2 L: (pJ.) var(lYi)
n i=l i
1n N 2 0-2 (N.-m)
= _ L: (-i) 1 i ..,...=.J.~n2 i=l Pi m (Ni -1)
,
where
(lYij= 2
2 ~ - 1Yi) = Ni 1Yij10-i = 1Yi = L:Ni
, Nii=l j=l
Hence
Next, we consider
124
From (5.4), we have
1 n Ni
2 m Ni . 2 _=2 L: (p) L: (..2:.sl) Vars (lY' j )
n i=l i j=l m 1,2 J.
where
Hence
Finally, it can be shown that
125
E [ES {ES ( VarS (1T)}}]"1 l 1,2 1,2,3
= 1 ~ Ni ~i Nij ~ij ~jk l(j~jk (Nijk-q)
n i=l Pi j=l m k=l r q (Nijk-l)
where
= ~ijk£=1
= 2(lYijk£ - lYijk)
Nijk
....Combining these results, we obtain the expression for Var (IT) as
I N lYi 2 lNi liiVar (IT) =- E P. (- - T) +- E -1.
n i=l J. Pi 1 n i=l Pi m
(Ni-m)
(N.-l)J.
1 N N. N.+ - E .2:. EJ.
n i=lPi j=l
Generally for k stages of sampling, the total variance of a linear....
estimator such as 1T will be made up of k parts or components.
To estimate
..
!.~., the total of the population characteristic of interest for the
second occasion when ~n first-stage units have been partially replaced
as described in Chapter 3, four types of estimators namely 2~' 2~t,
....*. A*2T and 2T , can be used, and the theory is the same as in the two-
stage sampling case discussed therein.
N r:f. N.+ .E ..1:. .EJ.
i=l Pi j=l
..
126
As can be seen from the derivation of Var (IT), all other variances
and covariances involved in the four estimators mentioned above, will
be made up of 4 components.
For example,
1 [N 2Y
i lYi= (l-~)n .E Pi (-p- - 2T)(-p- - IT)i=l i i
(2Yij - 2Yi)(lYij - lYi ) 1 (Ni-m)Ni m (Ni-l)
+ ~ Ni ~i ~j ~ij (2Yijk- 2Yij)(lYijk- lYij) 1 (Nij-r)
i=l Pi j=l m k=l Nij r (Nij-l)
+ ~ Ni ~j Nij ~ij ~.llii=l Pi j=l m k=l r
(5.13 )
To estimate the change between the first and second occasion,
!..~. ,
four estimators: can also be used. The theory will
again, be very similar to the theory for the two-stage sampling case.
The changes will be entirely in the structure of the estimators and the
addition of two extra components of variances or covariances as the
case may be.
5.3. More Than Two Occasions
t
•
The estimation theory will be extended to the general case when
multi-stage sampling is carried out on more than two successive occa-
sions and the partial replacement of first-stage units is as described
in Chapter 3. We will first review briefly the scheme of partial
replacement first-stage sampling units and also introduce the notation
to be used.
Suppose that the sampling is done for a successive occasions
where a > 2. On the a th occasion, the n first-stage units to be used
are those units which occur from order (a-l)lln+l to n+(a-l)lln in the
preliminary sample. Of those n first-stage units, there will be
(l-Il)n first-stage units occurring from order (a-l)lln+l to n+(a-2)lln
common to the (a-l)th occasion; the other Iln first-stage units which
occur from order n+(a-2)lln+l to n+(a-l)lln are newly selected to
replace those Iln first-stage units which Occur from order (a-2)lln+l
to (a-l)lln. (See diagram.)
order: (a-2)lln+l (a l)lln+l n+(a-2)fln
~~:onI--+----!----III---·_ath occasion ..
The sub-units in the succeeding stages are selected in the manner'
described previously •
, ~8
5.3.L Estimation of TotaL
If the sampling is carried out in two stages, the current total
is
T =aNr:
i=l
If the sampling is carried out in four stages, the current total is
= N Ni :s.j ~ijkaT ' r: r: 2,; &.. ci'iJ'kJ
i=l j=l k=l .8=1
As in Chapter 4, four types of estimators which utilize past as
well as present information from the constituent samples can be used.
Since the theory can be generalized to multi-stage sampling of any
degree as indicated in 5.2 , we will illustrate here only for the two-
stage case.
5.3.L1. A Linear Composite Estimator. The structure of the
linear composite estimator for the ath occasion is
~
where a-lT = the linear composite estimator of IT,a-
A
Ta,a-l
is the unbiased estimator of aT based on those (l-~)n first-stage
units which occur from order (a-l)~n+l to n+(a-2)~n, and common to
the (a_l)th occasion,
A 1 n+(a-2)~n Ni m a-1Yij
a-l,aT
= (l-~)n i=(a-f)~n+l Pi j~l m
•
J.
]29
is the unbiased estimator of a_1T based on the same (l-~)n first
stage units used on the ath occasion,
c1ijm
is the unbiased estimator of aT based on the set of n first-stage
units which occur from order (a-l)~n+l to~+(a-l)~n. This estimator
would be used to estimate T if the sampling is done only for onea
occasion, or if the sampler does not wish to utilize past information
from the previous occasions.
~ :~.~
Expe ct e. d val u eo f aT. ~ Intuitively, we would expect~
aT to be an unbiased estimator of aT. For example, when a = 3 !.~.,
when the sampling is carried out for 3 successive occasions, we will
have
(5.18)
and
~
E(3T)
~
Since E(2T) = 2T as shown in Chapter 4, we will have
A
E(3T) = Q[2T + 3T - 2T] + (l-Q) 3T .= 3T •
This will be true for a = 4, 5 and so on, so that it will be
generally true.
~
Variance of aT. From (5.14), 'we have
130
When the number of occasions ~ is not so large, for example, when
a =3 we will have
It is necessary to determine what may be termed, the sub-variances~
and sub-covariances which make up the total variance of 3T •~
Var (2T) is given previously.
It can be shown that
var(3,2T) = Xl-~)n (3Vy) , (where 3Vy= 3a~ + 3~)
= ~ P eYi _ T)2 + ~ ~ (3c{) (Ni-m)i=l i Pi 3 i=l Pi m (Ni-l)
131
Also,
where
N.(2Yij - 2Yi)(3Yij - 3Yi) (Ni-m)J. 1E - (N. -1)
,j=l N. m
J. J.
.... 1Var ( T) =- ( V ) ,
3. n 3 Y
"'.... 1COY (2,3T'3T) =n (2.3Vy) ,and
'" '" 1COY (3,2T, 3T) =n (3VY)
~ '" ~Now, to find the expressions for COY (2T, 3,2T) , COY (2T'g,3T) ,
*COY (2~'3T), we proceed as follows:
COY (2~)' 3,2T) = COY [{Q(lT + 2,lT -1,2T) + (1~)2T} , 3,2T]
= Q[COV (IT, 3,2T) + COY (2,lT, 3,2T)
- COY (1,2T , 3,2T)] + (l-Q) COY (2'r'3,2T) •
The following diagram shows the overlapping parts of the first-stage
samples on each of the three occasions. An inspection of this diagram will
help in the determination of covariances.
132
order
n+tJ,nn
o
./'--2,3Tr .....
I '"I T -,,- ............
Slon '"2nd 2T
occasion '" I3rd 2,1~ 3
TI
occasi n --- --'
1
1stoccs
It can be shown that
for
= 0
COY (IT'3,2T) =nfi:~)ln [1.3Vy]1
for 2" S Il <1
where
~i (lYij - 1~i)(3Yij- 3Yi)j=l Ni
1 (Ni - m)
m(Ni
- 1)
Next
1= 0 for 2" S Il < 1
where
'\
~i (2Yij - 2Yi)(3Yij - 3~i) 1j=l Ni m
(N. -m),~
133
for
1= 0 for 2 S ~ < 1 ,
and
1 [ J- v .n 2.3 y
Hence
Cov (2~' 3,.2T.) • Q[~( V) + (1-2~) ( V). ~ 1.3 Y (1_~)2n 2.3 Y
_ (1-2~) ( )J ( ) 1 () 12 1 3v + 1-Q - r'. 3V for 0 < ~ < 2- ,
(1-~) n· y n ~. Y
1= (l-Q) - ( V)n 2.3 y1
for 2' S ~ < 1.
The same procedure when applied to Cov (2~' 2,lT) and Cov (2~' 3T)yields:
~ ....Cov (2~' 2,3T) •
and
1= (l-Q) - ( V )2 2 Y
for
-~ ( V)J + (l-Q) (1-~) ( V), for 0 < ~ <-21~ 1.3 Y n 2.3 Y
= (1~) (1-~) ( V) f 1 < < 1-, n 2.3 Y or 2' - ~
and,
134
Combining these results, we will have from (5.20)
+ 2Q [~1-2» (~) + (1~2fl) (2.3Vy) _ (1-2fl) (1.3Vy)}
1-fl n (1_fl)2 n (1_fl)2 n
+ 2(1-Q) (2.3Vy ) _ 2Q { ~1-2» (1.2VY ) + (1-2fl) (2Vy )n 1-fl n (1_fl)2 n
_ (J.-2fl) (1.2Vy
)} _ 2(1-Q) ! ( V ) _ 2 ( V)](1_fl)2 n n 2 Y (l-fl)n 2.3 Y
+ (1_Q)2 ! ( V ) + 2Q(1-Q) [Q (1-2fl) ( V)n 3 Y t n 1.3 Y
+ ~1-2» ( V) _~ ( V)t + (l-Q) (l-fl) ( V)1-fl n 2.3 Y ~ 1.3 Y) n 2.3 Y
Var ( ~) = Q2 [Q2 [-E...} 1Vy + {1 + Q2 ..k.} 2Vy + {2Qfl2 _ 2Qf.l3 1-fl n l 1-fl n
V V V V_ 2Q2fl2} llx- + --lL.. +~ + 2(1-Q) (2.3 Y)
(l-fl)n (l-fl)n (l-fl)n n
V V V_ 2(l-Q) (U) _ 2 [~ + (1_Q)2 (U) +
n (l-fl)ru n
+ 2'1(1-'1) [(1-'1) (1;;) (2.3Vy) + ~ (3Vy) - ~ (2.3Vy )] ,
1for 2' =s fl < 1 (5.21)
1351 1
For example where I.l = 2' and Q = 2' ' we will have
Var ( ~) = ...l... (1VY) +...2... (2VY) + 20 (-D.)3 16 n 16 n Ib n
Under the assumptions that
l v = 2V = 3V = V ,Y Y Y Y
and
V - V1.2 Y - 2.3 y ,
also defining
Vp = 1.2 Y
-V1VY -v:}";we have
=V2.3 Y ,
If Var (3~) is compared with Var (3T), the variance of the
simple linear estimator
A 1 ni2l.ln Ni m 3Yij
3T = - L: -p L: ,
n i=2l.ln+l i j=l m
we will find that
(5.24 )
The percent gain in efficiency of the linear COIDP9site estimator
x 100 •
136
The percent gain in efficiency is tabulated in Table 5.1.
~ "Table 5.1. Percent gain in efficiency of 3T over 3T for
1 1f.l=2' Q=2
p %Gain in efficiency
.6 negative
·7 4.57
.8 21.21
·9 44.14
.95 59.20
~
Comparing the nature of the gain in efficiency of 3T with that of~. 1. 1
2T for Q = 2' f.l =2·' we see that the gain is much higher. This is
because information from both the first and the second occasions is used.-- ---If the saIIWler wants to use only the information from t.he second occa-
~
sion, the form of the estimator will be like 2T and the percent gain
in efficiency will be as in Table 4.1.
When the proportion of first-stage units partially replaced is
1 ~less than 2"' a more cOIIWlicated expression of Var (3T) will be in-
volved (see the first expression in (5.21».
For example, whenJ.l. = ~ and Q:: ~ , we have
...2... (1.3VY) ~ (~)- 48 n - 48 n (5.25 )
...
137~
To compare the efficiency of 3T with that of the simple estimator
3T for this case, we will again make the following assumptions:
(i) V = V = V = Vly 2y 3Y Y
and
(iii)
V1.3 Y11/
=V
2·3 Y ,
Under such assumptions, we will have
V.1L
n
*The percent gain in efficiency of 3T over
lOp2 + 5!tpl - 36
132 - 54Pl - 10P2x 100
We tabulate the percent gain in efficiency for a series of assumed
values of Pl and P2 in Table 5.2.
11/ 2- In practice, we would expect that P2 < Pl' or perhaps P2 ii Pl;
an ass'UlTlPtion similar to the latter one has been made by many authors.
-Table 5.2. *Percent gain in efficiency of 3T over
assumed values of Pl and P2, for ~
'"3T for some
= ~, Q =~
138
...
..
2 %Gain in efficiencyPl P2=Pl
.60 .36 0.00
·70 .49 7·50
.80 .64 16.50
.90 .81 26.16
·95 .90 33·89
Comparing with Table 4.1, we see that if the correlation pattern
is approximately as assumed, the use of the linear composite estimator
*3T which utilizes past information from~ the first and second
occasions will again be more efficient than the one based on the
second occasion alone.
The case when a = 4 or 5 can be treated similarly. However, when
the number of occasions (!.!., a) becomes larger and larger, the exact
*expression for Var (aT ) given by (5 .19) becomes too involved. An
approximate expression was obtained involving many assumptions on
variance and covariance stability. Discussion on this ~oint is omitted
in view of its algebraic complexity.
5.3.1.2. A Modified Linear Composite Estimator. The structure
of the modified linear composite estimator when the sampling is done
for a > 2 successive occasions is
"-where Tex,ex-l"-
and Tare as defined in (5.15) and (5.16)ex-l,ex
139
.. "- 1 n+(ex-l)~n Ni mT=- L: - L:
ex, ex ~n i=n+(ex-2 )~n+l Pi j=l
..
is the unbiased estimator of exT based on the ~n first-stage units~,
which occur from order n+(ex-2)~n+l to n+(ex-l)~n. ex_1T is the
modified linear composite estimator of ex_1T •~
Expected value of exT'. It can be shown by an argument similar to~ 1:::
that used in the case of exT that exT' is also an unbiased estimator
of exT •~
Variance of T'. From (5.26), we haveex
It can be shown that the three covariances in the last square
bracket are zero, so that
• 140
For example, when a = 3 we have
*Now, Var (2T') has been given previously.
(5.29 )
1IJ.n
Cov (T T) - "T::'""l~_. 3,2' 2,3 - (l-lJ.)n
III A
To obtain expressions for Cov (2T', 3,2T)
we proceed as follows:
,..", "" ,.."Cov (IT, 3,2T) , Cov (2,lT, 3,2T) , Cov (1,2T, 3,2T)
have been derived previously.
From the diagrain showing the overlapping parts of the first-stage
samples on the three occasions given below
141
order
A
T
,occasion
I2,2'"~
sion "-2nd 2,3T -
occasionI"3rd 2 IT
"-
1,2T~ .-A--_.......
1stocca
we see by inspection that
With the above results, we obtain
1for 0 < ~ < 2 '
Similarly,
1.2Vy] + (l-Q) rl-~)n (2Vy)'
for 0 < ~ < ~
1= (l-Q) ( )
l-~n
1for""2 :s ~ < 1 .
142
Using all these results in (5.29), we have
1for 0 < J..I. < 2" '
and
{
V V V V+ 2Q2 1.2 Y _ U _ U-L} + (1_Q)2 (~) + 1 (V)
n n {l-J..I.)n j.ln (l-J..I.)n 3 Y
For example, when 1 1J..I. = 2" ' Q = 2' we will have
Under the assumption that
and
v - V and defining1.2 y - 2.3 y ,
P =
we will have
V1.2 y =v2.3 Y ,
6 [ 2. 10 1 VVar (3T') = 1 + Ib - Ib PJ + .*If T'
3'" '"is compared to the simple estimator 3T where 3T is as
defined previously, we will have the relation:
The percent gain in efficiency of
( ) _ (10p-5)G12 P - 21-10p x 100 •
*T'3
over '"3T for this case is
The gain in efficiency ::i.s tabulated in Table 5.3.
Table 5.3. Percent gain in efficiency of1 1
J..l=2"' Q=2"
"over 3T for
P
.6
·7.8
·9·95
%Gain in efficiency
6.6614.29
23.07
33·33
39·13
• 144*, *,Comparing the gain in efficiency of 3T with 2T for this
case (see Table 4.7) we see again that the gain in efficiency is in-
creased when we use past information from both the first and second
occasions.
1 *When J.l. < 2' J a more complicated expression for Var (3T') is
involved (see the first expression for Var (3~') in (5.30». For the
1 1case J.l. = 3" J Q = 2' J we have
V V V VVar (3~f) = -l (l..1:) + 2.. (U) + 2.£ (U)_ ..£.. (1.3 Y)
32 n 32.n 32 n 32 n
Under the assumptions that
V = V = V =Vly 2y 3y y
and
V = V1.2 Y 2.3 y
and by writing
we have
• 145
~ [40 - ~~l - 2P2]
VVar (3T') = ..L.
n
[1 +8 - (6Pl + 2P2 )] v
= ..L.32 . n
and
* 1 1which implies that 3T' (for ~ =3 ' Q =2) is less efficient thanA ;$l,
the simple estimator 3T as in case of 2~ •
To make the estimator 3~' worthwhile, a proper choice of Q must
be made. For example, if we use Q =t keeping ~ =~ , we will have
for this case,
Under the assumptions given above, we will have
_ A + {286 - 37.8pl - 108P2} V- Var (3T) 512 +
The percent gain in efficiency of ~,3
for this case is
378pl + 108P2 - 286G13(P) = [798 - 378pl - 108P2] x 100.
The gain for some values of Pl and P2 is tabulated in Table
•Table 5.4.
APercent gain in efficiency of 3Tf
1.1. =~ , Q =t for some values of
for
146
2 %Gain in efficiencyPl P2=Pl
.6 .36 negative
·7 .49 6·56
.8 .64 20.04
.9 .81 38.25
.95 ·90 49.69
5.3.1.3. A General Linear Estimator. A general linear estimator
which can be used to estimate aT, when a > 2 is
"'T* - ( T"'*) + b ( T) + ( T) + d( T)a - a a-l a-l,a c a,a-l a,a (5.34 )
where a-lT* is an unbiased estimator of a-lT and is of the same
* '" A Atype as aT; a-l,aT, a,a-1T and a,rxT are as previously defined.
"'*As in the case of 2T, we find that we must choose the weights in
(5.34) such that
b = -a
and
d = l-c
for aT* to be an unbiased estimator of aT. Using these values we find
.."'*Further, by minimizing Var (aT ) with respect to a and c
simultaneously, we obtain the best values of the weights, viz.
147
a = [[var (a,aT)J[Cov (a,a-1T, a-1,aT) - Cov (a-1T*, a,a-1T)~ /
~var (a,aT) + Var (a,a-1T)] [Var (a_1T*) + Var (a-1,aT)
"'* '" A* A ) (A* A)]~- 2 Cov (a_IT, a-l,aT)] - [cov(a_lT , a,a-1T - Cov a_1T 'a-l,aT J
(5.36 )
and
o =[[var (a,aT)][var (a-1T*) + Var (a-1,i) - 2 Cov (a-1T*, a-1,aT)~ /
~var (a,aT) + Var (a,a_1T)][var (a_1T*) + Var (a-1,aT)
- 2 Cov (a_1T*, a-1,aT)] - [Cov (a_1T*, a,a-1T) - Cov (a-1T*'a-1,aT)]~
and consequently (for ready reference)
(1-0) = [[var (a,a-1T)l [Var (a-1T*) + Var (a-1,i) - 2 Cov(a_1T*'a_1,aT)]
- [Cov (a-1T*, a,a-1T) - Cov (a-1T*, a-1,aT)]~ / [[var (a,aT)
+ Var (a,a_1T)][var (a_1T*) + Var (a-l,aT) - 2 Cov (a_1T*, a-l,aT)]
- [Cov (a_1T*, a,a-1T) - Cov (a_1T*, a-1,aT)]~ .
(5.38)
With these values of a, c and l-c substituted in (5.35), we will
obtain the formula for aT*.. The expression however is omtted in view
of its length.
,
148A* A A A
Noting that Cov (a_1T , a,aT), Cov (a-l,aT, a,aT) andA A
Cov (a,a-1T, a,aT) are each equal to zero, we find variance of the
estimator aT* to be
where a and c are as defined in (5.36) and (5.37).
For example, when a =3, we will have
A A* A A* A
+ Var (3,2T)][var (2T ) + Var (2,3T) - 2 Cov (2T , 2,3T)]
- [Cov (2T*, 3,2T) - Cov (2T*, 2,3T)]'
c =[[var (3,3T)][var (2T*) + Var (2,3T) - 2 Cov (2T*, 2'3T)~/
[[var (3,3T) + Var (3,2T)][var (2T*) + Var (2,3T) - 2 COv (2T*'2,3T)]
"'* "'* A 21- [Cov (2T , 3,2@) - Cov (2T , 2,3T)] J ."'* A "'* '"All expressions except Cov (2T , 3,2T) and Cov (2T , 2,3T) have been
derived previously. To find the above two covariances, ·we proceed as
follows:
149
From (4.66) we have
(1-1J. ) A A
+ Cov (2,lT, 3,2T)(1_1J.2p2 )
1J.(1-lJ.p2 ) A A
+ Cov (2,2T, 3,2T) •(1_1J.2p2 )
~ ".With the aid of the diagram used in the discussion of 3T and 3T,
we find that
and
for ~:S IJ. < 1
for 0 < IJ. < ~ ,
150
and
Under the assumptions that V = V =V and V = V2 y 3 y y 2.3 Y 1.2 y ,
so that
V= _...,;;;1;,..;.2;;;....M.Y _
-V1Vy -V2Vy
1we will have for example, when j..L = 2'
,..
a =
and
c = 8
,
The variance of the estimator 3T* given by (5.39) is found to
be
A* AThe percent gain in efficiency of 3T over 3T as measured by
is
[Var (3T) - Var (3T*)]
Var (3T*)x 100
~ ~l
G14(P) = [(64 - l6p2 ) - [(4-2p2)(p-l)} 2]2 - 128(4-p2)(4-3p2)
- 2 [32 - 8p2 - {(4_2p2 )(p-l~ 2J x 100 / 128(4-p2 )(4-3p2)
+ 2 ~2 - 8p2 - t(4_2p2 )(p-l)}2J 2
The gain in efficiency for given values of p is tabulated in
Table 5.5.
Table 5.5. Percent gain in efficiency of /r* over /r for 1-1. = ~
..
/, %Gain in efficiencyp
.6 11.61
·7 16.76
.8 23·83
·9 34.11
·95 41.19
Comparing the nature of the gain in efficiency using this type
of estimator to the case where a = 2, we see again that there is a
slight increase in the gain in efficiency when the sampling is done on
three successive occasions (see Table 4.13).
1When 1-1. < 2' ' the expression for the weights a and c and
Var (3T*) will be more involved and will not be given here.
When a > 3, the problem can be treated similarly but the algebra
will become heavier as the number of occasions (!.~., a) increases.
152
5.3.1.4. A Ratio-Type Composite Estimator. The ratio-type
composite estimator which may also be used to estimate aT when a > 2
is
'"+ (l-Q) Ta
~*where a_1T is the ratio-type composite estimator of a_1T !.~.,
the total of population characteristic of interest for the (a_l)th
'" ."" '" '"occasion. lT, rv-l fVT, rvT , Q are as defined previously_a ,a- .....- ,..... .....~*The estimator aT is not an unbiased estimator of aT but when
properly used, may yield a more efficient estimator than the simple
'"estimator aT •
~*Variance of aT. From (5.40), we have
Var
Applying Lemma II and III, and using only the leading terms, we
~*obtain the approximate expression for Var (aT ) as
~* A **Assuming that ECa_1T) ~ a~lT C!.~.J the bias of a_1T , as an
estimate of IT, is relatively small) the approximate expression fora-**Vax CaT) is reduced to
.. 154
The expressions tor covariances such as
be worked out by applying Lemma III.
For example, when <l = .3 , we will have
(5.44)
'"All expressions except for Cov (3,2T,
llt* '"Cov (2T , 3T), have been given previously.
155llt* '" "'*
2T ), Cov (2,3T, 2T ),
To obtain the approximate
•
expressions for these covariances, we proceed as follows:
Consider
'"COy (2~' 3,2T) ~ COY [Q (2,1;)(1T)} + (l-Q) 2T, 3,2T]
1,2'"
2 IT '" '" '"= Q Cov t(~)(lT), 3,2T} + (l-Q) Cov (2T, 3,2T).
1,2T
By applying Lemma III, we will have
2T -Except for the ratio ("'T) which is a constant, the four sub-covariances
1 ~in the expression above are as previously given in the case Of
3T.
Similarly,
(5.4q)
•
..
156
and
.2'2.2. Estimation of Change in Total
To estimate the change in totals between the current and previous
occasions, !.~., aT - a_1T , where a > 2, four types of estimators
which are extensions of the case a = 2, can be used. We will
indicate the nature of the extension of the theory only briefly.
5.2 .2.1. The Estimator Based on the Linear Composite Estimator.
The estimator based on the linear composite estimator is
~ ~ ~
D = T - Ta-l a a-l
~ Awhere aT and a_1T are the linear composite estimators defined
previously. In view of the demonstration given in connection with
(5.18) it is unbiased.-
Variance: From (5.48), we have
(5.49)
The variances and covariances on the R.H.S. of (5.49) for specific
cases can be worked out as in the equation leading to (5.21).
5.3.2.2. The Estimator Based on the Modified Linear Composite
Estimator. The structure of this estimator is
~ ~ A.u' = T' - T'a-l a a-l
A A,where aT' and a-lT
as defined in (5.26).
157
(5 .50)
are the modified linear composite estimators
Variance: From (5.50), we have
(5.51)
The variances and covariances on the R.H.S. can also be worked out as
in the equation leading to (5.30).
5.3.2.3. A General Linear Estimator. This estimator is
By imposing the condition that
a '+ b' = -1
and
C ' + d' - 1- ,""*so that Da _l takes the form
shall be an unbiased estimator of
(5 .53)
""* I IFurther, minimizing Var (Da
_l ) with respect to a and c, yields
a 1 = [{Var (a,a.T) + COY (a-l,aT'a.,a-lT)} {COV (a.-l,aT'a,a.-lT)
"* ""} { ""- COV(a_lT 'a.,a-lT) - Var (a-l,aT)
- Cov (a-1T*'a-1,aT)} {var (a,a-1T) + Var(a)}J / [{Var (a-1T*)
and
- COY (a.~l~ 'a-l,aT)} {Cov(a._l,a.T'a,a_lT)
cov(a._lT:'a,a-lT)}] I [{var(a._l~) + var(a._l,a.T)
- 2 cov(a_lT~'a_l,aT)}{var(a.,a_lT) + var(a,aT)}
- {cov<CX_1,aT'a,a_1T) - cov(a_1T*'a,a-1'1!)}'
Variance. From (5.53) the variance of the estimator ~-l is
Var (B~_l) = a 12 var(a_lT*) + (l+a 1 )2var(a_l,a~) + c ,2Var (a,a_l;)
" (A* A)+ 2a c Cov a-lT_'a,a-lT
159
5.3.2.4. The Estimator Based on the Ratio-Type Estimator. This
estimator which is not an unbiased estimator of aT - a_1T is given by
A A
A * ~where aT and a_1T are the ratio~type composite estimators of aT
and a_1T respectively.
Variance. From (5.57), we will have
each component
A A*= var(a~) + var(a_l~_)
on the R.H.S. of (5.58) can be worked out for specific
cases as in the equation leading to (5.44).
160
6. SUMMARY AND CONCLUSIONS
6.1. SUlIIllI.8.ry
A multi-stage sampling design and the resulting estimation theory
particularly intended for large scale sample surveys on successive
occasions is developed. The sam;pling design is kept general in the
sense that the selection of the first-stage sampling units is done with
arbitrary (unequal) probabilities. A technique of partial replacement
of first-stage sampling units based on their order of occurrence in
the preliminary saIllJ?le is proposed. This technique is intended to
serve two purposes:
(i) To spread the burden of reporting among respondents which can
be expected to minimize response resistance.
(ii) To enable the sampler to utilize the correlation over time in
the reduction of the variance of several estimates of population totals
and change in totals.
The estimation theory is presented. Four types of estimator which
can be used to estimate the totals of the population characteristic of
interest and the changes in such totals are discussed. Of the four,
the first three estimators which are referred to in this thesis as
the linearcoIllJ?osite estimator~ the modified linear cOIllJ?osite estimator
and the general linear estimator are unbiased estimators •. The fourth
estimator which is referred to as the ratio-type cOIllJ?osite estimator is
a biased estimator but this bias is likely to be small. The expressions
for the variances of these estimators are given. The per cent gains in
.1
162
the Appendix, the estimation of variances and covariances in the rele
vant variance formulae will always be sim;ple in view of the mutual
statistical independence between the first-stage units brought about
by the sim;ple ex;pedient of sam;pling with replacement after each draw
in forming the preliminary sample.
• 16;
7 • LIST OF REFERENCES
1. Cochran, W. G. 196; • Sampling Techniques, 2nd Edition. JohnWiley and Sons, New York.
2. Des Raj. 1954. On sampling with varying probabilities in multistage designs. Ganita, 5:45..;51.
3 . Des Raj. 1965. On sampling over two occasions with probabilityproportionate to size. Annals of Mathematical Statistics,;6:327-3;0.
4. Eckler, A. R. 1955. Rotation sampling. Annals of MathematicalStatistics, ;6:664-685.
5. God.aIribe, V. P. 1955. A unified theory of sampling from fin,itepopulation. Journal of the Royal Statistical Society, SeriesB, 17:269-278.
6. Hansen, M. H., Hurwitz, W. N., Nisselson, H. and Steinberg, J.1955 • The redesign of the census current population survey'.Journal of the American Statistical Association, 50:701-719.
•
7· Hansen, M. H., Hurwitz, W. N., and Madow, W. G.Survey Methods and Theory, Vol. I and II.Sons, New York •
1953 • SampleJohn Wiley and
8. Jessen, R. J. 1942. Statistical investigation of a sample surveyfor obtaining farm facts. Iowa Agricultural Experiment StationResearch Bulletin, No. 304. Ames, Iowa.
9. Koop, J. c. 1963. On the axioms of sample formation and theirbearing on the construction of linear estimators in samplingtheory for finite universes. Metrika 7 (2 and 3): 81-114and 165-204.
10. Madow, W. G. 1949. On the theory of systematic sampling, II.Annals of Mathematical Statistics, 20:333-;54.
11. Onate, B. T. 1960. Development of multi-stage designs forstatistical surveys in the Philippines. M1meo-Multi1ithSeries, No. ;, Statistical Laboratory, Iowa State University,Ames, Iowa.
12 • Patterson, H. D. 1950. Sampling on successive occasions withpartial replacement of units. Journal of the Royal StatisticalSociety, Series B, 12:241-255.
..
164
13 . Rao, J. N. K. 1961. On sampling with varying probabilities andwith replacement in sub-sampling designs. Journal of· theIndian Society of Agricultural Statistics, 13:211-217.
14. Rao, J. N. K. and Graham, J. E. 1964. Rotation designs forsampling on repeated occasions. Journal of the AmericanStatistical Association, 59:492-509.
15. Sukhatme, P. V. 1954. Sampling Theory for Surveys with Applications. The Indian Society of Agricultural Statistics, NewDelhi, India, and the Iowa State College Press, Ames, Iowa.
16. Tikkiwal, B. D. 1955. Multiphase sampling on successive occasions. Unpublished Ph. D. thesis, Department of ExperimentalStatistics, North Carolina State University, Raleigh, NorthCarolina.
17. Tikkiwal, B. D. 1958. Theory of successive two-stage sampling.Abstract in Annals of Mathematical Statistics, 29:1291.
18. Ware, K. D. and Cunia, T. 1962. Continuous forest inventory withpartial replacement of samples. Forest Science, Monograph 3.The Society of American Foresters.
19. Woodruff, R. S. 1959. The use of rotating samples in the CensusBureau's monthly surveys. Proc. Social Statistics Section,American Statistical Association, 130-138.
20. Yates, F. 1960. Sampling Method for Censuses and Surveys.Charles Griffin and Co., London.
..8. APP~DICES
8.1. Note on the Estimation of Sa.rtW11ng Variances
To verify that 1(j~y + 1(j~ is unbiasedly est11!lated by
165
where
nL:
1=1
y'(1 1 ,,)2-- T
Pi 1
n-1,
and
m
1Yi' = N. L:J. j=l
Consider
... 1 n N1 mT=- L: L:
1 n i=l Pi j=l
nL:
1=1
n-11 n 1Yi' 2 2
= - E[ L: (-) - n ")n-1 i=l Pi 1
T
n Y' 2= 1: L: E(l i) - 1rr?- =
n 1=1 Pi
166
Now
N ~2
10"i= Z
Pi-
i=lm
N ~2
·lO"i= Z
Pii=lm
• •
.. Similarly" it can be verified that 20"~y + 20"~ is unbiasedly esti
mated by
and
n~n
Ei=J..m+l n-l where
161
And similarly, it can be verified that 1.2CTbyy + 1.2CTwyy is
unbiasedly estimated by
nE
'i=;.tn+l (l-IJ )n-l
'"2,lT)
8.20 Note on the Efficiency of the Estimators when lV ~ 2VY . Y
The nature of the gt:l.in in efficiency using the estimators discussed
in Chapter 4 has been examined under the assumption of equal varianceso
When the variances differ on each occasion, we would expect that the
gain in efficiency will be different from what has been tabulated.
We will indicate the nature of the change by the following exam;ples:
~I. The Linear Composite Estimator 2T
1 1For example, Q "" 2' ' IJ == 2" and assuming that lVy "" A. 2Vy
when A. > 0 and ~ 1-
From (4.35), we will have,
where
Table 8.2. '" '"The percent gain in efficiency of 2T over 2T when1 1 5Q=- IJ.=- and V =T.'(V)2' 2 ly Lj- 2y
p %Gain in Efficiency
.6 negative
·7 2.56
.8 12.04
·9 23.84
·95 30·55
169
Comparing with Table 8.1, it is interesting to note that, in
'"using this type of estimator (i.e. 2~)' the efficiency is increased
when lVy > 2Vy' otherwise it is decreased, value by value of p.
~f
II. The Modified Linear Com,posite Estimator 2T
For exgmple, Q = ~ IJ. = ~ and assuming again that lVy = A(2Vy)
where A > 0 and ~ 1-
From (4.53), we 'Will have by substitution
where
'" 2V
Var ( T) = U .:2 n
~I '"The per cent gain in efficiency of 2T over 2T is
[
2p Ii - A J.4 + A - 2p VA J x 100 •
NORTH CAROLINA STATE UNIVERSITY
INSTITUTE OF STATISTICS
(Mimeo Series available for distribution at cost)
423. Miller, H. D. Generalization of a theorem of Marcinkiewicz. 1965.424. Johnson, N. L. Paths and chains of random straight-line segment. 1965.425. Nasoetion, Andi H. An evaluation of two procedures to estimate genetic and environmental parameters in a simul·
taneous selfing and partial diallel test-crossing design. 1965.426. Bose, R. C. Error detecting and error correcting indexing systems for large serial numbers. 1965.427. Chakravarti, I. M. On the construction of different sets and their use in the search for orthogonal Latin squares and
error correcting codes. 1965.428. Bose, R. C. and Chakravarti, I. M. Hermitian varieties in a finite projective space PG (N, q"). 1965.429. Potthoff, R. F. and Maurice Whittinghill. Maximum-likelihood estimation of the proportion of non-paternity. 1965.430. Tranquilli, G. B. On the normality of independent random variables implied by intrinsic graph independence without
residues. 1965.431. Miller, H. D. Geometric ergodicity in a class of denumerable Markov chains. 1965.432. Braaten, M. O. The union of partial diallel mating designs and incomplete block environmental designs. 1965.433. Harvey, J. R. Fractional moments of a quadratic form in non-central normal random variables. 1965.434. Nixon, D. E. A study of the distributions of two test statistics for periodicity in variance. 1965.435. Basu, D. Problems related to the existence of maximal and minimal elements in some families of statistics (sub-fields).
1965.436. Gun, Atindramohan. The use of a preliminary test for interactions in the estimation of factorial means. 1965.437. Smith, W. L. A theorem on functions of characteristic functions and its application to some renewal theoretic random
walk problems. 1965.438. Nuri, Walid, Fourier methods in the study of variance fluctuations in time series analysis. 1965.439. Barlotti, A. Some topics in finite geometrical structures. 1965.440. Bhapkar, V. P. and Gary Koch. On the hypothesis of "no interaction" in the three dimensional contingency tables. 1965.441. Hall, W. J. Methods of sequentially testing composite hypotheses with special reference to the two-sample problem.
1965.442. Bohrer, R. E. On Bayes sequential design of experiments. Ph.D. Thesis. 1965.443. Hoeffding, Wassily. On probabilities of large deviations. 1965.444. Kanofsky, Paul B. Parametric confidence bands on cumulative distribution functions. 1965.445. Potthoff, R. F. A non-parametric test of whether two simple regression lines are parallel. 1965.446. Metzler, C. M., Gennard Matrone and H. L. Lucas, Jr. Estimation of transport rates by radioisotope studies of non-
steady-state systems. Ph.D. Thesis. 1965.447. Koch, Gary. A general approach to estimation of variance components. 1965.448. Assakul, Kwanchai. Testing hypotheses with categorical data subject to misclassification. Ph.D. Thesis. 1965.449. Bhapkar, V. P. and Gary Koch. On the hypothesis of "no interaction" in the four dimensional contingency tables. 1965.450. Bhapkar, V. P. Categorical data analogs of some multivariate tests. 1965.451. Chakravarti, I. M. Bounds on error correcting codes (non-random). 1965.452. Anderson, R. L. Non-balanced experimental designs for estimating variance components. Presented at Seminar on
Sampling of Bulk Materials, Tokyo, Japan. November 15-18, 1965.453. Quade, Dana. On analysis of variance for the k-sample problem. 1965.454. Ikeda, Sadao, JunjiroOgawa and Motoyasu Ogasawara. On the asymptotic distribution of F-statistics under the null
hypothesis in a randomized PBIB design and M associate classes under the Neyman model. December 1965.455. Ikeda, Sadao. On certain types of asymptotic equivalence of real probability distributions. I. Definitions and some of
their properties. December 1965.456. Yandle, David. A test of significance for comparing two different systems of stratifying the same population. Ph.D.
Thesis. 1965.457. Mesner, Dale M. The block structure of certain PBIB designs of partial geometric types. 1966.458. Sen, P. K. A class of permutation tests for stochastic independence. I. 1966.459. Koch, Gary. Some considerations about the effect of redundancies and restrictions in the general linear regression
model. 1966.460. van der Vaart, H. R. An elementary derivation of the Jordan normal form with an appendix on linear spaces. A
didactical report. 1966.461. Smith, Walter S. Some peculiar semi-Markov processes. 1966.462. Sen, P. K. On a class of non-parametric tests for bivariate interchangeability. 1966.