Computational Pool-Testing with Retesting Strategy

Cox Lwaka Tamba & Jesse Wachira Mwangi
International Journal of Scientific and Statistical Computing (IJSSC), Volume (3) : Issue (2), 2012 47
Computational Pool-Testing with Retesting Strategy
Cox Lwaka Tamba clwaka@yahoo.com
Faculty of Science/Department of Mathematics/Division of Statistics
Egerton University
P.o. Box 536-20115, Egerton, Kenya
Jesse Wachira Mwangi jdmwangi@egerton.ac.ke
Faculty of Science/Department of Mathematics/Division of Statistics
Egerton University
P.o. Box 536-20115, Egerton, Kenya
.
Abstract
Pool testing is a cost effective procedure for identifying defective items in a large population. It also
improves the efficiency of the testing procedure when imperfect tests are employed. This study
develops computational pool-testing strategy based on a proposed pool testing with re-testing
strategy. Statistical moments based on this applied design have been generated. With advent of
digital computers in 1980‘s, pool-testing with re-testing strategy under discussion is handled in the
context of computational statistics. From our study, it has been established that re-testing reduces
misclassifications significantly as compared to Dorfman procedure although re-testing comes with a
cost i.e. increase in the number of tests. Re-testing considered improves the sensitivity and specificity
of the testing scheme.
Keywords: Pool, Pooling, Re-Specificity, Sensitivity, Tests, Misclassifications.
1. INTRODUCTION
The idea of pool testing was initiated by Dorfman (1943) during World War II as an economical
method of testing blood samples of army inductees in order to detect the presence of infection.
Pooling procedures involve putting together individuals to form a group/pool and then testing the
group rather than testing each individual for evidence of a characteristic of interest. A negative
reading indicates that the group contains no defective items and a positive reading indicates the
presence of at least one defective individual in the group. There are two objectives of pool testing:
classification of the units of a population as either defective or non-defective (Dorfman, 1943) and
estimation of the prevalence of a disease in a population (Sobel and Elashoff, 1975). Pooling
procedures have proved to reduce the cost of testing when the prevalence rate is low. This is
because if a pool tests negative, it implies all its constituent members are non-defective and hence it
is not necessary to test each member of the pool. A procedure of classifying the population into
defective and non-defective when each unit i of the population has a different probability pi of being
defective (which is called a generalized binomial group test) problem has been studied (Hwang,
1975). The generalized binomial group test problem reduces to a binary pool testing problem which is
the Dorfman, (1943) procedure when all the units have the same probability p of being defective.
Hwang (1976) has also considered a pool testing model in the presence of dilution effect. Dilution
effect in this case refers to a situation where a pool which contains a few defective items may be
misidentified as containing no such items, especially when the size of the pool is large.
Johnson et al. (1992) has studied the cost effectiveness of pooling algorithm for the objective of
identifying individuals with a characteristic of interest using hierarchical procedures. In this procedure,
each pool that test positive is divided into two equal groups, which are tested, groups that tested
positive are further subdivided and tested and so on. This work has been extended by considering
pooling algorithms when there are errors and showed that some of these algorithms can reduce the
error rates of the screening procedures (the false positives and false negatives) compared to
individual testing (Litvak et al., 1994). Computational statistics has been used in pool testing to
compute the statistical measures when perfect and imperfect tests are used has been considered
(Nyongesa and Syaywa, 2011; Nyongesa and Syaywa, 2010; Tamba et al., 2012).

Pool testing has vast applications (Sobel and Groll, 1966). It has been applied industries, and recently
it has been applied in screening the population for the presence of HIV antibody (Kline et al., 1989
and Manzon et al., 1992). Pool testing has been used in screening HIV antibody to help curb the
further spread of the virus (Litvak et al., 1994).
In this study, we consider the computation of statistical measures based on a pool testing with re-
testing strategy via computer package MATLAB. The tests used in this procedure are assumed
imperfect i.e. the specificity and sensitivity are less than 100%. The rest of the paper is arranged as
follows: Section 2 formulates the problem while the pool-testing with re-testing strategy is discussed in
Section 3. Section 4 provides the number of tests and moments while the results in this design are
provided in Section 5. Misclassifications in the proposed testing design are discussed in Section 6.
Section 7 provides the discussion and conclusion to this study.
2. PROBLEM FORMULATION
Consider a population of size N pooled into n pools each of size k. Each pool is subjected to an initial
test. We re-test each pool irrespective of whether it tests positive or negative on the initial test. Pools
that test positive on re-testing of pools that initially tested positive and negative, their constituent
members are tested individually. We wish to establish the number of defective individuals in the
population N. For efficiency and cost effectiveness, pool the population N into n independent pools
each of equal sizes say k. The procedure is described diagrammatically below.
Pools
Figure 1: Diagrammatic description of the Pool Testing with Re-testing Strategy
The figure shows the n constructed pools and the test result on the i
th
pool, for i=1, 2, …, n. The result
on the test is binary: positive or negative. The analysis in this study will require the following indicator
functions:
Let
1;
0;
th
i
if thei pool tests positiveonthetest kit
T
otherwise

= 

' 1;
0;
th
i
if thei pool test is positive on the re test on the test kit
T
otherwise
 − −
= 

1;
0;
th
i
if thei pool is positive
D
otherwise

= 

1;
0;
th th
ij
if the j individual inani pool tests positiveonthetest kit
T
otherwise

= 

and
1;
0;
th th
ij
if the j individual inthei pool is positive with probability p
otherwise
ζ

= 

1 2 i n
, , … … … … … … … … , , … … … … … … … ,
+ve - ve
- ve+ve+ve - ve

The indicator functions provided above are essential in the subsequent developments. The
constituent members of the i
th
pool will be represented by 1 2( , ,..., ,..., )i i ij ikζ ζ ζ ζ or simply 1{ }k
ij jζ = .
Clearly,
1 2Pr( 0) Pr( 0, 0,..., 0,..., 0)i i i ij ikD ζ ζ ζ ζ= = = = = = (1)
by definition. For analysis purposes, we shall assume that the constituent member of a pool act
independently of each other, hence
Pr( 0) (1 )k
iD p= = − (2)
where p is the prevalence rate.
3. POOL TESTING WITH RE-TESTING STRATEGY
Let N be a universal set and ξ be a δ − field on N . Let 11X , 12X , 21X and 22X be random
variables defined on N . Now subdivide N into n partitions representing pools each of size k . We
perform a test on each pool as discussed above. Basically, let 11X be the number of pools that test
positive on re-testing initial declared positive pools, 12X be the number of pools that test negative on
re-testing initial declared positive pools, 21X be the number of pools that test positive on re-testing of
pools that initially tested negative test and 22X be the number of pools that test negative on retesting
initially declared negative pools. Let p be the probability measure on ξ such that an individual is
positive (prevalence rate). We derive new set functions 1π , 2π , 3π and 4π onξ , where
'
1 Pr( 1, 1)i iT Tπ = = = ,
'
2 Pr( 1, 0)i iT Tπ = = = ,
'
3 Pr( 0, 1)i iT Tπ = = = and
'
4 Pr( 0, 0)i iT Tπ = = = . To obtain these probabilities, we require the application of the law of total
probability (Ross, 1997). First,
2
1 (1 ) (1 ) [1 (1 ) ] .k k
p pπ φ η= − − + − − (3)
where η is the sensitivity of the test kit and φ the specificity of the test kit. By sensitivity, we mean the
probability of correctly classifying a defective pool or defective individual while φ is specificity of the
test kits and by specificity here means the probability of correctly classifying a non-defective pool or
non-defective individual. Ideally, we have introduced the error element in our model. The error
component will be assumed to be based on the manufacturers’ specifications and will remain constant
in the entire experiment. That is, sensitivity and specificity will remain constant at group level and
individual level. Clearly [ ]0,1p∈ and so
2
11 φ π η− ≤ ≤ which implies that 1π is a continuous
function bounded below by 1 φ− and above by
2
η . Next,
2 (1 )(1 ) (1 )[1 (1 ) ].k k
p pπ φ φ η η= − − + − − − (4)
The probability
'
Pr( 0, 1)i iT T= = is,
3 (1 )(1 ) (1 ) 1 (1 )k k
p pπ φ φ η η  = − − + − − −  (5)
and the probability
'
Pr( 0, 0)i iT T= = has been derived as,
2 2
4 (1 ) (1 ) 1 (1 )k k
p pπ φ β  = − + − − −  (6)
The probabilities 1 2 3, ,π π π and 4π will enable us to compute the joint probability distribution of 11X
, 12X , 21X and 22X . The joint probability density function of 11X , 12X , 21X and 22X is a multinomial
probability density i.e.,
( ) 11 12 1211 12 21
11 12 21 22, , , 2 11 12 11 12 12 22 1 2 3 1 2 3( , , ) 1 .
n
n x x xx x x
X X X Xf x x x x x x x π π π π π π
− − − 
= − − − 
 
(7)
Utilizing the probabilities 1π , 2π , 3π , 4π and Model (7) above, we propose a computer intensive
pool-testing with re-testing strategy. With the advent of the digital computers in early 1980’s,

computational statistics has evolved (Martinez and Martinez, 2002; L’Ecuyer, 2004). In a similar
format we wish to develop a computational pool-testing with re-testing model. The next section
discusses the moments of the number of tests in the testing scheme.
4. THE NUMBER OF TESTS AND MOMENTS
Model (7) is of major interest in this study. The overall number of tests in this design is
11 211 2Z n kX kX= + + + , (8)
where n is the number of pools and k is the pool size. It then follows that,
[ ] [ ]1 31 2 .E Z n kn π π= + + + (9)
In the field experiments, the sensitivity and specificity of the test kits are normally provided by the
manufacturers’ specifications. Therefore, for given η andφ , we can compute (9) and similarly, the
variance of the number of test is
[ ]2
1 1 3 3 1 3( ) (1 ) (1 ) 2Var Z k n n nπ π π π π π= − + − − (10)
from which, the standard deviation is given by [ ]2
1 1 3 3 1 3(1 ) (1 ) 2k n n nπ π π π π π− + − − .
We shall utilize Equations (9) and (10) to generate the mean, standard deviation in the proposed pool-
testing with re-testing strategy. Next, we consider misclassification arising from this testing design.
5. MISCLASSIFICATIONS
Note that since we allowed testing with errors in our design as is the case real life problem i.e. the test
kits in use are not 100% perfect, two possible misclassifications can arise in practice; false- positive
and false- negatives. A false- positive refers to a non- defective item being classified as defective
whereas a false- negative means that a defective item is classified as non-defective. The probability of
correctly classifying a defective individual is referred to as sensitivity. The sensitivity of the testing
procedure is derived as,
Sensitivity=
' '
Pr( 1, 1, 1| 1) Pr( 0, 1, 1| 1)i i ij ij i i ij ijT T T T T Tδ δ= = = = + = = = =
=
2
η . (12)
The probability of false positives arising from this model is
2
1 .pf η= − (13)
We know that η <1 in practice, this implies that
2
η <η , hence pool-testing scheme lowers the
sensitivity in general. The sensitivity of this procedure is the same as that of the pool testing without
re-testing. (c.f Tamba et al., 2012). We derive the probability of correctly classifying a non-defective
individual herein referred as specificity of the testing procedure.
( ) ( ) ( )
' '
' '
1 12
Pr( 1, 0 | 0) Pr( 0, 0 | 0)
Pr( 1, 1, 0 | 0) Pr( 0, 1, 0 | 0)
2 1 1 1 1
i i
i i
i ij i ij
i ij ij i ij ij
k k
Specificity T T T T
T T T T T T
p p
δ δ
δ δ
φ φ η ηφ
− −
= = = = + = = = +
= = = = + = = = =
      = − − + − + + − −      
(14)
One minus the specificity of the testing scheme yields the probability of false negative as
( ) ( ) ( )
1 12
1 2 1 1 1 1 .
k k
nf p pφ φ η ηφ
− −
    = − − − + − + + − −       
(15)
To investigate the performance of this design we shall utilize Equation (13) and (15), in computing the
false-positive and false-negative. The pool size in (14) appears to be 1k − since the
th
j individual is
known to be non- negative and we have employed sampling without replacement.
6. RESULTS
To this end, we have presented formulas that can be used to compute the central moments of the
number of tests in group testing with re-testing scheme. We illustrate the procedure by computing the
central moment measures for various sensitivity and specificity. In the tables we have used the word
total testing cost to imply the overall cost of performing the procedure.

Table 1: Various characteristics for pool testing with re-testing strategy with 1000 runs, N
=100, k=10, 99η φ= = %
Characteristics p=0.01 p=0.05 p=0.1
µ σ µ σ µ σ
Number of
defectives
9.7590 2.1833 29.5240 4.6742 54.0840 6.2065
Number of
defective groups
4.7710 1.8682 15.9250 2.4625 21.7910 1.4881
Number of group
tests
51.0000 - 51.0000 - 51.0000 -
Number of
individual tests
95.5420 37.3640 318.5000 49.2500 435.8200 29.7620
Total number of
tests
146.4200 37.3640 369.5000 49.2500 486.8200 29.7620
Total testing cost 29.2840 37.3640 73.9000 49.2500 97.3640 29.7620
Percentage
savings
70.7160 37.3640 26.1000 49.2500 2.6360 29.7620
=500, k=20, 99η φ= = %
µ σ µ σ µ σ
Number of
defectives
6.0190 1.0691 9.4710 2.0904 13.8480 2.7201
Number of
defective groups
1.3370 0.7994 4.0610 1.4019 6.3180 1.3284
Number of group
tests
21.0000 - 21.0000 - 21.0000 -
Number of
individual tests
13.3700 7.9944 40.6100 14.0190 63.1800 13.2840
Total number of
tests
34.3700 7.9944 61.6100 14.0190 84.1800 13.2840
Percentage
savings
65.6300 7.9944 38.3900 14.0190 15.8200 13.2840
=100, k=10, 95η φ= = %
µ σ µ σ µ σ
Number of
defectives
2.0010 0.9232 6.0020 2.2183 10.8000 3.0205
Number of
defective groups
1.0060 0.9172 4.0700 1.4931 6.5840 1.4860
Number of
group tests
21.0000 - 21.0000 - 21.0000 -
Number of
individual tests
10.0600 9.1720 40.7000 14.9310 65.8400 14.8600
Total number of
tests
31.6600 9.1720 61.7000 14.9310 86.8400 14.8600
Total testing
cost
31.6600 9.1720 61.7000 14.9310 86.8400 14.8600
Percentage
savings
68.3400 9.1720 38.3000 14.9310 13.1600 14.8600

µ σ µ σ µ σ
Number of
defectives
29.4820 2.2064 47.3720 4.27323 70.2500 6.1414
Number of
defective groups
5.2730 1.8166 15.5170 2.2012 21.0340 1.4323
Number of group
tests
51.0000 - 51.0000 - 51.0000 -
Number of
individual tests
105.4600 36.3320 310.3400 44.0240 420.6800 28.6460
Total number of
tests
156.4600 36.3320 361.3400 44.0240 471.6800 28.6460
Percentage
savings
68.7080 36.3320 27.7320 44.0240 5.6640 28.6460
=500, k=20, 95η φ= = %
Table 5: Number of false positives in the pool testing strategy for different pool sizes
99η φ= = %
Table 6: Number of false positives in the pool testing strategy for different pool sizes
95η φ= = %
Probability,
p
N =100, k=10 N =500, k=20 N =1000, k=20
µ σ µ σ µ σ
0.01 0.5652 0.7142 2.8714 1.6098 5.7724 2.2825
0.02 0.6572 0.7701 3.3189 1.7307 6.6502 2.4499
0.03 0.7399 0.8172 3.7829 1.8477 7.4711 2.5967
0.04 0.8429 0.8722 4.2054 1.9482 8.4221 2.7570
0.05 0.9141 0.9083 4.6254 2.0431 9.2276 2.8838
0.1 1.3625 1.1089 6.8062 2.4784 13.6571 3.5105
0.15 1.7953 1.2729 9.0324 2.8551 18.0663 4.0379
N =100, k=10 N =500, k=20 N =1000, k=20
µ σ µ σ µ σ
0.01 0.0388 0.1949 0.1964 0.4387 0.3941 0.6215
0.02 0.0593 0.2410 0.2940 0.5368 0.5912 0.7612
0.03 0.0767 0.2741 0.3915 0.6194 0.7825 0.8757
0.04 0.0987 0.3110 0.4902 0.6932 0.9809 0.9805
0.05 0.1171 0.3387 0.5805 0.7543 1.1705 1.0711
0.1 0.2147 0.4587 1.0749 1.0264 2.1483 1.4510
0.15 0.3143 0.5552 1.5672 1.2393 3.1239 1.7498

Table 7: Number of false negatives in the pool testing strategy for different pool sizes
99η φ= = %
Probability,
p
N =100, k=10 N =500, k=20 N =1000, k=20
µ σ µ σ µ σ
0.01 0.6007 0.7726 4.8543 2.1918 9.7127 3.1004
0.02 0.9303 0.9597 7.8444 2.7771 15.6995 3.9287
0.03 1.2270 1.1003 10.2817 3.1709 20.5605 4.4836
0.04 1.4941 1.2123 12.2386 3.4512 24.4819 4.8812
0.05 1.7317 1.3033 13.8131 3.6594 27.6312 5.1757
0.1 2.5874 1.5842 17.8083 4.1317 35.6026 5.8419
0.15 3.0186 1.7049 18.5098 4.2034 39.0662 5.9485
Table 8: Number of false negatives in the pool testing strategy for different pool sizes
95η φ= = %
Remark : In all the above tables we have; , tanmean s dard deviationµ σ= =
In the next section, we provide the discussion of our findings and the conclusion to the study.
7. DISCUSSION AND CONCLUSION
This study has presented a computational pool testing strategy with re-testing. It has been shown
from the results; Tables 1, 2, 3 and 4 that when the pool size and prevalence rate are small,
significant savings are realized. This is an empirical result since pool testing is only feasible when the
prevalence rate is small otherwise individual testing is preferred. Similarly large pools are prone to
increase the dilution effect and hence increase the misclassifications. It has been established that re-
testing pools increases the cost of testing however, the misclassifications significantly reduce as
compared to the Dorfman procedure when imperfect tests are used (Tamba et al., 2011). The results
in Tables 5, 6, 7 and 8 show that the higher the efficiency of the tests, the lower the misclassifications.
This implies that pool testing should be carried out when specificity and sensitivity of the testing
procedure are high. It has also been noted that this re-testing strategy improves the specificity and
sensitivity of the testing procedure. Misclassifications are high when the prevalence rate is high and
the efficiency of the test kits is low.
8. REFERENCES
1. C. L Tamba, K. L. Nyongesa, J. W. Mwangi, (2012). “Computational Pool-Testing Strategy”.
Egerton University Journal, 11:51-56.
2. E. Litvak, X. M. Tu, and M. Pagano, (1994). “Screening for the presence of a disease by pooling
sera samples.” Journal of the America statistical Association, 89, 424-434.
3. F.K. Hwang, (1975). “A Generalized Binomial Group Testing Problem”. Journal of the American
Statistical Association, 70,923- 926.
4. F.K. Hwang, (1976). “Group testing with a dilution effect”. Biometrika 63, 611-613.
5. L. K. Nyongesa and J. P. Syaywa, (2010). “Group Testing with Test Errors Made Easier.”
International Journal of Computational Statistics. Volume (1): Issue (1).
Probability,
p
N =100, k=10 N =500, k=20 N =1000, k=20
µ σ µ σ µ σ
0.01 0.0929 0.3046 0.8839 0.9393 1.7676 1.3283
0.02 0.1679 0.4093 1.5640 1.2486 3.1289 1.7660
0.03 0.2356 0.4848 2.1158 1.4514 4.2329 2.0529
0.04 0.2957 0.5430 2.5609 1.5960 5.1248 2.2577
0.05 0.3507 0.5911 2.9186 1.7031 5.8332 2.4077
0.1 0.5438 0.7352 3.8264 1.9477 7.6527 2.7545
0.15 0.6447 0.7999 3.9813 1.9859 7.9683 2.8094

6. L. K. Nyongesa and J. P. Syaywa, (2011). “Block Testing Strategy with Imperfect Tests and its
Improved Efficient Testing Model for Donor Blood.” Communication in Statistics-
Computational Statistics. (Submitted).
7. L. K. Nyongesa, (2005). “Hierarchical Screening with Retesting in a low Prevalence Population.”
The Indian Journal of Statistics.66, 779-790.
8. L. K. Nyongesa, (2010). “Dual Estimation of Prevalence and Disease Incidence in Pool-Testing
Strategy.” Communication in Statistics Theory and Method. Vol. (1): Issue (1).
9. M. Sobel, and P.A. Groll, (1966). “Binomial Group-Testing with an Unknown Proportion of
Defectives.” American Statistical Association and American Society for Quality, 8,631-656.
10. M. Sobel, and, R.M. Elashoff, (1975). “Group-testing with a new goal, Estimation.” Biometrika, 62,
181-193.
11. N. L. Johnson, S. Kotz, and, X Wu, (1992). “Inspection errors for attributes in quality control.”
London; Chapman and Hall.
12. O.T. Monzon, F.J.E Palalin, E. Dimaal, , A.M. Balis, C. Samson, and S. Mitchel, (1992).
“Relevance of antibody content and test format in HIV testing of pooled sera.” AIDS, 6, 43-48.
13. P. L’Ecuyer,, (2004). Uniform Random Number Generation. Handbook of Computational
Statistics. Springer-Verlag, Berlin, pp. 35–70.
14. R. B. Hunt, L. L. Ronald, M. R. Jonathan (2004). A Guide to MATLAB for Beginners and
Experienced Users. Cambridge University Press, pp.101-119.
15. R. Dorfman, (1943). “The detection of defective members of large population”. Annals of
Mathematical Statistics 14, 436-440.
16. R.L. Kline, T. Bothus, R. Brookmeyer, S. Zeyer,, and T. Quinn, (1989). “Evaluation of Human
Immunodeficiency Virus seroprevalence in population surveys using pooled sera.” Journal of
clinical microbiology, 27, 1449-1452.
17. S.M. Ross, (1997). Introduction to Probability Models. 6th
Edition, Academic Press, pp 4-7.
18. W.L. Martinez,and A.L. Martinez,, (2002). Computational Statistics Handbook with MATLAB.
Chapman & Hall/CRC, pp. 96-115.

Computational Pool-Testing with Retesting Strategy

More Related Content

What's hot (17)

Viewers also liked (10)

Similar to Computational Pool-Testing with Retesting Strategy (20)

More from Waqas Tariq (20)

Recently uploaded (20)

Computational Pool-Testing with Retesting Strategy