DOI: 10.1007/s10967-008-0503-3 Journal of Radioanalytical and Nuclear Chemistry, Vol. 276, No.

2 (2008) 305–312

Bayesian analysis for very-low-background counting of short-lived isotopes:

Lowest minimum detectable quantity
K. A. Mathews,* D. W. Gerts
Air Force Institute of Technology, WPAFB, 2950 Hobson Way, Bldg 640, Wright-Patterson AFB, OH 45433-7765, USA

(Received March 2, 2007)

Conventional statistical analyses of counting measurements with a paired-count blank and sample yield unacceptably-large estimates of uncertainty
that reduce measurement sensitivity when applied to very-low-background detection systems. An alternative is presented here: Bayesian analysis
using longer-duration background measurements, appropriate modeling of the background, and a binomial distribution of decay-induced counts
valid even for short-lived isotopes. We develop the needed formulae and demonstrate how the decision level and sample measurement duration are
optimized jointly to produce the lowest minimum detectable quantity subject to constraints of specified acceptable risks of false detection and false
failure to detect. A frequentist’s interpretation is maintained by using equal-likelihood prior distributions.

Introduction independent of Sj.” We refer to this assumption as a flat

prior and use it here. In this way, Bayes theorem
Background provides all the information about the unknown to be
measured that is provided by the data and physical
In many applications, it is customary to perform constraints, and no more. Here, for example, the
paired sample and blank counting measurements. constraint of non-negative counts appears in the limits of
Decision rules for determining whether a sample sums and integrals. Gaussian approximations violate this
contains radioactivity have been based on a comparison constraint.
of the sample and blank counts (or count rates). STROM Bayesian and maximum likelihood estimation
and MACLELLAN1 provide an excellent review of the analyses have been performed3–5 using Poisson, rather
development and variety of such decision rules, and an than Gaussian, distributions for the background and
extensive bibliography (not duplicated here due to space sample decay-induced counts, which extends their
constraints). Of eight such rules, seven contain a applicability to small counts, but still limits them to
constant (kD) that is the number of standard deviations counting times that are very short compared to the half-
from the mean of a normal distribution to contain the life of the isotope being measured. Our work is original
fraction 1–D of events. Thus, in one way or another, in several respects: we use a binomial distribution for
each of those seven schemes assumes a normal the decay-induced counts, which eliminates the latter
distribution, so they should be expected to fail for low restriction; we treat the counting time for a short-lived
counts. STROM and MACLELLAN found that with very isotope as a variable to be adjusted during the
low background, the achieved false detection rates using experiment design stage so as to maximize sensitivity;
the rules were substantially different from the design and our analysis explicitly accounts for decay between
value used in calculating the decision levels. We blame the time the sample is drawn or created and the start of
the uncertainty in the background caused by the the counting measurement.
inappropriate statistical assumptions in the decision
rules as underlying their observation that “our lowest The detection problem and sensitivity
background detectors may not be the most sensitive”.
This uncertainty is a consequence of using the paired A sample containing n atoms of an isotope with
blank to estimate the background count rate – it does not decay constant O is obtained at time t0, at which time it
contain sufficient information. Worse, for very low contains n atoms (unknown) of the radionuclide that is
background, the blank count is often zero. Then a point- to be measured. The count (measurement) commences at
estimator method yields “no background with no time t1, after a delay time 'td = t1–t0, and ends at time t2,
uncertainty”. Our analysis provides more sensitivity for after a counting time (measurement duration) 'tc = t2–t1,
lower background at all background levels. resulting in c total counts (known), consisting of k
BLACK and THOMPSON2 provide a concise foreground counts (unknown) that were caused by the
introduction to Bayesian data analysis, and observe that radioactive material and b = c–k (unknown) background
“Conventional ‘frequentist’ statistics assumes that [the counts. We consider the detection problem (is c large
prior probability of each possible solution] P(Sj) is enough to indicate that n was not zero?) and the
sensitivity (how large does n need to be to give an
adequate probability of detection?).
* E-mail: [email protected]

0236–5731/USD 20.00 Akadémiai Kiadó, Budapest

© 2007 Akadémiai Kiadó, Budapest Springer, Dordrecht
K. A. MATHEWS, D. W. GERTS: BAYESIAN ANALYSIS FOR VERY-LOW-BACKGROUND COUNTING

Objective P(c | n, Bdata)

c
The objective of this work has been to develop ¦ P (b c k | Bdata) P(k | n) (3)
methods based on appropriate small-sample exact k 0
statistics and use them to adjust a measurement’s c
parameters to obtain the lowest minimum detectable ¦ P(b | Bdata)P(k c b | n).
quantity. b 0

Analysis
Probability distribution of foreground counts
The analysis uses the terminology of hypothesis
testing and conditional probabilities. For detection; the Each atom of the isotope to be detected that is
null hypothesis is that n = 0; the alternate hypothesis is initially present in the sample has some probability p of
that n>0; Dcritical is the probability of detection given being detected. Assuming that all have the same
n = 0; D is the maximum acceptable value (design value) probability, each atom constitutes a Bernoulli trial and
of Dcritical (typically D = 0.05); Ecritical is the probability the number of atoms that are detected (the foreground
of failing to detect given n>0; E is the maximum count k) is binomially distributed:
acceptable risk of this error (often E = 0.05). §n· k
Detection is declared if the number of counts equals P ( k | n, p ) ¨¨ ¸¸ p (1 p) n k . (4)
or exceeds the detection limit, cDL, which is the smallest ©k ¹
value of c for which: For an atom to be detected, it must (1) not decay
before the sample measurement starts, then (2) decay
ª c 1 º
D critical «1 ¦ P(b | Bdata )» d D , (1) during the measurement (3) by the decay branch that
¬« b 0 ¼» emits the radiation that is to be detected, and (4) that
radiation must be detected, so that:
where “Bdata” is the information about the background
available before the sample is counted. The challenge p e O't d (1 e O't c ) f branchH
(5)
lies in estimating P(b|Bdata). One could introduce a
p0 (1 (1/ 2)W c ),
dilution of a well-characterized standard solution to
provide a calibrated background, but this would not help where fbranch is the branching fraction for the decay that
matters because the variability of the sum of natural and emits the radiation that is counted, H is the overall
introduced background counts would be greater than the efficiency of the measurement system, and Wc = 'tc/T1/2
variability of the natural background counts alone, and is the non-dimensionalized sample counting duration.
greater variability in b reduces sensitivity.
Probability distribution of p – systematic errors
Minimum detectable quantity
Most systematic errors will offset p0, while errors in
For short-lived isotopes (those for which decay timing and in the half-life will effect Wc. A constant error
during the delay time and/or count duration is not in 'tc will appear as an offset in p0, whereas a rate error
negligible), the minimum detectable level is a quantity is equivalent to an error in T1/2. Thus, the consequences
rather than an activity. Given a detection level and are readily evaluated by repeatedly calculating Dcritical,
sample count duration for which the probability of Ecritical, and MDQ using the planned measurement and
falsely detecting is not greater than D, the minimum decay times using various values of p0 and T1/2 that
detectable quantity (MDQ, or, in equations, nMDQ) is the characterize possible systematic errors. Ultimately, one
smallest n such that the probability of falsely failing to might adjust the planned counting time from the
detect, Ecritical, is no greater than E, where: nominal optimal value in order to improve robustness as
E critical (c DL , n, Bdata) a result of the investigation of possible systematic errors.
Alternatively, one could add a design margin by
P (c c DL | n, Bdata) (2) reducing D and E below the desired design values in
c DL 1 order to ensure that the desired values are not exceeded
¦ P(c | n, Bdata). throughout the range of credible systematic errors. A
c 0 formal propagation of uncertainties can be achieved by
The conditional distribution of c is obtained by distributing p0 and T1/2 with appropriate probability
accounting for all the possible combinations of b and k density functions f(p0) and f(T1/2) using Eq. (7).
for which c = b+k: For example, with a Marinelli detector system, the
*
probability that an emitted particle is counted, p0( r ),

306
K. A. MATHEWS, D. W. GERTS: BAYESIAN ANALYSIS FOR VERY-LOW-BACKGROUND COUNTING

can depend strongly-enough on the location of the Analysis with Poisson-based background distribution
emission (due to attenuation in transporting into the
detector’s active volume) that the spatial distribution of Parametric background probability
*
the source to be measured, s( r ), should be mapped to
the corresponding distribution: Presuming the background process does not vary
* * 3 * with time, and that each background count is
f ( p0 ) ³V s(r ) p0 (r )d r / ³ s ( r )d 3 r . (6) independent of all others, it is appropriate to assume that
V
background counts are Poisson distributed:
In other cases, uncertainties result in p being
unknown but distributed with some probability density P b e P
P(b | P) , (8)
function f(p), which an experimenter would approximate b!
in some form appropriate to the particular application. In
any event, one would account for this distribution by where P is the expected number of background counts in
integrating over it, presumably by using numerical a count of duration 'tc. Rather than using a point
quadratures: estimator for P, a Bayesian approach infers the
probability density function, f(P), of its distribution on
1 the basis of available information on the background,
P ( k | n) ³ P(k | n, p( p0 , T1 / 2 )) f ( p0 ) f (T1 / 2 )dp. (7) presumably obtained from background measurements,
0 as follows: a background measurement of duration 'tB is
For convenience in implementing and testing the taken with a blank sample and results in B counts. If
results of this analysis, and without loss of generality, several background counts (or many previous blank
uncertainty in p is neglected so that P(n|k) | P(n|k, p) in counts) are taken, they are pooled to produce an
the remainder of this paper. equivalent single count. Denote the ratio of the counting
times as:
Probability distribution of background counts a 't c / 't B . (9)

To obtain P(b|Bdata), it is necessary to use prior The sample count duration should not be longer than
background measurements to estimate the distribution of the background count duration; so normally, 0dad1.
b. This, together with the observed count, is sufficient The distribution of B is also Poisson, but with
for the detection objective. One could use nonparametric expected counts P̃ = P/a, because both observe the same
approaches such as the empirical distribution of counts background process. Assuming the uncertainty in the
from a large number of identical blank measurements, or count durations to be negligible, the distribution of P̃
by bootstrapping from a long record of the time at which given a and B, obtained by applying Bayes’s theorem,
each background count occurred during a long and using a flat prior (to ensure objectivity and maintain
background measurement. The approach we demonstrate a frequentist’s interpretation of probability), yields the
here assumes a parameterized form for the distribution conditional distribution of P:
of background counts, specifically a Poisson ~
dP
f (P | a, B) f (P~ | a, B )
distribution. We apply Bayes’s theorem to estimate the
dP
distribution of each parameter from the background
1 P( B | P ~)
measurements. (10)
A Poisson distribution becomes narrow when very af ~ ~
long measurements are used. One could assume some ³ P ( B | P ' )dP '
other distribution that has additional parameters (shape 0
parameters) that can be adjusted to broaden the ~ B e P~
1P P BeP / a
.
distribution. We see difficulties with this approach: a a B! a B 1B!
solid theoretical foundation for selecting the alternative
distribution is hard to find and the process of estimating If B = 0, a point estimator approach, P = 0, would
the shape parameters is ad hoc. (For example, MILLER et yield f(P|a, B = 0) = G(P), hence P(b = 0|B = 0) = 1 and
al.,3 use the (conjugate) gamma distribution and estimate P(b>0|B = 0) = 0. This is the flaw mentioned in the
the shape parameters by fitting empirical data of introduction. Bayes’s theorem yields f(P|a, B = 0)=
repeated background counts.) Consequently, we = e–P/a/a and, hence, a positive probability for every b.
recommend nonparametric alternatives when the The distribution of b can be conditioned on the
Poisson’s assumptions are not justified. background measurement data as:
f
P(b | a, B) ³ P(b | P ) f (P | a, B)dP. (11)
0

307
K. A. MATHEWS, D. W. GERTS: BAYESIAN ANALYSIS FOR VERY-LOW-BACKGROUND COUNTING

Substituting Eqs (8) and (10) and performing the (More careful coding is needed to avoid underflows
integration gives: and overflows when the count is long and the
background is high. We have developed such routines in
b 1 B
§ a · § 1 · § B b· Fortran-95 and will endeavor to get them released for
P (b | a , B ) ¨ ¸ ¨ ¸ ¨¨ ¸¸. (12)
©1 a ¹ ©1 a ¹ © b ¹ public distribution in some form. Please contact
Professor MATHEWS for information.) It is useful to non-
(Note that the last factor is a binomial coefficient.) dimensionalize the counting times by expressing them in
Substituting this result and the definition of a as the ratio units of half lives:Wc { 'tc/T1/2, WB { 'tB/T1/2, so a = Wc/WB.
of counting durations into Eq. (1) yields a formula for
the critical level with explicit dependence upon the Optimizing for greatest sensitivity
sample count durations:
c Optimization is possible by changing the sample
¦ P (b | a , B ) 1 D critical (c | 't c , 't B , B ) count duration, which changes the distribution of b. The
b 0 methodology and logic for this is demonstrated in the
(13) figures and described in the captions. Longer count
c 1 b 1 B
§ 't · § 't B · § B b· times yield higher expected background counts,
¦ ¨¨ 't c't ¸
¸
¨
¨ 't 't
¸
¸ ¨¨ ¸¸.
b 0© c B ¹ © c B ¹ © b ¹ monotonically increasing Dcritical for a specified decision
level cDL, as is seen in Fig. 1. Therefore, there are
This can be evaluated efficiently by starting with the optimum values aoptimum(D, B, cDL) at which Dcritical = D
b = 0 term: exactly (the intersection points in Fig. 1). These optima,
B 1 for D = 0.05, are listed in Table 1 for various values of B
§ 1 · and decision level. To the accuracy of the Bayesian
P (0 | a, B ) ¨ ¸ , (14)
©1 a ¹ background distribution estimate, measurements with
these counting time ratios and decision levels should
and calculating each term from the proceeding one using
achieve the design false detection probability. This
the recurrence formula:
solves the problem demonstrated by STROM and
a Bb MACLELLAN.
P (b | a , B ) P (b 1 | a, B). (15)
1 a b

Fig. 1. The curves are Dcritical for the labeled decision levels as functions of the count time ratio a in the case B = 3 and D = 0.05, which is shown by
the horizontal line. The intersections determine the optimum counting time ratio for each decision level. The decision level for each curve is
labeled on the top or right edge of the plot. Longer counts increase the likelihood of false detection by increasing the expected background count

308
K. A. MATHEWS, D. W. GERTS: BAYESIAN ANALYSIS FOR VERY-LOW-BACKGROUND COUNTING

Table 1. Count duration ratio, aoptimum, at which Dcritical = D = 0.05

B
0 1 2 3 4 5
1 0.0526 0.0260 0.0172 0.01290 0.0103 0.0086
2 0.2880 0.1565 0.1082 0.0828 0.0671 0.0564
3 0.5833 0.3309 0.2334 0.1809 0.1478 0.1250
4 0.8971 0.5211 0.3724 0.2909 0.2390 0.2030
5 1.2187 0.7188 0.5181 0.4069 0.3358 0.2861
DL
6 1.5443 0.9205 0.6675 0.5266 0.4358 0.3722
7 1.8722 1.1246 0.8194 0.6485 0.5380 0.4604
8 2.2016 1.3303 0.9728 0.7719 0.6417 0.5499
9 2.5320 1.5370 1.1273 0.8964 0.7464 0.6405
10 2.863 1.7445 1.2826 1.0216 0.8518 0.7318

For the same reason, Ecritical is monotone decreasing described in its caption. The smallest of the resulting
with count duration, as is shown in Fig. 2, and with MDQ values is LMDQ and it identifies the decision
quantity, n, as is shown in Fig. 3. Consequently nMDQ is level and sample count duration that achieve this most
monotone non-increasing (n is discrete) with count sensitive measurement.
duration, as shown in Fig. 4. However, only a portion of Table 3 shows the optimum detection parameters for
this curve is relevant for selecting the sample count various values of the background count, B, and
duration. If Wc is too large, Dcritical>D, and if Wc is small background count duration (in half-lives), WB, for the
enough, a smaller DL (with smaller MDQ) is sufficient case of a perfect detector and no delay time (p0 = 1.0).
to give DcriticaldD. Figure 5 superimposes the usable The results show several trends that one should
portions of the MDQ curves for the example problem anticipate. Increases in B (for fixed WB) cause increases
defined in the caption. The dots mark the points in the optimal decision level and the MDQ, and
(a = aoptimum) where the jumps from one DL to the next decreases in count duration (both a and Wc). Increases in
occur. Note that the overall curve (ignoring the jumps) is WB with fixed observed background rate (B/WB) improve
concave up and has a minimum. The lowest minimum the characterization of the background, narrowing the
detectable quantity, (LMDQ), occurs at the bottom of distribution of b, and hence decreasing DL and MDQ by
one of the jumps. It is not necessary to construct a plot increasing count duration Wc. (Compare B = WB =1 with
such as the one shown in Fig. 5 in order to find LMDQ; B = WB =5 and B = WB =10.)
it is only necessary to create a table such as Table 2 as

Fig. 2. The curve is E critical as a function of Wc with decision level nDL = 4, for the case of p0 = 1/4, WB = 10, B = 3, and n = 20.
Longer counts decrease the probability of false failure to detect by increasing both the expected foreground and expected background counts

309
K. A. MATHEWS, D. W. GERTS: BAYESIAN ANALYSIS FOR VERY-LOW-BACKGROUND COUNTING

Fig. 3. The dots are E critical at integer values of n with decision level nDL = 4, for the case of p0 = 1/4, WB = 10, B = 3, and Wc = 2.
Larger numbers of atoms initially present increase the expected foreground counts thereby decreasing the probability of false failure to detect

Fig. 4. The plotted points are MDQ as a function of Wc (at values of Wc = 0.25, 0.30, }, 10.00) with decision level nDL = 2 for the case of p0 = 1/4,
WB = 10, B = 3, and E = 0.05, but with no constraint by D. For any given count duration, there is some number of atoms initially present that is
just sufficient to lower the probability of false failure to detect down to or below the design value. Longer counts are more sensitive
(but increase the probability of false detection)

Table 2. Intersection points in Fig. 5: Wc = aoptimumWB where WB = 10 and aoptimum for D = 0.05 is taken for each DL
from the B = 3 column of Table 1. MDQ is then calculated for each DL and its Wc for E = 0.05 and p0 = 1/4

DL 1 2 3 4 5 6 7 8 9 10
Optimum
0.1290 0.828 1.809 2.909 4.069 5.266 6.485 7.719 8.964 10.22
Wc
Optimum 29
218 39 30 31 34 37 40 44 47
MDQ LMDQ

310
K. A. MATHEWS, D. W. GERTS: BAYESIAN ANALYSIS FOR VERY-LOW-BACKGROUND COUNTING

Fig. 5. The MDQ curves here apply to the same case as in Fig. 4. The leftmost curve, marked DL = 2, is the portion of the MDQ curve in Fig. 4
for the same case as in Fig. 4, but here it is terminated (at the open circle) where Dcritical has increased to equal the design value D = 0.05. At this
point, a further increase in count duration requires increasing DL to 3 to satisfy the false detection limit, which then requires an increase in MDQ
(from 39 to 53) to reduce the probability of false failure to detect down to its design value E = 0.05. Each curve is marked with its corresponding
decision level. Each circle marks the best sensitivity that can be achieved for its decision level. The filled circle, for DL = 4, is the lowest of these
local optima, thus determining the LMDQ which is 29 for this case and the count duration required to achieve this global optimum

Table 3. Global optimum decision level, count time ratios a and Wc, and as possible, without delay while the optimal count
the resulting LMDQ for various values of B and WB, for design risks duration is calculated.
D = 0.05 and E = 0.05, and with p0 = 1.0 (no delay time and a perfect
detection system)

WB Conclusions and recommendations

1 5 10 20
DL 5 3 2 2
a 1.219 0.583 0.288 0.288
The Bayesian approach used here is always valid: for
0 large background and sample counts the results
Wc 1.129 2.92 2.88 5.76
LMDQ 12 5 3 2 converge to the conventional Gaussian-model results.
DL 7 4 3 2 This has been confirmed by implementing the Bayesian-
a 1.125 0.521 0.331 0.157 Poisson-binomial model presented here in Fortran-95
1
Wc 1.125 2.61 3.31 3.14 and applying it to large counts. However, the Gaussian
LMDQ 16 6 4 3
models fail for very low backgrounds, while our analysis
DL 9 4 3 2
a 1.127 0.372 0.233 0.108
accommodates such backgrounds and properly accounts
B 2 for short half-lives.
Wc 1.127 1.86 2.33 2.16
LMDQ 19 7 5 4 Substantial decreases in sample measurement times
DL 13 5 4 3 and increases in sensitivity can be obtained using the
a 1.009 0.286 0.203 0.125 optimization presented here. For the example problem, if
5
Wc 1.009 1.43 2.03 2.50 the sample is counted for 10 half-lives, the same
LMDQ 26 10 7 5
duration as the background measurement, the MDQ is
DL 19 9 5 4
a 0.945 0.337 0.165 0.116
43 (with DL = 9). The optimized DL is 4 and count
10 duration is 2.9 half-lives (a 71% decrease) and the
Wc 0.945 1.69 1.65 2.32
LMDQ 34 12 9 6 decrease in counting time improves sensitivity by
reducing the MDQ to LMDQ, which is 29 (a 33%
The most sensitive measurement for one isotope will decrease). In practice, a cost-benefit tradeoff can be
not be optimal for other isotopes. For mixed samples, a made; for the example, decreasing DL to 3 further
record of counts vs. time (and for example, energy bin) shortens the count duration to 1.8 half-lives while
should be used to find the count in each isotope’s bin(s) increasing MDQ only from 29 to 30. However, between
at the most sensitive count duration for that isotope. This 1.8 and 2.9 half-lives, sensitivity can be significantly
is also advantageous when the isotope is especially less that at either optimum. If the actual mean
short-lived, in which case counting should begin as soon background count rate were 2.5 counts in 10 half-lives,

311
K. A. MATHEWS, D. W. GERTS: BAYESIAN ANALYSIS FOR VERY-LOW-BACKGROUND COUNTING

different results for the background when counted, B, of the background counts) is questionable, we suggest
would result in the use of different foreground count that the use of nonparametric methods for estimating the
durations and optimal decision levels. Our Monte Carlo distribution of background counts within sample
simulation of this situation, using a design value of 5% measurements be considered. The methods
for both type I and type II errors yields an overall MDQ demonstrated here still apply: the empirical or
of 30 atoms, with an achieved risk of false detection of bootstrapping estimate replaces the Bayesian estimate of
only 3.1%, and with an achieved risk of false failure to Eq. (12).
detect of 4.5%. This is conservative in both risks (less
than the design 5%) for a slight decrease in overall Final thoughts
sensitivity. Further work is needed to fully characterize
this behavior of the method. Bayesian methods are often distrusted because of the
For short-lived isotopes, small improvements in risk of including wishful thinking in the prior
system efficiency, often achieved at substantial costs, distribution. We have used non-informative priors here
only compensate for a small additional delay time before in order to ensure against such abuses.
measurement. Depending on the half-life of the isotope, For screening testing, where quantification is not
investments in faster sample transportation and required, the test design we present here provides an
processing may be more cost effective, and application optimal choice of critical level and the corresponding
of the optimization here would be very cost effective. count duration. If the actual count reaches the critical
On the other hand, investments to decrease background level before that count duration has passed, one may
improve sensitivity for all isotopes. stop the count and declare detection. This can save
In any case, optimizing the measurement duration is considerable counting time. However, if the critical level
important. It is not the 1 atom difference between the is not reached, it would be invalid to continue to count
best count duration at DL = 4, MDQ = 29, and at DL = 3, and then declare detection if a higher critical level is
MDQ = 30 that is the important benefit of the reached before its count duration, because the second
optimization, but rather the difference between those and test would be censored by the first.
the worst times for those decision levels, yielding
MDQ = 35 and 38, respectively. Indeed, without this *
analysis, one might use a count duration that is too long
yielding an MDQ of 50 or more, or too short yielding an This work was performed at the Air Force Institute of Technology
MDQ of hundreds. For this example, the optimization as part of the official duties of the authors. The views expressed in this
article are those of the authors and do not reflect the official policy or
could be equivalent to efficiency improvements of at position of the United States Air Force, Department of Defense, or the
least 15%. US Government. This document has been officially reviewed and
We recommend that paired blank measurements not assigned the following distribution notice: Cleared for public release –
be used to estimate the background, but instead be used distribution unlimited.
for the originally-intended purpose: detecting sample
contamination. If the blank tests positive (cblanktcDL), References
new samples should be obtained.
Long background counts improve sensitivity by 1. D. J. STROM, J. A. MACLELLAN, Health Phys., 81 (2001) 27.
2. T. C. BLACK, W. J. THOMPSON, Computing Sci. Eng., 3 (2001)
reducing the uncertainty in b. Examine the LMDQ rows 86.
in Table 3; increasing the background count from one 3. G. MILLER, H. F. MARTZ, T. T. LITTLE, R. GUILMETTE, Health
half-life to twenty half-lives improves sensitivity by a Phys., 83 (2002) 512.
factor of 5 to 6 for the cases in the table. However, if 4. M. D. HANNAM, W. J. THOMPSON, Nucl. Instr. Meth. Phys. Res.,
such long background counts are needed that constancy A431 (1999) 239.
5. R. MICHEL, J. Radioanal. Nucl. Chem., 245 (2000) 137.
of the mean background count rate (and/or independence

312