Sampling How Big a Sample?

Transcription

1 C. G. G. Aitken, 1 Ph.D. Sapling How Big a Saple? REFERENCE: Aitken CGG. Sapling how big a saple? J Forensic Sci 1999;44(4): ABSTRACT: It is thought that, in a consignent of discrete units, a certain proportion of the units contain illegal aterial. A saple of the consignent is to be inspected. Various ethods for the deterination of the saple size are copared. The consignent will be considered as a rando saple fro soe super-population of units, a certain proportion of which contain drugs. For large consignents, a probability distribution, known as the beta distribution, for the proportion of the consignent which contains illegal aterial is obtained. This distribution is based on prior beliefs about the proportion. Under certain specific conditions the beta distribution gives the sae nuerical results as an approach based on the binoial distribution. The binoial distribution provides a probability for the nuber of units in a saple which contain illegal aterial, conditional on knowing the proportion of the consignent which contains illegal aterial. This is in contrast to the beta distribution which provides probabilities for the proportion of a consignent which contains illegal aterial, conditional on knowing the nuber of units in the saple which contain illegal aterial. The interpretation when the beta distribution is used is uch ore intuitively satisfactory. It is also uch ore flexible in its ability to cater for prior beliefs which ay vary given the different circustances of different cries. For sall consignents, a distribution, known as the beta-binoial distribution, for the nuber of units in the consignent which are found to contain illegal aterial, is obtained, based on prior beliefs about the nuber of units in the consignent which are thought to contain illegal aterial. As with the beta and binoial distributions for large saples, it is shown that, in certain specific conditions, the beta-binoial and hypergeoetric distributions give the sae nuerical results. However, the beta-binoial distribution, as with the beta distribution, has a ore intuitively satisfactory interpretation and greater flexibility. The beta and the beta-binoial distributions provide ethods for the deterination of the iniu saple size to be taken fro a consignent in order to satisfy a certain criterion. The criterion requires the specification of a proportion and a probability. KEYWORDS: forensic science, drugs, statistics, sapling, Bayesian inference 1 Departent of Matheatics and Statistics, The King s Buildings. The University of Edinburgh, Mayfield Road, Edinburgh EH9 3JZ, United Kingdo. Received 29 May 1998; and in revised for 16 Oct. 1998; accepted 19 Oct Copyright 1999 by ASTM International 750 Introduction Consider a population or consignent, which consists of discrete units, such as individual tablets in a consignent of tablets or individual coputer disks in a consignent of disks. Each unit ay, or ay not, contain soething illegal, such as drugs or pornographic iages. It is of interest to an investigating scientist to deterine the proportion of the consignent which contains soething illegal. This ay be done exactly (assuing no istakes are ade) by exaination of every unit in the consignent. Such an exaination can be extreely costly. Considerable resources can be saved if inforation, sufficient to satisfy the needs of the investigators, ay be gained fro exaination of a saple fro the consignent. When a saple is considered, uncertainty is introduced when inference is ade fro the saple to the population, because the whole population is not inspected. However, this uncertainty ay be quantified probabilistically. It is shown that if two nubers are specified in advance of the inspection of the consignent, then a saple size ay be specified. The first of these nubers is the iniu proportion of units in the consignent which contain soething illegal that the exaination is to be designed to find. The second is the required probability with which the true proportion of illegal units exceeds this iniu proportion. With reasonable assuptions, a probability distribution for the true proportion of units in the consignent is derived, based on the scientist s prior beliefs (i.e., prior to the inspection of individual units) and the outcoe of the inspection of the saple. The prior beliefs ay be based on presuptive tests. For exaple, the physical appearance of the suspected illegal aterial ay be siilar to that of other known illegal aterial fro the scientist s experience. Alternatively, the results of an initial exaination of soe of the aterial by color spot tests ay affect the prior beliefs. It is possible to choose a function to represent the strength of the scientist s prior beliefs. It ay be thought inappropriate that the scientist s prior beliefs should have any effect on the decision to be ade regarding the saple size. In such a case, it is possible to choose the function in such a way that the effect is very sall. (It is also possible to choose the function such that the effect is very large.) It is not possible for the scientist s prior beliefs to have no effect on the analysis. For exaple, the choice of the odel which is used to represent the uncertainty introduced by the sapling process is a subjective choice. The binoial odel described here requires assuptions about independence of the probability for each unit being illegal and the choice of a constant value for this probability. The function representing the scientist s prior beliefs is then cobined with a function which accounts for the observation of the nuber of units in the saple which contain soething illegal. The cobination of these two functions provides a third function which represents the probability distribution for the true proportion of illegal units in the consignent. It is shown here that this function is of the sae for as the one representing the scientist s prior beliefs. Fro this so-called posterior distribution (i.e., posterior to the inspection of individual units) it is possible to deterine the probability with which the true proportion exceeds any specified proportion. This approach provides a probability stateent about the true proportion. This is in contrast to the inference obtainable fro an approach which provides a confidence interval for the true proportion. It is no accident that the word probability is not used to describe this

2 AITKEN SAMPLING 751 interval. A confidence interval derives its validity as a ethod of inference on a long-run frequency interpretation of probability (1). For exaple, consider specifically, the 95% confidence interval for a proportion. The probability with which this interval contains the true proportion is not known. However, suppose the experient which generated the 95% confidence interval is repeated any ties (under identical conditions, a theoretical stipulation, which it is ipossible to fullfil in practice) and on each of these occasions a 95% confidence interval for the true proportion is calculated. Then, it can be said that 95% of these (95%) confidence intervals will contain the true proportion. This does not provide inforation concerning the one 95% confidence interval which has been calculated. It is not known whether it does or does not contain the true proportion, and it is not even possible to deterine the probability with which it contains the true proportion. The ethod which uses the scientist s prior beliefs and enables a probability stateent to be ade is known as the Bayesian ethod, after the Rev. Thoas Bayes (2 4). The procedure by which prior beliefs and a function of the observations ay be cobined to provide posterior beliefs is known as Bayes Theore. The ethod which relies on the idealized long-run frequency for its validity is known as the frequentist ethod. It is the purpose of this paper to copare the results obtained fro the Bayesian and frequentist approaches to assessing uncertainty, to clarify the assuptions ade in the two approaches, to contrast the clarity of the inferences obtainable fro the Bayesian approach with the lack of clarity associated with the frequentist approach, and to illustrate the greater flexibility of the Bayesian approach with the inflexibility of the frequentist approach. The ethods are illustrated with reference to sapling fro consignents of drugs. However, they apply equally well to sapling in other forensic contexts, for exaple, glass (5) and pornographic iages. Frequentist procedures are described in (6) for choosing a saple size fro a consignent. Distinction is drawn between an approach based on the binoial distribution and an approach based on the hypergeoetric distribution. It is argued in (6) that the forer can be used for large consignents in which the sapling of units ay be taken to be sapling with replaceent. For sall saples, the sapling units ay be taken to be sapling without replaceent and the hypergeoetric approach is used. The Bayesian approach also has different ethods for analyzing large and sall saples. As reported in (7), various ethods for selecting the size of a rando saple fro a consignent have been accepted by the U.S. courts. An approach based on the hypergeoetric distribution is proposed in (7). A suary of different procedures used in 27 laboratories around the world in given in (8). These procedures include ethods based on the square root of the consignent size, a percentage of the consignent size, and a fixed nuber of units regardless of the consignent size, as well as the hypergeoetric distribution. The authors in (8) propose the forula %(N 20) (for N 20) where is the saple size and N is the consignent size. As well as being siple to ipleent, this approach, as the authors rightly clai, provides the opportunity to discover heterogeneous populations before the analysis is copleted. According to (7), it should be sufficient to deonstrate with good probability that ost of the exhibit contains the controlled substance. Yet, suaries are given as confidence liits using a frequentist approach (as described above) and not in probabilistic ters. For exaple, fro (7), a stateent of the for that at the 95% confidence level, 90% or ore of the packages in an exhibit contain the substance is suggested as being sufficient proof in cases of drug handling. The procedures to be described here provide suaries in probabilistic ters. In general, an answer is provided to the question: How big a saple should be taken for it to be said that there is a 100 p% probability that the proportion of units in the consignent which contain drugs is greater than %? For an exaple of a particular instance, an answer is provided to the question: How big a saple should be taken for it to be said that there is a 95% probability that the proportion of units in the consignent which contain drugs is greater than 50%? Here, the value 0.95 has been substituted for p and the value 0.5 has been substituted for 0. This requireent ay not see very stringent but ay be sufficient to satisfy certain legal requireents. A Coparison of Two Methods of Measuring Uncertainty Before coparing two ethods of estiating the saple size necessary to ake a stateent about the proportion of units in a consignent which contain drugs, it is useful to consider further the two ethods of easuring uncertainty on which these two ethods are based, and which were outlined in the Introduction. The Frequentist Method This assues that the proportion of the consignent which contains drugs is unknown but fixed. The data, that is the nuber of units in the saple which contain drugs, are variable. A so-called confidence interval or level is calculated. The nae confidence is used since no probability can be attached to the uncertain event that the interval contains. The frequentist approach derives its nae fro the relative frequency definition of probability. The probability that a particular event, A say, occurs is defined as the relative frequency of the nuber of occurrences of event A copared with the total nuber of occurrences of all possible events, over a long run of observations, conducted under identical conditions. For exaple, consider tossing a coin n ties. It is not known if the coin is fair and the outcoes of the n tosses are to be used to deterine the probability of a head occurring on an individual toss. There are two possible outcoes, heads (H ) and tails (T ). Let n(h ) be the nuber of H and n(t ) be the nuber of T such that n(h ) n(t ) n. Then the probability of tossing a head on an individual toss of the coin is defined as the liit as n of the fraction n(h )/n. The validity of the frequentist approach, however, relies on a belief in the long-run repetition of trials under identical conditions. This is an idealized situation, seldo, if ever, realized in practice. The Bayesian Method: Subjective Probability The Bayesian approach represents the uncertainty concerning knowledge of (the proportion of interest) with a probability distribution. The data are taken as fixed, in contrast to the frequentist ethod. A saple of a particular size fro a consignent has been taken and the nuber of units z which contain drugs noted. These data are considered fixed. There is no consideration for the long-run repetition of trials under identical conditions. Data which ay have been observed but have not are not allowed to affect the analysis. In the frequentist ethod the probability of z given and is a function represented by the binoial distribution. In the Bayesian ethod, the sae func-

3 752 JOURNAL OF FORENSIC SCIENCES tion is expressed as a function of, known as the likelihood function. It has the sae atheatical for as the binoial distribution but takes the data as fixed and expresses the for as a function of : L (, z) ( z ) z (1 ) z (0 1). (1) Uncertainty about can be expressed as a probability distribution. Probability intervals ay be deterined for with a uch clearer interpretation than with confidence intervals. Choice of Saple Size Large Consignents A large consignent will be taken to be one which is sufficiently large that sapling is effectively with replaceent. This can be as sall as 50, though in any cases it will be of the order of any thousands. A consignent of drugs containing N units will be considered as a rando saple fro soe super-population of units which contain drugs. Let (0 1) be the proportion of units in the super-population which contain drugs. For consignent sizes of the order of several thousand all realistic values of will represent an exact nuber of units. For sall saple sizes less than 50, can be considered as a nuisance paraeter (5 and Appendix 1) and integrated out of the calculation leaving a probability distribution for the unknown nuber of units in the consignent which contain drugs as a function of known values. For interediate calculations, can be treated as a continuous value in the interval (0 1), without any detrient to the inference. As before, let be the nuber of units sapled and let z be the nuber which are found to contain drugs. Frequentist Approaches to the Estiation of The saple proportion p z / is an unbiased estiate of. The variance of p is given by (9) (1 ) N N 1. The factor (N )/(N 1) is known as the finite population correction (fpc). Provided that the sapling fraction /N is low, the size of the population has no direct effect on the precision of the estiate of. For exaple, if is the sae in the two populations, a saple of 500 fro a population of 200,000 gives alost as precise an estiate of the population proportion as a saple of 500 fro a population of 10,000. The estiated standard deviation of in the second case is 0.98 ties the estiated standard deviation in the first case. Consider the following exaple. To siplify atters, the fpc is ignored and the saple proportion p is assued to be norally distributed. Assue that is thought to be about 75%. It is stipulated that a saple size is to be taken to estiate to within 25%, i.e., in the interval (0.50, 1.00) with 95% confidence. (This ay be thought a very wide interval but it is consistent with the for of question posed at the end of the Introduction that a saple size be deterined such that it can be said that, if all of the saple are found to contain drugs, then there is a 95% probability that is greater than 50%.) The criterion for the saple size is that there should be a confidence of 0.95 that the saple proportion p lies in the interval Fro known results of the noral distribution, this iplies that two standard deviations equal The standard deviation of p, ignoring the fpc, is (1 ) (see also equation (5) of (5)). Setting two standard deviations equal to 0.25, and solving for, gives the following expression for : 4 (1 ) When 0.75, 12. Thus a saple of size 12 is sufficient to estiate to be greater than 0.5 with confidence Siilar calculations are reported in (6), where one standard deviation is set equal to 0.1 and the saple size is chosen to aintain this value of the standard deviation for various values of N and estiated values of. Later, it will be shown using Bayesian techniques, that, if all the inspected saples are found to contain drugs, the required saple size to enable one to say that 0.5 with probability 0.95, is 4. An alternative approach, based on the binoial distribution, is discussed in (5). Consider a specific value 0 of. The forensic scientist wishes to show that 0. Another probability,, is selected as the probability that the null hypothesis 0 is rejected based on an inspection of a saple of units fro the consignent. It is also assued that the inspection of the saple reveals that all the units in the saple contain drugs. An exaple when not all the units which are exained contain drugs is given in (6). This analysis ay be considered as a hypothesis test. The null hypothesis is that 0 ; the alternative is that 0. As stated above, it is assued that z. Standard statistical theory shows that the null hypothesis is rejected in favor of the alternative if 0. The probability is the probability that the hypothesis 0 is rejected. The copleentary probability (1 ) is the probability that the hypothesis 0 is not rejected. Failure to reject the null hypothesis is not the sae as saying it is true. There will be any occasions in which data are such that the null hypothesis is not rejected but for which the data provide stronger support for an alternative hypothesis. The probability that the hypothesis 0 is not rejected is not the sae as the probability that 0. For this reason the word confidence is used instead of probability when referring to the first situation. It is said that one has 100(1 )% confidence that 0. Table 1, a subset of Table 1 of (6) shows the values of (1 ) for various values of 0 and. The eaning of the results in Table 1 can be illustrated by consideration of Assue Five units of the consignent are exained and all are found to contain drugs. The probability of this happening, when 0 0.7, is which equals Thus, there is 83% confidence that, the true proportion of drugs in the consignent, is greater than 0.7. Siilarly, there is 67% confidence that is greater than 0.8 and 41% confidence that is greater than 0.9. However, these stateents are not probability stateents about the value of. The stateent concerning the first result, written ore fully, is as follows. Suppose 0.7. The probability that, when 5 units are exained, all are found to contain drugs is less than TABLE 1 Values of (1 ) for various values of 0 and, fro (6)

4 AITKEN SAMPLING 753 What is taken to be known in the probability calculation is that 0.7. The probability stateent is concerned with the probability that all the exained units contain drugs, if 0.7. The data are variable, is fixed. This is the transpose of what is required. In practice, it is known that a saple has been exained and all of the units in the saple have been found to contain drugs. It is then the probability that 0.7, say, which is of interest. In the first case, illustrated in Table 1, probability stateents of the for Pr(z units exained, 0.7) are considered. In the transpose, probability stateents of the for Pr( 0.7 units exained and z ) are considered. In both cases, z is the nuber of units exained which are found to contain drugs. It is one of the ain purposes of this paper to deonstrate that this second approach provides a ore intuitively satisfactory ethod of deterining the saple size. A Bayesian Method for the Estiation of In order to ake probability stateents about, it is necessary to have a probability distribution for, to represent the variability in. This variability ay siply be uncertainty in one s knowledge of the exact value of, uncertainty which ay arise because the consignent is considered as a rando saple fro a super-population. However, the Bayesian philosophy perits one to represent this uncertainty as a probability distribution. The ost coon distribution for is the so-called beta distribution (10). Its use in another forensic context, that of sapling glass fragents, is described in (5). A continuous rando variable has a beta distribution with paraeters (,, 0, 0), denoted Beta(, ), if its probability density function ƒ(, ) is ƒ(, ) 1 (1 ) 1 /B(, ), 0 1, (2) where and ( ) ( ) B(, ) ( ) (z) 0 t z 1 e t dt is the gaa function. Integer and half-integer values of the gaa function are found fro the recursive relation (x 1) x (x) and the values (1) 1 and ( 1 / 2 ) Note that the beta distribution odels characteristics which only takes values in the range (0,1), which is particularly appropriate for proportions. Graphs of the beta distributions with paraeters (3,2), (3,1) and (10,1) are shown in Fig. 1 (a), (b), and (c). The graph of Beta (3,1) is proportional to 2. This reflects belief that the ost likely outcoe is that all units in the consignent contain drugs with a belief that reduces as a quadratic as decreases fro 1 to 0. The graph of Beta (10,1) is proportional to 9. The graph of Beta (3,2) has a ode at 2 / 3, reflecting a belief that it is quite likely that there are units in the consignent which do not contain drugs. The beta distribution is technically convenient in the context of sapling fro a discrete consignent because it is a so-called conjugate prior distribution for the binoial distribution. It cobines with the binoial distribution to provide a posterior distribution which is also a beta distribution. See Appendix 1 for further details. Thus, if units are exained and z are found to contain drugs then the probability density function which cobines this inforation with the prior distribution is given by ƒ( z,,, ) z 1 (1 ) z 1 /B(z, z ), 0 1, denoted Be(z, z ). In the particular case where z, the density function is given by ƒ(,,, ) 1 (1 ) 1 /B(, ), 0 1. A Copyright FIG. by 1 Probability ASTM Int'l (all rights density reserved); functions Thu Oct fro 10 (2) 07:09:46 for the EDT beta 2013 distribution with paraeters (a) 3, 2, (b) 3, 1, (c) 10, 1.

5 754 JOURNAL OF FORENSIC SCIENCES B C FIG. 1 (Continued) There is an interesting coparison with the frequentist forulation of the proble, when a liiting case is considered. Let 0 and let 1. This liiting case gives rise to an iproper prior distribution, so-called because it is not a probability density function. However, the analysis is justified in that the posterior distribution is a probability density function, Thus, ƒ(,,, ) 1 /B(, 1) 1. Pr( 0,,, ) 0 1 d 0, the sae expression as used in the frequentist case but with a uch 0 ore coherent interpretation, since it is now a probability stateent about. For exaple, if 5, Pr( 0.7 5, 5,, ) , the sae nuerical value as before. Another coonly used prior distribution is one in which 1. Then ƒ( 1, 1) This distribution is the so-called unifor distribution and is often used to represent axiu uncertainty about. Another representation of uncertainty is the case where 1 / 2 where greater belief is placed at the ends of the range, in favor of all ites or no

6 AITKEN SAMPLING 755 ites containing drugs, than in the iddle. Then ƒ( 0.5, 0.5) 1 1/2 (1 ) 1/2. In practice, a criterion has to be specified in order that the saple size ay be deterined. Consider the criterion fro the Introduction that the scientist wishes to be 95% certain that 50% or ore of the consignent contains drugs when all units sapled contain drugs. Then the criterion ay be written atheatically as Pr( 0.5,,, ) 0.95 or 1 1 (1 ) 1 d /B(, ) (3) 0.5 The general question in which p and 0 are specified at the end of the Introduction ay be answered by finding the value of which solves the equation 1 1 (1 ) 1 d /B(, ) p. (4) 0 Such integrals are easy to evaluate using standard statistical packages (e.g., SPLUS (11)) given values for, and. It is then a siple atter to substitute specified values for 0 and p and given values for and and then select by trial and error to solve (4). TABLE 2 The probability that the proportion of drugs in a large consignent is greater than 50% for various saple sizes and prior paraeters and For exaple: Consider the following pairs of values for and : (1,1), (0.5,0.5), and (0.065,0.935) (the last pair suggested by Professor T. Leonard, personal counication). For the first two pairs there is a prior probability of 0.5 that 0.5; for the third pair there is a prior probability of 0.05 that 0.5. This third choice was ade since 0.05 is the copleent of Table 2 shows the results for (3) for various values of. This gives the rearkable result that, for large consignents, of whatever size, the scientist need only exaine 4 units, in the first instance. If all are found to contain drugs, there is a 95% probability that 50% of the consignent contains drugs. Copare this with the result derived fro a frequentist approach using a noral approxiation to the binoial distribution which gave a value of 12 for the saple size. These saple sizes are not large. However, there is not very uch inforation gained about the exact value of. It is only deterined that there is probability of 0.95 that 0.5. This is a wide interval (fro 0.5 to 1) within which the true proportion ay lie. Figures 2 and 3 illustrate how varying prior beliefs have little influence on the conclusions once soe data have been observed. Figure 2 shows the prior probability that 0 for 0 0 1, for the values of (, ) given in Table 2, decreasing fro a value of 1 when 0 0 to a value of 0 when 0 1. There are considerable differences in the curves. Figure 3 shows the corresponding posterior probabilities for 0 given four units have been observed and all have been found to contain drugs, with the values for ephasized. There is very little difference in these curves. There ay be concerns that it is very difficult for a scientist to foralize his prior beliefs. However, if and are sall, large differences in the probabilities associated with the prior beliefs will not lead to large differences in the conclusions. The ethodology can be extended to allow for units which do not contain drugs. For exaple, if one of the original four units inspected is found not to contain drugs then three ore should be inspected. If they all contain drugs, then it can be shown that the probability that 0.5, given that six out of seven contain drugs, is FIG. 2 The prior 1-F ( 0 ) probability that the proportion of units in a consignent which contain drugs is greater than 0, for various choices of and : 1( ), 0.5( ), 0.065, 0.935( ).

7 756 JOURNAL OF FORENSIC SCIENCES FIG. 3 The posterior probability 1-F ( 0 ) that the proportion of units in a consignent which contain drugs is greater than 0, after inspection of 4 units has shown the all to contain drugs, for various choices of and : 1( ), 0.5( ), 0.065, 0.935( ). The solid lines show the probabilities that 0.5 for the various choices of and (fro Table 2, with 4, the probabilities are 0.97, and 0.95, respectively, for (, ) (1, 1), (0.5, 0.5) and (0.065, 0.935)). TABLE 3 The saple size required to be 100p% certain that the proportion of units in the consignent which contain drugs are greater than 0, when all the units inspected are found to contain drugs. The prior paraeters 1. p TABLE 4 The probability p that the proportion of a large consignent which contains drugs is greater than 0 when a saple of size is inspected and the entire saple is found to contain drugs ( 1.) The dependency of the saple size on the values of p and 0 is illustrated in Table 3. The prior paraeters and are set equal to 1. Consider p 0.90, 0.95 and 0.99 and consider values of 0.5, 0.6, 0.7, 0.8, 0.9, 0.95, The saple size required to be 100p% certain that is greater than the specified value is then given by the value of which satisfies the equation Pr( 0,, 1, 1) p, a special case of (4). The value of is thus given by the sallest integer greater than [log(1 p)/log( 0 )] 1. Obviously, when considering the results in Table 3, the consignent size has to be taken into account in order that the saple size ay be thought sall with respect to the size of the consignent. Thus, for the last row in particular to be useful, the size of the consignent fro which the saple is to be taken will have to be of the order of several tens of thousands. An alternative representation, given in Table 4, considers the value of p which is obtained for various values of 0 and for given saple sizes, when the entire saple of size is found to contain drugs. The first row of Table 2 is a special case of Table 4. There ay be situations in which different choices of and ay be wanted. The outcoe for sall saples of soe different choices is shown in Table 8. It ay be the scientist has soe substantial prior beliefs about the proportion of the consignent which ay contain drugs. These beliefs ay arise fro previous experiences of siilar consignents (fro what ay be considered to be the sae super-population), for exaple. In such cases, use can be ade of various properties of the beta distribution to assist the scientist in choosing values for and. The ean of the beta distribution is /( ) and the variance is /( ) 2 ( 1). Thus, a prior belief of the proportion of the consignent which contains drugs would set that proportion equal to the ean of the distribution and a belief about how precise that belief was, would provide a value for the variance. Alternatively, if it was felt that

8 AITKEN SAMPLING 757 could be set equal to 1, so that the shape of the probability density function is siilar to those in Fig. 1(b) and 1(c) (i.e., onotonic increasing with respect to ), and that there was a prior belief about a lower bound for the proportion, say that Pr(Proportion 0, ) p then use could be ade of the result that log(1 p)/log( 0 ). Sall Consignents Suppose now that the consignent size N is sall. A saple of units fro the consignent is exained and z( ) units are found to contain drugs. Denote the nuber, N, of units not exained by n, so that n N. A Frequentist Approach Consider a frequentist approach based on the hypergeoetric distribution (6). Let R Z Y be the total nuber of units in the consignent which contain drugs, where Z is the nuber of units in the saple of size and Y is the nuber of units in the reainder which contain drugs. Then the distribution of Z is hypergeoetric with Pr(Z z) ( z R )( N R z ) N, ( ) z 0, 1,..., in(r, ). To satisfy a given confidence level (1 ), the axiu value of R is needed, (6), such that ax ( z, N R) x in (o, R) When z this reduces to ( x R )( N R x ) N. ( ) R!( N )!. N! ( R )! Table 5, part of Table 3 in (6), shows the confidence levels when z and 0.7 for N 10, 20 and 30 (and hence R 7, 14, 21) and for 5, 10, 15. Consider the value 0.92 when N 10 and 5. This is the probability that, if 0.7, 5 units in a consignent of size 10, when exained, will be found to all contain drugs. Again this assues to be known to be 0.7 and gives a value for the probability that z. This is translated in frequentist ters to read that one is 92% confident that, when 5 units in a population of 10 are exained and all are found to contain drugs, 0.7. The Beta-Binoial Distribution The interpretation in the previous section is not so clear as that obtained fro the Bayesian ethod which uses a so-called beta-binoial distribution (10). The beta-binoial distribution provides a probability stateent about the nuber of units in the consignent which contain drugs. TABLE 5 Confidence levels when the true proportion of drugs in the consignent is 0.7, for consignents of size N and saples of size. N As before, let (0 1) be the proportion of units in the super-population which contain drugs. The probability distribution of z, given and, ay be taken to be binoial. For each unit, independently of the others in the consignent, the probability it contains drugs is taken to be equal to. The posterior distribution of is another beta distribution with paraeters ( z) and ( z) (see Appendix 1). Since the consignent size is sall, a better representation of the variability of the nuber of unexained units in the consignent which contain drugs is obtained by considering a probability distribution for this nuber, Y, explicitly. There are n units in the reainder of the consignent ( n N ) which have not been exained. Then Y (unknown and n) is the nuber of units in this reainder which contain drugs. Given, the distribution of (Y n, ), like that of (Z, ), is binoial. However, has a beta distribution and the distribution of (Y n, ) and the distribution of (, z,, ) can be cobined to give what is known as a Bayesian predictive distribution for (Y, n, z,, ), known as a beta-binoial distribution (10). Pr(Y y, n, z,, ) ( )( n y ) (y z ) ( n z y ), (z ) ( z ) ( n ) (y 0, 1,..., n), (5) (see Appendix 1). Fro this distribution, inferences can be ade about Y, such as probability intervals or lower bounds for Y. Note the flexibility given by the ability to vary and to incorporate prior beliefs about the proportion of units in the consignent which contain drugs. Also, for integer values of and, (5) reduces to a function of factorials. Coparison of Beta-Binoial and Hypergeoetric Approaches For large saples, it has been shown that in a liiting case, the beta distribution and the binoial distribution give the sae nuerical answers, though with different interpretations. A corresponding coparison can be ade with the beta-binoial and hypergeoetric distributions. Suppose units are exained and z are found to contain drugs; n units are not exained. Let y be the nuber of unexained units which contain drugs and let r z y. Then y is unknown. It can be shown (see Appendix 2) that n ( 1)( z )( y) Pr(R r, n, z) ( n 1) ( n (6) z y ) Pr(Y y, n, z, 1, 1), the beta-binoial (5) distribution with 1, where r can take values z,..., n z, and y can take values 0,..., n. This result depends on a theore known as Vanderonde s theore (12,13). This is not a new result. Todhunter (14) credits Condorcet in 1785 and Prevost and Lhuilier in 1799 with recognizing it. Todhunter (14) coents that the coincidence of the results obtained on the two different hypotheses is rearkable. This result has also been used ore recently in forensic science in the context of glass sapling (5) Use of the Beta-Binoial Distribution As an exaple, consider a consignent of size N 10, where five units are inspected and all five are found to contain drugs ( z 5). For the pro-

9 758 JOURNAL OF FORENSIC SCIENCES portion of units in the consignent which contain drugs to be at least 0.7 ( 0.7), it is necessary for the nuber of units Y in the five units not inspected to be at least 2(Y 2). The beta-binoial probability, with a unifor prior 1, is given by Pr(Y 2 5, 5, 5, 1, 1,) 5 y ( 5) ( y) , 1( ) 5 y the suation of ters in (6), with n z 5 and for y taking values fro 2 to 5. The hypergeoetric distribution has the interpretation that if z 5, one is 92% confident that 0.7. The beta-binoial approach enables one to assign a probability of to the event that 0.7. As with large consignents, values for and ay be chosen subjectively to represent the scientist s prior beliefs before inspection about the proportion of the units in the consignent (as a rando saple fro the super-population) which contain drugs. When considering large consignents, the criterion was stipulated that a saple size was to be chosen such that if all the units contained drugs then there was to be a probability of 0.95 that the proportion of units in the whole consignent which contains drugs was greater than 0.5. This criterion was used to investigate saple sizes required for sall consignents in which N 50. The results are shown in Table 6, where 1. Thus, for consignents no greater than 50 in size, saples of size 3 need to be exained. For saples greater than 50, theory relating to large consignents ay be used and saples of size 4 ay be used. For convenience, saples of size 4 can be used for all saple sizes. Results can also be obtained for proportions of N different fro 50%. For exaple, in Table 7, R 0.8N is chosen. The probability that R 0.8N is given for saple sizes 1, 2, 3 and 4 in which all units contain drugs ( 1). The results in Tables 6 and 7 are for the special circustances in which 1 and all units inspected contain drugs. More general results can be obtained. The proble is then to choose such that, given n,, and, (and possible values for z, consequential on the choice of and the outcoe of the inspection), a value for y can be deterined to satisfy soe probabilistic criterion, e.g., the value y 0 such that Pr(Y y 0, n, z,, ) p. Soe results are given in Table 8 for p 0.9, where the consignent size N is taken to be 30. Note fro the last two rows that if one or two of the six inspected units do not contain drugs then the nuber of units in the reainder of the consignent which can be said, with probability 0.9, to contain drugs drops fro 17 to 12 to 9. Note also that even if 16 units (out of 30) are inspected and all are found to contain drugs, then it can only be said, with probability 0.9, that 12 of the reaining 14 contain drugs (and this is so even with 4, 1). Suary The following suary of the ain results of the paper is phrased in the context of inspecting a consignent of drugs. However, the ideas expressed in the paper are just as applicable to other forensic contexts, such as the inspection of coputer disks for pornographic iages. In such a situation sapling ay be beneficial as it exposes the investigators to as little stress as possible. For sall consignents, the beta-binoial distribution provides a probability distribution for Y and hence for the total nuber of units R (and hence the proportion of units) in the consignent which contain drugs. For large consignents, the beta distribution is used. These probability distributions can then be used to ake inferences for the nuber of units in a consignent which contain drugs. There are no probles of interpretation as the uncertainty TABLE 6 For consignents of size N in which R( N and unknown) units contain drugs, the probability that R N/2 is given for various saple sizes, in which all units in the saple contain drugs. ( 1.) N TABLE 7 For consignents of size N in which R ( N and unknown) units contain drugs, the probability that R 0.8N is given for various saple sizes in which all units contain drugs. ( 1.) N 0.8N TABLE 8 Deterination of the saple size required fro a consignent of 30 units, to satisfy certain criteria. Paraeters and are representative of prior beliefs about the proportion of units which contain drugs. The nuber of units inspected equals, the nuber of those which contain drugs is z. The nuber of units not inspected is n (equals 30 ). y 0 is the largest nuber of those units not inspected for which it can be said that the probability is 0.9 or greater that y 0 or ore units contain drugs. Pr( Y n) is the probability that all the units not inspected contain drugs. z n y 0 Pr(Y n)

10 AITKEN SAMPLING 759 concerning the nuber of units in the consignent which contain drugs ay be expressed by a probability distribution. The cost, however, lies in the choice of the paraeters and. This choice is ade subjectively. If a sall change in their values leads to a large change in the outcoes then considerable care has to be exercised in that choice so that it ay be fully justified. However, if this were the case, this would be indicative that little inforation had been gained fro inspection of the consignent. Care would be needed in the interpretation, regardless of the statistical input to the investigation. For large consignents, once choices of,, 0 and p have been ade, it is a siple atter to deterine the value of which provides a solution for (4). For sall consignents the inferences which can be ade are illustrated in Tables 6 and 7 (for 1) and in Table 8 for other values of and. The general forula is given by (5). Finally, consideration has to be given to the effect of the discovery that soe units in the inspected saple do not contain drugs. Such a discovery can have quite an effect on Pr(Y n) and on y 0 as illustrated in Table 8. Further units can be inspected in a sequential process. An exaple for large consignents has been given. Appendix 1 Derivation of the Beta-Binoial Distribution A saple of size is taken fro a consignent which contains n units ( n N). Let be the proportion of the total nuber of units in the consignent which contain drugs and let Z be the nuber of units in the saple of size which contain drugs. Then Pr(Z z, ) z (1 ) z, z 0, 1,...,, z and! z z!( z)! is the binoial coefficient, where the! notation denotes factorials (e.g., for integer x, x! x(x 1)(x 2)...1) and the distribution for Z is a binoial distribution. The binoial distribution, written as a function of, is the likelihood function (1) which can then be cobined with a prior beta distribution (2) for to give a posterior beta distribution for. This follows fro Bayes Theore which states that ƒ( z,,, ) Pr(Z z, ) ƒ(, ). The posterior beta distribution for ( z,,, ) can be shown to be ƒ( z,,, ) z 1 (1 ) z 1 /B(z, z ), 0 1. However, for sall consignents, the ost interesting quantity is the nuber Y of units in the reaining n unexained units which contain drugs. The distribution of (Y n, ) is binoial. When this distribution is cobined with the posterior beta distribution for the resulting distribution is known as a beta-binoial distribution (10). Pr(Y y, n, z,, ) ( )( n y ) (y z ) ( n z y ), (z ) ( z ) ( n ) Appendix 2 Siilarity of Results fro a Beta-Binoial Distribution with 1 and a Hypergeoetric Distribution. First, define a to be equal to. k k! This in turn is equal to ( 1) a k 1 (a k 1)! ( 1) k a k 1 ( 1) a 1 (a 1)!k! k, (12). Then, by coparing the coefficients of t k in the two sides of the equation it can be shown that k (1 t) a (1 t) b (1 t) a b a k j 1 j k, j 0 k j (13). With a suitable change of notation this result can be written as n z k z z n. k z Consider the beta-binoial distribution (5) with 1. Then it can be shown that ( 1) ( z ) ( n y ), Pr(Y y, n, z, 1, 1) ( n 1) ( n (7) z y ) for y 0,1,2,..., n. Let R be the total nuber of units which contain drugs in a consignent of size N. Thus R takes a value in {0,1,2,..., N}. A unifor prior distribution for R assigns equal probability 1/(N 1) to each of these (N 1) integers. Pr(R r N) 1/(N 1). The distribution of Z, the nuber of inspected units which contain drugs, given, n and R, is hypergeoetric. For ease of notation, let N n and R Z Y. The distribution of Y, given, n and z and given the unifor prior for R can be written as Pr(Y y, n, z) Pr(Y z r, n, z) n k Pr(Z z, n, r)pr(r r N ) Pr(Z z, n) Pr(Z z, n, r)pr(r r N ) n z Pr(Z z, n, k)pr(r k N ) k z ( z z y ) ( n z z y ) n z k (z ) ( n k z ) k z ( a)( a 1)...( a k 1) b j 1 ( z z y )( n z z y ) ( n 1 n ) n 1 ( 1) ( z ) ( n y ) ( n 1) ( n. z y ) a b k 1 (y 0,1,..., n), expression (5). for y 0,1,2,..., n, which equals the beta-binoial probability (7).

11 760 JOURNAL OF FORENSIC SCIENCES Acknowledgent I wish to thank Dr. Freda Kep for helpful coents concerning the beta-binoial and hypergeoetric distributions. References 1. Barnett V. Coparative statistical inference. 2nd edition. Chichester, John Wiley and Sons Ltd., 1982; Bayes T. An essay towards solving a proble in the doctrine of chances. Philosophical Transactions of the Royal Society of London for 1763, 1764;53: Reprinted with Barnard (1958) in Pearson and Kendall 1970; Barnard GA, Bayes T. A biographical note (together with a reprinting of Bayes, 1764), Bioetrika 1958;45: Reprinted in Pearson and Kendall 1970; Pearson EG, Kendall MG, editors. Studies in the history of statistics and probability. London: Charles Griffin, Curran JM, Triggs CM, Buckleton J. Sapling in forensic coparison probles. Science and Justice 1998;38: Tzidony D, Ravreby M. A statistical approach to drug sapling: a case study. J Forensic Sci 1992;37: Frank RS, Hinkley SW, Hoffan CG. Representative sapling of drug seizures in ultiple containers. J Forensic Sci 1991;36: Colón M, Rodriguez G, Diaz RO. Representative sapling of street drug exhibits. J Forensic Sci 1993;38: Cochran WG. Sapling techniques, 3rd edition. Chichester: John Wiley and Sons Ltd., Bernardo JM, Sith AFM. Bayesian Theory. Chichester, John Wiley and Sons Ltd., 1994; Venables WM, Ripley BD. Modern applied statistics with S-Plus. 2nd edition. New York, Springer Verlag, Johnson NL, Kotz S, Kep AW. Univariate discrete distributions. 2nd edition. Chichester, John Wiley and Sons Ltd., 1992; Feller WF. An introduction to probability theory and its applications. 3rd edition, New York, John Wiley and Sons Ltd., 1968;1: Todhunter I. A history of the atheatical theory of probability. Cabridge and London, Macillan and Co., Reprinted 1965; New York, Chelsea. Additional inforation and reprint requests: Dr. C.G.G. Aitken Departent of Matheatics and Statistics The King s Buildings The University of Edinburgh, Mayfield Road Edinburgh EH9 3JZ, United Kingdo E-ail: cgga@aths.ed.ac.uk

12 ERRATUM Erratu/Correction of Aitken CGG, Sapling How Big a Saple? J Forensic Sci 1999 Jul;44(4): On page 750, in the second colun, second paragraph. An alternative approach, based on the binoial distribution, is discussed in (6). Consider a specific value..... should read: An alternative approach, based on the binoial distribution, is discussed in (5). Consider a specific value..... The Journal regrets this error. Note: Any and all future citations of the above-referenced paper should read: Aitken CGG. Sapling How Big a Saple? [published erratu appears in J Forensic Sci 2000 May;45(3)] Forensic Sci 1999 Jul;44(4): Copyright 2000 by ASTM International 751