A Summarization of Some Concepts Associated with the R.W. McMullen Paper, "Confidence Limits for a Maximum Likelihood Estimate of the Mean Number of Objects Per n Cells Based on the Fraction of Cells Unoccupied"
Dublin Core
Title
A Summarization of Some Concepts Associated with the R.W. McMullen Paper, "Confidence Limits for a Maximum Likelihood Estimate of the Mean Number of Objects Per n Cells Based on the Fraction of Cells Unoccupied"
Subject
biological weapons
bacterial cultivation
Description
Summarizing a paper about bacterial replication and asserting that tested cultivated sample data follows the Poisson Distribution, that is, bacteria is probabilistically likely to reproduce at an estimated frequency over time.
Creator
Booz-Allen Applied Research, Inc. (BAARINC)
Source
National Archives and Records Administration at College Park. RG 544, Box 2
Publisher
NARA
Date
02/01/1960
Contributor
Theodore W. Horner
Rights
Free
Format
PDF
Language
English
Type
Text
Identifier
rg544_box2_summarization_of_concepts.pdf
Coverage
1960s
United States. Army
Fort Detrick, Frederick, Maryland
Text Item Type Metadata
Text
BAARINC- PRO-TN-28 A SUMMARIZATION OF SOME CONCEPTS ASSOCIATED WITH THE R. W. MCMULLEN PAPER Theodore W. Horner February 1, 1960 PROGRAM RESEARCH OFFICE operated by Booz-Allen Applied Research, Inc. 4921 Auburn Avenue Bethesda 14, Maryland under Contract DA-18-064-404-CML-206 Biological Warfare Laboratories Fort Detrick, Maryland 1304-09 BAARING-PRO-TN-28 "A Summarization of Some Concepts DECLASSIRIED Authority NND 79758 BAARINC Technical Notes are informal documents for limited distribution. Their contents are entirely the responsibility of the individual authors. w h i s material has been prepared for the use of the United States Government and may not be used in whole or in part for any purpose by anyone else without the written permission of Booz, Allen Applied Research, Inc. 1304-09 ! BAARING-PRO- TN-28 DECLASSIRIED Authority NND 79758 SUMMARY This technical note summarizes and extends some of the concepts in the unpublished paper entitled "Confidence Limits for a Maximum Likelihood Estimate of the Mean Number of Objects Per n Cells Based on the Fraction of Cells Unoccupied" by R. W. McMullen of Stanford University. As shown in the tech note, sampler counts in the local region of a sampler follow the Poisson distribution with parameter, say 0. The sampler count divided by the volume, v , of aerosol passing through the sampler is an estimate of , which in turn es- timates the density of particles passing through the local region. As an alternative to the sampler count as an estimate of e , R. w. MeMullen proposed that the filter of the sampler be divided into n cells and that the function n In (1/p) of the fraction, P, of empty cells be used as an estimator of 0 . It is shown in the tech note that follows the bionomial distribution with expectation Ep = ₽ = e-0//n and variance P(I - P)/n for Laxge n. Expressions are developed for the mean and variance of Ô = n In (1/p) and it i s found that the estimator has a slight bias. The relative standard deviation for s is approximately "A Summarization of Some A s s o c i a t e d w i t h . t h e 1304-09 BAARING-PRO- TN-28 when P > 1 and n > 100, which griphs against f o r a fixed as a concave upward curve. When ® is small, the relative standard error is large, since only a few particles are falling on the filter. When o is large, there are not very many enpty cells. The relative efficiency of 02 compared to , the complete sampler count, 15 approximately P In (2/ ₽ ) Sor P> 2 and n > 100. This efficiency decreases to zero as the fraction of empty cells P d e c r e a s e s t o z e r o a n d a p p r o a c h e s l a s P a p p r o a c h e s 1 . I t i s also shown that the confidence limits proposed by McMullen are open to serious question because of difficulty concerning his conceptual pop- ulation of confidence statements. Alternative methods of forming confidence intervals for êz are given. A s s o c i a t e d w i t h t h e 60-4081 BAARING-PRO- TN-28 DECLASSIEIED Authority NND 79758 A SUMMARIZATION OF SOME CONCEPTS ASSOCIATED WITH THE R. W. MCMULLEN PAPER Theodore W. Horner I . INTRODUCTION The objective of the present discussion i s to summarize and slightly extend some of the concepts of the paper entitled, "Confidence Limits for a Maximum Likelihood Estimate of the Mean Number of Objects Per n Cells Based on the Fraction of Cells Unoccupied," by R. W. McMullen of Stanford University. Properties of aerosol clouds are frequently studied by particle counts on fi l t e r s of samplers strategically located with reference to cloud travel. The sampler collects on a filter all the particles of that small portion of the cloud which passes through i t . Within the local region of the sampler, which is quite large relative to the size of the sampler, i t is probably safe to assume that the particles are distributed a t random. The sampler count describes the small portion of the cloud which passes through the sampler. Primary interest in the sampler count, how- ever, centers in its use to infer the properties of the local region. Under the assumptions that the particles are randomly distributed in the local region and the cloud volume passing through the sampler is small relative to the volume of the cloud passing through the local region, the distribution of sampler counts obtained from random loca- tions of the sampler should approximate (see Appendix) the Poisson distribution: "A Summarization of Some Concepts McMullen 1304-09 BAARING-PRO- TN-28 DECLASSIRIED Authority NND 79758 f(x) = x ! where x is the sampler count and e i s t h e P o i s s o n p a r a m e t e r . t h e s t a t i s t i c where v is the volume of aerosol which passes through the sampler, can be interpreted as an estimate of where i s t h e number o f p a r t i c l e s a n d i s the volume of aerosol which passes through the local region. Thus 즙 e s t i m e t e s t h e d e n s i t y o f particles passing through the local region. I t i s well-known that the v a r i a n c e o f and that approximate 1 - & c o n fi d e n c e limits for can be constructed for large a s * + Ta Vx, where the probability that the absolute value of a normal deviate ex- ceeding i s Q. Exact, though asymmetrical, confidence limits are also easily constructed following standard procedures. I.1 Alternative Estimator for ® Obtaining a sampler count can often be quite laborious, since such counts often run into the hundreds. The purpose of McMullen's paper was to consider possible alternatives to an actual count. The proposed procedure i s to divide the fi l t e r of the sampler into c e l l s of equal size. The observed fraction, say of empty cells, perhaps obtained through a scanning device, can be used to estimate a through the equation w h e r e t h e s u b s c r i p t from the estimate o f ö distinguishes the alternative estimate = x = the sampler count.m h e o e s e t o r - 2 . "A Summarization of Some Concepts Associated with t h e McMullen 1304-09 BAARING-PRO- TN-28 1s that the expected value of i 1e cO/n and thus $ = c-/n 1.2 Properties of the Alternative Estimator The usefulness of depends upon its properties, and to develop these, we note that the fraction of empty cells, has a binomial aistribution with a mean of e-0/n, (call this P) and a variance of P(1 - P). For large values of and reasonable values o f P, s a y P > .1, the distribution of will be approximately normal (see Appendix). The rel-variance of p, which we will symbolize as V . i s "A Summarization of Some Concepts 1304-09 BAARING- PRO- TN-28 DECLASSIRIED Authority NND 79758 v_ B(1 = P1 = 1/4 (e0/n - 2). w i l l also be Now 18 e do dofined by the quation p = P+6/1-2), but approximately normally distributed with variance with a mean of zero. Thus õ = - n (In p) = - n In (P + E) = - n In P - n In (1 + $) - . . . ) 3p3 In order for the series to converge and 1. When lel < 30, we have lel < 3VSnp P) must be between - 1 -3- For P> .10 and. n > 100, ve have e<.9 and thus the series will converge for reasonably large values o f P a n d n. If In (1 +$) is approximated by employing only part of the terms of the series expansion, the absolute value of the approximation tends to be too small. This effect becoming more serious as i n c r e a s e s . S i n c e 1 8 normaly distributed, large values of Le will be rare and hence the series should have reasonable convergence p r o p e r t i e s w i t h r e s p e c t t o e x p e c t e d v a l u e s . Making use of the moments of theG S . we find 5 (.+10-0 + 10/2 + 342 + 1543 + 2034 + a n d 658 + (. +*2032 - 12 (+ 272 + 33 13 F o r p = 1 a n d n = 100, we have approximately E02 = 230.259 + 200 (.04500 + .00608 + .00182 + .00086 + .00056) =230.259 + 5.432 = 235.691 and 5.0900 + .02025 + .00778 + .00426 = 100 = 100 V.12229 = 34.970 -4- "A Summarization of Some Concepts 1304-09 BAARING-PRO- TN-28 DECLASSIFIED. Authority NND 79758 o, Thus we note the 2i s biased, although the bias is negligible for l a rg en a n d P> .1. Unbiased estimates could be obtained by computing V = Z (e°2/n - 1) and then following an iteration process using - n. Of course, the more terme of tie sories for that are used, the more unbiased will be the estimate at In the limit a s n → 00, 02 i s unbiased and its variance is which is the same as the variance for the complete count. I t might also be noted that efficiency of relative to l 1s approximately ٢٥٩ ١٠ - = P113 This relative efficiency, which is graphed against i n Fig. 2, decreases as the fraction of empty cells decreases. In choosing the number of cells into which the filter should be divided, i t is helpful to examine the relative standard error of Õ which can be approximated by 1 e / - 1 . 5 1 0 The relative standard error graphed against e f o r n = 500 is shown in Fig. 1, along with a similar curve for the relative standard error of the complete count . For the volume flowing through the sampler should be adjusted so that will be between 400 and 1300 for greatest efficiency. In order to - 5- "A S u m m a r i z a t i o n o f Some Concept: 1304-09 BAARING- PRO-TN-28 DECLASSIRIED Authority NND 79758 guard against the high relative errors associated with very high values o f 0 , i t i s suggested t h a t should be around1 0 0 0 . Using an of the order of 1000 will also tend to m a k e normally distributed and hence approximate (1 - a) symmetrical confidence limits can then be constructed for 0 - + O S T BAARINC- PRO- TN- 28 DECLASSIRIED. Authority NND 79758 "A Summarization of Some Concepts Associated with the R. W. - 6- APPENDIX A THE DISTRIBUTION OF SAMPLER COUNTS IN THE LOCAL REGION OF A SAMPLER S u n n a c e t h e l o n e l r e c i o n o f a n a e r o s o l s a m p l e r i s d i v i d e d i n t o n s e c t i o n s w i t h e a c h s e c t i o n r e p r e s e n t i n g t h e volume o f a e r o s o l t h a t would pass through a sampler located at that section. It appears reasonable t o assume that the actual sampler i s randomly located with respect t o the various sections. Further, any particle that enters the local region should have an equal chance of entering into any section. We wish to make inferences from a sampler count from a single t r i a l on a sampler located in a random section of the local region about a population of sampler counts resulting from an infinite number of trials in which N p a r t i c l e s a r e r e l e a s e d a n d i n w h i c h t h e s a m p l e r l o c a t i o n w i t h i n t h e local region i s varied from t r i a l t o t r i a l . The variable of interest can be represented by ( i = 1 , 2 , . . . n t = 1 . . . N ) gi (1 when the +th particle falls into the th cel, O otherwise. -7- *A Semarization of So W. NoMulen B A A R I N G - P R O - I N - 21 DECLASSIRIED Authority NND 79758 a n d I when the sampler is located in the 1th section of the local region, O otherwise. A property of S 1 8 B s = P ( s - 1 ) - 2 = E ( รู ) ? . Similarly, BS tS, = P(S1 - 1, 87, = 1) =P P = 1) P P(84, = 1) = n3 for t ' t , since the section into which particle t falls is independent of the section into which particle t ' falls. As regards Si, we have = ES?, since each section has an equal chance of being selected. Also BS,S,, = O, since the section selected in a trial cannot be section i and section 1 ' a t the same time. The procedure for finding the distribution of y will be to derive its moment generating function, and to show that it is the same as the moment generating function of a binomial random variable. density function of y turns out to be f(x) = (N) (E))* (1 - 1) M-Y. As the number of particles and the number of sections both become large, W= 0, the distribution of y approaches the Poisson distri- Since the volume of aerosol passing through the sampler is expected to be quite small relative to the volume passing through the local region and since the number of particles passing the local region i s expected to be -8- Associated with the R. W. McMulle 1304-09 BAARING-PRO- I N - 2 0 DECLASSIEIED Authority NND 79758 quite large, the population of interest is a Poisson population and the observed sample event is an estimate of の - - 兴 s k , where v= the volume of aerosol passing through the sampler and the volume of aerosol in the local region. The moment generating function of Y is V i s m, (h) = E OhY = E eh fS1 ESt. where E indicates expectation taken with respect to the possible po- sitions of the sampler for a given trial and taken over all trials. In obtaining BehY, it wil be convenient to let #= ) st, then RelY = I efS1"1 waich can be weiten a8 RI eneSs E I1 [2 + (c1x1 - 2) 5,30 This ts of the form BI (1 + b,) where B= (hy - 1) • Since I (1 + b. ) can be written as 1 + 도 0 + 도 로 , 0 0 , + 포 도 , 도 D D O + • • • i>i i>i - 9 - DECLASSIRIED A Summarization of So W. Mulen BAARING - PRO- TN-28 Authority NND 79758 = 2 + E (ohx1 - 3) (1/n) since ES S" = 0, along with the other terms of the series. Thus The moment generating function is then 7 ( 3 ) = 5 2 / 6 ) 2 0 4 4 ( 2 / 2 ) 5 5 0 4 N o w = B M (2 + (6" - 2) 87 1 which is of the form E I (1 + O,), where ba = (e" - 1) st. Employing the expansion for I (27 ), we have - 10- *A Sumarization of So W. Mullen 60-708T BAARING- PRO- TN- 28 DECLASSIRIED Authority NND 79758 = 2 + 5 ( - 2) (2/n) = E & (e - 1) (1/n) + . . . = 1 + Mle" - 2) (1/n) + M(I= 12 (e" - 2)2 (2/n)2 + . . . *- (*) ('-»)' (/a)' ,me 0" - 2) (2/0) +2) ] = [ Thus % = (a/n) [ [ (3 - M/n) + (2/a) o" [' - [ (3 - 2/n) + (2/n) e" J' The moment generating function of T, a binomial variable with frequency density mp (h) = [2 - P + PehTix. -11- P a d e r "A Summarization of Some Concepts the R. W. McMullen 1 3 0 4 - 0 : BAARING-PRO- TN-28 Since moment generating functions uniquely characterize the frequency density function, Y must have the same distribution as T where N = N* and P = ( 1 / n ) . Hence "A S u m m a r i z a t i o n o f Some Concept: 1304-09 DECLASSIEIED Authority NND 79758 A s s o c i a t e d w i t h t h e R . W . M C M u l l e n - 12- APPENDIX B DISTRIBUTION OF THE OBSERVED FRACTION OF EMPTY CELIS This distribution will be derived under the following assumptions: 1. The total number, k, of particles on the filter, which is divided into n cells, follows the Poisson distribution E(x) = 2-° * 2. Any particle is equally likely to be found in any cell and hence the probability that a particular particle falls into a particular cell is 1/n. Under the second assumption, the numbers of particles a , a , • • • a (Ea, = k) which fall into cells, 1, 2, . . . n follow the multi- nomial distribution for a fixed k. 5 (90 92 ... 9 ( *) = (k:/1 ay :)(3/n) E04 = (2/n)'* : -13- "A S u m m a r i z a t i o n o f Some Concept: 1304-09 BAARINC-PRO-TN-28 DECLASSIRIED Authority NND 79758 T h e s t a t i s t i c o f i n t e r e s t i s P = (2/n) E 091w h e r e 0 đ 1 = w h e n = 0. This occurs when no particles are found in t h e i - t h cell 0 otherwise. p is the observed fraction of empty cells and varies because the a's follow a multinomial distribution for a given k, while follows the Poisson distribution. The joint distribution of the a ' s i s The moment generating function of p is ™ ( t ) = B e p 0/0) ] )/( أ - } ص •* 4 "A Summarization o f Some Concept: Associated with the R. W. 60-40ET DECLASSIFIED Authority NND 79758 w e r e Q means summation over all possible combinations of the ai s u b i e c t . t o t h e r e s t r i c t i o n t h a t ia, = k. Thu: 3 (0)- * E F aMoses mp (t) = [1 - 6-(0/n) + e(t/n) 60/n]" which we recognize as the moment generating function of a binomial variable. Thus Bp = e0/n= P, say, 2 P(1 - P), and P is normally distributed for a given P and an n sufficiently large. To obtain the moment generating function above, we note that c a n b e w r i t t e n o o [ 2 + (o(4/n) - 2) 091]. Hence, 1e(/n) 6P4 -n [ 2 + Co(e/n) - 2) 001] which is of tihe form I (2 + b;), Mhexe D, = (et/n - 1) 021. The expression I (1 + L) can be willen as A S u m m a r i z a t i o n o f Some Concepts Associated with the R. W. McMullen DECLASSIRIED Authority NND 79758
Files
Citation
Booz-Allen Applied Research, Inc. (BAARINC), “A Summarization of Some Concepts Associated with the R.W. McMullen Paper, "Confidence Limits for a Maximum Likelihood Estimate of the Mean Number of Objects Per n Cells Based on the Fraction of Cells Unoccupied",” Pacific Asia War Archive, accessed November 18, 2026, https://pacificatrocitiesedu.reclaim.hosting/items/show/1464.