ML20062K151

From kanterella
Jump to navigation Jump to search
Forwards Description of How Confidence Interval Can Be Determined for Contours in Tradeoff Plots.Analysis Indicates Probabilities Shown on Tradeoff Plots within 12% of Actual Probabilities,W/Confidence of 90%
ML20062K151
Person / Time
Issue date: 07/29/1982
From: Silling S
NRC OFFICE OF NUCLEAR MATERIAL SAFETY & SAFEGUARDS (NMSS)
To: Knapp M
NRC OFFICE OF NUCLEAR MATERIAL SAFETY & SAFEGUARDS (NMSS)
References
REF-WM-1 NUDOCS 8208170065
Download: ML20062K151 (6)


Text

/

~

PC)R.

JUL 2 91932 g7 4 n %

3105.3.2/SAS/82/07/28/0 Distribution:

gv,,WtfHL file WMHL r/f I W,,~

WM r/f NMSS r/f WMiiL:

3105.3.7 JBMartin REBrowning tidBell MEIORANDUM FOR:

Malcolm R. Knapp SASilling & r/f High-Level Waste Licensing HJMiller Management Branch J0 Bunting Division of Waste Management PDR FR74-Stewart A. Silling' High-Level Waste Licensing Management'6 ranch Division of Waste Management 3UBJECT:

ERROR ANALYSIS FOR TRADE-OFF PLOTS I

The statistical techniqup/1 sed in constructing the trade-off plots presented in the 10<CFR 6b rationale document predicts the probability of fa'ilure of the assumed standard based on the number of failures in a finite samp?e.

The e,uestion of what confidence to place in these estimates arises.

The attached dccument describes how a confidence interval can be j

determined for the contours in the trade-off plots.

The analysis indicates that the probabilities shown on the trade-off plots are within 12% of the actual probabilities, with a confidence of 90%.

ORIGI::n sic:a rz Stewart A. Silling High-Level Waste Licensing Managment Branch Division of Waste Management

Enclosure:

As stated WMHL {;g :

0FC :

NAME :SASilling:1mc DATE : 7/E7 /82 8208170065 820729

[

PDR WASTE WM-1 PLR

ERROR ANALYSIS FOR TRADE-0FF PLOTS A trade-off plot is constructed by connecting points on a plane which have the same frequency of failure within a finite sample.

If this frequency of failure is to be interpreted as a measure of the probability of failure, it is of interest to examine the uncertainties introduced by the finiteness of the sample.

The failure or non-failure of a vector is a Bernoulli trial (a random event with two possible outcomes) with parameter p, which is the proba-bility of failure.

This probability is the quantity which the contours on the trade-off plot attempt to measure.

Let the sample size be N vectors.

If the parameter is p, the probability of obtaining exactly F failures in the sample may be derived from classical probability theory (ref.1):

P(Flp)=pg p (1-p)N-F (1)

F C

where P(Fjp)=conditionalprobabilityofFfailuresifpisgiven C = binomial coefficent pN N!

~ (N-F)!F!

i Equation 1 may be approximated by a normal distribution (ref. 2):

N(f-p) y

,2p(1-p)

P(flp)=

~

2) e

/2np(1-p)/N where P(flp) = conditional probability of failure fraction f if p is given f = failure fraction in sample = F/N Equation 2 gives the probability of having a certain fraction of failures for a given parameter p.

However, what is needed is the probability of having a certain parameter p given an observed failure fraction f.

This inversion of the problem may be accomplished with the help of Bayes' theorem (ref. 3), which may be applied as follows:

1

..m9

P(plf)=

(3) el P(p')P(flp')dp' 0

where P(plf)=probabilityofhavingBernoulliparameterpgiven failure fraction f in a sample P(p) = assumed probabilit/ density function for p based on a priori knowledge of system (also called the " prior distribution")

The prior distribution for p will be assumed to be a uniform distribution, since past experience with trade-off plots has shown a fairly even range of contours between 0 and 1.

In other words, it is assumed that nothing is known about the behavior of p before the sample is taken.

Substituting Equation 2 into Equation 3 using P(p) = 1 and cancelling the normalization constants from numerator and denominator, N(f-p)2 P(plf)=

p)fDII-DI)

(4) 8

,1 2p'(1-p') ge(1_p))-1/2 dp' i

e

<0 Equation 4 may be evaluated numerically. The resulting curves of P(plf) behave as shown in Figure 1 for the special case f = 0.30 for a few values of N.

Note that the curves become narrower for larger N.

This makes sense because the larger the sample size, the more confidence one has

~

that the observed failure fraction is a good approximation to the true probability of failure p.

The lopsidedness of the curves may be under-.

stood by considering the extreme case p = 0.

In that case, there is no possibility.that even a single vector in the sample would fail, hence f=0.

The same is not true of the point p = 2f, which is the same distance from f as 0 is, but could possibly be the parameter.

A confidence interval is any pair of numbers a and b such that there is a 1-a probability that the random variable is in the interval (a,b) where 1-a is the degree of confidence. One way of finding such a pair of numbers is to take a such that the probability of having the random variable in the interval (-sa) is a/2 and b such that the probability of having the random variable in the interval (b,=) is also a/2. This procedure was used to find confidence intervals for the probability distribution 2

c.

it l

f s

1 7

g)g gg;;4;gjg.

i..

x ajgi;r i bi:

.J Fi: 3p-

glj

.if; ;

0'i JIE

~ h:-

l 6-

!!};js 5[=.f i4:-

- p^-

=

5 E

f".i

- F

!t:

il:i tfi N =i25-4

fy 4-
e. i.\\;.
.y E! -

. T.

iH m.

t.

,. ; ! mit vi iF-i:h r

lW h-d.

I:-

!!j.!

til/ !::i:

W I

.(#

\\

T 1

~

~"

. %.5%::.

F./;;.#:

.fi p

4 y7

=. M9-

.3 pf y

--r..

0

.1~

.2

.3

.4

.5

.6

.7

.8

.9 1

P t

4 Figure 1.

, Conditional probability of p given f = 0.3.

(Unnormalized i

probability density function.)

t

?

i b

a f

i i

k i

3-a t

i 2

--n-wm

,-..a e

,v-

h for p described by Equation 4 with N = 50. Table 1 shows the resulting confidence intervals for various values of f and various degrees of confidence.

~

i Along the 90% confidence column in Table 1, a confidence interval of l

0.12 is sufficient to include any of the predicted intervals. Hence one may conclude that the probabilities plotted on the contour lines of a trade-off plot have at least a 90% chance of being within 12%

t of the actual probabilities.

It should be noted that the above is a conservative estimate for the confidence interval when La" tin Hypercube Sampling is used. However, it is difficult to quantify how much the use of Latin Hypercube i

Sampling improves the confidence interval for a given sample size.

' References l

(1)

H. J. Larson, Introduction to Probability Theory and Statistical l

Inference, Second Edition, John Wiley & Sons, New York (1974),

j page 125.

i (2)

Ibid., page 217.

(3)

Ibid., page 56.

e P

I E

l i

i I

i I

4 j

I i

I

--n r,,y.,

1 o.

i i

Observed Degree of Confidence (1-a)

Fraction (f) 99%

95%

90%

80%

-70%

i a

.037

.048

.054

.063

.070 0.10 b

.266

.221

.199

.176

.161 4

a

.098.

.117

.129

.143

.153 0*20 b

.379

.334

.312

.286

.270 a

.169

.195

.210

.229

.242-0*30 b

.482

.439

.416

.391

.373 a

.247

.280

.297

.319

.334 0*40 b

.578

.538

.516

.491

.474 a

.332

.369

.389

.412

.429 0*50 b

.668

.631

.611

.588

.571 a

.422

.462

.484

.509

.526 0.60 b

.753

.720

.703

.681

.666 i

a

.518

.561

.584

.609

.627 i

0*70 b

.831

.805

.790

.771

.758 a

.621

.666

.688

.714

.730 0*80-b

.902

.883

.871

.857

.847 a

.734

.779

.801-

.824

.839 0.90 t

b

.963

.952

.946

.937

.930 1

1 Table 1.

Confidence intervals for N = 50. There is a 1-a probability that the Bernoulli parameter p lies in the interval (a,b) if a fraction f of the l'

vectors fail.

I i

5 t

-,