American Society for Quality

A Discussion of All-or-None Inspection Policies

Author(s): Scott A. Vander Wiel and Stephen B. Vardeman
Source: Technometrics, Vol. 36, No. 1 (Feb., 1994), pp. 102-109
Published by: Taylor & Francis, Ltd. on behalf of American Statistical Association and
American Society for Quality
Stable URL: https://www.jstor.org/stable/1269203
Accessed: 24-09-2018 16:19 UTC

REFERENCES
Linked references are available on JSTOR for this article:
https://www.jstor.org/stable/1269203?seq=1&cid=pdf-reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

American Statistical Association, American Society for Quality, Taylor & Francis, Ltd.
are collaborating with JSTOR to digitize, preserve and extend access to Technometrics

This content downloaded from 193.140.253.41 on Mon, 24 Sep 2018 16:19:05 UTC
All use subject to https://about.jstor.org/terms
? 1994 American Statistical Association and
TECHNOMETRICS, FEBRUARY 1994, VOL. 36, NO. 1
the American Society for Quality Control

A Discussion of All-or-None Inspection Policies

Scott A. VANDER WIEL Stephen B. VARDEMAN

AT&T Bell Laboratories Department of Statistics,

Murray Hill, NJ 07974 Department of Industrial and Manufacturing
Systems Engineering and
Center for Nondestructive Evaluation
Iowa State University
Ames, IA 50011

An optimal inspection policy will inspect either every item produced or no item when (a) product
characteristics are well modeled as iid and (b) overall inspection cost is a sum of individually
and identically determined costs for each of the items encountered. This result is widely
known for special cases such as iid Bernoulli product characteristics with single-sample lot
acceptance-sampling plans. We show that the result holds true much more generally and over
a much wider class of inspection plans, even when independent inspection errors are possible.
We examine the assumptions that lead to all-or-none optimality and discuss the practical
meaning of all-or-none results to practitioners. Examples are given to demonstrate that both
"other" cost structures and "informative" inspections (i.e., lack of independence) can lead
to optimal policies that are not of the all-or-none type.

KEY WORDS: Acceptance sampling; Cost optimality; Imperfect inspection; Sequential de-
cision making.

Analysts have proposed and discussed hundredsfects such as internal cracks. These inspections are
of plans for sampling inspection of parts or materialrelatively expensive and imperfect. Sometimes cracks
prior to further processing or assembly. Using par- are not detected, and other times false alarms are
ticular probability and cost models Deming (1986,generated. In this setting it has been natural to ask
chap. 15) and others have demonstrated that ex- several questions:
pected costs are minimized (over particular classes
1. Which, if any, parts should be subject to non-
of inspection plans) by treating all parts identically.
destructive inspections?
That is, either every part should be inspected or all
parts should be passed for further processing without
2. Does the all-or-none result apply here?
inspection. Which alternative is best depends on the 3. Does it matter that the inspection process is
imperfect?
values of parameters in the cost structure and prob-
ability model. On the other hand, the optimal single-
This article discusses all-or-none inspection poli-
sample acceptance-sampling plans derived by Lor-cies. Our primary aim is to make clear the properties
enzen (1985) demonstrate that neither 0% nor 100%
of probability models and cost structures that lead
inspection need minimize expected costs undertoa all-or-none optimality. We do this by abstracting
Bayesian probability model. Thus the all-or-none re-
the problem from its usual grounding in single-
sult depends on the type of probability model em- sample acceptance-sampling of Bernoulli lots. The
ployed. Similarly, it depends on the cost structure.
mathematical result from our more general approach
To design an appropriate inspection plan one must is almost trivial but shows why all-or-none optimality
consider which (if any) cost structure and probabilityfollows from iid probability models and additive cost
model are appropriate for a given situation. Al-structures without being encumbered by the details
though these issues have been previously discussed
of a specific class of sampling plans. Because the aim
in many contexts over a period of 50 years, our recent of this article is primarily expository, the main results
interactions in the field of nondestructive evaluation may not seem profound to some readers. Our ex-
have convinced us that they continue to be poorly perience, however, indicates that engineers respon-
understood.
sible for determining inspection policies are often
When producing critical metal components for air- unsure whether all-or-none optimality has any rele-
craft engines, it is common to use nondestructive vance to their specific manufacturing processes and
inspection techniques based, for example, on eddy inspection methods. This discussion is intended to
currents or ultrasonics to detect manufacturing de-help analysts better understand the meaning and ap-

102

This content downloaded from 193.140.253.41 on Mon, 24 Sep 2018 16:19:05 UTC
All use subject to https://about.jstor.org/terms
 ALL-OR-NONE INSPECTION 103

plicability of all-or-none results. We also try to pro- Table 1. Deming's Cost Structure
vide some insight into the kind of cases in which the
Condition
all-or-none result does not hold and an inspection
policy can be developed with lower expected cost Good Defective
than both 0% and 100% inspection.
Inspected parts
Deming (1982, chap. 13) helped renew interest in
some arguments regarding acceptance sampling that k, k, + k
date at least to Barnard (1954) and Mood (1943) and Uninspected parts
probably beyond. Vardeman (1986) referred to sev-
0 k2 +k
eral more recent works and discussed the role of
sampling in light of modern statistical process im-
provement methods. See also Papadakis (1985) and cost of an uninspecte
Milligan (1991). duction, and k measu
Section 1 extends Deming's (fixed p) Bernoulli part to replace a defe
probability model and associated cost model to in- a generalization of th
corporate imperfect inspections. Section 2 shows that structure to include i
under the extended model inspecting either 0% or We consider a scenar
100% (depending on p) is optimal among random or are either good or def
fixed-interval inspection plans. Determining the cor- or sent directly to th
rect policy requires comparing p to a critical value assembly without ins
p, determined by various cost parameters. We dis- pass or fail the inspe
cuss the practical meaning of needing to know p. states and pass or fail
Section 3 generalizes the all-or-none result as far as iid from part to p
as possible. The broader result allows multivariate is a corresponding cost
part characteristics and imperfect measurements on on the part's condition
continuous scales with a broadly defined cost struc- The probability mod
ture. All-or-none schemes are shown to be optimal picted in Tables 2 and
within a large class of fully sequential inspection pol- the probability model
icies. The general result eliminates any need for fur- perfect inspection mo
ther all-or-none proofs for special classes of inspec- that a part is defective
tion plans. For example, it implies that either 0% or fail inspection with p
100% inspection is optimal not only in lot acceptance pass inspection with p
sampling but also among the Type 1 Continuous The cost structure in
Sampling Plans (CSP-1) of Dodge (1943). one possible in this s
If either the model supporting the all-or-none re- constants to describe six cells. The cost associated
sult is not appropriate or its parameters are unavail- with one cell has arbitrarily been set to 0 with no
able to the analyst, the best realizable policy can fail loss of generality. The costs are interpreted (relative
to be of the all-or-none type. Section 4 discusses to the zeroed cell) as follows: k, = cost to inspect a
Bayesian formulations of the inspection problem that part (this is Deming's k1), kDU = cost when a de-
address the matter of imprecisely known parameters fective part is sent on uninspected, kGF = cost when
through the use of prior distributions. Section 5 out- a good part fails inspection, kDF = cost when a de-
lines how Lorenzen's (1985) optimal Bayesian in- fective part fails inspection, and kDp = cost when a
spection policies can be extended to handle imperfect defective part passes inspection.
inspections. Section 6 motivates a cost structure under
which it is optimal to keep inspecting until the first Table 2. Probability Model With Inspection Errors
good part in a shipment is found. Section 7 gives
Condition
some concluding remarks. Inspection
result Good Defective

1. BERNOULLI PROCESS MODEL WITH Inspected parts

INSPECTION ERRORS
Pass inspection (1 - p)( - WG) PWD
Fail inspection (1 - P)WG p(1 - WD)
Deming's (1982, 1986) all-or-none argument is based
on modeling successive parts as independent, each Uninspected parts
with probability p of being defective. He used a cost 1 -p p
structure depicted in Table 1 with three constants:
NOTE: Parts are defective
k1 is the cost to inspect one part, k2 measures the probability WG. Bad parts p

TECHNOMETRICS,

This content downloaded from 193.140.253.41 on Mon, 24 Sep 2018 16:19:05 UTC
All use subject to https://about.jstor.org/terms
104 SOCTT A. VANDER WIEL AND STEPHEN B. VARDEMAN

Table 3. Cost Structure With Inspection Errors thermore, we h

using broad wor
Condition
Inspection spection scenari
result Good Defective process inspect
along with the
Inspected parts
help remove un
Pass inspection k, k, + kDp previous definit
Fail inspection k, + kGF k, + koF
critical cost rat
Uninspected parts viously importa
0 kou constants in th
end of Section 2 will illustrate this.
NOTE: The five constants in six cells give the m

2. RANDOM AND FIXED-INTERVAL

It seems likely that in any practical application of INSPECTIONS
this model all k's will be nonnegative, and we will
henceforth assume this to be the case. (This means All-or-none optimality is probably most easily
that uninspected good parts are cheapest and, among demonstrated for what we will call the class of ran-

inspected parts, good ones are cheapest.) Although dom-inspection policies. These policies advocate in-
it is not necessary for anything that follows, often specting each item independently with probability Tr.
kDu > kDF because if the inspection cost k, were Under the imperfect inspection model, the average
negligible it would often be better to inspect and fail cost per part is
a defective part than to simply send it on for further r[k, + (1 - p)wGkGF + p(l - WD)kDF + pWDkDP]
processing. Furthermore, usually kD > kDF because
it should be better to fail defective parts than pass + (1 - IT)pkDu = pkDU + Tr(k, + WGkGF- pK),
them. Finally, in many situations one would have
where K = (1 - WD)(kD - kDF) + WD(ku -
kDu = kDp because the inspection process will not
kDp) + wGkGF. If K 0, this is minimized by rr =
usually change the negative impact of a defective
0 (no inspection). Often K is positive, however, be-
part.
cause often kDU > kF and wD(kDU - kDP) is small
The specification of the cost structure here is in-
in magnitude (because kDu and kDp will not differ
tentionally general because we want it to apply in a
by much and wD will be small.) With K > 0, the per-
variety of situations, including but by no means lim-
unit cost is minimized by rr = 0 (no inspection) if p
ited to the inspection of incoming lots of parts shipped
< (k, + wGkGF)/K and by rr = 1 (inspect all) other-
from outside vendors. As an example of another pos-
wise. By defining
sible application consider the wave solder stage in
circuit-board assembly. Incoming parts at this stage PC = x, if K 0
are printed wiring boards populated with integrated-
= (k, + wGkGF)/K, otherwise,
circuit chips whose leads have been inserted through
holes in the boards. The wave solder machine solders
the optimal rule is to inspect nothing if p ' pc
the leads to copper traces onto the wiring board. inspect everything otherwise. The critical value
Sometimes, however, the chip inserter fails to geta generalization of Deming's critical cost ratio k
some leads into the proper holes. If an inspection which is obtained by setting wG = WD = 0 and i
were being contemplated to check for boards withtifying k, with his k1 and the denominator K =
improperly inserted chips, the constants in the cost- kDF with his k2.
model might have interpretations as follows: k, = If, in the preceding situation, the inspection sch
cost for an inspector to check a single board for proper is a fixed-interval inspection plan in which every
insertion of all leads, kDu = kDP = cost to repair a item is inspected, essentially the same analysis
board with improperly inserted chips after soldering be performed. In this case the role of IT is take
(and possibly further downstream assembly), and 1/m.
kGF = kDF = cost (parts and labor) for replacing Probably the most troubling aspect of this result
chips judged to be improperly inserted. is that determining whether it is 0% or 100% in-
It is interesting to compare the setup here withspection that is optimal requires knowing the rela-
that of Deming (1982, 1986). Formally, his is a spe-tionship ofp to p,. Of course, if p, _ 1, no inspection
cial case in which wG = WD = 0, k1 = k,, k2 = kDUis optimal regardless ofp. If, however, it is not known
- kDF, and k = kDF. Including the inspection errorwhether p exceeds pc then one does not know whether
probabilities (wG and wD) and associated costs (kGF to do 0% or 100% inspection. This may make the
and kDP) obviously represents a generalization. Fur-all-or-none result seem to be of little practical value.

TECHNOMETRICS, FEBRUARY 1994, VOL. 36, NO. 1

This content downloaded from 193.140.253.41 on Mon, 24 Sep 2018 16:19:05 UTC
All use subject to https://about.jstor.org/terms
 ALL-OR-NONE INSPECTION 105

Common sense and generally accepted statistical (> c2) is the labor cost incurred if a defective part
process control philosophy, however, would not al- is detected only after further downstream assembly
low a process to operate without some degree ofof the product. In this case the correct denominator
monitoring to help assure that unexpected processto use in calculating the critical cost ratio is K = kDU
upsets can be detected. Thus, in practical terms,- kDF = c3 - c2. We fear that Deming's description
"knowing p" means that there is upstream processof k2 would lead one to use either just c3 or c1 + c3
control and that process-monitoring data are avail- in the denominator.
able for deciding what kind of inspection is appro- A continuation of the wave solder example will
priate. If the upstream data indicate that p does notfurther emphasize the point. Suppose, in the ex-
reliably lie to one side of pc, then all-or-none opti-ample, that K is small (possibly even negative) for
mality results are not very helpful. If p is likely tothe following reason. Chips inserted improperly tend
be far below Pc at some times and far above Pc at to have leads that are too far apart, causing the chips
others, then the proper place to focus one's effortsto "stick" in their holes. Manually removing these
is on improving the process and bringing it into achips to insert replacements is difficult (large kDF)
better state of statistical control with (one hopes) abecause the remaining "loose chips" are easily dis-
small value of p. Deming (1986, chap. 15) also madelodged. On the other hand, it may be relatively easy
these points. to detect and replace improperly inserted chips after
If upstream monitoring shows that p varies onlywave soldering (small kDu). Thus for the pre-solder
slightly but from one side of Pc to the other, theninspection decision K = kDU - kDF is small. This
the all-or-none result is, strictly speaking, not rele-implies that Pc is large (say larger than 1) and no
vant. We should point out, however, that, althoughinspection is optimal. Failing to subtract the high pre-
the optimal r is discontinuous in p, it is clear fromsolder labor costs in the denominator would lower
the expression for per-unit cost that if p is near Pc Pc, possibly below p, leading one to recommend 100%
the per-unit cost changes very little as a function of inspection. But the cost of inspecting defective boards
rr and hence all, none, or anything in between are is too high relative to the cost of letting them through
about equivalent in terms of expected cost. to the next stage of production.
The possibility of inspection errors does not affect This reasoning is palatable in the context of one
the all-or-none nature of the optimal policy. The step in a sequential production process. When it is
functional form of pc, however, shows that inspectionapplied in the context of inspecting product to be
errors do typically push one in the direction of doing shipped to customers, however, the implication be-
no inspection. Specifically, where p, is mapped intocomes, "We cannot afford inspections due to the
(0, 1), it is an increasing function of wG and (assum-high cost of dealing with the defective product we
ing kDp > kDF) of WD. This can be understood in- make." Such reasoning is heretical to modern quality
tuitively as follows. Suppose that p = p, (i.e., inimprovement doctrine but can result from making a
terms of expected cost, one is indifferent to the amountdecision on whether to inspect by minimizing ex-
of inspection). If wG were to increase, then if partspected costs. If minimizing costs of the type included
were inspected, more good ones would fail and ex-in this analysis is not the only concern, one should
pected costs would increase because of kGF. If partsnot use only a cost-based decision criterion.
were not inspected, however, the change in wG would
have no effect on costs. Hence an increase in WG
3. GENERALIZATION
makes inspections less desirable. Likewise, if wD were
to increase, then if parts were inspected more de- Using the imperfect inspection model, Section 2
fective ones would pass and the associated costs would showed optimality of 0% or 100% inspection for ran-
increase (assuming kDP > kDF.) The expected costs dom or fixed-interval inspection policies. This sec-
for uninspected parts, however, are not affected bytion shows that the all-or-none result holds much
WD and thus no inspection would become optimal. more broadly. The Bernoulli models for good-
We have said that in Deming's critical cost ratio or-defective part states, and pass-or-fail inspection
kl/k2 (for perfect inspections), the denominator k2results are generalized to multivariate continuous part
corresponds to the difference kDu - kDF in our setup. states and inspection results, respectively. The as-
In some applications, as we hinted in Section 1, in-sociated cost model is similarly generalized. Fur-
spection costs have not been correctly allocated tothermore, all-or-none optimality is established in a
the cost constants. To illustrate this, suppose that wide class of fully sequential inspection plans. Proofs
inspections are perfect and kDF = Cl + c2, where c1for special cases can be found in the literature. For
is the cost of a good part used to replace a defectiveexample, Hald (1960), van der Waerden (1960), Lor-
one found at inspection and c2 is the associated laborenzen (1985), and Deming (1986, chap. 15) treated
cost. Suppose further that kDU = cl + C3, where c3 single-sampling lot acceptance-sampling plans; Ans-

TECHNOMETRICS, FEBRUARY 1994, VOL. 36, NO. 1

This content downloaded from 193.140.253.41 on Mon, 24 Sep 2018 16:19:05 UTC
All use subject to https://about.jstor.org/terms
 106 SOCTT A. VANDER WIEL AND STEPHEN B. VARDEMAN

combe (1958) and Wetherill (1991, chap. 14) treated E[A(X,)IZ,, ... Z,_t] = E[A(Xt)]. The fact that
CSP-1 plans. Vander Wiel and Vardeman (1992) 95 and s do not depend on t or (Z1, ..., Z,_1)
generalized these special-case proofs to include im- allows one to use an extremely straightforward back-
perfect inspections. ward induction argument to prove the following:
The general theorem that follows applies to all
sampling plans found in the work of Duncan (1986) Theorem 1. C is minimized by u, = 1 if 95 < s,
and all others of which we are aware. Mathemati- or ut = 0 otherwise (t = 1, . . ., N). That is, one
should inspect all-or-none with the choice depending
cally, it is almost trivial. It does, however, eliminate
any need to work out all-or-none results for other only on the sign of 9f - sA.
special-case sampling plans. Furthermore, because The proof is in the Appendix.
the general result is more abstract than the special The theorem holds for any fixed run size N. Since
cases treated in the literature, one can more easily the solution is independent of N, it may also be re-
see why an iid model and additive cost structure im- garded as a solution to the infinite-horizon problem
plies that each item should be treated identically. of inspecting a neverending stream of items.
Random and fixed-interval inspection plans (Sec. Qualitatively, the reason that the all-or-none result
2) might be termed nonsequential because all of the holds so universally under the iid probability model
units to be inspected can be determined before any with additive identical costs is that any information
of the inspection results are known. On the other gained on any number of units has no effect on one's
hand, CSP-1 and traditional lot acceptance-sampling assessment of the likely characteristics of the re-
plans are carried out sequentially. Decisions are made
maining units. Hence allowing even sequential de-
"on the run" regarding which units are to be in-cision making is of no extra value.
spected. We consider the most general class of (se- In Section 2, a major consideration for the use-
quential) sampling plans in which potentially a de- fulness of the Bernoulli all-or-none result was the
cision on how to proceed could be made after each need to know the sign of p - Pc. Now we need to
inspection result becomes available. Nonsequentialknow the sign of 95 - A, but the discussion of Section
plans are included as degenerate sequential plans. 2 still applies; namely, in practice upstream process-
To introduce the general all-or-none result, let {(X,,control and process-monitoring data are needed to
Yt)} be a sequence of jointly independent and iden- lend continued support to a decision of doing 0% or
tically distributed random pairs, where X, is possibly 100% inspection. (In this sense, "no inspection" does
vector valued and denotes the physical characteristics not really mean no inspection.)
of the tth unit and Y, is possibly vector valued and The imperfect inspection model is obtained from
denotes the inspection results of the tth unit. Let the
the general model by taking (X,, Y,) as iid with X,
sequence {u,} denote a general sequential inspection - Bernoulli(p) and Yt[X, ~ Bernoulli with
policy with u, = 1 indicating that the tth item is to
be inspected and u, = 0 otherwise. Each u, is allowed P[Yt 1 = w] if X, = 0
to depend on inspection outcomes obtained before
= 1 - WD ifX, = 1.
the tth item is presented for possible inspection. That
is, u, is potentially a function of the variables Zi =The cost structure of Table 3 is obtained by taking
(ui, uiY), i = 1, 2, . . . , t - 1. S(X, Y) = k, + (1 - X)YkGF + XYkDF + X(1 -
We seek an inspection policy to minimize the ex- Y)kDp and A(X) = XkDU. In this setup, we have Y
pected cost for a run of N parts given by = k, + (1 - p)wGkGF + p(l - WD)kDF + pWDkDP
and sQ = pkDU, and the condition 5f < s is equiv-
alent to p > Pc.
C = E [u,S(Xt, Yt) + (1 - ut)A(Xt)] , (1)
t=l Since random and fixed-interval inspection plans
(Sec. 2), as well as CSP-1, ..., CSP-5 (Duncan
where S(X, Y) denotes the cost of sampling, in- 1986) and many other attribute acceptance-sampling
specting, and either rejecting or accepting for use plans, are contained within the general class of se-
(according to the rules of the policy) an item with quential policies (and have both 0% and 100% in-
characteristics X and inspection measurements Y; spection as special cases), the theorem shows that an
and A (X) denotes the cost of accepting for use with optimal policy for each class of plans is all-or-none
no inspection an item with characteristics X. One
depending, respectively, on whether p > Pc or p <
could also allow A to depend on Y,, but this seems pc. Furthermore (under the imperfect inspection
unnecessary because the measurements Y, are taken model), no ingenious use of sequential inspection
only if the item is sampled. results can possibly reduce the optimal expected cost
Now define expectations Y and s by yf = E[S(X,, below that of the appropriate all-or-none policy. Under
Yt)IZ1, ... , Z,_1] = E[S(X,, Y,)] and s = non-iid probability models or cost structures differ-

TECHNOMETRICS, FEBRUARY 1994, VOL. 36, NO. 1

This content downloaded from 193.140.253.41 on Mon, 24 Sep 2018 16:19:05 UTC
All use subject to https://about.jstor.org/terms
 ALL-OR-NONE INSPECTION 107

ent from (1), however, the situation is not so clear- ing the fraction of items judged defective among those
inspected. In these cases optimal inspection plans
cut. We discuss this matter in the following sections.
need not be of the all-or-none type, and globally
4. MORE COMPLEX PROBABILITY MODELS optimal plans will generally be of a fully sequential
nature. In Section 5, we will discuss the problem of
In this and the remaining sections, we returnfinding
to optimum "single-sampling" plans in this
the case in which product characteristics and inspec-
context.
tion results are treated (possibly after classification
A much more complex situation is met if t, i
based on continuous quantities) as binary quantities.
modeled as varying unit to unit according to a serially
Under iid models, initial inspection results have no
dependent stochastic process like, for example,
bearing on one's assessment of the remaining parts,
Markov process. Wenner (1991) studied cost-optima
and thus all-or-none is optimal provided that CSP-1
costs inspection plans when pt can shift between a
are assessed as the sum of identically determined "in-control" level p, and an "out-of-control" level
costs for each part. This suggests that it may be rea-
po. We know of no other clean statement and op-
sonable to inspect a fraction of the items if doing so solution to a nontrivial inspection problem o
timal
can improve one's information about the condition this more general time-varying type. McShane and
of the remaining items. This is true and can be under- Turnbull (1991) and others referenced therein, how
stood by considering the extreme case of a manu- ever, studied properties of standard sampling plan
facturer who ships perfect lots of parts yet occasion- (e.g., CSP-1) under Markov models.
ally makes shipments of the wrong stock number
(i.e., 100% defective lots). If inspection is perfect
and not too costly, yet there is a cost associated with
5. BAYESIAN ACCEPTANCE SAMPLING OF
"defective" parts being passed on to assembly, then LOTS
it is obvious that an optimal plan is to inspect one
item from the lot to determine whether the lot is If p, the probability of being defective in the inde-
entirely good or entirely defective and replace the pendent Bernoulli setup, is the same for each item
whole lot if needed or route it to assembly otherwise. and unknown, there is no practical way to implement
More generally, one may wish to describe impre- the minimum-cost policy (that depends on p). One
cise knowledge of p, or a time-varying p (successive might consider, however, describing p with a prob-
values of which we denote as pi, P2, .. .), by means ability or density function g(p) and searching for a
of a joint probability structure forp1, P2, . Doing minimum-cost policy among those that do not de-
so can lead to many varied and interesting problems pend on p. In this case the optimal initial inspection
in sequential decision making. We will describe two policy need not be all-or-none (unless g puts all of
very narrowly defined problems of this type for which its probability to one side of p).
solutions are straightforward. In the context of lot acceptance-sampling, a single-
First, we note that it is possible to apparently gen- sample rule is a particular kind of two-stage decision
eralize the fixed p situations described by the im- rule. An initial sample of n is inspected from a lot,
perfect inspection model by adding the possibility and the remainder of a lot is inspected only if the
that each unit produced has a defective probability initial sample has more than some number c of de-
p, distributed independently according to a proba- fective units. Thyregod (1974) developed optimal ac-
bility or density function g(p) with mean po. (The ceptance-sampling plans of this type in a general con-
states of the units are then modeled as independent text, allowing the state of each part to be a
given the defective probabilities {pt}.) Such a model measurement on a continuous scale. Lorenzen (1985)
might be contemplated, for example, if before the specialized Thyregod's arguments to the Bernoulli
manufacture of each part a machine were loaded with situation that we are considering. For convenient
a cutting tool drawn randomly from a bin of tools. densities g(p), it is possible, with the aid of computer
This probability structure for the states, however, is programs, to find values of n and c that minimize
really no different from that obtained when each item the expected total cost associated with inspection.
has a fixed defective probability p0, which except for For example, Lorenzen's (1985) program uses, for
a slight notational change is the case we have already g(p), a beta(a, 8) density on the interval [0, 1] and
examined.
assumes perfect inspection. Suppose in this situation
On the other hand, if we model all units as having that inspection errors occur according to the imper-
a common but unknown defective probability, Pi fect = inspection model of Table 2, and let Y, = 1 if
P2 = . = p, where p is described by a probability the part fails inspection. Otherwise, let Y, = 0. Given
density or function g(p), then it is possible to im- p, the variables Y1, . .., YN are iid Bernoulli(py),
prove one's knowledge of uninspected units by know-where py = (1 - p)WG + p(l - wD). The prior on

TECHNOMETRICS, FEBRUARY 1994, VOL. 36, NO. 1

This content downloaded from 193.140.253.41 on Mon, 24 Sep 2018 16:19:05 UTC
All use subject to https://about.jstor.org/terms
108 SOCTT A. VANDER WIEL AND STEPHEN B. VARDEMAN

p implies a beta(a, 3) prior on py over the interval inspection results need not be Bernoulli but can
[WG, 1 - WD]. vector-valued continuous variables. Moreover, t
Vander Wiel and Vardeman (1992) showed that class of inspection policies considered can be
incorporating possible inspection errors into Lor- tended from random or fixed-interval policies to fu
enzen's problem leaves it mathematically un- sequential policies. Whether it is all or none that
changed, aside from the preceding rescaling of the optimal depends on the joint distribution of prod
prior distribution and a transformation of the cost characteristics and inspection results and the par
parameters. The proof of this involves showing that ular cost structure being used. The practical imp
the total cost functions for both problems have the cation of this is that upstream process-monitori
same form after taking appropriate conditional ex- data are needed to assure that either 0% or 100%
pectations. The random variables remaining in these inspection continues to be optimal. In all of these
conditional cost functions have identical probability arguments, we have allowed for the possibility that
structures apart from the different supports of the inspection results may not perfectly reflect the states
priors for p and py. Even with this equivalence, how- of the inspected parts.
ever, we must point out that the best of these two- If the appropriate inspection cost structure is not
stage rules is not generally optimal in the larger class a sum of identically determined component costs or
of fully sequential inspection policies. if by inspecting some parts one can obtain an im-
proved assessment of the remaining parts (as in non-
6. OTHER OVERALL COST STRUCTURES
trivial Bayesian models), then optimal inspection rules
So far we have used only an additive cost structure
need not be of the all-or-none type. Lorenzen's (1985)
that treats each part individually and identically as
Bayesian optimal single-sampling plans demonstrate
in Equation (1). To understand the importance
this. of
We have discussed the fact that adding the pos-
the cost structure to the all-or-none conclusions of sibility of inspection errors to Lorenzen's formula-
this article, consider a hypothetical scenario from tion of this problem leaves it unchanged except for
Vardeman (1986) in which a single working com- a rescaling of the prior distribution on the probability
ponent out of an early shipment of several such com- of judging a part defective and a transformation of
ponents will later be needed to complete a high- the cost parameters.
priority project (e.g., a space shuttle). Suppose that The implications of these results for practitioners
inspection of the components can be done on receipt should be clear. One should consider which proba-
at a moderate per-item cost and should the lot fail bility and cost models (if any) are appropriate to a
to contain at least one functioning component it can given situation before determining an "optimal" (or
be returned to the supplier for replacement at min- even "near optimal") inspection plan. In cases in
imal cost. On the other hand, if no incoming testing which the imperfect inspection model of Section 1 is
is done and it turns out at the time of attempted valid, our results show the effect of inspection errors
installation that the lot contains no good compo- on the all-or-none decision. Increasingly poorer in-
nents, a huge monetary penalty associated with proj- spections make doing inspections less and less cost
ect delay will be suffered. Even if a fixed p Bernoulli- effective relative to not inspecting. Larger inspection
process model is appropriate and p is known and error probabilities increase costs associated with in-
large, an optimal incoming inspection policy will spected parts but do not change costs associated with
sample until the first good component in the lot is uninspected parts.
found. Mathematization of this scenario involves a
cost structure that cannot be expressed additively as ACKNOWLEDGMENTS
in Equation (1) and produces something other than
We express our thanks to William Meeker for many
an all-or-none optimal rule.
useful suggestions on the content and presentation
7. CONCLUSION of this article. We also thank Emmanuel Papadakis
for commenting on an earlier version of the article
Under a cost structure that assesses costs as a sum
and the referees and editor (Vijay Nair) for several
of identically determined costs for each item of pro- suggestions.
helpful
duction and under an iid model for part states and
inspection results, the best inspection policy is eitherAPPENDIX: PROOF OF THEOREM 1
to inspect every item or to let all items pass with no
Let V* be the optimal cost obtainable from item
inspection. We first gave a simple proof of this result
t to item N. Then
for the special case of random or fixed-interval in-
spections and Bernoulli part characteristics whenVN
in-= min E{UNS(XN, YN) + (1 - UN)A(XN)}
UNE{O,1}
spection errors are possible. The result holds true
much more generally though. The part states and= min{Y, s}.

TECHNOMETRICS, FEBRUARY 1994, VOL. 36, NO. 1

This content downloaded from 193.140.253.41 on Mon, 24 Sep 2018 16:19:05 UTC
All use subject to https://about.jstor.org/terms
 ALL-OR-NONE INSPECTION 109

Hence the optimal fial l period decision is u = 1 Deming, W. E. (1982), Quality Productivity and Competitive Po-
sition, Cambridge, MA: MIT Center for Advanced Engineering
(inspect item N) if Y < .sa and U = 0 otherwise. Study.
Note that uN and VN do not depend on the observed (1986), Out of the Crisis, Cambridge, MA: MIT Center
history Zi, . . . , ZN- . for Advanced Engineering Study.
Now we use induction. Suppose that V*+1 = (N Dodge, H. F. (1943), "A Sampling Inspection Plan for Contin-
- t)min{S, sl}. Then uous Production," The Annals of Mathematical Statistics, 14,
264-279.

V* = min E{u,S(X,, Y,) + (1 - ut)A(X,) Duncan, A. J. (1986), Quality Control and Industrial Statistics,
,tE {0,1} Homewood, IL: Richard D. Irwin.
Hald, A. (1960), "The Compound Hypergeometric Distribution
+V,*+|z,.. ,z,_-} and a System of Single Sampling Plans Based on Prior Distri-
butions and Costs," Technometrics, 2, 275-340.
= min {u,Y + (1 - u,)s + (N - t)min{Yf, s}}
Lorenzen, T. J. (1985), "Minimum Cost Sampling Plans Using
u{Eo0,1} Bayesian Methods," Naval Research Logistics Quarterly, 32,
57-69.
= [N - (t - l)]min{Sf, os}. McShane, L. M., and Turnbull, B. W. (1991), "Probability Limits
on Outgoing Quality for Continuous Sampling Plans," Tech-
Hence the optimal decision for all periods t = 1, nometrics, 33, 393-404.
. ., Nis Milligan, G. W. (1991), "Is Sampling Really Dead?" Quality
Progress, 24, 77-81.
u* = 1 if y < .
Mood, A. M. (1943), "On the Dependence of Sampling Inspec-
tion Plans Upon the Population Distributions," The Annals of
= 0 otherwise. Mathematical Statistics, 14, 415-425.
Papadakis, E. P. (1985), "The Deming Inspection Criterion for
We see, in general, that u* and V* do not depend
Choosing Zero or 100 Percent Inspection," Journal of Quality
on the observed history Z, . . ., Z,_-. ThusTechnology,
the 17, 121-127.
optimal (potentially sequential) inspection policy is P. (1974), "Bayesian Single Sampling Acceptance Plans
Thyregod,
for Finite Lot Sizes," Journal of the Royal Statistical Society,
to inspect all items if F < s or no items otherwise.
Ser. B, 36, 305-319.
van der Waerden, B. L. (1960), "Sampling Inspection as a Min-
[Received August 1990. Revised August 1993.1 imum Loss Problem," The Annals of Mathematical Statistics,
31, 369-384.
Vander Wiel, S. A., and Vardeman, S. B. (1992), "A Discussion
of 'All or None' Inspection Policies," Statistical Research Re-
port 92-1, AT&T Bell Labs, Murray Hill, NJ.
REFERENCES Vardeman, S. B. (1986), "The Legitimate Role of Inspection in
Modern SQC," The American Statistician, 40, 325-328.
Anscombe, F. J. (1958), "Rectifying Inspection of a Continuous
Wenner, S. R. (1991), "The Economic Design of CSP-I Contin-
Output," Journal of the American Statistical Association, 53,uous Sampling Plans for Manufacturing Processes Subject to
702-719.
Occasional Upsets," unpublished paper presented at the Joint
Barnard, G. A. (1954), "Sampling Inspection and Statistical De-Statistical Meetings, Atlanta, August 18-22.
cisions," Journal of the Royal Statistical Society, Ser. B, 16,
Wetherill, G. B., and Brown, D. W. (1991), Statistical Process
151-174.
Control: Theory and Practice, London: Chapman and Hall.

TECHNOMETRICS, FEBRUARY 1994, VOL. 36, NO. 1

This content downloaded from 193.140.253.41 on Mon, 24 Sep 2018 16:19:05 UTC
All use subject to https://about.jstor.org/terms