See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/24098508

Effects of Quality and Quantity on Decision Effectiveness

Article  in  Journal of Consumer Research · February 1987

DOI: 10.1086/209106 · Source: RePEc

CITATIONS READS

528 4,973

2 authors, including:

Richard Staelin
Duke University
143 PUBLICATIONS   17,940 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

The Role of Paid, Earned and Owned Media in Building Entertainment Brands: Reminding Informing and Enhancing Enjoyment View project

A “Meta-Analysis” of Multibrand, Multioutlet Channel Systems View project

All content following this page was uploaded by Richard Staelin on 02 June 2014.

The user has requested enhancement of the downloaded file.

Journal of Consumer Research, Inc.

Effects of Quality and Quantity of Information on Decision Effectiveness

Author(s): Kevin Lane Keller and Richard Staelin
Reviewed work(s):
Source: Journal of Consumer Research, Vol. 14, No. 2 (Sep., 1987), pp. 200-213
Published by: The University of Chicago Press
Stable URL: http://www.jstor.org/stable/2489411 .
Accessed: 14/02/2013 10:45

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
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].

The University of Chicago Press and Journal of Consumer Research, Inc. are collaborating with JSTOR to
digitize, preserve and extend access to Journal of Consumer Research.

http://www.jstor.org

This content downloaded on Thu, 14 Feb 2013 10:45:33 AM

All use subject to JSTOR Terms and Conditions
Effects of Quality and Quantity of
Information on Decision Effectiveness

KEVINLANE KELLER
RICHARDSTAELIN*

Providingattributeinformationabout alternatives can both help and hinderconsumers

evaluate products. We separate attribute information into two components: infor-
mation quality and information quantity. We hypothesize, all else being equal, that
the former component improves decision effectiveness while the latter component
decreases decision effectiveness. The results of an experimental investigation de-
signed to estimate the marginal effects of these factors were compatible with these
premises. In addition, evaluations were also found to suffer in situations where high
levels of information quality were made available. Insights into these and other
observed effects are suggested after exploring the process by which subjects used
the available information.

One major concern of policy makers and research- effectiveness. Second, using this definition, we develop
ers interested in how individuals use available in- a model of how decision effectiveness is affected by two
formation is the determination of the maximum different dimensions of information: information qual-
amount of information that a consumer can effectively ity, which can be thought of as the information's in-
process before experiencing decreased accuracy in as- herent usefulness to consumers in assessing the utility
sessing the value (utility) of available choice alternatives. of an alternative; and information quantity, defined as
Currently, advocates for providing substantial amounts the number of items or attributes describing an alter-
of information often act as if this maximum is a large native (e.g., brand) (Summers 1974; Wilkie 1974). The
number. Although it is difficult to dispute the concept basic assertions to be tested are that in situations where
of the consumers' "right to know," some researchers a consumer is given a particular set of information, and
have warned that policies aimed at increasing infor- thus external search costs are not relevant, (1) increases
mation can lead to poorer decisions, citing numerous in the total quality of information made available,
laboratory experiment findings that show humans to be holding fixed the quantity of information, increase the
cognitively limited (e.g., Kelly and Fiske 1951; Miller likelihood of effective decisions (to be defined more
1956; Simon 1969; Streufert and Driver 1965). precisely below), and (2) for a fixed total quality level
In this article, we explore the effects of the available of information, increases in the quantity of information
information environment on the consumer's ability to decrease the likelihood of effective decisions.
make an accurate assessment of the best alternative in
a choice set. In doing so, we deviate from prior mar- CONCEPTUAL BACKGROUND
keting studies of this topic in a number of ways. First,
we explicitly define what we mean by a consumer mak- About a decade ago, Jacoby and his associates (Ja-
ing an accurate assessment-a concept we call decision coby, Speller, and Berning 1974; Jacoby, Speller, and
Kohn 1974) presented the results of two experiments
*Kevin Lane Keller is Assistant Professor of Marketing, Graduate designed to ascertain the influence of the amount of
School of Business, Stanford University, Stanford, CA 94305. Richard information available to a consumer on his/her ability
Staelin is Edward and Rose Donnell Professor of Business Admin- to make a correct choice among food products. This
istration, Fuqua School of Business, Duke University, Durham, NC work, which claimed to establish the existence of "in-
27706. This research was supported in part by a Duke University
formation overload," was questioned immediately by
Research Council grant and an Office of Naval Research grant. The
authors would like to acknowledge the helpful comments of Jim a number of researchers (Russo 1974; Staelin and Payne
Bettman, Dipankar Chakravarti, Mike Hagerty, Eric Johnson, J. Ed- 1976; Summers 1974; Wilkie 1974) on both conceptual
ward Russo, the Duke Marketing Seminar participants, and four and methodological grounds. One criticism concerned
anonymous reviewers, one of whom suggested the specific formulation the question of what is the best choice, i.e., how does
for Equation 1. The cooperation of Donn Saunders, Placement Di-
rector at the Fuqua School of Business, is also acknowledged. one determine the alternative a consumer would have
chosen if the information environment were optimal.

200
?3 JOURNAL OF CONSUMER RESEARCH* Vol. 14 * September 1987

This content downloaded on Thu, 14 Feb 2013 10:45:33 AM

All use subject to JSTOR Terms and Conditions
INFORMATIONAND DECISION EFFECTIVENESS 201

A second criticism was that these studies varied not Given this uncertainty, we decided to address the prob-
only the number of attributes provided, but also the lem again.
choice set size, thereby varying the probability of getting The approach taken here differs from the previously
a correct choice by chance alone. Russo, Wilkie, and discussed studies in a number of ways. First, we center
Staelin and Payne all attempted to circumvent this sec- our attention on the construct "decision effectiveness,"
ond concern by reanalyzing the Jacoby et al. data after which is concerned with how consumers assess the value
controlling for the effects of guessing. Although these of each alternative. This is in contrast to the prior stud-
analyses differed slightly from one another, each found ies' measure of the "best" choice. Second, we explicitly
little or no decrease in choice accuracy associated with acknowledge that processing additional information has
increases in the amount of attribute information pro- both benefits and costs. This is done by considering the
vided, controlling for choice set size. effects of the two components of information, quality
Following these exchanges (see Jacoby, Speller, and and quantity, on decision effectiveness. Third, we de-
Berning 1975; Jacoby 1977 for replies), Scammon velop a conceptual model of how decision effectiveness
(1977) published the results of a study that fixed the affects choice. This model influenced the types of mea-
number of alternatives among food products at two sures taken in our study, the sequencing of these mea-
while varying both the number of attributes for which sures, and the forms of analyses used subsequent to the
information was provided and the format of the infor- data collection. Given the importance of these consid-
mation. Subjects were asked to identify the brand that erations, we discuss each in detail in the next sections.
was nutritionally superior instead of choosing their most
preferred brand. She reported no increase in the selec- MODEL OF DECISION CHOICE AND
tion of the most nutritious brand resulting from in- DECISION EFFECTIVENESS
creases in attribute information.
Malhotra, Jain, and Lagakos (1982) presented yet In order to determine what is a best choice, it is either
another analysis of the Jacoby et al. data base as well necessary to have a universally accepted standard of
as an analysis of the Scammon data base. These anal- what is best or to establish how each consumer ideally
yses, which used a variation of the multivariate analysis evaluates alternatives. The studies by Scammon (1 977)
model originally used by Staelin and Payne (1976), si- and Muller (1984) took the former route using a choice
multaneously controlled for the number of alternatives task where presumably there was a "right" answer, i.e.,
and the number of attributes per alternative available. choose the most nutritious brand. However, many nu-
They found, in general, no decrease in accuracy as the tritionists (and consumers) believe what is best (i.e.,
amount of attribute information increased. Malhotra most nutritious) for one person may not be best for
(1982) also collected a new data base utilizing the same another person. Given this difficulty of finding a uni-
basic design first introduced by Jacoby et al. He found versally accepted standard for best choice independent
that when consumers were asked to choose among of individual consumer's tastes, we decided instead to
houses, choice accuracy decreased when the number of develop a general conceptual model of how highly mo-
attributes increased from 5 to 15 or more (up to 25) tivated consumers would ideally value alternatives in a
and when the number of alternatives increased from 5 choice task. This model assumes that under ideal con-
to 10 or more (also up to 25). ditions consumers want to review all of the important
More recently, Muller (1984) reported the results of attribute information available for each alternative and
a field study that varied the amount of nutritional in- assign an overall utility score to each alternative in their
formation provided to consumers (via point-of-pur- choice set by summing the attribute level utilities. They
chase signs hung in the aisles of grocery stores) and then choose the alternative with the largest overall util-
measured purchase behavior for a two-week period. He ity score. Said more directly, we assume (1) motivated
found that varying the information load on these signs consumers like to be compensatory decision makers,
produced no systematic shifts in purchase behavior over and (2) compensatory decision models are appropriate
the two-week period analyzed. Finally, Best and Ursic standards to measure decision effectiveness (Hagerty
(1986) following the basic methodology of Malhotra and Aaker 1984; Johnson and Payne 1985).
(1982), reported that choice accuracy was more sensitive The above conceptualization is appropriate for "per-
to the degree of preference similarity among brands than fect" informational environments and processing con-
to the number of attributes or brands provided. ditions, i.e., situations where the information is relevant
The preceding results provide no unified picture of and consumers can accurately process all available in-
the relationship between the amount of information formation and are willing to incur the "cost of thinking"
available at the time of choice and choice accuracy. (Shugan 1980). Notationally, u* represents this "ideal
This inconsistency may be due to methodological compensatory utility" for alternative i based on the
problems and/or differences in statistical power across available information, where the cost of processing the
studies. However, it also may be due to the fact that information is not included in the value of u*.
there is no such phenomenon as information overload. In reality, the informational environment and pro-

This content downloaded on Thu, 14 Feb 2013 10:45:33 AM

All use subject to JSTOR Terms and Conditions
202 THE JOURNAL OF CONSUMER RESEARCH

cessing conditions are often less than perfect. Such im- Specifically, our conceptual model of how quality and
perfections can decrease consumers' motivation, lead quantity affect decision effectiveness rests on the fol-
to processing "short cuts," and cause confusion and/ lowing premise:
or inaccurate inferencing. Define uzito be the consumer's
Pla: Increasing the quantity of attribute infor-
utility assessment for alternative i formed under such
imperfect environment and/or processing conditions mation while fixing the total quality of the
available information leads to a decrease, on
and let Ei= u - uzi.Small absolute values of Eiindicate
average, in decision effectiveness, although
a person is effective in assessing the ideal or true utility
value of the alternative while large absolute differences this decrease occurs at a decreasing rate.
are associated with ineffective assessments. In this way Plb: Increasing the total quality of information
lEilis inversely related to our concept of decision effec- while retaining the same quantity of infor-
tiveness. mation results in an increase, on average, in
More formally, we operationalize our concept of de- decision effectiveness, although this increase
cision effectiveness, yv*',as follows: occurs at a decreasing rate.
Mathematically,
=
(1)
* +* lei
y*= -g1(N) + g2(TQ) + 2i (2)
We use this definition for two reasons. First, larger val- where N is the quantity of information, TQ is the total
ues of y*~imply greater decision effectiveness. Second, quality of the information, E2i is an error term, gj( )
y*' is independent of the metric chosen for u* and uzi. are the functions that reflect the effects of quantity and
quality on decision effectiveness, with first derivatives,
MODEL OF INFORMATION g'( *) > 0, and second derivatives, g'( *) < 0, for i = 1,
2. The first partial derivatives capture our assumption
QUANTITY AND QUALITY EFFECTS of the direction of the effect while the second partial
Another criticism of the original Jacoby et al. meth- derivatives reflect our assumption of the decreasing
odology was that it did not acknowledge that each at- marginal influence of both factors.
tribute varies in the amount of evaluative information Although the constructs quantity and total quality
that it provides to people. For example, information are conceptually distinct, these two constructs, as de-
on the nutrient content of a food might be very relevant fined, are related. One way of making this relationship
to one consumer but not to another. We capture this explicit is to define AQ to be the average quality level
concept with the theoretical construct, quality of in- of the attributes available. Then TQ = AQ N, and .

formation, which refers to the usefulness of the available Equation 2 becomes

attribute information in aiding a decision maker to
yiv =-gl (N) + g2(AQ*N) +E2i, (3)
evaluate his/her true utility associated with an alter-
native. One measure of this construct might be the cu- i.e., decision effectiveness is a function of quantity and
mulative importance of the attribute information, i.e., the average quality level of the available information.
the total relevance of the available information. Since Equation 3 provides a natural mechanism for ex-
higher quality, more relevant information should en- ploring "information overload," since it operationalizes
hance a person's motivation and/or ability to process the question of when (if ever) does decision effectiveness
that information and reduce a person's need to infer, decrease with increases in quantity, N, assuming the
we hypothesize that, all else being equal, decision ef- average quality level, AQ, of the informational environ-
fectiveness should increase as the quality level of the ment is held fixed. Said differently, is Equation 3
information provided increases. monotonically increasing in N (i.e., there is no such
In normal situations, however, increases in the qual- phenomenon as information overload) when AQ is held
ity level of information do not occur with all else con- fixed, or does it first increase and then decrease as pre-
stant. Thus, when more attribute information is pro- dicted by information overload?
vided to increase the quality level of the informational Three points arise from such a question. First, the
environment, so is the quantity of information in- answer must be determined empirically. Second, no
creased. Increases in the number of attributes provided matter which answer proves to be true, both views im-
increase the cognitive costs associated with processing plicitly assume that decision effectiveness initially in-
this information. In fact, a useful way of thinking about creases with increases in N. This puts a logical constraint
the concept of information overload is that it arises on gl(*) and g2(*), i.e., the decreasesdue to gl( *) must
when the incremental decreases in decision effectiveness be less than the increases due to g2( * ) for small N.
due to additional information quantity are greater than Mathematically, this implies g'(N) <g2(AQ. N) for
the incremental increases in decision effectiveness due small N. Finally, if information overload occurs at some
to additional information quality. value of N, say N*, then g'(N) must be greater than

This content downloaded on Thu, 14 Feb 2013 10:45:33 AM

All use subject to JSTOR Terms and Conditions
INFORMATIONAND DECISION EFFECTIVENESS 203

g2(AQ. N) for values of N greater than N*. The latter served values uis. This requires us to assess the degree
two points imply that any functional form used to rep- of correspondence between the uis and the ^is. We do
resent Equation 3 must be flexible enough to incorpo- this by noting that the two variables are linked by the
rate both conditions. We address this issue again when construct u* . More precisely, the probability that BEST
we operationalize Equation 3. equals one increases as the value of uziapproaches ui.
The major objectives of this study are to test the This will occur as uziapproaches uO'holding fixed ui or
Proposition and to determine if information overload ui approaches u* holding fixed uzi.The former corre-
exists in at least one type of choice task. One way of spondence is increased with increases in decision effec-
accomplishing these objectives is to estimate gl( * ) and tiveness, while the latter is increased with decreases in
g2(*) of Equations 2 and 3 directly. However, such an the magnitude of the variance of measurement error
approach requires measuring the dependent variable E3i, Var(E3i) In addition, the probability that BEST
yV*',which is derived from two unobserved variables, equals one should be influenced by the relative values
u0' and uzi.Measuring u0"may not pose too much of a of the uO's, i.e., the degree to which the alternative with
problem since in a high-involvement situation there are the highest utility dominates the other available alter-
established measurement techniques to assess a person's natives (in particular, the alternative with the second
utility, e.g., conjoint analysis (Green and Srinivasan highest true (ideal) utility).
1978) or multiattribute utility models (Huber 1974; Ja- This discussion can be quantified as follows. First,
coby et al. 1974a; 1974b). Using any one of these pro- operationalize the task environment, TE, to be the dif-
cedures would produce a fallible but unbiased measure, ference between the true compensatory utilities of the
ui, of the construct up'. Mathematically, first and second ranked alternatives, i.e.,
+ E3i,
Ui= U1*' (4) TE=u*_U2*+E4 (5)
where E3i reflects the inability to measure precisely a = U1 - U2 + (4 + E32 - E31
person's ideal compensatory utility function for alter-
native i. The construct ui, on the other hand, may be where u* is the true utility for the highest ranked alter-
more difficult to measure since it represents the person's native, u2*is the true utility for the second highest ranked
utility assessment in less than ideal informational or alternative, and (4 is the error term, which captures the
processing conditions. Any attempt to measure uticould fact that the task difficulties can be caused by other
easily cause a subject to be "more careful," thereby alternative utility differences, e.g., ut' - uT, uT - U4,
biasing the assessment of ut?. and so on. Large values of TE imply an easy task, while
Fortunately, there is another approach that can be small values imply a difficult task.
used to estimate the effects gl( *) and g2( * ) as specified Second, operationalize our opinions on the relation-
in Equations 2 and 3. In this approach, we postulate a ship between yi* and N and TQ using a functional form
relationship between Y1*and an observable outcome that reflects the premises set forth in the Proposition,
measure. We then substitute for y* (after assuming a Equation 2, and Equation 3, i.e.,
specified form of gl( * ) and g2( )), thereby replacing the yj* = -log(N) + al log(TQ)-a2TQ2 + E2i (6)
unobserved variable yi* with observed variables (i.e., N
and TQ or equivalently N and AQ). Interestingly, this where al and a2 are positive constants that are restricted
resulting "reduced form" equation approach still en- to insure that the relationship between y'* and TQ is
ables us to identify (and estimate) the effects of N, TQ, monotonically increasing over the relevant range of
and AQ on decision effectiveness as defined in Equations TQ.1Next, relate yi* to a measure such as the probability
2 and 3. that variable BEST equals one by specifying how we
We limit our remaining discussion in this section to believe decision effectiveness, task difficulty, and mea-
one particular observable variable, BEST, defined to surement errors affect this variable, i.e.,
equal one if the person chooses the alternative deemed
best from all of the alternatives under consideration, P(BEST= 1) = bi + b2yI#+ b3TE-b4Var(E3i) + (5, (7)
based on the fallible utility measure ui, and zero oth-
erwise. In later sections, we develop and discuss a second where bi > 0, i = 2, 3, 4 and (5 is an error term. Sub-
observable measure, PAIR, which is identical to BEST
except the choice set is limited to two possible alter-
natives. 'Equation6 is no less generalthany;* = co- clog(N) + c21og(TQ)
The variable, BEST, as defined, is identical to the - c3TQ2 + Ecisince coand cl can be absorbedby rescalingthe unob-
dependent measure used by Jacoby et al. and Malhotra servedconstructy;* The specificforms of g, and g2 were chosen in
(1982). However, the development differs from prior recognitionof the fact that g'(N) had to be less than g2(AQ. N) for
small N and that the overall function should be flexibleenough to
analyses by explicitly acknowledging that the person's allow for Yi*to be eithera monotonicallyincreasingfunctionof N or
actual choice is based on the unobserved utilities uzis an increasingandthendecreasingfunction,i.e., g'(N) couldbe greater
while the determination of BEST is based on the ob- thang2(AQ Q N) for values of N greaterthan N*.

This content downloaded on Thu, 14 Feb 2013 10:45:33 AM

All use subject to JSTOR Terms and Conditions
204 THE JOURNAL OF CONSUMER RESEARCH

stituting Equation 6 into 7 and simplifying yields the a = (61)/f2 and

reduced form equation
P(BEST= 1) = b- b210g(N)+ b2a,logTQ
a2= (62)/ba2. (9)
-b2a2TQ2+b3TE+E6 (8) In summary, it is possible to identify and estimate
where E6 captures the cumulative effects of three unob- the underlying structural coefficients of Equation 6 even
served terms, E2i, the variance of f3i, and E5. Finally, though we never observe the underlying construct, de-
estimate P(BEST = 1) with the observed variable cision effectiveness. In addition, such a procedure can
BEST. yield multiple estimates of the structural coefficients
The above development, in addition to being explicit provided one can hypothesize multiple relationships
about choice based on ui and ui, also makes explicit the between y?*and different observable variables. In fact,
need to assume at estimation that the unobserved vari- we do just this in the results section. Finally, the coef-
able Var(E3i) is not correlatedwith any of the indepen- ficients of the reduced form equation (e.g., Equation 8)
dent variables, TQ, N, and TE; otherwise their reported have a direct interpretation if one is interested only in
coefficient estimates would be biased. the input-output relationship between the informational
Since it is impossible to determine any relationship and task environments (as denoted by N, TQ, and TE)
directly (we never observe Var(E3i)), we next determine and the (fallible) output measure, correct choice. We
the sign of any bias under two different assumptions. discuss such relationships in some depth in the results
Under the first assumption, we assume the variance of section.
f3i increases with increases in u*. More specifically, if
we assume that the standard deviation of 3i iS propor- RESEARCHF1 Ms3ETH[ODOgLOGY
tional to u0, then it is easy to show that TQ (holding N
fixed)is positivelyrelatedto Var(E3i). This relationship We test our proposition and the veracity of the phe-
implies that omitting the Var(E3i) term results in the nomenon of information overload with a three-phase
coefficient (b2aj) being negatively biased, a direction job choice experiment using second-year MBA students.
that runs counter to the hypothesized sign. Under this In Phase I, as part of a survey conducted by the Place-
scenario our test on (b2aj) is conservative. ment Office at the Fuqua School of Business, Duke
The second assumption is based on the premise that University, subjects were asked a series of questions
the variance of the measurement error (3i increases with aimed at assessing their "part-worth" utility, (Uki for
N. This might be true if one uses a compositional attribute k, alternative i) for 17 different attributes often
method to measure Ui (e.g., add up fallible part-worth), associated with a job. These attributes, which were cho-
since variance of a sum is positively related to the sum sen through a review of the relevant job preference lit-
of the individual variances. Such an assumption should erature (e.g., Jurgensen 1978) and discussions with
have no effect on the bias for the variable log( TQ) since placement officers and students, along with average im-
N is assumed to be fixed. However, the positive rela- portance ratings in parentheses, are: firm size (53), op-
tionship between the Var(E3j)and N could cause a neg- portunity to increase job responsibilities (86), indepen-
ative bias for the coefficient for log(N). On the other dence (80), influence over firm's performance (75), firm
hand, if one assumes that the percent error is constant prestige (66), length of work week (52), location (75),
(i.e., the Var(E3i) increases with increases in u*), then industry growth (58), travel required (59), opportunity
it is easy to show that N is independent of the Var(E3i) for training (95), initial salary (78), relative industry
once TQ is held fixed. Which assumption better mirrors performance of firm (62), expected pay increases (74),
reality is an empirical question and has significant im- job security (55), length of time firm has been in exis-
plications on the interpretation of the results. We ad- tence (34), "line" or "staff" (50), and location within
dress this question in more detail in the results section. firm (58).
Finally, within a cell (i.e., for a fixed N and TQ) there Part-worth utilities were estimated using the com-
is no logical systematic relationship between the mag- positional methodology recommended by Huber and
nitude of TE and Var(E3i). In sum, the only possible his colleagues (Hoepfl and Huber 1970; Huber 1974;
bias is the effect on the coefficient for log(N) caused by Huber, Daneshgar, and Ford 1971). This "self-expli-
omitting Var(E3i). cated" approach suggested by Huber requires each re-
Since Equation 8 is stated in terms of observable spondent to specify (1) the importance, Wk, of each of
variables, it is possible to obtain individual coefficient the 17 attributes, anchoring on 100 points for the most
estimates (i.e., bl, b2, (b2aj), and (b2a2).Although these important attribute, and (2) for the kth attribute, the
coefficients do not provide direct estimates of the un- attribute level utility (value) on level j, ekj(i), by an-
derlying structural coefficients of interest (i.e., a, and choring on 100 points for the most desired level and
a2 as defined in Equation 6), it is possible to estimate zero points for the least desired level. Using the standard
these parameters from the reduced form estimates, i.e., additive "part-worth" model it was possible to use this

This content downloaded on Thu, 14 Feb 2013 10:45:33 AM

All use subject to JSTOR Terms and Conditions
INFORMATIONAND DECISION EFFECTIVENESS 205

information to calculate for each respondent their ui TABLE 1

for any given alternative i, i.e., AVERAGE AND CUMULATIVE IMPORTANCE SCORES
n
Ascending Average Cumulative Cumulative Descending
Ui= Uki, (10) attribute importance (ascending (descending attribute
k= I
rank score rank) rank) rank

where Uki = wkekj(i)and n signifies the number of at- 1 99.5 99.5 1099.1 17
tributes given. Thus, the overall utility of an alternative 2 92.0 191.5 999.6 16
equals the sum of the associated attribute part-worth 3 88.0 279.5 907.6 15
utilities, where these part-worths reflect the attribute's 4 85.0 367.5 819.6 14
importance and the evaluation of its particular level. 5 82.0 446.5 734.6 13
6 79.0 525.5 652.6 12
Of course, other compensatory models could have been 7 75.0 600.5 573.6 11
used. For example, one could hypothesize a compen- 8 71.5 672.0 498.6 10
satory model that incorporates interactions. However, 9 67.7 739.7 427.1 9
since the additive, noninteractive model has been shown 10 63.3 803.0 359.4 8
11 58.9 861.9 296.1 7
to be very robust in a wide variety of circumstances, 12 53.5 915.4 237.2 6
we use this form. 13 48.3 963.7 183.7 5
In Phases II and III, subjects were asked to choose 14 43.1 1006.8 135.4 4
between pairs of hypothetical job alternatives in addi- 15 36.9 1043.7 92.3 3
16 31.5 1075.2 55.4 2
tion to choosing the best alternative among a set of five
17 23.9 1099.1 23.9 1
different hypothetical job alternatives in informational
environments that were precontrolled for specific values
of information quantity, information quality, and task
ease. The control in these choice tasks was accomplished The two major design variables were information
by using the Phase I utility data to construct the specific quantity with four levels (4, 8, 10, or 12 attributes) and
job alternatives presented for each subject according to information quality, also with four levels (cumulative
the following procedure. importance scores respectively of 365, 510, 655, and
Information quantity was operationalized as the 800-plus or minus 35). The information quality levels
number of attributes provided for each job alternative. represent approximately 33, 46, 60, and 73 percent, re-
Information quality was operationalized as the cumu- spectively, of the average cumulative importance scores
lative score of the individual's importance weights (i.e., for all of the 17 attributes. Although crossing the two
the sum of the WkS) for the attributes provided. Table factors would yield a design with 16 possible quantity-
1, for example, displays the importance scores for the quality combinations (cells), it was not feasible to con-
17 attributes obtained by first ranking the attributes for struct choice sets that fit in all of these cells for two
each individual and then averaging the ranked attribute reasons. First, since the maximum quality score for four
scores over all individuals. If we were to assume for the attributes was 400 (=4 X 100), it was not theoretically
moment that these data reflect one particular subject's possible to provide a medium quality informational
importance ranking, then a choice set of jobs where the environment (a quality score of 510) or higher with only
first eight ranked attributes were given would reflect an four attributes in our job choice set, thus negating the
informational environment of eight attributes with a use of three cells. Second, it was usually empirically
total quality level of 672. Similarly, if the last eight impossible to find individuals whose importance scores
ranked attributes were provided, the informational en- allowed them to be placed in four other cells (e.g., when
vironment would have a total quality level of 359. quantity was set at 12, almost no individuals had a cu-
By selectively choosing attributes for each individual, mulative importance score of 510 or less for any com-
given the Phase I data, it was possible to construct choice bination of attributes).
sets of approximately equal cumulative importance for Therefore, we were able to potentially develop choice
each subject. The Phase I data were also used to choose sets of job alternatives that fell into nine of the 16 cells
attribute levels for each of the individual's five job al- (see Table 2). Even then, many individuals had utility
ternatives to insure that the average utility difference functions that only allowed us to construct choice sets
between alternatives was approximately held constant, that placed them either into cells 1, 3, 4, 6, and 7 or
i.e., (Uj- (u
-Uj -uj+2) for j = 1, 2, 3 where j cells 2, 5, 6, 8, and 9 as defined in Table 2. Since subjects
orders the alternatives and uj is based only on those placed in the former group received a given quality of
attributes provided in the choice task. In this way we information using fewer attributes, the two groups dif-
controlled for task ease across different job choice sets. fered with respect to the average quality level per at-
In addition, the attribute levels were chosen so that no tribute. In this way the group identity captures any
one alternative was dominated on every attribute, en- between-subject differences that might be due to the
couraging compensatory processing. contextual influence of receiving informational envi-

This content downloaded on Thu, 14 Feb 2013 10:45:33 AM

All use subject to JSTOR Terms and Conditions
206 THE JOURNAL OF CONSUMER RESEARCH

TABLE 2 satisfaction,(b) confusion, (c) certainty,and (d) like-

EXPERIMENTAL DESIGN SHOWING NINE QUANTITY-QUALITY
lihood of regret)
COMBINATIONS USED 4. Desire for more information(measuredby a seven-
point Likert-typescale and an open-ended question
Qualitya instructingsubjectsto list desiredattributesnot pro-
vided in their choice task)
Low Medium High Very high
Quantityb (360 ? 35)C (510 ? 35)C (655 ? 35)C (800 ? 35)C 5. Usefulnessof the informationprovidedmeasuredon
a seven-pointLikert-typescale.
4 1
8 2 3 4 In addition, for the last choice task that a subject con-
10 5 6 7 sidered, the following measures were taken in an at-
12 8 9
tempt to capture different decision processes or strat-
Quality = cumulative sum of attribute importance ratings.
egies across the cells:
b Quantity = number of attributes.
c Cumulative importance scores. 6. Two self-reportedmeasuresof the decision procedure
used (measuredby a written open-ended, post-deci-
sion protocol and agreement among six different
ronments that varied on the average quality level per statementsdesignedto reflectwhetherthe personused
attribute (holding cumulative quality level fixed). a compensatorystrategy,tradingoff attributes,or used
Since pretests indicated few if any order effects as- a noncompensatorystrategy,i.e., any other strategy)
sociated with asking a subject to do more than one job 7. Two measuresof the amount of provided attribute
choice task, each subject was assigned five choice tasks, informationactually used in choosing the most pre-
(i.e., either cells 1, 3, 4, 6, and 7 or 2, 5, 6, 8, and 9) ferred alternative, (measured by a four-point scale
where the order of presentation of the cell was rotated anchoredby "just a few" and "all of them" and an
across subjects within a group. We explicitly account open-endedquestioninstructingsubjectsto list those
for any order effect in our analysis by including an order attributesthey used in choosing the most preferred
variable as a covariate in all analyses. alternative).
All Phase I subjects who qualified for either group Finally, one multiple choice measure of the subjects'
were contacted and asked if they would like to partic- general strategy of dealing with "missing" information
ipate in a job choice experiment. About 52 percent of (i.e., extent of inferencing) was taken.
the 105 eligible subjects signed up for Phase II. Subjects The choice task settings for Phase II and Phase III
were run in groups about two months subsequent to were identical except in terms of the format of infor-
the Phase I measurement. The task took about 45 min- mation presented and the timing: the format for Phase
utes to complete; each subject received $6 compensation II job descriptions was an attribute by alternative grid
for his/her time. No subject mentioned any connection (matrix) format, while the format for Phase III was or-
between Phases I and II. Phase III subjects were con- ganized around descriptive job paragraphs. Thus, the
tacted about three months after Phase II. In this in- two phases represent a format manipulation (con-
stance, 43 subjects agreed to participate, of whom 29 founded with the three-month time delay).
had also participated in Phase II. Coinpensation was
again $6 per subject.
METHODOLOGICAL ASSESSMENT
The Task-Phases II and III Before discussing the results, we explore the following
three key questions concerning our measure of infor-
After reading the instructions, Phase II and Phase III mation quality:
subjects participated in five choice tasks. Each choice
task required the subjects to make decisions among five 1. Did subjects perceive differences between our different
levels of information quality, i.e., did our manipu-
hypothetical job alternatives where each job alternative
lation work?
was individually constructed to insure specific levels of
quantity and quality and to control for task ease. During 2. How reliablewas our measureof informationquality?
each choice task, subjects provided the following mea- 3. Were any errors in measurement related to the levels
sures in the order indicated: of information quality or quantity?
1. Choiceof preferredalternativefor fivedifferentpaired
comparisonsselected from the five alternatives Manipulation Checks
2. Choice of most preferredalternativeamong the five We used three different measures to assess the veracity
alternatives of our,quality measure: subjects' perception of the use-
3. Confidencein the most preferredchoice (measured fulness of the information provided (1 = not at all use-
by seven-pointLikert-typescales assessing(a) choice ful, . . . , 7 = very useful), the amount of additional

This content downloaded on Thu, 14 Feb 2013 10:45:33 AM

All use subject to JSTOR Terms and Conditions
INFORMATIONAND DECISION EFFECTIVENESS 207

information desired (1 = none, . . . , 7 = a lot), and measures, however, was adjusted for ties. Thus, in gen-
the number of specific attributes subjects listed that were eral the upper bound for the correlation coefficient was
not provided but were desired. The first provides a direct not one, nor was the lower bound for average absolute
measure of subjects' perceptions of the quality of the differences in rank equal to zero. Moreover, no allow-
informational environment. The latter two provide in- ance could be made for the fact that the Phase I question
direct measures. concerned ratings while Phase III concerned rankings.
There were either 10 or 5 observations per subject Therefore, our reliability measures should be viewed as
depending on whether they took part in both Phase II the ceiling of the change over time of subjects' rankings
and III. Given these multiple observations, a repeated of attribute importance.
measures main effects model was run for each of the Given this instability, the question becomes, are these
three above-mentioned dependent variables with the measures of instability associated with any of our in-
following independent variables: quantity (measured at dependent variables (particularly our quality and
four levels), quality (measured at four levels), subject, quantity measures)? Two separate analyses were con-
group (high or low quality/attribute) format (matrix or ducted to address this question. First, for the 43 Phase
paragraph), order (1st, 2nd, . . . 5th), and task ease as III subjects the average rank of the attributes used in
defined in Equation 5. In each of the three multivariate constructing the job alternatives for the choice task in
analyses, quality was significant at the 0.001 level after each cell was calculated assuming both Phase I and
adjusting for all other factors. Moreover, the adjusted Phase III rankings. Then, assuming the importance
means of the four levels of quality indicated an almost weights were distributed as displayed in Table 1, cell
linear relationship with lower quality levels associated means for the quality levels were calculated for both
with (1) a greater desire for more information, and (2) the Phase I and Phase III data by combining the average
a perception that the information provided was less rank orders with the average importance weights. Al-
useful. though there was more variation in the information
Because our design did not fully cross TQ and N, it quality means within a quality level based on Phase III
was impossible to check for possible interaction effects data as compared to Phase I data, there never was an
of TQ and N on the dependent variables over the total overlap for the four quality levels used. Thus, instability
range of TQ and N. However, we could test for an in- in the importance weights over time did not impede us
teraction between TQ and N using the data from cells from designing job tasks with significantly different in-
3 through 6 and 6 through 9 (see Table 2) since both formational quality levels.
of these groups of cells represented a fully crossed de- A second analysis investigated whether the magnitude
sign. In both groups of cells, the interaction effect was of the error in measuring utility, E3i (i.e., from Equation
nonsignificant for all three dependent variables, indi- 4), was associated with N. This is important since any
cating that the effects of TQ and N were noninteractive positive relationship between Var(E3i) and N would
with each other, at least within the range tested. cause our results to overestimate the effects of N on
decision effectiveness and consequently would cause us
Reliability of Importance Weights to overestimate the effects of information overload. The
Over Time analysis was conducted by first regressing our measure
of information quality (i.e., IWk) against the three ma-
One possible concern about our design is the time nipulation check variables. Then the squared residuals
lag between the Phase I measurement of the importance were calculated for each of the nine different quality-
weights and the Phase II and III job choice tasks. Al- quantity cells and averaged over subjects. For the 15
though such a lag helps insure these Phases are not con- possible cell comparisons where TQ is fixed and N in-
nected in the minds of the subjects, it also increases the creases (five comparisons for each of the three manip-
possibility that the subjects revised their importance ulation check variables), only two showed an increase
weights over this time period, introducing error into in the squared error, holding TQ fixed, for increases in
our quality manipulation and our estimation of u*, (i.e., N. Thus, there does not seem to be a positive relation-
increasing the variance of E3i). To assess the degree to ship between the Var(E3i)and N (holding TQ fixed).2
which subjects adjusted their importance weights, Phase In summary, we interpret the above analyses as strong
III subjects were asked, after completing the basic ex- evidence that our manipulation of information quality
periment, to rank all original 17 attributes in terms of is highly related to the subjects' perception of usefulness
their importance. These rankings were then compared and desire for more information and that there is little
to rankings derived from the Phase I importance ratings.
Reliability (stability) was measured two ways for each
subject: rank order correlations (Mansfield 1980) and 2A third and more complex analysis was performed to assess re-
absolute differences in ranks. Both measures are lower versals in rankings from Phase I and Phase III. Proportionally more
reversals were found among attributes with large weights than small
than most reported reliability figures. The average rank weights (holding fixed the initial difference in weight), yielding results
correlation across subjects was 0.57 and the average compatible with the premise that the Var(E3E)increases with ur, not
absolute difference in rank was 3.6. Neither of these N. The details of this analysis are available from the authors.

This content downloaded on Thu, 14 Feb 2013 10:45:33 AM

All use subject to JSTOR Terms and Conditions
208 THE JOURNAL OF CONSUMER RESEARCH

TABLE 3
DECISION EFFECTIVENESS AND CONFIDENCE AS A FUNCTION OF INFORMATION QUANTITY AND QUALITY, TASK EASE,
AND COVARIATES: COEFFICIENT ESTIMATES (P VALUES)

BEST PAIR CONF

Indepent variables df Coefficient p value Coefficient p value Coefficient p value

SUBJ 62 * (p<.19) * (p <.03) * (p <.0001)

FORMATa 1 -.0846 (p <.11) -.022 (p <.21) -1.71 (p <.0005)
GROUPb 1 .169 (p < .43) .027 (p < .70) 5.17 (p < .009)
ORDER (p < .96) (p < .07) (p < .0001)
First-fifth 1 .004 (p < .95) .037 (p < .20) 2.67 (p < .0001)
Second-fifth 1 .034 (p < .59) .043 (p < .05) 1.61 (p < .008)
Third-fifth 1 .019 (p < .77) -.005 (p < .81) .27 (p < .65)
Fourth-fifth 1 .041 (p < .53) .039 (p < .07) 1.70 (p < .005)
TE 1 .230 X 10-3 (p <.0001) .102 X 10-3 (p <.0001) 1.2 X 10-4 (p <.012)
Log N 1 -.361 (p <.015) -.181 (p<.0007) -5.13 (p <.0004)
LOG TQ 1 .74 (p <.04) .411 (p <.002) 3.38 (p <.195)
(TQ)2 1 -.84 X 10-4 (p < .10) -.731 X 10-4 (p <.0007) .8 X 10-3 (p <.091)
F(72,397) 1.39 (p < .03) 4.02 (p <.0001)
F(72,2254) 4.33 (p <.0001)
R2 .20 .12 .42

0 = matrix, 1 = paragraph.
bO = high average quality, 1 = low average quality.
NOTE: * indicates that individualsubject coefficients are not reported.

or no interaction effect between TQ and N on these attributeinformation (GROUP), and format of infor-
manipulationchecks. Also, although our measureof a mation (FORMAT). (A similar LOGIT analysis of
subject'simportanceweightsmay have been unreliable BEST was also conducted,yieldingresultsalmost iden-
(or unstable)over time, the effect of this instability on tical to those from OLS, and thus is not reported.)The
our ability to establishdistinct quality levels was min- quantity variable,N, was coded as 23, 47, 59, or 71 to
imal. In addition,the magnitudeof the errorin assessing reflect the percent of available attribute information.
importanceweightsappearsto be unrelatedto N holding The total quality variable, TQ, was coded 33, 46, 60,
TQ fixed. Thus, the instabilityin attributeimportance and 73 to reflectthe percent of the averagecumulative
weights across phases does not invalidate any finding importancescores made available.
of statisticalsignificance,exceptformatperhaps,in that The results from the two OLS covariance analyses,
the only possible effect of instabilityis to decreasethe shown in Table 3, indicate the significanceof the coef-
statisticalsignificanceof any true relationshipthrough ficients in the reduced form equations. For BEST we
the introduction of random error. Moreover, we ac- had five observationsper subject.For PAIR we had 25
count for individual specific variance attributableto observationsper person(i.e., five pairedchoices for five
factorssuch as importanceweightinstabilitythrougha differentchoice tasks).Using the estimatesof b2, (b2a1),
repeated measure analysis that covaries out any sys- and (b2a2), it is possible from Equation 9 to derive the
tematic subject-specificeffectsof instability on the de- structuralmodel coefficientestimatesof interest,a, and
pendent measuresof interest. a2, and to sign the coefficientfor the log(N) term. The
estimates of a, are 2.05 and 2.27 for the BEST and
PAIR equations, respectively, and for a2 are -2.33
RESULTS X 10-4 and -4.04 X 10-4. In both cases the coefficient
for log(N) was negative.
To test the Proposition,we need estimatesof the pa- The above estimates are compatible with the Prop-
rametersof the structuralequationas statedin Equation osition. The negative coefficient for log(N) indicates
6. As shown earlier, an estimate can be obtained by that increasesin information quantity, holding infor-
relatingy*' to some observed measure such as BEST mation quality fixed, lead to decreasesin decision ef-
and then derivinga reducedform equation (e.g., Equa- fectiveness.Similarly,both estimatesof a, areconsistent
tion 8). Two such measuresused wereBEST and PAIR, with the premisethat increasesin informationquality,
the lattermeasuringthe "correct"selection from a pair holdinginformationquantityfixed, initially lead to in-
of alternatives.We now presentthe resultsof these two creasesin decision effectiveness.When the estimatesof
analyses. a2aretakeninto consideration,the resultsindicatethat
The two reducedformequationswereestimatedusing decision effectivenessincreasesat a decreasingrate and
OLS after controlling for repeated measures (SUBJ), eventually decreases with increases in total quality,
order of presentation (ORDER), high or low average holdingquantityfixed. This is equivalentto sayingthat

This content downloaded on Thu, 14 Feb 2013 10:45:33 AM

All use subject to JSTOR Terms and Conditions
 INFORMATIONAND DECISION EFFECTIVENESS 209

when the average quality level is increased (holding the in our study in situations where the average quality per
number of attributes available fixed), the model results attribute available is held constant.
imply an eventual decrease in decision effectiveness. At Finally, Table 3 also provides estimates of the influ-
first blush, such a finding may appear counterintuitive. ence of task ease and the covariates on the two observ-
However, the reader should note that such a finding is able choice measures. The results indicated that in-
a direct consequence of the implications derived from creases in task ease led to statistically significant in-
the Proposition, Equations 2 and 3, and the functional creases in BEST and PAIR. Only two of the eight
forms assumed. Thus, by assuming the functional forms possible control variables were found to affect actual
found in Equation 6 (thereby insuring the conditions choice; both of these were found in the PAIR reduced
as specified in footnote 1), it is easy to see that form equation. They indicated that the individual sub-
g'2(AQ.N) must eventually decrease in AQ (holding N ject and order of presentation affected the probability
fixed) if information overload exists, i.e., g'2(AQ N) that the alternative deemed best was chosen.
< g'1(N) for N > N*. Said differently, decreases in de- We previously noted that SUBJ captured the degree
cision effectiveness with increases in average quality of instability of the subject's importance weights. When
(holding N fixed) go hand in hand with decreases in SUBJ was replaced with a stability measure such as the
decision effectiveness with increases in N holding AQ rank order correlation of importance weights from
fixed. Phases I and III, the coefficients on quantity and quality
Although the above discussion is mathematically were only minimally affected (changes occurring in the
precise, it provides little insight into the behavioral rea- second or third significant digit), again indicating that
sons for such a finding. We postpone such an investi- importance weight instability was unrelated to our
gation until we look at the process measure results. It quality and quantity manipulation. Finally, although
should be noted, however, that this nonlinear effect of signs were consistent with our priors, there was little
TQ is not due to a perceived variation in TQ dependent statistical proof that subjects were more likely to choose
on the level of N because, as stated previously, the sub- the alternative deemed best when given the matrix as
jects' perceptions of the manipulation of TQ and N in- opposed to paragraph format.
dicated no interaction between these two variables. We also conducted an analysis of covariance using
We next determine the point where adding attribute as the dependent variable the subject's perceived con-
information of identical quality to the average quality fidence (CONF) in making the correct choice. The four
of the attributes already available begins to degrade de- items that were summed to measure confidence in the
cision effectiveness. To do this we substitute AQ * N for most preferred choice (choice statisfaction, confusion,
TQ in the reduced form equations for BEST and PAIR certainty, and likelihood of regret) had a coefficient al-
and then use the coefficient estimates found in Table 3 pha (Cronbach 195 1) of 0.87, exceeding the research
to derive the maximum as a function of N. This max- standard suggested by Nunnally (1978). Although we
imum is derived by differentiating the reduced form initially thought that such a variable also could be used
equations with respect to N and solving, i.e., to identify the structural equation parameters, we soon
came to believe that subjects' perceptions of confidence
*(*)/N= -b2/N+ (b2a )/N-2(b2a2)AQ2N= O, (11) also are directly affected by the informational environ-
ment, i.e., quality and quantity. This implies that either
or N or TQ (or both) directly enter into the reduced form
-1 +a, -2a2AQ2N2 = 0. (12) as well as enter through y*', making it impossible to
identify the structural effects a, and a2 without knowing
Solving for the maximum N yields N* = ((a, - 1)/ a priori the magnitude of this direct effect. Therefore,
2a2AQ2)"12, which can be solved once a value for AQ is in presenting the analysis of covariance results for this
specified. In our scaling, TQ increases almost identically dependent measure we restrict our discussion to reduced
to N (i.e., when N = 23, TQ = 33, when N = 47, TQ form implications, i.e., the aggregate effect of Nand TQ
= 46, when N = 59, TQ = 60, and when N = 71, TQ on the subjects' confidence in making a correct choice
= 73). Thus since AQ = TQ/N, we let AQ = 1 for our without specifying how these effects occur.
situation. The results of this analysis are shown in the last col-
The implication of this analysis is that the maximum umn of Table 3. Four things are evident. First, increases
for BEST and PAIR occurred when N = 40 and 47 in the quantity of information, holding total quality
percent, respectively. Said slightly differently, using our fixed, decrease the subjects' perceived confidence. Sec-
model and the estimated coefficients, we can infer that ond, unlike the decision effectiveness results, increases
when average quality levels are held fixed, maximum in total quality holding quantity fixed increase the sub-
decision effectiveness is reached in our information en- jects' confidence over the total range of quality values.
vironment when the number of attributes available is (The linear term is highly significant if the quadratic
between 6.8 (=.40 X 17) and 8 (=.47 X 17) attributes. term is dropped from the analysis.) Moreover, by con-
In this way, we have roughly determined the point where ducting an analysis similar to that outlined in Equations
information overload occurs for the 'average subject' 1 1 and 12, we note that subjects showed increased con-

This content downloaded on Thu, 14 Feb 2013 10:45:33 AM

All use subject to JSTOR Terms and Conditions
210 THE JOURNAL OF CONSUMER RESEARCH

fidence over the total range of quantity (allowing total mation. However, individuals accustomed to receiving
quality to vary). Thus, consumers' expressed confidence higher levels of quality information (i.e., Group 1) use
does not reflect the fact that increases in the average less information than individuals who previously re-
quality level, holding fixed the quantity of information, ceived lower quality information (i.e., Group 2) when
eventually lead to decreases in their decision accuracy; both groups receive information of equal quality (i.e.,
nor do consumers appear to perceive that increases in cell 6). No other variables were found to be statistically
attribute information eventually cause them confusion significant. We interpret these results to indicate that
even when the average quality level is held fixed. Third, individuals altered the amount of available information
the difficulty of the task was significantly related to con- that they used according to the perceived usefulness of
fidence; the easier our measure of the choice task, the the informational environment.
more confidence the subject reported. Finally, all co- We next investigated how the percentage of infor-
variates were statistically significant. Subjects expressed mation used affects decision effectiveness, as reflected
more confidence in their choice when they were given by the measures BEST and PAIR. This was done by
the matrix form, when for equivalent informational en- postulating a relationship where PUSED directly affects
vironments (e.g., identical values of N and TQ) they y*'. This relationship was then substituted into Equation
were members of group 2, and when they were making 7 for BEST and its equivalent for the measure, PAIR,
the earlier choices. The group effect is interesting be- along with two process-related dummy variables. The
cause it suggests that, given the same informational en- first dummy variable reflects whether subjects said they
vironment (i.e., cell 6), those subjects who had been tended to incorporate attributes for which no infor-
previously exposed to lower average information qual- mation was given into their decision process (INFER
ity, all else being equal, were more confident when they = 1) or not (INFER = 0) during the course of the choice
made their choice in cell 6 than were those previously tasks. As such, it is a general measure of the subject's
exposed to the higher average quality information. Such tendency to infer and can be thought of in terms of an
contextual effects will also be shown to be associated individual difference variable. The second variable,
with how subjects actually process the available attribute STRA T, indicated whether subjects reported using a
information. compensatory (STRA T = 0) or noncompensatory
(STRA T = 1) decision strategy in choosing their most
PROCESS EXPLANATION preferred alternative. Table 4 contains the results of the
appropriate multiple regressions.
OF MANIPULATION EFFECTS
Overall, the reported percentage of information used
All of the results discussed are compatible with the had a statistically significant impact on choice accuracy
Proposition, common sense (e.g., easier tasks result in as measured by BEST and PAIR. As Table 4 shows, an
greater decision effectiveness), or past empirical findings inverted U relationship was found with respect to
(e.g., confidence is monotonic in N). Although the find- PUSED for BEST and PAIR, solving the quadratic
ing that decision effectiveness eventually decreased with yields maxima when subjects used, respectively, 68 and
increases in average quality, holding quantity fixed, can 75 percent of the available attribute information. Thus,
be derived from the fact that information overload was it appears that subjects made more effective decisions
observed, we next look more closely at what might have when they were somewhat selective, using most but not
caused such a result and, in the process, gain some in- all the available attribute information.
sights as to how people went about using the available To check the adequacy of our process model, we con-
information to reach their final choices. We do this by ducted analyses identical to those shown in Table 4 but
analyzing the fallible, but potentially insightful, retro- that also included N and AQ directly. In both cases,
spective process measures reported by subjects after they these variables were not statistically significant; nor was
completed their fifth and final choice set task. the maximum altered more than one or two percent.
To explore how subjects went about their choice We take this as evidence that our process model captures
tasks, we constructed a process measure, PUSED, the the time path of events, i.e., N and AQ do not directly
percent of available information used by the subject in affect decision effectiveness, but only the mediating
making his/her choice. This variable was obtained by variable, PUSED.
dividing the number of available attributes used as de- We find these reported process results satisfying in
termined by the question, "List the attributes used in that they help explain why decision effectiveness even-
making your choice," by the number of attributes tually decreases with increases in average quality. The
available for that choice situation. first process model analysis indicated that as the average
We then conducted a multiple regression analysis quality level of available information is increased people
with N, A Q, TE, GROUP and FORMA T as independent tend to report using more of the available information.
variables and PUSED as the dependent measure. The However, the next analyses indicate that decision ef-
results indicate that when the informational environ- fectiveness first increases and then decreases as the per-
ment is changed by increasing the average quality level, centage of information used increases. Thus, increases
individuals increase their use of the available infor- in average quality motivate individuals to incorporate

This content downloaded on Thu, 14 Feb 2013 10:45:33 AM

All use subject to JSTOR Terms and Conditions
INFORMATIONAND DECISION EFFECTIVENESS 211

TABLE 4 2. Our analysis pertains to situations where all attribute

information is presented to the decision maker; thus
DECISION EFFECTIVENESS AS A FUNCTION OF REPORTED
AMOUNT OF INFORMATION USED AND DECISION there are no relevant external search costs. Moreover,
STRATEGY: REGRESSION COEFFICIENT the "cost" of processing information is not explicitly
ESTIMATES (P VALUES) included in our model. We do attribute the observed
decrease in decision effectiveness, however, to the ef-
Independent fects of these cognitive costs.
variables BEST PAIR
3. Our model of utility formation assumes uncorrelated
attributes with no attribute interactions. Attribute
Intercept -.56 (p < .14) .20 (p < .22)
values are treated with certainty. In particular, we use
TE .00019 (p < .04) .00010 (p < .0001) a linear additive utility model, although we acknowl-
PUSED 6.68 (p < .0002) 1.34 (p < .003) edge that other compensatory models might also be
PUSED2 -4.97 (p < .0001) -.88 (p < .008) appropriate.
STRAT .043 (p < .65) .03 (p < .231)
INFER -.21 (p < .10) -.02 (p < .34) 4. The experimental design by which we test our prop-
F(6,85) 3.87 (p < .001) ositions is nonorthogonal. Although we were able to
F(6,452) 10.24 (p < .0001) test for interactions within a limited range using both
R2 .24 .10 the manipulation variables and the observed choice
variables as dependent measures, the restricted range
of the independent variables employed provides a
weaker test of the nonexistence of quantity-by-quality
more of the available information and lead to "infor- interactions than a complete factorial design. Specif-
mation overload." Said differently, just as we observed ically, we caution against generalizing the results be-
that increases in the available information, holding fixed yond our experimental domain.
the average quality, eventually lead to poorer decision 5. The model formulated explicitly stated that the cor-
effectiveness, we also note that holding fixed the number rectness of the observed choice was a function of the
of attributes available but increasing the average quality unobserved variable, Var(E3E), i.e., the variance as-
of the attribute information causes consumers to use sociated with estimating a person's true utility. Nu-
more of the available information, and thus eventually merous factors could affect the size of this variable.
results in poorer decision accuracy. Since we could not directly measure this variable, it
was necessary for us to assume that this variable was
not associated with N, TQ, and/or TE. Although we
SUMMARY AND DISCUSSION took considerable effort to ascertain no relationship
We developed a model that rests on the premise that (especially with N), it was not possible to guarantee
that no relationship existed. If the error increases with
consumers vary how they assess the value of alternatives
N, then our estimate of the effects of quantity and
in a choice set depending on the conditions in the in- information overload are biased.
formational environment. This model is then used to
define the concept of decision effectiveness (i.e., a con- With these cautions in mind, we conclude the following:
sumer's ability to accurately assess the utility (value) of 1. Consumer's perceptions of the usefulness of an in-
available alternatives assuming assessment uses an ad- formational environrnent are strongly associated with
ditive compensatory rule). We also postulated that in- our measure of information quality, TQ, the consum-
formation has two components, information quantity er's cumulative or total importance weights for the
and information quality, which differentially affect de- attributes made available.
cision effectiveness, and that in any given setting these 2. Decision effectiveness is adversely affected by in-
two components are usually empirically related. creases in the quantity of information made available
A choice task experiment was conducted and our (holding quality fixed) and fostered by increases in
model was used to test our premises. Before summa- the quality level, at least up to a point (holding quan-
rizing our main conclusions, we note five limitations tity levels fixed). In this way we have shown that in-
or cautions to the model formulation and/or experi- formation has two different components and that each
mental findings of this research. component has a different effect.
3. Decision effectiveness first increases and then de-
1. The settingto whichwe applyour model is one where creases as the amount of information available in-
motivatedconsumersare interestedin examiningall creases, holding fixed the average quality level of the
attributeinformation,suggestingthat decision effec- informational environment.
tiveness is an appropriatestandardof comparison.
Thus, our results are more relevantto high-involve- 4. Consumers' confidence in how good a choice they
ment consumer products than, say, frequentlypur- made is affected directly not only by the unobserved
chased consumer nondurables. Also, it should be construct decision effectiveness, but also by a number
noted that subjectswere "role-playing"in that they of contextual factors such as the average quality level
did not actuallyget to consumetheirchoices. Finally, of previous information provided, and the quantity,
our subjectswere young and highly educated. quality, and format of the present information pro-

This content downloaded on Thu, 14 Feb 2013 10:45:33 AM

All use subject to JSTOR Terms and Conditions
212 THE JOURNAL OF CONSUMER RESEARCH

vided.Increasesin quality,holdingquantityfixed,led being overloaded when "too much" information is

consumersto be moreconfidentin theirdecisionover made readilyavailableto the decision maker;too much
the total relevantrangeof the informationalenviron- being relatedto the quantity and averagequality level
ment used in our task. Similarly, increases in the of the availableinformation.This statementonce again
quantity, holding fixed the quality, resulted in de- points out the necessity of making explicit the two
creasesin confidence.Increasesin quantity, holding
fixed average quality, however, never caused con- components of informationand their respectiveinflu-
sumersto feel lessconfident,a findingcompatiblewith ences on consumerdecision making.The public policy
that reportedby Jacobyand his associates. implications of this study are clear: greaterattention
must be devoted to the consumerperceptionof quality
5. The informationalenvironment affects the manner of information provided as well as the quantity of in-
in which subjects process the available information
which, in turn, affects their decision effectiveness.
formationsince the qualityof the availableinformation
Higher levels of information quality are associated also affectsusageand ultimatelydecision effectiveness.
with higherrelativeamountsof attributeinformation Given the importanceof informationqualityin con-
used, holdingquantityof availableinformationfixed. sumers' ability to make effective decisions, future re-
Relatingusage of attributeinformationto choice ac- search,in additionto recognizingthe two components,
curacy,it appearedthat subjectswere more likely to could profitablybe directedat betterunderstandinghow
reachthe correctchoice when they used most but not each component of information affects the decision-
all of the availableinformation. making process, especially the consumers' decision
strategies.Analyses of the effects of the informational
We draw two main conclusions from these results. environmenton our measuresof the decision strategies
First, information has at least two dimensions, quantity providedsome insightsbut werehinderedby the crude-
and quality, and these two dimensions have opposing ness of our measures. Subsequent researchcould in-
influences on decision effectiveness. Second, points 2, creasethe insights of the effectsof informationquality
3, and 5 in conjunction with our model are compatible on decisionmakingby developingand/or implementing
with "information overload" in that both the presence richer strategy measures. One approach would be to
of too much available information and too much high- build upon the work of Payne (1976) and Olshavsky
quality information appears to have caused consumers (1979) who looked at the effectsof informationquantity
to exhibit a decrease in decision effectiveness. In this on decision strategy.
way our results are similar to Malhotra (1982) in that Greater conceptual development of information
both studies demonstrate that large amounts of attribute quality could also be useful. Our operationalizationof
information degrade choice accuracy, as measured by informationquality as the cumulativesum of attribute
an ideal additive compensatory decision model. Our importanceweights,althoughintuitivelyappealingand
research, however, is more specific in that by assuming theoretically a good starting point, could be embel-
our model of decision effectiveness we were able to em- lished. Such work might attemptto uncover consumer
pirically show the functional relationship between in- attitudestowardsinformationquality.Uncertaintywith
formation quantity and decision effectiveness to be an respectto attributevalues,the perceivedvarianceacross
inverted U. As shown in Equation 12, the optimal brandson an attribute'svalues,and informationformat
number of attributes to be made available is a decreasing are probablyother areas that deserve greaterresearch
function of average quality. Assuming the average emphasisin terms of informationquality. More atten-
quality level used in our choice task, we also found the tion also needs to be paid to how inferencing affects
optimal number of attributes to make available to be decision processesand outcomes.
in the neighborhood of seven. Our research also extends Anotherline of researchmight explore other means
previous thinking by providing a detailed conceptual of defining choice accuracy or quality. Possible ap-
model to guide estimation of information overload ef- proachesmightinvolve specifyingan objectivefunction
fects. This model emphasizes the importance of the for subjectsto follow or having subjectsdescribetheir
construct, decision effectiveness, as well as the concept, "ideal"decision strategy,which could be later used to
information quality, and permits the more specific model the accuracyof their decisions.
findings reported above. Finally, given the possible bias estimate problemas-
Malhotra (1984) and Jacoby (1984) disagree on the sociated with not observinga measurementerror, 63i,
importance of the information overload paradigm, with futureworkcould concentrateon minimizingthe mag-
Jacoby criticizing use of the paradigm due to his belief nitude of this error. One possible approachwould be
that the key issue is not can consumers be overloaded to reducethe time between measuringthe importance
but will consumers be overloaded. In adopting this po- weight and presenting the subjects with the choice
sition, Jacoby maintains that consumers are selective scenarios.Anotherapproachwould take multiple mea-
in the amount and nature of the information that they sures on the weight, thereby providing a more direct
obtain, suggesting that factors that impact this selectiv- assessmentof the measurementerror.
ity should be examined. The results of this study imply
that consumers are not able to shield themselves from [Received July 1985. Revised February 1987.]

This content downloaded on Thu, 14 Feb 2013 10:45:33 AM

All use subject to JSTOR Terms and Conditions
 INFORMATIONAND DECISION EFFECTIVENESS 213

REFERENCES Malhotra, Naresh K. (1982), "Information Load and Con-

sumer Decision Making," Journal of Consumer Re-
search, 8 (March), 419-430.
Best, Roger J. and Michael Ursic (1986), "The Impact of In-
(1984), "Reflections on the Information Overload
formation Load and Variability on Choice Accuracy,"
Paradigm in Consumer Decision Making," Journal of
in AMA Proceedings, Vol. 52, 1-3.
Consumer Research, 10 (March), 436-440.
Cronbach, Lee J. (1951), "Coefficient Alpha and the Internal
Structure of Tests," Psychometrika, 16 (September), , Arun K. Jain, and Stephen W. Lagakos (1982), "The
Information Overload Controversy: An Alternative
297-334.
Viewpoint," Journal of Marketing, 46 (Spring), 27-37.
Green, Paul E. and V. Srinivasan (1978), "Conjoint Analysis
Mansfield, Edwin S. (1980), Statistics for Business and Eco-
in Consumer Research: Issues and Outlook," Journal of
nomics. Methods and Applications, New York: W.W.
Consumer Research, 5 (September), 103-23.
Norton.
Hagerty, Michael R. and David A. Aaker (1984), "A Nor-
Miller, George A. (1956), "The Magical Number Seven, Plus
mative Model of Consumer Information Processing,"
or Minus Two: Some Limits of our Capacity for Pro-
Marketing Science, 3 (Summer), 227-246.
cessing Information," Psychological Review, 63, 81-97.
Hoepfl, Robert T. and George D. Huber (1970), "A Study of
Muller, Thomas E. (1984), "Buyer Response to Variations
Self-Explicated Models," Behavioral Science, 15 (Sep-
in Product Information Load," Journal of Applied Psy-
tember), 708-714.
chology, 69 (May), 300-306.
Huber, George P. (1974), "Multiattribute Utility Models: A
Nunnally, Jum C. (1978), Psychometric Theory, New York:
Review of Field and Field-like Studies," Management
McGraw-Hill.
Science, 20 (June), 1393-1402.
Olshavsky, Richard W. (1979), "Task Complexity and Con-
, Rahman Daneshgar, and David L. Ford (1971), "An
tingent Processing in Decision Making: A Replication
Empirical Comparison of Five Utility Models for Pre-
and Extension," Organizational Behavior and Human
dicting Job Preferences," Organizational Behavior and
Performance, 24 (3), 300-316.
Human Performance, 6 (May), 267-282.
Payne, John W. (1976), "Task Complexity and Contingent
Jacoby, Jacob (1977), "Information Load and Decision
Processing in Decision Making: An Information Search
Quality: Some Contested Issues," Journal of Marketing
and Protocol Analysis," Organizational Behavior and
Research, 14 (November), 569-573.
Human Performance, 16 (August), 366-387.
(1984), "Perspectives on Information Overload,"
Russo, J. Edward (1974), "More Information is Better: A Re-
Journal of Consumer Research, 10 (March), 432-435.
evaluation of Jacoby, Speller, and Kohn," Journal of
, Donald E. Speller, and Carol Kohn Berning (1974a),
Consumer Research, 1 (December), 68-72.
"Brand Choice Behavior as a Function of Information
Scammon, Debra L. (1977), "Information Load and Con-
Load: Replication and Extension," Journal of Consumer
sumers," Journal of Consumer Research, 4 (December),
Research, 1 (June), 33-42.
148-155.
, Donald E. Speller, and Carol A. Kohn (1974b),
Shugan, Steve (1980), "The Cost of Thinking," Journal of
"Brand Choice Behavior as a Function of Information
Consumer Research, 7 (September), 99- 111.
Load," Journal of Marketing Research, 11 (February),
Simon, Herbert A. (1969), The Sciences of the Artificial,
63-69.
, Donald E. Speller, and Carol A. K. Berning (1975), Cambridge, MA: M.I.T. Press.
Staelin, Richard and John W. Payne (1976), "Studies of the
"Constructive Criticism and Programmatic Research:
Information-Seeking Behavior of Consumers," in Cog-
Reply to Russo," Journal of Consumer Research, 2 (Sep-
nition and Social Behavior, eds. John S. Carroll and John
tember), 154-156.
W. Payne, Hillsdale, NJ: Lawrence Erlbaum, 185-201.
Johnson, Eric J. and John W. Payne (1985), "Effort and Ac- Streufert, Sigfried and Michael J. Driver (1965), "Conceptual
curacy in Choice," Management Science, 31 (April), 395- Structure, Information Load, and Perceptual Complex-
414. ity," Psychonomic Science, 3, 249-250.
Jurgensen, Clifford E. (1978), "Job Preferences (What Makes Summers, John 0. (1974), "Analysis of Effects of Information
a Job Good or Bad?)," Journal of Applied Psychology, Load," Journal of Marketing Research, 11 (November),
63 (June), 267-276. 467-468.
Kelly, Everett L. and Donald W. Fiske (1951), The Predictions Wilkie, William L. (1974), "Analysis of Effects of Information
of Performance in Clinical Psychology, Ann Arbor: Uni- Load," Journal of Marketing Research, 11 (November),
versity of Michigan Press. 462-466.

This content downloaded on Thu, 14 Feb 2013 10:45:33 AM

All use subject to JSTOR Terms and Conditions
View publication stats