Job Market Signaling
Michael Spence
The Quarterly Journal of Economics, Vol. 87, No. 3. (Aug., 1973), pp. 355-374.
Stable URL:
http://links.jstor.org/sici?sici=0033-5533%28197308%2987%3A3%3C355%3AJMS%3E2.0.CO%3B2-3
The Quarterly Journal of Economics is currently published by The MIT Press.
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/about/terms.html. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained
prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in
the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/journals/mitpress.html.
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
The JSTOR Archive is a trusted digital repository providing for long-term preservation and access to leading academic
journals and scholarly literature from around the world. The Archive is supported by libraries, scholarly societies, publishers,
and foundations. It is an initiative of JSTOR, a not-for-profit organization with a mission to help the scholarly community take
advantage of advances in technology. For more information regarding JSTOR, please contact support@jstor.org.
http://www.jstor.org
Thu Apr 3 07:17:58 2008
JOB MARKET SIGNALING * 

1. Introduction, 355. -2. Hiring as investment under uncertainty, 356. -
3. Applicant signaling, 358. -4. Informational feedback and the definition of
equilibrium, 359. -5. Properties of informational equilibria: an example, 361.
- 6. The informational impact of indices, 368. -Conclusions, 374.
The term "market signaling" is not exactly a part of the welldefined,
technical vocabulary of the economist. As a part of the
preamble, therefore, I feel I owe the reader a word of explanation
about the title. I find it difficult, however, to give a coherent and
comprehensive explanation of the meaning of the term abstracted
from the contents of the essay. In fact, it is part of my purpose to
outline a model in which signaling is implicitly defined and to explain
why one can, and perhaps should, be interested in it. One
might accurately characterize my problem as a signaling one, and
that of the reader, who is faced with an investment decision under
uncertainty, as that of interpreting signals.
How the reader interprets my report of the content of this essay
will depend upon his expectation concerning my stay in the market.
If one believes I will be in the essay market repeatedly, then both
the reader and I will contemplate the possibility that I might invest
in my future ability to communicate by accurately reporting the
content of this essay now. On the other hand, if I am to be in the
market only once, or relatively infrequently, then the above-mentioned
possibility deserves a low probability. This essay is about
markets in which signaling takes place and in which the primary
signalers are relatively numerous and in the market sufficiently infrequently
that they are not expected to (and therefore do not) invest
in acquiring signaling reputations.
*The essay is based on the author's doctoral dissertation ("Market
Signalling: The Informational Structure of Job Markets and Related Phenomcna,"
Ph.D. thesis, Harvard University, 1972), forthcoming as a book entitled
Market Signaling: Information Transfer in Hiring and Related Screening
Processes in the Harvard Economic Studies Series, Harvard University Press.
The aim here is to present the outline of the signaling model and some of its
conclusions. Generalizations of the numerical examples used for expositional
purposes here are found in ibid. and elsewhere.
I owe many people thanks for help in the course of the current study,
too many to mention all. However, I should acknowledge explicitly the magnitude
of my debts to Kenneth Arrow and Thomas Schelling for persistently
directing my attention to new and interesting problems.
356 QUARTERLY JOURNAL OF ECONOMICS
I shall argue that the paradigm case of the market with this
type of informational structure is the job market and will therefore
focus upon it. By the end I hope it will be clear (although space
limitations will not permit an extended argument) that a considerable
variety of market and quasi-market phenomena like admissions
procedures, promotion in organizations, loans and consumer credit,
can be usefully viewed through the conceptual lens applied to the
j oh market.
If the incentives for veracity in reporting anything by means
of a conventional signaling code are weak, then one must look for
other means by which information transfers take place. My aim is
to outline a conceptual apparatus within which the signaling power
of education, job experience, race, sex, and a host of other observable,
personal characteristics can be determined. The question, put
crudely, is what in the interactive structure of a market accounts
for the informational content, if any, of these potential signals. I
have placed primary emphasis upon (i) the definition and properties
of signaling equilibria, (ii) the interaction of potential signals, and
(iii) the allocative efficiency of the market.
In most job markets the employer is not sure of the productive
capabilities of an individual at the time he hires him.l Nor will
this information necessarily become available to the employer immediately
after hiring. The job may take time to learn. Often
specific training is required. And there may be a contract period
within which no recontracting is allowed. The fact that it takes
time to learn an individual's productive capabilities means that
hiring is an investment decision. The fact that these capabilities
are not known beforehand makes the decision one under uncertainty.
To hire someone, then, is frequently to purchase a 10ttery.~In
what follows, I shall assume the employer pays the certain monetary
equivalent of the lottery to the individual as wage.3 If he is
1. There are, of course, other informational gaps in the job market. Just
a;r employers have less than perfect information about applicants, so also will
applicants be imperfectly informed about the qualities of jobs and work environmenks.
And in a different vein neither potential employees nor employers
know all of the people in the market. The resulting activities are job search
and recruiting. For the purpose of this essay I concentrate upon employer
uncertainty and the signaling game that results.
2. The term "lottery" is used in the technical sense, imparted to it by
decision theory,
3. The certain monetary equivalent of a lottery is the amount the individual
would take, with certainty, in lieu of the lottery. I t is generally
thought to be less than the actuarial value of the lottery.
357JOB MARKET SIGNALING
risk-neutral, the wage is taken to be the individual's marginal contribution
to the hiring organization.
Primary interest attaches to how the employer perceives the
lottery, for it is these perceptions that determine the wages he offers
to pay. We have stipulated that the employer cannot directly observe
the marginal product prior to hiring. What he does observe
is a plethora of personal data in the form of observable characteristics
and attributes of the individual, and it is these that must
ultimately determine his assessment of the lottery he is buying.
(The image that the individual presents includes education, previous
work, race, sex, criminal and service records, and a host of other
data.) This essay is about the endogenous market process whereby
the employer requires (and the individual transmits) information
about the potential employee, which ultimately determines the
implicit lottery involved in hiring, the offered wages, and in the
end the allocation of jobs to people and people to jobs in the market.
At this point, it is useful to introduce a distinction, the import
of which will be clear shortly. Of those observable, personal attributes
that collectively constitute the image the job applicant
presents, some are immutably fixed, while others are alterable. For
example, education is something that the individual can invest in at
some cost in terms of time and money. On the other hand, race and
sex are not generally thought to be alterable. I shall refer to observable,
unalterable attributes as indices, reserving the term signals
for those observable characteristics attached to the individual that
are subject to maniplllation by him.4 Some attributes, like age, do
change, but not at the discretion of the individual. I n my terms,
these are indices.
Sometime after hiring an individual, the employer will learn
the individual's productive capabilities. On the basis of previous
experience in the market, the employer will have conditional probability
assessments over productive capacity given various combinations
of signals and indices. At any point of time when confronted
with an individual applicant with certain observable attributes, the
employer's subjective assessment of the lottery with which he is
confronted is defined by these conditional probability distributions
over productivity given the new data.
From one point of view, then, signals and indices are to be re-
4. The terminological distinction is borrowed from Robert Jervis (The
Logic of Images in International Relations (Princeton, N.J.: Princeton University
Press, 1970)). My use of the terms follows that of Jervis sufficiently
closely to warrant their transplantation.
358 QUARTERLY JOURNAL OF ECONOMICS
garded as parameters in shifting conditional probability distributions
that define an employer's beliefs."
For simplicity I shall speak as if the employer were risk-neutral.
For each set of signals and indices that the employer confronts, he
will have an expected marginal product for an individual who has
these observable attributes. This is taken to be the offered wage
to applicants with those characteristics. Potential employees therefore
confront an offered wage schedule whose arguments are signals
and indices.
There is not much that the applicant can do about indices.
Signals, on the other hand, are alterable and therefore potentially
subject to manipulation by the job applicant. Of course, there
may be costs of making these adjustments. Education, for example,
is costly. We refer to these costs as signaling costs. Notice that the
individual, in acquiring an education, need not think of himself as
signaling. He will invest in education if there is sufficient return
as defined by the offered wage s c h e d ~ l e . ~Individuals, then, are
assumed to select signals (for the most part, I shall talk in terms of
education) so as to maximize the difference between offered wages
and signaling costs. Signaling costs play a key role in this type of
signaling situation, for they functionally replace the less direct
costs and benefits associa.ted with a reputation for signaling reliability
acquired by those who are more prominent in their markets
than job seekers are in theirs.
A Critical Assumption
It is not difficult to see that a signal will not effectively distinguish
one applicant from another, unless the costs of signaling are
negatively correlated with productive capability. For if this condition
fails to hold, given the offered wage schedule, everyone will
invest in the signal in exactly the same way, so that they cannot be
distinguished on the basis of the signal. In what follows, we shall
make the assumption that signaling costs are negatively correlated
with productivity. It is, however, most appropriately viewed as a
5. The shifting of the distributions occurs when new market data are
received and conditional probabilities are revised or updated. Hiring in the
market is to be regarded as sampling, and revising conditional probabilities as
passing from prior to posterior. The whole process is a learning one.
6. There may be other returns to education. It may be a consumption
good or serve as a signal of things other than work potential (status for example).
These returns should be added to the offered wage schedule.
JOB MARKET SIGNALING 359
Employer's Gonditional
Probabilistic Beliefs
r 2
Offered Wage Schedule
as a Function of
A Signals and Indices
Hiring, Observation of 1 ISignaling Decisions by 

Relationship between
HApplicants; Maximization 

Marginal Product and of Return Net of Signal- 

Signals ing Costs
Signaling Costs
LFIGUREI
Informational Feedback in the Job Market
prerequisite for an observable, alterable characteristic to be a persistently
informative signal in the market. This means, among other
things, that a characteristic may be a signal with respect to some
types of jobs but not with respect to other^.^
Signaling costs are to be interpreted broadly to include psychic
and other costs, as well as the direct monetary ones. One element
of cost, for example, is time.
At this point it is perhaps clear that there is informational
feedback to the employer over time. As new market information
comes in to the employer through hiring and subsequent observation
of productive capabilities as they relate to signals, the employer's
conditional probabilistic beliefs are adjusted, and a new round
starts. The wage schedule facing the new entrants in the market
generally differs from that facing the previous group. The elements
in the feedback loop are shown in Figure I.
It is desirable to find a way to study this feedback loop in the
7. The reason is that signaling costs can be negatively correlated with
one type of productive capability but not with another.
360 QUARTERLY JOURNAL OF ECONOMICS
market over time. To avoid studying a system in a continual state
of flux, it is useful to look for nontransitory configuration of the
feedback system. The system will be stationary if the employer
starts out with conditional probabilistic beliefs that after one round
are not disconfirmed by the incoming data they generated. We shall
refer to such beliefs as self-confirming. The sense in which they are
self-confirming is defined by the feedback loop in Figure I.
A Signaling Equilibrium
As successive waves of new applicants conie into the market,
we can imagine repeated cycles around the loop. Employers' conditional
probabilistic beliefs are modified, offered wage schedules
Ere adjusted, applicant behavior with respect to signal choice
changes, and after hiring, new data become available to the employer.
Each cycle, then, generates the next one. In thinking about
it, one can interrupt the cycle at any point. An equilibrium is a set
of components in the cycle that regenerate themselves. Thus, we
can think of employer beliefs being self-confirming, or offered wage
schedules regenerating themselves, or applicant behavior reproducing
itself on the next round.8
I find it most useful to think in terms of the self-confirming
aspect of the employer beliefs because of the continuity provided
by the employer's persistent presence in the market.Q Thus, in these
terms an equilibrium can be thought of as a set of employer beliefs
that generate offered wage schedules, applicant signaling decisions,
hiring, and ultiinately new market data over time that are consistent
with the initial beliefs.
A further word about the definition of equilibrium is in order.
Given an offered wage schedule, one can think of the market as
generating, via individual optimizing decisions, an empirical distribution
of productive capabilites given observable attributes or
signals (and indices). On the other hand, the employer has subjectively
held conditional probabilistic beliefs with respect to productivity,
given signals. In an equilibrium the subjective distribution
and the one implicit in the market mechanism are identical,
8. In pursuing the properties of signaling equilibria, we select as the object
for regeneration whatever is analytically convenient, but usually employer
beliefs or offered wage schedules.
9. The mathematically oriented will realize that what is at issue here is
a fixed point property. A mapping from the space of conditional distributions
over productivity given signals into itself is defined by the market response
mechanism. An equilibrium can he thought of as a fixed point of this mapping.
A mathematical treatment of this subject is contained in Spence, op. cit.
361JOB MARKET SIGNALING
over the range of signals that the employer actually 0bserves.l Any
other subjective beliefs will eventually be disconfirmed in the market
because of the employer's persistent presence there.
Indices continue to be relevant. But since they are not a matter
of individual choice, they do not figure prominently in the feedback
system just described. I shall return to them later.
I propose to discuss the existence and properties of market signaling
equilibria via a specific numerical e ~ a m p l e . ~For the time being,
indices play no part. The properties of signaling equilibria that
we shall encounter in the example are general.3
Let us suppose that there are just two productively distinct
groups in a population facing one employer. Individuals in Group
I have a productivity of 1,while those in Group I1 have a productivity
of 2.4 Group I is a proportion ql of the population; Group I1
is a proportion of 1-ql. There is, in addition, a potential signal, say
education, which is available at a cost. We shall assume that education
is measured by an index y of level and achievement and is subject
to individual choice. Education costs are both monetary and
psychic. It is assumed that the cost to a member of Group I of y
units of education is y, while the cost to a member of Group I1 is
~ / 2 .
We summarize the underlying data of our numerical example in
Table I.
TABLE I
Marginal Proportion of Cost of educaGroup
product populat~on tion level L!
1. In a multi-market model one faces the possibility that certain types
of potential applicants will rationally select themselves out of certain job
markets, and hence certain signal configurations may never appear in these
markets. When this happens, the beliefs of the employers in the relevitnt
market are not disconfirmed in a degenerate way. No data are forthcoming.
This raises the possibility of persistent informationally based discrimination
against certain groups. The subject is pursued in detail in ibid.
2. Obviously, an example does not prove generality. On the other hand,
if the reader will take reasonable generality on faith, the example does illustrate
some essential properties of signaling equilibria.
3. See Spence, op. cit.
4. For productivity the reader may read "what the individual is worth
to the employer." There is no need to rely on marginal productivity here.
QUARTERLY JOURNAL OF ECONOMICS
Offered Wages as a Function of Level of Education
To find an equilibrium in the market, we guess at a set of selfconfirming
conditional probabilistic beliefs for the employer and
then determine whether they are in fact confirmed by the feedback
mechanisms described above. Suppose that the employer believes
that there is some level of education, say yHsuch that if y <yH,then
productivity is one with probability one, and that if y>yK, then
productivity will be two with probability one. If these are his conditional
beliefs, then his offered wage schedule, W(y), will be as
shown in Figure 11.
Given the offered wage schedule, members of each group will
select optimal levels for education. Consider the person who will
set y <yH. If he does this, we know he will set y = O because education
is costly, and until he reaches, y*, there are no benefits to increasing
y, given the employer's hypothesized beliefs. Similarly, any
individual who sets y>yH will in fact set y =yH,since further increases
would merely incur costs with no corresponding benefits.
Everyone will therefore either set y =O or set y =yH. Given the employer's
initial beliefs and the fact just deduced, if the employer's
beliefs are to be confirmed, then members of Group I must set y =0,
while members of Group I1 set y =y*. Diagrams of the options fating
the two groups are shown in Figure 111.
Superimposed upon the wage schedule are the cost schedules for
the two groups. Each group selects y to maximize the difference between
the offered wages and the costs of education. Given the level
of y* in the diagram, it is easy to see that Group I selects y =0, and
Group I1 sets y =y*. Thus, in this case the employer's beliefs are
confirmed, and we have a signaling equilibrium. We can state the
JOB MARKET SIGNALlNG
A Group I1
'i J
Optimal Choice of y Optimal Choice of y
Optimizing Choice of Education for Both Groups
conditions on behavior by the two groups, in order that the employer's
beliefs be confirmed, in algebraic terms. Group I sets y=O if
1>2-y*.
Group I1 will set y =y* as required, provided that
2-y*/2>1.
Putting these two conditions together, we find that the employer's
initial beliefs are confirmed by market experience, provided that the
parameter y* satisfies the inequality,
1<yH<2.
It is worth pausing at this point to remark upon some striking
features of this type of equilibrium. One is that within the class
of employer expectations used above, there is an infinite number of
possible equilibrium values for yH. This means that there is an infinite
number of equilibria. In any one of the equilibria the employer
is able to make perfect point predictions concerning the
productivity of any individual, having observed his level of education.
The reader will realize that this property is special and depends,
at least in part, upon the assumption that education costs are
perfectly negatively correlated with productivity. However, even
in this case, there are equilibria in which the employer is uncertain,
as we shall shortly see.
The equilibria are not equivalent from the point of view of welfare.
Increases in the level of y* hurt Group 11,while, a t the same
time, members of Group I are unaffected. Group I is worse off than
it was with no signaling at all. For if no signaling takes place, each
364 QUARTERLY JOURNAL OF ECONOMICS
person is paid his unconditional expected marginal product, which
is just
Group I1 may also be worse off than it was with no signaling. Assume
that the proportion of people in Group I is 0.5. Since yH>1
and the net return to the member of Group I1 is 2-yH/2, in equilibrium
his net return must be below 1.5, the no-signaling wage.
Thus, everyone would prefer a situation in which there is no signal-
ing.
No one is acting irrationally as an individual. Coalitions might
profitably form and upset the signaling eq~ilibrium.~The initial
proportions of people in the two groups ql and 1- ql have no effect
upon the equilibrium. This conclusion depends upon this assumption
that the marginal product of a person in a given group does not
change with numbers hired.
Given the signaling equilibrium, the education level y*, which
defines the equilibrium, is an entrance requirement or prerequisite
for the high-salary job -or so it would appear from the outside.
From the point of view of the individual, it is a prerequisite that
has its source in a signaling game. Looked a t from the outside,
education niigllt appear to be productive. It is productive for the
individual, but, in this example, it does not increase his real marginal
product at
A sophisticated objection to the assertion that private and
social returns differ might be that, in the context of our example,
the social return is not really zero. We have an information problem
in the society and the problem of allocating the right people
to the right jobs. Education, in its capacity as a signal in the
model, is helping us to do this properly. The objection is well
founded. To decide how efficient or inefficient this system is, one
must consider the realistic alternatives to market sorting procedures
in the society.' But notice that even within the confines of the market
model, there are more or less efficient ways of getting the sorting
accomplished. Increases in y* improve the quality of the sorting
not one bit. They simply use up real or psychic resources. This is
5. Coalitions to change the patterns of signaling are discussed in Spence,
op. cit.
6 . I am ignoring external benefits to education here. The assertion is
simply that in the example education does not contribute to productivity. One
might still claim that the social product is not zero. The signal cost function
does, in principle, capture education as a consumption good, an effect that
simply reduces the cost of education.
7. This question is pursued in Spence, op. cit.
365JOB MARKET SIGNALING
just another way of saying that there are Pareto inferior signaling
equilibria in the market.
It is not always the case that all groups lose due to the existence
of signaling. For example, if, in the signaling equilibrium, y* <2ql,
then Group I1 would be better off when education is functioning
effectively as a signal than it would be otherwise. Thus, in our
example if ql >&so that Group I1 is a minority, then there exists
a signaling equilibrium in which the members of Group I1 improve
their position over the no-signaling case. Recall that the wage in
the no-signaling case was a uniform 2-ql over all groups.
We may generalize this bit of analysis slightly. Suppose that
the signaling cost schedule for Group I was given by aly and that
for Group I1 by a , ~ . ~ a calculation,Then with small amount of
we can show that there is a signaling equilibrium in which Group
I1 is better off than with no ~ignaling,~provided that
41>a2/al.
How small a "minority" Group I1 has to be to have the possibility
of benefiting from signaling depends upon the ratio of the marginal
signaling costs of the two gr0ups.l
Before leaving our education signaling model, it is worth noting
that there are other equilibria in the system with quite different
properties. Suppose that the employer's expectations are of the
following form:
If y <y" : Group I with probability ql,
Group I1 with probability 1-ql ;
if y>y": Group I1 with probability 1.
As before, the only levels of y that could conceivably be selected are
8. It is assumed that au<al.
9. Notice that the statement is that there exists a signaling equilibrium
in which Grour, I1is better off. It turns out that there alwavs exists a signaling--equilibrium in'which Group I1 is worse off as well.
1. The calculation is straightforward. Given these signaling costs groups
will make the requisite choice to confirm the employer's beliefs provided that
1>2-aly*
and
2-azy*>l.
These translate easily into the following condition on y* :
1 1
-<v*<- .az a1
Now, if Group I1 is to be better off for some signaling equilibrium, then
This is what we set out to show.
366 QUARTERLY JOURNAL OF ECONOMICS
Return Return
t t
Optimal y for Group I Optimal y for Group 11
Optimal Signaling Decisions for the Two Groups
y=O and y= y". The wage for y=O is 2-ql, while the wage for
y =y* is simply 2. From Figure IV it is easy to see that both groups
rationally set y =0, provided that y" >2ql. If they both do this,
then the employer's beliefs are confirmed, and we have an equi-
librium.
It should be noted that the employer's beliefs about the relationship
between productivity and education for y>yH are confirmed
in a somewhat degenerate, but perfectly acceptable, sense. There
are no data relating to these levels of education and hence, by logic,
no disconfirming data. This is an example of a phenomenon of much
wider potential importance. The employer's beliefs may drive certain
groups from the market and into another labor market. We
cannot capture this situation in a simple one-employer, one-market
model. But when it happens, there is no experience forthcoming to
the employer to cause him to alter his belief^.^
Education conveys no information in this type of equilibrium.
In fact, we have reproduced the wages and information state of the
employer in the no-signaling model, as a signaling equilibrium.
Just as there exists a signaling equilibrium in which everyone
sets y=O, there is also an equilibrium in which everyone sets y= y"
for some positive y". The requisite employer beliefs are as follows:
If y <y": Group I with probability 1;
if y>y" : Group I with probability ql,
Group I1with probability 1-ql.
2. This is discussed in detail in Spence, op. cit.
367JOB MARKET SIGNALING
Following our familiar mode of analysis, one finds that these beliefs
are self-confirming in the market, provided that
y*<l-q1.
Again, the education level conveys no useful information, but in this
instance individuals are rationally investing in education. If they
as individuals did not invest, they would incur lower wages, and
the loss would exceed the gain from not making the educational investment.
The implication of this version of the signaling equilibrium
is that there can be stable prerequisites for jobs that convey
no information by virtue of their existence and hence serve no func-
tion.
It is interesting to note that this last possibility does not depend
upon costs being correlated with productivity at all. Suppose
that the signaling costs for both groups were given by the one
schedule y. And suppose further that employer beliefs were as described
above. Then everyone will rationally select y = yH,provided
that
y*<l-41.
The outcome is the same. But the interesting thing is that, because
of the absence of any correlation between educational costs and
productivity, education could never be an effective signal, in the
sense of conveying useful information, in an equilibrium in this
market.
We have dwelt enough upon the specifics of this model to have
observed some of the effects the signaling game may have upon the
allocational functioning of the market. The numerical example is
not important. The potential effects and patterns of signaling are.
An alterable characteristic like education, which is a potential
signal, becomes an actual signal if the signaling costs are negatively
correlated with the individual's unknown productivity. Actually,
the negative correlation is a necessary but not sufficient condition
for signaling to take place. To see this in the context of our model,
assume that the only values y can have are one and three. That
is to say, one can only get units of education in lumps. If this is
true, then there is no feasible value of y* that will make it worthwhile
for Group I1to acquire an education. Three units is too much,
and one unit will not distinguish Group I1 from Group I. Therefore,
effective signaling depends not only upon the negative correlation
of costs and productivities, but also upon there being a "sufficient"
number of signals within the appropriate cost range.3
3. In ibid. it is argued that many potential signals in credit and loan
368 QUARTERLY JOURNAL OF ECONOMICS
An equilibrium is defined in the context of a feedback loop, in
which employer expectations lead to offered wages to various levels
of education, which in turn lead to investment in education by individuals.
After hiring, the discovery of the actual relationships
between education and productivity in the sample leads to revised
expectations or beliefs. Here the cycle starts again. An equilibrium
is best thought of as a set of beliefs that are confirmed or at least
not contradicted by the new data at the end of the loop just described.
Such beliefs will tend to persist over time as new entrants
into the market flow through.
Multiple equilibria are a distinct possibility. Some may be
Pareto inferior to others. Private and social returns to education
diverge. Sometimes everyone loses as a result of the existence of
signaling. In other situations some gain, while others lose. Systematic
overinvestment in education is a distinct possibility because of
the element of arbitrariness in the equilibrium configuration of the
market. In the context of atomistic behavior (which we have assumed
thus far) everyone is reacting rationally to the niarket situation.
Information is passed to the employer through the educational
signal. In some of our examples it was perfect information.
In other cases this is not so. There will be random variation in signaling
costs that prevent the employer from distinguishing perfectly
among individuals of varying productive capabilities.
In our examples, education was measured by a scalar quantity.
With no basic adjustment in the conceptual apparatus, we can think
of education as a multidimensional quantity: years of education,
institution attended, grades, recommendations and so on. Similarly,
it is not necessary to think in terms of two groups of people. There
may be many groups, or even a continuum of people: some suited to
certain kinds of work, others suited to other kinds. Nor need education
be strictly unproductive. However, if it is too productive relative
to the costs, everyone will invest heavily in education, and education
may cease to have a signaling function.
In the educational signaling model we avoided considering any
observable characteristics other than education. In that model
education was a signal. Here we consider what role, if any, is played
markets effectively become indices because the "signaling" costs swamp the
gains, so that characteristics that could be manipulated in fact are not. House
ownership is an example of a potential signal that, in the context of the loan
market, fails on this criterion and hence becomes an index.
JOB MARKET SIGNALING 369
by indices. For concreteness I shall use sex as the example. But
just as education can stand for any set of observable, alterable
characteristics in the first model, sex can stand for observable,
unalterable ones here. The reader may wish to think in terms of
race, nationality, size, or in terms of criminal or police records and
service records. The latter is potentially public information about
a person's history and is, of course, unalterable when viewed retrospectively
from the p r e ~ e n t . ~
Let us assume that there are two groups, men and women. I
shall refer to these groups as W and M. Within each group the distribution
of productive capabilities and the incidence of signaling
costs are the same. Thus, within M the proportion of people vith
productivity one and signaling (education) costs of y is gl. The
remainder have productivity two and signaling costs y/2. The
same is true for group W. Here m is the proportion of men in the
overall population of job applicants.
TABLE IV
Education Proportion Proportion of
Race Productivity costs within group total population
W 1 ?d ql s1(l-m)
W 2 ?dl2 1-q1 ( 1 - q 1 ) ( l - m )
nl 1 Id 41 qlm
M 2 ?d/2 1-ql (1-ql)m
Given the assumptions the central question is, "how could sex
have an informational impact on the market?" The next few paragraphs
are devoted to arguing that indices do have a potential impact
and to explaining why this is true. We begin by noting that,
under the assumptions, the conditional probability that a person
drawn at random from the population has a productivity of two,
given that he is a man (or she is a wornan), is the same as the unconditional
probability that his productivity is two. Sex and productivity
are uncorrelated in the population. Therefore, by itself,
sex could never tell the employer anything about productivity.
We are forced to the conclusion that if sex is to have any informational
impact, it must be through its interaction with the
educational signaling mechanism. But here again we run up against
an initially puzzling symmetry. Under the assumptions, men and
women of equal productivity have the same signaling (education)
4. I t is, or ought to be, the subject of policy decisions as well.
370 QUARTERLY JOURNAL OF ECONOMICS
costs. It is a general maxim in economics that people with the same
preferences and opportunity sets will make similar decisions and end
up in similar situations. We may assume that people maximize
their income net of signaling costs so that their preferences are the
same. And since signaling costs are the same, it would appear that
their opportunity sets are the same. Hence, again we appear to be
driven to the conclusion that sex can have no informational impact.
But the conclusion is wrong, for an interesting reason.
The opportunity sets of men and women of comparable productivity
are not necessarily the same. To see this, let us step back to
the simple educational signaling model. There are externalities in
that model. One person's signaling strategy or decision affects the
market data obtained by the employer, which in turn affect the employer's
conditional probabilities. These determine the offered wages
to various levels of education and hence of rates of return on education
for the next group in the job market. The same mechanism
applies here, with a notable modification. If employers' distributions
are conditional on sex as well as education, then the external
impacts of a man's signaling decision are felt only by other men.
The same holds for won~en.
If a t some point in time men and women are not investing in
education in the same ways, then the returns to education for men
and women will be different in the next round. In short, their opportunity
sets differ. I n what follows, we demonstrate rigorously
that this sort of situation can persist in an equilibrium. The important
point, however, is that there are externalities implicit in
the fact that an individual is treated as the average member of the
group of people who look the same and that, as a result, and in spite
of an apparent sameness the opportunity sets facing two or more
groups that are visibly distinguishable may in fact be different.
The employer now has two potential signals to consider: education
and sex. At the start he does not know whether either education
or sex will be correlated with productivity. Uninformative
potential signals or indices are discarded in the course of reaching
an equilibrium. As before we must guess at an equilibrium form for
the employer's expectations and then verify that these beliefs can be
self-confirming via the niarket informational feedback mechanisms.
We will try beliefs on the following form.
If W and y <yHW,productivity =1with probability 1.
If W and y>yHW,productivity =2 with probability 1.
If M and y <ySM,productivity =1with probability 1.
If M and y>yHN,productivity =2 with probability 1.
371JOB MARKET SIGNALING
FIGUREV
Offered Wages to W and M
These lead to offered wage schedules W w( y ) and W M ( y )as shown
in Figure V.
Because groups W and M are distinguishable to the employer,
their offered wages are not connected a t the level of employer expectations.
Applying the reasoning used in the straightforward educational
signaling model, we find that the required equilibrium conditions
on yHwand yHMare
1<yHw<2
and
1<yH&f<2.
No logical condition requires that yHwequals yHMin an equilibrium.
Essentially we simply have the educational signaling model
iterated twice. Because sex is observable, the employer can make
his conditional probability assessments depend upon sex as well as
education. This has the effect of making signaling interdependencies
between two groups, W and M, nonexistent. They settle into signaling
equilibrium configurations in the market independently of each
other. But in the first model there was not one equilibrium, there
were many. Therefore, there is a t least the logical possibility that
men and women will settle into different stable signaling equilibria
in the market and stay there.
As we noted earlier, the signaling equilibria are not equivalent
from the point of view of social welfare. The higher that yHw (or
yHM)is, the worse off is the relevant group or, more accurately, the
high-productivity portion of the group. One example of an asymmetrical
equilibrium would be given by ySM=1.1and y"M= 1.9. I11
this case high-productivity women have to spend more on education
372 QUARTERLY JOURNAL OF ECONOMICS
Returns Returns
t 3.
FIGUREVI
Market Equilibrium with Sex as an Index
and have less left over to consume in order to convince the employer
that they are in the high-productivity group.
Notice that the proportions of high- and low-productivity
people in each group do not affect the signaling equilibrium in the
market. Hence, our initial assumption that the groups were identical
with respect to the distribution of productive characteristics and
the incidence of signaling costs was superfluous. More accurately,
it was superfluous with respect to this type of equilibrium. As we
saw in the educational signaling model, there are other types of
equilibrium in which the proportions matter.
Since from an equilibrium point of view men and women really
are independent,they might settle into different types of equilibrium.
Thus, we might have men signaling y = yWAqI=1.1if they are also in
the higher productivity group, while other men set y=O. On the
other hand, we may find that all women set y=O. In this case all
women would be paid 2-ql, and the upper signaling cutoff point
yHMwould have to be greater than 2q,. Notice that all women, including
lower productivity women, would be paid more than lowproductivity
men in this s i t ~ a t i o n . ~High-productivity women
would, of course, be hurt in terms of wages received. It is conceivable,
however, that returns net of signaling would be higher for
women with productivity of two. In other words, it is possible that
2-ql>w-y*,/,.
5. I have not assumed that employers are prejudiced. If they are, this
differential could be wiped out. Perhaps more interestingly laws prohibiting
wage discrimination, if enforced, would also wipe it out.
JOB MARKET SIGNALING
Y
Women
FIGUREVII
Another Equilibrium Configuration in the Market
This will occur when
2q1< Y",.
Looking at this situation from outside, one might conclude that
women receive lower wages than some men because of a lack of
education, which keeps their productivity down. One might then go
looking outside the job market for the explanation for the lack of
education. In this model the analysis just suggested would be wrong.
The source of the signaling and wage differentials is in the informational
structure of the market itself ."
Because of the independence of the two groups, M and W , at
the level of signaling, we can generate many different possible
equilibrium configurations by taking any of the educational signaling
equilibria in our first model and assigning it to W and then taking
any education equilibrium and assigning it to M. However, an
exhaustive listing of the possibilities seems pointless at this stage.
We have here the possibility of arbitrary differences in the
equilibrium signaling configurations of two or more distinct groups.
Some of them may be at a disadvantage relative to the others. Subsets
of one may be a t a disadvantage to comparable subsets of the
others. Since the mechanism that generates the equilibrium is a
feedback loop, we might, following Myrdal and others, wish to
refer to the situation of the disadvantaged group as a vicious cycle,
albeit it an informationally based one. I prefer to refer to the situ-
6. Differential signaling costs over groups are an important possibility
pursued in Spence, op. cit.
374 QUARTERLY JOURNAL OF ECONOMICS
ation of the disadvantaged group as a lower level equilibrium trap,
which conveys the notion of a situation that, once achieved, persists
for reasons endogenous to the model. The multiple equilibria of the
education model translate into arbitrary differences in the equilibrium
configuration and status of two groups, as defined as observable,
unalterable characteristics.
We have looked at the characteristics of a basic equilibrium
signaling model and at one possible type of interaction of signals
and indices. There remains a host of questions, which can be posed
and partially answered within the conceptual framework outlined
here. Among them are the following:
1. What is the effect of cooperative behavior on the signaling
game?
2. What is the informational impact of randomness in signaling
costs?
3. What is the effect of signaling costs that differ systematically
with indices?
4. How general are the properties of the examples considered
here?
5. In a multiple-market setting, does the indeterminateness of
the equilibrium remain?
6. Do signaling equilibria exist in general?
7. What kinds of discriminatory mechanisms are implicit in,
or interact with, the informational structure of the market, and what
policies are effective or ineffective in dealing with them?
I would argue further that a range of phenomena from selective
admissions procedures through promotion, loans and consumer
credit, and signaling status via conspicuous consumption lends itself
to analysis with the same basic conceptual apparatus. Moreover,
it may be as important to explain the absence of effective signaling
as its presence, and here the prerequisites for effective signaling are
of some use.
On the other hand, it is well to remember that the property of
relative infrequency of appearance by signalers in the market, which
defines the class signaling phenomena under scrutiny here, is not
characteristic of many markets, like those for consumer durables,
and that, as a result, the informational structures of these latter are
likely to be quite different.