The Peter Principle: A Theory of Decline

Edward P. Lazear
Hoover Institution and Stanford University

Some have observed that individuals perform worse after being pro-
moted. The Peter principle, which states that people are promoted
to their level of incompetence, suggests that something is fundamen-
tally misaligned in the promotion process. This view is unnecessary
and inconsistent with the data. Below, it is argued that ability appears
lower after promotion purely as a statistical matter. Being promoted
is evidence that a standard has been met. Regression to the mean
implies that future ability will be lower, on average. Firms optimally
account for the regression bias in making promotion decisions, but
the effect is never eliminated. Rather than evidence of a mistake, the
Peter principle is a necessary consequence of any promotion rule.
Furthermore, firms that take it into account appropriately adopt an
optimal strategy. Usually, firms inflate the promotion criterion to offset
the Peter principle effect, and the more important the transitory com-
ponent is relative to total variation in ability, the larger the amount
that the standard is inflated. The same logic applies to other situations.
For example, it explains why movie sequels are worse than the original
film on which they are based and why second visits to restaurants are
less rewarding than the first.

The Peter principle states that workers are promoted to their level of
incompetence (Peter and Hull 1969). One interpretation is that firms

Sherwin Rosen was my most important teacher, my valued colleague, and dear friend.
Sherwin served on my thesis committee and taught me much of what I know. Throughout
the 30 years that we were friends, Sherwin was a constant source of inspiration, wisdom,
and kindness. A deep thinker who opened up a number of areas of research, Sherwin
was interested in hierarchies and promotion, so this paper is very much in keeping with
his research agenda and derives from it. This research was supported in part by the National
Science Foundation. Useful comments were provided by many. Po-Han Fong, Thomas P.
Gherig, Kevin J. Murphy, Torsten Persson, Paul Pfleiderer, Kathryn Shaw, Eskil Wadensjo,
and Michael Waldman were particularly helpful.

[Journal of Political Economy, 2004, vol. 112, no. 1, pt. 2]

䉷 2004 by The University of Chicago. All rights reserved. 0022-3808/2004/11201S1-0012$10.00

S141
S142 journal of political economy
systematically make mistakes in their promotion decisions. Another, fa-
vored here, is that the decline in ability that is seen after promotion is
the natural outcome of a statistical process that displays regression to
the mean. Workers are promoted on the basis of having met some
standard. To the extent that ability is the sum of both permanent and
transitory components, those who meet the standard have expected
transitory components that are positive. The expectation of the post-
promotion transitory component is zero, implying a reduction in ex-
pected ability. Firms that understand the statistical process take this
phenomenon into account by adjusting the promotion standard, but
the result remains: Expected ability for those promoted is lower after
promotion than before.1
There is substantial evidence of the Peter principle.2 In addition to
papers from the marketing and organizational behavior literature (see,
e.g., Anderson, Dubinsky, and Mehta 1999), there are a number of
findings in empirical labor economics that support the claim. In an
early paper that used subjective performance evaluation, Medoff and
Abraham (1980) reported that workers’ subjective evaluation scores fell
the longer they were on the job. In Lazear (1992), it was found that
the coefficient of job tenure in a wage regression was actually negative.
The longer a worker was in a particular job, given his tenure in the
firm, the lower his wage. The reason presumably is that the better work-
ers are promoted out of the job, so those with a given number of years
of firm experience who have fewer years in a job are less likely to have
gotten stuck in that job. Baker, Gibbs, and Holmstrom (1994) replicate
this finding in their data, and Gibbs and Hendricks (2001) find that
raises and bonuses fall with tenure.
The tone of the literature outside of economics is that there is some-
thing wrong with promotion dynamics, and this anomaly shows up as
the Peter principle. (Indeed, the book written by Peter and Hull is
entitled The Peter Principle: Why Things Always Go Wrong.) The approach
taken here is different. The Peter principle results from optimal ad-
justment to decision making under uncertainty. It is argued that even
when firms use exactly the right promotion rule, the Peter principle
1
It is also true that those who are denied promotion do better after they are turned
down than they did before the decision was made, for the same reason.
2
Two recent papers (Faria 2000; Fairburn and Malcolmson 2001) on this topic use an
approach very different from this paper and from each other. The Peter principle is a by-
product of using promotion to solve a moral hazard problem in Fairburn and Malcolmson.
Rather than motivate through money, which induces influence activity, firms choose pro-
motion because then managers must live with the consequences of their decisions. Too
many workers are promoted under certain circumstances, resulting in a Peter principle
effect. In Faria’s paper, workers have two skills. Those who are good at one are necessarily
less good at another when on the frontier. Faria argues that this is what is meant by the
Peter principle.
the peter principle S143
effect will be observed. The fact that reversals of promotion decisions
are rare is more compatible with the view that they were correct in the
first place than that something went wrong.
More often, the Peter principle is interpreted in a multifactor context.
Individuals who are good in one job are not necessarily good in the job
into which they are promoted. As a result, individuals appear incom-
petent in the job in which they settle. To obtain this result, it is merely
necessary to make a slight modification in the regression to the mean
structure. Here, general ability is combined with a job-specific ability to
produce output. Regression to the mean results because positive read-
ings on the job-specific component prior to promotion are uncorrelated
with the job-specific component after promotion.3
The fact that promoted individuals are less able than their apparent
prepromotion ability induces firms to adjust in two respects. First, firms
select their promotion rule with the understanding that the prepro-
motion ability is a biased estimate of true ability for those who exceed
some standard. Second, as the variance in the transitory component of
ability rises relative to the variance in the permanent component, the
adjustment factor becomes greater. In the typical case, the standard that
one must exceed to obtain a promotion increases with the relative im-
portance of the transitory component because the regression effect is
larger when the variance of the transitory component is large.
The model presented below yields the following results:

1. Promoted individuals’ performance falls, on average, relative to

their prepromotion performance.
2. Firms that take the decline into account adjust their promotion rule
accordingly, but this does not negate the observation that ability
declines after promotion.
3. The importance of the Peter principle depends on the amount of
variation in the transitory component relative to the permanent.
The Peter principle is most pronounced when the transitory com-
ponent is large.
4. The length of the prepromotion period depends on the ratio of
transitory variation to permanent variation. As the transitory com-
ponent becomes more important, firms lengthen the prepromotion
period.
5. Movie sequels are systematically worse than the original on which
they are based.
6. Follow-up visits to good restaurants provide poorer meals than the
first sampling.
3
The structure is a variant of the Jovanovic (1979a, 1979b) model that was modified
and used in a context closer to this structure in Lazear (1986).
S144 journal of political economy
7. In the absence of learning effects, second-term elected officials are
less effective than they were during the first term.

I. Model
Let there be two periods. Each worker has a time-invariant component
of ability, denoted A ∼ f(A), and a time-varying component of ability,
denoted e1 for period 1 and e 2 for period 2. Let the time-varying com-
ponents be independently and identically distributed with density g(e).
The firm can observe A ⫹ et in each period but cannot disentangle the
time-varying component of ability from the permanent component.
There are a variety of interpretations that are consistent with this spec-
ification. One can think of the et as being a true transitory aspect of
ability or just measurement error. Later, the interpretation of different
jobs will be considered.
There are two jobs (two are sufficient), which we denote difficult and
easy. An individual’s productivity in the easy job is given by
a ⫹ b(A ⫹ et)
and in the difficult job is given by
g ⫹ d(A ⫹ et),
where a 1 g and d 1 b. Thus it pays to assign a worker to the difficult
job if and only if
A ⫹ et 1 x,
where x { (a ⫺ g)/(d ⫺ b). The situation and the crossing point that
correspond to x are shown in figure 1.4 The setup seeks to capture the
idea that the most able have a comparative advantage in the difficult
job.
Assume that individual ability f(A) is such that, in the absence of
information, it pays to assign everyone to the easy job in period 1.5
Intuitively, this assumption amounts to saying that most people are not
well suited to the difficult job and that, in the absence of countervailing
information, individuals are assigned to the easy job. With one excep-
4
This production structure is similar to that used in a comprehensive analysis by Gibbons
and Waldman (1999b), who also allow for transitory and permanent components with
regression. The focus of their paper is earnings and promotion. Neither optimal decision
making by firms given the transitory component nor strategic effort in response to pro-
motion rules is central to their discussion.
5
This amounts to assuming that

冕冕 冕冕
⬁ ⬁ ⬁ ⬁

[a ⫹ b(A ⫹ e)]dGdF 1 [g ⫹ d(A ⫹ e)]dGdF.

⫺⬁ ⫺⬁ ⫺⬁ ⫺⬁
the peter principle S145

Fig. 1

tion, noted in the Appendix, it is sufficient to assume symmetric ig-

norance, so that workers are no better informed about their abilities
than firms.
After the first period, firms obtain an estimate of A, namely  p
A ⫹ e1. Since e1 is the period 1 transitory component (either measure-
ment error or transitory ability), it is A and not  on which a promotion
decision should be made. But A is not observed, so firms are forced to
base their decision on Â.

A. Workers Perform Worse after Being Promoted

Firms must select some criterion level, A∗, such that if Aˆ 1 A∗, the worker
is promoted to the difficult job. If  ! A∗, the worker remains in his
current job. It is now shown that workers who are promoted have levels
of ability in period 1 that are higher, on average, than their ability in
period 2.
First, note that the expectation of e1 given that an individual is pro-
moted is

冕冕
⬁ ⬁
1
E(e1FA ⫹ e1 1 A∗) p eg(e)f(A)dedA
⫺⬁ A∗⫺A 1 ⫺ G(A∗ ⫺ A)

冕
⬁

p E(eFe 1 A∗ ⫺ A)f(A)dA,
⫺⬁
S146 journal of political economy
which is positive since f(A) is positive and the conditional expectation
of e given e greater than any number is positive (because the uncon-
ditional expectation of e is zero). Thus the conditional expectation of
e1 is positive among those who are promoted.
Now, in period 2, the expectation of the transitory component is
E(e 2FA ⫹ e1 1 A∗) p 0
because e 2 is independent of A and of e1. As a result, for any promoted
individual with ability A,
A ⫹ E(e1FA ⫹ e1 1 A∗) 1 A ⫹ E(e 2FA ⫹ e1 1 A∗).
Thus expected ability falls for promoted individuals from period 1 to
period 2.
Individuals who are promoted are promoted in part because they are
likely to have high permanent ability,6 but also because the transitory
component of their ability is high. One of the reasons that academics
tend to write better papers before they receive tenure is that they would
not have received tenure had they not written the better-than-average
papers. The point is obvious but is made graphic by the following ex-
ample. Suppose that a firm promotes all individuals who can obtain
three heads on three consecutive coin tosses. Only one in eight will be
promoted. But when the firm asks the promoted individuals to repeat
the feat, only one in eight will measure up. Seven out of eight will do
worse than they did before being promoted. The reason is that all of
the “performance” on the coin toss is transitory since tosses are
independent.
As a general matter, the larger the transitory component is relative
to the permanent component, the more important the Peter principle
effect. If there were no transitory component, there would be no re-
gression to the mean. Thus the importance of “luck” is positively as-
sociated with the force of the Peter principle.

B. The Promotion Rule

Firms know in advance that there will be some expected fall in pro-
ductivity among those promoted and adjust their promotion standard
accordingly. Below, the general optimization problem for the firm is
presented. Then, the way in which the rule operates is demonstrated
by an example.
The firm’s problem is to maximize profits (or worker utility) by se-
lecting the job for each candidate with the highest expected value. Recall
6
The notation e1 and A could be swapped in the discussion above to show that
E(AFA ⫹ e1 1 A ∗) 1 E(A).
the peter principle S147
that individuals who have period 2 ability greater than x, defined above,
would be assigned to job 2 were second-period ability known. The firm
does not see A, but only Â, and must choose some criterion, A∗, such
that it promotes workers whose observed ability in period 1 is greater
than A∗. This is equivalent to promoting individuals when A 1 A∗ ⫺ e1.
Thus the firm wants to choose A∗ so as to maximize

冕冕 冕
⬁ ⬁ ⬁

max [g ⫹ d(A ⫹ e 2)]f(A)g(e1)g(e 2)de 2dAde1

A∗ ⫺⬁ A∗⫺e1 ⫺⬁

冕冕 冕
⬁ A∗⫺e1 ⬁

⫹ [a ⫹ b(A ⫹ e 2)]f(A)g(e1)g(e 2)de 2dAde1. (1)

⫺⬁ ⫺⬁ ⫺⬁

Because the expectation of e 2 is zero, (1) can be written as

冕冕 冕冕
⬁ ⬁ ⬁ A∗⫺e1

max (g ⫹ dA)dFdG ⫹ (a ⫹ bA)dFdG. (2)

A∗ ⫺⬁ A∗⫺e1 ⫺⬁ ⫺⬁

The choice of A∗ depends on the distribution. However, two examples

reveal that A∗ does not equal x as a general rule. Instead, in typical
cases, firms adjust A∗ upward. Knowing that worker ability in period 2
will differ from worker ability in period 1, firms usually set the bar higher
than they would, where ability observed in period 1 carried over directly
to period 2.
Actual solutions that provide intuition are available for given distri-
butions. Consider, for example, the case in which A, e1, and e 2 are all
distributed normally, with mean zero and variance equal to one. Let
a p 1, b p 0.5, g p 0, and d p 1. Then x, the ability level at which
jobs produce equal value, is two since
a ⫹ b(A ⫹ e) p g ⫹ d(A ⫹ e)
for A ⫹ e p 2. However, A∗ p 4.01. The firm sets its promotion standard
more than two standard deviations higher than the crossing point in
figure 1 because it understands that the worker’s ability in period 2 is
likely to be lower than it was in period 1 for the promoted group. As
a result, the firm insists on a very high level of observed ability in period
1 in order to warrant promotion. Statements such as “tenure requires
that the faculty member be the best in his or her field, having produced
outstanding research” is a manifestation of the upward adjustment.
Consider the same example with a twist. Let the distribution of A
remain the same, namely, normal with mean zero and a standard de-
viation of one, but let the standard deviation of e1 fall to 0.1. Then A∗
drops from 4.01 to 2.08. Although the firm still adjusts its promotion
criterion upward from x, the adjustment is much smaller because the
S148 journal of political economy
importance of the transitory component has been diminished. There
is regression to the mean, but the regression that takes place is small
relative to the amount in the prior example. When the standard devi-
ation of e is zero, the promotion standard is two, which is exactly x as
expected. Then the distribution of e1 is degenerate, so that all observed
ability in period 1 is permanent ability. The problem in (2) becomes

冕 冕
⬁ A∗

max (g ⫹ dA)f(A)dA ⫹ (a ⫹ bA)f(A)dA,

A∗ A∗ ⫺⬁

which has first-order condition

⭸
p ⫺(g ⫹ dA∗)f(A∗) ⫹ (a ⫹ bA∗)f(A∗) p 0.
⭸A∗

The solution is

g ⫹ dA∗ p a ⫹ bA∗,

which is the crossing point, that is, x, in figure 1. When there is no

transitory component, the firm simply promotes those whose permanent
ability places them better in the difficult job than in the easy job.
It is possible to derive the relation between A∗ and x in more general
terms.7 The first-order condition to (2) is

冕
⬁

(g ⫹ dA∗ ⫺ a ⫺ bA∗) f(A∗ ⫺ e1)g(e1)de1 p

⫺⬁

冕
⬁

(d ⫺ b) e1 f(A∗ ⫺ e1)g(e1)de1.
⫺⬁

The integral on the left-hand side is always positive, so whether A∗ ex-

ceeds x or not depends on the sign of the integral on the right-hand
side. Assume that both f(7) and g(7) are symmetric densities, and let
g(7) be symmetric around zero and f(7) be symmetric around A ¯ . Write
the integral on the right side as

冕 冕
0 ⬁

e1 f(A ⫺ e1)g(e1)de1 ⫹
∗
e1 f(A∗ ⫺ e1)g(e1)de1.
⫺⬁ 0

Use a change of variable in the first integral of u p ⫺e1, and in the

second, allow u p e1. Because of symmetry, g(u) p g(⫺u), so one can
7
I am indebted to Wing Suen for this derivation.
the peter principle S149
write the two integrals as one:

冕
⬁

u[f(A∗ ⫺ u) ⫺ f(A∗ ⫹ u)]g(u)du.

0

Suppose that the firm wants to promote fewer than 50 percent of the
¯ . (Recall that x is the value such that a ⫹
people, that is, that x 1 A
bx p g ⫹ dx. If x p A, then because of symmetry of the density func-
¯
tions, half of the population would have A 1 x and half would have
A ! x.) Under such circumstances, f(x ⫺ u) 1 f(x ⫹ u) (because f is uni-
modal around Ā). Thus the right-hand side of the first-order condition
is positive for A∗ p x, which implies that

a ⫹ bA∗ 1 g ⫹ dA∗

or that A∗ 1 x. Thus the firm adjusts the cutoff level upward when fewer
than 50 percent of the workers are better suited to the difficult job than
to the easy job.
The same reasoning applies in reverse. If more than 50 percent are
to be promoted, then x ! A ¯ , which means that f(x ⫺ u) ! f(x ⫹ u). As a
result, the right-hand side of the first-order condition is negative at x,
which means that A∗ must be less than x to satisfy optimality. Thus, when
more than 50 percent are better suited to the difficult job than to the
easy job, the firm reduces the promotion cutoff below the one that
would be optimal were there no error in period 1. (Actually, under such
circumstances, workers would initially be assigned to the difficult job,
and the standard would be one such that workers who fell below it
would be demoted after the first period.)
The intuition is this. Although there is always regression to the mean,
adjusting the promotion level upward reduces the probability that the
firm will make a bad promotion decision. However, at the same time,
the adjustment increases the probability that it will fail to promote a
qualified worker; that is, it reduces the false positive while increasing
the false negative error. Conversely, lowering the promotion cutoff re-
duces the probability that someone who erroneously was observed to
be a poor worker is not promoted but increases the probability that the
firm promotes too many bad workers. Thus there is a trade-off. When
fewer than 50 percent are to be promoted, the expected cost of making
a false positive error exceeds that of making a false negative error so
that the criterion must be adjusted upward. To the extent that most
hierarchies are narrower at the top than at the bottom, A∗ 1 x is probably
the more typical case. Standards are adjusted upward.
S150 journal of political economy
C. Another Interpretation of the Peter Principle
Rosen (1986) presents a model of sequential promotions in which in-
dividuals are sorted by ability at each stage such that members of the
entering class at each round of a tournament have equal (ex ante)
ability. Rosen uses the model to determine the optimal compensation
at each level to motivate workers. Sorting is also an issue because the
pool of workers at successive rounds have higher ability than those at
earlier rounds. The Rosen model is in some ways more general than
the one described here because it allows for effort as well as ability
differences. The focus is quite different, however, because neither the
optimal promotion rule nor the worker’s output over time is an im-
portant part of the analysis. It is likely that many of the Peter principle
results that come out of this paper could have been derived in that
important paper on sequential promotions.8
Still, the Rosen model does not fit one of the most common inter-
pretations of the Peter principle, which is that workers are promoted
to their level of incompetence because a worker who is good in one
job is not necessarily good in a job one level up. Fine professors do not
necessarily make good deans (although not all would interpret moving
to the dean’s job as a promotion).9 A slight modification of the defi-
nitions above and some of the formulas permits this interpretation.
To see this, allow e1 to be defined as the job-specific component of
ability associated with the easy job and e 2 as the job-specific component
of ability associated with the difficult job. Individuals are assigned to
the easy job in period 1 for the reason given before: Most are better at
the easy job, and in the absence of information, the easy job is the right
assignment. After evaluation, Â is observed and the worker is promoted
or not. If he is not promoted, then his ability after promotion is A ⫹
e1. If he is promoted, then his ability after promotion is A ⫹ e 2.
Under this interpretation, workers who are not promoted have output
that remains constant over time and equal to a ⫹ b(A ⫹ e1). Those who
are promoted have output equal to g ⫹ d(A ⫹ e 2). The argument of
subsection A holds:

冕冕
⬁ ⬁
1
E(e1FA ⫹ e1 1 A ) p
∗
eg(e)f(A)dedA
⫺⬁ A∗⫺A 1 ⫺ G(A ⫺ A)
∗

冕
⬁

p E(eFe 1 A∗ ⫺ A)f(A)dA,
⫺⬁

8
Gibbons and Waldman (1999a) and Prendergast (1999) survey the large literature on
careers and promotions.
9
Anderson et al. (1999) claim that the data reveal a Peter principle for sales managers
because the skills needed by salespeople are generally distinctly different from those
needed by sales managers.
the peter principle S151
which is positive since f(A) is positive and the conditional expectation
of e given e greater than any number is positive (because the uncon-
ditional expectation of e is zero). But the expectation of e 2 is zero for
promoted workers because e1 and e 2 are uncorrelated. As a result, ex-
pected ability is higher before promotion than for promoted workers
after promotion.
This does not necessarily imply that output is lower after promotion
because workers are in different jobs. On the contrary, if A∗ is chosen
optimally, it must be the case that expected output for the promoted
workers is higher in the difficult job than in the easy job. If it were not,
it would be better to raise A∗ until expected output were higher. Rather
the point is that after promotion, the average promoted worker is not
as able in the difficult job as he was in the easy job, that is,
E(A ⫹ e 2Fpromoted) ! E(A ⫹ e1Fpromoted).
Also true is that within any job, those left behind and not promoted
have lower average ability than those of their cohorts entering into that
job. If there were a series of promotion rounds, then at every level,
those who were not promoted would have a job-specific component that
is negative. This can be seen simply by examining the first round, which
can be thought of as a “promotion” from being out of the firm to being
hired as a worker. (Individuals must exceed some standard in order to
be hired.) Since it has already been shown that
E(e1FA ⫹ e1 1 A∗) 1 0
and since E(e1) p 0, it must be true that
E(e1FA ⫹ e1 ≤ A∗) ! 0.
They appear “incompetent” because, within any given job, the actual
ability of those who are not promoted out of the job is lower than the
unconditional expectation of ability for that job. Those who are left
behind and become the long-termers are worse than those who come
into the job. They are incompetent relative to the entry pool because
the competent workers are promoted out of the job. In a tournament
with enough steps, each competent worker would continue to be pro-
moted until he too was incompetent, that is, until E(et) ! 0 for those
whose highest job attained is job t. This is the Peter principle: Workers
are promoted to their level of incompetence. Those who are “compe-
tent” are promoted again.10

10
One difference between this interpretation of the Peter principle and the one used
in the rest of this paper is that the output of those not promoted does not rise under the
job-specific interpretation. Since e1 is a job-specific effect and not a transitory component,
those who are not promoted have ability A ⫹ e1 in both periods.
S152 journal of political economy
D. Mistake or Optimal Adjustment?
The original book was entitled The Peter Principle: Why Things Always Go
Wrong. The implications of the view that promotion decisions are biased
are different from those of this model. The view in this model is that
if A∗ is chosen optimally, most of the time firms will not want to undo
their decision. Even though ability is below that predicted by a naive
use of the first-period estimate, the promoted worker’s ability is still
above x, at least on average. If it were not, then the choice of A∗ would
have been suboptimal. Ability falls but does not fall below x in most
cases, so the firm does not want to reverse its decision. The behavioral
view that “things go wrong” is different. If the firm really made a mistake,
then it would want to demote or fire workers in most cases, which leads
to the inevitable question, “Why are demotions so rare?” This analysis
provides an answer. The promotion rule is chosen optimally so that
ability is not as high as it was before promotion, but it is still high enough
to justify the promotion.

II. Strategic Behavior by Workers

So far, worker effort has been assumed to be given. In this section, I
relax the assumption that effort is given in order to determine how
workers may game the system to alter their promotion possibilities.
Whether workers overproduce during the probationary period depends
crucially on the compensation scheme.

A. Efficient Effort with Worker Job Choice

The first result is that if workers are paid piece rates and allowed to
choose their own jobs, all is efficient, even if workers know their ability
and firms do not. In order to examine incentives, it is necessary to
define more terms: let me be the amount of effort that an individual
chooses if he is in the easy job and md be the amount of effort he chooses
if he is in the difficult job. Let C(m) be the cost incurred for any given
level of effort m. Then if A is known but e is not, individuals for whom

冕
⬁

[g ⫹ d(A ⫹ e ⫹ md)]g(e)de ⫺ C(md) 1

⫺⬁

冕
⬁

[a ⫹ b(A ⫹ e ⫹ me)]g(e)de ⫺ C(me) (3)

⫺⬁

choose the difficult job. Those for whom the condition in (3) does not
the peter principle S153
hold choose the easy job. Effort levels in (3) are merely the optimal
levels, given the job chosen.
Because the expectation of e is zero, (3) can be rewritten as

a ⫹ b(A ⫹ me) ⫺ C(me) ! g ⫹ d(A ⫹ md) ⫺ C(md). (4)

If the condition in (4) holds, a worker prefers the difficult job. If it

does not, the easy job is selected.11 This is the same as the efficiency
condition, so workers choose jobs and effort efficiently under these
conditions. There is no distortion in effort choice. The worker inter-
nalizes everything. This is a simple problem of occupational choice with
effort. By contrast, were the firm to choose the job for the worker, then
effort would be distorted, although the surprising result is that the
distortion is as likely to take the form of underwork as it is of overwork.
This is shown in the Appendix.

B. Tournaments
The usual intuition that most have about promotions inducing atypically
high effort in period 1 comes from a tournament-like payment struc-
ture.12 When period 2 wages depend on the job rather than on the
output in the job, all workers put forth more effort than they would in
the absence of period 2 promotion concerns.
The intuition holds whether the tournament is against another player
or against a standard. For the purposes here, there is little difference
between competing against another player and competing against a
standard. In the Lazear-Rosen tournament structure, any level of effort
can be implemented for any standard simply by choosing the wage
spread appropriately. In the case of a standard, wages in period 2 are
fixed in advance and depend only on promotion. Even if workers receive
no wage prior to promotion, workers who are ignorant of A put forth
effort in order to maximize

max
m1
冕 {Wd Pr (A ⫹ m 1 ⫹ e1 1 A∗) ⫹ We[1 ⫺ Pr (A ⫹ m 1 ⫹ e1 1 A∗)]

⫺ C(m 1)}g(A)dA,

where Wd is the difficult job’s wage and We is the easy job’s wage.
11
In a competitive market with a rising supply price for workers (because they are
distinguished by ability), firms earn zero profit. The marginal worker is the one for whom
ability A0 is low enough that
a ⫹ b(A0 ⫹ m) ⫺ C(m) p 0.
12
Here again, Rosen is instrumental. The first paper on the subject is Lazear and Rosen
(1981).
S154 journal of political economy
The first-order condition is

冕[ (Wd ⫺ We)
⭸ Pr (A ⫹ m 1 ⫹ e1 1 A∗)
⭸m 1 ]
⫺ C (m 1) dG p 0

or
(Wd ⫺ We)g(A∗ ⫺ m 1 ⫺ A) p C (m 1).
The firm can obtain any level of effort, m 1, simply by setting the spread
between the difficult job wage and easy job wage appropriately. Then
it is necessary only to set the expected wage sufficiently high to attract
the marginal worker.
What is clear, however, is that effort in period 1 exceeds that in period
2. The tournament structure induces individuals to work at some positive
level in period 1 but to reduce effort in period 2. In this stylized model,
since there is no contingent reward in period 2, effort falls to zero. But
the general point is that the tournament against a standard creates
incentives to perform better in the prepromotion period than in the
postpromotion period.
Firms understand that their compensation schemes induce strategic
behavior by workers and set A∗ accordingly. Although this may mitigate
the effects of the behavior, it in no way changes the results of this section.
Since all derivations hold for any given A∗, they hold for the A∗ chosen
to take these effects into account.
As is the case of the tournament against a standard, workers put forth
more effort before the promotion decision than after the promotion
decision in a tournament against another player. This follows directly
from Lazear and Rosen (1981), where effort during the contest period
exceeds effort after the contest period. Worker effort during the contest
period is monotonically increasing in the spread between the winner’s
wage and the loser’s wage. After the contest has been decided, effort
falls off.
In both the tournament story and the regression to the mean story,
worker output declines after the promotion decision. In the tournament
context, the reason is that effort declines. In the regression to the mean
version, output declines because of the statistical proposition that en-
sures that winners do worse after promotion. There is a difference,
however. In tournaments, even losers reduce effort after the promotion
has been decided, so expected output for all workers falls over time. In
the statistical version, winners’ output falls and losers’ output rises above
their prepromotion levels on average.
To amplify this point, just as those who are promoted have higher-
than-average prepromotion transitory error, e1, so do those who fail to
be promoted have lower-than-expected transitory components. Other
the peter principle S155
things equal, this implies that those who do not get a promotion should
do better after being turned down than they did before. Thus faculty
who are denied tenure and move to other schools should do better, on
average, at those other schools than they did when they were assistant
professors at the first institution.
Observing this effect may be difficult for a number of reasons. For
example, a worker’s output might depend on the individuals with whom
he works. In an up-or-out context (Kahn and Huberman 1988), those
who fail to be promoted may find that the complementary factors in
the new job are not as productivity-enhancing as those in the first job.
Furthermore, motivation is an issue. To the extent that an individual
believes that he is in the running for promotion, tournament effects
are present, inducing effort. After the promotion has been denied, the
incentives vanish, reducing effort and output.

III. Other Examples of the Principle

The regression to the mean phenomenon that is observed as the Peter
principle in the labor market has other manifestations. For example, it
is often observed that sequels are rarely as good as the original movie
on which the sequel is based.13 If each movie is thought of as having a
theme-constant component, A, and a transitory component, et (e.g.,
actors, specific story, or direction), associated with each particular film,
then the same analysis holds. In order for a sequel to be made, the
value of the original film, A ⫹ et, must be estimated to be greater than
A∗. But given that the value exceeds the threshold level, A∗, the expec-
tation of the value of the sequel will be less than that of the original
simply because

E(e1Fsequel is made) 1 0,

but

E(e 2Fsequel is made) p 0.

As a result, an original film must be sufficiently good to generate a

sequel because studios, knowing that the second film is likely to be
inferior to the first, adjust upward their cutoff levels.
It is straightforward to test this proposition. Among other things, it
13
Data from Teichner and Luehrman (1992) establish clearly that revenues are lower
and costs higher on sequels than they are on the original on which they are based. See,
e.g., their exhibits 4 and 5.
S156 journal of political economy
implies that measures of film quality such as academy awards or ticket
sales should be higher on the original film than on the sequels.14
The fact that sequels do worse than originals is not evidence that the
studio made a mistake. Once again, if A∗ is chosen optimally, then the
average sequel is profitable, even if not as profitable as the original. A
decline in profitability does not imply that something went wrong in
decision making.
Similarly, the first meal in a good restaurant is often the best, followed
by less satisfying repeat visits. Just as above, think of A as the restaurant
component of the first meal (recipes, management) and e1 as the tran-
sitory component associated with the meal itself (that night’s chef, spe-
cific ingredients, dinner companion). A second visit to the restaurant
is made only if A ⫹ e1 1 A∗. Once again, the expected value of the second
meal lies below that of the first, conditional on the decision to make a
second visit to the restaurant. The larger the transitory component, the
larger the discrepancy between the first and second meals and the
higher the standard set to merit a second visit. As before, the fact that
the second visit is not as good as the first does not imply that a mistake
was made. The second meal, on average, is good enough to justify a
repeat visit to the restaurant, even if it is not as good as the first. The
point can also be used to explain why favorite restaurants go out of
fashion. A restaurant becomes a favorite in part because of the per-
manent component (e.g., good recipes and an insightful owner) and
in part because of potentially transitory components (e.g., the current
chef and the service of the staff). A favored restaurant can be thought
of as one that has gotten a draw of A ⫹ e1 1 A∗. It is favored precisely
because the value of the output exceeds some standard. Over time, e1
is replaced by transitory effects, the expectation of which is zero. The
quality falls and the restaurant goes out of fashion.
The same logic can be applied to the “Sports Illustrated effect.” It is
claimed that it is a curse to be on the cover of Sports Illustrated because
athletes’ performance falls thereafter. Again this reflects regression to
the mean, but it does not imply that Sports Illustrated chose the wrong
athlete for the cover.
In the political arena, reelections of elected officials occur when con-
stituents view performance as having exceeded some standard. The post-
election performance should be worse without learning and seniority
effects. If the learning effect is not too strong, second-term presidents
will do less well than they did during their first term. But this does not
imply that voters made a mistake. Even when voters know that there is
14
A countervailing effect is the notoriety that is created by the first film, which makes
it easier to sell tickets on the sequel than on the original. Even if consumers understand
that the sequel is worse than the original, more tickets might be sold on the sequel if,
say, the actors and director are not well known before the first film is made.
the peter principle S157
regression, voting for the incumbent is rational if he exceeds a suffi-
ciently high standard.
Finally, there is a close relation of this analysis to that of the winner’s
curse (Wilson 1969).15 If a bidder submits a bid based on a naive esti-
mate, the winner will lose money because, on average, the winner has
obtained a higher-than-average error. As a result, he shades his bid, so
in equilibrium he adopts the rent-maximizing strategy. The same is true
here. By adjusting the cutoff criterion, A∗, the decision maker maximizes
rent.

IV. Length of Probationary Period and Relative Importance of the

Transitory Component

The longer a firm waits to make a promotion decision, the better the
information. One would expect that transitory components that bias a
decision could be reduced or eliminated if the firm waited long enough
to make a promotion decision. The cost of waiting, however, is that
workers are in the wrong job for more of their lifetimes. For example,
suppose that it were possible to get a perfect reading on A by waiting
until the date of retirement. The information would have no value
because the worker would have spent his entire working career in the
easy job, even if he were better suited to the difficult job. The trade-off
is modeled. The conclusion is that as the variance of e1 rises, it becomes
more valuable to wait on a promotion decision.
To see this, let us add one period to the previous model (without
effort). Now, e1, e 2, and e3 refer to the transitory component in periods
1, 2, and 3, and assume that they are distributed independently and
identically and, to reduce notation, that E(A) p 0. Suppose that by
waiting two periods, an employer can obtain a perfect reading of A.
Under those circumstances, the optimum is simply to promote those
for whom A 1 x. The cost is that when the firm delays its promotion
decision to the end of period 2, all workers are in the easy job during
period 2 even though it might be better to place some in the difficult

15
Winner’s curse usually relates to a reading relative to others’ readings rather than
the time dimension of taking multiple readings, sometimes in different settings. The
transitory vs. permanent component is central to the discussion of this paper, but not the
theme of most of the winner’s curse literature. Actually, loser’s curse is as important to
the assignment problem as winner’s curse. In the job context, the goal is to assign a worker
to the right job. Workers who do not satisfy the promotion criterion are, on average,
undervalued just as those who are promoted are overvalued. The optimal selection of the
cutoff point trades off the two kinds of errors.
S158 journal of political economy
job in period 2. Expected output over the lifetime is then
expected output if wait

冕 冕
⬁ x

p 2[a ⫹ bE(A ⫹ e1)] ⫹ (g ⫹ dA)d ⫹ (a ⫹ bA)dFF

x ⫺⬁

冕 冕
⬁ x

p 2a ⫹ (g ⫹ dA)dF ⫹ (a ⫹ bA)dF. (5)

x ⫺⬁

The alternative is to make a decision after one period, using imperfect

information and recognizing that sorting will be imperfect. To make
things simple, assume that a firm that makes a promotion decision at
the end of period 1 cannot reevaluate at the end of period 2. The gain
is that workers are sorted early so that very able people can be put in
the difficult job more quickly. The cost is that more errors are made in
assigning workers to jobs. Then expected output over the three periods
is
expected output early

冕冕
⬁ ⬁

p [a ⫹ bE(A ⫹ e1)] ⫹ 2 (g ⫹ dA)dFdG

⫺⬁ A∗⫺e

冕冕
⬁ A∗⫺e

⫹2 (a ⫹ bA)dFdG
⫺⬁ ⫺⬁

冕冕 冕冕
⬁ ⬁ ⬁ A∗⫺e

pa⫹2 (g ⫹ dA)dFd ⫹ 2 (a ⫹ bA)dFdG. (6)

⫺⬁ A∗⫺e ⫺⬁ ⫺⬁

In an extreme case, it is clear that it pays to decide early. If the distri-

bution of e is degenerate so that there is no error, then (6) becomes

冕 冕
⬁ x

expected output early p a ⫹ 2 (g ⫹ dA)dFdG ⫹ 2 (a ⫹ bA)dF.

x ⫺⬁

(7)
The right-hand side of the expression in (7) must exceed the right-
hand side of (5) because
g ⫹ dA 1 a ⫹ bA for A 1 x
since that is how x is defined. Thus, when the variance in e shrinks to
zero, it always pays to promote early.
The example used earlier shows that it sometimes pays to defer the
promotion decision until the end of period 2. As before, let a p 1,
the peter principle S159
b p 0.5, g p 0, and d p 1, where the distributions of A and et are nor-
mal with variance equal to one. As shown earlier, the optimal cut point
is A∗ p 4.01. Then the right-hand side of (5) equals 3.004. The right-
hand side of (6) is 3.000. Thus deferring the promotion decision until
the second period pays when the variance in e is one. Other numerical
examples show that the advantage of deferring the promotion decision
becomes larger for higher variances in e.
The general point is that when the distribution of e is sufficiently
tight, it pays to make the promotion decision early. When it is sufficiently
diffuse, it pays to make the promotion decision later. Later promotion
decisions are more accurate but result in workers’ spending a longer
proportion of their work life in the wrong job.

V. Conclusion
Workers who are promoted have been observed to have exceeded some
standard. Part of the observation is based on lasting ability, but part is
based on transitory components that may reflect measurement diffi-
culties, short-term luck, or skills that are job-specific. As a result, there
is regression to the mean, creating a “Peter principle.” Workers who are
promoted do not appear to be as able as they were before the promotion.
Firms take this phenomenon into account in setting up their pro-
motion rule. Under general conditions, when fewer than 50 percent of
the workers are better suited to the high-level job, the firm adjusts the
promotion standard upward to compensate for the regression to the
mean. The amount of the adjustment depends on the tightness of the
error distribution. When the prepromotion error has high dispersion,
promotion standards are inflated by more than they are when the error
dispersion is low.
The statistical argument has been contrasted with incentive argu-
ments. Whether workers overproduce because they are gaming the sys-
tem depends on the payment structure. If, for example, output were
observable so that workers could be paid on the basis of output both
before and after the promotion decision, then it would be optimal to
allow workers to make their own job choice. Under these circumstances,
there is no distortion in effort; all is efficient. When a tournament
structure is chosen because of inability to observe output, workers pro-
duce more before promotion than they do after promotion. Although
tournaments result in declining output after promotion, the implica-
tions of tournaments for losers and winners are different from those of
the statistical argument. In particular, in tournaments, the output of
both losers and winners falls after promotion. The statistical argument
implies that losers’ output rises and winners’ output falls after
promotion.
S160 journal of political economy
The Peter principle can be interpreted to mean that workers are not
as able as perceived before promotion or that they were better in their
prior job relative to their peers than they are in their current one. In
a multilevel firm, the typical worker who remains at a given level is
“incompetent” in that he is not as good as the average worker coming
into the job, nor is he as good as he was in his previous assignment
relative to the comparison set.
One way to offset the Peter principle is to wait for a longer time
before making a promotion decision. The advantage is that the job
assignment is better than it would have been had the decision been
made earlier. The disadvantage is that able workers remain in the wrong
job for a longer period of time.
The logic of the Peter principle applies in other contexts as well. The
regression to the mean phenomenon implies that movie sequels are
lower-quality than the original films on which they are based and that
excellent restaurant meals are followed by ones that are closer to
mediocre.

Appendix

Effort Is Distorted When Firms Assign Jobs but Workers Know Ability

Assume asymmetric information in which workers know their abilities, A, but

firms do not. Define m 1 as effort in period 1, m 2e as effort in period 2 if the
worker is not promoted, and m 2d as effort in period 2 if the worker is promoted.
Note that effort in period 1 is determined before the promotion decision is
made, so period 1 effort is independent of promotion. The cost of effort is given
by C(m). For simplicity, C(m) is assumed to be independent of ability and the
same across periods.
The worker is paid a piece rate, so in period 2, a worker who has not been
promoted chooses effort m 2e so as to solve

max a ⫹ bE(A ⫹ m 2e ⫹ e 2) ⫺ C(m 2e)

m 2e

or

max a ⫹ b(A ⫹ m 2e) ⫺ C(m 2e).

m 2e

The first-order condition is

C (m 2e) p b. (A1)

An analogous problem can be solved for those who are promoted. Their
problem is

max g ⫹ d(A 2 ⫹ m 2d) ⫺ C(m 2d),

m 2d
the peter principle S161
which has first-order condition
C (m 2d) p d. (A2)
Equations (A1) and (A2) define m 2e and m 2d . Promoted workers put forth more
effort in period 2 because the marginal return to effort is higher in the difficult
job than in the easy job, that is, d 1 b . Given this, the worker solves a two-period
problem in period 1, knowing that he will choose m 2d and m 2e depending on
whether or not he is promoted.
The worker who knows his own ability has a first-period problem given by16
max a ⫹ bE(m 1 ⫹ A ⫹ e1) ⫺ C(m 1)
m1

⫹ Prob(A ⫹ m 1 ⫹ e1 1 A∗)E[g ⫹ d(A ⫹ m 2d ⫹ e 2) ⫺ C(m 2d)]

⫹ Prob(A ⫹ e1 ⫹ m 1 ≤ A∗)E[a ⫹ b(A ⫹ m 2e ⫹ e 2) ⫺ C(m 2e)]
or
max a ⫹ b(m 1 ⫹ A) ⫺ C(m 1) ⫹ [1 ⫺ G(A∗ ⫺ m 1 ⫺ A)][g ⫹ d(A ⫹ m 2d)
m1

⫺ C(m 2d)] ⫹ G(A∗ ⫺ m 1 ⫺ A)[a ⫹ b(A ⫹ m 2e) ⫺ C(m 2e)]. (A3)

The first-order condition is
b ⫺ C (m 1) ⫹ g(A∗ ⫺ m 1 ⫺ A){[g ⫹ d(A ⫹ m 2d) ⫺ C(m 2d)]
⫺ [a ⫹ b(A ⫹ m 2e) ⫺ C(m 2e)]} p 0. (A4)

Efficient effort is supplied when workers set C (m 1) p b . According to the first-
order condition in (A4), this occurs only when the last term on the left-hand
side is equal to zero. In general, it will not be zero. In fact, the last term is
positive, implying overinvestment, when
g ⫹ d(A ⫹ m 2d) ⫺ C(m 2d) 1 a ⫹ b(A ⫹ m 2e) ⫺ C(m 2e).
Sufficiently high-ability workers prefer job 1 because they earn more in job 1.
As a result, they overwork in period 1 to enhance the probability that they will
be promoted. Because the firm cannot distinguish effort from ability, workers
who want to be promoted have an incentive to work too hard in order to fool
the firm into believing that their ability levels are higher than they actually are.
Less intuitive, the converse is also true. Low-ability workers, that is, those for
whom A is sufficiently low so that
g ⫹ d(A ⫹ m 2d) ⫺ C(m 2d) ! a ⫹ b(A ⫹ m 2e) ⫺ C(m 2e),
17
underwork. These workers underachieve because they do not want to take the
chance of being promoted. From their point of view, a promotion is bad because
they are likely to earn less in the difficult job than in the easy job.
The workers who are most likely to distort their effort in period 1 are those
for whom g(A∗ ⫺ m 1 ⫺ A) is high (see eq. [A4]) and for whom
F[g ⫹ d(A ⫹ m 2d) ⫺ C(m 2d)] ⫺ [a ⫹ b(A ⫹ m 2e) ⫺ C(m 2e)]F

16
The discount rate is assumed to be zero.
17
Since m2 is independent of A, there is always an A sufficiently low to make this condition
hold.
S162 journal of political economy

Fig. A1

is high. Under standard assumptions about the distribution of e, in particular

that
lim g(e) p 0,
er⫺⬁,⬁

very high- and very low-ability workers choose the efficient level of effort in
period 1. They have little to fear in terms of incorrect promotion decisions. The
extremely able are almost certain to be promoted, so that extra effort has very
little effect on the probability of promotion. Conversely, the totally inept are
almost certain to avoid promotion, so that reducing effort has almost no effect
on lowering the probability of promotion.
Also true is that those whose underlying ability is very near the efficient job
switch point (x in fig. 1) do not distort effort much. Even if they are misclassified,
they have little to lose. Define A 0 such that
g ⫹ d(A 0 ⫹ m 2d) ⫺ C(m 2d) p a ⫹ b(A 0 ⫹ m 2e) ⫺ C(m 2e).
Then excess effort is zero at both extremes and also at A 0 , which is likely to be
close to x. The pattern of distortion is shown in figure A1. Those at A 0 do not
distort at all. Those at the ability extremes do not distort. Those with ability less
than the switch point underwork and those with ability more than the switch
point overwork.

References
Anderson, Rolph E., Alan J. Dubinsky, and Rajiv Mehta. 1999. “Sales Managers:
Marketing’s Best Example of the Peter Principle?” Bus. Horizons 42 (January/
February): 19–26.
Baker, George, Michael Gibbs, and Bengt Holmstrom. 1994. “The Internal Eco-
nomics of the Firm: Evidence from Personnel Data.” Q.J.E. 109 (November):
881–919.
Fairburn, James A., and James M. Malcomson. 2001. “Performance, Promotion,
and the Peter Principle.” Rev. Econ. Studies 68 (January): 45–66.
the peter principle S163
Faria, Joao Ricardo. 2000. “An Economic Analysis of the Peter and Dilbert Prin-
ciples.” Manuscript. Sydney: Univ. Technology.
Gibbons, Robert, and Michael Waldman. 1999a. “Careers in Organizations: The-
ory and Evidence.” In Handbook of Labor Economics, vol. 3B, edited by Orley
C. Ashenfelter and David Card. Amsterdam: North-Holland.
———. 1999b. “A Theory of Wage and Promotion Dynamics inside Firms.” Q.J.E.
114 (November): 1321–58.
Gibbs, Michael, and Wallace Hendricks. 2001. “Are Formal Salary Systems a
Veil?” Manuscript. Chicago: Univ. Chicago.
Jovanovic, Boyan. 1979a. “Firm-Specific Capital and Turnover.” J.P.E. 87 (De-
cember): 1246–60.
———. 1979b. “Job Matching and the Theory of Turnover.” J.P.E. 87, no. 5, pt.
1 (October): 972–90.
Kahn, Charles, and Gur Huberman. 1988. “Two-Sided Uncertainty and ‘Up-or-
Out’ Contracts.” J. Labor Econ. 6 (October): 423–44.
Lazear, Edward P. 1986. “Raids and Offer Matching.” In Research in Labor Eco-
nomics, vol. 8, pt. A, edited by Ronald G. Ehrenberg. Greenwich, Conn.: JAI.
———. 1992. “The Job as a Concept.” In Performance Measurement, Evaluation,
and Incentives, edited by William J. Bruns, Jr. Boston: Harvard Bus. School
Press.
Lazear, Edward P., and Sherwin Rosen. 1981. “Rank-Order Tournaments as Op-
timum Labor Contracts.” J.P.E. 89 (October): 841–64.
Medoff, James L., and Katharine G. Abraham. 1980. “Experience, Performance,
and Earnings.” Q.J.E. 95 (December): 703–36.
Peter, Lawrence J., and Raymond Hull. 1969. The Peter Principle: Why Things Always
Go Wrong. New York: Morrow.
Prendergast, Canice. 1999. “The Provision of Incentives in Firms.” J. Econ. Lit-
erature 37 (March): 7–63.
Rosen, Sherwin. 1986. “Prizes and Incentives in Elimination Tournaments.”
A.E.R. 76 (September): 701–15.
Teichner, William A., and Timothy A. Luehrman. 1992. “Arundel Partners: The
Sequel Project.” Harvard Business School Case no. 9-292-140. Boston: Harvard
Bus. School Pub.
Wilson, Robert B. 1969. “Competitive Bidding with Disparate Information.” Man-
agement Sci. 15 (March): 446–48. Reprinted in The Economics of Information, vol.
1, edited by David K. Levine and Steven A. Lippman. Aldershot, U.K.: Elgar,
1995. Also reprinted in The Economic Theory of Auctions, vol. 2, edited by Paul
Klemperer. Cheltenham, U.K.: Elgar, 2000.