Complements?
Xu Jiang*
Duke University
September 2016
*: Accepted by Haresh Sapra. Xu Jiang can be reached at xu.jiang@duke.edu.The paper develops one of
the ideas but is substantially different from an earlier paper Accounting Conservatism and Debt Contract
Efficiency: The Role of Soft Information, which was presented at workshops and conferences in various
places. I thank participants at those workshops for helping me to develop this paper. I appreciate valuable
discussions with Qi Chen, Frank Gigler, Chandra Kanodia, Mingzhu Li, Baohua Xin and Ming Yang and
valuable comments from an anonymous referee on various versions of this paper. I acknowledge financial
support from Duke University Fuqua School of Business. All errors are my own.
Complements?
Abstract
This paper studies how bias in non-accounting and in accounting information should be
in the relative information content of accounting numbers. The optimal bias in one type of
information is shown to be a complement of the bias in the other type. This result can be applied
This paper examines the optimal relation between the properties of accounting information
specifically, the nature of biases in it and the properties of other information. 1 This question is
as long as they are not completely uninformative, remain incrementally beneficial to decisions,
depending in part on the properties of the information, including the potentially non-trivial bias
embedded in it. Thus, it is not clear, a priori, how the bias in accounting and in non-accounting
information should be related from a decision-making standpoint. This is the theme of this paper.
I examine a setting in which the decision-making role of information arises naturally. In the
accounting and non-accounting information signals. Based on those signals, the investor decides
whether to continue or to liquidate the project. This problem thus resembles a stylized depiction
of decision-making under uncertainty in which the investors expected payoff depends crucially
on the quality of the information on which the decision is based. This in turn depends on the
properties, including biases, of the systems that generate the accounting and the non-accounting
information signals. Consequently, the relationship between the bias in the accounting
information and that in the non-accounting information is non-trivial. One can think of
accounting information as earnings numbers that indicate future cash flows and of non-
accounting information as another noisy signal of future cash flows that is free of the bias
generated by the accounting system but may contain bias of its own.
among others, by Gigler and Hemmer (2001), Venugopalan (2004), Chen et al. (2007) and
1
In this paper other information and non-accounting information are interchangeable.
2
Gigler et al. (2009). An information system is considered to be conservatively biased if the
favorable signals are more informative about the underlying fundamental than the unfavorable
ones, aggressively biased if the reverse holds. 2 In seeking to determine optimal bias in the
accounting information system, one may conjecture that its signals should counterbalance the
differential biases in the non-accounting signals. Such a conjecture would imply that optimally,
I show that from the standpoint of social efficiency, when the accounting information is
conservative, then conservative non-accounting information too is desirable; that is, favorable
non-accounting signals should be more informative than unfavorable ones. Similarly, when the
aggressive; i.e. its unfavorable signals should be more informative than the favorable ones. In
other words, the bias in accounting information and the bias in non-accounting information
should be complements.
To understand the intuition underlying this central result, note that the function of additional
information is to reduce the decision errors that the investor would have made relying on a single
information system. To start, suppose the decision whether to liquidate or to continue a project is
based on accounting information only. If this accounting signal is informative, it will be optimal
to liquidate upon observing a low (unfavorable) signal and continue when there is a high
(favorable) signal. Therefore, if the accounting information is conservatively biased, so that low
signals are less informative than high signals, decision errors are more likely when the
accounting signal is low. Now consider supplementing the accounting signal with non-
2
In this paper I use the terms favorable and high interchangeably to refer to good news as per Milgrom (1981)
and unfavorable and low interchangeably to refer bad news.
3
accounting information. This is most useful when it leads to the largest reduction in decision
errors. Since decision errors stemming from the accounting information are more likely when the
accounting signals are low (i.e., leading to inefficient liquidation decisions), the non-accounting
information is most valuable when it induces the investor to continue the project despite the low
accounting signals. This corresponds to the case when high non-accounting signals are more
informative, the defining characteristic of a conservatively biased system. The same intuition
implies that when the non-accounting information is conservatively or aggressively biased, the
This result, applied in different settings, carries interesting implications. For example, if the
non-accounting information signals are subject to managerial manipulation, then they are likely
to be aggressive, since favorable signals are less informative. On the basis of the conclusions
reached here, then, in this case accounting information should be aggressively rather than
conservatively biased. This runs counter to LaFond and Watts (2008) who argue that
conservative accounting is needed to overcome the optimistic bias of other information that is
accounting information and that other information signals are aggressive, in the sense that high
signals are less informative than low signals. Our conclusion is that if the accounting information
is conservatively biased, investors are less likely to process it, i.e. they under-react to it. This
This paper contributes to the accounting and information economics literature along several
dimensions. First, it sheds light on the basic question of the optimal relation between bias in the
accounting and non-accounting information. For example, Chen et al. (2007) and Gao (2013)
3
See section III for more discussions of the implications.
4
show that conservative accounting standards are desirable when managers can manipulate the
signal to which the accounting standards apply. My paper shows that when accounting signals
manipulate other signals in a similar manner, as in Chen et al. (2007) and Gao (2013).
Second, the focus on the bias instead of the accuracy of information generates different,
more general implications than previous works that posit information as unbiased noisy signals
of some fundamental and focus on their optimal precision. 4 The focus on bias, which is one of
the qualitative properties of information, also connects with the growing literature on rational
inattention and flexible information acquisition (e.g., Sims [1998, 2003], Yang [2015]), whose
main insight is that to economize on the use of information, agents will focus on the information
that is relevant to their welfare and be rationally inattentive to other aspects. This paper generates
similar insights in an accounting setting. Given biased accounting information, investors will
prefer the bias in the non-accounting information that is most relevant to reducing decision errors
induced by the bias in the accounting information. The contribution to the literature consists in
The remainder of the paper is organized as follows. Section II sets up the model, solves for
the optimal decision rule, and presents the main result on the optimal relation between the bias in
accounting information and the bias in non-accounting information. Section III discusses the
implications of the main result and Section IV concludes. Technical details, including all proofs,
4
A non-exhaustive list of exceptions includes Gigler and Hemmer (2001), Kwon et. al. (2001), Chen et. al. (2007),
Gigler et. al. (2009), Bertomeu et. al. (2014), Li (2013) and Gao (2013).
5
II. The Model
Consider a risk-neutral investor at date 0 with an investment project that will return an
uncertain cash flow two periods later. 5 The investors prior on has binary support { , }
1. 6 For notational ease, throughout the paper refers to non-accounting information, although
both and can be more broadly interpreted as two noisy signals of the fundamentals, the noise
being reciprocally independent. As is made clear below, I allow the noise in and to be biased
and study how the property of the noise in affects that of the noise in .
At date 1, after observing the accounting signal and the non-accounting signal , the
investor chooses whether to liquidate the project for proceeds or to continue it until date 2. In
the latter case, the cash flow is realized at date 2. To make the problem interesting, I assume that
() > > () , where is the support of which, in the binary case, reduces to
> > .
As noted earlier, I focus on a single investors decision problem, in which the liquidation
and continuation decisions can be interpreted more broadly as actions that affect the investors
payoff (e.g., selling shares in an intermediate period versus holding them until the end). The fact
that liquidation generates a certain rather than an uncertain payoff can be justified by a debt
contracting setting. However, the stylized decision-making setting used here reflects an
5
Assuming risk aversion would not qualitatively change the results.
6
In the binary case, subscript refers to favorable realization of any signal while subscript refers to unfavorable
realization.
6
important characteristic of any setting with a continuation/termination decision, including selling
and ( , ).
I also assume that the project has positive net present value, i.e. [] . In the binary
( )
case, it is equivalent to ( ) ( ). Further, denote 1. 8
( )
Following prior literature, such as Gigler and Hemmer (2001), Venugopalan (2004), Chen et al.
(2007), Bertomeu et al. (2014) and Li (2013), the bias in accounting information, denoted by ,
( | ) = + , ( | ) = 1 and
( | ) = 1 + , ( | ) = .
where (+1 , 1), [0 , 0 ] (+1 , 1 ) for some 0 > 0.
The probability structure and the assumptions imply that is good news relative to in
( | ) ( | )
the sense of Milgrom (1981) as > . and have the usual interpretation of the
( | ) ( | )
7
I thank an anonymous referee for making this point. In addition, the assumption of a certain payoff of upon
termination, while appropriate for a debt-contracting setting, may not be for other settings. When is uncertain but
its distribution is independent of the bias in both the accounting and the non-accounting signal, it is straightforward
to show that the main conclusion is unaltered. Whether it still holds when is uncertain and the distribution
depends on the signal biases is an interesting question left for future work.
8
The assumption of 1 has also been used in prior work, e.g. Gigler et al. (2009). The justification for this
assumption is that the project must be so promising that absent other information, a rational decision maker will not
want to start the project with the intention of liquidating it in the intermediate period. Results when < 1 mirror
those for > 1 and will be mentioned in footnotes wherever relevant.
9
Throughout the paper is used to denote probabilities.
7
( | )
informativeness and bias of the accounting signals. To see this, note that both and
( | )
( | )
are increasing in . Thus, when increases, both and become more informative.
( | )
( | ) (| )
On the other hand, is decreasing but is increasing in . Thus, when increases
( | ) ( | )
and accounting becomes more aggressive (or less conservative), becomes less informative but
is more informative.
The assumptions imposed on and warrant some discussion. First, as is noted in Lemma
A.1 in the appendix, the investor, upon seeing a signal , will continue (liquidate) the project
(| ) 1 ( | )
when (| )
> (<) . The assumptions that > +1 and > +1 result in ( | )
>
1 ( | ) 1
and ( | ) < . In other words, these two assumptions ensure that regardless of the
bias in the accounting information, the investor will liquidate on seeing and continue on
seeing if only accounting information is available. The assumptions rule out the uninteresting
cases in which the investor will always continue or always terminate regardless of , implying
that the accounting information is always informative. Second, when = 0 this captures the case
of neutral accounting with no bias when the informativeness of the favorable signal is equal to
( | ) ( | )
that of the unfavorable signal, i.e. = 1 = . The case of < 0 accordingly
( | ) ( | )
( | ) + ( | )
greater than that of the unfavorable, i.e. = > 1 > = .
( | ) 1 + 1 ( | )
Similarly, > 0 corresponds to the case of aggressively biased accounting, where the
( | ) + ( | )
unfavorable signal is more informative, i.e. = > 1 > = .
( | ) 1 1 ( | )
Following the notion in Gigler et al. (2009), I therefore define an accounting system as
8
conservatively (aggressively) biased if < (>)0. Finally, the bias is assumed to lie in a
symmetrical interval around zero, ensuring that the favorable signal in the most conservatively
biased accounting system is equally informative as the unfavorable signal in the most
aggressively biased accounting system and conversely, i.e. that the unfavorable signal in the
most conservatively biased accounting system is equally uninformative as the favorable signal in
the most aggressively biased accounting system. In other words, the focus is on the effect of bias
( | ) = + , ( | ) = 1
and ( | ) = 1 + , ( | ) =
where (+1 , 1) and [0 , 0 ] (+1 , 1 ) for some 0 > 0
Similarly, the probability structure and the assumptions imply that is good news
( | ) ( | )
relative to in the sense of Milgrom (1981) as > . and have the usual
( | ) ( | )
interpretation of the informativeness and bias of the non-accounting signals, respectively. To see
( | ) ( | )
this, note that both and are increasing in . Thus, when increases, both
( | ) ( | )
( | ) ( | )
and become more informative. On the other hand, is decreasing but is
( | ) ( | )
10
Note that while the literature has generally interpreted as informativeness and as bias, distributions with same
but different in fact have different levels of informativeness as gauged by common measures from information
theory, such as mutual information (see, e.g., Sims 2003 for more details on mutual information). Therefore I control
for the sum of + to ensure that the informativeness of the most conservatively biased signal is neither greater nor
less than that of the most aggressively biased signal. An alternative way of modelling the bias is to impose a
different probability structure ensuring that the variation in does not alter the informativeness of the distributions,
which is outside the scope of this paper.
9
increasing in . Thus, when increases and non-accounting information becomes more
aggressive (or less conservative), becomes less informative but is more informative.
Again, the assumptions on and ensure that the non-accounting information is always
useful to the decision if only the non-accounting information is available. Analogously with the
discussion of the bias in accounting information, = 0 represents the case in which the non-
accounting information is neutral, i.e. the favorable signal is equally informative as the
unfavorable. The case < 0 ( > 0) occurs when the non-accounting information is
conservative (aggressive), i.e. the favorable signal is more (less) informative than the
unfavorable.
The subsequent discussions focus on the case = in order to isolate the effect of
= , , which are functions of both and . We now address the question of the optimal
We first look at the optimal decision rule based on and , which is specified by {(, )}
11
Relaxing this assumption does not alter the results qualitatively, but it makes the algebra much messier.
10
Program A:
{(,)} (, ) + [ ( |, ) + ( |, )](, )
(,)(,) (,)(,)
In Program A, the investor chooses the optimal decision rule to maximize his payoff.
The solution for the optimal relation between the bias in accounting and non-accounting
information is in three steps. First, solve for Program A and, correspondingly, the optimal
decision rule. Second, based on the optimal decision rule, write the maximized payoff, which is a
function of both and , in terms of the probabilities of terminating and continuing the project
explicitly. Finally, solve for the optimal that maximizes this maximized payoff given and the
optimal that maximizes this maximized payoff given . The results of the first step are
summarized in Proposition 1.
Proposition 1 The optimal contract that solves Program A results in termination of the
Proposition 1 is intuitive. Any deviation from the efficient decision rule will lower the
expected payoff insofar as otherwise the rule will not be efficient. Although the decision rule is
always efficient, its effectiveness depends on how effective the available information is in
reducing decision errors, which in turn depends on its informational properties, i.e. and .
Given the optimal decision rule and the foregoing distribution assumptions, we can write out
the respective probabilities of termination and continuation, and hence the objective function,
explicitly as the second step. Lemma 1 sets out an intuitive characterization of the objective
12
The project can be either terminated or continued when (|, ) = . Assuming either decision does not affect
any of the results.
11
where (, ) = {(,):[(|,)<]} ( )(, | , , )
Lemma 1 says that the optimal decision rule must necessarily minimize the sum of decision
errors due to signal noise. The first term of represents the error of terminating a good project
(the false alarm error); the second term represents the error of continuing a bad one (undue
optimism error). Clearly, is a function of and . The final step examines how changes
From Proposition 1 and the imposed assumptions, when only one of them is available the
unfavorable ( or ). This implies that when both accounting and non-accounting information is
available, the project is always continued when ( , ) is observed (i.e. good news from both
systems) and terminated in case of ( , ), i.e. bad news from both systems, regardless of bias.
Before presenting the main results, let us discuss a special case that better illustrates the
1
underlying intuition. In this case = 1, i.e. the project has zero net present value and (2 , 1);
the following lemma characterizes the decision rule when either ( , ) or ( , ) occurs.
Lemma 2. Assume that K=1. If + < (>)0, then the project is continued (terminated)
The lemma shows that in the two cases with conflicting signals the decision whether to
continue or terminate depends on the net bias of the accounting and the non-accounting
information, i.e. + . To see this, consider the case of < 0 , i.e. conservatively biased
accounting information. When < 0 or at most slightly positive, so that + < 0, then Lemma
2 implies that the project is continued when either ( , ) or ( , ) occurs. This is because
12
is highly informative when generated by a conservatively biased accounting system and is not
very informative when the non-accounting system is either conservatively biased (i.e < 0) or
just slightly aggressively biased (i.e. 0 < < ), so that the investor uses the more informative
and ignores the less informative . Similarly, since is not very informative when generated
biased (i.e. < 0) or only slightly aggressive (i.e. 0 < < ), the investor ignores the less
informative and continues the project on the basis of the more informative . But when is
very large, so that + > 0, the non-accounting information is so aggressively biased that
becomes more informative than and less informative than . In this case, the investor
relies on the more informative signals and , resulting in termination given either ( , ) or
( , ).
Once the decision rule has been determined, the benchmark case set forth in Lemma 3
below shows that when accounting (non-accounting) information is neutral, the investor is
Lemma 3. Assume that K=1. When = 0 , the optimal bias of the non-accounting
Lemma 3 illustrates that when one information system is neutral, then from the decision-
making perspective the other system can be either conservatively biased or aggressively biased.
The intuition is that when one system is neutral, its good-news signals are just as informative as
its bad-news signals. The errors generated by the high and low signals are therefore the same,
resulting in indifference between the other signal being more informative for high or for low
13
realizations. An interesting result from Lemma 3 is that the optimal bias of the other information
system is either maximally conservative or maximally aggressive when the existing information
system is neutral. To see the underlying intuition, note that from Lemma 2 when K=1 and
implying that lots of decision errors will be generated in those scenarios. Adding bias into either
direction will result in investors more certain about which direction to take when facing ( , )
and ( , ), thus reducing decision errors. The more bias added, the more confident investors
Lemma 4 shows the result when one of the two information systems is biased.
Lemma 4. Assume K=1. If the accounting system is conservatively biased, i.e. < 0, it is
accounting system is aggressively biased, i.e. > 0, it is optimal for to be aggressive (optimal
it is optimal for the accounting system too to be conservatively (aggressively) biased. In other
Only the intuition of the first part need be explained here; that behind the second part is
analogous. Note that is most valuable when it generates information that conflicts with that
generated by , information that ends up changing the decision that the investor would otherwise
informative than . Thus, when is the only information, the decision to terminate (based on )
is likely to contain larger errors than the decision to continue (based on ). When is available,
Lemma 2 shows that it can be optimal either to continue or to terminate upon observing ( , )
14
because any improvement in the efficiency of the original decision thanks to adding the
observation of to must be relatively minor, in that the original decision based on the
informative has relatively less error to begin with. By contrast, a decision based on the less
informative is more likely to contain errors; that is, a more informative significantly
Similarly, with an aggressive accounting system the additional information signal is most
valuable in the case ( , ) where it changes the decision from continuation (without ) to
termination (observing ). Since is less informative with aggressive accounting, this change
should be aggressive.
We can now present the main result, demonstrating that the result set out in Lemma 4
< 0, then it is optimal for too to be conservative; that is, the optimal is negative. If the
accounting system is aggressively biased, i.e. > 0, then it is optimal for to be aggressive
when 0 and are sufficiently large. Similarly, when the non-accounting system is
is aggressively biased, then it is optimal for accounting to be aggressive when 0 and are
sufficiently large. 13
Note that the conclusion of Proposition 2 is essentially the same as that of Lemma 4, with
just one difference: the conditions under which complementary bias is optimal are stronger in
13
When < 1, if > 0, then it is optimal for to be aggressive. If < 0, then it is optimal for to be conservative
when the magnitude of 0 and are sufficiently large. Similarly, when > 0, it is optimal to have an aggressively
biased accounting system, while if < 0, then it is optimal for accounting to be conservative when 0 and are
sufficiently negative.
15
the case of aggressive than of conservative bias, in that in the former case the range of bias must
be sufficiently large. To understand the difference, look again at the case in which one of the
information systems is neutral. The following lemma shows that when > 1 and accounting is
Lemma 5. Assume that K>1. If the accounting system is neutral, i.e. = 0, then it is
socially optimal for to be conservative, i.e. the optimal is negative. If the non-accounting
system is neutral, i.e. = 0, then it is socially optimal for to be conservative, i.e. the optimal
14
is negative.
The intuition for lemma 5 15 is that when > 1, the prospects are so good that it is more
costly to terminate the project than to let it continue. Thus a socially optimal decision rule will let
the project continue as much as possible, i.e. also when and transmit conflicting signals.
Since is neutral, and are equally informative; the decision therefore depends on the
Note that Lemma 5 does not contradict the conclusion of Gigler et al. (2009), who find that
if only accounting information is available, then when > 1 the optimal bias must necessarily
be aggressive, because continuation generates fewer decision errors. In the present setting their
result still holds. When only one piece of information is available, the optimal bias can be shown
14
When < 1, if = 0, then it is socially optimal for to be aggressive. If = 0, then it is socially optimal for
to be aggressive.
15
Again we explicate the intuition for the first part only; that for the second part is essentially the same.
16
to be aggressive; but when both pieces are available and one is neutral, the other piece should be
conservatively biased for the same reason: it maximizes the likelihood of continuation, thus
Lemma 5 says that conservative is preferable when > 1 and is unbiased. So when is
aggressive, the benefit from aggressive has to be large enough to overcome the advantage of
conservative bias driven by > 1 , which explains why and have to be sufficiently
The main result is for a decision problem where the bias has to be chosen ex-ante, i.e. before
observing the realization of any signals. But the information choice problems can also be posed
ex-post, i.e. after observing some realizations. This sub-section shows that the main intuition
For algebraic simplicity, let us take the case where = 1 (the insight does not change
substantially when > 1). Assume that the investor can observe the realization of before
Corollary 1. Assume that K=1. If the investor observes , he will prefer an aggressively
biased non-accounting system (i.e. choose 0 ) if 0 > . If the investor observes , he will
The conditions concerning 0 and guarantee that the bias in is great enough to change
the decision when is observed; for otherwise the observation of changes nothing and the bias
in is irrelevant to the decision. Note that while Corollary 1 may look different from Lemma 4
16
I thank an anonymous referee for suggesting this extension.
17
or Proposition 2, the intuition behind them all is the same. When is observed, an unfavorable
that is not very informative will not be much help in informing decision makers to terminate,
because the realization of itself informs investors that the project is more likely to succeed.
A similar intuition would suggest that conservative is preferred when is observed. Ex-ante,
given that a conservatively biased accounting system generates more often than , a
be preferred more often. Corollary 1 is thus consistent with the main result.
In the main analysis, the property of is chosen by an investor in such a way as to maximize
the social efficiency of decision-making. In reality, though, we may have different parties with
conflicting interests, and the bias may not be chosen for that purpose. Specifically, a manager
may have interests that conflict with the shareholders say, private benefits from continuing the
project yet still be able to choose the property of . LaFond and Watts (2008) argue that
managers tend to be overly optimistic in disclosing other information, so the authors advocate
conservative accounting to help investors discount that bias. Here I show that this argument is
not necessarily correct, because the manager has an incentive to manipulate and add an
aggressive bias to non-accounting information, which means that accounting information should
In particular, I assume a risk-neutral manager in charge of the project. The investor can sign
a debt contract with the manager, specifying not only the face value of the contract but also the
conditions under which the project is to be continued or liquidated in the intermediate period.
18
The manager enjoys a non-pledgeable private benefit > 0 from continuation. 17 I also assume
that + > and that + < , so that it is socially optimal to continue when the true
cash flow is and terminate when it is . I further assume that the manager can manipulate
at no cost by shifting its distribution to the right, in the sense of first order stochastic dominance
up to some maximum amount. 18 I use [0, ] to denote the managers ex-ante manipulation,
with higher values of representing sharper upward manipulation. Thus, a high value of
makes it more likely that low realizations of will be informative and high realizations
First let us look at the optimal contract based on and . The contract is specified by
the investor takes control when (, ) (, , ) and the manager takes control otherwise. Note
that the choice of (, , ) also depends on the investors conjecture concerning the managers
choice of . 20 Without loss of generality I assume that the investor has all the bargaining
power. 21
Given the assumptions on the parameters, when the investor is in control he liquidates and
obtains , while the manager gets nothing. When the manager is in control, she continues the
project and receives + if it succeeds and if it fails. The investor gets if the project
17
This modeling approach follows Aghion and Bolton (1992) and Gao (2013).
18
The type of ex-ante manipulation technology posited in this paper has been used in Dye (2002), Chen et. al. (2007)
and Gao (2013), among others. Chen et. al. (2007) also assume zero cost for earnings manipulation up to some
maximum amount.
19
The ability to verify with some probability does not change the result qualitatively, as long as the penalty that
can be imposed on the manager is not too heavy.
20
The dependence of on is suppressed, as in the optimization problem is fixed.
21
Assuming that the manager has all the bargaining power or postulating a Nash bargaining outcome does not
change any of the results. As is shown below, the contract results in a Pareto-optimal decision rule and is therefore
renegotiation-proof.
19
succeeds and if it fails. Thus the managers payoff is
manager chooses = .
Program B:
{,(,,),} (, , )
(,)(,,)
+ [( |, , ) + ( |, , )] (, , ) . .
(,)(,,)
= (Rational Expectations)
In Program B, the investor chooses the contract that maximizes his payoff, subject to three
constraints: the (IR) constraint, requiring that the manager receive at least his reservation utility;
the (IC) constraint, that the manager chooses his optimal action based on the investors
conjecture; and the (Rational Expectations) constraint, i.e. that the conjecture correspond to the
Proposition 3 The optimal contract that solves Program B results in termination of the
manager chooses = .
22
The project can be either terminated or continued when ( + |, ) = . Assuming either decision will not
affect any of the results.
20
The logic underlying Proposition 3 is as follows. The manager always wants to continue and
the investor always wants to terminate, regardless of the signal. Since the socially optimal
solution is to continue when the signal realization is sufficiently good and terminate when it is
sufficiently bad, the optimal decision rule would give the power of decision to the manager when
the signals are good enough, so that he would continue. In the same way, when the signals are
sufficiently bad it would give the power of decision to the investor, in order to terminate.
Proposition 3 shows that the manager would manipulate the non-accounting signal. The
intuition is that whatever the investors belief about the distribution of , the manager has an
incentive to shift the distribution of rightward, because this increases the chances of
continuation. This means that the only sequential equilibrium is for the investor to conjecture
that the manager will choose = and the manager in fact does. 23 This is consistent with the
thesis of LaFond and Watts (2008) that managers tend to overstate information that they can
manipulate. It is also consistent with Gao (2013), who argues that managers have an incentive to
The fact that managerial manipulation results in an aggressively biased , combined with
Proposition 2, gives us Corollary 2, the proof of which is omitted as it follows directly from the
proposition itself.
In other words, the managers tendency to overstate other information makes less
managerial manipulation conservative accounting might have other benefits, my results suggest
23
Contracting on a report of will not solve the problem, as the manager can always claim to choose = 0. If the
investor naively believes that = 0, the manager still has the incentive to choose = 0 . Knowing this, the
investor will never believe the managers report and we are back to Program B.
21
that when managers overstate the non-accounting information conservative accounting may be
inefficient.
First, Lemma 4 and Proposition 2, if applied in a different setting, also provide a rational
explanation for investors under-reaction to earnings announcements (e.g., Bernard and Thomas
1989, 1990) as well as for their inability to understand the accrual component of earnings (Sloan
1996).
To see this, consider a setting in which investors have some private information around the
earnings announcement date (say, reports from market analysts) and their information decision
problem is to whether to spend costly time and effort to digest/interpret the earnings
announcement data based on . 24 Assume that is neither too high nor too low, so that the
investors sometimes choose to process the accounting data and sometimes not. Depending on the
nature of the bias in the two sources of information, investors may rationally choose not to
process earnings data, i.e. to ignore the information given in the earnings announcement. This of
earnings. 25 Investors will choose to process the accounting data only when the error reduction
thanks to the additional accounting information is large. If the private information signals are
aggressive, Proposition 2 says that a conservatively biased accounting signal generates more
errors than an aggressively biased one, so investors will be less likely to incur the processing cost
hence, more likely to ignore earnings numbers when the accounting data are more
24
This interpretation is adopted by Indjejikian (1991).
25
This is because the conservative bias of accounting earnings numbers is entirely embodied in the accrual
component.
22
aggressive, in that favorable signals are less informative than unfavorable ones, investors may
only have access to information that is aggressively biased. When this is the case, they are less
likely to sustain the processing cost for firms or industries that report more conservatively. My
model thus predicts that under-reaction to earnings announcements and/or the inability to process
the accrual component will be the greater, the more conservative the accounting practices of a
firm, an industry or a country. The literature offers some empirical evidence in support of this
conjecture. Kaserer and Klingler (2008) document that the accrual anomaly only exists for
German firms that use IFRS or U.S. GAAP, not for those that use German GAAP. To the extent
that German GAAP is less conservative than the other two accounting standards, these findings
are consistent with my results. Another example is Narayanamoorthy (2006) who finds that the
markets tend to ignore the information in earnings due to accounting conservatism, which again
Secondly, the result also carries some implications concerning the behavior of investors in
acquiring private information. Consider a setting where investors have access to accounting
information and the possibility of acquiring private information at a cost . Assume that is
neither too high (so that the investor would never acquire the information) nor too low (so that
the investor always would). Private information acquisition then depends on whether the
consequent reduction in decision error is great enough to cover the cost . Proposition 2 then
implies that investors who have more conservatively biased accounting information will only
acquire other information that is conservative and will not acquire aggressively biased
more sharply. On the assumption that the non-accounting information tends to be aggressive, in
23
that favorable signals are less informative than unfavorable, 26 informed investors may have only
the option of acquiring information that is aggressively biased. When this is the case, they are
less likely to acquire it for firms or industries that report more conservatively. To the extent that
private information acquisition is the source of information asymmetry among investors (see
Verrecchia 1982, and Bushman 1991, among others), this implies that more conservative
accounting should be associated with less information asymmetry. By the same token, for firms
or industries that report more aggressively (or less conservatively), informed investors are more
the findings of earlier literature that information asymmetry increases around earnings
announcement (Lee et. al. 1993, Krinsky and Lee 1996), although the focus has mostly been on
the relationship between the accuracy of earnings information and information asymmetry (Gow
et. al. 2012), not on whether the increased information asymmetry is a function of the bias in
accounting information. More recently, DAugusta et. al. (2012) have found that firms with more
conservative accounting practices tend to produce less disagreement among investors around
earnings announcement dates. Insofar as less disagreement stems from the lesser likelihood of
informed investors acquiring private information, hence producing less information asymmetry,
analysis shows that firms that report more conservative accounting numbers will also make more
discussion and analysis (MD&A) section in the annual report, which includes discussion of
factors that would affect future performance and consists mostly in non-accounting information,
26
According to LaFond and Watts (2008), this is the key reason why from the capital market perspective accounting
should be conservatively biased.
24
although in principle it can be applied to all types of disclosure of non-accounting information. It
is also plausible that managers find it difficult to manipulate the hard accounting numbers but
may well have more leeway to manipulate the mostly forward-looking and qualitative
information of the MD&A section. To the best of my knowledge, no study has conducted a direct
empirical test of the relationship between the conservative bias of financial reports and the bias
in non-accounting disclosures like MD&A. Assuming that habitual optimism is associated with
aggressiveness and pessimism with conservatism, 27 one may use the disclosure tone of past
MD&A as a proxy for the degree of conservatism, with a more optimistic tone representing more
IV. Conclusion
This paper examines the optimal relation between the nature of biases in non-accounting and
part of information is more informative and which less. In this sense the paper differs from most
previous works, which focus on the overall informativeness (i.e. precision) of the information
signal. My paper shows that from the decision-making perspective, bias in accounting
substitutes. That is, the effort to counteract the aggressive bias of non-accounting information
efficiency. Future research could explore the optimal relation between the biases of different
information sources in various settings in greater detail. The insights provided by the present
27
When disclosures tend to be optimistic, optimistic disclosures would not be very informative while pessimistic
ones would be highly informative, corresponding to disclosures being aggressive. By the same logic, habitually
pessimistic disclosures are associated with conservatism.
28
See, e.g., Rogers et al. (2011), for an example in which the disclosure tone in mandatory disclosures can be
used to capture optimism and pessimism.
25
REFERENCES
26
ROGERS, J.; A. VAN BUSKIRK AND S. Zechman. Disclosure Tone and Shareholder Litigation. The
Accounting Review 86 (2011): 2155-2183.
SIMS, C. A. 'Stickiness.' Carnegie-Rochester Conference Series on Public Policy 49 (1998): 317-356.
North-Holland.
SIMS, C. Implications of Rational Inattention. Journal of Monetary Economics 50 (2003): 665-690.
SIMS, C. Rational Inattention: A Research Agenda. Discussion Paper Series 1: Economic Studies 2005,
34. Deutsche Bundesbank, Research Center.
YANG, M. Coordination with Flexible Information Acquisition. Journal of Economic Theory, 158
(2015): 721-738.
27
APPENDIX: PROOFS
Before proceeding, let us prove the following technical lemma, which will be used
(,| )
().
(,| )
(,| )
By Bayes rule, ( |, , ) = for = , . Insert this into the above expression,
(,)
Proof of Proposition 1:
The investor chooses the decision rule that maximizes his expected payoff conditional on all
the information he may have. Since the investor always have the option to liquidate, then if he
wants to continue, the expected continuation payoff must be greater than . Thus, based on his
information set at date 1, he will terminate any project that results in (|) < and continue
Proof of Lemma 1:
Following the optimal decision rule of Proposition 1, we can write the objective function in
Program A as
(, , , )
{(,):[|,]<}
+ {(,):[|,]>} ( |, , , )
28
+ {(,):[|,]>} ( |, , , ) (A-1)
(,| ,,)
Using Bayes Rule, we can write ( |, , , ) = (,)
for = , . Note that we
[(, | , , ) + (, | , , ) ]
{(,);[|,]<}
+ {(,);[|,]>} (, | , , )
+ {(,):[|,]> (, | , , ) (A-2)
{(,):[|,]<} (, | , , ) + {(,);[|,]>} (, | , , ) =
= (,) (, ) = 1
for = , .
Thus
Inserting this into the first and second terms of (A-2), we now have the objective as
{(,):[|,]< (, | , , ) +
{(,);[|,]>} (, | , , )
{(,):[|,]<} (, | , , ) + {(,):[|,]>} (, | , , )
= +
= + (, ) (A-3)
29
Since the first two terms are independent of and , maximizing the objective function is
equivalent to minimizing .
Proof of Lemma 2:
Lemma A.1 says that the decision maker should terminate (continue) at ( , ) if
( , | ) + 1
( , | )
= 1+
< (>)1 and the decision maker should terminate (continue) at ( , )
( , | ) 1 +
if ( , | )
= 1+
< (>)1 . Solving for the two inequalities gives + > (<)0 and
+ > (<)0 , respectively. Hence the decision maker should terminate (continue) at both
Proof of Lemma 3:
We prove the case when = 0; the proof when = 0 is analogous. Since = 0, we know
from Lemma 2 that the decision maker should terminate (continue) at both ( , ) and ( , ) if
> (<)0. When > 0, the project is continued at ( , ) and terminated otherwise. The
() = ( )( , | ) + ( )1 ( , | )
= ( )[( , | ) + 1 ( , | )]
= ( )[(1 )(1 + ) + 1 ( + )]
= ( )[2(1 ) (2 1)]
(0 ) = ( )[2(1 ) 0 (2 1)].
When < 0, the project is terminated at ( , ) and continued otherwise. The summation
() = ( )(1 ( , | )) + ( )( , | )
30
= ( )[1 ( , | ) + ( , | )]
= ( )[1 ( ) + (1 )(1 )]
= ( )[2(1 ) + (2 1)]
Proof of Lemma 4:
We prove the first part only; the proof of the second part is analogous. When < 0 and
case is possible. When this condition is satisfied, the summation of errors then becomes
(, ) = ( )( , | ) + ( )1 ( , | )
= ( )[( , | ) + 1 ( , | )]
= ( )[(1 + )(1 + ) + 1 ( + )( + )]
= ( )[2(1 ) (2 1) (2 1)]
in (0 , ) = ( )[2(1 ) 0 (2 1) (2 1)].
When + < 0, which is always possible when < 0, the project is terminated at ( , )
(, ) = ( )(1 ( , | )) + ( )( , | )
= ( )[1 ( , | ) + ( , | )]
= ( )[1 ( )( ) + (1 )(1 )]
= ( )[2(1 ) + (2 1) + (2 1)]
31
Since (, ) is increasing in , (, ) achieves the minimum when = 0 , resulting
the case + > 0 is possible, (0 , ) > (0 , ). Thus the decision maker will choose
Proof of Proposition 2:
Lemma A.1 says that the decision maker should terminate (continue) at ( , ) if
( , | ) + 1 1 ( , | )
( , | )
= 1+
< (>) and should terminate (continue) at ( , ) if ( , | )
=
1 + 1
1+
< (>) . Solving for the two inequalities yields
and
respectively.
We only prove the first part, as the second part can be proved analogously.
Before we proceed, note that the left-hand sides of (A-4) and (A-5) are all strictly increasing
in . To see this, take the derivative of the left-hand side of (A-5) with respect to ; this results in
( + 1) + ( 1) > ( + 1) + ( 1) = 2 +1 > 0.
+1
foregoing.
When = 0, both the left-hand sides of (A-4) and (A-5) are negative, which is smaller than
their right-hand sides. Thus when < 0, both the left-hand sides of (A-4) and (A-5) are smaller
32
than the right hand side. Since the left-hand side of (A-4) is strictly greater than that of (A-5)
when = 0 and it increases with at a faster pace, we know that when > 0, the left-hand side
of (A-4) is strictly greater than that of (A-5). This leaves us with the following cases:
Case a2: when > 0, depending on the parameters, it is possible that only the left-hand side
of (A-4) or else both the left-hand sides of (A-4) and (A-5) are greater than the right-hand sides.
( , ) and ( , ).
(, , ) = ( )(1 ( , | )) + ( )( , | )
= ( )[1 ( , | ) + ( , | )]
= ( )[1 ( )( ) + (1 )(1 )]
= ( ){2(1 ) + ( 1)(1 )2
+[( + 1) ] + [( + 1) ] + ( 1)}
(0 , , ) = ( ){2(1 ) + ( 1)(1 )2 0 [( + 1) ] +
[( + 1) ] ( 1)0 }.
(, , ) = ( )( , | ) + ( )(1 ( , | ))
= ( )[( , | ) + (1 ( , | ))]
= ( ){2(1 ) + ( 1)(1 2 )
33
[( + 1) 1] [( + 1) 1] ( 1)}
( 1), is negative, as shown above. Therefore, achieves its minimum at = 0 (if both
the left-hand sides of equations (A-4) and (A-5) are larger than the right-hand sides when =
0 ) , resulting in
(0 , , ) = ( ){2(1 ) + ( 1)(1 2 ) 0 [( + 1) 1]
[( + 1) 1] ( 1)0 }.
(, , ) = ( )( | ) + ( )( | )
= ( )( | ) + ( | )
= ( )[1 + + (1 )]
Since > 1, (, , ) is decreasing in and thus achieves its minimum when = 0 (if
(0 , , ) (0 , , )
The right-hand side of equation (A-6) is negative. The reason is that, first, the term ( +
(, , ).
34
We then compare (0 , , ) and (0 , , ).
(0 , , ) (0 , , )
2+1
Since [0 , 0 ] (+1 , 1 ), when (+1 ) < 1 , or, equivalently, < 2(+1),
2+1
0 < +1. Otherwise when 2(+1), 0 < 1 .
2+1
When < 2(+1),
[( + 1) ] ( 1)0
= [( + 1) ( 1)0 ]
The coefficient of ,
(1)
( + 1) ( 1)0 > ( + 1) ( 1) = 2 + =
+1 +1
2 +1 > 0. Therefore [( + 1) ] ( 1)0 < 0.
( 1)(1 ) + 0 [2 1 ( + 1)]
< ( 1)(1 ) + +1 [2 1 ( + 1)]
(2+1)
= 2( ) +1
.
(0 , , ) < 0.
2+1
When 2(+1), The coefficient of ,
35
( + 1) ( 1)0 > ( + 1) ( 1)(1 ) = 2 (2 1) > 0 as
2+1 21
2(+1) > 2
.
( 1)(1 ) + 0 [2 1 ( + 1)]
21
= 2(1 ) 2
< 0
2+1 21
as 2(+1) > 2
. Therefore we have (0 , , ) (0 , , ) < 0.
We now consider the case > 0. Recall that the two inequalities concerning the termination
and
respectively.
Case b1: When 0 < . Note that the left-hand side of (A-4) minus the left-hand side of (A-
5) is equal to ( 1)( ). Thus, if 0 < the left-hand side of (A-4) is always smaller than
the left-hand side of (A-5). We further divide the discussion into subcases.
). In this case, the left-hand side of (A-5) is always smaller than the right-hand side. Since the
left-hand side of (A-4) is always smaller than the left-hand side of (A-5), it must be that the left-
36
hand side of (A-4) is always smaller than the right-hand side. Thus, the decision will always be
5) is greater than the right-hand side at = 0 but the left-hand side of (A-4) is smaller than the
right-hand side at = 0 . Thus, the left-hand side of (A-4) is always smaller than the right-hand
than the right-hand side, resulting in termination at ( , ). However, when becomes smaller,
it is possible that the left-hand side of (A-5) is smaller than the right-hand side, resulting in
(0 , , ). Recall that
(0 , , ) (0 , , )
+[( + 1) ] ( 1)0 } .
Note that since > +1, 2 1 ( + 1) < ( + 1) 1. Thus
37
where the last inequality is due to the fact that ( 1)(1 ) + 0 [( + 1) 1] +
this case the optimal solution is = 0, i.e. conservative bias if the left-hand side of (A-5) is
). In this case both the left-hand sides of (A-4) and (A-5) are greater than their right-hand sides
Also note that the left-hand side of (A-4) is always smaller than that of (A-5). Thus there are
three possibilities when becomes smaller: 1) both the left-hand sides of (A-4) and (A-5) are
greater than their right-hand sides, resulting in termination at both ( , ) or ( , ); 2) both the
left-hand sides of (A-4) and (A-5) are smaller than their right-hand sides, resulting in
continuation at both ( , ) or ( , ); 3) only the left-hand side of (A-4) is smaller than the
(0 , , ) (0 , , )
38
( 1)2(1 ) + ( 1)0 + ( + 1)(2 1) 2(1 )(2 1) > 0. This implies
(1)2(1)(1)0
(+1)(21)
. Clearly we also need 0 to be sufficiently large.
Under this possibility there also exists an , [0 , 0 ] such that the project is
(, , ) = ( )( | ) + ( )( | )
= ( )( | ) + ( | )
= ( )[1 + + (1 )]
(0 , , ) (, , )
when = 0 is optimal.
39
( 1)0 < ( 1)(1 ) < [( + 1) ]0 + [( + 1) 1] + ( 1)0 or
requires and 0 to be sufficiently small. For all other cases, aggressive bias is optimal.
Case b2: When 0 . Note that in this case the left-hand side of (A-4) is greater than the
left-hand side of (A-5) when [, 0 ]. We further divide the discussion into sub-cases.
). In this case, both left-hand sides of (A-4) and (A-5) are always smaller than their right-hand
sides for all . This results in the optimal decision rule being always to continue under both
conservative.
4) is greater than the right-hand side at = 0 but the left-hand side of (A-5) is smaller than the
right-hand side at = 0 . Thus, the left-hand side of (A-5) is always smaller than the right-hand
than the right-hand side, resulting in termination at ( , ). However, when becomes smaller,
it is possible for the left-hand side of (A-4) to be smaller than the right-hand side, resulting in
continuation at ( , ). In fact, since the left-hand side of (A-4) is smaller than the left-hand side
of (A-5) when < , the project will be continued at both ( , ) and ( , ) at least when
< . We thus need to compare the decision errors of (, , ) and (, , ). Note that
40
(, , ) is a constant of whereas (, , ) goes to its minimum at = 0 . We thus
(0 , , ) (0 , , )
= ( ){( 1)(1 ) 0 [( + 1) ]
( 1)0 is the left-hand side of (A-5) minus the right-hand side of (A-5) evaluated at
= 0 . Since the left-hand side of (A-5) is always smaller than the right-hand side of (A-5),
(0 , , ) < (, , ).
Therefore in this case the optimal solution is = 0 , i.e. conservative bias if the left-hand
). In this case both the left-hand sides of (A-4) and (A-5) are greater than their right-hand sides
Also note that the left-hand side of (A-4) is greater (smaller) than that of (A-5) when > (<).
Hence, there are four possibilities when becomes smaller: 1) both the left-hand sides of (A-4)
and (A-5) are always greater than their right-hand sides, resulting in termination at both ( , )
or ( , ) for all ; 2) both the left-hand sides of (A-4) and (A-5) are smaller than their right-
hand sides, resulting in continuation at both ( , ) or ( , ); 3) only the left-hand side of (A-
( , ); 4) only the left-hand side of (A-5) is smaller than the right-hand side, resulting in
41
For possibility 1) we only have (, , ), which has its minimum at = 0 , so that in
conclusion will be the same as above. So under this possibility, either > (0 , ) when
or
]0 + [( + 1) 1] + ( 1)0
or
1) ]0 + [( + 1) 1] + ( 1)0 )
or if
[( + 1) ]0 + [( + 1) 1] + ( 1)0
42
or if
( 1)(1 ) <
0 and to be sufficiently small, or, equivalently, < (0 , ) and 0 < for some and .
Proof of Lemma 5:
Again we prove only the first part. When = 0, equations (A-4) and (A-5) now become
and
since ( + 1) > ( + 1) 1. We can divide our discussion into the following three
cases:
Case 1, when [( + 1) ] > ( 1)(1 ). This results in both the left-hand sides
of (A-8) and (A-9) being greater than the right-hand sides, thus termination when observing both
left-hand side of (A-8) being greater than its right-hand side but the left-hand side of (A-9) being
smaller than its right-hand side. As a result, the project is terminated when observing ( , ) but
43
)[( + 1)(1 ) ( 1)]. Since (, 0, ) is increasing in 0 it is minimized when
Case 3, when [( + 1) 1] < ( 1)(1 ). This results in both the left-hand sides
of (A-8) and (A-9) being smaller than their right-hand sides. Hence, the project is continued
(, 0, ) = ( ){2(1 ) + ( 1)(1 )2 + [( + 1) ]}
2+1
0 < +1. Otherwise when 2(+1), 0 < 1 .
2+1
When < 2(+1),
( 1)(1 ) + 0 [2 1 ( + 1)]
< ( 1)(1 ) + +1 [2 1 ( + 1)]
(2+1)
= 2( ) +1
.
44
1 2+1 (2+1) 2+1 1 (2+1) 2+1
Since 2 < < 2(+1), 2( ) < 2 2(+1) 2 = 2(+1) < 0.
+1 +1
(0 , , ) < 0.
2+1
When 2(+1),
( 1)(1 ) + 0 [2 1 ( + 1)]
21
= 2(1 ) 2
< 0
2+1 21
as 2(+1) > 2
. Therefore we have (0 , , ) (0 , , ) < 0.
Proof of Corollary 1:
We prove only the first half; the proof of the second half is analogous. Suppose investor
observes . Then he will continue when observing ( , ) and may choose to terminate when
1 (, ) = ( )[( , | ) + ( , | )]
= ( )( | ) = ( )(1 + )
+ ).
45
1 (, ) = ( )( , | ) + ( )( , | )
= ( )[( , | ) + ( , | )]
0 (2 1)].
Proof of Proposition 2:
Program B:
{,(,,),} (, , )
(,)(,,)
+ [( |, , ) + ( |, , )](, , )
(,)(,,)
. .
= (Rational Expectations)
First note that (IR) must be binding, as otherwise we could decrease all payments to the
manager by the same amount without violating any of the constraints. Thus {(,)(,,)[ +
46
( )( |, , )](, , ) = . Inserting this into the objective function, the objective
Given , for any and , the investor either chooses to liquidate and get or chooses
(, ) s.t. [ + |, ] < (>) . This implies that the optimal decision rule results in
Finally, we look at the (IC) constraint to solve for the optimal . From the optimal
decision rule we can write the objective function in the (IC) constraint as
{(,):[+|,]>}[ + ( )( |, , )](, , )
= {(,):[+|,]>}[( |, , ) + ( + )( |, , )](, , ).
(,| ,)
Using Bayes Rule, ( |, , ) = for = , . Inserting this into the above
(,,)
expression, we have
{(,):[+|,]>}[( |, , ) + ( + )( |, , )](, , )
and ( | ), it can be verified that for any (, , ) that excludes ( , ), both terms of the
last expression are increasing in and the first term is strictly increasing. We thus have the
47
As a summary, the optimal decision rule is such that the project is continued (terminated)
whenever the accounting report and the managers report are such that [( + )|, ] < (>
48
Figure 1. Timeline of the model
49