Tel. : 32 - (0)9 – 264.34.61
Fax. : 32 - (0)9 – 264.35.92
A Critical Note on Empirical Comprehensive Income Research
Philippe Van Cauwenberge 1
Ignace De Beelde 1
1 Department of Accountancy and Finance, Ghent University. Kuiperskaai 55E, 9000 Ghent, Belgium.
This paper presents a critical analysis of empirical comprehensive income research, the majority of
which consists of value relevance studies. The first part of the analysis focuses on the functional form
of the value relevance regressions. The paper distinguishes between the informational and the
measurement approach to value relevance research and systematically traces the origin of the
specifications under both approaches. The informational approach is characterized by diversity in
functional form specifications, which can be traced to differences in expectations models for earnings
or price/earnings ratios. In the light of this diversity, it is remarkable to observe that most of the
authors provide little or no argumentation for the choice of their particular functional specification.
While the authors under the measurement approach try to legitimize their regression specifications by
referring to the work of Ohlson, in fact, one of the contributions of the residual income model is that it
demonstrates the restrictive nature of these specifications. The second part of the paper performs a
detailed investigation of the empirical findings. This analysis reveals several peculiarities that defy
economic intuition. These peculiarities are nontrivial since they, together with the low explanatory
power of the regressions, affect the credibility of the main research findings. The conclusion of the
paper is that there is danger in taking the empirical findings of this research domain at face value and
that the potential for informing standards setters is limited.
Keywords: Comprehensive Income; Dirty Surplus; Value Relevance; Residual Income Valuation
JEL Descriptors: M41
Ackowlegdements: This paper has benefited from the comments and suggestions of Walter Aerts
(University of Antwerp), Ann Tarca (University of Western Australia) and Guy Weyns (Morgan
Stanley and Co, London)
This paper presents a critical analysis of empirical comprehensive income research. The purpose
thereby is to assess the credibility of the research findings. Broadly defined, empirical comprehensive
income research considers statistical relations between market data and different income measures,
with the intention of providing information that is relevant to standard setters’ deliberations of their
performance-reporting projects. Section 2 starts out by briefly discussing the rational behind the
performance-reporting projects of FASB and IASB.
It should be noted from the outset that there is no intention of providing an exhaustive literature
review or a summarization of research findings. Such an overview can be found for instance in
(Thinggaard et al., 2006).1 Rather, the intention is to scrutinize the underlying methodological aspects
and to highlight some peculiar empirical findings. The majority of empirical comprehensive income
studies consists of value relevance research. Hence, a significant part of the analysis focuses on the
functional form specification of the value relevance regressions that are estimated therein. Section 3
provides an introduction to value relevance research and explains the distinction between the
informational and the measurement approach.
The original starting point for this analysis was the observation that, although comprehensive
income value relevance studies are, more or less, all intended to inform standard setters on the same
issue, they employ a wide variety of functional forms. In contrast to this diversity, most researchers
seem to take their functional form specification for granted, providing little or no argumentation for
their particular choice. Given that the functional specification obviously influences the statistical
findings, has implications for the inferences that can be drawn from it (Beaver, 2002), and potentially
has an impact on the credibility of the research, the variety of functional specifications might not be an
innocuous issue and seemed worthy of consideration.
Section 4 analyzes comprehensive income value relevance studies that are conducted under the so-
called informational approach. It is shown that the diversity in functional specifications therein can be
traced to differences in expectations models for earnings or price/earnings ratios. Since empirical
dominance among these different expectations models is not established, the choice of a functional
specification under the informational value relevance approach is arbitrary.
Section 5 considers value relevance tests under the so-called measurement approach, i.e. value
relevance test that are derived from a valuation model. The particular valuation model that is referred
to is the residual income valuation model (RIM). After closer inspection, it turns out that the
restrictions needed to attain congruence between RIM and the value relevance regressions are very
demanding. In fact, rather than legitimizing these particular value relevance tests, RIM actually reveals
their limiting character.
Section 6 considers in detail the nature of the empirical findings. A first general observation is that
the explanatory power of the regressions is exceedingly low, suggesting danger of model
misspecification and bias in coefficient estimation. Furthermore, a deliberately chosen selection of
findings shows several outcomes that defy economic reasoning. Generally, the researchers themselves
do not mention these - somewhat embarrassing - findings. They are also not discussed in the overview
paper of (Thinggaard et al., 2006), which takes the research findings at face value. However, these
counterintuitive findings are nontrivial since they affect the credibility of the main empirical findings
Section 7 presents an alternative approach to value relevance research (Isidro et al., 2006), which
is more in the spirit of RIM and is less restrictive than the aforementioned value relevance studies.
To our knowledge, this is the first paper that brings to the surface and makes an inventory of the
variety of functional form specifications in comprehensive income value relevance research and
systematically traces the origin of these functional forms. Through this analysis, it becomes clear that
the value relevance regressions are both arbitrary and underspecified. Another contribution is the
detailed analysis of the empirical results, which reveals several peculiarities that defy economic
intuition. Together with the low explanatory power of the regressions, these findings impact
negatively on the overall credibility of the main conclusions.
An old and unresolved issue in accounting is whether income should be determined according to the
principle of clean surplus accounting (Brief and Peasnell, 1996). Clean surplus income includes all
value changes in equity, except those resulting from transactions with the owners. Standard setters
have departed from clean surplus accounting on many occasions, allowing certain value changes to
bypass the income statement and be booked directly into equity. Typical examples of these so-called
dirty surplus flows are unrealised gains and losses on available-for-sale securities, additional minimum
pension liability adjustments, currency translations, gains and losses of cash flow hedges and asset
An important concept related to dirty surplus accounting is recycling or reclassification. Under
recycling, dirty surplus flows are booked into income against equity when, at a later date, certain
criteria regarding realisation and uncertainty are met (G4+1, 1999).
The practice of dirty surplus accounting has grown over the years and has been applied in an ad
hoc manner, mostly as a political way out of controversial accounting issues (Barker, 2004).
International evidence indicates that dirty surplus flows are potentially material, often not centred
on zero and subject to substantial cross-country variation (Isidro et al., 2004).
In response to this state of affairs, concerns arose about the increasing lack of transparency of dirty
surplus flows. Smith and Reither (1996) report that companies tend to obscure dirty surplus flows by
combining them with each other or with other categories of equity, resulting in significant search cost
and inefficiencies. Especially from the users’ side grew the demand for a statement of comprehensive
income, which would transparently present all income flows in one statement (AIMR, 1993).
In the US, the FASB responded to these concerns by issuing SFAS No. 130 in 1997, requiring
companies to report comprehensive income in a primary financial statement. However, SFAS 130, in
contrast to the exposure draft that preceded it, allows for the presentation of comprehensive income in
either a performance or a non-performance statement. The significance of this admission is reflected
by the number of companies that took advantage of it. In a survey of US companies, Mazza and Porco
(2004) find that 83 percent report comprehensive income in a statement of stockholders equity.
A particular problem of not presenting comprehensive income in a performance statement arises in
the context of earnings management. Lee et al. (2007) find that insurers who manage earnings through
selective realisation of gains and losses are more likely to report comprehensive income in the
statement of changes in equity. The underlying mechanism that allows companies to manage net
income through such gains trading is recycling. Thereby, previously unrealised gains and losses
appear in net income when they are realised. However, due to double-entry accounting, recycling of
income needs to be matched by a balancing cross-entry in equity. Of crucial importance for the
possibility of gains trading is the display of this cross-entry. When presentation of other
comprehensive income is required, one possibility is that this cross-entry is not only conducted in
equity, but also shown in other comprehensive income. In fact, other comprehensive income is nothing
more than a visually explicit elaboration on the dirty surplus changes in equity (i.e. changes in equity
not arising from retained earnings). In that sense, the inclusion of these so-called reclassification
adjustments in other comprehensive income is self-evident. These adjustments avoid that, over time,
other comprehensive income is double counted in comprehensive income. However, when other
comprehensive income is not included in the performance statement, the recycling entry that appears
in the performance statement is not matched in the same statement by the accompanying cross-entry of
opposite sign. Experimental research (Hirst and Hopkins, 1998) has shown that professional analysts
are less able to detect earnings management when other comprehensive income – and the cross-entry
therein, is not included in a performance statement, but rather in the statement of changes in equity.
For this and other reasons, FASB is currently reconsidering SFAS No. 130.
The IASB is currently also engaged in an investigation of the presentation of comprehensive
income (Barker, 2004). IAS 1, Presentation of Financial Statements, permits but does not require a
single comprehensive income statement. Since 2004, the IASB and the FASB joined forces in a
performance-reporting project (IASB, 2004).
Currently, some of the more pertinent issues are whether a single or a double performance
statement of comprehensive income should be required, whether earnings per share (EPS) should be
displayed for comprehensive income and whether recycling should be required, allowed or prohibited
2 (IASPlus, 2007). In other words, the debate is not so much about an exclusive choice between net
and comprehensive income, but rather about the relative prominence of display that should be
allocated to net income versus comprehensive income (Van Cauwenberge and De Beelde, 2007).3
When comprehensive income is presented as the most prominent income number, the risk is that users
might not be able to unscramble the different analytical properties of its components (Tarca, 2006).
When other comprehensive income is relegated to less prominent financial statements, its visibility is
reduced, which increases the chance that it might be overlooked (Robinson, 1991).
Obviously, an important consideration for standard setters’ deliberations is the actual relevance of
comprehensive income and other comprehensive income for valuation purposes, as is evidenced by
stock market data. Accordingly, it is no surprise that a considerable number of empirical research
papers has appeared that explicitly try to answer this question. This field of research is referred to in
this paper as empirical comprehensive income research. Since the majority of empirical
comprehensive income papers are value relevance studies, the next section presents a brief
introduction to value relevance research.
3. The nature and specifications of value relevance tests
Value relevance studies examine associations between security price-based dependent variables and a
set of accounting variables (Barth et al., 2001; Holthausen and Watts, 2001; Kothari, 2001). These
associations are investigated by applying linear regression techniques. An accounting number is
termed “value relevant” if its estimated regression coefficient reveals a statistically significant
association with the dependent variable.
Value relevance tests are either relative or incremental. In the context of comprehensive income
research, “relative” association studies compare the association between stock market prices (or
returns) and alternative bottom-line measures - mostly net income versus one or more measures of
comprehensive income. The income number yielding the most significant earnings response
coefficient (ERC) or the highest R² is considered to be the most value-relevant. In other words, relative
association tests aim to rank different income measures in terms of value relevance. On the other hand,
“incremental” value relevance studies examine whether, in addition to net income, one or more
components of OCI are useful in increasing the explained portion of the dependent variable.
Incremental value relevance is judged by whether the earnings response coefficients of these OCI
components are significantly different from zero or whether the adjusted R² increases after inclusion of
the OCI component. Several comprehensive income value relevance studies perform both relative and
incremental association tests.
An issue of obvious importance concerns the extent to which value relevance research has the
potential to allow for standard-setting inferences, an issue which already raised considerable
controversy (Barth et al., 2001; Holthausen and Watts, 2001). Implicitly, all value relevance papers
presume that value relevance implies superiority in some normative sense (Lee, 1999). Also in the
context of comprehensive income research, value relevance studies typically start out by pointing to
the comprehensive income reporting projects of the IASB and the FASB as the motivation for their
research, thereby suggesting that the empirical findings of the performed association tests are relevant
to standard setters’ deliberations.
To scrutinize the question of potential policy relevance, it is necessary to be more specific about
the meaning of the term value relevance. In other words, one needs to consider the underlying rational
behind the investigation of these associations. An important distinction in that regard is between the
measurement and the informational perspective (Barth, 2000; Barth et al., 2001; Holthausen and
Watts, 2001; Beaver, 2002; Walker, 1997)
Under the “measurement” perspective, the specification of a value relevance regression equation is
derived from a valuation model. A valuation model expresses firm value as a function of accounting
variables – mostly income, book value or a combination of both. At the same time, the underlying
valuation model also delivers predictions for the values of the estimated regression coefficients.
Association studies based on the measurement approach potentially allow for inferences with regard to
“relevance” and “reliability” of accounting numbers, properties that the standard setters have adopted
as their goals (Barth, 2000). In other words, under the presumption that the valuation model is
descriptive, failure to find the predicted coefficients may indicate that the inputs to the valuation
model, i.e. the accounting data, are irrelevant, unreliable or both. Therefore, in principle, standard
setters can use the findings of association studies under the measurement approach to judge the
effectiveness of their recognition and measurement policies.
An important assumption of the measurement approach is that accounting variables directly feed
into a firm valuation function. In other words, it needs to be assumed that a valuation function in terms
of accounting variables can be specified. Since the end of the 1960s, this approach has been subject to
severe criticism (Beaver et al., 1968). The most important source of objection was that, due to
uncertainty and the existence of imperfect and incomplete markets for a company’s assets and
liabilities, it is impossible to entertain the assumption that accounting variables bear any simple, direct
relationship to valuation (Beaver, 1998). The informational perspective, which grew out of this belief,
relies on less ambitious assumptions and considers only whether accounting data are useful to
investors (Lev 1989). Accordingly, with regard to the potential of informing standard setters, its
findings pertain only to the more abstract purpose of usefulness of accounting data.4 The only
assumption, that is entertained by the informational association studies, is that, if accounting variables
are useful to investors, then they should show up as price revisions (Ball and Brown, 1968).5
Accordingly, no predictions with respect to the sign and size of the coefficients are made. Usefulness
is assumed when the estimate of a coefficient of an accounting variable statistically differs from zero.
The distinction between the measurement approach and the informational approach also has
implications for whether the estimated relation between market and accounting data is specified in
levels or in flows. The informational approach states that a signal is informative only if the signal can
alter beliefs conditional upon the other information available (Beaver, 1998). In other words,
accounting information can only be relevant if it comes as a surprise. The determination of the surprise
component requires the formulation of an expectations model. With respect to income, a standard
approach has been to assume a naive random walk model, i.e.
Bernard, 1989; Kothari and Zimmerman, 1995; Lev and Ohlson, 1982). The surprise component is
then the change in earnings or .
According to the same reasoning, the corresponding dependent variable is the change in price (or
return) and not the price itself since the latter reflects both expected and unexpected information.
Accordingly, the econometric specification that corresponds naturally to the informational approach is
a flow equation. On the other hand, the measurement approach features a levels specification as a
logical starting point. The reason is that the valuation models, which underlie the estimated
regressions, are defined in levels (Barth, 2000).
Whether a value relevance test is performed in levels or in differences also determines the type of
inferences that are allowed for. The levels specification under the measurement approach addresses the
question of how well accounting variables reflect or summarize the economic condition of the firm, as
reflected by its stock price, irrespective of whether this information is timely or not. The informational
perspective explicitly addresses the question of timeliness: Accounting data can only be relevant in an
informational sense if they are not pre-empted by other information that is relevant for forming
However, when an association is uncovered under the informational approach, due to the lack of
the relation with a valuation model, the underlying economic rational for this association cannot
immediately be inferred from the value relevance study itself and remains an open question. The result
is merely an association, without direct implications or reference potential for standard setters
(Holthausen and Watts, 2001). In contrast, when an association is found under the measurement
approach, there is a valid presumption to conclude that the related accounting variable is both reliable
and relevant and that the valuation model has descriptive power.
It is possible that accounting data are relevant in a measurement sense, but not in an informational
sense. The reason is that they are not timely. A potential critique is that timeliness is not an overriding
issue for accounting standards setters. According to Ohlson and Penman (1992), the construct of
unexpected earnings, which researchers employ to explain returns, is irrelevant to the practicing
accountant who implements accounting principles. An important function of accounting data could be
to provide a legally liable and publicly available confirmation of company events (Beaver, 2002;
Watts and Zimmerman, 2001). This said, in practice, most value relevance studies that follow a
measurement approach are also estimated in first difference form. Typical reasons to depart from a
levels specification are concerns of heteroscedasticity and correlated omitted variables such as scale
(Barth, 2000). Note that, also in this case, the use of returns as a dependent variable requires the
formulation of an expectations model for the right-hand-side variables. However, under the
measurement approach, this issue is typically ignored.
4. Comprehensive income value relevance under the informational approach
Table 1 presents an overview of the different functional form specifications that appear in
comprehensive income value relevance research, together with a categorization of these specifications
as either derived from a valuation model or not. With the exception of (Brimble and Hodgson, 2005)
and (Cahan et al., 2000), all studies follow an informational perspective. In other words, they do not
provide a link to a valuation model.
As mentioned before, value relevance studies under the informational perspective are inferior, in
terms of standard setting inferencing potential, vis-à-vis valuation theory-based association studies.
Another consequence of the absence of an underlying valuation model is that the informational
perspective lacks the necessary rigor to discriminate among different functional specifications (Ohlson
and Skroff, 1992). A priori, one functional specification is as good as another and dominance can only
be settled by goodness-of-fit considerations. This ad hoc character is reflected by the diversity of
econometric specifications that characterizes the informational approach. This section intends to
investigate this diversity.
As explained in section 3, according to the basic random walk expectations model, the
specification that accords naturally to the informational perspective is a first difference relation
between returns and changes in income.
When relative informational value-relevance is investigated, the estimated value relevance
equation has the following form:
In the context of comprehensive income research, income is measured either as net income (NI) or
as comprehensive income (CI). When incremental value-relevance is investigated, the corresponding
is the vector of changes of other comprehensive income (OCI) components and
the associated vector of coefficients. When the incremental value relevance of aggregate OCI is
reduce to a scalar.
Returns are measured either as raw (R) or as abnormal returns (AR). Abnormal returns are
calculated as the out-of-sample forecast errors of an estimated market return model (i.e.
it Mt it
, where is the return of the reference market).
It is standard procedure in value relevance studies that income, or the change therein, is
normalized by beginning of period stock price (Christie, 1987; Ohlson, 1991).
Equations (1) and (2) are typically considered as the straightforward articulations of the research
question under the informational approach (Easton and Harris, 1991; Ohslon, 1991; Ohlson and Skroff
,1992). Note however that, although the use of the linear form for these equations is standard practice,
the restriction of linearity on the relation between unexpected earnings and returns is not implied as
such by the informational approach. Consider for instance (Ball and Brown, 1968), which is
uncontroversially regarded as the seminal paper for the informational approach to accounting research
(Beaver, 1998). Ball and Brown only partition a sample of firms according to whether they have a
negative or a positive earnings surprise and investigate whether this partition has discriminatory power
with respect to unexpected stock returns. No linearity is assumed. However, over time, correlation has
become a standard measure in empirical information content studies (Lev, 1989). Correlation confines
the discovery of co-movement between two variables to the linear form.
After consideration of part A of table 1, it is clear that none of the estimated functional forms
exactly corresponds to the basic format of equations (1) and (2). Nonetheless, it can be shown that the
right-hand sides of these equations, in some way or another, all proxy for unexpected earnings and that
they are variations to the basic formulations in (1) and (2). More in particular, the difference in
specifications is due to alternative expectations models for earnings or price/earnings ratios.
A first class of functional specifications is the one used by (Chambers et al., 2006) – henceforth
CLSS, (Dhaliwal et al., 1999) – henceforth DST, and O’Hanlon and Pope (1999) – henceforth OP (see
table 1). In each of these studies, returns are regressed over levels of income, to wit net income or
comprehensive income. Note that first differencing of a balance sheet valuation model (i.e.
, where MVE
u BVE MVE
it = market value of equity, BVEit = book value of equity and uit =
error term) also leads to this specification. However, none of these papers provide a link to a valuation
model. In fact, the motivation of the functional form generally receives scarce attention. With regard
to DST for example, it is only in a footnote that there is an (implicit) indication that the authors follow
an informational perspective: DST state that the use of income proxies for unexpected income (p. 50).
However, no further explanation is provided, apart from a reference to work of Easton and Harris
(1991) and Ohlson and Skroff (1992). DST also provide no link between the functional form
specification and the research question. Actually, the fact that DST essentially trace the relation
between returns and unexpected income is at odds with the central motivation of their paper, which is
described as ‘to investigate whether comprehensive income or net income better summarizes firm
performance as reflected in stock returns’ (p. 46). However, as mentioned in section 3, an association
study under the informational approach, which investigates the surprise component of earnings, is ill
suited to address this question.
CLSS provide even less explanation and defend their particular functional specification only by
the desire to be consistent with DST. Accordingly, CLSS also follow an informational perspective. It
is remarkable that CLSS invoke the work of Ohlson –more in particular Ohlson (1999), to explain
their empirical findings. So, while CLSS are clearly conscious of the work in Ohlson, they are silent
about the implications that this work has for the specification of value relevance regressions (see
To see how the specification in DST-CLSS-OP accords with equation (1), the latter is rewritten as:
it it itit itit
P INCP INCreturn
The only difference between (1) and (3) is that the normalisation by previous period price is
written explicitly in equation (3) instead of implicitly as in equation (1). The advantage is that the
formulation in equation (3) facilitates an alternative interpretation, which is that
P INCP INC
represents scaled unexpected earnings, with
serving as a
measure of expected
In an equilibrium model under certainty, the expected ratio for income to price is always equal to a
constant, more specifically, the risk free rate of return r (Ohlson, 1991). In other words,
. Accordingly then, the effect of adjusting
for its expected value –
the constant, is irrelevant in a regression context because the regression intercept picks up the constant.
In other words, when investigating the relation between returns and unexpected income, one might just
as well regress returns on income, which provides a rational for the specification of DST-CLSS-OP.
Notice that (Ohlson, 1991) is actually an attempt to reintroduce the measurement perspective into
the informational approach. More in particular, the equality between the price/earnings ratio and 1/r
applies only when accounting income equals economic income in a Hicksian sense.
Another specification is the one by (Cheng et al., 1993) – henceforth CCG. CCG regress both raw
and abnormal returns on income and the change in income. Notice that this specification corresponds
to the first difference form of a valuation model expressed in book value and income (i.e.
u INCk BVEk MVE
, were k indicates the level of rents and θ is related to
the discount rate). However, the use of abnormal returns as the dependent variable in CCG precludes
this possibility. CCG cite the conflict in the accounting profession between all-inclusive
comprehensive income and net income as the motivation for their work (p. 195), but do not explain
how this functional specification fits in addressing this policy issue. To justify their specification,
CCG only refer to work of Easton and Harris (1991), without any further explanation.
Returning to the specification of DST-CLSS-OP, the property of a constant ratio between expected
earnings and prices (Ohlson, 1991) was derived under specific conditions and is probably too stylized
a description of reality. A less extreme assumption is that earnings/price ratios are mean reverting.
Beaver and Morse (1978) have presented evidence supporting this assumption. The implication of this
mean-reverting behaviour is that
is a suboptimal indicator of expected
it itP INC
. As a
practical remedy towards this problem, the change in is often included as an additional
variable to the level of . Easton and Harris (1991) have provided evidence indicating that the
inclusion of both levels and changes in income increases the explained portion of returns. This
inclusion is however not supported by any formal modelling, but rather, is defended by goodness-of-fit
considerations. The above reasoning explains the specification of CCG as a variant of equations (1)
15 Download full-text
Finally, another functional form under the informational approach is the one by (Biddle and Choi,
2006) – henceforth BC, (Kubota et al., 2006) – henceforth KST and (Kanagaretnam et al., 2007) –
henceforth KMS. As a motivation, BC merely provide a reference to (Biddle et al., 1995). Instead of
considering the time series properties of the price/earnings ratio, (Biddle et al., 1995) models the time
series properties of earnings themselves. In particular, they propose an AR(1) pattern as a proxy for
the earnings expectations process. Furthermore, their approach to modelling the expectations process
is less restrictive in the sense that they leave the determination of the persistence parameter to be
settled by the data. In other words, the markets expectations parameter is estimated jointly with the
earnings response coefficient. The link to the specification of BC-KST-KMS can be seen as follows.
According to the AR(1) process,
where ρ is the persistence parameter. Equation (1) between returns and unexpected income becomes:
is the surprise component in income. Or alternatively,
In conclusion, behind the diversity in functional form specifications that appears in informational
value relevance comprehensive income studies, lies a variety in expectations models regarding
earnings or price-earnings ratios. In the light of this variety, it is remarkable to observe that most
authors seem to take their specification for granted, providing no explanation for their particular
Another problem concerning the formulations of earnings expectations models relates to the
presumed uniformity over different income components. Note that this is even a problematic
assumption for the different components of net income (Sloan, 1996). However, as a rough
approximation, it might be defensible. For instance, prior research (Bernard, 1993) has shown that
most components of net income follow a random walk. However, in the context of comprehensive
income, this assumption becomes more difficult to uphold. Components of other comprehensive