Accounting Research Center, Booth School of Business, University of Chicago

Earnings As an Explanatory Variable for Returns

Author(s): Peter D. Easton and Trevor S. Harris
Source: Journal of Accounting Research, Vol. 29, No. 1 (Spring, 1991), pp. 19-36
Published by: Wiley on behalf of Accounting Research Center, Booth School of Business,
University of Chicago
Stable URL: http://www.jstor.org/stable/2491026
Accessed: 19-08-2015 08:51 UTC

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

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact [email protected].

Wiley and Accounting Research Center, Booth School of Business, University of Chicago are collaborating with
JSTOR to digitize, preserve and extend access to Journal of Accounting Research.

http://www.jstor.org

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
Journal of Accounting Research
Vol. 29 No. 1 Spring 1991
Printed in U.S.A.

Earnings as an Explanatory
Variable for Returns
PETER D. EASTON* AND TREVOR S. HARRISt

1. Introduction
In this paper we investigate whether the level of earnings divided by
price at the beginningof the stock returnperiodis relevant for evaluating
earnings/returns associations.' The primary model motivating this re-
search relies on the idea that book value (owners' equity) and market
value are both "stock"variablesindicatingthe wealth of the firm'sequity
holders. The related "flow"variables (after adjustingfor dividends) are,
respectively, earnings divided by price at the beginning of the return
period (A/P-1) and market returns. It then follows that earnings divided
by beginning of period price should be associated with stock returns.
Although models based on a relation between market value and book
value are used occasionally in the accountingresearchliterature (see, for
example, Landsman [1986], Harris and Ohlson [1987], and Barth
*
Macquarie University and University of Chicago; tColumbia University. The paper is
a revised version of working papers entitled "An Empirical Evaluation of Accounting
Income Numbers: Further Evidence" and "Evidence of Accounting Earnings as an Index
of Change in Value." The authors would like to acknowledge comments from workshop
participants at the following universities: Arizona, Auckland, California at Berkeley,
California at Los Angeles, Columbia, CUNY-Baruch College, Harvard, Macquarie, Michi-
gan, New South Wales (AGSM), and Southern California. Special thanks are due to Vic
Bernard, Jim Haggard, Robert Lipe, Jim Ohlson, Eric Noreen, Stephen Penman, Ram
Ramakrishnan, Jake Thomas, and anonymous referees. The study was partly funded by
the Faculty Research Fund, Columbia Business School, Columbia University and by the
Institute of Professional Accounting, Graduate School of Business, University of Chicago.
' The variable of interest is not the standard earnings-to-price ratio which is based on
contemporaneous, or past, earnings divided by contemporaneous, or past, price (Al/P-,).
19
Copyright ?, Institute of Professional Accounting 1991

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
20 JOURNAL OF ACCOUNTING RESEARCH, SPRING 1991

[1989]), another frequently used model expresses price as a multiple of

earnings. This latter model is generally used to motivate empirical studies
of the relation between security returns and change in earnings or
between abnormal returns and unexpected earnings (see, for example,
Beaver, Lambert, and Morse [1980] (henceforth BLM) and Collins and
Kothari [1989] (henceforth CK)). However, we show that the assumption
that price is a multiple of earnings also implies that the earnings level
variable (A/P-1) is a relevant explanatory variable for returns.
Although all the valuation models discussed in this paper indicate the
potential relevance of the level of current earnings divided by beginning-
of-period price, they do not exclude the relevance of change in earnings
divided by beginning-of-period price (AA/P_1). For example, differences
between the book value and stock price could be a function of earnings.
Thus, while the primary motivation of the empirical analysis is to
evaluate the relevance of the earnings level variable (A/P-1), the analysis
also considers and tests the relevance of the change in earnings variable
(zA/P_1) for explaining stock returns.
The results provide evidence that current earnings divided by begin-
ning-of-period price (A/P-1) is associated with stock returns. Univariate
regressions show that A/P-1 and zA/P-1 are each associated with returns.
In multiple cross-sectional regressions of annual returns on both the
levels (A/P-1) and the changes (AA/P_1) variables, the coefficient on
earnings levels is statistically significant (at 1% or better) in all years,
while the coefficient on earnings changes is significant (at this level) in
less than half the years. Additional analysis suggests that the relevance
of the earnings level variable (A/P-1) for assessing the association
between returns and earnings does not derive merely from a correlation
between the two earnings variables.
The primary empirical analyses focus on raw returns as the dependent
variable because it is the return measure defined in the valuation models
which we invoke. However, much of the empirical literature evaluating
the association between earnings and security prices considers the rela-
tion between unexpected (abnormal) returns and unexpected earnings.
A measure of unexpected earnings frequently used in association studies
involving annual return windows is change in earnings, sometimes de-
flated by the beginning-of-period price. Given the measurement error
inherent in any unexpected earnings measure and the relevance of the
earnings levels variable (A/P-1) in the return association tests, it is
pertinent to consider whether earnings levels are useful when considering
the unexpected return/unexpected earnings relation. In multiple regres-
sions of annual cumulative abnormal returns-based on a (monthly)
market model-on both A/P-1 and zA/P-1, the earnings levels variable
has statistically significant explanatory power over the earnings change
variable, and vice versa. Given that both earnings variables measure
unexpected earnings with error, we suggest, in the spirit of Brown et al.

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
 EARNINGS AS AN EXPLANATORY VARIABLE 21

[1987], that studies which use the residual from a regression of annual
abnormal returns on unexpected earnings might mitigate the effect of
measurement error by including both earnings level and earnings change
variables as measures of unexpected earnings.
The relevance of the earnings levels variable suggests new opportuni-
ties in our search for a better understanding of the associations between
accounting earnings and returns. For example, Lev [1989], who expresses
concern about the pervasiveness of low R2 statistics in returns/earnings
association studies, cites the focus on earnings levels as a potential
direction for improvement.
The models relating earnings variables and security returns are pre-
sented in section 2. After describing the data and sample selection
procedure in section 3, empirical analyses of the relation between the
earnings variables and security returns are documented in section 4.
Section 4 also addresses the issue of whether earnings levels and earnings
changes capture the same or different information. The use of earnings
divided by beginning-of-period price as an instrument for unexpected
earnings is addressed in section 5. Section 6 contains a summary of the
results and some conclusions.

2. The Relation Between Earnings and Returns

2.1 RETURNS AND EARNINGS ASSOCIATIONS BASED ON A
BOOK VALUE VALUATION MODEL
The idea that price and book value are both measures of the "stock"
value of the shareholders' equity may be expressed more formally as:

pit = B Vjt + ujt (1)

where Pjt is the price per share of firm i at time t, BVjtis the book value
per share of firm j at time t, and ujt is the difference between Pjt and
B Vjt.
The difference between market and book values (ujt) can result from
many factors including the choice of conservative accounting practices
and other information incorporated in price but not yet reflected in
accounting values. The relation between the "flow" variables-account-
ing earnings and security returns-may be obtained by taking first
differences of the variables in equation (1). This yields:

A\Pjt= ABXVjt
+ t. (2)
But, in general:

AB Vjt = Aft - djt (3)

where Aft is accounting earnings per share of firm j over the time period

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
22 P. D. EASTON AND T. S. HARRIS

t - 1 to t, and dj, is dividends paid per share of firm j over time period
t - 1 to t.2
Substituting (3) into (2), rearranging, and dividing by Pjtj yields:

(APjt+ djt)/Pjt-l = Ajt/Pjt-l + UjIt. (4)

That is, if stock price and book value are related, as we might expect,
then earnings divided by beginning-of-period price should be an appro-
priate variable for explaining returns.3

2.2 RETURNS AND EARNINGS ASSOCIATIONS BASED ON AN

EARNINGS VALUATION MODEL

Because some of the empirical literature focuses on an earnings-based

valuation model, we consider an alternative model which expresses price
as a multiple of earnings. That is:

Pit = pA1t+ Vjt. (5)

Ohlson [1989a] demonstrates that the Miller and Modigliani [1961]
dividend irrelevance proposition requires that if a dividend is paid on
security j at time t, then equation (5) must be written as:4

Pjt + djt = pA1t+ Vjt. (6)

It follows that:5

(APjt + djt)/Pjt-l = p[XAjt1/Pjt-] + Vjt. (7)

That is, there is a linear relation between change in earnings divided by

beginning-of-period price and security returns over that period.6

2 In the empirical analysis all data are adjusted to reflect the implicit assumption that
prices, dividends, and accounting variables are calculated based on the shareholding at a
particular point in time. Further, in principle, dj, reflects net withdrawals by the firm's
owners.
' The factors included in
ui, and uj, are not germane to this paper. The variable uj',is
relevant for the empirical analysis and is considered in section 4.
4 The coefficient p is frequently assumed to be constant across firms and time periods

(see, for example, BLM in an empirical analysis and Ohlson [1989a] in a theoretical
framework).
'Changes in earnings have been used in a way similar to equation (7) in BLM which
uses equation (5) to test the following form of the earnings/return relation:

,APjtPj,-, = AAj,/Aj,-l + jt.

While CK examine change in earnings via the relation:

AAjt/Pj,-, = (l/pj,-1)(,APj, + dj,)/Pj,-.

in a reverse regression framework.
6 Implicit in equation (7) is the assumption that a dividend is paid at time t but there is

no dividend paid at time t - 1. If a dividend is paid at t - 1, equation (7) becomes:

(zAPj,+ dj,)/Pj, = p[AAjtl/Pj,] - djt1l/P1t- + Vj't. (7a)

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
 EARNINGS AS AN EXPLANATORY VARIABLE 23

Of particular relevance to this study is the fact that dividing equation

(6) by beginning-of-period price yields:
(Pjt + djt)/Pjt-l = p[Ajt/Pjt1l] + vJt'. (8)
This equation suggests that from an earnings valuation perspective,
earnings levels (divided by beginning-of-period price) will be associated
with returns. The return variable (APj, + djt)/Pjt-i can be obtained from
(8) by subtracting one from each side of the equation and then expanding
the left-hand side.7
2.3 COMBINING BOTH VALUATION PERSPECTIVES
As a practical matter, for most companies the stock price is likely to
be a function of both book value and earnings. For example, Ohlson
[1989a] presents a model that suggests that the variable ujt in equation
(1) is partly a function of earnings. By combining a "book value only"
model (similar, in principle, to equation (4)) and an "earnings only"
model (summarized, in principle, by equation (7)) he proposes a valuation
relation in which price is a weighted function of book value and earnings.8
In a similar manner, equations (4) and (7) may be combined to give:
(APjt + djt)/Pjt-l = kp[A1jtl/Pjt-] + (1 - k)[AJt1/Pjt11+ wjt (9)
where k is a factor for weighting the contribution of change in earnings
versus earnings levels in the explanation of stock returns.
In the empirical analyses, we examine the relations between earnings
and returns implied by (i) equations (4) or (8) for earnings levels,
(ii) equation (7) for earnings changes, and (iii) equation (9) for both
earnings measures together.9
2.4 POTENTIAL CONFOUNDING EFFECTS
Prior empirical research has also examined the relation between earn-
ings/price ratios and returns. Basu [1977], for example, finds a positive

For practical purposes, both dj1-1/Pj,-1and, in particular, the cross-sectional variation in

this variable are sufficiently small so as to have no effect in the empirical analyses which
include an intercept term.
7 Equations (4) and (8) both express returns as a function of earnings levels divided by

beginning-of-period price. Reconciliation of these equations requires:

Ajt= (p I)-'Pj,-,.

This is a familiar definition of earnings under certainty with p equal to the reciprocal of
the required rate of return plus one (Ohlson [1989b]).
8 Ohlson's [1989a] model provides a role for earnings and book value in a dynamic

uncertainty environment that relies on the simple "clean-surplus" relation (that is,
ABVj, = Aj- djt) and the Miller and Modigliani [1961] propositions. While we provide a
more heuristic analysis, it should be noted that relations (4), (7), and (9) are consistent
with the Ohlson model in which earnings and book value are defined as "primitive" and
fundamental economic variables.
9 The variables ujt, vji,and wjtare important in this paper only as they affect the empirical
analyses. Thus, we leave discussion of these terms to section 4.

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
24 P. D. EASTON AND T. S. HARRIS

association between returns and historical earnings/price ratios. This

result potentially affects the interpretation of the empirical results. Since:
AJt1Pjt_1= AAJ.t/Pjt-1+ AJt_1/Pjt_1 (10)
it is not possible to distinguish between (i) the combined explanatory
power (for returns) of change in earnings divided by beginning-period
price and prior-period earnings (i.e., for period t - 2 to t - 1) divided by
the price at the beginning of the return period (At_1/P1)jt_,and (ii) the
explanatory power of current earnings (i.e., for period t - 1 to t) divided
by beginning-of-period price. Inclusion of any two of the variables in
equation (10) as explanatory variables in. a regression model precludes
inclusion of the third. Therefore, we consider the explanatory power of
each of the three earnings variables represented in equation (10) and
then compare these to the explanatory power of the empirical counterpart
of equation (9).
It has also been shown that returns vary with firm size (Banz [1981]
and Reinganum [1981]); size is just one proxy for variables which may
explain security returns but are omitted from equation (9). For example,
size could proxy for a variable which causes ujt (in equation (1)) to differ
from zero. Given past empirical accounting research and the question
being addressed in this paper, we choose not to pursue a search for
potential additional explanatory variables.

3. Data and Sample Selection

The sample is selected from the period 1969-86 using the criteria:
(i) annual earnings per share and the factor to adjust for stock splits and
stock dividends are available on the 1987 Compustat Primary, Secondary,
Tertiary and Full Coverage Annual Industrial File; (ii) security price and
the factor to adjust for stock splits and stock dividends are available on
the Center for Security Prices (CRSP) Daily Returns File for the first
trading day of the ninth month prior to the fiscal year-end; and
(iii) monthly security return data are available on the CRSP Monthly
Returns File for 69 months prior to and 3 months after the fiscal year-
end.
This selection procedure results in a sample of 20,188 firm-year obser-
vations. An additional 192 firm-year observations are deleted because
either A t/Pjt-1, Ajt1/Pjt_1,or Ajt_1/Pjt_1is not between +1.5 and -1.5.1o
Earnings and price variables are adjusted for stock splits and stock
dividends."
10This truncation rule was imposed to ensure that the results were not unduly affected
by a few outlying (unusual) observations. However, truncation did not change the substance
of the results.
All analyses are based on a return period extending from 9 months prior to 3 months
after the fiscal year-end, corresponding roughly with the period between earnings announce-
ments. The analyses were repeated for the subsample of 2,947 firms with no change in
fiscal year-end and the date of announcement of fourth-quarter earnings available on the

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
 EARNINGS AS AN EXPLANATORY VARIABLE 25

4. Empirical Analyses
To better understandthe empiricalvalidity of the models describedin
section 2, and the role of the current level of earnings, we consider first
the correlationsbetween stock returnsand each of the earningsvariables.
The correlations are describedvia univariate regressions to facilitate a
comparisonwith the multivariateregressionwhich empiricallyestimates
the relation expressed in equation (9). The analyses facilitate consider-
ation of the incremental explanatory power of the levels and changes
variablesand the extent to which overall explanatorypower is improved
by the incorporationof the levels variable (Lev [1989]).

4.1 UNIVARIATE ANALYSES

We begin with univariateregressionsof returnsand the earnings levels
and changes variables.The regressionmodels are:
t= ato + ati[Ait/Pjt-l] + EJt (11)
and:
Rj= to + q it A~jt/Pjti] + E2t (12)
where:
Rjt= (APjt + djt)/Pjt-1.
Althoughthe presence of an intercept is not implied by the theoretical
relations that underpin these regressions (equations (4) and (7)), the
omitted variables that may explain security returns (incorporatedin uJt
or vjt)may have, on average,a nonzero effect implyingnonzero intercept
terms (ato and qto).
These and all other regression models are estimated for the pooled
cross-section and time-series sample as well as for each year (t) of
available data. The results from regressions of the models in (11) and
(12) are reportedin table 1. In the regressionusing the pooled sample of
all 19,996 firm-yearobservationsas well as in the annual cross-sectional
regressions, the coefficients at, and Mt, are significantly different from
zero at the 0.01 level. The R2 from the pooled regression based on the
levels model in equation (11) is 7.5% compared to the R2 of 4% from
the equivalent regression for the changes model in equation (12).
For the year-by-yearregressions,the R2 from the levels model is higher
than the R2from the changes model in 14 of the 19 years and is at least
twice as high in 7 of these years. While the R2 from the changes model
(12) is higher than the R2 from the levels model (11) in 5 of the years,

Compustat Quarterly Industrial File. Returns (and abnormal returns) were calculated for
the 12 months up to and including the earnings announcement month. Since the results
for the analyses of this subsample are qualitatively similar to those of the larger sample,
only the results based on the larger sample are reported.

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
 TABLE 1
Simple Regressions of Annual Security Returns on Deflated Earnings Levels and
Earnings Changes
Levels Model:' R,,= cato+ atAjt/Pjt-l + fjt
Changes Model: Rjt= to+ OtiAAjt1Pjt-j+ f2t

Year AjtlPjt-, R2 -"jtlPjt-i

R2 N2
ato at, Oto On

ALL 0.11 0.82 0.075 0.18 0.55 0.040 19,996

(30.8)** (40.3)** (57.0)** (29.0)**
1986 0.19 0.72 0.104 0.21 0.39 0.037 1,459
(20.5)** (13.1)** (22.2)** (7.6)**
1985 0.23 1.08 0.182 0.29 0.61 0.074 1,414
(21.3)** (17.8)** (26.2)** (10.7)**
1984 0.03 0.98 0.195 0.09 0.41 0.036 1,368
(3.5)** (18.2)** (10.9)** (7.2)**
1983 0.27 0.81 0.042 0.31 0.73 0.043 1,333
(14.7)** (7.7)** (18.2)** (7.8)**
1982 0.40 0.55 0.034 0.46 0.62 0.046 1,288
(24.2)** (6.8)** (29.6)** (8.0)**
1981 -0.08 0.86 0.117 0.00 0.50 0.044 1,279
(-6.6)** (13.0)** (0.4) (7.8)**
1980 0.32 0.91 0.084 0.45 0.71 0.070 1,249
(18.1)** (10.8)** (33.2)** (9.7)**
1979 0.02 0.73 0.055 0.11 0.67 0.063 1,212
(1.0) (8.4)** (9.3)** (9.1)**
1978 0.18 0.73 0.050 0.26 0.72 0.075 1,162
(9.9)** (7.9)** (20.2)** (9.8)**
1977 0.04 0.69 0.098 0.10 0.80 0.137 1,105
(3.2)** (11.0)** (11.0)** (13.3)**
1976 0.05 1.05 0.189 0.16 0.63 0.105 1,039
(4.3)** (15.6)** (15.5)** (11.1)**
1975 0.33 0.90 0.133 0.45 0.53 0.066 977
(19.7)** (12.3)** (31.0)** (8.4)**
1974 -0.20 0.78 0.230 -0.12 0.65 0.157 917
(-20.6)** (I16.6)* (-13.8)** (13.1)**
1973 -0.23 1.21 0.099 -0.14 0.70 0.057 833
(-14.7)** (9.6)* (-12.4)** (7.1)**
1972 -0.05 0.70 0.033 -0.02 0.30 0.008 788
(-3.7)** (5.3)** (-1.4) (2.7)**
1971 0.08 1.29 0.094 0.13 0.88 0.045 750
(4.9)** (8.9)** (9.5)** (6.0)**
1970 0.01 1.39 0.098 0.09 1.12 0.054 678
(0.4) (8.6)** (8.0)** (6.2)**
1969 -0.20 1.46 0.072 -0.12 1.33 0.055 612
(-12.2)** (6.9)** (-10.2)** (6.1)**
1968 0.12 2.52 0.090 0.27 1.66 0.049 533
(4.3)** (7.3)** (16.4)** (5.2)**
Mean3 1.02 0.74
(10.0)** (9.7)**
(t-statistics are provided in parentheses.)
* Significant at 0.01 c ayc 0.05.
** Significant at a c 0.01.
'Description of regression variables: Rj, is the return on a share of firm i over the 12 months
extending from 9 months prior to the fiscal year-end to 3 months after the fiscal year-end, Aj, is
accounting earnings per share of firm I for period t, and Pj,, is the price per share of firm j at time
t - 1.
2N is the number of observations in the regression.
3This is the mean of the yearly coefficients, estimated to test for the effect of cross-sectional
correlations in the error terms.

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
 EARNINGS AS AN EXPLANATORY VARIABLE 27
there are no years in which it is twice as high.12 These results indicate
that, as the models described in section 2 suggest, the current earnings
level variable,Aj1/Pj1,-,is correlatedwith stock returns.These univariate
results suggest also that we can expect both levels and changes variables
to be associated with returns.13
Given the equality expressed in equation (10), we also consider the
association between Aj,-,IPj,-, and returns by estimating the following
regressionmodel:
Rjt = Oto+ Ot[Ajt-1/Pjt_1]+ t (13)
The results from the regressions based on model (13) are reported in
table 2. The coefficients from the pooled regression are statistically
significant at the 0.01 level. However,the R2 is 0.003 as comparedto R2
of 0.075 and 0.040 from the pooled regressionsof equations (11) and (12),
respectively. The results from the annual cross-sectional regressions
indicate statistically significant Ot,coefficients (at the 0.01 level) in 7 of
the 19 years. In all years the R2 from regressions of equation (13) are
lower than the R2 from regressions of the levels model (11) by several
multiples. While this is generally also true of comparisonsbetween the
R2from regressionsof the models in equations (13) and (12), in one year
(1984) the regression based on Ajt-1/Pjt-lhas an R2 of 0.059 while the
regressionbased on Ajt1/Pjtjhas an R2of 0.036.
As a result of the equality expressed in equation (10) it is necessary to
interpretthe univariateregressionswith caution. But, it is useful to note
that there are years (for example, 1973 and 1976) in which the R2 from
regressionson AjtllPjt-l are zero while the R2sfrom the regressionbased
on A1t/P1t-,are at least 80% higher than the R2s from the equivalent
regressionsbased on' AAtl/Pjt-.Thus, overall, these results suggest that
the difference in associations between security returns and the earnings
levels (Ajt/Pjt-,)and earningschanges (SAAt/Pjt-l)variablesreflects more
than the earnings/priceeffect documentedby, for example, Basu [1977].
Another reason to interpret the results with caution is the potential
bias in the coefficients due to cross-sectional correlation in the error
terms of the regressions. Bernard [1987] suggests that for regressions
based on annual returns, if it is assumed that each annual regressionis
independent, then the mean and standard error of the coefficients ob-
tained from the annual regressions may be used to test whether this
mean is statistically different from zero. If it is, then the bias from any
cross-sectionalcorrelationwill not be sufficient to negate the statistical
12 There is no statistical reason to use the comparison of R2 being twice as high. We use

this simply to provide a sense of the order of magnitude of the difference. Our focus on R'
stems from the emphasis on this statistic in recent review papers by Bernard [1989] and
Lev [1989] who note that this statistic has been consistently low in earnings/returns
association studies.
13 We also calculated the Spearman correlations between stock returns and each of the

earnings variables (not reported). The results were qualitatively the same as the correlations
which can be imputed from the R2s reported in table 1.

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
28 P. D. EASTON AND T. S. HARRIS

TABLE 2
Regressionsof Annual SecurityReturnsand Prior-PeriodEarningsDividedby
Beginning-of-PeriodPrice
Prior Earnings Model:' Rjt = Oto+ OtAjt- /P1t- + Et

Year OtO Oti R2 N2

ALL 0.17 0.19 0.003 19,996
(45.9)** (8.4)**
1986 0.21 0.25 0.011 1,459
(20.7)** (4.2)**
1985 0.26 0.35 0.016 1,414
(21.7)** (4.9)**
1984 0.07 0.55 0.059 1,368
(8.2)** (9.3)**
1983 0.33 -0.10 0.000 1,333
(18.4)** (-1.0)
1982 0.45 -0.14 0.002 1,288
(24.3)** (-1.5)
1981 -0.03 0.37 0.016 1,279
(-2.1)** (4.7)**
1980 0.46 -0.09 0.000 1,249
(22.4)** (-0.9)
1979 0.16 -0.26 0.005 1,212
(9.3)** (-2.5)*
1978 0.32 -0.34 0.011 1,162
(19.9)** (-3-7)**
1977 0.13 -0.14 0.004 1,105
(11.2)** (-2.2)*
1976 0.18 0.06 0.000 1,039
(14.8)** (0.9)
1975 0.42 0.16 0.003 977
(22.5)** (2.0)*
1974 -0.15 0.34 0.018 917
(-11.8)** (4.2)**
1973 -0.12 0.01 0.000 833
(-8.6)** (0.1)
1972 -0.02 0.21 0.004 788
(-1.5) (1.8)
1971 0.12 0.38 0.007 750
(7.4)** (2.5)*
1970 0.04 0.66 0.013 678
(2.1)* (3.2)**
1969 -0.15 0.60 0.005 612
(-6.6)** (1.7)
1968 0.26 0.45 0.003 533
(10.0)** (1.3)
Mean3 0.18
(2.5)*
(t-statisticsare providedin parentheses.)
* Significantat 0.01 c a c 0.05.
**Significantat a - 0.01.
1Description of regressionvariables:Rjtis the return on a share of firm j over the 12 months
extending from 9 months prior to the fiscal year-end to 3 months after the fiscal year-end,Ajt is
accountingearningsper share of firmj for period t, and Pj,_1is the price per share of firmj at time
t - 1.
2 N is the numberof observationsin the regression.
This is the mean of the yearly coefficients, estimated to test for the effect of cross-sectional
correlationsin the errorterms.

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
 EARNINGS AS AN EXPLANATORY VARIABLE 29
relevance of the variable. This calculation is reportedin the last line of
tables 1 and 2. Both coefficients a,, and /t, are statistically different
from zero at the 0.01 level while Ot,is statistically different from zero at
the 0.05 level. Thus, the significance of the earnings coefficients is
unlikely to be a result of potential cross-sectionalcorrelations.Although
in this paperwe are not concernedwith the magnitudesof the coefficients,
it is interesting to observe, from the test for the effect of cross-sectional
correlations in the errors, that the coefficient ait is, on average, not
significantly different from one. This result is consistent with the book
value valuation model in equation (4).14

4.2 MULTIVARIATE ANALYSES

The model of returns incorporating both the earnings levels and
changes measures, as summarizedby equation (9), is examined empiri-
cally via the following cross-sectional regressions:
Rjt= yot + -yt[At1/Pjt_1]+ y2t[AjtA/Pit_1]+ Ejt. (14)
In moving to the multivariateregressionwith both earnings levels and
changes as explanatory variables, we choose to use the current levels
variable (Ajt/Pjt-,) because it is the levels variable which is consistent
with the theory. Given the equality expressed in equation (10), we can
obtain the same amount of explanatorypowerin the multivariateregres-
sions by includingany two of the three earningsvariables.The discussion
in section 2.2 and the theoretical model in Ohlson [1989b] justify the
inclusion of the two variables we have chosen. On the other hand,
we know of no theoretical model that justifies the inclusion of both
Ajt_11/Pt_1and either Ajt/Pjt-lor AAjt1/Pjt1. The results based on (14) are
reportedin table 3.
The regressionusing the pooled (all years) sample yields an estimated
coefficient yutof 0.71 (t-statistic = 28.3). Further, Yit is significant at the
0.05 level or better in all 19 years. The coefficient Y2tis significant in the
pooled regression(t-statistic = 7.1) and in 8 of the 19 years (one of these
significant coefficients is negative).15 Comparingthe R2s from the mul-
tivariate regressions and the univariate regressions discussed in the
previous subsection (reportedin table 1), we naturally obtain a similar
picture. In 11 of the years when AAjt1/Pjt_1 is added to the regression
model (11) the change in R2 is insignificant; that is, the addition of
14
Equation (9) suggests the existence of correlated omitted variables in the univariate
regression models so that we expect these models to be misspecified. However, in the
interest of facilitating interpretation of these univariate regressions, we conducted addi-
tional tests of the properties of the error terms. We could not reject (at the 0.01 level) a
null hypothesis that the error terms were normally distributed in all cases. Also, relevant
scatter plots did not reveal any nonlinearities.
15 All reported results are based on the sample of 19,996 firm-year observations. This

sample excludes the 192 observations with unusual earnings numbers (see section 3). For
the pooled regression using the sample of 20,188 observations, jit is 0.20 (t-statistic = 16.4)
while i2t is 0.04 (t-statistic = 4.3).

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
 TABLE 3
EarningsLevelsand MultipleRegressionsof Annual SecurityReturnsand
DeflatedChanges
Model:' Rjt= yot+ 'yt [Ajt/Pjt-l] + Y2d[AAt1/Pjtj]+ Et
Year Yot 'ylt 72t H2 N2
ALL 0.12 0.71 0.16 0.077 19,996
(31.6)** (28.3)** (7.1)**
1986 0.20 0.70 0.03 0.104 1,459
(20.4)** (10.4)** (0.5)
1985 0.23 1.05 0.04 0.182 1,414
(20.6)** (13.7)** (0.5)
1984 0.03 1.06 -0.13 0.197 1,368
(3.1)** (16.6)** (-2.2)*
1983 0.28 0.55 0.51 0.058 1,333
(15.2)** (4.7)** (4-9)**
1982 0.43 0.24 0.47 0.049 1,288
(23.9)** (2.3)* (4.7)**
1981 -0.09 0.88 -0.03 0.116 1,279
(-6.2)** (10.3)** (-0.3)
1980 0.36 0.65 0.32 0.091 1,249
(16.8)** (5.5)** (3.2)**
1979 0.19 0.34 0.02 0.069 1,212
(3.1)** (2.9)** (4.5)**
1978 0.22 0.32 0.57 0.081 1,162
(11.5)** (2.9)** (6.3)**
1977 0.06 0.38 0.61 0.158 1,105
(9.4)** (5.3)** (8.9)**
1976 0.06 0.88 0.26 0.201 1,039
(27.3)** (5.1)** (4.0)**
1975 0.34 0.83 0.09 0.133 977
(2.6)** (18.2)** (1.1)
1974 -0.19 0.71 0.09 0.231 917
(-16.9)** (9.4)** (1.2)
1973 -0.22 1.06 0.19 0.101 833
(-13.2)** (6.5)** (1.5)
1972 -0.05 0.69 0.01 0.032 788
(-3.6)** (4.5)** (0.1)
1971 0.08 1.13 0.31 0.096 750
(5.2)** (6.6)** (1.8)
1970 0.02 1.23 0.26 0.097 678
(1.0) (5.8)** (1.1)
1969 -0.19 1.22 0.31 0.071 612
(-8.1)** (3.4)** (0.8)
1968 0.14 2.14 0.68 0.094 533
(4-7)** (5.4)** (1.9)
Mean3 0.85 0.24
(8.4)** (4.4)**
(t-statisticsare providedin parentheses.)
* Significantat 0.01 c a c 0.05.
**Significantat a c 0.01.
Description of regressionvariables:Rjt is the return on a share of firm j over the 12 months
extending from 9 months prior to the fiscal year-end to 3 months after the fiscal year-end,Ajt is
accountingearningsper share of firmj for period t, and Pjt-l is the price per share of firmj at time
t - 1.
2 N is the numberof observationsin the regression.

3This is the mean of the yearly coefficients, estimated to test for the effect of cross-sectional
correlationsin the errorterms.
30

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
 EARNINGS AS AN EXPLANATORY VARIABLE 31

SAjA/Pj,_1 will yield a ("partial")F-statistic which is not significant at

the 0.05 level.16On the other hand, there is not one year in which the
inclusion of Aj1/Pj,-1in the regression model (12) does not yield a
significant (at the 0.05 level) improvementin R2.
An obvious question which arises is the extent to which the results
might be affected by collinearity among the variables in (14). For each
regressionwe computed the condition indexes which Belsley, Kuh, and
Welsch [1980] (henceforth BKW) advocate as the primary measure for
detecting collinearity. BKW suggest that mild multicollinearityexists if
the maximumcondition index is between 5 and 10, and potentially severe
multicollinearityexists if the condition index is over 30.1' The highest
condition index we obtained was 4, and in most yearly regressions the
maximum condition index was below 2.5.18 Thus, collinearity does not
seem to influence our results.
As an additional specification test we performedseveral analyses on
the residuals from each multivariate regression.These included checks
for normalityand considerationof variousscatterplots. A null hypothesis
of normality could not be rejected at the 0.01 level in all cases, and the
plots revealed some heteroscedasticity but no other obvious problems.
To ensure that the inferences made are not affected by any inefficiency
caused by the heteroscedasticitywe calculated the t-statistics after cor-
recting for the heteroscedasticity in the manner described by White
[1980].These t-statistics are not reportedas they are qualitativelysimilar
to those reportedin table 3, and it is easier to make comparisonsbetween
the univariate and multivariate regressionsusing the standardordinary
least squaresestimates.
In a manner similar to that describedin subsection 4.1, we tested for
the effect on inferences about the coefficients from potential cross-
sectional correlations in the error terms. The results of these tests are
reported at the bottom of table 3 and indicate that any such cross-
sectional correlationdoes not bias against concludingthat each variable's
coefficient is statistically significant.
Overall, the evidence suggests that both the current earnings levels
variable (Ajt/Pjt-,) and the earnings changes variable (Ajt1/Pjt_1)are
relevant for explaining returns, and the two variables are not just sub-
stitutes. They are complements in the sense that, for the pooled sample
and for several individualyears, significantly more of the cross-sectional
variation in returns is explained by both earnings levels and earnings
changes than is explained by either variable consideredalone.
16 The R2 reported for the multivariate regressions are adjusted for degrees of freedom.

"Partial" F-statistics may be calculated directly from the t-statistics since the "partial"
F-statistic is the square of the t-statistic on the added variable.
17 Warga [1989] shows that the condition index is a valid diagnostic for financial return

data and that the bounds established by BKW are robust.

18 We also calculated a variance inflation factor (BKW [1980, p. 93]) for each regression.

The critical value for indicating severe collinearity from this statistic is 10. In no case was
the factor above 3 in any of our regressions.

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
32 P. D. EASTON AND T. S. HARRIS

5. Earnings Levels as a Measure of Unexpected Earnings

The precedinganalysis considersthe associationbetween earningsand
returns based on the valuation models described in section 2. The
empirical results reported in section 4 suggest current earnings as an
explanatoryvariablefor returns.Since the focus of much of the literature
on the relation between earnings and returns has been on unexpected
earnings (sometimes measuredas zAjt1/Pjt1)and unexpected returns,we
evaluate whether the current earnings variable, AjtlPjt-,, is relevant as
an explanatoryvariable for unexpected returns.
One model of the unexpected earnings/unexpectedreturns relation is:
Rjt- E[Rjt] = a, + a2l(Ajt-Ajt-)IPjt-ll + ejt (15)

where a, and a2 are regressioncoefficients, Rjtis the return on securityj

for time t, and ejtis the regressiondisturbanceterm.
One interpretation of this model is that Ajtj represents expected
earnings (if earnings follow a randomwalk) and Ajt- Ajt-l is then scaled
by Pjt-l to conformwith the left-hand-sidereturnvariable.An alternative
interpretation of (15) is that lAjt/Pjt-1- Ajt11/Pjt-1lrepresents scaled
unexpected earnings with Ajt11/Pjt-1serving as the measure of expected
Ajt/Pjt-1.If, as the evidence in Beaver and Morse [1978] suggests, earn-
ings/price ratios are mean-reverting,Ajt-llPjt-l may be a poor measure
of expected Ajt/Pjt-1.Consider the extreme situation where earnings/
price ratios revert immediatelyto a mean and the reversionis driven by
the earnings variable. In this circumstance, a cross-sectional constant,
Kt, would be a more precise measure of expected Ajt/Pjt-1.That is, an
alternativeto (15) would be:

{Rjt- E[Rjt]l = a' + a' lAjt/Pjt-1- Kt} + ejt (16)

where Kt is a cross-sectionalconstant.
In this case, current earnings levels divided by beginning-of-period
price will have more explanatorypower for unexpected security returns
(equation (16)) than do earnings changes dividedby beginning-of-period
price (equation (15)).'9 Although this extreme situation may not be
observedas a practical manner, it demonstratesthe point that earnings
levels divided by beginning-of-periodprice may be a useful alternative
measureof unexpected earnings.
We do not wish to imply that the relations in equations (15) and (16)
are derived from formal models. Rather they are representationsof the
presumed association between unexpected earnings and unexpected re-
turns first hypothesized by Ball and Brown [1968]. Consequently, in
contrast to the previous analysis in which the relevance of the earnings
19We are grateful to a referee for suggestions that considerably improved the articulation
of this argument. A similar argument is made in Biddle and Seow [1990].

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
 EARNINGS AS AN EXPLANATORY VARIABLE 33

levels variable was based on a valuation model, in the current context,

Aj1/Pj,1-is more appropriately interpreted as an alternative construct for
the generic notion of unexpected earnings.
Brown et al. [1987] demonstrate that multiple proxies for unexpected
earnings may reduce the measurement error bias in regression estimates
of the coefficients relating unexpected earnings and unexpected returns.
In view of the fact that change in earnings has been a dominant measure
of unexpected earnings, we consider both earnings changes (equation
(15)) and earnings levels (equation (16)) as measures of unexpected
earnings using the procedure described by Brown et al.
Our measure of unexpected return is the residual from the market
model:
Rjt1= 30j + /3ijRmt+ zjt (17)
where Rmtis the CRSP equally weighted market index for month t and
(17) is estimated over the 60 months prior to the month of the accumu-
lation of returns.
Similarly to Brown et al. [1987], we consider the following regression
model:
CARjt = 'hOt + *1t UE1jt + *2t UE2jt + n7it (18)
where CARjtis the cumulative abnormal return (monthly residuals from
the monthly market model (equation (17)) calculated over the 12 months
extending from 9 months prior to through 3 months after the fiscal year-
end, UE1jtis measure 1 of unexpected earnings Ajtl/Pjtl, and UE2jtis
measure 2 of unexpected earnings -Ajt1Pjt-j .
These regressions are conducted for the pooled sample and each year
t of available data. Results are reported in table 4. The coefficients (*j')
on the earnings levels variable (Ajt/Pjt-1)are significantly different from
zero (at the 0.01 level) in 12 of the 19 years while the coefficients (0J2t)
on the earnings changes variable (\Ajt1/Pjt-l)are significant in 15 of the
19 years at the same significance level. In the regression on the pooled
sample, J'lthas a t-statistic of 14.0 and *J2thas a t-statistic of 23.0. Thus,
the level of earnings is associated with raw returns, as anticipated by the
valuation models described in section 2, and with unexpected returns. In
the spirit of Brown et al. [1987], earnings levels might be used to mitigate
the effects of error when unexpected earnings are measured using earn-
ings changes (that is, UE2t).

6. Summary and Conclusions

This study demonstrates an association between the level of current
accounting earnings divided by beginning-of-period price (Ajt/Pjt-1) and
stock returns. These results are consistent with more detailed theoretical
models in Demski and Sappington [1989] and, in particular, in Ohlson

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
34 P. D. EASTON AND T. S. HARRIS

TABLE 4
MultipleRegressionsof MarketModelCumulativeAbnormalReturnsand Deflated
EarningsLevelsand EarningsChanges
Model:' CARjtgot + TIt[Ajt1/Pjt-]+ T2t [Ajtl/Pjt11] + 7ljt

Year Tot Tit T2t K2 N2

ALL -0.02 0.27 0.41 0.078 19,996
(-7.5)** (14.0)** (23.0)**
1986 -0.02 0.8 0.26 0.027 1,459
(-2.3)* (1.2) (4.5)**
1985 -0.02 0.59 0.18 0.125 1,413
(-2.4)* (9.0)** (3.2)**
1984 -0.03 0.63 0.21 0.140 1,368
(-3.8)** (10.4)** (3.6)**
1983 0.01 0.17 0.63 0.098 1,333
(0.4) (2.3)* (9.5)**
1982 0.05 0.12 0.43 0.061 1,288
(3.9)** (1.6) (6.0)**
1981 -0.03 0.46 0.37 0.108 1,279
(-2.0)* (5.7)** (4.9)**
1980 -0.02 0.26 0.36 0.089 1,249
(-1.4) (3.3)** (5.3)**
1979 -0.02 0.17 0.55 0.069 1,212
(-1.2) (1.5) (5-7)**
1978 0.03 -0.09 0.73 0.087 1,162
(1.7) (-0.9) (9.1)**
1977 -0.06 0.13 0.68 0.125 1,105
(-5.4)** (1.9) (9.9)**
1976 -0.09 0.55 0.35 0.153 1,039
(-7.5)** (7.4)** (5.9)**
1975 -0.08 0.41 0.13 0.086 977
(-6.2)** (5.9)** (2.2)*
1974 -0.18 0.60 0.14 0.172 917
(-14.7)** (7-3)** (1.7)
1973 -0.10 0.67 0.90 0.139 833
(-5.4)** (3.5)** (6.3)**
1972 0.06 0.27 0.57 0.046 788
(4.1)** (1.7) (4-4)**
1971 -0.04 0.53 0.69 0.091 750
(-3.0)** (3.7)** (4.9)**
1970 -0.05 1.26 0.14 0.111 678
(-3-1)** (6.6)** (0.7)
1969 0.03 0.45 1.14 0.048 612
(1.0) (1.0) (2.5)*
1968 -0.18 2.62 0.40 0.172 533
(-7.5)** (8.3)** (1.4)
(t-statisticsare providedin parentheses.)
* Significantat 0.01 c a c 0.05.
** Significantat a < 0.01.
1 Descriptionof regressionvariables:CARjt is the cumulativeabnormalreturnon a share of firmj
over the 12 months extendingfrom 9 months priorto the fiscal year-endto 3 months after the fiscal
year-end,Ajtis accountingearningsper share of firmj for period t, and Pjti is the price per share of
firm j at time t - 1.
2
N is the numberof observationsin the regression.

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
 EARNINGS AS AN EXPLANATORY VARIABLE 35

[1989a; 1989b], which suggest earnings divided by beginning-of-period

price as an explanatory variable for returns. In multivariate regressions
of security returns on both the current earnings level and the earnings
change variables, both coefficients are generally significantly different
from zero. This result suggests that both earnings variables play a role
in security valuation. Similar results are obtained from multiple regres-
sions of abnormal returns on the two earnings variables included as
proxies for unexpected earnings.

REFERENCES
BALL, R., AND P. BROWN. "An Empirical Evaluation of Accounting Income Numbers."
Journal of Accounting Research (Autumn 1968): 159-78.
BANZ, R. W. "The Relationship Between Return and Market Value of Common Stocks."
Journal of Financial Economics (March 1981): 3-18.
BARTH, M. E. "Assessing Financial Accounting Measurement Alternatives for Assets and
Liabilities." Working paper, Stanford University, January 1989.
BASU, S., "Investment Performance of Common Stocks in Relation to Their Price-Earnings
Ratio: A Test of the Efficient Market Hypothesis." Journal of Finance (June 1977):
663-82.
BEAVER, W., AND D. MORSE. "What Determines Price-Earnings Ratios?" Financial Ana-
lysts Journal (July/August 1978): 65-76.
BEAVER, W.; R. LAMBERT; AND D. MORSE. "The Information Content of Security Prices."
Journal of Accounting and Economics (March 1980): 3-28.
BELSLEY, D. A.; E. KUH; AND R. E. WELSCH. Regression Diagnostics: Identifying Influential
Data and Sources of Collinearity. New York: Wiley, 1980.
BERNARD, V. L. "Cross-Sectional Dependence and Problems in Inference in Market-Based
Accounting Research." Journal of Accounting Research (Spring 1987): 1-48.
. "Capital Market Research in Accounting During the 1980s: A Critical Review."
Paper prepared in honor of the University of Illinois Accountancy Ph.D. Program Golden
Jubilee Symposium, July 1989.
BIDDLE, G. C., AND G. S. SEOW. "The Information Content of Earnings: Evidence from
Cross-Industry Comparisons." Working paper, University of Washington, July 1989.
BROWN, L.; P. GRIFFIN; R. HAGERMAN; AND M. ZMIJEWSKI. "An Evaluation of Alternative
Proxies for the Market's Assessment of Unexpected Earnings." Journal of Accounting
and Economics (July 1987): 159-94.
COLLINS, D. W., AND S. P. KOTHARI. "An Analysis of Intertemporal and Cross-Sectional
Determinants of Earnings Response Coefficients." Journal of Accounting and Economics
(July 1989): 143-81.
DEMSKI, J., AND D. SAPPINGTON. "Fully Revealing Income Measurement and Cost Allo-
cation." Working paper, Yale University, January 1989.
HARRIS, T. S., AND J. A. OHLSON. "Accounting Disclosures and the Market Valuation of
Oil and Gas Properties." The Accounting Review (October 1987): 651-70.
LANDSMAN, W. "An Empirical Investigation of Pensions and Property Rights." The
Accounting Review (October 1986): 953-1001.
LEV, B. "On the Usefulness of Earnings and Earnings Research: Lessons and Directions
from Two Decades of Empirical Research." Journal of Accounting Research (Supplement
1989): 153-92.
MILLER, M., AND F. MODIGLIANI. "Dividend Policy, Growth, and the Valuation of Shares."
Journal of Business (October 1961): 411-33.
OHLSON, J. A. "Accounting Earnings, Book Value, and Dividends: The Theory of the Clean
Surplus Equation (Part I)." Working paper, Columbia University, March 1989a.

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions
36 P. D. EASTON AND T. S. HARRIS

. "The Theory of Value and Earnings and an Introduction to the Ball-Brown

Analysis." Working paper, Columbia University, March 1989b.
REINGANUM, M. R. "Misspecification in Capital Asset Pricing: Empirical Anomalies Based
on Earnings Yields and Market Values." Journal of Financial Economics (March 1981):
19-46.
WARGA, A. "Experimental Design in Tests of Linear Factor Models." Journal of Business
& Economic Statistics (April 1989): 191-98.
WHITE, H. "A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct
Test for Heteroskedasticity." Econometrica (May 1980): 817-38.

This content downloaded from 139.184.14.159 on Wed, 19 Aug 2015 08:51:24 UTC
All use subject to JSTOR Terms and Conditions