5- Anticipatory income smoothing a re-examination.pdf

7/27/2019 5- Anticipatory income smoothing a re-examination.pdf

http://slidepdf.com/reader/full/5-anticipatory-income-smoothing-a-re-examinationpdf 1/18

Journal of Accounting and Economics 35 (2003) 405–422

Anticipatory income smoothing:

a re-examination$

Pieter T. Elgers, Ray J. Pfeiffer Jr*, Susan L. Porter

Department of Accounting and Information Systems, Isenberg School of Management,University of Massachusetts, Amherst, MA 01003-4915, USA

Received 1 August 2002; accepted 28 May 2003

Abstract

This paper reassesses evidence of anticipatory income smoothing reported in DeFond and

Park (DP) (J. Accounting Econom. 23 (1997) 115) in light of knowledge about measurement

error in discretionary accrual estimates. We argue that the method DP use to measure un-

managed earnings mechanically biases the evidence in a manner consistent with anticipatoryincome smoothing. Using an approximate randomization approach, we find that DP’s results

cannot be distinguished from those achieved when discretionary accruals are randomly

assigned to firm-years in our sample. Overall, these results show that the ‘backing out’

approach to measuring un-managed earnings is ineffective in testing earnings management

hypotheses.

r 2003 Elsevier B.V. All rights reserved.

JEL classification: M41

Keywords: Earnings management; Smoothing

1. Introduction

Earnings management through the use of discretionary accruals to achieve target

levels of reported earnings has been the subject of considerable public press,

$We are grateful to Bill Brown, Mark DeFond, the workshop participants at the University of

Connecticut, and an anonymous reviewer for helpful comments and suggestions and for research support

provided by the Isenberg School of Management at the University of Massachusetts. We thank I/B/E/SInternational, Inc. for providing earnings per share forecast data at an academic rate.

*Corresponding author. Tel.: +1-413-545-5653; fax: +1-413-545-3858.

E-mail address: [email protected] (R.J. Pfeiffer Jr).

0165-4101/$ - see front matterr 2003 Elsevier B.V. All rights reserved.

doi:10.1016/S0165-4101(03)00039-9



regulatory attention, and academic study. One particular manifestation of earnings

management is ‘smoothing,’ or managing reported earnings to achieve earnings

targets across accounting periods. A recent study, DeFond and Park (1997) (DP),

draws upon Fudenberg and Tirole (1995) to predict and test hypotheses thatmanagers smooth earnings to meet earnings targets. Specifically, they propose that

managers consider both current-year earnings and expected year-ahead earnings

when making decisions about current-year discretionary accruals.

A fundamental issue in assessing earnings management is the unobservability of

the managed and un-managed components of reported earnings. Much of prior

research relies upon the Jones (1991) model of unmanaged earnings. Recent

evidence, however, calls into question the precision and power of discretionary

accruals estimates using the Jones model (e.g., Dechow et al., 1995; Guay et al.,

1996).

The paper’s primary purpose is to reassess whether the ‘backing-out’ approach to

defining un-managed earnings influences the DP results. Our analysis indicates that

classifying firm-years in this manner, where un-managed earnings is defined as the

difference between reported earnings and an error-prone discretionary accrual

estimate, creates patterns of measured discretionary accruals that appear to support

the anticipatory income smoothing hypothesis, even in the absence of earnings

management. Lim and Lustgarten (2000) examine this problem in a regression-based

setting and illustrate similar biases in other earnings-management studies. DeFond

and Park acknowledge this self-selection problem throughout their paper and

conduct several empirical tests to address the self-selection issue as an alternativeexplanation. However, DP are unable to rule out the self-selection bias as an

alternative explanation.

To address the problem of measuring managed and unmanaged components of

earnings, we use an approximate randomization procedure to generate an empirical

distribution of the discretionary accrual characteristics for each pair-wise classifica-

tion of current and expected future relative earnings performance. We then compare

the actual statistics obtained using the DP methodology to the empirical distribution

as a means of assessing whether the observed behavior is sufficient in magnitude to

exceed the thresholds using conventional confidence intervals.1 We find that the

results using the DP methodology are statistically indistinguishable from the randomresults. We conclude that the ‘backing out’ method is not capable of providing

evidence of smoothing.

We next extend our analysis in two ways. First, Abarbanell and Lehavy (2000)

hypothesize that if financial analysts are not motivated to or able to perfectly predict

the extent of earnings management, their forecasts may be proxies for the un-

managed component of earnings (and thus their forecast errors may be proxies for

the managed component of earnings). For purposes of illustration, we show that

even though analysts’ forecast errors are potentially less error-prone than modified-

Jones-model estimated discretionary accruals, replacing discretionary accruals with

forecast errors does not solve the ‘backing out’ problem in this context. We obtain

1For a more detailed description of applications of approximate randomization tests, see Noreen (1989).

P.T. Elgers et al. / Journal of Accounting and Economics 35 (2003) 405–422406



patterns of forecast errors that, like those from the DP methodology, appear to be

supportive of the hypothesized smoothing behavior. However, the analysts’ forecast

results are also indistinguishable from the results obtained when randomly assigning

forecast errors across firm-years.In the second extension of our analysis, we investigate the use of operating cash

flow as a proxy for un-managed earnings. As one of their robustness tests, DP

employ this approach. Also, more recently, Ahmed et al. (2001) use cash flows in an

attempt to avoid the problems created by the ‘backing out’ approach. Importantly,

DP and Ahmed, Lobo, and Zhou (ALZ) implement this idea slightly differently: DP

compute operating cash flow as the difference between earnings and measured total

accruals, while ALZ use disclosed operating cash flow from firms’ Statements of

Cash Flows.

We show that DP’s means of defining operating cash flow, referred to by ALZ and

by Collins and Hribar (1999) as the ‘balance sheet approach,’ is in itself a form of

‘backing out’ that potentially leads to the same sort of mechanical pattern of

discretionary accruals as we illustrate for the DP methodology. Using cash flows

defined in this manner yields results that cannot be discriminated from those using a

random assignment of discretionary accruals across firm-years.

In addition to problems caused by measurement error, a second issue in

addressing earnings management is that economic theory is at odds about whether

managing earnings is a rational economic decision on the part of managers. While

Fudenberg and Tirole (1995) are motivated by the fact that earnings is a readily

available and often-used basis for contracting, their model is of a profit centermanager, for whom there are no available measures of equity-market-based

performance. For Chief Executive Officers, other relevant information about firms

is available for use in contracting, including equity market performance. The

presence of these other means of solving contracting problems weakens the

applicability of the earnings smoothing hypothesis because earnings manipulation

has at best a very indirect effect on the termination decision.

The remainder of the paper is organized as follows. Section 2 develops

our hypotheses, explains the research design, describes the sample selection

criteria, and presents the empirical results. Section 3 summarizes and concludes

the paper.

2. Research design, sample selection and empirical results

We seek to replicate and re-evaluate the evidence of anticipatory income

smoothing presented in DP. For this reason, we adopt the same the research

design, sample selection and variable definitions. We begin with a replication of the

results in DP, which are based on the modified-Jones model of discretionary

accruals. We next apply a randomization approach to assessing the statistical

significance of the observed sample statistics in the presence of measurement error inthe discretionary accrual proxy, and we use that approach to show that it is not

possible to test the smoothing hypothesis with this methodology.

P.T. Elgers et al. / Journal of Accounting and Economics 35 (2003) 405–422 407



2.1. Research design and sample selection

Based on the predictions of their theory, DP classify firm-years into three groups:

(1) firm-years where managers have incentives to manage current earningsdownward; (2) firm-years where managers have incentives to manage earnings

upward; and (3) firm-years for which the theory makes no prediction regarding

managers’ incentives. The grouping is based on current-year and expected next-year

un-managed earnings relative to annual industry medians. For example, in firm-

years where current-year un-managed earnings are relatively low but expected year-

ahead earnings are relatively high, managers are hypothesized to have the incentive

to ‘borrow’ from year-ahead earnings to avoid reporting of sub-standard current

year earnings.

DP examine descriptive statistics of the measured discretionary accruals for the

firm-years in each group. For firm-years where managers are predicted to manage

earnings upward (downward), average discretionary accruals will be positive

(negative). Their results indicate that discretionary accruals fit this prediction very

well: upward (downward) smoothing appears in 87% (92%) of the predicted cases.

The key research design issues are the measurement of discretionary accruals, the

measurement of management’s expectation of year-ahead earnings, and the

classification of current and expected year-ahead un-managed earnings as ‘high’ or

‘low.’ We follow DP’s sample selection and measurement choices to ensure

comparability of our results, as described in the following sections.

2.1.1. Measuring discretionary accruals

Discretionary accruals are computed using the modified Jones model as was

employed in DP. Un-managed earnings in a given year are equal to reported

(asset-scaled) earnings before extraordinary items less estimated discretionary

accruals.

2.1.2. Measuring management’s expectation of year-ahead earnings

Using the I/B/E/S Summary History data updated through March of 1999, we

obtain consensus analysts’ forecasts of year t þ 1 earnings made during March of year t þ 1 for all December fiscal year firms. These forecasts serve as management’s

expectation of period t þ 1 un-managed earnings as of the time that they are making

their discretionary accrual decision for year t; as in DP.2 In each year, firms’ un-

managed earnings are classified as ‘good’ or ‘poor’ relative to the 2-digit SIC-code

industry median un-managed earnings.

2DeFond and Park (1997, pp. 120–121) acknowledge that this design choice presumes (1) that managers

choose their discretionary accruals for year t just prior to their announcement of period t earnings; and (2)

that analysts’ March t þ 1 forecasts of year t þ 1 earnings are proxies for managers’ expectations of

year t þ 1 earnings. DP’s sensitivity results indicate that their inferences are not affected by the use

of actual earnings in t þ 1 as proxies for managers’ expectations, suggesting that this is not a critical

design choice.




2.1.3. Sample selection

We duplicate DP’s sample selection. We begin with 152,883 firm-years on the 1998

Standard and Poors’ Compustat PC-Plus database of active and research companies

having a CUSIP identifier and an SIC code. We then retain firms in industries with atleast 20 members in each year. We further require that all variables needed for our

subsequent analyses be available in the I/B/E/S and Compustat databases (including

the items needed to compute the modified-Jones-model discretionary accrual

estimates). This results in a sample of 14,684 firm-years. We eliminate the extreme

1% of firm-years based on scaled discretionary accruals, non-discretionary accruals,

and operating cash flows, and we eliminate firm-years having less than $1 million in

assets. Finally, we exclude all financial institutions and unclassified firms (SIC

between 5999 and 7000 and SIC=9999). These selection criteria yield a sample of

14,194 firm years, which compares with DP’s sample of 10,167 analysts’ forecast

errors based on 1994 Compustat and I/B/E/S data.

We add one last criterion not included by DP. We exclude all non-December

fiscal-year-end firm-years for the following reason: management is posited to act

strategically based on its projection of the firm’s position relative to the industry

median. When all industry members do not share the same fiscal year, it is difficult

for the management to identify the median firm, for it would require comparisons of

industry members across many months of the year. This last criterion results in a

sample of 8,535 firm-years, and as shown later in Table 2, does not appear to affect

our ability to replicate the results in DP. Table 1 contains descriptive statistics for

our sample and shows that the means, medians and inter-quartile distributions of thevariables in our sample correspond closely to those in DeFond and Park (1997,

Table 1, p. 123).

2.2. Results

We present the results in the following sequence. We begin with a straightforward

replication of the main results supportive of anticipatory income smoothing reported

in DP. Next, we demonstrate that we obtain essentially the same results by randomly

assigning firm-years to sample partitions where anticipatory income smoothing

incentives are hypothesized to exist. We re-assess the statistical significance of theseresults using our approximate randomization tests. We then proceed in parallel

fashion to examine results using two alternative means of partitioning firms based on

current performance: first using analysts’ forecasts and then using operating cash

flows as proxies for un-managed earnings.

2.2.1. Replication of Defond and Park (1997)

Table 2, Panel A replicates DeFond and Park’s (1997, Table 2) main results. The

table presents a two-by-two classification of the sample firm-years based on current

un-managed earnings and expected next-year earnings above/below annual industry

medians. The columns partition the cases by current-year un-managed relativeperformance, and the rows partition the cases by expected next-year un-managed

performance. Current-year un-managed performance is measured by subtracting the


http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-



modified-Jones-model discretionary accrual estimate from reported earnings (income

before extraordinary items). Expected next-year un-managed performance is

measured as the March t þ 1 consensus forecast of year t þ 1 annual earnings.

Poor or good performance is defined relative to annual industry medians.

The anticipatory smoothing hypothesis developed in DP predicts earnings

management via discretionary accruals in the second and third quadrants of theclassification matrix in Table 2, panel A. In the second quadrant, discretionary

accruals are expected to be negative, as managers attempt to defer current year

Table 1

Descriptive statistics (n=8,535)

Variablesa Mean s Lower

quartile

Median Upper

quartile

Net earnings (NE ) 0.069 0.064 0.036 0.059 0.098

Operating cash flows (OCF ) 0.114 0.093 0.063 0.111 0.165

Assets ($millions) 3085 6294 246 865 3886

Total accruals (TA) À0.046 0.079 À0.088 À0.047 À0.006

Discretionary accruals (DA) 0.000 0.069 À0.032 0.001 0.033

Pre-managed earnings (NDE ) 0.069 0.091 0.019 0.064 0.116

Median net earnings, by industry and year (MNE ) 0.066 0.027 0.048 0.065 0.085

Current pre-managed earnings minus sample

median earnings, by industry and year (RP )

0.003 0.087 À0.043 0.001 0.047

Expected earnings (ENE ) 0.085 0.069 0.043 0.068 0.113

Median expected earnings, by industry and year

(MENE )

0.077 0.030 0.045 0.077 0.098

Earnings forecast errors (FE ) 0.001 0.026 À0.004 0.001 0.006

Expected earnings minus sample median expected

earnings, by industry and year (FD)

0.008 0.061 À0.020 0.000 0.025

Earnings forecast (AF ) 0.068 0.059 0.034 0.055 0.095

Median earnings forecast, by industry and year

(MAF )

0.063 0.025 0.043 0.061 0.079

Earnings forecast minus sample median expected

earnings, by industry and year (AFD)

0.005 0.053 À0.019 0.000 0.022

aVariable definitions: NE : Net income before extraordinary items scaled by lagged assets; OCF : Cash

flow from operations scaled by lagged assets; Assets ($millions): Total assets; TA: Net income beforeextraordinary items minus operating cash flows, scaled by lagged assets; DA: Discretionary accruals are

prediction errors from fitted values of the variant of the Jones (1991) model as was employed by DeFond

and Park (1997), where total accruals are defined as TAit=DCAit À DCLit À DCashit þ DSTDit À Depit

(DCAit=change in current assets, DCLit=change in current liabilities, DCashit=change in cash and cash

equivalents, DSTDit=change in debt included in current liabilities, and Depit=depreciation and

amortization expense); NDE : Non-discretionary earnings=NE ÀDA; MNE : Median net earnings in the

sample firm’s industry (2-digit SIC), measured in the year of interest using data from Compustat; RP :

Current pre-managed earnings minus median net earnings, by industry and year; RP =NDE – MNE ; ENE :

I/B/E/S consensus forecast of NE tþ1 as of March t þ 1; scaled by lagged assets; MENE : Median expected

future earnings (ENE ) in the sample firm’s industry (2-digit SIC); FE : NE 2AF ; FD: Expected future

earnings minus sample median earnings, by industry and year; FD=ENE ÀMENE ; AF : Analysts’ forecast

of E t as of December of year t; MAF : Expected future earnings (AF ) in the sample firm’s industry based on

the December forecasts; AFD: Expected future earnings based on the December forecast, less median net

earnings; AFD ¼ AF À MNE :


http://-/?-

http://-/?-



Table 2

Discretionary accruals partitioned by current relative performance and expected relative performance

Current relative performance

Poor Good All

Panel A: Modified-Jones-model discretionary accruals for the current period

Expected future relative performance

Poor (i) (ii)

Mean 0.026 À0.062 À0.002

Median 0.017 À0.044 0.001

% positive 72% 3% 49%

N 2726 1316 4042

Good (iii) (iv)

Mean 0.060 À0.025 0.003

Median 0.040 À0.017 0.002

% positive 96% 30% 52%

N 1473 3020 4493

All

Mean 0.038 À0.037 0.000

Median 0.027 À0.026 0.001

% positive 80% 22% 51%

N 4199 4336 8535

Panel B: Randomly assigned modified-Jones-model discretionary accruals for the current period


Poor (i) (ii)

Mean 0.028 À0.063 À0.001

Median 0.018 À0.045 0.000

% positive 72% 3% 50%

N 2756 1286 4042

Good (iii) (iv)

Mean 0.060 0.027 0.001Median 0.044 0.017 0.001

% positive 95% 30% 51%

N 1443 3050 4493

All

Mean 0.039 À0.037 0.000

Median 0.027 À0.026 0.001

% positive 80% 22% 51%

N 4199 4336 8535

Discretionary accruals for the current period are measured using the modified-Jones model; un-managed

earnings are measured as: reported earnings less estimated discretionary accruals (Panel A); reported

earnings less randomly assigned estimated discretionary accruals (Panel B); expected earnings are based onMarch consensus financial analysts’ earnings forecasts; ‘Good’ and ‘Poor’ partitions are determined

relative to annual industry medians.




income to offset expected poor performance in the subsequent year. In the third

quadrant, discretionary accruals are expected to be positive, as managers attempt to

shift subsequent year income to the current year to offset poor current performance.

The anticipatory smoothing hypothesis does not predict smoothing behavior for themain diagonal (quadrants one and four) observations because relative un-managed

performance is expected to persist over adjacent years.

The results of the replication correspond closely to those in DP. In cell ii (the

upper right cell of the table), mean discretionary accruals are À6.2% of total assets,

and 97% of the firm-years in that cell have negative discretionary accruals. Similarly,

in cell iii (lower left), mean discretionary accruals are 6.0% of total assets, and 96%

of the firm-years have positive discretionary accruals.3 We thus replicate the DeFond

and Park results (Table 2, panel A, p. 126) although they perform significance tests

based upon pooled rather than annual statistical tests. The comparative statistics in

their sample are À4.4%of assets (and 92% negative) for cell ii and 3.8% of assets

(and 87% positive) for cell iii.

2.2.2. Re-evaluation of Defond and Park (1997) results

Potential errors in measuring discretionary accruals and the ‘backing out’

approach to determining the classifications of current performance suggest a

cautious interpretation of the DeFond and Park results and the replicated results in

Table 2. Note that current relative performance, RP ; which is used to assign the

sample firm-years to the columns in panel A of Table 2, is

RP ¼ NE À MNE À D #A; where D #A ¼ DAtrue þ e; ð1Þ

where NE is the reported earnings, MNE the annual industry median earnings, D #A

the estimated discretionary accruals, DAtrue the amount of unobservable ‘true’

discretionary accruals, and e the error in measuring discretionary accruals.4 Because

the discretionary accrual proxy, D #A; is subtracted in measuring RP ; positive

(negative) measurement errors in D #A increase the likelihood that a given firm-year

will be classified in the ‘Poor’ (‘Good’) column of the matrix in panel A of Table 2.

Under the null hypothesis of no smoothing, the true discretionary accruals are

zero (that is, DAtrue 0). We assume that e is unbiased (that is, E½e ¼ 0), but that

measurement error in estimating discretionary accruals causes e for individual firm-

years to be non-zero. Stated equivalently, estimated discretionary accruals are by

definition random noise ðD #A ¼ eÞ: We also assume under the null that NEt À MNEt

is persistent.5

3The mean and median discretionary accruals differ significantly from zero in both cells based on t-tests

constructed using the means and standard errors of the 17 annual means and medians. The percentages of

positive errors differ significantly from 50% in both cells based on binomial tests of the annual percentages

relative to a null of 50%.4Note that in general, NE ÀMNE is a measure of the departure of actual, un-managed earnings from

target earnings. MNE can be thought of as an arbitrary target, and as such it has no direct role in our

analysis. Our focus is on the effects of subtracting DA from NE in the measurement of RP .5Under the null hypothesis of no smoothing, we expect reclassification across the industry median to

occur infrequently.


http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-



Table 3 illustrates expectations for D #A under these assumptions. As in DP, the

column classifications are based on year t relative performance above and below

annual industry medians, and the column classifications are based on relative

expected year t þ 1 performance based on analysts’ expectations above and belowindustry medians. Eq. (1), which defines relative performance, can be rewritten as

follows for firm-years whose current performance is below the industry median:

RP o03NE À MNE À D #Ao03D #A > NE À MNE : ð2Þ

Similarly, for firm-years whose current performance is above the industry median,

D #AoNE À MNE : The column headings reflect these expressions of Eq. (1). The row

headings are as defined in DP, and we include in the headings the implications of our

assumption regarding the persistence of relative performance.

Focusing for the moment on cell iii; Eq. (2) implies that E D #AtÂ Ã

> E½NE t À MNE t:

And using the properties of firm-years in the ‘Good’ row and the assump-

tion of persistence of NE t À MNE t; it follows that E NE tþ1 À MNE tþ1½ > 0

) E NE t À MNE t½ > 0: Combining these two inequalities yields E D #At

Â Ã>

E NE t À MNE t½ > 0 ) E D #At

Â Ãc0: A similar analysis for cell ii leads to the

prediction that for those firm-years, E D #At

Â Ã{0:

6 And thus, under the null

hypothesis of no smoothing behavior, we nonetheless obtain predicted results that

appear to support the alternative hypothesis of anticipatory smoothing. This is the

essence of the problem with the ‘backing out’ approach to measuring un-managed

earnings.

Table 3

Cell membership and expected discretionary accruals

Expected future relative performance (t þ 1) Current relative performance (time t)

Poor Good

D #A > NE À MNE D #AoNE À MNE

Poor

E NE tþ1 À MNE tþ1½ o0 E D #At

Â ÃoE NE t À MNE t½

) E NE t À MNE t½ o0 E NE t À MNE t½ o0

) E D #At

Â Ã{0

Good

E½NE tþ1 À MNE tþ140 E D #At

Â Ã> E NE t À MNE t½

) E½NE t À MNE t40 E½NE t À MNE t40

) E D#

AtÂ Ã

c0

Note: Column classifications are determined (as in DP) by comparing ‘un-managed earnings’ ðNE t À D #AtÞ

to industry median net earnings (MNE t). Row classifications are determined by comparing expected

earnings (ENE tþ1) to expected median net earnings (MENE tþ1). Current relative performance (RP t) is

defined as NE t À D #At À MNE t:

6We do not make predictions for cells i and iv, because the row and column inequalities are opposite,

and thus the sign of the expected discretionary accruals is indeterminate.




The foregoing discussion and analysis assume that discretionary accruals are

entirely error. Note, however, if we instead assume that firms do manage earnings as

hypothesized, and that discretionary accruals are measured perfectly, the expecta-

tions for cells ii and iii are identical to those in Table 3. Thus, estimating un-managedearnings by subtracting estimated discretionary accruals when such discretionary

accruals are measured with error produces discretionary accrual distributions in the

cells of interest (ii and iii) that are observationally equivalent to those if the

discretionary accruals were entirely random errors. Therefore, valid inferences

cannot be drawn from sample statistics in the 2 Â 2 classifications. The ‘backing out’

of error-prone discretionary accrual measures biases the statistics toward rejection of

the null of no smoothing. And it is important to note that this will be true regardless

of the discretionary accrual proxy chosen.

We next evaluate whether the results of tests such as in DP can be ascribed entirely

to the problems caused by the ‘backing out’ approach. We employ a randomization

procedure whereby we generate an empirical distribution of the means, medians, and

percentage of positive cases in Table 2, panel A. We compare the characteristics of

the discretionary accruals from each candidate discretionary accruals proxy to the

empirical distribution. We test whether the number of cases from our randomized

iterations of test statistics exceeds the corresponding statistics from the actual

analysis.

2.2.3. Results based on random assignment of discretionary accruals

To gauge the effects of errors in measuring discretionary accruals on the results in

Table 2, Panel A, we induce the no smoothing condition suggested in the foregoing

discussion. We randomly assign the discretionary accrual estimates generated by the

modified-Jones model to the sample firm-years (Noreen (1989)). This re-assignment,

in effect, creates a random measurement error variable that is distributed identically

to the discretionary accrual measure.7

This random assignment is repeated 1,000 times, each time capturing the mean

and median discretionary accrual and the percentage of cases in each cell where

measured discretionary accruals are positive. We interpret the resulting statistics

from the 1,000 iterations as representative of the distribution that would exist under

the null of no income smoothing. We then repeat the tests of the anticipatory

income-smoothing hypothesis by comparing the characteristics of the distribution of

measured discretionary accruals with the empirical distribution. p-values are then

computed as the proportion of the 1000 iterations in which the mean and median

exceed the actual statistics.

Table 2, panel B presents a representative outcome from the 1,000 iterations, and

the first three columns of Table 6 summarize and test the differences between panels

A and B of Table 2. The results show that there is no reliable, statistically discernable

7As an alternative approach, we also replaced the discretionary accrual estimates with random numbers

drawn from a normal distribution with the same mean and standard deviation as the vector of Jones-

model-based estimates. The results from this procedure were virtually identical to those reported in

Table 2, Panel A.


http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-



difference between the actual discretionary accrual measures and the random error

variable. While the median in cell ii and the percentage positive in cell iii are

statistically different, there is no general pattern of reliable differences, either

statistically or economically. These results indicate that the observed pattern of discretionary accruals does not provide a test of management’s earnings smoothing

behavior because the result is guaranteed by errors in the column classifications that

are related mechanically to the magnitudes of the discretionary accrual estimates.

2.2.4. Results using analysts’ forecast errors as discretionary accrual proxies

Abarbanell and Lehavy (2000) argue that analysts’ forecast errors might proxy for

the managed portion of annual earnings. They assume that late in the earnings year,

for reasons of motivation or ability, analysts may not forecast the managed

component of earnings. If analysts’ forecasts are to be a useful proxy for un-

managed earnings, we must also assume that forecast errors caused by reasons other

than management’s late in the year earnings management are reasonably well

diversified across the sample. To the extent that these assumptions are descriptive,

analysts’ forecasts have several characteristics that are desirable relative to modified-

Jones-model estimates, including lower variance, the ability to capture earnings

management of all types (not only accruals management), and no need to require

that parameters of a model be constant over extended periods.

To implement this approach, we repeat the analysis in Table 2, dividing firms into

‘Poor’ and ‘Good’ current relative performance based on December t forecasts of

year t earnings relative to annual industry median earnings, rather than usingreported income less discretionary accruals as in DP. Analysts’ forecast errors then

become proxies for the managed component of earnings. We present the results of

this test in Table 4, panel A. As indicated therein, there is apparent support for the

income-smoothing hypothesis; cells ii and iii have mean (median, % positive)

forecast errors of À0.009 (À0.004, 25%) and 0.015 (0.006, 82%), respectively,

suggesting downward earnings management in cell ii and upward earnings

management in cell iii:

However, this too is a ‘backing out’ approach. The column classifications in Table

4, while nominally based on analysts’ forecasts, can equivalently be described as

earnings less forecast error since forecast error is defined as actual earnings lessanalysts’ forecasts. Thus large forecast errors will likely lead to classification in the

‘Poor’ column and vice-versa, and hence the sample statistics for forecast errors may

be mechanically induced as in Table 2.

To investigate whether this potential problem renders the approach useless in

testing the smoothing hypothesis, we again employ a randomization procedure. We

randomly re-assign all forecast errors across the sample firm-years, re-compute the

implied consensus analyst forecast as reported earnings less the re-assigned forecast

error, and re-classify the firm-years relative to current-year annual industry median

earnings. Panel B of Table 4 reports a representative outcome from the 1000

iterations of this procedure. The second three columns of Table 6 show thecomparisons of mean and median forecast errors from the actual and random

versions of the test, together with a p-value that indicates the number of the 1000


http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-



Table 4

Analysts’ forecast errors partitioned by current relative performance and expected relative performance


Poor Good All

Panel A: Actual forecast errors

Expected future Relative Performance

Poor (i) (ii)

Mean 0.000 À0.009 À0.001

Median 0.001 À0.004 0.001

% positive 58% 25% 55%

N 3694 348 4042

Good (iii) (iv)Mean 0.015 À0.001 0.003

Median 0.006 0.000 0.001

% positive 82% 50% 57%

N 964 3529 4493

All

Mean 0.003 À0.002 0.001

Median 0.002 À0.000 0.001

% positive 63% 48% 56%

N 4658 3877 8535

Panel B: Randomly Assigned analysts’ forecast errors


Poor (i) (ii)

Mean 0.005 À0.021 0.000

Median 0.002 À0.010 0.001

% positive 64% 17% 55%

N 3285 757 4042

Good (iii) (iv)

Mean 0.018 À0.003 0.002

Median 0.007 À0.000 0.001

% positive 85% 48% 57%

N 1053 3440 4493

All

Mean 0.008 À0.006 0.000

Median 0.003 À0.001 0.001

% positive 69% 43% 56%

N 4338 4197 8535

Discretionary accruals for the current period are represented by December consensus analysts’ forecast

errors; un-managed earnings are measured as: reported earnings less analysts forecast errors (Panel A);

reported earnings less randomly assigned forecast errors (Panel B) ‘expected earnings’ are based on March

consensus financial analysts’ earnings forecasts; ‘Good’ and ‘Poor’ partitions are determined relative toannual industry medians.




cases where the sample statistics from the random iterations exceeded the actual

statistics.

The results show that none of the actual statistics can be distinguished from the

results of the randomized iterations, and thus we infer that the ‘backing out’procedure inherent in this approach renders the evidence useless as a test of the

smoothing hypothesis. Note that the very large p-values have (at least) one likely

explanation. In contrast to the modified-Jones-model discretionary accrual

estimates, December forecast errors are very tightly distributed around actual

earnings. As a result, randomly re-assigning a forecast error to another firm-year is

more likely than not to result in an implied forecast that lies further from actual

earnings than the actual forecast. Under the mechanics underlying the column

partitions, this means that a large number, or perhaps all, of the random iterations

will result in forecast errors that exceed the forecast errors using the actual data.

2.2.5. Results based on partitions of current performance using operating cash flows

Recognizing the problems inherent in the ‘backing out’ approach, both DP and

ALZ investigate the use of operating cash flows as a proxy for un-managed earnings.

Note that using cash flows as a proxy for un-managed earnings introduces its own

challenges. For example, assuming that cash flows equal un-managed earnings is

equivalent to assuming that all accruals represent earnings management, which is not

likely to be descriptive. And for cash flows to make sense as a proxy for un-managed

earnings, we would have to assume that it is relative levels of cash flows that cause

managers to manage earnings, even though the theory clearly refers to managers’sensitivity to industry earnings, not cash flows. An additional challenge is that there

is a strong negative correlation between operating cash flows and accruals (Dechow

et al., 1998). The negative correlation raises the concern that partitions based on cash

flow are highly overlapping with partitions based on accruals. Nevertheless, in this

section, we investigate whether the use of cash flows mitigates the inference problems

caused by the ‘backing out’ approach.

One important distinction between the use of cash flows in each of the two studies

in question is that DP define cash flows as earnings less measured total accruals,

while ALZ use firms’ disclosed operating cash flow from the Statement of Cash

Flows. Collins and Hribar (1999) point out that accruals computed using the‘balance sheet approach,’ whereby changes in selected current asset and liability

accounts are used to measure accruals, are subject to potentially important

measurement error, especially as the result of business combinations, but more

generally when balance sheet changes reflect phenomena other than accounting

accruals. Together with the well-documented negative correlation between accruals

and cash flows and the low explained variance of the modified Jones model, there is

reason to be concerned that results using cash flows obtained in this manner are

subject to a somewhat different, but equally important ‘backing out’ problem.

Specifically, if total accruals are measured with error, and given that the R2 from

estimating the discretionary accrual model is typically quite low, discretionaryaccruals (the residual portion of total accruals) will reflect the error in total accruals.

And given that accruals and cash flows are strongly negatively correlated, errors in




Table 5

Discretionary accruals partitioned by current relative performance (based on cash flows) and expected

relative performance


Poor Good All

Panel A: Modified-Jones-model discretionary accruals


Poor (i) (ii)

Mean 0.043 À0.027 À0.002

Median 0.037 À0.016 0.001

% positive 79% 33% 49%

N 1433 2609 4042

Good (iii) (iv)

Mean 0.081 À0.013 0.003

Median 0.071 À0.006 0.002

% positive 92% 44% 52%

N 724 3769 4493

All

Mean 0.056 À0.019 0.000

Median 0.046 À0.010 0.001

% positive 83% 40% 51%

N 2157 6388 8535

Panel B: Randomly assigned modified-Jones-model discretionary accruals


Poor (i) (ii)

Mean 0.043 À0.026 0.002

Median 0.033 À0.016 0.001

% positive 77% 33% 51%

N 1631 2411 4042

Good (iii) (iv)

Mean 0.076 À0.015 À0.001Median 0.067 À0.007 0.001

% positive 91% 43% 50%

N 691 3802 4493

All

Mean 0.053 À0.019 0.000

Median 0.043 À0.011 0.001

% positive 81% 39% 51%

N 2322 6213 8535

Discretionary accruals for the current period are measured using the modified-Jones model; un-managed

earnings are measured as operating cash flows: reported earnings less measured total accruals (Panel A);

reported earnings less randomly assigned total accruals (Panel B) expected earnings are based on Marchconsensus financial analysts’ earnings forecasts; ‘Good’ and ‘Poor’ partitions are determined relative to

annual industry medians.




Table 6

Statistical comparisons of discretionary accrual magnitudes with empirical (randomized) distributions

Table 2: Partitions based on ‘backing out’

modified-Jones-model discretionary

accruals

Table 4: Partitions based on analysts’

forecasts; forecast errors as discretionary

accrual proxies

Table

modifi

accrua

Actual

(Panel A)

Random

(Panel B)

p-value Actual

(Panel A)

Random

(Panel B)

p-value Actual

(Panel

Cell ii

Mean À0.062 À0.063 0.341 À0.009 À0.021 1.000 À0.02

Median À0.044 À0.045 0.031 À0.004 À0.010 1.000 À0.01

Cell iii

Mean 0.060 0.060 0.157 0.015 0.018 1.000 0.08

Median 0.040 0.044 0.261 0.006 0.007 0.971 0.07

‘Actual’ represents the statistics from Tables 2, 4, and 5 from employing each of the alternative specifications of

description in the second row of this table. ‘Random’ indicates the statistics obtained after randomly re-assigning th

analysts’ forecast error (Table 4), or both the discretionary and non-discretionary accruals (Table 5) across firm-years a

indicate the number of cases (out of 1000 iterations) where the indicated statistic was more negative (positive) i

corresponding actual condition for cell ii (iii).



accruals are likely to coincide with errors in cash flow in the opposite direction. As a

result, firm-years classified as having had ‘Good’ current performance based on large

operating cash flow will be on average those firms who have had relatively small

accruals. If the given firm-year had small accruals because of a large negative error inmeasuring accruals, and the negative error flows through the discretionary accrual

model to be included in discretionary accruals, once again, there will be a mechanical

explanation for observed signs and magnitudes of discretionary accruals that are

obtained using the type of analysis under scrutiny in this paper.

Table 5 presents the results of our investigation of the use of cash flows obtained

from the ‘balance sheet approach’ as a proxy for current un-managed earnings

performance. We partition firm-years into ‘Poor’ and ‘Good’ columns based on

operating cash flows, where operating cash flows are defined as reported earnings less

balance-sheet-derived total accruals, as in DP. The rest of the methodology in Table

5 is similar to that in the preceding tables.

Panel A shows that, as DP find in un-tabulated results, this approach yields

patterns in discretionary accruals that are supportive of the income smoothing

hypothesis. Panel B shows the results obtained after randomly re-assigning

discretionary accrual estimates across firm-years, computing an implied cash flow

from operations based on the sum of income, discretionary accruals, and implied

non-discretionary accruals, and using the resulting cash flow from operations to

partition firms based on current performance. As before, this procedure essentially

enforces the null of no income smoothing by removing the economic correspondence

between cash flow (the partitioning variable) and discretionary accruals (the variableof interest).

The results in Panel B are strikingly similar to those in Panel A. The last three

columns of Table 6 summarize the comparisons and provide statistical tests of

differences based on 1000 iterations of the randomization. As in Tables 2 and 4,

there is no reliable statistical difference between the results in Panel A and the results

obtained from a random classification of firms. Once again, this evidence leads to the

conclusion that using cash flows measured in this manner lead to empirical tests that

are incapable of providing evidence about the income smoothing hypothesis.

3. Conclusion

This paper re-examines evidence on the multi-year smoothing hypothesis proposed

by Fudenberg and Tirole (1995) and tested by DP. The hypothesis predicts that

managers consider both relative current-year and expected year-ahead performance

when contemplating their current-year discretionary accrual decisions. They do so to

reduce the threat of dismissal caused by under-performance in either year.

We first demonstrate analytically that the mechanical relationship between

measured discretionary accruals and un-managed earnings that is used to identify

firms’ earnings management incentives guarantees the results in DP. We thenintroduce an alternative approach to testing for earnings management in the

presence of errors in the measurement of discretionary accruals. We use a


http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-

http://-/?-



randomization approach to construct an empirical distribution of discretionary

accruals under the null of no earnings smoothing. Summary statistics from the

original analysis are then compared to the empirical distribution as a means of

assessing their statistical significance. The results of this approach indicate thatevidence consistent with the theory of earnings smoothing is not distinguishable

from that obtained when discretionary accrual estimates are purely random

numbers.

We next investigate analysts’ December forecast errors as an alternative managed

earnings proxy, but demonstrate that despite the possibility that analysts’ forecast

errors contain less measurement error than modified-Jones-model estimates, the

relationship between forecasts as the measure of un-managed earnings and forecast

errors as the measure of managed earnings causes this to also be a ‘backing out’

approach, and as such it is not possible to draw conclusions from this alternative

test.

We also investigate the use of cash flows as a proxy for un-managed earnings.

While on the surface, it might appear that cash flows are not subject to the ‘backing

out’ problem because they are not directly related to the magnitudes of discretionary

accruals, we show that when operating cash flows are measured using as the residual

of income and total accruals computed using a balance sheet approach, a different

sort of ‘backing out’ problem arises that renders the cash-flow-based analysis useless

in assessing the smoothing hypothesis.

Overall, our results do not provide evidence supportive of anticipatory income

smoothing that is incremental to the results obtained by a simple random assignmentof discretionary accrual estimates to the analysis firms. Our results do not reject the

underlying theory proposed in Fudenberg and Tirole (1995). Rather, we show only

that tests using the framework applied in DP are not capable of testing the theory.

More generally, the randomization approach illustrated in this paper provides a

threshold that is appropriate in assessing the results of studies that use error-prone

measures of managed earnings in order to partition reported earnings into its

managed and un-managed components. Given the prevalence of the ‘backing out’

approach in both classification tests as used in this context and in regression-based

tests, this paper highlights the difficulties inherent in measuring un-managed

earnings.

References

Abarbanell, J., Lehavy, R., 2000. Biased forecasts or biased earnings? The role of earnings management in

explaining apparent optimism and inefficiency in analysts’ earnings forecasts. Working Paper,

University of North Carolina.

Ahmed, A., Lobo, G., Zhou, J., 2001. Job security and income smoothing: an empirical test of the

Fudenberg and Tirole (1995) model. Working paper, Syracuse University.

Collins, D., Hribar, P., 1999. Errors in estimating accruals: Implications for empirical research. Working

paper, University of Iowa.

Dechow, P., Sloan, R., Sweeney, A., 1995. Detecting earnings management. The Accounting Review 70,

193–225.




Dechow, P., Kothari, S.P., Watts, R., 1998. The relation between earnings and cash flows. Journal of

Accounting and Economics 25 (2), 133–168.

DeFond, M., Park, C., 1997. Smoothing income in anticipation of future earnings. Journal of Accounting

and Economics 23, 115–139.Fudenberg, K., Tirole, J., 1995. A theory of income and dividend smoothing based on incumbency rents.

Journal of Political Economy 103, 75–93.

Guay, W., Kothari, S., Watts, R., 1996. A market-based evaluation of discretionary-accrual models.

Journal of Accounting Research 34, (Suppl. 1996), pp. 83–105.

Jones, J., 1991. Earnings management during import relief investigations. Journal of Accounting Research

29, 193–228.

Lim, S., Lustgarten, S., 2000. Testing for income smoothing using the backing out method. Texas

Christian University Working Paper.

Noreen, E., 1989. Computer-Intensive Methods for Testing Hypotheses: an Introduction. Wiley, New

York, NY.


5- Anticipatory income smoothing a re-examination.pdf

Documents