Professional Documents
Culture Documents
Accounting Choice and Earnings Quality The Case of Software Development
Accounting Choice and Earnings Quality The Case of Software Development
Mustafa Ciftci
To cite this article: Mustafa Ciftci (2010) Accounting Choice and Earnings Quality: The
Case of Software Development, European Accounting Review, 19:3, 429-459, DOI:
10.1080/09638180.2010.496551
MUSTAFA CIFTCI
SUNY at Binghamton, School of Management, Binghamton, New York, USA
ABSTRACT In this study I explore how accounting choice affects earnings quality in the
software development industry. SFAS No. 86, which requires capitalization of software
development costs (SDC), is the only exception in the US to SFAS No. 2, which
requires immediate expensing of all research and development (R&D) expenditures.
Aboody and Lev (1998) suggest that capitalized SDC are value-relevant. Thus,
expensing of these costs might introduce noise into earnings. However, it has been
suggested that future benefits associated with SDC are highly uncertain (Software
Publishers Association). Consequently, capitalization might introduce noise into
earnings by capitalizing unproductive expenditures. Hence, it is not clear how managers’
choice between capitalization and expensing will affect earnings quality. I first find that
there is a decline in the quality of earnings in the software industry after the adoption of
SFAS No. 86, whereas no such decline is observed in other high-tech industries. Second,
I find that, within the software industry, the quality of earnings for expensers is greater
than for capitalizers. Finally, I find that, among the capitalizers, those with a large
increase in software capital have lower earnings quality than others. Overall, the results
suggest that capitalization of software costs does not improve earnings quality.
1. Introduction
In this study I investigate how accounting choice affects the quality of earnings.
SFAS No. 2, which requires immediate expensing of all R&D expenditures,
does not allow any choice in reporting. SFAS No. 86 is the only exception to
the universal expensing of R&D in the US. Under this regime, SDC must be
capitalized after the attainment of technical feasibility if the future benefits are
probable. While capitalization is required under SFAS No. 86, some companies
never capitalize, due to the discretion allowed in application of the standard.
Hence, in practice, managers have a choice between capitalization and expensing.
Consequently, SFAS No. 86 provides us with an opportunity to see the impact of
managers’ reporting choice between expensing and capitalization on the quality of
earnings in the US setting.
Aboody and Lev (1998) find that capitalized SDC are value-relevant. If these
costs are value-relevant, expensing of them might introduce noise into earnings,
thus reducing their quality. On the other hand, the Software Publishers Associ-
ation (SPA) states that the future benefits associated with soft assets are so
uncertain that capitalization of these assets may not be reliable (Aboody and
Lev, 1998). A high level of uncertainty associated with future benefits might
lead to managerial errors in the amount of capitalized SDC. In addition, investors
cannot verify whether capitalized SDC correctly reflects productive expenditures.
Consistent with this view, the Association for Investment Management Research
(AIMR) states that ‘It is usually next to impossible to determine in any sensible or
codifiable manner exactly which costs provide future benefits and which not’
(AIMR, 1993). Thus, because of the high uncertainty associated with future
benefits, and the lack of verifiability associated with SDC, capitalization of
these costs might reduce earnings quality.
Hence, the impact of capitalization of soft assets on the quality of earnings is
unclear. I use the earnings response coefficient (ERC) – the association between
unexpected earnings and cumulative abnormal returns around earnings announce-
ments – to measure the quality of reported earnings. The ERC shows investors’
perceptions of earnings quality. Following Mohd (2005), I first compare the
quality of reported earnings in the software development industry, where SDC
are capitalized after SFAS No. 86, with that in other high-tech industries, where
all R&D expenditures are expensed. The results indicate that there is a decline in
the quality of reported earnings for the software industry after the adoption of
SFAS No. 86, but no such decline for other high-tech industries in the same period.
Second, I compare the quality of earnings among software firms based on their
reporting choice. Specifically, I compare the quality of earnings for firms that capi-
talize some or all of their SDC with that for firms that expense all of these costs
after SFAS No. 86. I find that the earnings quality of the firms that expense all
R&D is greater than that of those that capitalize. Moreover, I find that 10.72%
(10.30%) of capitalizers avoid reporting loss (decline in earnings) by capitalizing
SDC, raising the possibility that capitalization might be used for earnings manage-
ment. Third, I compare the quality of earnings among the capitalizers based on the
magnitude of change in software capital. More specifically, I compare the quality
of earnings for firms that have a large increase in software capital with that for
other capitalizers. I find that the quality of earnings for firms with a large increase
in SDC is significantly lower than for other capitalizers. Moreover, I find that
22.14% (19.15%) of large-increase firms avoid reporting loss (decline in earnings)
by capitalizing SDC, compared with 2.48% (5.88%) for other capitalizers.
Accounting Choice and Earnings Quality 431
reduce the quality of reported earnings. Accordingly, financial analysts argue that
‘software capitalization would reduce the quality of earnings. It would make it
more difficult to assess which companies are doing well or not.’9
Consistent with the greater information asymmetry associated with capitaliza-
tion, Aboody and Lev (1998) document that capitalization leads to greater absolute
analyst forecast errors and dispersion in analyst forecasts.10 However, Mohd
(2005) finds that the mean spread and volume over the whole year are lower for
capitalizers than for expensers, and interprets these results as a decline in infor-
mation asymmetry due to capitalization of SDC. Hence, there is an inconsistency
between the evidence presented by Mohd (2005) and that presented by Aboody and
Lev (1998). There are two possible explanations for this inconsistency. The first is
that the sample used by Mohd does not seem to properly represent the population of
capitalizers and expensers (i.e. the market value of equity for capitalizers is six
times larger than for expensers in their sample: $150 million for expensers
versus $913 million for capitalizers). His sample is not consistent with the
sample composition of most of the studies in this topic.11 Furthermore, Aboody
and Lev (1998) suggest that capitalization is negatively related to firm size.
Given that firm size is an important determinant of the information environment
(Atiase, 1985), Mohd’s results might be driven by his sample characteristics.12
Second, Mohd explores information asymmetry over a whole year, and attri-
butes the difference between capitalizers and expensers to SFAS No. 86.
However, it is not clear whether the decline in information asymmetry is
driven by capitalization per se, because he does not document whether this
decline is due to the disclosure of specific accounting information regarding capi-
talization. There might be many confounding events over an entire year that
might lead to a decline in information asymmetry, rather than capitalization
itself. In fact, Mohd claims that investors are pessimistic about capitalizers,
and narrates anecdotal evidence suggesting that capitalizers show every effort
to demonstrate that their judgments are well justified within the boundaries of
generally accepted accounting principles (GAAP).13 If investors are pessimistic
about capitalization, it is not clear how information asymmetry declines as a
result of capitalization.14
3. Research Design
I use ERC to measure the quality of financial information.18 The research design
in this study is similar to Mohd (2005) in several aspects. I first compare the
quality of reported earnings in the software development industry, where SDC
are required to be capitalized after SFAS No. 86, with that in other high-tech
industries, where all R&D expenditures are expensed. Mohd (2005) suggests
that the software development industry and other high-tech industries are both
R&D-intensive, but that they differ in their accounting treatment of R&D expen-
ditures. Second, I compare software firms after the adoption of SFAS No. 86,
based on their decision to capitalize. Third, I compare the capitalizers based on
the magnitude of change in capitalized SDC. The following three subsections
describe the research method for each research question.
The definition of variables is in the Appendix. Unexpected earnings are the change
in earnings in year t relative to year t 2 1, scaled by the market value of equity. The
436 M. Ciftci
Software firms might have low earnings quality due not to differential account-
ing treatment per se but earnings management. Hence, I need to disentangle the
impact of earnings management on earnings quality from that of accounting treat-
ment. Managers might use capitalization to avoid reporting loss or a decline in
earnings, or to smooth earnings. To control for earnings management, I include
three variables in the regression: AVLOSSit, AVDECit, and SMOOTHit. The first
variable, AVLOSSit, controls for managers’ use of capitalization to avoid reporting
loss. The second variable, AVDECit, is included to control for managers’ use of
capitalization to avoid reporting a decline in earnings. To the extent that these
variables capture earnings management, there should be a decline in the quality
of earnings, and hence I expect a negative sign on both b12 and b13. Consistent
with Francis et al. (2004), I also add SMOOTHit to control for earnings smoothing
in equation (1A). Francis et al. (2004) define smoothness in earnings as the ratio of
the standard deviation of earnings to that of cash flows. If managers are using
accruals to reduce the standard deviation in cash flows, the standard deviation
of earnings should be smaller than that of cash flows. In the same spirit, if man-
agers are smoothing the volatility of earnings by using capitalization, then the
standard deviation of the reported earnings under capitalization will be smaller
than that of earnings under expensing (from hypothetical expensing of SDC).
Given that earnings smoothing for earnings management would reduce earnings
quality, I expect a negative sign on b15 in equation (1A). To control for disclosures
in the pre-announcement period, consistent with Christensen et al. (2004), I add
PREDISCit in equation (1). Christensen et al. (2004) use cumulative returns as
a proxy for predisclosure information environment quality. I use average abnor-
mal returns over the pre-announcement period, scaled by the standard deviation
of returns in that period. To the extent that information about a firm’s earnings
is disclosed prior to earnings announcements, the reaction around earnings
announcement will be smaller. Hence, I expect a negative sign on b14.
However, as suggested by Mohd (2005) and Aboody and Lev (1998), the
decision to capitalize SDC might be endogenous, and hence the coefficient esti-
mate of CAPITALit in equation (2) might be biased. To eliminate the impact of
endogeneity, consistent with Mohd (2005) and Aboody and Lev (1998), I esti-
mate a two-stage model (2SLS), instead of OLS. The variables that are related
to the capitalization decision are firm size, leverage, firm beta, development
intensity, and income. Consistent with Mohd (2005) and Aboody and Lev
(1998), I estimate the following logit model and generate the predicted value,
PCAPITALit, from the equation below:
After generating the predicted values of CAPITALit from equation (3), I rank
these estimates into two groups, and estimate equation (2) using the ranked
PCAPITALit. It is possible that there might be differences in earnings quality
between capitalizers and expensers, even in the pre-SFAS No. 86 period:
hence, it would be more appropriate to estimate ERC for expensers and capitali-
zers in the pre-SFAS No. 86 period. However, I cannot extend this analysis to the
pre-SFAS No. 86 period, because I cannot estimate the corresponding
PCAPITALit in that period.
3.3. Earnings Quality for Firms with Large Increase in Capitalized SDC
To investigate the third research question, I explore earnings quality among capi-
talizers only. The test involves a comparison of ERC between capitalizing firms
that have a large increase in software capital and other capitalizers. I estimate the
following model.
The coefficient estimate of b2 in equation (4) shows the difference in ERC for
firms that have a large increase in software capital relative to other capitalizers. If
there is a significant decline in the quality of earnings for firms with a large
increase in software capital, the coefficient estimate of b2 should be negative.
Accounting Choice and Earnings Quality 439
4. Empirical Results
4.1. The Results for Software versus Other High-tech Firms
4.1.1. Sample selection and descriptive statistics
The capitalized SDC are not available in machine-readable form in Compustat.
Therefore, following a procedure similar to Mohd (2005), they are hand-collected
from financial statements using LexisNexis. Similar to Mohd (2005), the software
firms in my sample are those in four-digit SICs 7370 – 7374. Consistent with
Francis and Schipper (1999), other high-tech firms are 283 (drugs), 357 (compu-
ter and office equipment), 360 (electrical machinery and equipment), 361 (elec-
trical transmissions and distribution equipment), 362 (electrical industrial
apparatus), 363 (household appliances), 364 (electrical lighting and wiring equip-
ment), 365 (household audio, video equipment, audio receiving), 366 (communi-
cations equipment), 367 (electronic components, semiconductors), 368
(computer hardware), 481 (telephone communications), 873 (research develop-
ment, testing services), and firms other than the software industry in three-digit
SIC code 737.
The sample period is 1981 – 1990. The pre-SFAS No. 86 period is 1981– 1985,
and the post-SFAS No. 86 period is 1986– 1990. To be included in the sample
each firm should have at least one observation in both the pre-SFAS No. 86
and post-SFAS No. 86 periods. The sample period ends at 1990. If I extend
the analysis into future years such as 1995, I can have more observations,
which might increase the power of the tests. However, over a 15-year time
period there might be structural changes in the economic characteristics of the
industries being compared, which might raise questions about whether the
results are due to a structural change or to SFAS No. 86. I delete extreme obser-
vations in the top and bottom 1% of UNEXit. The breakpoints for extreme obser-
vations are determined using all R&D firms. In addition, following Collins et al.
(1997), I check for the extreme observations with studentized residuals greater
than 4 in the regression of equation (1). No observation is deleted by this cri-
terion. There are 429 firms and 2694 firm – year observations in the other high-
tech group, and 80 firms and 451 firm– year observations in the software industry.
Panel A of Table 1 presents the descriptive statistics for software and other
high-tech firms before the adoption of SFAS No. 86. There are significant differ-
ences in EPit (earnings-to-price ratio), BMit (book-to-market ratio), BETAit (beta),
MVit (market value of equity), and PERSISTit (persistence of earnings) between
other high-tech industries and the software industry. Overall, the descriptive stat-
istics show that there are systematic differences in firm characteristics between
the software industry and other high-tech industries, indicating the importance
of controlling for these characteristics.
Panel B reports the descriptive statistics after the adoption of SFAS No. 86. The
differences in EPit between software and other high-tech firms are reversed after
SFAS No. 86: the mean EPit for software firms is significantly greater than that
for other high-tech firms. This suggests that, after SFAS No. 86, software firms
440 M. Ciftci
become more profitable than other high-tech firms. Additionally, even though there
was not a significant difference between the two groups in the frequency of report-
ing loss prior to SFAS No. 86, the mean loss frequency for other high-tech firms is
significantly greater than that for software firms after SFAS No. 86. More interest-
ingly, the mean value of AVLOSSit is 0.0696, suggesting that 6.96% of software
Accounting Choice and Earnings Quality 441
firms avoid reporting loss by capitalizing SDC. Moreover, the mean value of
AVDECit is 0.0633, suggesting that 6.33% of software firms avoid reporting a
decline in earnings by capitalizing SDC. Finally, the mean value of SMOOTHit
is 0.8739 for the software industry, whereas by definition it is 1.00 for other
high-tech industries, suggesting that software firms smooth earnings by capitaliz-
ing SDC. Overall, the results in Panel B raise the possibility that software firms use
capitalization of SDC to achieve performance targets or to smooth earnings.
4.1.3. Robustness
In this section I perform several robustness checks. I first estimate equation (1)
using only the firms that (a) capitalize SDC, (b) expense all SDC, rather than
442 M. Ciftci
Table 2. Comparison of earnings quality between software and other high-tech industriesa
Equation (1) without
Variable controlb Equation (1)b Equation (1A)b
UNEXit 0.1370∗∗∗ 0.3761∗∗∗ 0.3975∗∗∗
(3.96) (3.79) (3.06)
UNEXit∗ PERIODit 20.0285 20.0332 20.0356
(20.69) (20.60) (20.55)
UNEXit∗ SOFTit 0.0321 0.0423 0.0521
(0.47) (0.51) (0.62)
UNEXit∗ PERIODit∗ SOFTit 20.1432∗∗∗ 20.1695∗∗∗ 20.1521∗∗
(22.73) (23.21) (22.21)
UNEXit∗ SIZEit 20.0278∗∗ 20.0232∗
(21.95) (21.81)
UNEXit∗ BMit 20.1424∗∗∗ 20.1152∗∗
(22.95) (22.35)
UNEXit∗ BETAit 20.0269 20.0141
(20.75) (20.36)
UNEXit∗ LOSSit 20.1345∗∗∗ 20.1643∗∗∗
(22.68) (22.83)
UNEXit∗ LEVit 20.0290 0.0221
(21.43) (0.86)
UNEXit∗ PERSISTit 0.0671∗ 0.0772∗
(1.72) (1.68)
UNEXit∗ FOLLOWit 20.0312∗ 20.0225
(21.65) (21.17)
UNEXit∗ AVLOSSit 20.1311 20.0361
(20.52) (20.14)
UNEXit∗ AVDECit 20.2541 20.2241
(20.69) (20.55)
UNEXit∗ PREDISCit 20.0402 20.0287
(20.95) (20.49)
UNEXit∗ SMOOTHit 20.3482
(21.51)
Adjusted-R2 0.0105 0.0172 0.0174
N 3,145 3,145 2,435
a ∗ ∗∗
, , and ∗∗∗ denote significance at 0.10, 0.05 and 0.01 respectively. The t-values are in parenthesis.
Before (after) SFAS No. 86 is fiscal years 1981–1985 (1986– 1990). Industry definitions are in Table 1.
Variable definitions are in the Appendix.
b
The model used in estimation:
CARit =b0 + b1 UNEXit + b2 (UNEXit∗ SOFTit ) + b3 (UNEXit∗ PERIODit )+
b4 (UNEXit∗ PERIOD∗it SOFTit ) + b5 (UNEXit∗ SIZEit ) + b6 (UNEXit∗ BMit )+
b7 (UNEXit∗ LOSSit ) + b8 (UNEXit∗ BETAit ) + b9 (UNEXit∗ LEVit )+
b10 (UNEXit∗ PERSISTit ) + b11 (UNEXit∗ FOLLOWit ) + b12 (UNEXit∗ AVLOSSit )+
b13 (UNEXit∗ AVDECit ) + b14 (UNEXit∗ PREDISCit ) + 1it (1)
CARit = EQUATION(1) + b15 (UNEXit∗ SMOOTHit ) + 1it (1A)
Accounting Choice and Earnings Quality 443
including all software firms.20 Panel A of Table 3 reports the estimation result
when I include only capitalizers in the software industry. The first column
reports the estimation results without control variables. The coefficient estimate
for (UNEXit∗ PERIODit) is 20.0316 and for (UNEXit∗ PERIODit∗ SOFTit) is
20.1499 (p-value , 0.01), similar to the results in Table 4. The next column
reports the estimation results with control variables. The coefficient estimate
for (UNEXit∗ PERIODit∗ SOFTit) is 20.1975 (p-value , 0.01), which is even
more negative than the results with all software firms. The last column shows
the estimation results when I also control for earnings smoothness. The coeffi-
cient estimate for (UNEXit∗ PERIODit∗ SOFTit) is 20.1683 (p-value , 0.05).
Overall, the results in Panel A indicate that, when I consider only the firms
that capitalize SDC in the software industry, the results are similar to those
with all firms in this industry. In Panel B of Table 3, I include only expensers
in the software industry. The coefficient estimate for (UNEXit∗ PERIODit∗ SOFTit)
is 20.0291, which is statistically insignificant, and suggests that there is not a sig-
nificant difference in ERC between expensers in the software industry and other
high-tech firms. This result confirms that low earnings quality in the software
industry is driven by capitalizers rather than by expensers, and that the differ-
ences in industry characteristics between the two industry groups are not respon-
sible for the results documented in Table 2.
In Tables 2 and 3, I delete extreme observations in the top and bottom 1% of the
distribution. The breakpoints are generated using all R&D firms. To explore
whether the above results are robust to alternative breakpoint choices, I also try
2%, 3% and 5% as breakpoints for extreme observations. The results get stronger
at higher breakpoints. This result is consistent with Freeman and Tse (1992),
which suggests that ERC has a nonlinear S-shape, being largest around the
center and getting smaller at the tails. However, I report the results with 1% break-
points because it is the most commonly used benchmark, and other choices are ad
hoc. UNEXit is calculated as change in earnings assuming that last year’s earnings
is a reasonable proxy for investors’ earnings expectations. To check the robustness
of the results I use a seasonal random-walk model and calculate UNEXit as earn-
ings in the fourth quarter in year t minus earnings in the same quarter of year t 2 1.
I find that (UNEXit∗ PERIODit∗ SOFTit) is still significantly negative.
Moreover, I calculate CARit over two days around earnings announcements, at
days 21 and 0, instead of the three days over 21 to +1. The magnitude of coef-
ficients is smaller, because CARit is accumulated over only two days, but the
main conclusions do not change. In addition, I scale the change in earnings by
the book value of equity (Compustat item #60) rather than the market value of
equity to calculate UNEXit, and use market-adjusted returns or market-model-
adjusted returns instead of size-adjusted returns. The main conclusions are not
altered by these changes.21 Furthermore, I use industry-adjusted unexpected
earnings, instead of raw unexpected earnings. The industry-adjusted unexpected
earnings are calculated by deducting the mean unexpected earnings in the firm’s
industry from the unexpected earnings of the firm. The coefficient estimate of
444 M. Ciftci
Table 3. Comparison of earnings quality for expensers and capitalizers in software with
other high-tech industriesa
(Continued)
Accounting Choice and Earnings Quality 445
Table 3. Continued
Equation (1) without
Variable controlb Equation (1)b Equation (1A)b
UNEXit∗ PERIODit∗ SOFTit 20.0143 20.0291
(20.21) (20.47)
UNEXit∗ SIZEit 20.0271∗∗
(22.01)
UNEXit∗ BMit 20.1529∗∗∗
(23.03)
UNEXit∗ BETAit 20.0303
(20.83)
UNEXit∗ LOSSit 20.1401∗∗∗
(22.76)
UNEXit∗ LEVit 20.0281
(21.41)
UNEXit∗ PERSISTit 0.0709∗
(1.67)
UNEXit∗ FOLLOWit 20.0312∗
(21.64)
UNEXit∗ PREDISCit 20.0481
(20.99)
Adjusted-R2 0.0103 0.0161
N 2,863 2,863
a ∗ ∗∗
, , and ∗∗∗ denote significance at 0.10, 0.05 and 0.01 respectively. The t-values are in parenthesis.
Before SFAS No. 86 is fiscal years 1981– 1985 and after SFAS No. 86 is fiscal years 1986– 1990.
Variable definitions are in the Appendix. Industry definitions are in Table 1.
b
The model used in estimation:
before 1986. I estimate PCAPITALit from the logistic regression of equation (3).
The pseudo-R2 from this estimation is 17.12%. The procedure for data collection
and elimination of extreme observations is the same as that in the earlier analysis
in section 4.1.1. A firm is considered as a capitalizer if it capitalizes some or all of
its SDC, and as an expenser if it expenses all of its R&D expenditures. The
sample consists of 208 firm-year observations for expensers and 392 firm-year
observations for capitalizers in the sample in the period 1986 – 1995.
Table 4 reports the descriptive statistics for expensers and capitalizers. The
mean EPit for capitalizers is not significantly different from that for expensers.
The mean capitalization intensity, CPINTit, for capitalizers is 0.2926, indicating
that capitalizers capitalize 29% of their total R&D spending. The mean and
median DEVINTit, development intensity, for capitalizers is 17% significantly
greater than for expensers. This is consistent with the argument in Aboody and
Lev (1998) that firms that spend more on R&D are more likely to capitalize.
The mean and median of UNEXit, LEVit, BMit FOLLOWit, PREDISCit and
BETAit for expensers are not significantly different from those for capitalizers.
The mean value of PERSISTit for capitalizers is greater than for expensers
(p-value ¼ 0.08). Interestingly, the mean of LOSSit for expensers is significantly
Accounting Choice and Earnings Quality 447
greater than for capitalizers, raising the possibility that capitalizers use the capi-
talization of SDC to avoid reporting loss. In fact, the mean value of AVLOSSit is
0.1072, suggesting that 10.72% of capitalizers avoid reporting loss by capitaliz-
ing SDC. The mean value of AVDECit is 0.1030, suggesting that 10.30% of capi-
talizers avoid reporting a decline in earnings by capitalizing SDC. In addition, the
mean value of SMOOTHit is 0.7760, suggesting that capitalizers smooth earnings
by capitalizing SDC. This result is consistent with the earnings smoothing by
capitalization of R&D documented by Markarian et al. (2008) in the Italian
setting. The mean MVit (market value of equity) for expensers is greater than
that for capitalizers, indicating that capitalizers are larger firms than expensers.
This result is consistent with Aboody and Lev (1998), who suggest that capitali-
zation is negatively related to firm size. Moreover, the results here also suggest
that my sample is quite different from that in Mohd (2005) (the market value
of capitalizers is six times larger than that of expensers in his sample).
4.2.3 Robustness
In the above analysis I use predicted values (PCAPITALit) instead of actual
choice to capitalize: hence, I could not extend the estimation into the pre-
SFAS No. 86 era. One concern with the above analysis is that there might be sys-
tematic differences between expensers and capitalizers, even in the pre-SFAS No.
86 period. To investigate this issue, I use the actual decision to capitalize or
expense instead of predicted values. Specifically, I consider a firm as a capitalizer
448 M. Ciftci
a
Table 5. Comparison of earnings quality for expensers and capitalizers
Variable Equation (2) without control b Equation (2)b Equation (2A)b
UNEXit 0.1464∗∗∗ 0.7002∗∗∗ 0.7400∗∗∗
(2.96) (5.09) (3.98)
UNEXit∗ PCAPITALit 20.1515∗∗∗ 20.1792∗∗∗ 20.1640∗∗
(22.77) (23.41) (22.38)
UNEXit∗ SIZEit 20.0421∗∗ 20.0541∗∗∗
(22.38) (22.95)
UNEXit∗ BMit 20.2557∗∗∗ 20.2570∗∗∗
(23.21) (23.35)
UNEXit∗ BETAit 20.0292 20.0345
(20.92) (20.89)
UNEXit∗ LOSSit 20.0651 20.0882
(20.72) (20.98)
UNEXit∗ LEVit 20.0744 20.0749
(21.01) (20.98)
UNEXit∗ PERSISTit 0.0751∗ 0.0895∗
(1.65) (1.70)
UNEXit∗ FOLLOWit 20.0515 20.0521
(21.49) (21.51)
UNEXit∗ AVLOSSit 20.0965 20.1834
(20.48) (20.93)
UNEXit∗ AVDECit 20.2496 20.2250
(20.36) (20.34)
UNEXit∗ PREDISCit 20.0739∗ 20.0435
(21.68) (20.89)
UNEXit∗ SMOOTHit 0.1079
(1.26)
Adjusted-R2 0.0113 0.0398 0.0405
N 600 600 489
a ∗ ∗∗
, , and ∗∗∗ denote significance at 0.10, 0.05 and 0.01 respectively. The t-values are in parenthesis.
Expensers are firms that expense all R&D expenditures after SFAS No. 86. Capitalizers are firms that
capitalize some or all R&D expenditures after SFAS No. 86. Variable definitions are in the Appendix.
b
The models used in estimation:
(expenser) in both the post- and pre-SFAS No. 86 periods if it chooses to capita-
lize (expense) in all the post-SFAS No. 86 era. The results are slightly weaker, but
the coefficient estimate of (UNEXit∗ PCAPITALit) is still significantly negative in
the post-SFAS No. 86 era. I also run the robustness checks used in Section 4.1.3.
The tenor of the results does not change for most of the robustness checks, except
Accounting Choice and Earnings Quality 449
that the difference between expensers and capitalizers is insignificant when I use
industry-adjusted UNEXit, probably because industry adjustment introduces noise
into earnings.
Table 6. Descriptive statistics for firms with large increase in capitalized SDC and other
capitalizers
Mean Median
Other Large Other Large
Variable capitalizersa increaseb Differencec capitalizers a
increaseb Differenced
CARit 0.0088 0.0048 0.63 0.0050 20.0023 0.37
EPit 0.0046 20.0120 0.42 0.0380 0.0414 0.04
CH_SOFTit 20.0031 0.0503 0.00 0.0013 0.0309 0.00
CPINTit 0.2279 0.3849 0.00 0.1592 0.3092 0.00
DEVINTit 0.2048 0.1667 0.00 0.1814 0.1566 0.03
UNEXit 20.0440 0.0009 0.02 0.0105 0.0074 0.65
BMit 0.3902 0.5039 0.00 0.3043 0.3860 0.00
LEVit 0.0915 0.1424 0.00 0.0235 0.0599 0.03
BETAit 1.2506 1.1204 0.11 1.2806 1.0816 0.07
LOSSit 0.2712 0.2065 0.10 0.0000 0.0000 0.14
PERSISTit 0.8983 0.7871 0.00 1.0000 1.0000 0.00
MVit 707.84 281.04 0.00 186.75 72.07 0.00
FOLLOWit 1.3142 1.3341 0.42 0.9845 0.9867 0.59
AVLOSSit 0.0248 0.2214 0.00 0.0000 0.0000 0.00
AVDECit 0.0588 0.1915 0.00 0.0000 0.0000 0.00
PREDISCit 1.0544 1.0607 0.54 1.0845 1.0751 0.41
SMOOTHit 0.8012 0.7342 0.02 0.7413 0.6351 0.01
a
Other capitalizers are firms in the bottom three quartiles of change in software capital, CH_SOFTit.
Variable definitions are in the Appendix.
b
Large-increase firms are those in the top quartile of change in software capital, CH_SOFTit.
c
p-value of t-statistics for the difference between large-increase firms and other capitalizers (two-
sided).
d
p-value of Wilcoxon rank sum test whether the populations of large-increase firms and other
capitalizers are equal (two-sided).
450 M. Ciftci
check does not affect the results. In addition, when I define large-increase firms as
those with a change in SDC greater than the median change, instead of the top
quartile, the results are weaker than those reported above, but still statistically
significant.
452 M. Ciftci
investors are pessimistic about capitalizers, they might make more voluntary dis-
closures to increase investor confidence (Mohd, 2005). This argument is also con-
sistent with Gigler and Hemmer (2001), who argue that conservative firms make
fewer voluntary disclosures. Therefore, it would be interesting to investigate the
disclosure practices for firms with different reporting choices.25
Appendix
Definition of Variables
Notes
1
While LaFond and Watts (2008) perform their analysis on conditional conservatism, they do
not restrict their arguments to conditional conservatism. Therefore it is reasonable to expect
that their arguments will also apply to unconditional conservatism (i.e. expensing of SDC).
2
There are several reasons why intangible-intensive firms have more information asymmetry.
First, soft assets are unique: therefore investors cannot get any information about the economic
performance of a specific product by looking at similar products. For example, the market
success of a drug by Pfizer will not say much about the market success of a similar drug by
Merck. In addition, there are no established markets for soft assets that can provide information
about the value of these assets, whereas there are markets for tangible assets, from which inves-
tors can infer the prices of those assets (Lev, 2001).
3
The primary criticism against expensing is that R&D is expensed immediately, whereas the
benefits are realized much later, which violates the matching principle (Lev and Zarowin,
1999). Amir and Lev (1996) suggest that earnings and book values do not represent firm
value and performance, owing to expensing of R&D. However, Collins et al. (1997) (Francis
and Schipper, 1999) did not find any evidence that accounting information is less value-relevant
in intangible-intensive (high-tech) industries.
4
In certain situations capitalization might lead to lower reported earnings than expensing;
however, these situations should be quite rare, because on average R&D expenditures grow
not only nominally but also in real terms (e.g. Skinner, 2008, suggests that inflation-adjusted
R&D has grown by 150% over the past 20 years). Hence amortization expense under capita-
lization, on average, would be smaller than a change in capitalized SDC, leading to greater
earnings and book values under capitalization than under full expensing. I find that in 71%
of the cases in my sample the change in capitalized SDC was larger than amortization
expense.
5
‘The members of SPA CFO committee . . . have indicated the substantial majority of their inves-
tors, underwriters and financial analysts believe financial reporting by software companies is
improved when all SDC are charged to expense as incurred. These users of financial statements
do not believe that recording of soft assets for software being developed is particularly relevant
and does not aid the user of financial statements. The users of financial statements . . . have a
high degree of skepticism when it comes to soft assets resulting from capitalization of SDC.’
SPA letter, March 1996 (Aboody and Lev, 1998).
6
‘An asset should be recognized . . . only if ultimate realization of the assets is reasonably assured
. . .. Due to factors such as the ever-increasing volatility in the software marketplace, the com-
pression of product cycles, the heightened level of competition and divergence of technology
platforms, realization of software assets has become increasingly uncertain even at the point
of technological feasibility . . . We do not believe that software development costs are useful
predictive factor of future product sales.’ This statement is quoted from a letter written by
Ken Walsch, president of SPA, to the FASB (Aboody and Lev, 1998).
7
Consistent with this view, Kothari et al. (2002) suggest that variability of future earnings
associated with R&D expenditures is almost four times greater than for capital expenditures.
Uncertainty of future benefits is also important for standard setters. FASB provides the uncer-
tainty of future benefits as the main reason to justify the requirement to expense the R&D in
other all industries. In SFAS No. 4, FASB states that ‘. . . the relationship between current
456 M. Ciftci
research and development costs and the amount of resultant future benefits to an enterprise is so
uncertain that capitalization of any R&D costs is useful to investors.’
8
Capitalization of SDC: A Survey of Accounting Practices in the Software Industry Sponsored
By Georgia Tech Financial Analysis Lab. CFO Magazine, December 18, 2007.
9
James Mandelson, software analyst at Morgan Stanley, testifying at the FASB public hearing,
May 1985 (Aboody and Lev, 1998).
10
In fact, Mohd (2005) suggests that an alternative proxy to measure information asymmetry is
dispersion in analyst forecasts. Thus, the evidence in Aboody and Lev (1998) suggests that
there is an increase in information asymmetry, rather than a decrease.
11
Cazavan-Jeny and Jeanjean (2006) suggest that capitalizers are the smallest firms (the market
capitalization of expensers in their sample is $6.9 billion – 22 times larger than the $310
million for capitalizers). Similarly, Oswald (2008) reports that the market value of equity for
expensers is seven times larger than for capitalizers ($1529 million versus $216 million).
12
While he controls for size, it may not eliminate the size effect completely in a pooled regression
setting. His sample covers 1986–1995 – ten years, like ours. A large firm in 1986 might seem
small in 1995. A better approach would be to adjust the relative size comparison every year.
13
Consistent with this, Mohd (2005) reports from Davey (1992) that ‘Sulucus Computer has made
every effort to demonstrate that its judgment to [capitalize] is well justified . . . The company
even included, with each investors relations packet it sent out, a cogent and favorable examin-
ation prepared by an analyst from a big, fancy brokerage house.’
14
As long as management is not providing more specific information that will help investors better
understand the relationship between R&D investments and future revenues, disclosing a lump-
sum number for capitalized SDC can hardly be informative. This argument is consistent with
Skinner (2008), who suggests that in industries where full expensing is required there are no
voluntary disclosures for estimated R&D capital, indicating that there is no demand for such
information. Thus, the lack of voluntary supply of estimated R&D capital in industries with
full expensing suggests that there is no such demand.
15
Lipe (1990) defines earnings persistence as autocorrelation in earnings. Schipper and Vincent
(2003) define earnings predictability as the ability of past earnings to predict future earnings,
and earnings smoothness as the lack of variability in earnings across time. Accrual quality is
mapping of accruals into cash flows.
16
Persistence reflected in ERC is different from autocorrelation in earnings (another measure of
persistence used in prior literature), such that the latter might be affected by earnings manage-
ment, whereas the former would not (assuming market efficiency, investors should be able to
see through earnings management).
17
Accounting rules might also affect investment decisions. Oswald and Zarowin (2005) suggest
that expensers engage in real earnings management by cutting R&D, whereas capitalizers
engage in accrual earnings management using capitalization. However, Seybert (2008) suggests
that capitalization of R&D might lead to overinvestment.
18
ERC indicates the extent of a security’s abnormal market return in response to an unexpected
component of reported earnings. Kim and Verrecchia (1991) show analytically that the price
response to a signal is positively related to signal quality, suggesting that the greater the
ERC, the greater the earnings quality. ERC shows investors’ assessment of earnings quality.
It is widely used to measure the quality of information (Christensen, 2002). Ghosh and Moon
(2005) investigate the impact of auditor tenure on earnings quality by using ERC. Collins
and Slatka (1993) use ERC to investigate how the adoption of SFAS No. 52 affects the
quality of earnings. Similarly, Teoh and Wong (1993) use ERC to measure the impact of
auditor quality on earnings quality.
19
Consistent with this argument, Bhattacharya et al. (2003) suggest that actual earnings from I/B/
E/S are significantly different from earnings under GAAP, and argue that GAAP earnings
are more conservative than actual earnings in I/B/E/S. Similarly, Collins et al. (2009) use
I/B/E/S actual earnings as a proxy for street earnings, an earnings definition that is not
Accounting Choice and Earnings Quality 457
consistent with GAAP. Therefore it may not be reasonable to test the impact of accounting
choice under GAAP using analyst forecasts.
20
This analysis might suffer from selection bias, because we are selecting firms in the software
industry based on their accounting choice. However, consistent with Mohd (2005), I perform
this analysis as a robustness check to see whether the results are driven by expensers within
the software industry.
21
Size-adjusted returns are better than market- or market-model-adjusted returns because market
capitalization in the calculation of size-adjusted returns is updated every year. Given that market
capitalization grows over time, adding a control variable for firm size may not eliminate the size
effect completely in pooled regressions, because the sample period covers ten years.
22
Mohd (2005) suggests that DEVINTit increases the chance of success in software projects. Given
that large capitalizers have lower DEVINTit, they might have less success in these projects.
23
Oswald and Zarowin (2007) use capitalization data based on the UK. They argue that their find-
ings are comparable to the US, because UK capital markets are well-developed, liquid markets,
similar to those of the US and other developed nations. Following their arguments, it is reason-
able to assume that European data, at least in some countries, are comparable to those for the
US.
24
There are several studies that explore the impact of capitalization with European data; however,
the results are also mixed. Cazavan-Jeny and Jeanjean (2006) find that capitalized costs are not
value-relevant. Oswald (2008) did not find any economically significant improvement from the
hypothetical capitalization of R&D for firms that expense it in the UK. However, Oswald and
Zarowin (2007) find that stock prices are more informative capitalizers than expensers,
suggesting that capitalization is informative.
25
Oswald (2008) investigates the determinants of the decision to capitalize or expense. However,
to my knowledge, no study investigates the impact of capitalization versus expensing on volun-
tary disclosures.
References
Aboody, D. and Lev, B. (1998) The value relevance of intangibles: the case of software capitalization,
Journal of Accounting Research, 36, pp. 161–191.
Ali, A. and Zarowin, P. (1992) The role of earnings levels in annual earnings-regressions, Journal of
Accounting Research, 30, pp. 286–296.
Amir, E. and Lev, B. (1996) Value-relevance of nonfinancial information: the wireless communi-
cations industry, Journal of Accounting & Economics, 22, pp. 3–30.
Association for Investment Management and Research (1993) Financial Reporting in the 1990s and
Beyond (Charlottesville, VA: AIMR).
Atiase, R. (1985) Predisclosure information, firm capitalization and security price behavior around
earnings announcements, Journal of Accounting Research, 23, pp. 21–36.
Ball, R. and Shivakumar, L. (2008) Earnings quality at initial public offering, Journal of Accounting &
Economics, 45, pp. 324–349.
Bhattacharya, N., Black, E. L., Christensen, T. E. and Larson, C. R. (2003) Assessing the relative
informativeness and permanence of pro forma earnings and GAAP operating earnings,
Journal of Accounting & Economics, 36, pp. 285–319.
Cazavan-Jeny, A. and Jeanjean, T. (2006) The negative impact of R&D capitalization: A value rel-
evance approach, European Accounting Review, 15, pp. 37–61.
Christensen, T. E. (2002) The effects of uncertainty on the informativeness of earnings: Evidence from
the insurance industry in the wake of catastrophic events, Journal of Business Finance &
Accounting, 29, pp. 223–255.
Christensen, T. E., Smith, T. Q. and Stuerke, P. S. (2004) Public predisclosure information, firm size,
analyst following and market reactions to earnings announcements, Journal of Business Finance
& Accounting, 31, pp. 951–984.
458 M. Ciftci
Oswald, D. R. (2008) The determinants and value relevance of the choice of accounting of research
and development expenditures in the United Kingdom, Journal of Business Finance & Account-
ing, 35, pp. 1–24.
Oswald, D. R. and Zarowin, P. (2005) Capitalization vs expensing of R&D and earnings management,
Unpublished Working Paper, New York University.
Oswald, D. R. and Zarowin, P. (2007) Capitalization of R&D and the informativeness of stock prices,
European Accounting Review, 16, pp. 703–726.
Petersen, M. A. (2009) Estimating standard errors in finance panel data sets: comparing approaches,
Review of Financial Studies, 22, pp. 435–480.
Ryan, S. G. (2006) Identifying conditional conservatism, European Accounting Review, 15,
pp. 511–525.
Schipper, K. and Vincent, L. (2003) Earnings quality, Accounting Horizons, 17, pp. 97–110.
Scott, W. R. (2006) Financial Accounting Theory (Prentice Hall: Toronto).
Seybert, N. (2008) R&D capitalization and overinvestment in continuing projects. Unpublished
Working Paper, University of Texas at Austin.
Skinner, D. J. (2008) Accounting for intangibles: a critical review of policy recommendations,
Accounting & Business Research, 38, pp. 191–204.
Teoh, S. H. and Wong, T. J. (1993) Perceived auditor quality and the earnings response coefficient,
The Accounting Review, 68, pp. 346–366.