You are on page 1of 24

Journal of Accounting and Economics 63 (2017) 75–98

Contents lists available at ScienceDirect

Journal of Accounting and Economics


journal homepage: www.elsevier.com/locate/jae

Conditionally conservative fair value measurements$


Marc Badia a, Miguel Duro a, Fernando Penalva a,n, Stephen Ryan b
a
IESE Business School, University of Navarra, Ave. Pearson, 21, Barcelona 08034, Spain
b
Leonard N. Stern School of Business, New York University, 44 West, 4th Street, 10-73, New York, NY 10012, United States

a r t i c l e i n f o abstract

Article history: Firms measure fair values using Level 2 or 3 inputs when items do not trade in liquid
Received 4 March 2015 markets, limiting market discipline over the measurements. We provide evidence that
Received in revised form firms holding higher proportions of financial instruments measured at Level 2 and 3 fair
8 August 2016
values report more conditionally conservative comprehensive income attributable to fair
Accepted 17 October 2016
Available online 21 October 2016
value measurements, consistent with firms trying to mitigate investors' discounting of the
measurements. We further predict and find that this conditional conservatism (1) in-
JEL Classifications: creases with governance mechanisms that increase the strength and persistence of firms'
C23 incentives to report conservatively and (2) decreases with firms’ earnings management
D21
incentives.
G32
& 2016 Elsevier B.V. All rights reserved.
M41

Keywords:
Conditional conservatism
Fair value measurement
Disclosure

1. Introduction

We investigate whether firms' recurring measurements of the fair values of recognized financial instruments (hereafter
“fair value measurements”) exhibit conditional conservatism when the instruments do not trade in liquid markets, thereby
enabling them to exercise discretion over the measurements.1 Absent such discretion, fair value measurements should be


We gratefully acknowledge helpful comments and suggestions from Michelle Hanlon (editor), Sugata Roychowdhury (referee), Manuel Cano-Rodríguez
(discussant), Dmitri Byzalov (discussant), Shira Cohen, Ethan Rouen, David Elliott and Carrie Nermo from the Alberta Securities Commission, John Lee from
the Securities Exchange Commission, and participants at the 2015 European Accounting Association Annual Conference, the 2015 American Accounting
Association Annual Meeting, and the XI Workshop on Empirical Research in Financial Accounting. We greatly appreciate the Alberta Securities Commission,
CanOils Database Ltd., and Toronto Stock Exchange for providing access to their databases, as well as Brad Badertscher for providing the frequencies of
other-than-temporary impairments of investment securities. Marc Badia and Miguel Duro acknowledge financial support from the Spanish Ministry of
Science and Innovation, Grant ECO2010-19314, and from Columbia University CIBER, respectively. Marc Badia and Fernando Penalva acknowledge financial
support from the Spanish Ministry of Economics and Competitiveness, Grants ECO2013-48328-C3-1-P, and ECO2016-77579-C3-1-P.
n
Corresponding author.
E-mail address: penalva@iese.edu (F. Penalva).
1
Following Ball and Shivakumar (2005) and Beaver and Ryan (2005), the accounting literature refers to two distinct forms of conservatism: conditional
and unconditional. Conditional conservatism involves the timelier recognition of unrealized losses than of unrealized gains. That is, recognition is con-
ditional on the receipt of news and the timeliness of recognition is conditional on the sign of the news received. Unconditional conservatism involves the
predetermined (i.e., news-independent) understatement of net assets. Ball and Shivakumar (2005) argue that conditional conservatism has a clear role in
contracting, whereas unconditional conservatism yields uninformative biases. Beaver and Ryan (2005) show that unconditional conservatism preempts
conditional conservatism, so that the two types of conservatism, while conceptually distinct, are empirically interrelated.

http://dx.doi.org/10.1016/j.jacceco.2016.10.006
0165-4101/& 2016 Elsevier B.V. All rights reserved.
76 M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98

unbiased, i.e., symmetrically incorporate unrealized gains and unrealized losses. Policymakers and academics often make
statements or research design choices that suggest fair value measurements generally are unbiased, ignoring biases in-
troduced by discretion. For example, SEC (2008) states that the “objective of SFAS No. 157 and of existing fair value standards
is to provide transparent unbiased information about fair value.” Shaffer (2011) describes fair value measurement as “an
anchor for”…“an asset/liability approach… [in which] valuations must be objective and neutral. This requirement overrides
the constraint of conservatism, which historically assumed a downward bias for asset valuations.” (Emphasis added in both
quotes.) Empirical research examining fair value measurements' predictive power and their value-, returns-, and risk-re-
levance often sums fair values or fair value gains and losses, which is appropriate only if the summed items are measured
comparably (e.g., are unbiased or have the same proportional bias).2
Ostensibly in contrast to the view that fair values are unbiased, some influential observers claim that fair value mea-
surements for instruments that do not trade in liquid markets represent optimistic or noisy “mark-to-myth” (Buffett, 2003)
or “mark-to-make-believe” (Weil, 2007) valuations that are discounted by investors. Empirical research provides evidence
supporting these claims (Kolev, 2009; Song et al., 2010; Goh et al., 2009). To mitigate such discounting, we expect firms to
take steps to increase the perceived reliability of these measurements. The step we examine is suggested by the extensive
literature that documents the contracting and other benefits of conditional conservatism (e.g., Basu, 1997; Watts, 2003).
Roychowdhury (2010) emphasizes the critical role of explicit and implicit governance mechanisms, which increase the
strength and persistence of firms' incentives to report conditionally conservatively, to obtain these benefits. Although firms
make reporting decisions each period, when these incentives are sufficiently strong and persistent, they effectively con-
stitute a “credible commitment” to conditional conservatism (e.g., Zhang, 2008).
We identify the extent of firms' discretion over fair value measurements using disclosures provided under Statement of
Financial Accounting Standards (SFAS) 157, Fair Value Measurements (FASB, 2006). This standard requires firms to disclose,
for each major category of asset and liability, the fair value measurement amounts that are based on active market prices for
the items being measured (Level 1 inputs), other observable information (Level 2 inputs), or unobservable information
(Level 3 inputs).3 These three input levels differ in their observability and individual completeness as measures of fair value,
and thus in the extent of discipline they provide over firms' fair value measurements. Level 1 inputs are both observable and
individually complete measures of fair value, and thus provide a very high level of discipline over fair value measurements.
Level 2 inputs are observable but individually incomplete measures of fair value, requiring either adjustments to observed
market prices for the illiquidity or dissimilarity of the items being fair valued or model-based calculations involving other
inputs to determine fair values. Level 3 inputs are both unobservable and individually incomplete measures of fair value. We
refer to Level 2 or 3 inputs or fair value measurements based on these inputs as “lower-level”.
We predict that firms mitigate investors' discounting of lower-level fair value measurements by reporting the mea-
surements conditionally conservatively, i.e., by recognizing unrealized losses in a timelier fashion than unrealized gains. We
further predict that firms report more conditionally conservative fair value measurements when the measurements are
evaluated by more knowledgeable investors, verified by more independent third parties, and disclosed more fully in fi-
nancial reports, governance mechanisms that increase the strength and persistence of firms' incentives to report con-
ditionally conservatively. Lastly, we predict that firms that narrowly meet or beat earnings targets, an earnings management
incentive that typically conflicts with conditional conservatism, report less conditionally conservative fair value
measurements.
To test these predictions, we measure conditional conservatism using two expanded versions of Basu's (1997) approach.4
The dependent variables in these expanded models are the portions of comprehensive income (CI) attributable to fair value
measurements of financial instruments (FVCI) and the remainder of CI (CI–FVCI). We interact the proportions of fair value
measurements at the various levels with the explanatory variables in recent implementations of Basu's (1997) approach.
Consistent with the discussion above, we expect firms with higher proportions of Level 2 and 3 fair value measurements,
but not those with higher proportions of Level 1 measurements, to exhibit more conditionally conservative FVCI.
In contrast, we do not expect any association between the fair value measurement level proportions and the conditional
conservatism of CI–FVCI. Although CI–FVCI includes any conditionally conservative fair-value-based impairment write-
downs of tangible and intangible assets (Ramanna and Watts, 2012; Roychowdhury and Martin, 2013), there is no apparent
reason why the frequency or size of these non-recurring write-downs should vary with the proportions of recognized fi-
nancial instruments measured at the various fair value levels on a recurring basis. In this regard, the analysis of CI–FVCI
constitutes a placebo test.
To test our predictions about variation in incentives to report conditional conservative Level 2 and 3 fair value mea-
surements across firms and time, we compare the conditional conservatism of FVCI across four partitions of the full sample.

2
For example, the studies examining the value-relevance of SFAS 157 fair value measurement-level disclosures discussed in Section 2 exhibit such
features.
3
SFAS 157 was incorporated into the FASB's Accounting Standards Codification (ASC), Topic 820, in 2009. The FASB amended Topic 820 in 2010 through
Accounting Standards Update No. 2010-06 – Fair Value Measurements and Disclosures (Topic 820): Improving Disclosures about Fair Value Measurements, and
in 2011 through Accounting Standards Update No. 2011–04 – Fair Value Measurement Topic 820): Amendments for Common Fair Value Measurement and
Disclosure Requirements in U.S. GAAP and IFRSs. These amendments involved fairly subtle changes to fair value measurements and disclosures that do not
affect our hypotheses or empirical analysis.
4
We discuss the assumptions and limitations of Basu's (1997) approach and why we apply this approach to our research question in Section 3.
M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98 77

First, to capture the ability of firms' investors to evaluate lower-level fair value measurements, we partition based on the
proportion of the firm's equity that is owned by non-transient institutional investors. Second, to capture the verification of
firms’ lower-level fair value measurements by independent third parties, we partition based on an indicator for audit
quality. Third, to capture firms' earnings management incentives that conflict with reporting conditionally conservative
lower-level fair value measurements, we partition based on an indicator for the firm narrowly meeting or beating earnings
targets versus clearly missing these targets. Fourth, to capture investors' improved ability to evaluate firms' Level 2 and
3 measurements under SFAS 157's expanded disclosure requirements, we examine the sample in the first year before the
required adoption of SFAS 157, 2007, versus the first year after that adoption, 2008.5 In addition, we expect SFAS 157's
expanded disclosures to strengthen the effects of the three firm-level partitioning variables. To capture these interactive
effects, we compare the conditional conservatism of FVCI for the sets of subsamples formed based on each of these proxies
in 2007 versus 2008.
Our empirical findings are largely consistent with these predictions. As expected, in the full sample we find that firms
with higher Level 2 and 3 proportions exhibit more conditionally conservative FVCI but not CI–FVCI. In the subsamples
formed partitioning on a single variable, we find that these results are more pronounced for firms with higher non-transient
institutional ownership, with higher audit quality, and that do not narrowly meet or beat earnings thresholds, as well as in
2008. The results are mostly but less uniformly consistent with expectations in the sets of subsamples formed partitioning
on both a firm-level proxy and 2007 versus 2008, possibly due to the smaller sample sizes in these two-way partitions.
To demonstrate the generalizability of our results to balance sheet items other than financial instruments, we conduct
similar analysis on pro forma earnings measures calculated using Canadian and U.S. oil and gas firms' disclosed measures of
reserves, which are akin to Level 2 or 3 fair value measurements. We find that the conditional conservatism of these
earnings measures increases with the inclusion of these reserve measures.
To the best of our knowledge, ours is the first paper to examine firms' exercise of discretion to report conditionally
conservative recurring Level 2 and 3 fair value measurements for recognized financial instruments. Our paper contributes to
ongoing debates about the relative merits of unbiased versus conditionally conservative accounting measurement and,
relatedly, of fair value versus (impaired) amortized cost measurement. Accounting standard setters, securities regulators,
and some investors and academics argue that expanded requirements for firms to measure financial instruments and si-
milar items at (unbiased) fair value would increase the usefulness of financial reports (e.g., FASB, 2011; SEC, 2008; Barth,
2006, 2014; Barth and Taylor, 2010; Ryan, 2008). Consistent with these arguments, the FASB recently issued Statement of
Financial Accounting Concepts No. 8, which eliminated conservatism from the conceptual framework.6 The Board's rationale
is that conservative biases prevent accounting numbers from being representationally faithful, an essential qualitative
characteristic of useful financial information. Taking an opposing view, other investors and academics argue that eliminating
conditional conservatism without understanding the reasons for its longstanding existence and pervasive use likely would
generate unintended adverse economic consequences (e.g., Watts, 2003; Nissim and Penman, 2008; Kothari et al., 2010).
These authors also point out that, despite the FASB's disavowal of conservatism, over its life the Board has repeatedly and
even increasingly issued standards that require conditional conservatism (Watts, 2003). Our findings indicate that in
practice firms report conditionally conservative Level 2 and 3 fair value measurements, potentially mitigating concerns
about unintended adverse consequences of fair value accounting requirements.
Our paper complements prior research that finds firms exercise discretion over the timing and amount of (fair-value-
based) impairment write-downs, typically for non-financial assets (Ramanna and Watts, 2012; Roychowdhury and Martin,
2013). The settings examined in this prior research differ from our setting in the following crucial respect. Asset impairment
standards require conditional conservatism (i.e., asset write-ups are prohibited). In contrast, the FASB's intent in SFAS 157 is
to yield unbiased reporting, consistent with CON 8's revisions to the conceptual framework. We provide evidence that,
despite this intent, in practice reporting firms exercise discretion to meet the market's demand for conditionally con-
servative fair value measurements. We believe this evidence is likely to generalize to other settings involving lower-level fair
value measurements, because it is difficult to legislate against market demand.
Our paper exhibits three more specific but still notable differences from this prior research. First, asset impairments are
non-recurring, being recorded only when specified, and sometimes hard to meet, conditions are satisfied. For example, firms
may have to determine that assets are other-than-temporarily impaired or have less than fully recoverable amortized cost
bases to record impairments. This feature implies that firms' exercise of discretion over asset impairments is substantially
about whether to record impairments in given periods. In contrast, we examine firms' exercise of discretion over fair value
measurements every period. Second, impairments often involve tangible or intangible assets for which fair value mea-
surements rarely are based on Level 1 inputs. This feature constrains variation in the extent of firms' discretion over im-
pairments, which is generally fairly high. In contrast, in our setting firms' discretion varies between high for Level 2 and
3 measurements and low for Level 1 measurements. Third, firms record asset impairments only when news is sufficiently
bad. Hence, studies documenting an association between impairment and conditional conservatism using Basu's (1997)

5
Following Riedl and Serafeim (2011), we proxy for the undisclosed fair value measurement level proportions pre-SFAS 157 using the disclosed
proportions in the first post-SFAS 157 year, 2008 for regular adopters.
6
Statement of Financial Accounting Concepts No. 8, Conceptual Framework for Financial Reporting – Chapter 1, The Objective of General Purpose
Financial Reporting, and Chapter 3, Qualitative Characteristics of Useful Financial Information (a replacement of FASB Concepts Statements No. 1 and No. 2)
(2010, BC3.19, BC3.27–29).
78 M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98

approach essentially show differential response to bad news captured in (fair-value-based) impairments versus good news
captured in non-fair-value-based measurements. In contrast, we document differential response to bad news conveyed by
different levels of fair value measurements.
Lastly, our results suggest that empirical research examining fair value measurements' predictive power and their value-,
returns-, and risk-relevance would exhibit improved model specification and statistical power by distinguishing these
measurements based on their extent of conditional conservatism.

2. SFAS 157 and related prior research

SFAS 157, which became effective in 2008 with early adoption allowed in 2007, defines fair value and provides com-
prehensive guidance regarding fair value measurement. As a central part of that guidance, the standard specifies a three-
level hierarchy of fair value measurement inputs. Level 1 (highest-level) inputs are quoted prices in active markets for
identical items. SFAS 157 requires reporting firms to measure fair values using Level 1 inputs, generally unadjusted, if they
exist. Level 2 inputs are most other directly or indirectly observable inputs. Level 2 inputs include quoted prices for similar
assets in active markets or identical items in less than active markets. Quoted priced-based Level 2 inputs generally require
adjustment for the dissimilarity or illiquidity of the item being fair valued. Level 2 inputs also include published interest
rates, yield curves, volatilities, prepayment speeds, default rates, loss severities, credit ratings, et cetera. These rawer Level
2 inputs generally must be plugged into models to generate fair value estimates, and the accuracy of these estimates
depends critically on the validity of the models used. Level 3 (lowest-level) inputs are based on the reporting entity's
internal data or are otherwise unobservable, but they should reflect the inputs that market participants would use in pricing
the item. Generally, these inputs must also be plugged into models. SFAS 157 “prioritizes” Level 2 over Level 3 inputs but
allows reporting firms to use Level 3 inputs rather than available Level 2 inputs if the former inputs reflect “the assumptions
that market participants would use in pricing the asset or liability.” The level of a fair value measurement is determined by
the lowest significant input into the measurement.
SFAS 157 requires firms to disclose the amounts of each level of recurring recognized fair value measurements for each
major category of asset or liability (including major types of investment securities). The standard and its amendments
require expanded disclosures for Level 3 (and to a lesser extent Level 2) measurements due to their greater subjectivity.
Two sets of prior papers examine SFAS 157-required fair value measurement-level disclosures for financial assets and
liabilities recognized at fair value. First, Kolev (2009), Song et al. (2010), and Goh et al. (2009) examine the value-relevance
of these disclosures. These studies provide evidence that financial instrument fair values based on higher level inputs are
more value-relevant. Specifically, all three studies find that Level 1 and 2 fair values are more value-relevant than Level 3 fair
values. Goh et al. (2009) find that Level 1 fair values are more value-relevant than Level 2 fair values. Song et al. (2010) find
that Level 3 fair values are more value-relevant for banks with better corporate governance.
Second, Lev and Zhou (2009) and Riedl and Serafeim (2011) examine the liquidity- and information-risk relevance of
SFAS 157 fair value measurement-level disclosures. These studies provide evidence that firms with more financial assets
with fair values measured using lower-level inputs exhibit higher risk. Specifically, Lev and Zhou (2009) argue that the three
fair value measurement levels indicate liquidity risk. Consistent with this argument, they find that, for events during the
recent financial crisis that impaired market liquidity, the share price reaction is most negative for financial institutions
holding Level 3 financial assets, followed by Level 2 assets and then by Level 1 assets. The reverse holds for events that
improved market liquidity. Riedl and Serafeim (2011) argue that the three levels indicate information risk. They provide
evidence that firms with more Level 3 financial assets exhibit higher betas than those with Level 1 or Level 2 assets, with
these differences being more pronounced for firms with poorer information environments. They also show that SFAS 157
adoption reduces differences in betas across the fair value levels for firms with richer information environments.

3. Research design

We measure conditional conservatism using expanded versions of Basu's (1997) approach. Basu regresses a measure of
income deflated by beginning market value of equity on share returns (hereafter “returns”), an aggregate proxy for news
from all sources, allowing for separate slope coefficients on positive returns (good news) and negative returns (bad news).
This model captures the fact that conditional conservatism imposes a higher verification threshold to recognize unrealized
gains than to recognize unrealized losses, so that income incorporates bad news faster than good news. Consistent with this
fact, Basu expects and finds that income is more positively associated with returns when they are negative than when they
are positive.
Dietrich et al. (2007), Givoly et al. (2007), Patatoukas and Thomas (2011), and others criticize Basu's (1997) approach
primarily based on its measure of news, arguing that returns are endogenous and their variance decreases with firm size.
Ball et al. (2013a, 2013b) and others counter that these criticisms are addressable through adjustments to the model
specification.
We use Basu's approach in this study for two primary, non-mutually exclusive reasons. First, non-news-based measures
do not capture conditional conservatism as directly as news-based measures (Ryan, 2006). Second, the markets for debt
M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98 79

instruments, the vast majority of firms’ financial instruments measured at fair value, were highly illiquid during the financial
crisis. In contrast, equity markets remained quite liquid during the crisis.7 Hence, returns likely constituted the best available
summary news measure with which external parties could verify firms' fair value measurements during this period.
This point highlights the fact that, in using Basu's (1997) approach, we assume that returns is a good proxy for external
parties' information about changes in the fair values of firms' financial instruments. An analogous assumption is made in
prior studies using Basu's approach. Two aspects of our results provide comfort that this assumption does not drive our
results. First, we find predictable differences in the conditional conservatism of Level 2 and 3 fair value measurements across
sample partitions based on proxies for investors' ability to evaluate these measurements, the verification of these mea-
surements by independent third parties, and firms' incentives not to report conditionally conservatively. There is no obvious
reason why any limitation of returns as a measure of news would vary across these partitions in this fashion. Second, in
Section 6 we analyze oil and gas firms' disclosures of measures of their reserves that are akin to Level 2 and 3 fair value
measurements. In untabulated analyses, we verify that these reserve measures exhibit conditional conservatism using a
measure of news that is tailored to the setting: the portion of returns explained by oil and gas price index returns.
A now common expansion of Basu's (1997) model interacts the beginning market-to-book ratio (MTBt-1) with the two
returns variables. Accounting standards that govern asset impairments typically require assets to be written down when
their fair values drop sufficiently below their book values. MTBt-1 captures the amount of slack that firms have built up as of
the beginning of the period to absorb bad news before impairment write-downs are required (Beaver and Ryan, 2005;
Roychowdhury and Watts, 2007). Relatedly, Lawrence et al. (2013) view MTBt-1 as an inverse proxy for non-discretionary
conditional conservatism. To ensure that our fair value measurement level proportions do not capture these previously
documented effects, we also include interactions of MTBt-1 with the returns variables in the model. Following the re-
commendations of Roychowdhury and Martin (2013, pp. 142), we do not include other control variables, although our
inferences are robust to controlling interactively for leverage and size.
To examine the distinct impacts of the fair value measurement levels on conditional conservatism, we expand Basu's
(1997) model further by incorporating the fair value measurement level proportions and their interactions with the returns
variables:
FVCIt = β0 + β1Dt + β2Rett + β3Dt * Rett
+β4 MTBt − 1 + β5Dt * MTBt − 1 + β6Rett * MTBt − 1 + β7Dt * Ret * MTBt − 1
+β8Lev1t + β9Dt * Lev1t + β10Rett * Lev1t + β11Dt * Rett * Lev1t
+β12Lev23t + β13Dt * Lev23t + β14 Rett * Lev23t + β15Dt * Rett * Lev23t + εt (1)

Appendix 1 defines all model variables. The dependent variable, FVCIt, is the portion of comprehensive income attri-
butable to financial instruments recognized at fair value divided by the firm's market value of equity at the beginning of the
year. The numerator of FVCIt contains all significant unrealized and realized gains and losses on these instruments iden-
tifiable on COMPUSTAT, including: (1) net realized gains on trading, available-for-sale (AFS), and held-to-maturity (HTM)
investment securities, net unrealized gains on trading securities, and net hedge ineffectiveness gains that are recorded in
net income; and (2) net unrealized gains on AFS securities, HTM securities (post-April 2009), and cash-flow-hedge deri-
vatives that are recorded in other comprehensive income. We include both unrealized and realized gains and losses in FVCIt
because realized gains and losses mechanically reduce net unrealized gains and losses dollar for dollar, and we want FVCI to
be unaffected by the likely asymmetric realization of gains and losses (due to tax minimization and other reasons). In-
tuitively, FVCI equals what unrealized gains and losses would have been had the firm not realized any gains and losses
during the period,8 which is often referred to as “total” gains and losses (e.g., Ryan, 2007). Rett is the firm's stock return for
the 12 months ending three months after the fiscal year end. Dt is an indicator variable that takes a value of 1 when Rett is
negative and 0 otherwise.
Lev1t (Lev23t) is the fair value of recognized financial assets and liabilities that are measured at fair value on a recurring
basis using Level 1 (Level 2 and 3) inputs divided by the carrying value of total assets.9 In most analyses, we combine Level
2 and Level 3 fair value measurements because the latter constitute a very small proportion of financial instruments re-
cognized at fair value, although this choice does not alter our inferences. We calculate the fair value measurement level
proportions deflating by total assets so that these proportions reflect the percentages of firms' balance sheets that are
measured at these levels. In Eq. (1), these are the most natural proportions to mediate the relation between Rett, which
reflects all firm-level news, whether attributable to balance sheet items measured at fair value or other exposures, and FVCIt,
which reflects only income attributable to items measured at fair value.10,11 Henceforth, we suppress time subscripts except

7
For example, Longstaff (2010) refers to equity markets as “highly liquid” during the financial crisis. He provides evidence that the less liquid ABX
index forecast S&P 500 share returns as much as three weeks ahead during 2007 as the crisis was unfolding, but not subsequently.
8
As discussed in Section 5.1, we conduct specification analyses separating FVCI into unrealized gains and losses versus realized gains and losses.
9
Our inferences are unaffected if we calculate the fair value measurement level proportions including only financial assets measured at fair value in
the numerator, because financial liabilities measured at fair value are immaterial.
10
Calculating the fair value measurement level proportions deflating by financial assets would instead be appropriate if, unlike Ret, the measure of
news reflected only financial instruments measured at fair value. We are unaware of the availability of a generally applicable measure of news of this type.
11
Readers may ask whether we should calculate the fair value measurement level proportions deflating by market value of equity, the deflator of the
80 M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98

when necessary for clarity.


Ideally, the numerator of FVCI would include only the comprehensive income attributable to recurring fair value mea-
surements of financial instruments. However, due to COMPUSTAT's aggregation of the data items used to calculate FVCI, this
variable includes one type of non-recurring fair value measurement: other-than-temporary (OTT) impairments of HTM
securities.12 This inclusion should not significantly affect our results for three reasons. First, OTT impairments of HTM
securities are infrequent and small compared to bank assets.13 Second, the inclusion of non-recurring impairments in FVCI
should bias against finding support for our predictions regarding the interactions of returns with Lev1 and Lev23, because
these variables only reflect the proportions of recognized financial instruments measured at fair value on a recurring basis.
Third, we conduct untabulated specification analyses estimating Eq. (1) using only unrealized gains and losses recorded in
other comprehensive income as the dependent variable and, alternatively, on a smaller sample of bank holding companies
(BHCs) with available regulatory Y-9C data that enables FVCI to be calculated purely in terms of recurring fair value mea-
surements. As discussed in Section 5.1 below, the results of these specification analyses are consistent with the tabulated
results.
Because Level 1 fair value measurements leave little room for managerial discretion, we expect β11 on D*Ret*Lev1 (re-
ferred to as the asymmetric timeliness coefficient for those measurements) to be zero. Because Level 2 and 3 fair value
measurements allow for discretion, we expect β15 on D*Ret*Lev23 (referred to as the asymmetric timeliness coefficient for
those measurements) to be positive. Whereas zero is the most appropriate null hypothesis (or benchmark) for β15, we also
expect this coefficient to be larger than β11. Although a less direct test of differential conditional conservatism, we also
expect the sum of β14 on Ret*Lev23 and β15 (referred to as the total bad news timeliness coefficient for Level 2 and
3 measurements), to be significantly larger than the sum of β10 on Ret*Lev1 and β11 (referred to as the total bad news
timeliness coefficient for Level 1 measurements).
As a placebo test, we also estimate Eq. (1) replacing FVCI as the dependent variable with the portion of comprehensive
income not attributable to recognized fair value measurements of financial instruments divided by the firm's market value
of equity at the beginning of the year, CI–FVCI. Although we expect CI–FVCI to exhibit conditional conservatism, this
conservatism should not vary with the fair value measurement level proportions.
All pooled regressions include firm and year fixed effects (Ball et al., 2013a, 2013b) and report standard errors calculated
clustering observations by firm (Petersen, 2009). Unless indicated otherwise, significance levels are calculated using two-
tailed tests.

4. Sample selection and descriptive statistics

We obtain accounting data, including recurring measurements of the fair values of recognized financial instruments
using the various input levels, from COMPUSTAT, stock market data from CRSP, and analyst data from IBES. The sample
period covers the eight years 2007–2014. This period begins in 2007, the year the few early adopters of SFAS 157 made the
required disclosures of fair value measurements of financial instruments in the three levels. Regular adopters first made
these disclosures in 2008. To be included in the full sample, we require observations to have non-missing values of the
variables in Eq. (1), which yields 27,904 firm-year observations.14 Analyses involving partitions of the full sample may lose
observations due to the non-availability of the partitioning variable. To mitigate the influence of outliers while not dis-
proportionately truncating observations in given years, we winsorize continuous variables at the 1st and 99th percentiles of
their distributions each year. Our inferences are unaffected by instead winsorizing at the same percentiles in the pooled
sample.
Table 1 reports the number of observations and the means of the fair value measurement level proportions for each of
the Fama-French 17 industry classifications.15 As expected, financial firms (industry #16) have the highest proportion of
assets that are fair valued financial instruments. The vast majority of fair valued financial instruments are assets, with the
ratio of fair value assets to fair value liabilities being over 7 to 1. The mean of Level 2 financial instruments is slightly higher
than the mean of Level 1 instruments, particularly for financial firms, and both far exceed the mean of Level 3 instruments.16

(footnote continued)
dependent variable in Eq. (1), FVCI. The use of this alternative deflator would introduce rather than resolve inconsistency, however, because these pro-
portions are interacted with Ret, which is already deflated by the market value of equity. Relatedly, deflating by market value of equity effectively would
multiply the proportions by firms' financial leverage, thereby conflating two distinct constructs.
12
In contrast, FVCI does not include OTT impairments of AFS securities, because these impairments reclassify unrealized losses on these securities
from other comprehensive income to net income, with no effect on comprehensive income.
13
For example, Brad Badertscher indicates in private communication that Badertscher et al. (2014) sample of banks from the first quarter of 2007 to
the fourth quarter of 2011 includes only 15 OTT impairments of HTM securities that summed to $6.3 billion, compared to commercial banking industry
assets of $11.9 trillion at the end of 2009. Hence, these impairments reduced the industry return on assets by approximately 0.01% per year (¼ $6.3 billion/
($11.9 trillion * 5 years)) during Badertscher et al.'s five-year sample period.
14
The full sample contains 38 observations of early adopters in 2007. Dropping these observations, which limits the full sample to 2008–2014, does not
affect our inferences.
15
See http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html for definitions of the classifications.
16
If Compustat reports a missing value for a fair value measurement level in a sample year from 2008 onwards, we set that value equal to zero. Our
inferences are not affected by this choice.
M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98 81

Table 1
Number of observations and means of fair value measurement level proportions Fama-French 17 industries.

Lev1 Lev2 Lev3 Lev1 Lev2 Lev3


N A A A L L L

1 Food 747 0.02 0.02 0.00 0.00 0.01 0.00


2 Mining and minerals 542 0.03 0.01 0.00 0.00 0.01 0.00
3 Oil and petroleum products 1,383 0.01 0.01 0.00 0.01 0.01 0.00
4 Textiles, apparel & footwear 374 0.03 0.02 0.00 0.00 0.01 0.00
5 Consumer durables 387 0.02 0.01 0.00 0.01 0.01 0.00
6 Chemicals 489 0.04 0.01 0.00 0.01 0.01 0.00
7 Drugs, soap, perfumes, tobacco 857 0.11 0.08 0.00 0.00 0.01 0.01
8 Construction and constr. materials 671 0.03 0.01 0.00 0.00 0.02 0.00
9 Steel works etc. 315 0.03 0.02 0.00 0.01 0.01 0.00
10 Fabricated products 165 0.03 0.01 0.00 0.00 0.00 0.00
11 Machinery and business equipment 3,410 0.07 0.05 0.01 0.00 0.01 0.00
12 Automobiles 339 0.03 0.01 0.00 0.00 0.01 0.00
13 Transportation 1,122 0.03 0.02 0.00 0.00 0.02 0.00
14 Utilities 849 0.01 0.02 0.00 0.00 0.01 0.00
15 Retail stores 1,354 0.03 0.02 0.00 0.00 0.00 0.00
16 Banks, insurers, and other financials 6,767 0.04 0.17 0.02 0.00 0.02 0.01
17 Other 8,133 0.07 0.05 0.02 0.00 0.01 0.01

Total 27,904 0.05 0.07 0.01 0.00 0.01 0.00

The sample covers the period 2007–2014. The fair value measurement level proportions are calculated at the end of the fiscal year.

Lev1 A ¼Level 1 fair value assets/total assets.


Lev2 A ¼Level 2 fair value assets/total assets.
Lev3 A ¼Level 3 fair value assets/total assets.
Lev1 L¼ Level 1 fair value liabilities/total assets.
Lev2 L¼ Level 2 fair value liabilities/total assets.
Lev3 L¼ Level 3 fair value liabilities/total assets.

See Appendix 1 for complete definitions of the variables.

Table 2
Descriptive statistics.

Variable Mean SD p10 p25 p50 p75 p90

FVCI 0.002 0.039  0.004 0.000 0.000 0.001 0.011


CI–FVCI 0.008 0.191  0.134  0.003 0.047 0.078 0.120
Ret 0.150 0.641  0.468  0.188 0.081 0.344 0.723
MTB 2.765 3.937 0.709 1.058 1.690 2.941 5.241
Lev1 0.053 0.125 0.000 0.000 0.000 0.035 0.179
Lev2 0.080 0.162 0.000 0.000 0.002 0.077 0.282
Lev3 0.016 0.088 0.000 0.000 0.000 0.001 0.019
Lev1 A 0.050 0.120 0.000 0.000 0.000 0.031 0.165
Lev2 A 0.069 0.148 0.000 0.000 0.000 0.052 0.247
Lev3 A 0.012 0.075 0.000 0.000 0.000 0.000 0.009
Lev1 L 0.003 0.025 0.000 0.000 0.000 0.000 0.000
Lev2 L 0.011 0.055 0.000 0.000 0.000 0.001 0.009
Lev3 L 0.004 0.030 0.000 0.000 0.000 0.000 0.002

The sample contains 27,904 firm-year observations for the fiscal years 2007–2014. All balance sheet variables are measured at the end of the fiscal year,
except for MTB which is measured at the beginning of the fiscal year.

CI ¼ (net income before extraordinary itemsþfair value components of other comprehensive income)/beginning-of-year market value of equity.
FVCI¼ fair value components of comprehensive income/beginning-of-year market value of equity.
Ret ¼12-month return ending three months after the fiscal year end.
MTB ¼beginning-of-year market to book ratio.
Lev1 ¼(Level 1 fair value assets þ Level 1 fair value liabilities)/total assets.
Lev2 ¼(Level 2 fair value assets þ Level 2 fair value liabilities)/total assets.
Lev3 ¼(Level 3 fair value assets þ Level 3 fair value liabilities)/total assets.
Lev1 A ¼Level 1 fair value assets/total assets.
Lev2 A ¼Level 2 fair value assets/total assets.
Lev3 A ¼Level 3 fair value assets/total assets.
Lev1 L¼ Level 1 fair value liabilities/total assets.
Lev2 L¼ Level 2 fair value liabilities/total assets.
Lev3 L¼ Level 3 fair value liabilities/total assets.

See Appendix 1 for complete definitions of the variables.


82 M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98

Table 3
Pearson correlations.

Panel A: Regression variables

FVCI CI–FVCI Ret MTB Lev1 Lev2 Lev3

CI–FVCI  0.063 1.000


Ret 0.161 0.066 1.000
MTB  0.030 0.056  0.049 1.000
Lev1 0.016  0.031 0.031 0.122 1.000
Lev2 0.124 0.021 0.006  0.041 0.103 1.000
Lev3 0.023 0.017 0.007  0.036 0.024 0.047 1.000
Lev23 0.118 0.026 0.009  0.052 0.100 0.883 0.510

Panel B: Partitioning variables

Institutional ownership Audit quality Earnings management incentives

Institutional ownership 1.000


Audit quality 0.277 1.000
Earnings management incentives 0.015 0.016 1.000

Regression variables (in Panel A):


FVCI¼ fair value components of comprehensive income/beginning-of-year market value of equity.
CI ¼ (net income before extraordinary itemsþ fair value components of other comprehensive income)/beginning-of-year market value of equity.
Ret¼ 12-month return ending three months after the fiscal year end.
MTB ¼beginning-of-year market to book ratio.
Lev1 ¼ (Level 1 fair value assets þLevel 1 fair value liabilities)/total assets.
Lev2 ¼ (Level 2 fair value assets þLevel 2 fair value liabilities)/total assets.
Lev3 ¼ (Level 3 fair value assets þLevel 3 fair value liabilities)/total assets.
Lev23 ¼ Lev 2 þLev 3.

Partitioning variables (in Panel B):


Institutional ownership¼ the proportion of the firms’ shares owned by non-transient (dedicated and quasi-indexer) institutional investors.
Audit quality ¼discrete variable that equals one if the firm has a Big 4 auditor with auditor tenure above the sample median (i.e., high audit quality), minus
one if the firm does not have a Big 4 auditor and the auditor tenure is below the sample median (i.e., low audit quality), and zero otherwise.
Earnings management incentives¼discrete variable that equals one if the firm reports either a change in basic EPS excluding extraordinary items between
zero and two cents or actual EPS exceeding the last analyst forecast consensus before the fiscal year end between zero and two cents (i.e., it is a narrow
beater), minus one if the firm reports both the absolute change in basic EPS excluding extraordinary items greater than five cents and the absolute
difference of actual EPS and the last analyst forecast consensus before the fiscal year end greater than five cents (i.e., it is a clear misser), and zero
otherwise.

See Appendix 1 for complete definitions of the variables. Correlations in boldface are significant at p o0.01 in two-tailed tests.

Table 2 reports descriptive statistics for the uninteracted continuous variables in Eq. (1), breaking out fair value mea-
surements into all three levels separately for assets and liabilities. These statistics are similar to those reported in other
studies. The distributions of fair value measurement levels are negatively skewed because non-zero observations are
concentrated in the upper quartiles, so that the means always exceed the medians.
Table 3, Panel A reports the Pearson correlations of these continuous variables, for brevity not breaking out fair value
measurements separately for assets and liabilities. FVCI is significantly positively associated with Lev2 and Lev3 but in-
significantly associated with Lev1; these differential associations likely reflect the tendency for financial instruments fair
valued using lower-level inputs to have greater credit and liquidity risk and the fact that firms require higher expected
returns to hold riskier items. MTB is significantly positively associated with Lev1 and significantly negatively associated with
Lev2 and Lev3, indicating the importance of controlling for MTB in our study.
Table 3, Panel B reports the Pearson correlations of the three firm-level partitioning variables: institutional ownership
percentage, the audit quality indicator, and the indicator for narrowly meeting or beating earning targets. The correlations
range from low values of 0.015 for institutional ownership and narrowly meeting or beating earnings targets and 0.016 for
audit quality and narrowly meeting or beating earnings targets to a moderate 0.277 for institutional ownership and audit
quality. Hence, the partitioning variables appear to capture largely distinct constructs.

5. Empirical results

5.1. Conditionally conservative Level 2 and 3 fair value measurements

We first test our prediction that Level 2 and 3 fair value measurements exhibit conditional conservatism by estimating
M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98 83

Table 4
Conditional conservatism of fair value measurements by level.

Panel A: Distinguishing Level 1 from Level 2 and 3 fair value measurements

(1) (2)
Exp. sign Dep. var: FVCI Dep.var: CI–FVCI

D 0.001  0.024***
(1.06) (  4.02)
Ret 0.002  0.008
(1.37) (  0.91)
D*Ret ?/þ 0.005 0.154***
(1.63) (7.23)
MTB 0.000 0.003***
(0.93) (3.10)
D*MTB  0.000 0.002*
(  1.23) (1.75)
Ret*MTB  0.000 0.005***
(  1.05) (3.61)
D*Ret*MTB –  0.002***  0.016***
(  3.13) (  4.65)
Lev1 0.004 0.065***
(0.61) (2.82)
D*Lev1  0.010  0.039
(  1.18) (  1.44)
Ret*Lev1 (β10)  0.009  0.050
(  1.22) (  1.57)
D*Ret*Lev1 (β11) ?  0.005  0.096
(  0.22) (  1.08)
Lev23  0.003 0.062**
(  0.46) (2.20)
D*Lev23 0.032***  0.052**
(3.17) (  2.03)
Ret*Lev23 (β14) 0.090***  0.038
(7.08) (  0.92)
D*Ret*Lev23 (β15) þ /? 0.119*** 0.029
(3.75) (0.41)

Significance of β15 4 β11: p-val ¼ 0.002 p-val ¼0.153


Significance of β14 þ β15 4 β10 þ β11: p-val o 0.001 p-val ¼0.224

R-squared 0.306 0.452


Observations 27,904 27,904
Firm and year fixed effects Yes Yes
Cluster by firm Yes Yes

Panel B: Distinguishing Level 1 from Level 2 from Level 3 fair value measurements

(1) (2)
Exp. sign Dep. var: FVCI Dep.var: CI–FVCI

D 0.001  0.024***
(1.06) (  3.93)
Ret 0.002  0.009
(1.37) (  0.94)
D*Ret ?/þ 0.005 0.155***
(1.64) (7.26)
MTB 0.000 0.003***
(0.93) (3.02)
D*MTB  0.000 0.002*
(  1.27) (1.77)
Ret*MTB  0.000 0.005***
(  1.05) (3.66)
D*Ret*MTB –  0.002***  0.016***
(  3.13) (  4.68)
Lev1 0.004 0.063***
(0.72) (2.76)
D*Lev1  0.009  0.033
(  1.11) (  1.23)
Ret*Lev1 (β10)  0.010  0.044
(  1.32) (  1.36)
84 M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98

Table 4 (continued )

D*Ret*Lev1 (β11) ?  0.001  0.092


(  0.06) (  1.03)
Lev2  0.004 0.095***
(  0.48) (3.14)
D*Lev2 0.031***  0.084***
(2.70) (  2.73)
Ret*Lev2 (β14) 0.093***  0.079*
(7.18) (  1.66)
D*Ret*Lev2 (β15) þ /? 0.098*** 0.033
(3.01) (0.42)
Lev3  0.003  0.031
(  0.12) (  0.52)
D*Lev3 0.034 0.044
(1.53) (1.25)
Ret*Lev3 (β18) 0.079* 0.090*
(1.89) (1.88)
D*Ret*Lev3 (β19) þ /? 0.198** 0.023
(1.99) (0.12)

Significance of β19 4β11: p-val ¼0.025 p-val ¼ 0.287


Significance of β15 4β11: p-val ¼0.011 p-val ¼0.168
Significance of β19 4β15: p-val ¼ 0.178 p-val ¼0.480
Significance of β18 þ β19 4 β10 þ β11: p-val ¼0.001 p-val ¼ 0.110
Significance of β14 þ β15 4β10 þ β11: p-val o 0.001 p-val ¼0.227
Significance of β18 þ β19 4 β14 þ β15: p-val ¼0.192 p-val ¼ 0.219

R-squared 0.306 0.453


Observations 27,904 27,904
Firm and year fixed effects Yes Yes
Cluster by firm Yes Yes

The table reports the estimation of Eq. (1), the expanded Basu (1997) reverse regressions that include Lev1 and Lev23 (Panel A) or Lev1, Lev2, and Lev3
(Panel B) interactively to capture the conditional conservatism of these fair value measurement levels.

The dependent variable in column (1) of Panels A and B is FVCI¼ fair value components of comprehensive income/beginning-of-year market value of
equity. The dependent variable in column (2) of panels A and B is CI–FVCI¼ (comprehensive income minus fair value components of comprehensive
income)/beginning-of-year market value of equity. The explanatory variables are:

Ret¼ 12-month return ending three months after the closing of the fiscal year.
D¼ indicator variable equal to 1 if Ret o 0 and 0 otherwise.
MTB ¼beginning-of-year market to book ratio.
Lev1 ¼ (Level 1 fair value assets þLevel 1 fair value liabilities)/total assets.
Lev23 ¼ (Level 2 & 3 fair value assets þLevel 2 & 3 fair value liabilities)/total assets.

See Appendix 1 for complete definitions of the variables. The regressions include intercepts that are not reported for parsimony.

* p o 0.10, ** p o0.05, *** p o 0.01 indicate the significance levels of individual coefficients in two-tailed t tests. t-statistics, reported in parentheses, are
based on robust standard errors calculated clustering observations at the firm level. p-vals indicate the significance levels of differences in (sums of)
coefficients across adjacent columns in one-tailed Wald tests.

Eq. (1), i.e., the expansion of Basu's (1997) model to include interactions of Ret with the fair value measurement level
proportions Lev1 and Lev23, on the full sample. Table 4, Panel A reports the primary estimation results. Column (1) reports
the estimation with FVCI as the dependent variable. Column (2) reports the estimation with CI–FVCI as the dependent
variable, i.e., the placebo test. The bottom of the panel reports the significance levels of one-tailed Wald tests of the dif-
ference of the asymmetric timeliness coefficients for Level 2 and 3 versus Level 1 fair value measurements (i.e., β15 on
D*Ret*Lev23 minus β11 on D*Ret*Lev1), as well as of the difference of the corresponding total bad news timeliness coef-
ficients (i.e., the sum of β14 on Ret*Lev23 and β15 minus the sum of β10 on Ret*Lev1 and β11).
We briefly discuss the coefficients on the variables other than those involving the fair value measurement level pro-
portions, which are the focus of prior research. The coefficient on Ret is insignificant in both the FVCI and CI–FVCI models.17
Hence, neither component of CI captures much contemporaneous good news. Conditional conservatism is strongly evident
in the large and highly significantly positive coefficient on D*Ret in the CI–FVCI model (t-stat ¼7.23). That is, CI–FVCI is more

17
Prior research often finds a negative coefficient on positive returns when net income is the dependent variable; e.g., see Lawrence et al. (2013) and
Patatoukas and Thomas (2011). Givoly and Hayn (2000) and Ryan and Zarowin (2003) explain this result as attributable to the greater presence of younger,
less profitable, and high growth firms in post- Basu (1997) sample years.
M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98 85

sensitive to contemporaneous bad news than to contemporaneous good news, i.e., exhibits asymmetric timeliness. In
contrast, this coefficient is insignificant in the FVCI model. The coefficient on D*Ret*MTB is significantly negative in both the
FVCI and the CI–FVCI models (t-stat¼ 3.13 and  4.65, respectively), but it is one-eighth the absolute magnitude in the
FVCI model than in the CI–FVCI model. These results are consistent with the prior research discussed above which shows
that MTB captures the amount of slack that firms have built up as of the beginning of the period to absorb bad news before
impairment write-downs are required. Collectively, the insignificant coefficient on D*Ret and the significant but small
negative coefficient on D*Ret*MTB in the FVCI model suggests that fair value measurements exhibit at most a low baseline
level of conditional conservatism that is unrelated to Lev23. This low baseline level could be explained by FVCI including the
few OTT impairments of HTM securities, or by Lev23 only partially capturing discretion over lower-level fair value
measurements.
We turn to the asymmetric timeliness and total bad news timeliness coefficients for the fair value measurement level
proportions. As predicted, in the FVCI model the asymmetric timeliness coefficient for Level 1 fair value measurements, β11
on D*Ret*Lev1, is insignificant, consistent with high market discipline forcing firms to record unbiased Level 1 fair value
measurements. Also as predicted, the asymmetric timeliness coefficient for Level 2 and 3 measurements, β15 on D*Re-
t*Lev23, is positive and significant (0.119, t-stat¼3.75) and significantly larger than β11 (p-valo0.002), consistent with
limited market discipline enabling firms to exercise discretion to record conditionally conservative lower-level fair value
measurements. Similarly, the total bad news timeliness coefficient for Level 2 and 3 measurements, β14 on Ret*Lev23 plus
β15, is significantly larger than the total bad news timeliness coefficient for Level 1 measurements, β10 on Ret*Lev1 plus β11
(p-valo0.001). In contrast, in the CI–FVCI model we observe insignificant asymmetric timeliness for both the fair value
measurement level proportions.
Table 4, Panel B reports the estimation of expanded versions of Eq. (1) that decompose Lev23 into Lev2 and Lev3. This
decomposition yields minimal changes in the coefficients on variables that do not involve Lev2 and Lev3 from the results
reported in Panel A. The bottom of Panel B reports the significance of one-tailed Wald tests of the differences of the
asymmetric timeliness and total bad news timeliness coefficients for the three possible pairings of each of these coefficients
across the three fair value measurement level proportions. As predicted, in the FVCI model the asymmetric timeliness
coefficients for Level 2 fair value measurements, β15 on D*Ret*Lev2, and for Level 3 measurements, β19 on D*Ret*Lev3, are
both positive and significant (0.098, t-stat¼3.01 and 0.198, t-stat ¼1.99, respectively). Although β19 is twice as large as β15,
consistent with the expected greater conditional conservatism in Level 3 than in Level 2 measurements, the difference of
these coefficients is insignificant due to the relatively high standard error of β19, which likely reflects the small mean and
variability of the Level 3 proportion. However, β15 and β19 are both significantly larger than the insignificant asymmetric
timeliness coefficient for Level 1 measurements, β11 on D*Ret*Lev1 (p-val¼0.011 and p-val¼0.025, respectively). Similarly,
the total bad news timeliness coefficient for Level 3 measurements, β18 on Ret*Lev3 plus β19, is considerably larger than the
total bad news timeliness coefficient for Level 2 measurements, β14 on Ret*Lev2 plus β15, but insignificantly so, and both of
these coefficients are significantly larger than the insignificant total bad news timeliness coefficient for Level 1 measure-
ment, β10 on Ret*Lev1 plus β11 (p-val¼0.001 and p-val¼0.001, respectively). In contrast, in the CI–FVCI model we observe
insignificant asymmetric timeliness coefficients for all three fair value measurement level proportions.
We conducted three untabulated specification analyses for the models whose estimations are reported in Table 4, Panel
A, all of which yield the same inferences as the tabulated results. The first analysis addresses the fact that the dependent
variable in the model in column (1) of Table 4, Panel A, FVCI, includes the sum of unrealized and realized gains and losses. As
discussed in Section 3, because realized gains and losses yield perfectly offsetting effects on the balance of net unrealized
gains and losses, FVCI is unaffected by whether unrealized gains and losses are realized in a period. Still, the question arises
whether the findings that Level 2 and 3 fair value measurements are conditionally conservative reported in this column are
attributable to the net unrealized gains and losses component of FVCI, the realized gains and losses component of FVCI, or
both. This question arises because firms' choices to realize gains and losses likely exhibit asymmetry due to tax mitigation or
for other reasons, and so realization of gains and losses may affect the estimated conditional conservatism of the gains and
losses that remain unrealized after such realization (i.e. net unrealized gains and losses).
To address this concern, we estimated this model replacing the dependent variable FVCI with two narrower but purer
measures of unrealized gains and losses versus realized gains and losses available on COMPUSTAT: (1) net unrealized gains
and losses on AFS and HTM securities and cash flow hedge derivatives recorded in other comprehensive income, which
should mostly capture unrealized gains and losses, since AFS and HTM securities are sold or OTT impaired relatively in-
frequently; and (2) net realized and unrealized gains and losses on the disposition of securities held for resale, which should
mostly capture realized gains and losses due to the typically short holding period for trading securities. We also estimated
the model for a much smaller sample of 1,829 observations for bank holding companies (BHCs) from 2007–2013, for which
less aggregated bank regulatory Y-9C filing data is available, replacing FVCI with narrower but purer unrealized gains and
losses versus realized gains and losses available on those filings: (1) unrealized gains and losses on AFS securities (the
change in data item BHCK8434) and (2) realized gains and losses on AFS securities (data item BHCK3196). A downside of the
Y-9C analysis is considerably lower statistical power and external validity due to the much smaller and single-industry
sample.
In all four of these estimations, the coefficient β11 on D*Ret*Lev1 is insignificantly different from zero, indicating that
Level 1 measurements are not associated with either the conditional conservatism of unrealized gains and losses or the
asymmetric realization of gains and losses. In the two estimations with net realized and unrealized gains and losses on the
86 M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98

disposition of securities held for resale from COMPUSTAT or realized gains and losses on AFS securities from the Y-9C filings
as the dependent variable, the coefficient on β15 on D*Ret*Lev23 is insignificantly different from zero, indicating that Level
2 and 3 measurements are not associated with asymmetric realization of gains and losses. In the two estimations with net
unrealized gains and losses on AFS and HTM securities and cash flow hedge derivatives recorded in other comprehensive
income from COMPUSTAT or unrealized gains and losses on AFS securities from the Y-9C filings as the dependent variable,
the coefficient on β15 on D*Ret*Lev23 is significantly positive at the 10% level or better (t-stat¼3.28 in the COMPUSTAT
sample and t-stat¼1.86 in the Y-9C sample), consistent with the results for the model with FVCI as the dependent variable
reported in the Table 4, Panel A, column (1).
The second specification analysis is intended to ensure that the inclusion of non-recurring OTT impairments of HTM
securities in FVCI rather than CI-FVCI does not drive the results in the models reported in columns (1) and (2) of Table 4,
Panel A. In this analysis, we again use the less aggregated Y-9C filing data to calculate the following analog to FVCI that
excludes OTT impairments of HTM securities: trading revenue (data item BHCKA220) plus net realized gains on AFS se-
curities (BHCK3196) plus net unrealized gains on AFS securities (the change in BHCK8434) divided by beginning-of-year
market value of equity. We define CI-FVCI accordingly. As in the prior test, the coefficient on β11 on D*Ret*Lev1 is insig-
nificantly different from zero in both the FVCI and CI-FVCI models. The coefficient on β15 on D*Ret*Lev23 is significantly
positive in the FVCI model (t-stat¼3.29) and insignificant in the CI-FVCI model, consistent with the results for the models
reported in the Table 4, Panel A.
The third specification analysis addresses the fact that many firms do not hold appreciable financial instruments. In this
analysis, we restrict the sample to the 12,173 observations with strictly positive values of both Lev1 and Lev23. Compared to
the tabulated analysis, this analysis sacrifices external validity for internal validity. Specifically, each observation in this
subsample potentially exhibits both the predicted conditionally conservative Level 2 and 3 fair value measurements and the
predicted unbiased Level 1 measurements. We estimate both models reported in Table 4, Panel A. As in Table 4, Panel A, we
find that the coefficient on β11 on D*Ret*Lev1 is insignificantly different from zero in both models. We find that the coef-
ficient on β15 on D*Ret*Lev23 is significantly positive in the FVCI model (t-stat¼3.32) and insignificant in the CI-FVCI model.
In summary, the results reported in Table 4 and of untabulated specification analyses indicate that Level 2 and 3 fair
value measurements on average exhibit conditional conservatism, but Level 1 measurements do not. These results are
consistent with limited market discipline over lower-level fair value measurements providing reporting firms with the
incentive to mitigate investors' discounting of these measurements by reporting conditionally conservative measurements.
In the following two subsections, we test whether these on-average effects vary across firms and time based on proxies for
firms’ incentives to report conditionally conservative lower-level fair value measurements.

5.2. Partitioning by firm-level proxies for firms' incentives to report conditionally conservative level 2 and 3 fair value
measurements

We partition the full sample based on three firm-level proxies for firms' incentives to report conditionally conservative
level 2 and 3 fair value measurements. First, more knowledgeable and skilled (hereafter “knowledgeable”) investors should
be better able to evaluate firms' lower-level fair value measurements and thus more likely to discount unreliable mea-
surements. Hence, we expect firms with more knowledgeable investor bases to have greater incentive to mitigate such
discounting and thus to be more likely to report conditionally conservative lower-level fair value measurements. We proxy
for the knowledgeableness of a firm's investor base using the proportion of its equity that is owned by non-transient
(dedicated and quasi-indexer) institutional investors, equity investors that prior research shows actively monitor firms
(Bushee, 2001; Ramalingegowda and Yu, 2012).18 Firms in the lowest (highest) institutional investor quintiles are classified
as having low (high) non-transient institutional ownership (hereafter “institutional ownership”).
Second, verification of financial report information by more independent third parties increases firms' incentives to
provide more reliable information, including lower-level fair value measurements. Given the limited market discipline over
these measurements, however, no amount of verification can render them perfectly reliable. Accordingly, we expect in-
dependent third parties verifying firms' lower-level fair value measurements to be particularly concerned with firms un-
derstating unrealized losses or overstating unrealized gains, and thus to be more amenable to verifying measurements that
exhibit conditional conservatism as acceptable. Based on prior research showing that larger and longer tenured auditors
provide higher quality audits (Basu, 1997; Basu et al., 2001; Watts, 2003; Ruddock et al., 2006; Bell et al., 2015), we proxy for
the independence of third-party verification using an indicator variable that categorizes firms employing Big 4 auditors with
above-median auditor tenure as high audit quality and firms employing non-Big 4 auditors with below-median auditor
tenure as low audit quality.
Third, prior research provides evidence that firms that narrowly meet or beat earnings targets exhibit incentives that
conflict with conditional conservatism (Givoly et al., 2007; Lobo et al., 2008; Park and Lee, 2014). We characterize firms that
narrowly meet or beat earnings targets (hereafter “narrow beaters”) as exhibiting incentives that conflict with conditional

18
We include quasi-indexers in part because prior research shows that these passive institutional investors use financial report information to monitor
firms' performance at low cost and are a key contributor to institutional demand for timely information dissemination (Bushee and Noe, 2000; D’Souza
et al., 2010) and in part because active investors' trading strategies rely on cooperation from these passive investors (Brav et al., 2008; Appel et al., 2014).
M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98 87

Table 5
Partitioning the sample on proxies for the demand and supply of conditional conservatism.

Panel A: Partitioning on institutional ownership, audit quality, and earnings management incentives

(1) (2) (3) (4) (5) (6)


Dep. var: FVCI Exp. High inst. Low inst. High audit Low audit Clear Narrow
sign ownership ownership quality quality missers beaters

D 0.000  0.000 0.003  0.004 0.010  0.001


(0.16) (  0.01) (1.52) (  0.86) (0.91) (  0.44)
Ret  0.002 0.004 0.000 0.001 0.014 0.002
(  0.94) (0.65) (0.02) (0.41) (0.67) (0.83)
D*Ret ? 0.005 0.011 0.007 0.005 0.002  0.007
(0.59) (0.69) (0.97) (0.53) (0.07) (  0.64)
MTB 0.001**  0.000 0.000  0.001 0.001  0.000
(2.43) (  0.57) (0.90) (  0.67) (0.20) (  0.12)
D*MTB  0.001**  0.000  0.001 0.001  0.002  0.000
(  2.11) (  0.38) (  1.25) (1.29) (  0.68) (  0.57)
Ret*MTB  0.001 0.000 0.000 0.001  0.003  0.000
(  0.74) (0.35) (0.04) (1.07) (  0.57) (  0.94)
D*Ret*MTB –  0.005**  0.002  0.003**  0.001  0.004 0.001
(  1.97) (  1.28) (  2.03) (  0.74) (  0.67) (0.45)
Lev1  0.011  0.034 0.014 0.002 0.017 0.007
(  0.51) (  1.17) (1.02) (0.14) (0.35) (0.90)
D*Lev1 0.017 0.004  0.026*  0.013  0.010  0.013
(0.64) (0.15) (  1.73) (  0.55) (  0.09) (  1.54)
Ret*Lev1  0.029 0.005  0.021  0.011 0.088  0.021*
(  0.97) (0.47) (  1.09) (  0.61) (0.61) (  1.94)
D*Ret*Lev1 ? 0.052 0.000  0.008 0.068  0.061 0.001
(0.49) (0.00) (-0.13) (1.46) (-0.18) (0.02)
Lev23  0.007 0.016  0.011  0.025 0.094*  0.000
(  0.46) (0.62) (  0.96) (  0.65) (1.65) (  0.00)
D*Lev23 0.102***  0.054 0.057***  0.013  0.015 0.033
(3.36) (  1.29) (3.91) (  0.32) (  0.23) (1.21)
Ret*Lev23 (β14) 0.160***  0.032 0.126*** 0.040  0.027 0.016
(4.57) (  0.85) (6.62) (0.93) (  0.39) (0.65)
D*Ret*Lev23 (β15) þ 0.342** 0.036 0.147**  0.077 0.252** 0.092
(2.55) (0.40) (2.49) (  1.00) (2.12) (1.04)

High IO – Low IO High AQ – Low AQ Clear miss – Narrow beat

Significance of difference in β15: p-val ¼ 0.029 p-val ¼ 0.011 p-val ¼0.139


Significance of diff. in β14 þ β15: p-val o0.001 p-val o0.001 p-val ¼ 0.168

R-squared 0.467 0.545 0.356 0.415 0.857 0.651


Observations 5,000 5,447 9,088 3,438 3,180 2,802
Firm & year fixed effects Yes Yes Yes Yes Yes Yes
Cluster by firm Yes Yes Yes Yes Yes Yes

Panel B: Combination of pair-wise partitions (only the coefficients of interest)

(1) (2) (3) (4) (5) (6)


High inst. Low inst. High inst. Low inst. High audit Low audit
ownership & ownership & ownership & ownership & quality & quality &
High audit Low audit Clear Narrow Clear Narrow
Dep. var: FVCI Exp. sign. quality quality misser beater misser beater

Ret*Lev1  0.061 0.025  0.022  0.019 0.008 0.345


(  1.16) (0.37) (  0.40) (  0.23) (0.21) (1.04)
D*Ret*Lev1 ?  0.002 0.009 0.083  0.045 0.030  1.130
(  0.02) (0.06) (0.32) (  0.14) (0.26) (  0.98)
Ret*Lev23 (β14) 0.204***  0.024  0.097  0.007 0.083***  0.259
(5.61) (  0.71) (  1.43) (  0.08) (3.06) (  0.82)
D*Ret*Lev23 (β15) þ 0.468**  0.005 0.748*** 0.210 0.240** 0.956
(2.24) (  0.08) (2.70) (0.45) (2.39) (1.17)

High IO & AQ – High IO & Clear misser – High AQ & Clear misser –
Low IO & AQ Low IO & Narrow beater Low AQ & Narrow beater
Significance of difference in β15: p-val ¼0.015 p-val ¼ 0.162 p-val ¼ 0.808
Significance of difference in β14 þβ15: p-val o 0.001 p-val ¼ 0.196 p-val ¼ 0.595
88 M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98

Table 5 (continued )

R-squared 0.545 0.555 0.591 0.905 0.434 0.929


Observations 2,318 1,028 2,393 777 4,365 279
Firm & year fixed Yes Yes Yes Yes Yes Yes
effects
Cluster by firm Yes Yes Yes Yes Yes Yes

Panels A and B of the table report the estimation of the model in the first column of Table 4, Panel A for partitions of the full sample based on three firm-
level variables.

In Panel A, Columns (1) and (2) report the estimations for partitions of the sample based on the proportion of non-transient (dedicated and quasi-indexer)
institutional ownership. A firm is defined as having high (low) non-transient institutional ownership if it is in the top (bottom) quintile of the distribution
of non-transient institutional ownership. The definition of dedicated and quasi-indexer institutional shareholders is based on Bushee and Noe (2000). See
http://acct.wharton.upenn.edu/faculty/bushee/IIclass.html. Columns (3) and (4) report the estimation for partitions of the sample based on audit quality. A
firm is defined as having high audit quality if it has Big 4 auditor with auditor tenure above the sample median. A low audit quality firm has a non-Big
4 auditor with auditing tenure below the sample median. Columns (5) and (6) report the estimation for partitions of the sample based on managerial
incentives to beat earnings targets. Firm-years with either the change in basic EPS excluding extraordinary items between zero and two cents, or actual EPS
exceeding the last analyst forecast consensus before the fiscal year end by from zero to two cents are classified as “narrow beaters.” These firms are
assumed to have had strong incentive to be less conditionally conservative, if necessary, to achieve the earnings target. Firm-years with both the absolute
change in basic EPS excluding extraordinary items greater than five cents, and the absolute difference of actual EPS and the last analyst forecast consensus
before the fiscal year end greater than five cents are classified as “clear missers.”

Panel B contains the estimation of pair-wise partitions based on complementary values of two of the partitioning variables in Panel A. For parsimony, only
the coefficients of interest are reported.

Ret¼ 12-month return ending three months after the closing of the fiscal year.
D¼ indicator variable equal to 1 if Ret o 0 and 0 otherwise.
MTB ¼beginning-of-year market to book ratio.
Lev1 ¼ (Level 1 fair value assets þLevel 1 fair value liabilities)/total assets.
Lev23 ¼ (Level 2 & 3 fair value assets þLevel 2 & 3 fair value liabilities)/total assets.

See Appendix 1 for complete definitions of the variables. The regressions include intercepts that are not reported for parsimony.

* p o 0.10, ** p o0.05, *** p o 0.01 indicate the significance levels of individual coefficients in two-tailed t tests. t-statistics, reported in parentheses, are
based on robust standard errors calculated clustering observations at the firm level. p-vals indicate the significance levels of differences in (sums of)
coefficients across adjacent columns in one-tailed Wald tests.

conservatism, and firms that clearly miss these targets (hereafter “clear missers”) as exhibiting incentives that do not
conflict with conditional conservatism. We define narrow beaters and clear missers using thresholds employed in prior
research (Roychowdhury, 2006; Zang, 2012). Narrow beaters are firm-years with either a change in basic EPS excluding
extraordinary items between zero and two cents or actual EPS exceeding the last analyst forecast consensus before the fiscal
year end between zero and two cents. Clear missers are firm-years with both of these differences being more than five cents
in absolute value.
Table 5, Panel A reports the results of estimating Eq. (1) with FVCI as the dependent variable (i.e., the model reported in
column (1) of Table 4, Panel A) for partitions of the full sample based on each of these three proxies. Columns (1) and (2)
[(3) and (4)] {(5) and (6)} report the results for the high versus low institutional ownership [high versus low audit quality]
{clear missers versus narrow beaters} subsamples, respectively. The bottom of the panel reports the significance of one-
tailed Wald tests of the difference of the asymmetric timeliness coefficient for Level 2 and 3 fair value measurements, β15,
and of the difference of the total bad news timeliness coefficient for those measurements, β14 þ β15, across the two sub-
samples formed by each partitioning variable.
In addition, to provide insight into the joint effects of the partitioning variables, Table 5, Panel B reports the estimated
coefficients of interest and the same Wald tests just described for partitions of the full sample based on the values of two
proxies with the same expected directional effect on the conditional conservatism of firms' lower-level fair value mea-
surements (hereafter “complementary” values of the proxies). High ability of investors to evaluate lower-level fair value
measurements, high third-party verification of these measurements, and low firm incentives that conflict with conditional
conservatism are complementary values of the proxies; similarly, low investor ability, low third-party verification, and high
firm incentives that conflict with conditional conservatism are also complementary. Columns (1) and (2) [(3) and (4)]
{(5) and (6)} report the results for the high institutional ownership and high audit quality subsample versus the low in-
stitutional ownership and low audit quality subsample [the high institutional ownership and clear missers subsample
versus the low institutional ownership and narrow beaters subsample] {the high audit quality and clear missers subsample
versus the low audit quality and narrow beaters subsample}, respectively.19

19
Because of the reduced number of observations per partition that would result from two-way partitioning, we do not attempt to test for the
incremental significance of each of the three firm-level partitioning variables controlling for the other two variables. Some sense for the incremental effects
of the partitioning variables can be gleaned from comparison of the estimated coefficients for the one-way partitions in Table 5, Panel A and the two-way
complementary partitions in Table 5, Panel B. The reader should be careful in making inferences about the significance of these effects, however.
M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98 89

As with the Table 4 results discussed above, for all partitions in both panels of Table 5 we find no evidence of conditional
conservatism in Level 1 fair value measurements. Hence, for brevity we discuss the asymmetric timeliness and total bad
news timeliness coefficients only for Level 2 and 3 measurements.
We first discuss the results partitioning by one proxy at a time reported in Table 5, Panel A. As expected, for the high
institutional ownership subsample in Column (1) the asymmetric timeliness coefficient for Level 2 and 3 fair value mea-
surements, β15, is positive and significant (0.342, t-stat¼2.55), whereas β15 is insignificant for the low institutional own-
ership subsample in Column (2). The Wald test of the positive difference of β15 across the two subsamples is significant (p-
val¼0.029). The Wald test of the positive difference of the total bad news timeliness coefficient for Level 2 and 3 mea-
surements, β14 þ β15, across the subsamples is also significant (p-valo0.001).
Also as expected, for the high audit quality subsample in Column (3) the asymmetric timeliness coefficient for Level
2 and 3 fair value measurements, β15, is positive and significant (0.147, t-stat¼2.49), whereas for the low audit quality
subsample in Column (2) this coefficient is insignificant. The Wald test of the positive difference of β15 across the two
subsamples is significant (p-val¼0.011). Similarly, the Wald test of the positive difference of the total bad news timeliness
coefficient for Level 2 and 3 measurements, β14 þ β15, across the subsamples is significant (p-valo0.001).
Again as expected, for the clear missers subsample in Column (5) the asymmetric timeliness coefficient for Level 2 and
3 fair value measurements, β15, is positive and significant (0.252, t-stat¼2.12), whereas for the narrow beaters subsample in
Column (6) β15 is positive and of reasonable magnitude but insignificant. The latter result may reflect an appreciable portion
of narrow beaters not having incentives that conflict with conditional conservatism. For this reason, the Wald test of the
positive difference of β15 across the two subsamples is insignificant, as is the Wald test of the positive difference of the total
bad news timeliness coefficient for Level 2 and 3 measurements, β14 þ β15.
We turn to the results partitioning by complementary values of two proxies reported in Table 5, Panel B. Consistent with
the results in Table 5, Panel A, this panel reports significant positive asymmetric timeliness coefficients for Level 2 and 3 fair
value measurements, β15, for the subsamples in columns (1), (3), and (5) for which we expect more conditionally con-
servative lower-level fair value measurements, and insignificant β15 in the other three columns. For the high institutional
ownership and high audit quality subsample versus the low institutional ownership and low audit quality subsample re-
ported in columns (1) and (2), the Wald tests of the positive differences in β15 and in the total bad news timeliness coef-
ficients for Level 2 and 3 measurements, β14 þ β15, are significant (p-val¼0.015 and p-valo0.001, respectively). The Wald
tests for the differences of these coefficients for the other subsample pairings in the remaining columns are insignificant,
however. This appears to be attributable to the large positive but insignificant β15 coefficients (as in Table 5, Panel A) and
relatively few observations (777 in column (4) and 279 in column (6)) in the subsamples involving narrow beaters.
In summary, aside from the insignificant Wald tests of differences in the coefficients of interest across the intersections of
the complementary subsamples that involve narrow beaters, the results reported in Table 5 provide strong evidence that the
conditional conservatism of Level 2 and 3 fair value measurements increases with two firm-level proxies for firms' in-
centives to report conditionally conservative measurements (institutional ownership and audit quality) and decreases with
a firm-level proxy for firms' earnings management incentives that conflict with conditional conservatism (narrower beaters
versus clear missers).

5.3. Partitioning pre- versus post-SFAS 157

By requiring firms to disclose their fair value measurements at each level for each major category of asset and liability
each period, SFAS 157 significantly improves the ability of investors to evaluate firms' fair value measurements. This re-
quirement should increase investors' discounting of measurements perceived to be unreliable. Hence, we expect the re-
quirement to increase firms’ incentive to report conditionally conservative measurements.
To test this expectation, we partition the full sample into the year before regular adopters' adoption of SFAS 157, 2007,
versus the subsequent year, 2008. To ensure a clean pre-SFAS 157 subsample, we drop the 38 early adopters of the standard.
Because SFAS 157 did not require firms to make disclosures for years prior to adoption, to conduct the 2007 analysis we
must make some assumption about the fair value measurement level proportions in that year. Similar to Riedl and Serafeim
(2011), we assume firms' 2007 proportions equal their disclosed proportions for 2008.20 While the single year 2007 con-
stitutes a relatively small sample for the pre-SFAS 157 analysis, we do not include 2006 or prior years in the pre-SFAS 157
sample, because we expect firms' fair value measurement level proportions to vary considerably more than usual from the
non-crisis year 2006 to the crisis year 2007. For comparability, we also include only the single year 2008 for the post-SFAS
157 analysis.
Columns (1) and (2) of Table 6, Panel A report the estimation of Eq. (1) with FVCI as the dependent variable for the 2007
and 2008 subsamples, respectively. The bottom of panel reports the significance of one-tailed Wald tests of the difference of
the asymmetric timeliness coefficient for Level 2 and 3 fair value measurements, β15, and of the difference of the total bad
news timeliness coefficient for those measurements, β14 þ β15, across the two subsamples. Unexpectedly, we find a small but
significantly negative asymmetric timeliness coefficient β11 for Level 1 fair value measurements in the 2007 subsample

20
This is reasonable assumption because the proportions of fair value measurements in each level vary slowly over our 2008–2014 sample period; the
first-order autocorrelation of Levels 1, 2, and 3 recognized fair value measurements are 0.77, 0.89, and 0.78, respectively.
90 M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98

Table 6
Partitioning the sample pre- and post SFAS 157.

Panel A: Year before (2007) and after (2008) SFAS 157 Implementation

(1) (2)
Exp. Pre-SFAS 157 Post-SFAS 157
Dep. var: FVCI sign (2007) (2008)

D 0.002*** 0.002
(3.01) (1.12)
Ret 0.001  0.001
(1.64) (  0.44)
D*Ret ? 0.005*** 0.012**
(2.66) (2.34)
MTB 0.000 0.000**
(0.39) (1.97)
D*MTB  0.000**  0.001*
(  2.21) (  1.81)
Ret*MTB  0.000  0.000
(  0.09) (  0.39)
D*Ret*MTB –  0.001***  0.002***
(  2.97) (  2.59)
Lev1  0.003 0.015*
(  0.74) (1.76)
D*Lev1  0.002  0.015
(  0.22) (  1.08)
Ret*Lev1 (β10) 0.014  0.010
(1.40) (  0.75)
D*Ret*Lev1 (β11) ?  0.036** 0.005
(  2.14) (0.14)
Lev23 0.009**  0.034
(2.31) (  1.18)
D*Lev23  0.006 0.041
(-0.84) (1.24)
Ret*Lev23 (β14)  0.020** 0.038
(  2.11) (0.81)
D*Ret*Lev23 (β15) þ 0.059*** 0.190***
(3.08) (2.89)

Post 157 – Pre 157

Significance of difference in β15: p-val o 0.015


Significance of difference in β14 þ β15: p-val o 0.001
R-squared 0.019 0.166
Observations 3,705 4,094
Cluster by firm Yes Yes

Panel B: Effect of SFAS 157 and Non-Transient Institutional Ownership (only the coefficients of interest)

Pre-SFAS 157 (2007) & Post-SFAS 157 (2008) &

(1) (2) (3) (4)


Dep. var: FVCI Exp. High inst. Low inst. High inst. Low inst.
sign ownership ownership ownership ownership

Ret*Lev1  0.002  0.013 0.000  0.001


(  0.15) (  0.62) (0.01) (  0.06)
D*Ret*Lev1 ?  0.024  0.038  0.034 0.100
(  0.71) (  0.91) (  0.49) (0.48)
Ret*Lev23 (β14) 0.002  0.006  0.068  0.030
(0.18) (  0.36) (  1.12) (  1.51)
D*Ret*Lev23 (β15) þ 0.009 0.055 0.424*** 0.107
(0.38) (1.62) (3.37) (0.95)

(High – Low)PRE (High – Low)POST


Sig. of diff. in β15: col (1) – col (2) or col (3) – col (4) p-val ¼0.137 p-val ¼0.031
Sig. of diff in β14 þβ15: col (1) – col (2) or col (3) – col (4) p-val ¼ 0.145 p-val ¼ 0.036
HighPOST – HighPRE
Sig. of diff. in β15: col (3) – col (1) p-val o0.001
Sig. of diff in β14 þ β15: col (3) – col (1) p-val o0.001
LowPOST – LowPRE
Sig. of diff. in β15: col (4) – col (2) p-val ¼ 0.328
Sig. of diff in β14 þ β15: col (4) – col (2) p-val ¼ 0.401
M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98 91

Table 6 (continued )

(High – Low)POST – (High – Low)PRE


Sig. of diff in diff in β15: [col (3) – col (4)] – [col (1) – col (2)] p-val ¼ 0.014
Sig. of diff in diff in β14 þβ15: [col (3) – col (4)] – [col (1) – col (2)] p-val ¼ 0.019

R-squared 0.050 0.030 0.303 0.080


Observations 733 705 736 704
Cluster by firm Yes Yes Yes Yes

Panel C: Effect of SFAS 157 and Audit Quality (only the coefficients of interest)

Pre-SFAS 157 (2007) & Post-SFAS 157 (2008) &

(1) (2) (3) (4)


Dep. var: FVCI Exp. sign High audit quality Low audit quality High audit quality Low audit quality

Ret*Lev1 0.005 0.026 0.008 0.156


(0.28) (1.36) (0.63) (1.25)
D*Ret*Lev1 ?  0.010  0.032 0.034  0.136
(  0.32) (  0.83) (0.39) (  1.03)
Ret*Lev23 (β14)  0.027*  0.031  0.009  0.022
(  1.70) (  1.03) (  0.26) (  0.08)
D*Ret*Lev23 (β15) þ 0.060** 0.064 0.219*** 0.204
(2.12) (1.05) (2.66) (0.72)

(High – Low)PRE (High – Low)POST


Sig. of diff. in β15: col (1) – col (2) or col (3) – col (4) p-val ¼ 0.482 p-val ¼0.268
Sig. of diff in β14 þβ15: col (1) – col (2) or col (3) –col (4) p-val ¼ 0.493 p-val ¼ 0.162

HighPOST – HighPRE
Sig. of diff. in β15: col (3) – col (1) p-val ¼0.026
Sig. of diff in β14 þβ15: col (3) – col (1) p-val ¼0.007

LowPOST – LowPRE
Sig. of diff. in β15: col (4) – col (2) p-val ¼0.136
Sig. of diff in β14 þβ15: col (4) – col (2) p-val ¼0.402

(High – Low)POST – (High – Low)PRE


Sig. of diff in diff in β15: [col (3) – col (4)] – [col (1) – col (2)] p-val ¼0.270
Sig. of diff in diff in β14 þ β15: [col (3) – col (4)] – [col (1) – col (2)] p-val ¼ 0.166

R- squared 0.068 0.026 0.275 0.097


Observations 1353 692 1342 823
Cluster by firm Yes Yes Yes Yes

Panel D: Effect of SFAS 157 and Managerial Incentives to Beat Earnings Targets (only the coefficients of interest)

Pre-SFAS 157 (2007) & Post-SFAS 157 (2008) &

(1) (2) (3) (4)


Dep. var: FVCI Exp. sign Clear missers Narrow beaters Clear missers Narrow beaters

Ret*Lev1  0.069 0.028  0.092*** 0.021


(  0.92) (0.86) (  3.68) (0.74)
D*Ret*Lev1 ? 0.085  0.017 0.074  0.086
(0.94) (  0.46) (0.63) (  1.63)
Ret*Lev23 (β14)  0.001  0.081  0.908***  0.011
(  0.02) (  0.59) (  5.14) (  0.65)
D*Ret*Lev23 (β15) þ 0.026 0.026 1.145*** 0.251*
(0.31) (0.17) (5.94) (1.77)

(Clear misser – Narrow beater)PRE (Clear misser – Narrow beater)POST

Sig. of diff. in β15: col (1) – col (2) or col (3) – col (4) p-val ¼0.499 p-val o 0.001
Sig. of diff in β14 þβ15: col (1) – col (2) or col (3) – col (4) p-val ¼ 0.161 p-val ¼0.491

Clear misserPOST – Clear misserPRE


Sig. of diff. in β15: col (3) – col (1) p-val o0.001
Sig. of diff. in β14 þβ15: col (3) – col (1) p-val ¼ 0.030
92 M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98

Table 6 (continued )

Narrow beaterPOST – Narrow beaterPRE


Sig. of diff. in β15: col (4) – col (2) p-val o 0.001
Sig. of diff. in β14 þβ15: col (4) – col (2) p-val o 0.001

(Clear misser – Narrow beater)POST – (Clear misser – Narrow beater)PRE

Sig. of diff in diff in β15: [col (3) – col (4)] – [col (1) – col (2)] p-val ¼0.001
Sig. of diff in diff in β14 þβ15: [col (3) – col (4)] – [col (1) – col (2)] p-val ¼0.317

R-squared 0.026 0.0.065 0.349 0.179


Observations 414 411 857 337
Cluster by firm Yes Yes Yes Yes

The table reports estimations of the model in the first column of Table 4, Panel A for regular adopters in the 2007 versus 2008 years of the full sample
period, i.e., immediately prior to and after the effective date of SFAS 157 (Panel A) or for those years further partitioned on the three firm-level variables
examined in Table 5, Panel A (Panels B-D). Panel B reports the estimations for 2007 versus 2008 with each year further partitioned based on the proportion
of non-transient (dedicated and quasi-indexer) institutional ownership. Panel C reports the estimations for 2007 versus 2008 with each year further
partitioned based on audit quality. Panel D reports the estimations on managerial incentives to beat earnings targets. Firm-years with either the change in
basic EPS excluding extraordinary items between zero to two cents, or the difference of actual EPS and the last analyst forecast consensus before the fiscal
year end between zero to two cents are classified as “narrow beaters.” Firm-years with both the absolute values of both the change in basic EPS excluding
extraordinary items greater than five cents, and the difference of actual EPS and the last analyst forecast consensus before the fiscal year end greater than
five cents are classified as “clear missers.” See the notes to Table 5 for more complete descriptions of the three firm-level partitioning variables.

The dependent variable is FVCI ¼(fair value components of comprehensive income)/beginning-of-year market value of equity. The explanatory variables
are:

Ret¼ 12-month return ending three months after the closing of the fiscal year.
D¼ indicator variable equal to 1 if Ret o 0 and 0 otherwise.
MTB ¼beginning-of-year market to book ratio.
Lev1 ¼ (Level 1 fair value assets þLevel 1 fair value liabilities)/total assets.
Lev23 ¼ (Level 2 and 3 fair value assets þLevel 2 and 3 fair value liabilities)/total assets.

See Appendix 1 for complete definitions of the variables. Firms' fair value measurements proportions in 2007 are assumed to equal the actual proportions
in 2008.

* p o 0.10, ** p o 0.05, *** po 0.01indicate the significance levels of individual coefficients in two-tailed t tests. t-statistics, reported in parentheses, are
based on robust standard errors calculated clustering observations at the firm level. p-vals indicate the significance levels of (differences in) differences of
the indicated coefficients in one-tailed Wald tests.

(  0.036, t-stat¼  2.14); we have no explanation for this coefficient, although it may result from using the 2008 Level
1 proportion to proxy for the 2007 proportion. As expected, β15 is positive and significant for the 2008 subsample (0.190, t-
stat ¼2.89). β15 is also positive and significant for the 2007 subsample (0.059, t-stat¼3.08), but less than one-third its
magnitude in the 2008 subsample. As expected, the positive difference of this coefficient across the 2008 versus 2007
subsamples is significant (p-val¼0.015), as is the positive difference of β14 þ β15 across these subsamples (p-valo0.001).
In summary, the results reported in Table 6, Panel A provide strong support for our prediction that the conditional
conservatism of Level 2 and 3 fair value measurements increased under SFAS 157's expanded disclosure requirements for
these measurements.

5.4. Partitioning by the firm-level proxies and pre- versus post-SFAS 157

We also expect SFAS 157's expanded disclosure requirements to increase the effects of the three firm-level proxies on the
conditional conservatism of lower-level fair value measurements. Specifically, these disclosures should: enable more
knowledgeable investors to better evaluate firms' lower-level fair value measurements; make independent third-party
auditors more aware of and thus concerned about firms' discretion over these measurements; and make firms' exercise of
discretion over these measurements more apparent. Although the last effect may reduce the number of firms that fall in the
narrow beater subsample due to their exercise of discretion over fair value measurements, in this case it should also sharpen
the distinction between the remaining narrow beaters and clear missers.
Table 6, Panels B-D report the estimated coefficients of interest in Eq. (1) for subsamples formed by further partitioning
the subsamples based on the three firm-level proxies into 2007 versus 2008. The bottoms of these panels report the sig-
nificance of three sets of one-tailed Wald tests of (differences in) differences of the asymmetric timeliness coefficient for
Level 2 and 3 fair value measurements, β15, and of (differences in) differences of the total bad news timeliness coefficient for
these measurements, β14 þ β15, across (pairs of) subsamples. The first set tests the difference of β15 and the difference of
β14 þ β15 across the two subsamples formed based on each of the firm-level proxies (e.g., high versus low institutional
ownership) in each of 2007 and 2008; these tests reveal whether each firm-level proxy has effects in each year. The second
M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98 93

set tests the difference of β15 and the difference of β14 þ β15 for subsamples formed based on a given value of the each of
three firm-level proxies (e.g., high institutional ownership) across 2007 versus 2008; these tests reveal whether SFAS 157
influences the effects of each value of each firm-level proxy. The third set tests the difference in the difference of β15 and the
difference in the difference of β14 þ β15 across the two subsamples formed based on each firm-level proxy (e.g., high versus
low institutional ownership) across 2007 versus 2008; these tests reveal whether SFAS 157 influences the effects of each
firm-level proxy. Because the sample sizes are considerably reduced from prior analyses, the significance levels of the
coefficients and Wald tests, particularly the demanding difference-in-difference tests, are reduced.
Table 6, Panel B reports the results for the high versus low institutional ownership subsamples in 2007 (columns (1) and
(2), respectively) and 2008 (columns (3) and (4), respectively). In 2007, the asymmetric timeliness coefficient for Level 2 and
3 fair value measurements, β15, is insignificant for both institutional ownership subsamples, as are the Wald tests across
these subsamples. In contrast, in 2008 β15 is significantly positive for the high institutional ownership subsample (0.424, t-
stat ¼3.37) and insignificant for the low subsample. The Wald tests of the positive differences in β15 and in the total bad
news coefficient for Level 2 and 3 fair value measurements, β14 þ β15, across the high versus low institutional ownership
subsamples in 2008 are both significant (p-val ¼0.031 and p-val¼0.036, respectively), consistent with the conditional
conservatism of lower-level fair value measurements increasing with institutional ownership in the post-SFAS 157 period.
The Wald tests of the positive differences in these coefficients for the high institutional ownership subsample across 2008
versus 2007 are also both significant (p-val o0.001 and p-valo0.001, respectively), consistent with the effect of high in-
stitutional ownership increasing as a result of SFAS 157. In contrast, the Wald tests of the differences in these coefficients for
the low institutional ownership subsample across 2008 versus 2007 are both insignificant. The Wald tests of the positive
differences in the differences of β15 and of β14 þ β15 across the high versus low institutional ownership subsamples across
2008 versus 2007 are both significant (p-val¼0.014 and p-val ¼0.019, respectively), consistent with the differential effect of
high versus low institutional ownership increasing as a result of SFAS 157.
Table 6, Panel C reports the results for the high versus low audit quality subsamples in 2007 (columns (1) and (2),
respectively) and 2008 (columns (3) and (4), respectively). In 2007, the asymmetric timeliness coefficient for Level 2 and
3 fair value measurements, β15, is significantly positive (although fairly small) for the high audit quality subsample (0.060, t-
stat ¼2.12) and insignificant in the low audit quality subsample. The Wald tests of the differences in β15 and in the total bad
news coefficient for Level 2 and 3 fair value measurements, β14 þ β15, across these subsamples are both insignificant. In
contrast, in 2008 β15 is significantly positive (and much larger) for the high audit quality subsample (0.219, t-stat ¼2.66) and
insignificant for the low subsample. The Wald tests across these subsamples are insignificant. The Wald tests of the positive
differences in these coefficients for the high audit quality subsample across 2008 versus 2007 are both significant (p-
val¼0.026 and p-val¼0.007, respectively), consistent with the effect of high audit quality increasing as a result of SFAS 157.
In contrast, the Wald tests of the differences in these coefficients for the low audit quality subsample across 2008 versus
2007 are insignificant. The Wald tests of the differences in the differences these coefficients across the high versus low audit
quality subsamples across 2008 versus 2007 are both insignificant.
Table 6, Panel D reports the results for the clear missers versus narrow beaters subsamples in 2007 (columns (1) and (2),
respectively) and 2008 (columns (3) and (4), respectively). In 2007, the asymmetric timeliness coefficient for Level 2 and
3 fair value measurements, β15, is insignificant for both earnings management incentives subsamples, as are the Wald tests
across these subsamples. In contrast, in 2008 β15 is significantly positive for the clear missers subsample (1.145, t-stat ¼5.94)
and significantly positive (but much smaller) for the narrow beaters subsample (0.251, t-stat¼1.77). The Wald tests of the
positive differences in β15 and in the total bad news coefficient for Level 2 and 3 fair value measurements, β14 þ β15, across
these subsamples in 2008 are significant and insignificant (p-val o0.001 and p-val¼0.491, respectively). The significant
difference of β15 is consistent with the conditional conservatism of lower-level fair value measurements being higher for
clear missers than for narrow beaters in 2008; the insignificant difference of β14 þ β15 is attributable to an anomalous
significantly negative (and large) good news response coefficient β14 for the clear missers in 2008. The Wald tests of the
positive differences in these coefficients for the clear missers subsample across 2008 versus 2007 are both significant (p-
valo0.001 and p-val¼0.030, respectively), as are the tests of the positive differences in these coefficients for the narrow
beaters subsample across these years (p-valo0.001 and p-valo0.001, respectively). The Wald test of the positive difference
in the difference of β15 across the high versus low institutional ownership subsamples across 2008 versus 2007 is significant
(p-valo0.001), although the test of the corresponding difference in the difference of β14 þ β15 is insignificant, providing
some evidence that the differential effect of low versus high earnings management incentives that conflict with conditional
conservatism increased as a result of SFAS 157.
In summary, the results for the two-way partitions of the full sample reported in Table 6, Panels B-D provide: (1) strong
support for our prediction that the effect of institutional ownership on the conditional conservatism of Level 2 and 3 fair
value measurements increased as a consequence of SFAS 157; (2) weak (i.e., directionally correct but mostly insignificant)
evidence that the corresponding effect of audit quality increased as a consequence of the standard; and (3) some (i.e.,
directionally correct but significant only for the asymmetric timeliness coefficient, not the total bad news coefficient)
evidence that the corresponding effect of incentives to meet-or-beat earnings targets that conflict with conditional con-
servatism increased as a consequence of the standard.
94 M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98

6. Extension to non-financial assets: oil and gas reserves

In Section 5, we examine firms' exercise of discretion over their fair value estimates in the particular context of re-
cognized fair values for financial instruments. This context enables powerful tests, because financial instruments are the
most commonly fair valued type of balance sheet item and SFAS 157 requires fair-value-input-level disclosures that enable
users of financial reports to assess firms’ discretion over their fair value estimates. In this section, we explore the external
validity of these findings to the informationally sparser setting of the disclosed fair values of Canadian and U.S. firms’ oil and
gas (O&G) reserves, a type of non-financial asset. Because the disclosure and related regulatory requirements differ in two
countries, we analyze Canadian and U.S. O&G firms separately.

6.1. Canadian O&G firms

Canadian O&G firms are required to recognize their reserves using the full (historical) cost method,21 and to disclose
estimates akin to the fair values of their proved reserves, which are estimated to be recoverable with at least 90 percent
probability, and their proved plus probable reserves, which are estimated to be recoverable with at least 50 percent
probability. The fair values of these two types of reserves are the discounted present values, at a 10% rate, of the estimated
future revenues from the reserves based on either constant or forecast O&G prices, minus royalties, operating, development,
and well-abandonment costs, after taxes. Because O&G prices are volatile and thus difficult to predict, O&G price forecasts,
when used, are the critical driver of fair values.22 To the extent that measurements of the fair value of reserves mainly are
based on observable O&G prices, they may be closer to Level 2 than Level 3 measurements, but they certainly are not Level
1 measurements. These measures are independently audited.
We hand collect Canadian O&G firms' earnings, book value of reserves, and disclosed fair values for proved reserves and
for proved plus probable reserves for the years 2004–2006 from their public annual information forms, annual reports, and
Forms 51-101F1 to F3 filed in the System for Electronic Document Analysis and Retrieval (SEDAR). We examine three (pro
forma) measures of earnings: net income before extraordinary items (NI), which incorporates changes in O&G reserves at
book value; ΔNAV 1P, which equals NI plus the difference between changes in O&G proved reserves at fair value and
changes in O&G reserves at book value; and ΔNAV 2P, which equals NI plus the difference between changes in O&G proved
plus probable reserves at fair value and the changes in O&G reserves at book value. Moving from NI to ΔNAV 1P and then to
ΔNAV 2P effectively trades a more observable valuation for a more complete valuation, akin to incorporating fair value
rather than cost-based measurements into earnings. We use each of these earnings measures of as numerator of the de-
pendent variable in equation (2). We obtain stock market data from Datastream, Bloomberg, and TSX Venture Summary
Trading Files and O&G prices and exchange rates from Bloomberg. Requiring all variables to be available yields a sample of
378 observations.
Table 7, Panel A, columns (1), (2), and (3) report the estimation of equation (2) with NI, ΔNAV 1P, and ΔNAV 2P as the
dependent variable, respectively, for this sample. The panel also reports the significance of Wald tests of differences in the
coefficients on the asymmetric timeliness coefficient β3 on D*Ret across the three columns. As expected, β3 increases from
0.231 (t-stat¼ 3.75) for the full cost-based reserves measurement in NI reported in Column (1) to 0.839 (t-stat¼2.14) in-
cluding the fair value measurement of proved reserves in ΔNAV 1P reported in Column (2), with the latter coefficient being
significantly larger than the former (p-val o 0.036). β3 increases further to 1.136 (t-stat ¼2.72) including the more complete
fair value measurement of proved plus probable reserves in ΔNAV 2P reported in Column (3), with this coefficient being
significantly larger than the coefficients in the first and second columns (p-val o 0.016 and 0.019, respectively).23
As discussed in Section 3, certain accounting researchers have criticized the Basu (1997) model's use of returns as the
measure of news. Ryan (2006) emphasizes that researchers could, but largely have not, address this criticism by exploiting
industry context to develop measures of news other than returns. In untabulated robustness test we follow this suggestion
and use the fitted values from a firm-level regression of annual firm share returns on annual O&G price index returns (WTI
[West Texas Intermediate] and HH [Henry Hub]) and firm-level reserve changes (discoveries and technical revisions) as the
proxy for news instead of returns. The correlation of Ret with this alternative proxy for news is 0.57. The results are broadly

21
Under the full cost method, essentially all exploration costs, even for unsuccessful wells, are capitalized. The idea is that all exploration costs are
necessary to locate and develop the company's O&G reserves.
22
Gary Finnis, a partner in Sproule, one of the main evaluator companies, explained to us that O&G prices, particularly short-term prices, are the single
biggest source of error in the valuation of reserves. The futures market provides more efficient forecasts of prices further in the future. He also explained
that Sproule uses various O&G prices when forecasting O&G revenues. For the first three years, they use prices derived from futures for the WTI (West
Texas Intermediate), the main international oil index. For subsequent years, they forecast WTI prices based on their assumptions about supply and demand,
as well as industry M&A activity, escalated into the future at a specific price inflation rate. For example, as of August 2011 Sproule forecast a price of $90 US/
BBL escalated at 2%/year. Sproule forecasts the prices of non-WTI oil (i.e., product with different qualities or local availability) using historical average
percentages of the WTI price.
23
Notably, when NI is the dependent variable, the coefficient on Ret is significantly negative. This may be due to the limited role of current earnings in
the evaluation of firms in this industry. JP Morgan Analyst Report (April 17, 2008) states that “EPS and CFPS growth do not tell the whole story as 1) they do
not reflect long-term capital efficiency, 2) they are strongly dependent on commodity prices, which makes us reluctant to use it as a primary metric of
success and 3) they do not take into account differences in timing of growth projects. In an industry with long lead times on projects, we think a focus on
near-term EPS growth might be detrimental to investment decisions and thus to longer-term growth”.
M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98 95

Table 7
Alternative expanded Basu (1997) regressions: Oil & Gas industry measurement of reserves.

Panel A: Canadian Oil & Gas industry

Dependent variable:

NI ΔNAV (1P) ΔNAV (2P)


Exp. Sign (1) (2) (3)

Coef. t-stat Coef. t-stat Coef. t-stat

D  2.592  0.82 13.571*** 12.84 12.952 0.55


Ret  0.014**  2.35 0.094* 1.71 0.097 0.64
D*Ret (β3) þ 0.231*** 3.75 0.839** 2.14 1.136*** 2.72
MTB  0.481  1.07  3.357  1.32  6.306  0.92
D*MTB 0.314 0.37  4.114***  2.95  3.794  0.45
Ret*MTB 0.005*** 4.32  0.015  0.44  0.005  0.08
D*Ret*MTB –  0.048***  6.05  0.071  0.86  0.132  1.07

Significance of NAV 2P β3 4NAV 1P β3: p-val ¼ 0.019


Significance of NAV 1P β3 4NI β3: p-val ¼0.036
Significance of NAV 2P β3 4NI β3: p-val ¼ 0.016

R-squared 0.15 0.11 0.09


Observations 378 378 378

Panel B: US Oil & Gas Industry

Dependent variable:

PRE-SEC O&G Regulation POST-SEC O&G Regulation

NI ΔNAV (1 P) NI ΔNAV (1 P)
Exp. Sign (1) (2) (3) (4)

Coef. t-stat Coef. t-stat Coef. t-stat Coef. t-stat

D  1.301  1.16  53.010  1.37 5.446 0.63 33.225*** 7.09


Ret  0.017  0.88  0.105  0.57  0.087**  2.04 0.405*** 15.06
D*Ret (β3) þ 0.302*** 5.33  0.750  0.56 0.321** 2.22 1.101** 3.76
MTB  1.618*** -2.85  7.420  1.46 0.427 0.50 6.257* 1.68
D*MTB 0.601 0.75 6.304 0.39  1.358  0.48  15.570***  5.73
Ret*MTB 0.019*** 2.61 0.096 1.41 0.026 1.12  0.201***  39.81
D*Ret*MTB –  0.064**  2.42  0.015  0.015  0.031  0.66  0.129**  2.04

Significance of NI-PRE β3 4NI-POST β3: p-val ¼0.313


Significance of NAV 1P-PRE β3 4NAV 1P-POST β3: p-val ¼ 0.013

R-squared 20% 7% 20% 45%


Observations 248 248 235 235

The table reports the estimation of expanded Basu (1997) regressions that include MTBt-1 interactively to control for non-discretionary conservatism man-
dated by extant accounting standards, and in which three distinct dependent variables incorporate different changes in the value of oil and gas firms' reserves
during the year. Oil and Gas Proved (Proved Plus Probable) Reserves at Fair Value is calculated as the discounted value of estimated future revenues from
Proved (Proved plus Probable) Reserves less royalties, operating, development, and well abandonment costs. Estimates of future revenues are based on
forecast oil and gas prices, after taxes, and are discounted using a 10% rate. Proved reserves (1P) are those reserves that can be estimated with a high degree of
certainty to be recoverable, such that there is at least a 90 percent probability that the quantities actually recovered will equal or exceed the estimated proved
reserves. Proved plus probable reserves (2P) are those reserves that are equally likely that the actual remaining quantities recovered will be greater or lower
than the sum of the estimated proved plus probable reserves, i.e. the median. Panel A (Panel B) reports results for the Canadian (US) oil and gas industry.

The dependent variables are:


NI ¼ net income before extraordinary items, which incorporates the change in Oil and Gas Reserves at Book Value during the year, divided by beginning-of-
year market value of equity.
ΔNAV 1P¼ the numerator of NI plus the difference between the change in Oil and Gas Proved. Reserves at Fair Value during the year and the change in Oil and
Gas Reserves at Book Value during the year, divided by beginning-of-year market value of equity.
ΔNAV 2P¼ the numerator of NI plus the difference between the change in Oil and Gas Proved plus. Probable Reserves at Fair Value during the year and the
change in Oil and Gas Reserves at Book. Value during the year, divided by beginning-of-year market value of equity.

The explanatory variables are:


Ret¼12-month return ending three months after the closing of the fiscal year.
D¼indicator variable equal to 1 if Reto0 and 0 otherwise.
MTB¼ beginning-of-year market to book ratio.

All balance sheet variables are measured at the fiscal year end except for MTB which is measured at the beginning of the fiscal year. The regressions include
intercepts that are not reported for parsimony.
* po0.10, ** po0.05, *** po0.01 indicate the significance levels of individual coefficients in two-tailed t tests. t-statistics are based on robust standard errors
calculated clustering observations at the firm and year level. p-vals indicate the significance levels of differences in coefficients across adjacent columns in
one-tailed Wald tests.
96 M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98

consistent with those reported in Table 7, Panel A, with the coefficients on D*Ret monotonically increasing when moving
from NI to ΔNAV 1P and then to ΔNAV 2P as the dependent variable, although the coefficients and coefficient differences
across the columns are somewhat less significant.
These results are consistent with Canadian O&G firms disclosing estimates of the fair value of their reserves in a con-
ditionally conservative fashion.

6.2. U.S. O&G firms

We exploit a regulatory change in the U.S. O&G industry in December 2008, when the Securities and Exchange Commission
(SEC) issued the Modernization of Oil and Gas Reporting final rule. The calculation of the fair value of O&G reserves and disclosure
requirements under this rule are virtually identical to those in described in Section 6.1 except that the rule does not require the
disclosure of proved plus probable reserves. Approximately 90% of U.S. O&G firms do not voluntarily disclose proved plus probable
reserves, and so we cannot reliably calculate ΔNAV 2P for these firms. We hand collect fair value of proved reserves data from U.S.
O&G firms' public financial reports for the years 2004–2011. We obtain other model variables from the same sources described in
Section 4. Using these data for U.S. Oil & Gas firms, we follow the same approach as before and calculate two earnings measures, NI
and ΔNAV 1P. Requiring all variables to be available yields samples of 248 observations in the 2004–2007 pre-SEC rule period and
235 observations in the 2008–2011 post SEC-rule period.
Table 7, Panel B reports the results of estimating equation (2) using these two measures as the dependent variable, both for the
pre- and post-SEC rule periods. When the dependent variable is NI, the asymmetric timeliness coefficient β3 is significantly positive
both in the pre-regulation period reported in Column (1) (0.302, t-stat¼5.33) and in the post-regulation period reported in Column
(3) (0.321, t-stat¼2.22). In contrast, when the dependent variable is ΔNAV 1 P, β3 is insignificant in the pre-regulation period
reported in Column (2) but significantly positive in the post-regulation period reported in Column (4) (1.101, t-stat¼3.76);
moreover, a Wald test indicates that β3 increases from the pre-SEC rule period to the post-SEC rule period (p-valo0.013). These
results suggest that, similar to SFAS 157's effect on fair value measurements, the SEC rule increased the conditional conservatism of
U.S. O&G firms' subsequently disclosed fair values of proved reserves.

7. Conclusion

In this paper, we predict that firms report conditionally conservative recurring Level 2 and 3 measurements of the fair
values of recognized financial instruments when low activity in the markets for the instruments provides leeway for firms to
exercise discretion over the measurements. This prediction is based on claims and prior research showing that firms' ability
to exercise discretion over lower-level fair value measurements causes investors to discount these measurements. This
discounting provides firms with the incentive to report conditionally conservative measurements.
We identify the extent of firms' discretion over fair value measurements using their required disclosures of the amounts
of financial instruments measured based on the three levels of fair value measurement inputs defined in SFAS 157 beginning
in 2008. These input levels differ in their observability and individual completeness as measures of fair value, and thus in the
extent of the discipline that they provide over firms’ fair value measurements. Specifically, Level 1 inputs are observable and
complete measures of fair value, Level 2 inputs are observable but individually incomplete measures, and Level 3 inputs are
unobservable and individually incomplete measures. We measure conditional conservatism using an expansion of Basu's
(1997) approach that uses the part of comprehensive income attributable to fair value measurements (FVCI) and the re-
maining part of comprehensive income (CI–FVCI) as the dependent variables and that separately and interactively includes
the proportions of assets that are financial instruments measured at fair value on a recurring basis using Level 2 and 3 versus
Level 1 inputs. As predicted, we find that firms with higher Level 2 and 3 proportions record more conditionally conservative
FVCI, but not CI–FVCI. We find no association of firms' Level 1 proportions with the conditional conservatism of FVCI. We
further predict and provide evidence that the conditional conservatism of Level 2 and 3 fair value measurements increases
when the measurements are evaluated by more knowledgeable investors, verified by more independent third parties, and
disclosed more fully in financial reports, as required by SFAS 157. Lastly, we predict and provide evidence that the condi-
tional conservatism of Level 2 and 3 fair value measurements is lower when firms narrowly meet or beat earnings targets,
suggesting they have incentives that conflict with conditional conservatism.
We demonstrate that these results generalize to balance sheet items other than financial instruments by examining pro
forma earnings measures calculated using Canadian and U.S. oil and gas firms’ disclosures of measures akin to Level 2 or
3 fair values of their reserves. We find that the conditional conservatism of the firms' (pro forma) earnings increases with
whether and the extent to which those earnings include changes in the fair value of their reserves.
As discussed in the introduction, our findings speak to the ongoing debates about the relative merits of unbiased versus
conditionally conservative accounting measurement and, relatedly, of fair value versus (impaired) amortized cost mea-
surement. Our findings have direct implications for empirical research examining fair value measurements' predictive
power and their value-, returns-, and risk-relevance.
We acknowledge that our evidence does not address the following important and difficult questions, which merit (re)
consideration by future researchers. First, do firms' fair value measurements exhibit an optimal level of conditional con-
servatism? We only show that firms report conditionally conservative lower-level measurements when they have incentives
M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98 97

to do so. Future researchers could attempt to specify and estimate the cost-benefit trade-offs of conditional conservatism
from the perspectives of reporting firms and users of financial reports. Second, do firms anticipate being in positions where
they want to exercise discretion over fair value measurements, and so elect to acquire potentially illiquid assets and li-
abilities? We take firms' exposures as given. Future researchers could attempt to model the accounting-related determi-
nants of firms' choices regarding asset and liability liquidity. Third, should accounting standard setters expand fair value
measurement or conditional conservatism to additional balance sheet items? The valuation and other economic con-
sequences of fair value measurement and conditional conservatism are examined in extensive, and to date mostly separate,
prior literatures. We provide evidence that firms report conditionally conservative Level 2 and 3 fair value measurements to
mitigate discounting of those measurements by investors. Future researchers could identify other reasons why fair value
measurement and conditionally conservatism overlap or interrelate.

Appendix 1. Variable definitions

CI The sum of net income before extraordinary items and fair value components of comprehensive income divided by beginning-of-year
market value of equity. Compustat items: (ib þ cisecgl þ cidergl)t/(prcc*csho)t-1
D Indicator variable that equals 1 if Ret o 0 and 0 otherwise
FVCI Compustat-based measure of fair value components of comprehensive income divided by beginning-of-year market value of equity. Fair
value components of comprehensive income include: net realized and unrealized gains (losses) on the disposition of securities held for
resale; net realized and unrealized gains (losses) from the disposition of investment securities; gains (losses) on hedge transactions that
exceed the risk faced by the company; total unrealized gains (losses) on investment securities reported in other comprehensive income; and
derivative gains (losses) reported in other comprehensive income. Compustat items: (tdsg þisgt þhedgeglþ cisecgl þ cidergl)t/(prcc*csho)t-1
CI-FVCI CI minus FVCI
Lev1 The sum of Level 1 fair value assets and Level 1 fair value liabilities divided by total assets. Compustat items: (aqpl1 þlqpl1)/at
Lev2 The sum of Level 2 fair value assets and Level 2 fair value liabilities divided by total assets. Compustat items: (aol1 þlol1)/at
Lev3 The sum of Level 3 fair value assets and Level 3 fair value liabilities divided by total assets. Compustat items: (aul1 þlul1)/at
Lev23 Lev2þ Lev3
MTB Beginning-of-year market value of equity divided by the beginning-of-year book value of equity. Compustat items: (prcc*csho/ceq)t-1
Ret 12-month return ending three months after the fiscal year end, calculated by compounding twelve monthly CRSP returns

References

Appel, I., Gormley, T., Keim, D., 2014. Passive investors, not passive owners. Working Paper. December, University of Pennsylvania.
Badertscher, B., Burks, J., Easton, P., 2014. The market pricing of other-than-temporary impairments. Account. Rev. 89 (3), 811–838.
Ball, R., Shivakumar, L., 2005. Earnings quality in UK private firms: comparative loss recognition timeliness. J. Account. Econ. 39, 83–128.
Ball, R., Kothari, S.P., Nikolaev, V., 2013a. Econometrics of the Basu asymmetric timeliness coefficient and accounting conservatism. J. Account. Res. 51,
1071–1097.
Ball, R., Kothari, S.P., Nikolaev, V., 2013b. On estimating conditional conservatism. Account. Rev. 88, 755–787.
Barth, M.E., 2006. Including estimates of the future in today's financial statements. Account. Horiz. 20, 271–285.
Barth, M.E., 2014. Measurement in financial reporting: the need for concepts. Account. Horiz. 28 (2), 331–352.
Barth, M.E., Taylor, D., 2010. In defense of fair value: weighing the evidence on earnings management and asset securitizations. J. Account. Econ. 49, 26–33.
Basu, S., 1997. The conservatism principle and the asymmetric timeliness of earnings. J. Account. Econ. 24, 3–37.
Basu, S., Hwang, L., Jan, C., 2001. Differences in conservatism between big eight and non-big eight auditors. Working Paper. Baruch College, City University
of New York.
Beaver, W.H., Ryan, S.G., 2005. Conditional and unconditional conservatism: concepts and modeling. Rev. Account. Stud. 10, 269–309.
Bell, T.B., Causholli, M., Knechel, W.R., 2015. Audit firm tenure, non-audit services, and internal assessments of audit quality. J. Account. Res. 53 (3), 461–509.
Brav, A., Jiang, W., Partnoy, F., Thomas, R., 2008. Hedge fund activism, corporate governance, and firm performance. J. Financ. 63, 1729–1775.
Buffett, W., 2003. Chairman's Letter, Berkshire Hathaway, Inc., 2002 Annual Report. 〈http://www.berkshirehathaway.com/letters/2002pdf.pdf〉.
Bushee, B.J., 2001. Do institutional investors prefer near-term earnings over long-run value? Contemp. Account. Res. 18 (2), 207–246.
Bushee, B.J., Noe, C.F., 2000. Corporate disclosure practices, institutional investors, and stock return volatility. J. Account. Res. 38, 171–202.
Dietrich, D., Muller, K., Riedl, E., 2007. Asymmetric timeliness tests of accounting conservatism. Rev. Account. Stud 12, 95–124.
D’Souza, J., Ramesh, K., Shen, M., 2010. The interdependence between institutional stock ownership and information dissemination by capital market data
aggregators. Account. Rev. 85 (1), 159–193.
Financial Accounting Standards Board (FASB), 2006. Fair value measurements. Statement of Financial Accounting Standards No. 157. Norwalk, CT.
Financial Accounting Standards Board (FASB), 2011. Accounting standards update no. 2011-04 – fair value measurement (topic 820): amendments to
achieve common fair value measurement and disclosure requirements in U.S. GAAP and IFRSs. Financial Accounting Standards Board (FASB). Norwalk,
CT.
Givoly, D., Hayn, C., 2000. The changing time-series properties of earnings, cash flows and accruals: has financial reporting become more conservative? J.
Account. Econ. 29 (3), 287–320.
Givoly, D., Hayn, C., Natarajan, A., 2007. Measuring reporting conservatism. Account. Rev. 82, 65–106.
Goh, B.W., Ng, J., Yong, K.O., 2009. Market pricing of banks’ fair value assets reported under SFAS 157 during the 2008 economic crisis. Working Paper.
Singapore Management University and MIT.
Kolev, K., 2009. Do investors perceive marking-to-model as marking-to-myth? Early Evidence from FAS 157 Disclosure SSRN Working Paper Series, No.
1336368. 〈http://ssrn.com/abstract¼ 1336368〉.
Kothari, S.P., Ramanna, K., Skinner, D., 2010. Implications for GAAP from an analysis of positive research in accounting. J. Account. Econ. 50, 246–286.
Lawrence, A., Sloan, R., Sun, Y., 2013. Non-discretionary conservatism: evidence and implications. J. Account. Econ. 56 (2-3), 112–133.
98 M. Badia et al. / Journal of Accounting and Economics 63 (2017) 75–98

Lev, B., Zhou, N., 2009 Unintended consequences: fair value accounting informs liquidity risk. Working Paper.
Lobo, G., Parthasarathy, K., Sivaramakrishnan, K., 2008. Growth, managerial reporting behavior, and accounting conservatism. Working paper. University of
Houston.
Longstaff, F., 2010. The subprime credit crisis and contagion in financial markets. J. Financ. Econ. 97 (3), 436–450.
Nissim, D., Penman, S., 2008. Principles for the application of fair value accounting. Center for Excellence in Accounting and Security Analysis, White Paper
No. 2. April.
Park, S., Lee, J., 2014. Practice of accounting conservatism in MBE firms. J. Appl. Bus. Res. 30 (6), 1785–1802.
Patatoukas, P.N., Thomas, J.K., 2011. More evidence of bias in the differential timeliness measure of conditional conservatism. Account. Rev. 86, 1765–1793.
Petersen, M.A., 2009. Estimating standard errors in finance panel data sets: comparing approaches. Rev. Financ. Stud. 22, 435–480.
Ramalingegowda, S., Yu, Y., 2012. Institutional ownership and conservatism. J. Account. Econ. 53, 98–114.
Ramanna, K., Watts, R., 2012. Evidence on the use of unverifiable estimates in required goodwill impairment. Rev. Account. Stud. 17 (4), 747–780.
Riedl, E.J., Serafeim, G., 2011. Information risk and fair values: an examination of equity betas. J. Account. Res. 49 (4), 1083–1122.
Roychowdhury, S., 2006. Management of earnings through the manipulation of real activities that affect cash flow from operations. J. Account. Econ. 42 (3),
335–370.
Roychowdhury, S., 2010. Discussion of: acquisition profitability and timely loss recognition. J. Account. Econ. 49 (1-2), 179–183.
Roychowdhury, S., Watts, R., 2007. Asymmetric timeliness of earnings, market to book and conservatism in financial reporting. J. Account. Econ. 44 (1-2),
2–31.
Roychowdhury, S., Martin, X., 2013. Understanding discretion in conservatism: an alternative viewpoint. J. Account. Econ. 56, 134–146.
Ruddock, C., Taylor, S., Taylor, S., 2006. Nonaudit services and earnings conservatism: is auditor Independence impaired? Contemp. Account. Res. 23 (3),
701–746.
Ryan, S.G., 2006. Identifying conditional conservatism. Eur. Account. Rev. 15, 511–525.
Ryan, S.G., 2007. Financial Instruments and Institutions: Accounting and Disclosure Rules, 2nd Ed. John Wiley & Sons, Hoboken, N.J..
Ryan, S.G., Zarowin, P.A., 2003. Why has the contemporaneous linear returns-earnings relation declined? Account. Rev. 78 (2), 523–553.
Ryan, S.G., 2008. Fair value accounting: understanding the issues raised by the credit crunch. White Paper Prepared for the Council of Institutional
Investors. New York.
Shaffer, S., 2011. Evaluating the impact of fair value accounting on financial institutions: implications for accounting standards setting and bank supervision.
Federal Reserve Bank of Boston Working Paper No. QAU 12-01.
Song, C., Thomas, W., Yi, H., 2010. Value relevance of FAS 157 fair value hierarchy information and the impact of corporate governance mechanisms.
Account. Rev. 85, 1375–1410.
U.S. Securities and Exchange Commission, Office of the Chief Accountant, Division of Corporation Finance, 2008. Report and Recommendations Pursuant to
Section 133 of the Emergency Economic Stabilization Act of 2008: Study on Mark-To-Market Accounting. December 30.
Watts, R., 2003. Conservatism in accounting part I: explanations and implications. Account. Horiz. 17 (3), 207–221.
Weil, J., 2007. Mark-to-make-believe accounting claims a victim. Bloomberg.com. October 27.
Zang, A.Y., 2012. Evidence on the trade-off between real activities manipulation and accrual-based earnings management. Account. Rev. 87 (2), 675–703.
Zhang, J., 2008. The contracting benefits of accounting conservatism to borrowers and lenders. J. Account. Econ. 45, 27–54.

You might also like