4.1 Ball et al. (2013a) model of accounting income recognition
In this section, we discuss the accounting income recognition model proposed by Ball et al. (
2013a).
11 The model distinguishes four unobserved (latent) components that capture the key characteristics of income recognition as practiced. The purpose of this section is to simulate artificial returns and earnings based on the Ball et al. (
2013a) model and a modified version that excludes accounting conservatism. The goal of the model is not to match all empirical properties of returns and earnings, but rather to generate artificial returns and earnings from underlying latent factors with different distributional properties. This simulation approach thus allows for comparisons between firms with different levels of skewness in returns and earnings and between conservative and non-conservative firms, facilitating a strict comparison that would be challenging using empirical data and bootstrapping methods.
12
The relation between unexpected returns (
\({R}_{it}\)) and unexpected earnings (
\({I}_{it}\)) in the Ball et al. (
2013a) model is described by Eqs. (
8) and (
9):
$${R}_{it}={x}_{it}+{y}_{it}+{g}_{it}$$
(8)
$${I}_{it}={x}_{it}+{w}_{it}{y}_{it}+\left(1-{w}_{it-1}\right){y}_{it-1}+{g}_{it-1}+{\varepsilon }_{it}-{\varepsilon }_{it-1},$$
(9)
where the subscripts
i and
t denote firm and year, respectively. Returns consist of three unobserved information components
\({x}_{it}\), \({y}_{it},\) and
\({g}_{it}\). The information component
\({x}_{it}\) is incorporated into accounting income contemporaneously. The second information component
\({y}_{it}\) is subject to conditional conservatism. It is incorporated into accounting income contemporaneously or with a lag, depending on its state. When
\({y}_{it}\) is negative, it is reported immediately in accounting income (
\({w}_{it}=1\) if
\({y}_{it}<0\)), while reporting
\({y}_{it}\) is delayed to the next period during positive states (
\({w}_{it}=0 \mathrm{if }{y}_{it}\ge 0\)).
13 The third information component
\({g}_{it}\) is always incorporated into accounting income with delay. The fourth component
\({\varepsilon }_{it}\) is an accounting error that is reversed in the next period.
Without loss of generalizations, we assume that the unobserved components
\({x}_{i,t}\), \({y}_{i,t},\) \({g}_{i,t},\) and
\({\varepsilon }_{i,t}\) have mean zero and are serially uncorrelated. However, we do allow, following Ball et al. (
2013a), for a positive contemporaneous correlation
ρ between
\({x}_{it} , {y}_{it},\) and
\({g}_{it}\). Given these properties, the covariance between
I and
R, conditional on a negative state (
\({w}_{it}=1\)), is:
$$\begin{aligned} \sigma_{RI|w = 1} & = Cov\left( {\left( {x_{it} + y_{it} + g_{it} } \right),\left( {x_{it} + y_{it} + \left( {1 - w_{it - 1} } \right)y_{it - 1} + g_{it - 1} + \varepsilon_{it} - \varepsilon_{it - 1} } \right)|w_{it} = 1} \right) \\ & = Cov\left( {\left( {x_{it} + y_{it} + g_{it} } \right),\left( {x_{it} + y_{it} } \right)|w_{it} = 1} \right) \\ & = \sigma_{x|w = 1}^{2} + \sigma_{y|w = 1}^{2} + 2\sigma_{xy|w = 1} + \sigma_{xg|w = 1} + \sigma_{yg|w = 1} , \\ \end{aligned}$$
(10)
while the covariance between
I and
R, conditional on a positive state (
\(w_{it} = 0\)), is:
$$\begin{aligned} \sigma_{RI|w = 0} & = Cov\left( {\left( {x_{it} + y_{it} + g_{it} } \right),\left( {x_{it} + \left( {1 - w_{it - 1} } \right)y_{it - 1} + g_{it - 1} + \varepsilon_{it} - \varepsilon_{it - 1} } \right)|w_{it} = 1} \right) \\ & = Cov\left( {\left( {x_{it} + y_{it} + g_{it} } \right),x_{it} |w_{it} = 1} \right) \\ & = \sigma_{x|w = 0}^{2} + \sigma_{xy|w = 0} + \sigma_{xg|w = 0} . \\ \end{aligned}$$
(11)
Ball et al (
2013a) assume that the unobserved components
x,
y, and
g follow a symmetric (non-skewed) distribution. In this case, since
\({\sigma }_{A|w=1}={\sigma }_{A|w=0}\) for
\(A\in \left\{x,y,g,xy,xg,yg\right\}\), it follows that
\({\sigma }_{RI|w=1}-{\sigma }_{RI|w=0}={\sigma }_{y}^{2}+{\sigma }_{yg}\), which is a strictly positive number implying that
\({\sigma }_{RI|w=1}\ge {\sigma }_{RI|w=0}\). In the absence of conservative reporting, (i.e. when the component
y is not included in the model), the difference between
\({\sigma }_{RI|w=1}\) and
\({\sigma }_{RI|w=0}\) reduces to zero. In summary, conservative accounting practice causes the covariance between returns and earnings to be higher in negative states than in positive states. Given this result, the asymmetric timeliness coefficient
\({\beta }_{1}\) in Eq. (
1) is expected to be positive in the presence of conservatism and zero in the absence of conservatism.
14 However, this result depends crucially on the assumption that
x,
y, and
g follow a symmetric distribution. When these components are skewed, such that
\({\sigma }_{A|w=1}\ne {\sigma }_{A|w=0}\), the inequality
\({\sigma }_{RI|w=1}\ge {\sigma }_{RI|w=0}\) is not necessarily valid, and hence the sign of
\({\beta }_{1}\) is unclear.
We illustrate this insight by a simulation exercise. We simulate the components
\(x, y, g\), and
\(\varepsilon\) for a cross-section of
N = 1000 firms with
t = 2 time-series observations per firm. For each firm
i, we thus draw eight random numbers
\(\left({x}_{i0}, {x}_{i1}, {y}_{i0},{y}_{i1},{g}_{i0},{g}_{i1},{\varepsilon }_{i0},{\varepsilon }_{i1}\right)\) from a Multivariate Standard Normal (i.e. non-skewed) distribution. From these simulated random numbers, we compute
\({R}_{i1} \mathrm{and} {I}_{i1},\) implied by Eqs. (
8)–(
9), respectively, for each firm
i. We also use the simulated variables
\(\left({x}_{i0}, {x}_{i1}, {g}_{i0},{g}_{i1},{\varepsilon }_{i0},{\varepsilon }_{i1}\right)\) to generate returns and earnings from a restricted specification of the model that does not feature accounting conservatism. This model is equal to the model by Ball et al. (Eqs.
8–
9), with the difference that the term
y is excluded from the model:
$${R}_{it}={x}_{it}+{g}_{it}$$
(12)
$${I}_{it}={x}_{it}+{g}_{it-1}+{\varepsilon }_{it}-{\varepsilon }_{it-1}.$$
(13)
We then use the simulated samples of
N observations of earnings and returns to estimate the coefficients of the Basu (
1997) regression Eq. (
1), both in the presence and absence of conservatism.
This process is repeated
r = 10,000 times, and we report the means of the 10,000 estimates in Table
4. We consider four different values of the correlation ρ between
x, y, and
g (
\(\rho \in \left\{0, 0.2, 0.5, 0.8\right\}\)). The top panel of Table
4 shows that in the presence of conservatism, the mean of the estimated AT coefficients (
\({\beta }_{1}\)) is indeed positive and significant. In the absence of conservatism (right panel), the mean of the estimated AT coefficients (
\({\beta }_{1}\)) is, as expected, insignificant and close to zero. With Normally-distributed data, the AT coefficient thus correctly identifies conservatism.
Table 4
Simulation of the Ball et al. (
2013a) model
No skewness | 0 | 0.12 | 0.00 | 0.41 | 0.18 | 0.00 | 0.00 | 0.00 | 0.00 | 0.50 | 0.00 | 0.00 | 0.00 | 0.17 |
0.2 | 0.17 | 0.00 | 0.40 | 0.21 | 0.00 | 0.00 | 0.00 | 0.00 | 0.50 | 0.00 | 0.00 | 0.00 | 0.21 |
0.5 | 0.26 | − 0.01 | 0.37 | 0.26 | 0.00 | 0.00 | 0.00 | 0.01 | 0.50 | 0.00 | 0.00 | 0.00 | 0.26 |
0.8 | 0.34 | 0.00 | 0.35 | 0.30 | 0.00 | 0.00 | 0.00 | 0.00 | 0.50 | 0.00 | 0.00 | 0.00 | 0.30 |
x skewed | 0 | − 0.05 | 0.07 | 0.53 | − 0.02 | 0.18 | 0.09 | − 0.21 | − 0.01 | 0.67 | − 0.38 | 0.33 | 0.12 | 0.36 |
0.2 | 0.01 | 0.05 | 0.49 | 0.04 | 0.22 | 0.10 | − 0.20 | − 0.03 | 0.65 | − 0.35 | 0.37 | 0.12 | 0.38 |
0.5 | 0.08 | 0.04 | 0.45 | 0.11 | 0.27 | 0.11 | − 0.17 | − 0.05 | 0.61 | − 0.29 | 0.42 | 0.12 | 0.39 |
0.8 | 0.17 | 0.02 | 0.42 | 0.17 | 0.31 | 0.13 | − 0.15 | − 0.06 | 0.59 | − 0.24 | 0.47 | 0.12 | 0.40 |
y skewed | 0 | 0.15 | − 0.07 | 0.36 | 0.17 | 0.18 | 0.06 | | | | | | | |
0.2 | 0.19 | − 0.08 | 0.36 | 0.19 | 0.22 | 0.09 | | | | | | | |
0.5 | 0.28 | − 0.10 | 0.34 | 0.24 | 0.27 | 0.13 | | | | | | | |
0.8 | 0.40 | − 0.16 | 0.32 | 0.28 | 0.31 | 0.16 | | | | | | | |
g skewed | 0 | 0.25 | 0.03 | 0.33 | 0.39 | 0.18 | 0.09 | 0.21 | 0.02 | 0.33 | 0.38 | 0.33 | 0.11 | 0.01 |
0.2 | 0.28 | 0.07 | 0.33 | 0.39 | 0.22 | 0.10 | 0.19 | 0.04 | 0.36 | 0.34 | 0.37 | 0.12 | 0.05 |
0.5 | 0.36 | 0.09 | 0.32 | 0.42 | 0.27 | 0.10 | 0.16 | 0.06 | 0.39 | 0.28 | 0.42 | 0.12 | 0.14 |
0.8 | 0.44 | 0.12 | 0.31 | 0.45 | 0.31 | 0.11 | 0.13 | 0.07 | 0.41 | 0.23 | 0.47 | 0.12 | 0.21 |
We repeat this experiment after transforming the Standard Normally distributed variables
x, y, and
g into right-skewed variables, using the same transformation as in Sect.
3, based on the Skew-Normal distribution by Azzalini (2003). This transformation of the underlying components
x, y, and
g induces skewness of the realized earnings and returns. The results are reported in the lower three panels of Table
4. After allowing the unobserved components of returns and earnings to be skewed, the mean AT coefficients
\({\beta }_{1}\) remain significant, but they are no longer uniformly positive. Hence, even if the data are generated by a model that features accounting conservatism, the Basu regression could indicate negative rather than positive asymmetric timeliness, due to the non-symmetric distribution of the underlying components.
Using the restricted model that is free of conservatism Eqs. (
12)–(
13), a statistically significant AT coefficients (
\({\beta }_{1}\)) appears once we introduce skewness to the latent variables
x and
g. This constitutes a clear example of “spurious conservatism”, in which the statistical distribution of the underlying components leads to an asymmetric relation between returns and earnings, which may be interpreted incorrectly as accounting conservatism. The results in Table
4 further indicate that skewness does not only affect the AT coefficient (
\({\beta }_{1}\)), but also the estimated individual effect of the positive-return dummy variable (
\({\alpha }_{1}\)) is spuriously different from zero in the presence of skewness.
The final column of Table
4 presents the difference between the estimated AT coefficient in the presence and absence of conservatism. Importantly, this difference is consistently positive across all cases. Specifically, when keeping skewness constant, introducing conservatism into the model leads to an increase in the estimated AT coefficient. This finding confirms that the Basu regression has the ability to distinguish between firms with and without conservative accounting practices, given that returns and earnings in both groups exhibit similar skewness properties.
In such a scenario where the skewness properties are comparable, the variation in the AT coefficient across groups provides stronger evidence of accounting conservatism, as it is not driven by differences in skewness alone. However, in practice, achieving similar skewness properties between the two groups can be challenging due to significant cross-sectional differences in the skewness of earnings and returns. In the next section, we compare empirical AT coefficients across firms sorted based on market capitalization and market-to-book ratio, further exploring this issue.
It is worth noting that the skewness of the simulated earnings, as reported in Table
4, does not appear to be influenced by the presence of conservatism. The skewness coefficients of earnings vary based on the skewness and correlations of the underlying components, but there is no systematic difference observed between the full model (left panel) and the restricted model without conservatism (right panel). This finding contrasts with the arguments made by Ball and Shivakumar (
2005) and Givoly and Hayn (
2000), who propose that negative earnings skewness may be a result of accounting conservatism. Within the framework of the Ball et al. (
2013a) model, it seems that conservatism does not have a direct effect on earnings skewness. The variation in skewness appears to be primarily driven by the characteristics of the underlying components in the model rather than the presence or absence of conservatism.
15
A tempting solution to address the skewness-induced component of the asymmetric timeliness (AT) coefficient is to transform the variables. For instance, a logarithmic transformation can reduce right skewness, while a rank transformation can eliminate both positive and negative skewness from a variable. We apply both of these transformations to the simulated observations of returns (
R) and earnings (
I) before estimating the Basu model (Eq.
1). However, the resulting AT coefficients, as shown in Table
5, indicate that these data transformations do not resolve the issue at hand.
Table 5
Skew-reducing transformations
No skewness | 0 | 0.17 | 0.13 | 0.16 | 0.04 | − 0.01 | 0.00 | 0.00 | 0.00 |
0.2 | 0.20 | 0.18 | 0.18 | 0.06 | − 0.01 | 0.00 | 0.00 | 0.00 |
0.5 | 0.25 | 0.24 | 0.24 | 0.10 | − 0.01 | 0.00 | 0.00 | 0.00 |
0.8 | 0.29 | 0.32 | 0.29 | 0.14 | − 0.01 | 0.00 | 0.00 | 0.00 |
x skewed | 0 | − 0.03 | − 0.07 | − 0.03 | − 0.01 | − 0.39 | − 0.35 | − 0.35 | − 0.08 |
0.2 | 0.03 | − 0.05 | 0.03 | − 0.01 | − 0.36 | − 0.38 | − 0.32 | − 0.09 |
0.5 | 0.09 | 0.00 | 0.10 | 0.02 | − 0.30 | − 0.39 | − 0.27 | − 0.11 |
0.8 | 0.15 | 0.04 | 0.17 | 0.03 | − 0.25 | − 0.39 | − 0.21 | − 0.12 |
y skewed | 0 | 0.16 | 0.08 | 0.15 | 0.03 | | | | |
0.2 | 0.17 | 0.08 | 0.16 | 0.03 | | | | |
0.5 | 0.22 | 0.13 | 0.22 | 0.07 | | | | |
0.8 | 0.27 | 0.17 | 0.27 | 0.09 | | | | |
g skewed | 0 | 0.37 | 0.24 | 0.34 | 0.07 | 0.38 | 0.18 | 0.36 | 0.05 |
0.2 | 0.38 | 0.25 | 0.35 | 0.10 | 0.33 | 0.14 | 0.32 | 0.05 |
0.5 | 0.41 | 0.28 | 0.38 | 0.12 | 0.27 | 0.07 | 0.26 | 0.04 |
0.8 | 0.43 | 0.31 | 0.42 | 0.16 | 0.22 | 0.00 | 0.21 | 0.02 |
Even when using log and rank transformed data, we consistently observe that the AT coefficient varies with the distributional properties of the latent factors in our simulation. Consequently, the AT coefficient does not exhibit a consistent positive relationship with conservatism when it is present, and it often remains significantly different from zero even in the absence of conservatism. The reason these data transformations do not help in identifying conservatism is that the skewness originates from the latent factors (x, y, and g) in our model. The skewness of these latent factors not only affects the skewness of the observable variables (R and I) but also introduces inherent nonlinearity in the relationship between these variables. Merely removing the univariate skewness from R and I post-transformation does not eliminate the nonlinearity in the relationship between them. This nonlinearity, which is unrelated to accounting conservatism in our simulation, leads to a statistically significant AT coefficient even if the variables in the regression are transformed to reduce skewness.
Table
5 also shows that the AT coefficient does not identify conservatism when the model is estimated using the Theil (
1950)–Sen (
1968) (TS) estimator, which mitigates the impact of outlier observations. Contrary to suggestions made by Kim and Ohlson (
2018), the TS estimator does not resolve the issue of the spurious AT coefficient caused by skewness, as the asymmetry in the earnings-returns relationship extends beyond the distribution's tails. Therefore, the TS estimator does not offer a solution to address the nonlinearity and asymmetry induced by skewness, preventing the reliable identification of accounting conservatism using the AT coefficient.
Finally, Table
5 presents the difference in R-squared values, denoted as
ΔR2 =
R2(−) −
R2(+), which represents the disparity between the R-squared obtained from regressing earnings on returns in the subsamples of negative and positive news. This measure is relevant to Basu's (
1997) hypothesis, suggesting that conservative reporting leads to a stronger relationship between earnings and returns, resulting in a higher R-squared during periods of bad news. However, the findings in Table
5 indicate that the differential R-squared is also influenced by skewness. While
ΔR2 generally demonstrates higher values in the presence of conservatism (Panel A), it does not necessarily equal zero in the absence of conservatism (Panel B). These results align with the simulation outcomes presented in Tables
1,
2, and
3, which show that the correlation (ρ) between earnings and returns is not consistent across positive and negative news samples. Consequently, the squared correlation, or R-squared, can vary between positive and negative news states for reasons unrelated to conservative reporting.