The Effect of Aggregation of Accounting Information via Segment Reporting on Accounting Conservatism

Abstract

In a sample of U.S. multiple-segment firms, we document a negative association between aggregation via segment reporting and timely loss recognition. A higher level of aggregation, as reflected in a firm’s reported organizational structure (the definition and characteristics of its segments), causes a multiple-segment firm to exhibit less cross-segment variation in profitability than a matched control portfolio of single-segment firms. We find that firms that engage in more aggregation report accounting numbers that provide less timely information about economic losses. We also observe that firms that provide more disaggregated segment data subsequent to adopting SFAS 131 experienced an increase in timely loss recognition. This result implies that higher quality segment reporting leads to an increase in timely loss recognition, which, per extant research, is associated with better governance. Our results complement results in Berger and Hann [2003. The impact of SFAS No. 131 on information and monitoring. Journal of Accounting Research, 41, 163–223] that show a decline in inefficient internal-capital-market transfers subsequent to the adoption of SFAS 131. Overall, we provide evidence supporting Beyer, Cohen, Lys, and Walther’s [2010. The financial reporting environment: Review of the recent literature. Journal of Accounting and Economics, 50, 296–343] contention that accounting conservatism is, in part, a function of managers’ aggregation choices.

Acknowledgements

This paper has benefitted from comments and suggestions provided by an anonymous reviewer, Peter Easton, Mozaffar Khan, Stephen Ryan, Florin Vasvari as editor, Terry Warfield and workshop participants at the HEC-INSEAD Accounting Colloquium, Columbia University, London Business School, New York University, Tilburg University, and the University of Arizona. We are solely responsible for any errors.

Notes

1 We are not the first to either note or study the variance effect. For example, Hayes and Lundholm (1996) develop an analytical model in which the variance effect is determined by managers' strategic decisions regarding the definition of a firm's segments.

2 R_SR is the acronym for Ranked_Standard deviation Ratio. For each multiple-segment firm in our sample, it is the ratio of the standard deviation of a set of single-segment firms' profit margins to the standard deviation of the profit margins of the segments of the multiple-segment firm. The single-segment firms are matched by industry and size to the segments of the multiple-segment firm. We provide a detailed description of how R_SR is calculated and the steps we take to increase its construct validity in Section 3. We discuss validity tests of R_SR in Section 6.

3 We assume that the Basu (1997) regression captures asymmetric timely loss recognition. However, we acknowledge that there is a debate about how to interpret estimates obtained from it. For example, studies by Patatoukas and Thomas (2011, 2016) and by Dietrich et al. (2007) provide evidence that the regression is mis-specified, whereas Ball, Kothari, and Nikolaev (2013a, 2013b) provide evidence that it is not. Given this debate, we conduct several robustness checks, which are discussed in Section 6.2.

4 The literature on conservatism and the confirmatory role of accounting is more developed with regard to debt contracting (e.g. Watts, 2003; Ball, Bushman, & Vasvari, 2008; Wittenberg-Moerman, 2008). A notable exception, however, is the paper by García Lara, García, Osma, and Penalva (2014), which takes an equity market perspective.

5 Evidence of the role of segment reports in determining the confirmatory role of accounting can be found in Berger and Hann (2007), Bens et al. (2011) and Blanco, García, Lara, and Tribó (2014). We complement these studies by focusing on conservatism.

6 Ball (2001) and Watts (2003a, 2003b) provide detailed explanations of how such communication arises endogenously because of governance needs and lending arrangements.

7 There is a sizeable body of literature on segment reporting. For example, Harris (1998) studies how segments are determined; Tse (1989) and Chen and Zhang (2003) evaluate how segment reporting affects the information content of financial reports; Berger and Hann (2007) and Bens et al. (2011) examine how incomplete segment disclosure is correlated with agency problems; Blanco et al. (2014) demonstrate that committing to report higher quality earnings leads to higher quality segment disclosures; and Franco, Urcan, and Vasvari (2016) demonstrate that firms with higher quality segment disclosures enjoy a lower cost of debt capital because their creditors are able to evaluate the co-insurance effect of line of business diversification (e.g. Lewellen, 1971) better. The two studies in the segment-reporting literature that are the most-closely related to our study are Givoly, Hayn, and D’Souza (1999) and Ettredge et al. (2005). Our research differs from Givoly et al. (1999) in two fundamental ways. First, they evaluate the information content of segment-level income whereas we focus on the conservatism of firm-level earnings. Second, in order to construct the key empirical proxy in Givoly et al. (1999) a long time series (on average 14 years) is required. Our measure, on the other hand, is constructed at the firm-year level and it can be computed using a single year of data. Our results complement the results in Ettredge et al. (2005), which document increases in the future earnings response coefficients of firms that report more segments after SFAS 131.

8 We assume that conservative reporting is valued by all capital-market investors. Although the vast majority of the literature focuses on creditors' demands (e.g. Wittenberg-Moerman, 2008; Ball, Bushman, et al., 2008), Watts (2003) argues that timely loss recognition is also important to shareholders. Evidence supporting this argument is provided by García Lara et al. (2014).

9 We match on industry at the most precise level possible using SIC codes. We first determine whether there is a valid match using the four-digit SIC code. If not, we determine whether there is a valid match using the three-digit code. Finally, if we still have not found a match, we determine whether there is a valid match using the two-digit code. If we cannot find a match using either four-digit, three-digit or two-digit SIC code, we exclude the segment and the corresponding AMS firm from our sample. 63%, 22% and 15% of our matches are made on the basis of four-digit, three-digit and two-digit SIC codes, respectively.

10 As discussed in Section 6.2, we draw the same inferences when we use the contemporaneous value of SR_it.

11 We also conduct our tests without adjustments for these factors – i.e. we use the RANK_SR variable. The untabulated results of these tests lead to the same conclusions as the results shown in the tables.

12 We also show the number of observations in each decile (i.e. N). The reason that N varies across deciles is that we remove extreme values of other variables after assigning observations to deciles.

13 When the X_it is a ratio (e.g. book-to-market ratio), we set X_it for the PMS firm equal to the sales-weighted average of the numerator of the single-segment firms matched to the segments of AMS firm i in year t divided by the sales-weighted average of the denominator of the single-segment firms. When X_it is a function of stock return (e.g. annual stock return, RET_it), we calculate stock return for PMS firm i by taking a sales-weighted average of the separate returns for the single-segment firms matched to the segments of AMS firm i in year t.

14 Although we match a single-segment firm to a segment of the AMS firm on the basis of industry sales, some differences in sales are significant. When we require that the sum of the sales of the matched portfolio of single-segment firms are within ±300% of the total sales of the multiple-segment firm, our sample size declines from 11,464 to 10,950. The conclusions we draw after conducting our empirical tests on this restricted sample are the same as those discussed in the main text.

15 EXVAL is drawn from Bens and Monahan (2004) who document an association between the excess value of diversification and aggregation. RELATE, NUMSEG and CAPINT are drawn from Givoly et al. (1999). Givoly et al. (1999) also include a relative segment size variable and an international operations measure. The former is not applicable in our model at the firm level; the latter is problematic as international segment reporting changed midway through our sample via SFAS 131 (Hope & Thomas, 2008). See Table 3 of Givoly et al. for more details of their approach.

16 Inclusion of these variables is partially redundant, as we have purged R_SR of its variation with the constructs with respect to their differences across single and multiple-segment firms. However, it is still possible that within the multiple-segment sample both R_SR and the asymmetric timeliness of earnings might vary with these factors. In addition, we note that the increase in explanatory power for the model that adds R_SR is quite low. Still, the statistical significance of the coefficient is high and suggests that allowing the bad news coefficient to vary with R_SR provides a better understanding of the conditions under which such news translates into earnings.

17 Although one might expect the effects to be concentrated around the adoption year (i.e. a two-year window with one year pre and post), we do not observe significant results when we use this time period. A possible explanation for this is that there is an adaptation process in the application of SFAS 131 and how it relates to the communication of bad news. Also, note that for our predicted result to be detectable, we need a sufficient number of firms from the post period and with previously hidden segments and with negative returns. So there are power issues involved. We believe that using three years before and after: (1) addresses these issues and (2) is a tight enough window around the reporting change.

18 This affects 9.28% of the sample. We obtain similar results when we do not replace negative values of HS_i with zero.

19 Note that all firms in this sample reported multiple segments in at least one year – i.e. there are no firms in this sample that are single-segment throughout the entire sample period.

20 We also explored other possible data sources where it appeared that more disaggregated data about firm operations were collected by independent sources, but we found that the promised disaggregated data was either nearly identical to the segments as defined by the company in its financial statements (Thomson Reuters) or not consistently measured over time (Lexis Nexis).

21 We also find a significant negative relation between range-based aggregation proxies and ATLR when we use the contemporaneous value of R_RR and when we use the unorthogonalized variable RANK_RR, which is the decile ranking of RR that we orthogonalize to obtain R_RR.

22 We obtain similar results when we exclude operating cash flows.

23 We are not the first to evaluate impairments, restructurings and special items within the context of accounting conservatism. In particular, Givoly and Hayn (2000) evaluate the time-series properties of the variable ‘non-operating’ accruals, which includes impairment charges, restructuring charges and other special items. They use the accumulated amount of aggregate non-operating accruals over time to evaluate temporal variation in the degree of accounting conservatism.

24 Impairments, restructuring charges and most special items are continuous variables censored at zero. Hence, we also estimate Equation (5) as a Tobit model. When doing this we set positive amounts (i.e. gains) equal to zero and we use the absolute values of negative amounts. The untabulated results lead to the same conclusions as the results discussed in the main text.

25 Because the dependent variable is a function of impairment losses, we add impairment losses back to net income before extraordinary items when calculating ROA.