The Impact of Accounting Standards on Pension Investment Decisions^‡

‡ Paper presented at the International Accounting Standards Board Research Forum, Brussels, Belgium, 28 November 2017.

Abstract

This study analyzes the ‘real’ effects of accounting standards in the context of defined benefit pension plans. Specifically, we examine the impact of IAS 19R, which increases expected pension-induced equity volatility by eliminating the so-called ‘corridor method’, a smoothing device for actuarial gains and losses. Supported by interview evidence, we show that IAS 19R leads firms to reconsider their pension investment decisions. Using matched samples of treatment and control firms, results from multivariate difference-in-differences tests indicate that firms affected by the adoption of IAS 19R significantly shift their pension assets from equities into bonds, relative to control firms. This effect is attenuated for firms with larger and better-funded pension plans. Supplementary analyses suggest that this shift in pension investment is mainly due to IAS 19R’s changes in the accounting for actuarial gains and losses (the ‘OCI method’). These results are robust to several sensitivity tests, although endogeneity concerns cannot be fully ruled out. Our study informs accounting standard setters and other stakeholders of potential shifts in firms’ real economic activities due to concerns about pension-induced equity volatility.

Acknowledgments

We sincerely thank Hervé Stolowy, the Editor, two anonymous reviewers, Nick Anderson (discussant), Saverio Bozzolan, Stefano Cascino (discussant), Katharina Hombach, Anne McGeachin, David Oesch, Eddie Riedl, Katherine Schipper, Harm H. Schütt, Tom Scott, workshop and conference participants at Boston University, Georg-August-Universität Göttingen, LMU Munich School of Management, LUISS Guido Carli, Hochschule St. Gallen, University of Parma, University of Zurich, Siemens AG, the 2016 Annual Congress of the European Accounting Association in Maastricht, the XII. Workshop on Empirical Research in Financial Accounting at the University of Exeter, and the IASB Research Forum 2017 for helpful comments and suggestions. All remaining errors are ours.

Supplemental Data and Research Materials

Supplemental data for this article can be accessed at http://dx.doi.org/10.1080/09638180.2018.1461670.

Notes

‡ Paper presented at the International Accounting Standards Board Research Forum, Brussels, Belgium, 28 November 2017.

1 We thank an anonymous reviewer for stressing that IAS 19R requires additional pension-related disclosures that could also influence affected firms’ de-risking behavior through a ‘real effect’ (e.g. Kanodia, 2006; Leuz & Wysocki, 2016). However, since these new disclosure requirements affect both treatment and control firms, our difference-in-differences research design should eliminate their effects. Hence, although we consider the de-risking effects of pension disclosures to be an important research question, the present setting is not particularly conducive to studying them.

2 Actuarial gains and losses arise for DBOs and plan assets if firms experience adjustments to actuarial assumptions, or change these assumptions for future periods. Actuarial assumptions relating to the DBO include the discount rate, mortality rates, and salary trends. For example, actuarial losses (gains) result when the pension discount rate decreases (increases), which increases (decreases) the DBO.

3 Method (3) was virtually nonexistent in the initial sample (see Table 1 and Glaum et al., 2018).

4 In 2012, treatment firms’ mean unrecognized actuarial losses were €981 million, or 10% of equity book value.

5 The ERR effect is difficult to calculate precisely from public disclosures. For example, Volkswagen reports separate discount rates and expected rates of return on plan assets for Germany versus all other countries, but does not provide the same disaggregation for the DBO and the fair value of plan assets. However, assuming that Volkswagen replaced the 2012 expected rate of return on plan assets for Germany (4.12%) with the 2013 discount rate for Germany (3.70%), and applied the difference (0.42 percentage points) to the fair value of plan assets at the end of 2012 (€7288b), this would have led to a pension expense that would have been higher by €31m, or 0.04 of book equity. Clearly, however, the magnitude of the ERR effect is driven by (1) a firm’s fair value of plan assets, and (2) its spread between the discount rate and the expected rate of return on plan assets.

6 In Section 4.3, we consider whether IAS 19R’s ERR effect is associated with shifts in firms’ pension asset allocations. This test amounts to a replication of Anantharaman and Chuk (2017) for the German setting.

7 These papers are complemented by several studies that document accrual earnings management in the pension context, typically using actuarial assumptions (e.g. see Glaum, 2009 for a review).

8 Mashruwala (2008) reports similar findings in the UK setting by documenting that UK firms reduce equity allocations by approximately 8 percentage points following the introduction of FRS 17.

9 As discussed in Section 4.3, we cannot cleanly separate the OCI and ERR effects in our main setting. However, we complement the main tests with alternative analyses that allow isolation of the ERR effect. (We discuss an empirical approach to isolating the OCI effect, but data availability constraints prevent us from implementing it.)

10 Refer to Online Appendix 2 for summary information on these interviews. We present this evidence in the spirit of Gow, Larcker, and Reiss (2016, p. 479), which argues ‘that evidence on the actions and beliefs of individuals and institutions can bolster causal claims based on associations, even absent compelling estimates of the causal effects’.

11 Whereas German corporate law restricts dividend payments to the amount of retained earnings calculated under German GAAP, shareholders may form expectations about dividends on the basis of published IFRS financial statements. Also, as indicated by one interviewee, corporate bylaws and charters occasionally make dividend distribution conditional on maintained minimum ratios of book value of equity to total assets.

12 Alternative (and potentially costlier, and longer-term) ways of reducing the financial statement risks of pension plans include settlement payments, termination/freezing of existing pension plans, and risk transfers to insurance companies.

13 Again, refer to Section 2.1 for a detailed comparison of the two methods of recognizing actuarial gains and losses, and Online Appendix 1 for a numerical example.

14 In Online Appendix 3, we further discuss the German regulatory environment in terms of three distinct institutional factors that prior studies have linked to variation in pension sponsors’ plan asset allocations: (1) restrictions on funding agencies’ investment strategies; (2) funding requirements and insurance; and (3) taxation. The purpose of that discussion is to show that these institutional factors are unlikely to explain the empirical patterns we observe in the context of IAS 19R adoption.

15 We suppress subscripts in the subsequent text. See Appendix A for detailed variable definitions.

16 Alternatively, we test for a non-linear relation between Fund and %EQ as well %BONDS in Section 4.4.4.

17 In Section 4.2, we discuss results for an alternative measure of pension plan exposure.

18 In Section 4.4.1, we discuss robustness tests using alternative time windows.

19 Both measures capture (i) the sensitivity of firms’ equity book values to pension-induced volatility and, simultaneously, (ii) the degree to which firms can mitigate that volatility by adjusting pension asset allocations.

20 Specifically, for results to be explained by an omitted variable, that variable would have to vary contemporaneously with IAS 19R adoption, and affect treatment and control firms differently.

21 Data constraints preclude us from including cumulative unrecognized actuarial gains and losses, another likely factor of influence. This is due to cumulative actuarial gains and losses not arising under the OCI method. However, our interview suggests that cumulative actuarial gains and losses were an important determinant of the choice of corridor versus OCI method only for early adopters switching to the OCI method in or shortly after 2005. Further, comprehensive evidence in Glaum et al. (2018) documents that cumulative actuarial gains/losses explain firms’ OCI adoption decisions in 2005 – but not between 2006 and 2013. We thus believe that omitting cumulative actuarial gains/losses from a selection model estimated in 2011 will not result in severe omitted variables problems.

22 We also carry out robustness tests using caliper matching in Section 4.4.3.

23 The Prime Standard segment comprises 340 listings representing 319 unique firms. (Several firms have both common stock and preferred stock outstanding; these count as separate listings.) We select this particular market segment, as it imposes transparency standards that go beyond EU minimum requirements. Besides reporting under IFRS, these firms published quarterly reports in German and English during our analysis period. However, as firms usually do not disclose the pension asset allocation in their interim reports, observations are restricted to one per year. Furthermore, firms listed outside of this segment often lack significant pension plans and typically do not provide sufficient information on pension asset allocations.

24 In robustness tests discussed in Section 4.4.1, we vary the event window between 2010–2014 and 2009–2013.

25 Note that, during the analysis period, mandated disclosure requirements on the composition of plan assets were limited to a disaggregation of the percentages or amounts of ‘equity instruments, debt instruments, property, and all other assets ’ (IAS 19.120A (j)) and ‘amounts included in the fair value of plan assets for  …  the entity’s own financial instruments; and  …  property occupied by, or other assets used by, the entity’ (IAS 19.120A (k)). See Online Appendix 4 for examples of the various degrees of granularity in pension asset disclosures.

26 To formally test this assumption, we estimate a model where %EQ is the dependent variable, and the independent variables are those in Equation (1). We include year dummies for the pre-treatment period (i.e. 2010) and the post-treatment periods (i.e. 2012 and 2013), and omit the dummy (i.e. 2011) for the year before the actual treatment date (i.e. 2012). In addition, we interact these year dummies with TREAT. Untabulated results reinforce the validity of the parallel-trends assumption, as the coefficient on the interaction term TREAT  × 2010 is statistically insignificant. We thank an anonymous reviewer for this suggestion.

27 Panel A of Table 2 is based on the matched sample. Note that matching eliminated previous significant covariate differences between the treatment and control firms. As Table OA.1 in the Online Appendix shows significant pre-matching differences between the treatment and control firms (particularly in the pre-treatment period), propensity-score matching achieves its main objective of covariate balance between the treatment and control observations.

28 Table 3 reports difference-in-differences results based on the matched sample. Online Appendix Table OA.2 provides the results for the probit model used to match treatment and control firms. These results are robust to including %EQ (contemporaneous and lagged) in the probit model as an additional explanatory variable (untabulated). We thank an anonymous reviewer for this suggestion. Difference-in-differences results without matching (untabulated) lead to the same inferences as those reported in Table 3. However, note that covariate balance is limited in the unmatched sample.

29 The pre-treatment difference between treatment and control firms is captured by TREAT, which is negative but not significantly different from zero (coefficient = –1.36; z‐statistic = −0.43).

30 These results also hold when we calculate Exp as defined benefit obligation (rather than by pension assets, as in the main analyses) divided by the book value of equity, to capture the exposure of firms’ book value of equity to the size of pension plans for firms with relatively low pension assets but – at the same time – relatively high pension obligations. Untabulated results are consistent with the findings reported in the main analysis.

31 In addition to the Table 3 analyses, we implement the test of H₂ by estimating Equation (3) after splitting the sample based on the median value of PP_CHAR, which is either Exp or Fund. Results reported in Table OA.3 in the Online Appendix are consistent with those reported in Table 3. Specifically, we find significantly negative (positive) coefficients on POST × TREAT for firms with below-median PP_CHAR where the dependent variable is %EQ (%BONDS). In contrast, coefficients on POST × TREAT are insignificant for firms with above-median PP_CHAR (except for PP_CHAR = Fund in the %BONDS model).

32 These results suggest that treatment and control firms are closely balanced along these variables.

33 We repeated the analysis reported in Table 3 after including year fixed effects and industry-by-year fixed effects, and the results are qualitatively similar.

34 Due to lack of granularity in firms’ pension asset disclosures, we are unable to test explicitly whether treatment firms adjust their derivatives-based pension de-risking activities around the adoption of IAS 19R. Inspection of pension footnotes indicates that only a small portion of our sample discloses the proportion of pension assets representing derivatives. Even for those firms, the specific types of derivatives, their attributes, and whether they are intended for pension de-risking, are ultimately opaque.

35 Overall, these findings are robust to including SHIFT, an indicator variable equal to one if firm i terminates or freezes existing defined benefit pension plans, or transfers risk to insurance companies, in year t, and zero otherwise, as an additional control variable, following Amir et al. (2010). However, the three-way interactions are insignificant when PP_CHAR is Fund (refer to Section 4.4.4). The purpose of including SHIFT is to control for firms altering or closing their defined benefit pension plans as an alternative means of mitigating the increase in pension-induced equity volatility brought about by IAS 19R.

36 For sample selection details, refer to Table OA.4 in the Online Appendix.

37 Full summary statistics are reported in Table OA.5 in the Online Appendix.

38 Untabulated results for H₁ suggest that Post × TREAT is insignificant in the %EQ model, whereas it is negatively significant in the %BONDS model.

39 We repeated the analysis reported in Table 4 after including year fixed effects and industry-by-year fixed effects, and the results are qualitatively similar, with the exception of the coefficient on Post × TREAT, which is positive and marginally significant in the %EQ model in column (1) when including industry-by-year fixed effects.

40 Note this test is largely comparable with the main test in Anantharaman and Chuk (2017) – with one exception: Whereas our treatment group in this test consists solely of OCI method firms, Anantharaman and Chuk’s (2017) main test includes corridor firms in the treatment group, effectively mixing the ERR and OCI effects. Anantharaman and Chuk (2017) include an indicator variable, e.g. CORRIDOR, to control for the alternative options for recognizing actuarial gains or losses under IAS 19; they note that the CORRIDOR indicator variable is insignificant, ‘suggesting that any balance sheet effects for firms using the corridor method for balance sheet recognition prior to IAS 19R are not substantial’ (p. 28).

41 Furthermore, Anantharaman and Chuk’s (2017) research design choices vary to some extent from ours, including in the choice of matching algorithm, measurement windows and control variables.

42 In untabulated analyses, we conduct another robustness test, where, we define 2009–2010 as the pre-treatment period and 2013–2014 as the post-treatment period. Overall, the results are very similar (with coefficients slightly larger) to those reported in the main analysis when PP_CHAR = Exp. However, the coefficients of interest are only significant for PP_CHAR = Fund in the %BONDS specification.

43 Pearson and Spearman correlation coefficients between Fund and Fund² are significant and above 0.96 and those between FF and FF² are significant and above 0.99.

44 However, the Post × TREAT and Post × TREAT × Fund coefficients are insignificant in column (3) where PP_CHAR is Fund. The Post × TREAT × Fund is also insignificant in column (4). Given the high correlation between Fund and Fund², this regression specification suffers from multicollinearity.