Criticisms of the Feltham-Ohlson and the Ohlson model

applicability in value-relevance research. Value-relevance studies do not use the Feltham-Ohlson model for equity valuation, but as basis for regression models whose purpose is to test the valuation usefulness of accounting numbers (e.g.

Barth et al. 2001). Value-relevance research is not motivated by equity valuation per se, but motivated to give standard-setting implications on the valuation usefulness of accounting numbers. The Feltham-Ohlson model does not give any implications for accounting standard-setting as any set of accounting methods meeting the clean-surplus assumption will encompass the model. This suggests that the strength of the model seen from a fundamental-analysis perspective is a limitation when seen from a value-relevance perspective.

This has led researchers to question the use of the Feltham-Ohlson model as justification for value-relevance regressions (Ohlson 1995, Bernard 1995, Holthausen and Watts 2001, Kothari 2001). Barth et al. (2001) acknowledge that the model itself does not give any implications for accounting-method choices.

Still, they do not think this undermines the usefulness of the model for standard-setting: “(…) none of the valuation models explicitly derive an optimal accounting system or even a demand for accounting information, this does not preclude use of such models to assess the value relevance of accounting amounts and to provide insights relevant to standard setters, as HW [Holthausen and Watts] claims”

(Barth et al. (2001:92). In a footnote, Holthausen and Watts (2001:61, footnote 20) give a response to this argument: “We agree that the model can be used to assess associations between equity value and accounting numbers, but that is not the point we are making. Our point is that the model itself has no implications for accounting methods and provides no direct inferences for accounting standards.”

Beaver (2002) claims that the criticism stated by Holthausen and Watts (2001) is misplaced and misdirected. He argues that “(...) the modelling could be

informative without including an endogenous demand for accounting. By analogy, the Capital Asset Pricing Model (CAPM) has no demand for financial institutions, yet we observe financial institutions empirically. What do we conclude? Do we conclude that the risk-return trade-off derived from the CAPM is of no interest or relevance to investors or to managers of financial institutions? I think not”

(Beaver 2002:458). The model provides a framework for valuation based on accounting numbers. As Beaver (2002:458) states: “This framework relates published accounting data to equity valuation (…).With contextual accounting arguments added to the general framework, researchers can predict how accounting numbers would relate to value (…).”

In order to derive the Ohlson model, additional assumptions regarding the information dynamics are needed to specify the time-series pattern of abnormal earning and non-accounting information. These information dynamics are also essential to the empirical applicability of the model beyond the Feltham-Ohlson model and the dividend-discount model. These additional assumptions make it possible to derive a link between current earnings and book-equity values and future abnormal earnings (Ohlson 1995, Lo and Lys 2000). Dechow et al. (1999) conduct an empirical analysis of the linear dynamics of abnormal earnings. Using a pooled regression of all the firm observations with one period lag, the persistence parameter equals 0.62. The persistence is far from its limits of 0 and 1, suggesting that stock prices are jointly explained by current net earnings and book equity. Thus, neither a balance-sheet model nor an earnings model seems appropriate to explain variation in stock prices. In the second part of the article they examine variables that may affect the persistence of abnormal earnings across firms and over time such as high levels of earnings, extreme accounting rates of return, high operating accruals, high payout ratios of dividends, high levels of

non-recurring items and industry-specific factors. The analysis reveals that all the determinants are statistically significant, suggesting that the persistence parameter varies cross-sectionally and time-serially as a result of firm-specific and industry-specific characteristics. Thus, the information dynamics are not completely captured by the simple autoregressive model presented by Ohlson (1995). Lo and Lys (2001), however, argue in the spirit of Roll’s critique (Roll 1977) that the test of the Feltham-Ohlson model and the Ohlson model is a joint test of the models’

assumptions on the one hand and whether the model is descriptive of the market pricing of stocks on the other. Kothari (2001) takes the same position and concludes that the evidence rejecting the information dynamics is weak.

Other aspects of the Ohlson model also question its applicability. First, the model and its regression counterparts are built on the assumption of linearity. This assumption is generally violated if there are omitted variables which are correlated with the independent variables. Potential candidates are variables affecting the persistence of abnormal earnings. Holthausen and Watts (2001) argue that nonlinearity could be due to growth options, abandonment options and conservatism. For instance, Hayn (1995) investigates the information content of positive and negative earnings. She reports that negative earnings are less informative than positive earnings and maintains that this is due to the abandonment option held by the shareholders. The shareholders can always liquidate the firm rather than suffer from indefinite losses. A similar point is made by Collins, Maydew and Weiss (1997) and Ball and Shivakumar (2006).

Barth et al. (2001) claim that potential nonlinearity problems due to growth options and abandonment options can be handled within the existing Ohlson model. The growth options, termed economic rents in their article, are captured by

the persistence parameter of earnings,Z, and the non-accounting information parameter,_X. In the regression counterpart of the Ohlson model the present value of future cash flows not attributed to book equity can be used as a proxy for future growth options. They also claim that intangible assets such as customer lists, brand names and research and development costs are attributable to growth options. These suggestions, however, do not seem to solve the problem addressed by Holthausen and Watts (2001). Expected future cash flows are generally uncertain and unobservable, and any allocation of cash flows between book equity and other net assets not recognised on the balance sheet will most likely be arbitrary. Another way to counter the problem of nonlinearity is to allow the regression coefficients to vary cross-sectionally and time-serially, using a fixed effects regression model. This approach will control for correlated omitted variables that are associated with particular firms or reporting periods and potentially maintain linearity within each partitioning.

A different approach might be used to control for growth options and abandonment options (Barth et al. 2001). Growth options will probably be associated with industry membership and the intensity of intangible assets such as goodwill. This suggests that the inclusion of industry dummies and proxies for growth might control for growth options. Similarly, abandonment options will be strongly associated with weak economic performance. Including proxies of financial health will potentially control for these options (Barth et al. 2001). Like growth and abandonment options, conservatism is another reason for a nonlinear relationship between accounting numbers and stock prices. Conservatism refers to the fact that losses are generally recognised before profits in the profit and loss account. For instance, Basu (1997) defines conservatism as an accounting principle making earnings reflect “bad news” more quickly than “good news”,

which has consequences both for the timeliness and persistence of net earnings.

Consistent with this, he reports that the earnings-response coefficients for positive earnings changes are higher than the earnings-response coefficients for negative earnings changes. This suggests that the association between earnings and stock prices is nonlinear, rather than linear. Barth et al. (2001) argue that the Ohlson model can handle conservatism. Subsequent refinements of the initial Ohlson model explicitly model the effects of conservatism (Feltham and Ohlson 1995, 1996). Moreover, the size of the coefficient on asset, liability and equity numbers might be interpreted as the degree of conservatism in those numbers. A lot of value-relevance studies try to explain why equity-market values exceed equity book values. These studies can be seen as attempts to examining conservatism in accounting (Barth et al. 2001).

A final concern is that value-relevance studies assume assets to be additively separable (Holthausen and Watts 2001). Lack of separability is one of the important characteristics of goodwill. As discussed in section 2.1, goodwill consists of economic assets that are inseparable from the firm. There is no active market where goodwill is traded, and hence, goodwill is not additively separable from other assets in the firm. Barth et al. (2001) argue that lack of separability does not lead to any problems. The regression coefficient on inseparable assets, such as goodwill, captures the incremental association with stock prices beyond that of other assets and liabilities (Barth et al. 2001).

In sum, it is debatable whether the value-relevance regressions can be justified by reference to the Feltham-Ohlson and Ohlson model. The reason is the weak link between the theoretical valuation models on the one hand and the regression specifications used to test value relevance on the other. If the Ohlson model is

used as theoretical justification, at least some caution should be exercised when it comes to potential correlated-omitted variables and nonlinearity problems. Such problems will potentially bias the ordinary-least-square regression coefficients, t-statistics and R-square estimates which may lead to misinterpretations of the regression results.