Self-sorting, incentive compensation and human-capital assets

Abstract

Skilled labour has gained significance as a production factor in the age of information technology, but accounting does not recognize human capital as an asset that contributes to the firm's earning power. This paper suggests a method to develop a latent index to proxy the managerial-skill component of human capital. The proposed index depends on the empirical validity of self-sorting theories for managerial tasks and the choice of the type of at-risk (i.e. outcome-contingent) compensation contract. The empirical analysis uses data on compensation of executive members of the board of directors, their personal attributes (experience, risk aversion and wealth), firm-specific variables (profitability growth rates, organizational complexity and operating risk), and type of industry. The extent to which equity markets value the predicted labour skills shows that investors in the marketplace recognize human capital even though accounting does not. The valuation coefficient on the variable imputed for human capital is significant for all years examined. This study contributes to the literature by showing that relative incentive compensation (incentive pay per dollar of fixed salary) is a viable surrogate for human capital defined as the skills embodied in people.

ACKNOWLEDGEMENTS

Special thanks go to Ira Horowitz and two anonymous reviewers for making valuable comments on an earlier draft of the paper. Numerous helpful comments were received from Bipin Ajinkya, Sasson Bar-Yosef, Allen Blay, Sanjeev Bhojraj, Keejae Hong, Adel Ibrahim, Ken Koga, Avi Kohl, Sri Ramamoorti, Theodore Sougiannis, Shyam Sunder, Hong Xie, David Ziebart, and workshop participants at the University of Florida, Florida State University, the University of Kansas, the University of Illinois at Urbana-Champaign, the University of Illinois at Chicago, and the Hong Kong University of Science and Technology. Earlier drafts of this paper were also presented at the Fourth European Conference on Capital Markets, Cyprus, October 2002; the Big-Ten Faculty Research Consortium held at the University of Iowa in June 2001; the European Accounting Congress held in Munich, Germany, in May 2000; and the Third Conference on Corporate Governance at the Chinese University of Hong Kong in 1999. Also special thanks go to the Editors of this Special Issue on Intangibles, Baruch Lev and Stefano Zambon.

Notes

¹The Cobb–Douglas production function was developed for macro (aggregated) data, although much of the literature has used this form for micro-data. Simon (1979) discusses the biases of estimating a Cobb–Douglas function and suggests that the homogeneity of the function is a statistical artefact. However, I use this form in this study as a starting point to develop an empirical model for relative incentive pay.

²The concept of relative incentive compensation might be illustrated by two celebrated real-life cases with each attempting to use the equivalent of RIC to induce executives to reveal their talents. The first is the case of Lee Iacoca at the Chrysler Corporation. In the 1980s he accepted a one-dollar annual salary and stacked all his expected compensation on stock options, a calculated decision that later earned him millions of dollars in stock-option awards. The second case is that of Michael Eisner at Disney. His first contract (Crystal, 1990: 354) consisted of the following components: a base salary of $750,000 a year; 2% of profits in excess of a 9% ROE (rate of return on equity) level; and an option on about 2 million common shares at a strike price of $14.00 a share with a ten-year exercise period. In his fourth year, Eisner earned an annual bonus of $7 million and the value of his options reached $104 million (Crystal, 1990: 355).

³Tangible assets are used as measures of physical capital. The accounting measures of assets, however, are measured by mixed attributes containing historical cost, current cost and estimated present values.

⁴The deferral of exercising vested and in-the-money options may be viewed as providing a joint signal of risk taking and expectations of future profitability. In this paper, I use only the former. The latter implication is the subject of another study.

⁵This analysis is replicated using random samples for the estimation and prediction periods T and P.

⁶I had expected that the risk preference scores would differ between the CEOs and Oexecs. However, the data do not bear out this expectation. Examination of this issue is deferred for a future study.

⁷I used STATA in all the empirical analyses reported in this study. The preferred test for collinearity is based on the degree to which the variance is inflated due to the collinearity of variables. This index is discussed in Chatterjee and Price (1991). I used STATA for all the analyses.

⁸It would be desirable to estimate the coefficients μ₁ and μ_2q for each firm separately. However, a long time series is required for estimation of firm-specific parameters. The cross-section analysis, therefore, is only a representation of average results.

⁹

• The reason for using the log-transformation to begin with in the previous drafts is due to the non-linearity of the models used in Barth et al. (1998) and others. Testing for non-linearity is based on the likelihood ratio of estimating models under different specifications. For example, consider the BoxCox regression:

  where for any of the explanatory variables, say it is x, the transformed value x ^(λ) is measured by x ^(λ) = (x ^λ − 1)/λ; if λ = 1, the model is linear, and if λ = 0, the model is log-linear; otherwise, the model is non-linear of a different form.

• To illustrate the test carried out in this paper, a comparison for the valuation model for P = 1999 using the predictions based on T = 1996 is shown below:

  Table

  Likelihood ratio test of λ = 0, χ² = 3.6 (p = 0.058); λ = 1, χ² = 2,317 (p = 0) (see Greene, 2003; all estimation and tests are carried out using STATA).

¹⁰The BoxCox transformation requires that all variables be greater than zero. To satisfy this property, I added a constant to each variable. Adding a constant does not alter the results of estimation.

¹¹The tables presenting the results of the random samples are available from the author on request.