Journal of Productivity Analysis

, Volume 33, Issue 1, pp 67–80

The productivity effects of stock option schemes: evidence from Finnish panel data

Authors

  • Derek C. Jones
    • Department of EconomicsHamilton College
  • Panu Kalmi
    • Department of EconomicsHelsinki School of Economics and HECER
    • Department of EconomicsHelsinki School of Economics and HECER
Article

DOI: 10.1007/s11123-009-0146-6

Cite this article as:
Jones, D.C., Kalmi, P. & Mäkinen, M. J Prod Anal (2010) 33: 67. doi:10.1007/s11123-009-0146-6

Abstract

In this study we investigate the productivity effects of employee stock option schemes. We estimate Cobb-Douglas production functions by using new panel data for all Finnish publicly listed firms during 1992–2002. The data enable us to distinguish broad-based option plans, for which all employees are eligible, from those selectively allocated to particular employees. For both type of schemes, our baseline fixed effects estimators consistently find statistically insignificant associations with firm productivity. When endogeneity of production inputs and option-schemes are taken into account we continue to find no evidence of a link with firm productivity. Our main findings are consistent with hypotheses that predict negligible effects of option plans for enterprise performance, such as those based on free riding, psychological expectancy theory, accounting myopia, or rent-seeking. We consider reasons why our empirical findings on the impact of broad-based options differ from those found in earlier studies.

Keywords

ProductivityPanel dataEmployee stock optionsCompensation and compensation methods and their effects

JEL Classification

D24C33J33M52

1 Introduction

During the 1990s, employee stock options became an increasingly popular compensation method in many countries (e.g., Hall 1998; Murphy 1999). Initially, option programs were typically allocated only to executives. But this association of options mainly with managerial compensation changed rapidly after other companies started to issue stock options to their workforce more broadly (e.g., Blasi et al. 2003).1

While the growth of options has also been accompanied by a mushrooming of literature, the economic impact of stock options remains a contentious issue. For example, an important survey on the impact of executive equity compensation on firm performance found that “[T]here is presently no theoretical or empirical consensus how stock options and managerial equity ownership affect firm performance” (Core et al. 2003, p. 34). In contrast, existing empirical work on broad-based stock options consistently finds that firm performance is enhanced by stock options. This has led some observers to conclude that it is economically desirable to extend stock options to a large segment of employees (Rosen 2006). However, previous empirical analyses of broad-based stock options have relied on survey data for a specific industrial sector, typically are for only short time periods and are mainly for the US and the UK (e.g., Sesil et al. 2002; Ittner et al. 2003; Conyon and Freeman 2004). Moreover, existing data may not enable a careful investigation of the productivity effects of different types of option plans (selective vs. broad-based).

In this paper we assemble new panel data including all Finnish publicly listed firms during a relatively long period, namely 1992–2002. This enables us to see if previous findings based mainly on evidence generated using US and UK data, are applicable to another country which once had a very different system of corporate governance but which has evolved recently in the direction of the Anglo-Saxon model.

We estimate a range of Cobb-Douglas production functions with different option program indicators. Whereas earlier empirical literature has used cross-section and fixed effects estimators, we also use GMM estimators to address the potentially important issue of endogeneity of production inputs and option schemes. For both type of stock option schemes, our baseline fixed effects estimators consistently find a statistically insignificant association with firm productivity. When endogeneity of production inputs and option-schemes are taken into account, our key finding persists—we continue to find no evidence of a statistically significant link with firm productivity. Hence our empirical findings do not provide support for theorists who predict that potentially powerful economic incentives effects will flow from options and dominate the effect of other factors such as free riding, accounting myopia, line-of-sight and managerial rent-seeking.

However, to some extent the divergence in results may reflect differences in institutional arrangements. For example, because of the Finnish collective bargaining agreements, employee base wages were protected from downward falls and employers could not use stock options as a substitute for other compensation, except possibly for top management. The risks employees incurred were further reduced by the facts that they usually received the options for free or for nominal compensation, and the options could be traded in secondary markets once they vested (this is not possible in the US). Thus, the Finnish option schemes have exposed employees to much smaller risks than the US option schemes, and this may have had behavioral implications. Further, while around 40% of Finnish stock option schemes included a dividend protection clause (the strike price for shares is adjusted downwards by the amount of dividends paid), in the US this was the case in only 1% of cases (Liljeblom and Pasternack 2006). Conceivably some of these differences might be expected to have a bearing on the effectiveness of options in the two contexts.

The paper is organized as follows. Section 2 provides the conceptual framework and surveys relevant research. Section 3 describes the data. Section 4 outlines empirical strategy. The final section concludes.

2 Conceptual framework and previous empirical work

By examining existing theoretical and empirical work we identify two key research hypotheses. The first concerns the impact of options in general on enterprise performance. The second relates to whether or not a plan being broad-based or selective is expected to have different implications for firm performance. While we continue by focusing on the first hypothesis, it is clear that many of the points we discuss apply to different degrees to both hypotheses.

The theoretical literature concerning the expected impact of options in general on business performance is quite ambiguous. In one camp are those who expect business performance to be enhanced, since options may align the interest of employees and shareholders—e.g., stock options may motivate employees to exert more effort and take actions that are mutually beneficial to both owners and employees.

Other papers view stock options as a substitute for fixed wage contracts. Inderst and Müller (2005) show that stock options may prevent inefficient firm closures, since they substitute for fixed wage contractual payments and hence save firms cash expenses in bad times. Arya and Mittendorf (2005) argue that when a manager is willing to accept stock options as a part of her compensation package (and, by assumption, a lower fixed wage), she is signaling her confidence in her ability (managerial quality) to raise firm value.

Stock options have also been deemed crucial in recruiting and retaining employees, especially in markets where employees are potentially highly mobile (Rousseau and Shperling 2003). In addition, stock options may also help to retain key employees, both because they adjust pay according to current market conditions (Oyer 2004) and because they are a deferred form of compensation.

However, there has also been criticism of stock options. For example, stock options, when exercised, entail a cost to shareholders in the form of dilution of ownership. In addition, others argue that the increasing popularity of options in part reflects firms’ mistakenly thinking that options are a cheaper form of compensation than their true cost, since the costs of options have not been included in income statements (e.g., Hall and Murphy 2003).

Others question the potential performance impact of options. One argument is payment schemes that reward collective performance suffer from the free-rider problem (e.g., Alchian and Demsetz 1972). Another criticism comes from psychological expectancy theory (Vroom 1995); according to the “line-of-sight” argument, rewards based on performance can only be motivating if, by their actions, employees can influence the measures on which performance-pay is based. This is typically not the case with stock option plans, where employees (with the possible exception of top executives) can hardly perceive any direct link between their actions and the share price performance.

Another cost related to equity schemes is that employees face increased risk, if they have a substantial proportion of their financial and human capital invested in one workplace. Since risk-averse employees do not value their options to the same extent as an outsider would, they may require higher total compensation (Meulbroek 2001). However, there can be important differences between top executives and employees in these respects. In the compensation of executives, the line-of-sight and diversification problems may not be as severe as with employees at lower levels within the organization (Hall and Murphy 2003). On the other hand, executive stock option compensation may be motivated by rent-seeking activities (Bebchuk and Fried 2003).

Thus, ultimately, the impact of stock options on firm performance is an empirical question. And most previous work has concentrated on the performance consequences of executive stock options. For instance, Hanlon et al. (2003) find that changes in Black-Scholes values of option grants are positively associated with future operating income of the firm. However, using similar data, Larcker (2003) find that results depend crucially on estimation strategy. Core et al. (2003) conclude that the research consensus is that there is no clear connection between executive equity compensation and firm performance.

Our second hypothesis concerns the expected impact of the different type of option plans. As with the first hypothesis the theoretical literature is ambiguous. For example, the free-rider and line-of-sight arguments might be expected to be especially convincing when a larger fraction of the workforce gains from a broad-based scheme. On the other hand employees are much more likely to act against shirkers when they have a financial interest in the company (Freeman et al. 2008). Also, when rewards are based on group performance, it is in the interest of employees to develop a group norm to monitor the performance of their peers and prevent free-riding behavior (Kandel and Lazear 1992). These arguments are especially applicable to broad-based schemes and are expected to produce anti-shirking behavior by employees and, ultimately, to result in enhanced firm productivity.

As before the issue of the impact of broad-based versus selective stock options schemes on firm performance is an empirical question. We are aware of two studies that analyze the effects of broad-based options with samples of listed firms from various industries, Conyon and Freeman (2004) for the UK and Sesil et al. (2000) for the US. Both of these studies find a significant positive association between the presence of broad-based stock option plans and firm productivity. Two other studies analyze broad-based options in US new economy firms. Sesil et al. (2002) find evidence that productivity is higher in firms with broad-based stock option plans, and Ittner et al. (2003) find that lower than expected option grants and/or existing option holdings are associated with lower accounting and stock price performance in subsequent years. In sum, the existing empirical evidence suggests a positive and often quite sizeable link between broad-based stock option plans and firm productivity.

While there are only a small number of empirical studies on the impact of stock options on firm productivity, the empirical literature on the productivity effects of other forms of employee financial participation, namely profit-sharing and employee stock ownership plans, is large. For profit-sharing schemes, several surveys (e.g., Weitzman and Kruse 1990; Jones and Pliskin 1991) conclude that usually there is a positive relationship between profit-sharing and firm productivity. This finding emerges from empirical studies that employ diverse methods to investigate profit sharing arrangements that exist in a variety of institutional settings (e.g., Cable and Wilson 1989; Wadhwani and Wall 1990; Kruse 1992). There is also empirical evidence that employee stock ownership can be positively associated with enhanced business performance in a variety of institutional settings (e.g., Kumbhakar and Dunbar 1993; Jones and Kato 1995).

3 The data

The data we assemble are for publicly listed firms traded on the Helsinki Stock Exchange (HEX). Our data on stock options were initially organized by an accounting professor from the Helsinki School of Economics, but we have used several sources to complement and update the original data. These include annual reports, stock market releases and option data obtained from an investment bank. Next we briefly describe the general evolution of stock option plans in Finland during 1992–2002 (see Jones et al. (2006) for more detailed evolution and institutional background in Finland).2

Column (1) in Table 1 gives the total number of firms at the HEX during the period. As can be seen, the number fluctuates considerably, which partially relates to the business cycle. Column (2) describes the number of new option plans. The early peak year is 1994 (21 new plans). Then the number increases from 1997 (22) until 2000 (61). Thereafter it drops to 33 for 2001 and 2002. Column (3) shows the development of new broad-based plans. The introduction of such schemes is concentrated during the years 1999–2000, when one-half of new plans were broad-based. Columns (4) and (5) provide data on the existence of option schemes. In Column (4), a quarter of firms had an option scheme in 1993. This proportion jumps to around 50% in 1994, where it stays until 1996. After a temporary drop in 1997, the proportion increases from 1998 (58%) until 2001 (77%). In 2002 74% of firms had an existing scheme.3 Column (5) shows the development for broad-based plans. Roughly 3–8% of firms had an existing broad-based scheme in 1992–1997. This fraction steadily increases until 1998–1999, and stays around 36% in 2000–2002.
Table 1

The evolution of Finnish stock option plans 1992–2002

Year

(1)

No. of firms in Helsinki stock exchanges

(2)

No. of new option plans

(3)

No. of new broad-based option plans

(4)

No. of firms having option plan

(5)

No. of firms having broad-based option plan

1992

65

1

0

11

(16.9%)

2

(3.1%)

1993

60

6

1

15

(25.0%)

2

(3.3%)

1994

68

21

2

34

(50%)

3

(5.0%)

1995

74

7

1

38

(51.4%)

3

(4.1%)

1996

73

9

3

36

(49.3%)

6

(8.2%)

1997

115a

22a

4a

46a

(40.0%)

7a

(6.1%)

1998

119

47

17

69

(58.0%)

21

(17.6%)

1999

137

42

23

91

(66.4%)

36

(26.3%)

2000

150

61

30

113

(75.3%)

54

(36.0%)

2001

145

33

11

112

(77.2%)

54

(37.2%)

2002

137

33

6

101

(73.7%)

49

(35.8%)

Total

 

282

101

  

Note that stock option data in Table 1 also includes firms that have less than four consecutive year observations

aBefore 1997 data are only for main list firms. From 1997 onwards, the data also include the new market and the investor list firms

In the analysis that follows, we are able to distinguish between broad-based and selective option schemes. Our classification is based on public stock exchange reports which, in turn, are based on the Finnish Law on Joint Stock Companies. Since the shareholders’ general meeting (and not the board of the directors) makes the final decision on adoption, the law requires listed firms to report publicly all relevant conditions about stock option schemes prior to adoption.4 Selective schemes are mostly managerial schemes, although they can also include other key employees. In order to qualify as a broad-based scheme, all employees (or at least a great majority) should be eligible, but the schemes do not have to be egalitarian in the sense of all participants having the same number of options. While a high rate of eligibility does not automatically guarantee a high participation rate, these features are closely connected. For one thing, employees usually face only small costs when they subscribe to options—e.g., by providing a zero-interest loan to the company, with the company repaying the loan at face value after a certain period, usually 1–3 years. Thus, while employees face a cost in terms of foregone interest and liquidity, typically this cost is far below the real value of the options. Moreover, not all companies use this procedure, but rather they essentially give options to employees for free. Finally, we interviewed a partner in an investment bank, who has enormous expertise in setting up dozens of option schemes. He confirmed that there are dramatic differences in the participation rates for option schemes, depending on eligibility.

Our option program indicators measure the presence or absence of a scheme, the size of the scheme and whether the scheme is selective or broad-based. Our first indicator is opt measuring the presence of a scheme in a firm in given year t. It equals one for the group of option firms and zero otherwise. Our second indicator also measures the presence or absence of a plan but it distinguishes between selective (ssopt) and broad-based (bbsopt) plans. Our third indicator potential dilution (dilu) allows us to examine the association between the size of effective schemes and firm productivity. This is a continuous variable—the ratio of the number of shares that may be awarded through effective stock option plans in a given year divided by the sum of total number of shares and the number of new shares that may be awarded through options at the end of a year. If a program ends in the middle of the year t, then the year t − 1 is the last year used in calculating dilution. We also distinguish a size indicator between selective (diluss) and broad-based (dilubb) plans.

For several reasons, we regard potential dilution (dilu) as the preferred measure for the size of a scheme, given the available data. Since we do not have access to information on stock option program details such as exercise prices, we cannot calculate Black-Scholes values. However, Hallock and Olson (2006) find that employees may value their stock options significantly below the Black-Scholes value (see also Bergman and Jenter 2007). Using actual dilution is rather questionable because this measure is strongly endogenous (actual dilution depends on realized performance.) Also we know actual dilution only ex-post, whereas potential dilution is known ex-ante. Another reason for using potential dilution is that other Finnish researchers have used similar measures.5

By combining the option data set with financial statement data obtained from a consulting firm, we assemble firm-level panel data for 117 publicly listed firms from 1992 to 2002. We exclude only a few firms with less than four consecutive observations.6 This is mainly because there is some entry and attrition of listed firms. Also, some firms merged during the period. In this case, we have included only merged firms and excluded all information prior to the merger. Finally, since potential outliers (and influential observations) in the data can bias our estimators, we use our judgment and knowledge based on modifications of the key criteria for listing on the Helsinki Stock exchange to delete observations where: employment is less than 50 (32 firm-year observations); fixed capital is less than €1,000,000 (23 firm-year observations); or employment is more than 50,000 (4 firm-year observations).7 All nominal monetary variables are deflated by using industry-specific gross output deflators, published by Statistics Finland. Table 2 summarizes the pattern of our panel data. Table 3 presents summary statistics.
Table 2

The pattern of firm-level panel data, 1992–2002

Frequency

Percent

Cumulative

Pattern

40

34.2

34.2

11111111111

21

18.0

52.1

01111111111

12

10.3

62.4

00111111111

10

8.6

70.9

00001111111

7

6.0

76.9

00000111111

4

3.4

80.3

00000011111

3

2.6

82.9

00000001111

3

2.6

85.5

00011111111

3

2.6

88.0

01111111110

14

12.0

100.0

(Other patterns)

117

100

  

The last column describes the pattern of data: 1 means we have an observation for this year, 0 we do not

The first digit (0 or 1) in the pattern column is year 1992. Thus the first row indicates that in 40 cases we have data for all years, whereas the second row indicates that in 21 cases there is no data for 1992 but that data are available in all other years

Table 3

Summary statistics

Variable

Name

Firm- year obs

Mean

SD

l

Employees

1,042

4,066

7,123

k

Fixed capital (tan. + intan.), €1,000

1,042

464,000

1,500,000

q

Value added, €1,000

1,042

255,000

574,000

dilu*

Potential dilution in the range of (0,1); a proxy of option program size

531

0.0547

0.0450

diluss*

Potential dilution for selective stock option programs

364

0.0286

0.0285

dilubb*

Potential dilution for broad-based stock option programs

167

0.0900

0.0533

opt

Option program dummy

1,042

0.5182

0.4999

ssopt

Selective option program dummy

1,042

0.3580

0.4796

bbsopt

Broad-based option program dummy

1,042

0.1603

0.3670

ln(l)

Natural logarithm of employees

1,042

7.10

1.62

ln(k)

Natural logarithm of deflated fixed capital

1,042

17.95

2.07

ln(q)

Natural logarithm of deflated sales

1,042

17.95

1.72

All value measures are deflated using an industry-specific gross output deflator at 2000 constant Euros obtained from Statistics Finland

* Summary statistics for dilu, diluss and dilubb variables are only for those firms that have a stock option program

The total number of firm-year observations is 1042 and data are for 117 firms

Table 4 presents summary statistics for some key variables by option program adoption status. First, on average, firms with option schemes have higher value added, bigger labor forces and they also use more fixed capital compared to firms without options. For example, the mean value added for selective scheme firms is €496 million, whereas for broad-based firms it is €166 million and for non-option firms only €105 million. Second, large firms (number of employees) have preferred selective schemes to broad-based schemes. Finally, selective scheme firms have about 3.4% higher average value added per employee compared to broad-based firms.
Table 4

Summary statistics: option vs. non-option firms

Variable

Broad-based option scheme

Selective option scheme

No option scheme

Value added, €1,000

    Mean

166,000

496,000

105,000

    (Standard deviation)

(366,000)

(834,000)

(238,000)

Employees

    Mean

2,215

7,542

2,100

    (Standard deviation)

3,273

9,625

4,371

Fixed capital (tan. + intan.), €1,000

    Mean

365,000

929,000

153,000

    (Standard deviation)

1,580,000

2,130,000

466,000

Value added/employees, €

    Mean

55,206

57,064

52,191

    (Standard deviation)

21,787

18,925

21,302

Firm-year obs.

167

373

502

Based on a firm’s option program adoption status in a given year, all firms are classified into three groups, namely broad-based, selective and non-option firms

All value measures are deflated using an industry-specific gross output deflator at 2000 constant Euros obtained from Statistics Finland

4 Empirical strategy

Our empirical strategy is to use a production function approach and panel data estimators to investigate our two key research hypotheses: (1) firm-level productivity is expected to be higher in option than in non-option firms; (2) the impact of options on firm productivity is expected to be dependent upon whether the plan is broad-based or selective. First, we estimate a series of baseline fixed effects estimators by assuming that all explanatory variables are strictly exogenous. Second, we estimate dynamic panel data GMM estimators and impose (and test) the common factor restrictions to account for the potential endogeneity of production inputs and stock option schemes.

Our baseline fixed effects specification is the following Cobb-Douglas production function:
$$ \begin{gathered} y_{it} = \beta_{k} k_{it} + \beta_{l} l_{it} + \beta_{op} op_{it} + \gamma_{t} + \eta_{i} + \varepsilon_{it} \hfill \\ \varepsilon_{it} \sim iid(0,\sigma^{2} );\quad i = 1,2, \ldots ,N;\;t = 1,2, \ldots ,T \hfill \\ \end{gathered} $$
(1)

In Eq. (1) subscripts i and t index firm and time, respectively, yit is the natural logarithm of value added, the natural logarithm of capital is kit, the sum of a firm’s tangible and intangible assets at the end of the year, and the natural logarithm of labor lit is the mean number of employees8 in a given year. Intangible assets, like other assets, are valued in ways that adhere strictly to the standard accounting methods that are used in Finland. The valuations are reported on balance sheets, and we have taken those as given. The option program indicator is denoted by opit (i.e., opt, ssopt, bbsopt, dilu, diluss, and dilubb), γt is a year-specific intercept for common technological and economic shocks, and \( \eta_{i} \) is an unobserved firm-specific effect.

If the assumption of strict exogeneity on capital and labor inputs as well as on option schemes is violated in Eq. (1), our baseline fixed effects estimator is potentially inconsistent. Therefore, to obtain asymptotically consistent parameter estimates, we estimate single equation dynamic GMM estimators by using a common factor representation for Eq. (1) (see e.g., Blundell and Bond 2000):
$$ \begin{gathered} y_{it} = \beta_{y\_1} y_{i,t - 1} + \beta_{k} k_{it} + \beta_{k\_1} k_{i,t - 1} + \beta_{l} l_{it} + \beta_{l\_1} l_{i,t - 1} \hfill \\ + \beta_{op} op_{it} + \beta_{op\_1} op_{i,t - 1} + \gamma_{t} + \eta_{i} + \varepsilon_{it} \hfill \\ \varepsilon_{it} \sim iid(0,\sigma^{2} );\quad i = 1,2, \ldots ,N;\quad t = 2,3, \ldots ,T \hfill \\ \end{gathered} $$
(2)

In Eq. (2) variables correspond to those used in Eq. (1). The key difference is that the option program indicator measures the size of a plan, i.e., dilu, dilubb and diluss. Note that Eq. (2) does not have a dynamic interpretation, but with a dynamic specification we can impose (and test) the non-linear common factor restrictions for lit, kit and opit by using minimum distance.9

To construct a GMM instrument matrix, we first investigate the potential endogeneity of inputs and options by using Difference-in-Sargan (DS) tests (based on the Hansen J statistic). Initially, within a static GMM model, we assume that opt (option program dummy indicator) is strictly exogenous and investigate whether l and k are predetermined or endogenous. The p-value of 0.16 does not lead us to reject Ho, the hypothesis of pre-determinedness. On the contrary, when the same model was estimated by using a two-step system GMM estimator, the p-value for the DS statistic is 0.05, a finding that goes against the pre-determinedness hypothesis (explanatory variables were lit, kit, diluit, year dummies). Since the set of moment conditions is smaller under endogeneity, and given the potential for measurement errors, we treat capital and labor as endogenous variables.

We also investigate whether we can treat the option program indicator dilu (the proxy for program size) as predetermined or endogenous. The p-values rejecting the pre-determinedness hypothesis are 0.39 for dilu (and 0.13 for opt), so we treat dilu as predetermined.10 However, our key finding, that options have no impact on productivity, remains even though we treat dilu as an endogenous variable.

In the GMM context, we use the lagged levels of explanatory variables as instruments for the corresponding first-differenced variables (the first-differenced GMM estimator). As an extended estimator we apply the system GMM estimator, where we use lagged differences of explanatory variables as instruments for the level equations, since under weak instruments the system GMM estimator parameter estimates can be much more reasonable than the differenced GMM estimates.11

The GMM estimators may also suffer from the problem of too many instruments. Though this does not make the estimators inconsistent, it can cause severe problems in finite samples (see Roodman 2006). Therefore, in the system GMM models, first we use all possible lags from period t − 2 as GMM instruments for the explanatory variables yi,t−1, kit and lit. Next we use only the lags dated t − 2 and t − 3, but not lags from earlier periods, and, finally, restrict the size of the instrument matrix Z by a -collapse- option after -xtabond2- in Stata as explained in Roodman (2006). For option program indicators, we use only the lags dated t − 1.

5 Empirical results

Table 5 reports the baseline contemporaneous fixed effects estimation results.12 Since the tests indicate autocorrelation, the error term is first-order autoregressive in Table 5 (the estimated ρ is about 0.4). Three main conclusions emerge from Table 5.
Table 5

Baseline fixed effects estimates: Cobb-Douglas production functions, 1992–2002

Column

(1)

(2)

(3)

(4)

ln(l)it

0.621***

(15.88)

0.624***

(15.87)

0.620***

(15.55)

0.623***

(15.64)

ln(k)it

0.145***

(6.16)

0.146***

(6.18)

0.145***

(6.16)

0.145***

(6.17)

optit

0.007

(0.30)

   

ssoptit

 

0.016

(0.59)

  

bbsoptit

 

−0.011

(0.31)

  

diluit

  

0.058

(0.19)

 

dilussit

   

0.614

(1.31)

dilubbit

   

−0.183

(0.53)

Firm-year obs.

925

925

925

925

Firms

117

117

117

117

Baltagi-Wu LBIa

1.31

1.32

1.31

1.32

Modified Bhargava et al.a

0.93

0.94

0.94

0.94

Rho_ar

0.39

0.39

0.39

0.39

Fraction of variance because of u_i

0.88

0.88

0.88

0.88

R2 within

0.57

0.58

0.57

0.58

The dependent variable is the logarithm of firm value added, i.e., ln(value added)

The estimator is a modified fixed effects estimator -xtregar-in Stata, where the error term is first-order autoregressive. Autocorrelation parameter (rho_theil) is calculated as (ee_(t − 1)/ee) × (N − k)/N, where e is the vector of residuals and e_(t − 1) is the vector of lagged residuals

Absolute values of t-statistics in parentheses. *** Significant at 1% level

Opt is a dummy variable for the presence of an option program, ssopt is a dummy variable for the presence of a selective program and bbsopt is a dummy variable for the presence of a broad-based program. dilu is a proxy variable for the size of a program. diluss is an interaction variable between dilu and ssopt dummy. dilubb is an interaction variable between dilu and bbsopt dummy

As a robustness check, we also estimated models in Table 5 by using a standard fixed effects estimator (-xtreg-, fe cluster(id) in Stata; not reported here). The results are much the same as reported in Table 5

aBaltagi-Wu LBI is the Baltagi and Wu (1999) locally best invariant (LBI) test for the hypothesis of ρ = 0. The test statistics far below 2 indicate positive serial correlation. Modified Bhargava et al. Durbin-Watson (Bhargawa et al. 1982) tests the hypothesis of ρ = 0. Substantial deviations from zero indicate serial correlation

First, the estimates for capital kit and labor lit are highly significant in all columns. The baseline elasticity of capital input is close to 0.15, whereas for labor it is about 0.62. While these estimates clearly imply strongly decreasing returns to scale, it is known that fixed effects estimators tend to underestimate capital coefficients (Griliches and Mairesse 1998). Second, in Columns (1) and (2), where our option program indicator is the presence of a plan, a contemporaneous association between option schemes and firm productivity is found to be statistically insignificant. Third, in Columns (3) and (4), where our option program indicator is the proxy for the size of a plan, again we find no evidence of a statistically significant link with firm productivity.

As a robustness check, we also estimated standard fixed effects estimators for Columns (1)–(4) without modeling the error term as first-order autoregressive, but the results (available upon request) are much the same as reported in Table 5, despite the number of observations increasing from 925 to 1,042.

In addition, we implemented the flexible modeling approach described in Cable and Wilson (1989) in their study of profit sharing and estimated fixed effects production functions for opt (option program dummy variable). However, unlike Cable and Wilson we do not find statistically significant evidence of indirect effects option schemes through labor and capital inputs (results available upon request). Consequently, we cannot reject the hypothesis of a Hicks-neutral shift in the production surface. One reason for the difference in findings might be that option and profit sharing schemes are rather different. Second, we do not have data on those organizational characteristics that Cable and Wilson (1989) had available for their estimates.

We also followed an anonymous referee’s suggestion and modeled option program dynamics as in Jones and Kato (1995). We estimated fixed effects estimators with standard errors that allow for intragroup correlation and used distinct lags for option dummy variables until the fourth year. Though there are challenges in interpreting convincingly distinct time lags, again the key finding is that we did not find statistically significant evidence of a positive productivity effect.

Tables 6 and 7 report the estimation results of the pooled OLS, the fixed effects, and the system GMM estimators for a Cobb-Douglas production function.13 The pooled OLS and fixed effects standard errors are corrected for intra-firm correlation. The system GMM estimates are based on the two-step estimator with heteroskedastic-consistent asymptotic standard errors and a finite-sample correction proposed by Windmeijer (2005). For all test statistics in GMM models, we rely on the two-step GMM estimator. Since in our data the i, t-variation of option program dummy indicators is less than the variation in the dilution indicators, we use the dilution option program indicators in GMM models.
Table 6

GMM estimates with an option program size indicator (dilu)

 

(1)

Pooled OLS

(2)

Fixed effects

(3)

System GMM

(4)

System GMM

(5)

System GMM

ln(va)i,t−1

0.758***

(18.17)

0.438***

(5.47)

0.600***

(9.64)

0.630***

(11.15)

0.582***

(6.22)

ln(l)i,t

0.647***

(11.44)

0.591***

(9.28)

0.709***

(5.10)

0.676***

(3.98)

0.496***

(3.35)

ln(l)i,t−1

−0.465***

(8.47)

−0.264***

(3.06)

−0.433***

(3.12)

−0.401***

(2.77)

−0.308**

(2.13)

ln(k)i,t

0.120***

(4.34)

0.131***

(4.06)

0.086

(1.24)

0.079

(0.98)

0.167**

(2.33)

ln(k)i,t−1

−0.063**

(2.29)

−0.007

(0.23)

0.020

(0.35)

0.005

(0.08)

−0.026

(0.39)

dilui,t

0.186

(0.64)

0.127

(0.42)

0.017

(0.03)

−0.266

(0.39)

0.004

(0.01)

dilui,t−1

−0.373

(1.00)

−0.336

(1.03)

−0.578

(0.90)

−0.349

(0.84)

−0.177

(0.55)

R2

0.99

0.80 (within)

m1 (test statistics neg.)

0.00

0.00

0.00

m2 (test statistics neg.)

0.87

0.93

0.91

Number of GMM instruments

181

97

49

Hansen J test for overid.restrictions (p-value)

1.00

0.12

0.55

Diff.-in-Hansen tests for exogeneity of GMM instrument subsets (p-values)

GMM instr. for levels

1.00

0.40

0.11

GMM instr. for ln(va)

1.00

0.46

0.57

GMM instr. for ln(l)

1.00

0.52

0.08

GMM instr. for ln(k)

1.00

0.17

0.70

GMM instr. for dilu

1.00

0.97

0.55

GMM parameter estimates imposing the common factor restrictions

ln(va)i,t−1

0.662***

(11.58)

0.660***

(12.81)

0.619***

(6.89)

ln(l)i,t

0.848***

(9.81)

0.820***

(8.88)

0.659***

(5.98)

ln(k)i,t

0.162***

(2.79)

0.182**

(2.48)

0.180***

(2.88)

dilui,t

0.020

(0.04)

−0.105

(0.26)

0.251

(0.69)

Comfac

0.01

0.01

0.17

CRS

0.87

0.96

0.05

The dependent variable is ln(value added)

Pooled OLS and fixed effects standard errors are corrected for intra-firm correlation. GMM standard errors are based on the two-step heteroskedastic-robust estimator with a finite-sample correction proposed by Windmeijer (2005). Absolute values of t statistics are in parentheses. *** Significant at 1% level, ** at 5% level, respectively

m1 and m2 are tests for first- and second-order autocorrelation. The statistics are asymptotically standard normal under the null of no serial correlation. p-values are reported

The number of firm-year observations is 925 (117 firms). Year dummies are included in all models

Comfac is a minimum distance test of the non-linear common factor restrictions. p-values are reported

CRS is a Wald test for the constant scale to returns hypothesis (βl + βk = 1) after imposing the common factor restrictions. p-values are reported

Table 7

GMM estimates with an option program size indicators dilubb (broad-based schemes) and diluss (selective schemes)

 

(1)

Pooled OLS

(2)

Fixed effects

(3)

System GMM

(4)

System GMM

(5)

System GMM

ln(va)i,t−1

0.758***

(18.02)

0.433***

(5.44)

0.605***

(9.47)

0.625***

(11.48)

0.617***

(6.74)

ln(l)i,t

0.646***

(11.51)

0.595***

(9.21)

0.671***

(5.15)

0.657***

(3.88)

0.507***

(3.63)

ln(l)i,t−1

−0.465***

(8.45)

−0.260***

(3.06)

−0.417***

(2.93)

−0.385***

(2.71)

−0.350**

(2.47)

ln(k)i,t

0.120***

(4.34)

0.130***

(4.01)

0.113*

(1.80)

0.115*

(1.67)

0.197***

(2.61)

ln(k)i,t−1

−0.063 **

(2.29)

−0.008 **

(0.25)

−0.007

(0.14)

−0.026

(0.41)

−0.054

(0.73)

dilussi,t

0.163

(0.28)

0.193

(0.37)

−0.389

(0.32)

−0.393

(0.25)

−0.875

(0.49)

dilussi,t−1

−0.300

(0.44)

0.483

(1.00)

−0.464

(0.58)

−0.211

(0.26)

−0.754

(0.81)

dilubbi,t

0.200

(0.70)

0.103

(0.3,)

0.177

(0.34)

−0.188

(0.30)

0.107

(0.15)

dilubbi,t−1

−0.398

(1.00)

−0.554

(1.44)

−0.575

(0.87)

−0.378

(0.96)

−0.196

(0.58)

R2

0.99

0.80 (within)

m1 (test statistics neg.)

0.00

0.00

0.00

m2 (test statistics neg.)

0.88

0.98

0.97

Number of GMM instruments

189

105

57

Hansen J test for overid. restrictions

1.00

0.22

0.39

Diff.-in-Hansen tests for exogeneity of GMM instrument subsets (p-values)

GMM instr. for levels

1.00

0.58

0.13

GMM instr. for ln(va)

1.00

0.83

0.97

GMM instr. for ln(l)

1.00

0.75

0.06

GMM instr. for ln(k)

1.00

0.47

0.54

GMM instr. for diluss

1.00

0.81

0.09

GMM instr. for dilubb

1.00

1.00

0.16

GMM parameter estimates imposing the common factor restrictions

ln(va)i,t−1

0.622***

(10.66)

0.654***

(13.34)

0.616***

(7.17)

ln(l)i,t

0.810***

(10.25)

0.796***

(9.67)

0.622***

(6.04)

ln(k)i,t

0.163***

(2.95)

0.190***

(3.02)

0.192***

(3.01)

dilussi,t

0.681

(0.70)

0.189

(0.21)

0.714

(0.84)

dilubbi,t

−0.223

(0.48)

0.025

(0.07)

0.174

(0.39)

Comfac

0.02

0.06

0.46

CRS

0.57

0.77

0.01

Notes

The dependent variable is ln(value added)

Pooled OLS and fixed effects standard errors are corrected for intra-firm correlation. GMM standard errors are based on the two-step heteroskedastic-robust estimator with a finite-sample correction proposed by Windmeijer (2005). Absolute values of t statistics are in parentheses. *** Significant at 1% level, ** at 5% level, * at 10% level, respectively

m1- and m2 are tests for first- and second-order autocorrelation. The statistics are asymptotically standard normal under the null of no serial correlation. p-values are reported

The number of firm-year observations is 925 (117 firms). Year dummies are included in all models

Comfac is a minimum distance test of the non-linear common factor restrictions. p-values are reported

CRS is a Wald test for the constant scale to returns hypothesis (βl + βk = 1) after imposing the common factor restrictions. p-values are reported

In Table 6 we relax the strict exogeneity assumptions on production inputs and option schemes, as explained in Sect. 5, and treat production inputs as endogenous and option schemes as predetermined. The following key findings emerge from Table 6. First, the findings suggest consistently that the option program indicators are statistically insignificant. In Column (3), where we use a very large GMM instrument set (181 instruments), i.e., all possible lags from period t − 2 for lagged dependent variable and production inputs, but only once lagged option scheme variables in levels as instruments for differences, we observe a clear overfitting bias in the overidentification and the Difference-in-Hansen tests. The Wald test suggests constant returns to scale, but the common factor restrictions are clearly rejected.

When we decrease the size of the instrument matrix by using lags only from the periods t − 2 and t − 3 for lagged dependent variable and production inputs (97 instruments), the findings in Column (4) again show statistically insignificant evidence of an association between option programs and firm productivity. Also, the overidentification and the Difference-in-Hansen tests seem to be more reasonable in Column (4) than in Column (3), but unfortunately the common factor restrictions are rejected. The constant returns to scale hypothesis is accepted.

In Column (5) we further reduce the size of instrument matrix by allowing all possible lags from the periods t − 2 for lagged dependent variable and production inputs, as in Column (3), but use -collapse- option (see Roodman 2006) to limit the size of the instrument matrix (49 instruments). As can be seen from Column (5), our key finding that there is no evidence of option plans affecting performance does not change. Now the common factor restrictions are accepted (p-value 0.17), but the constant returns to scale hypothesis is rejected (p-value 0.05).

Third, the system GMM parameter estimates that use a lagged dependent variable and which are reported in Columns (3)–(5) are all lower than the pooled OLS estimates but higher than the fixed effects estimates. This finding indicates that the system GMM estimator is consistent (see e.g., Blundell and Bond 2000).

Fourth, the autocorrelation tests, namely m1 and m2 reported in Columns (3)–(5), provide additional support for the use of the system GMM estimator (significant negative autocorrelation in the first-differenced residuals but not in the second-order). In addition, the Hansen J tests do not reject the over-identification restrictions in Columns (4) and (5), though the test statistic is quite high in Column (5) (p-value 0.55). The Difference-in-Hansen tests for exogeneity of GMM instruments are accepted in the results reported in both columns, indicating that our key finding is robust to the number of instruments. As a robustness check, we estimated a GMM model by using the smallest possible instrument matrix Z (25 instruments) under our assumptions that capital and labor are endogenous variables and that the option program proxy dilu for the size of the program is a predetermined variable. Unfortunately, we found substantial bias in the labor parameter estimate; it was highly insignificant and unexceptionally low 0.087. However, the finding of no link with productivity remained.

Another robustness check was to estimate all our GMM models by treating dilu as endogenous; the finding of no impact on productivity remained. As an additional robustness check, we estimated the dynamic panel data model in Column (5) of Table 6 as a static panel data GMM model, where the explanatory variables were k, l, dilu and year dummies and no lagged variables were included. Production inputs k and l were endogenous, and initially dilu was treated as endogenous and then subsequently as predetermined. The finding of no impact on productivity remained. We also repeated the same estimations with diluss and dilubb, and the finding of no impact on productivity remained.

Finally, when assessing the productivity effects of different types of option plans, Table 7 shows no convincing statistical evidence of a link between broad-based and selective schemes and firm productivity. Since the key findings and models in Table 7 are similar to those reported in Table 6, we do not discuss these more fully.

6 Conclusions

We assemble new panel data for all Finnish publicly listed firms during 1992–2002 and estimate Cobb-Douglas production functions with different option program indicators. These measures reflect the presence or absence of an option scheme, the size of the scheme and whether the scheme is selective or broad based. Furthermore, the long panel nature of our data allows us to address the potentially important issue of endogeneity of production inputs and options by estimating dynamic panel data models with a GMM estimator (and impose and test common factor restrictions).

The most important finding, consistently found in diverse specifications, is of a statistically insignificant association between option programs and firm productivity. This result is independent of which particular option program indicator is used. When endogeneity of production inputs and option-schemes are taken into account, we continue to find no evidence of a link with firm productivity.

Hence our empirical findings provide support for those theorists who hypothesize that the performance impact of options (incentive effect) will be limited because of reasons such as free-rider problems (e.g., Oyer 2004), accounting myopia (e.g., Hall and Murphy 2003), line-of-sight arguments (e.g., Vroom 1995), or rent-seeking (e.g., Bebchuk and Fried 2003). Our results are also consistent with much of the financial literature that does not find evidence of a link between options and business performance (e.g., Hall and Murphy 2003).

At the same time our findings differ from those contained in earlier productivity studies, especially those that tended to find evidence of a positive effect of broad-based options. One reason for the divergence might be that we have access to more reliable and extensive data that encompasses a period spanning both stock market boom and downturn. Also we use data that is much less likely than previous work to suffer from selection biases and employ econometric methods that address the potentially important issue of endogeneity of production inputs and employee stock option schemes. However, to some extent the divergence in results may reflect differences in institutional arrangements in Finland and arrangements elsewhere.

Footnotes
1

See Mäkinen (2001, 2007, 2008) and Jones et al. (2006) for studies of diverse dimensions of stock option schemes and CEO compensation in Finland.

 
2

Our data contains all firms listed in the main list of the Helsinki Stock Exchange (HEX) for the period 1992–2006 and for the two minor lists, I-list and NM-list, for the period 1997–2002.

 
3

In response to an anonymous referee’s comment we note that while there are certain industry specific patterns observed in our data, one should note that at the end of the observation period almost 74% of firms used stock options, so the phenomenon was common across many different industries.

 
4

In public stock exchange reports, the firm typically reports whether the scheme is targeted only to managers, to managers and a selected group of key personnel or to the workforce more broadly. Our classification is thus different from Sesil et al. (2000, 2002), who use 50% threshold as a criterion for broad-based schemes. Our data do not include this information, but have the important advantage of being derived from publicly reported sources that must be externally verifiable, rather than from confidential surveys.

 
5

See for example Ikäheimo et al. (2004) and Liljeblom and Pasternack (2006).

 
6

We omit 8 firms or 15 firm-year observations due to their having fewer than 4 consecutive observations. To utilize all possible firm-level data we also collected, if possible, data on income statements and option schemes prior to a firm’s listing on the Helsinki Stock Exchange.

 
7

One of a series of robustness checks we performed involved re-estimating many of the specifications to be discussed and reported in Tables 5, 6 and 7, but without excluding any outliers. Reassuringly, the key finding—that options did not affect productivity—was unaffected when this was done. However, if we adopted a rule of excluding firms who employed fewer than 18 persons, then we found mixed evidence on productivity effects.

 
8

Finnish firms report only mean number of employees per year in financial statements. Unfortunately, our data do not contain information on firms’ labor costs.

 
9

We kindly thank an anonymous referee for providing a program for the common factor restrictions.

 
10

Since option plans typically are introduced publicly in early spring, a few weeks before the annual general meeting of shareholders, this seems to be a reasonable assumption. Therefore, a potential correlation is likely to be with a previous period rather than being contemporaneous. We also estimated an IV estimator (i.e. -ivreg2- with -cluster- in Stata), where an option program indicator was treated as an endogenous variable in a Cobb-Douglas production function. As excluded instruments we used the HEX general index, ICT sector dummy indicator and the share ownership of a largest shareholder. The instrumented option dummy indicator was still insignificant (p-value 0.84), when the F-test for the excluded instruments in the first-stage regression was highly significant (p-value 0.00) and the Hansen J statistic supported the validity of the (all) instruments (p-value 0.12).

 
11

For more on dynamic GMM estimators and constructing an instrument matrix, see Arellano and Bond (1991), Arellano and Bover (1995) and Blundell and Bond (1998, 2000).

 
12

We use Stata/SE 10.1 in all estimations. We pooled the sectors in order to increase the number of observations, since GMM estimators can perform poorly in finite samples. However, in unreported regressions we estimated production functions separately for the three sectors; information technology (IT), service, and manufacturing. As expected, when comparing parameter estimates for production technology inputs, we find, in the fixed effects model context, that the labor coefficient was higher in the IT and service sectors than in manufacturing, whereas capital was clearly higher in the manufacturing sector compared to other two sectors. However, our key finding is preserved—we did not find statistically significant coefficients for our option program indicator variables opt, ssopt, bbsopt, dilu, diluss and dilubb in different model specifications. In the GMM context, however, the size of the instrument matrix can be large compared to the number of observations in the service and IT sectors, which seems to substantially bias estimation results.

 
13

We do not report the differenced GMM estimation results in Tables 6 and 7, since our unreported results (available upon request) suggest that the individual series are highly persistent (but not an exact unit root). That being the case, the estimator is very likely to suffer from severe finite sample bias or inconsistency (e.g., Blundell and Bond 1998; Blundell et al. 2000).

 

Acknowledgments

Earlier versions of this paper have benefited from comments by participants at the ASSA/ACES meeting in San Diego, January 2–5, 2004, the 12th IAFEP Conference in Halifax, July 8–10, 2004, the 16th EALE conference in Lisbon, September 9–11, 2004, the FPPE Industrial Organization Workshop in Helsinki, December 9–10, 2004, and the EALE/SOLE World Conference in San Francisco, June 2–5, 2005. We are grateful to comments from three anonymous referees as well as suggestions from Kari Hämäläinen, Kevin Hallock, Seppo Ikäheimo, Pekka Ilmakunnas, Uwe Jirjahn, Jeffrey Pliskin and Otto Toivanen. Also we acknowledge Mikael Katajamäki for his outstanding research assistance. Kalmi and Mäkinen gratefully acknowledge funding from the Academy of Finland, the Finnish Work Environment Fund, the Marcus Wallenberg Foundation and the Helsinki School of Economics Research Foundation. In addition, Mäkinen thanks the Yrjö Jahnsson Foundation and the Foundation of Kluuvi for financial support. Support from the Research Institute of the Finnish Economy (ETLA) is gratefully acknowledged. Jones’ work was supported in part by a Foundation for Economic Education Fellowship for which he is grateful.

Copyright information

© Springer Science+Business Media, LLC 2009