# «Three essays on corporate boards R. Øystein Strøm A dissertation submitted to BI Norwegian School of Management for the degree of Dr.Oecon SERIES ...»

4.4 Variable deﬁnitions I need to deﬁne CEO turnover and board turbulence. In table 4.1 three measures of CEO turnover emerge. First of all, I follow Yermack (2004) in considering all departures as relevant. Second, I supplement with a proxy of forced dismissals constructed as the product of CEO and chairman turnover. As is evident from table 4.1, CEO and chairman turnovers are simultaneous in more than 50 per cent of the CEO turnover cases. I suppose that these cases of coincidence signify dismissals. Third, a measure of forced departures for CEOs is obtained from previous work on CEO turnover in Norwegian companies (Bøhren et al., 2002).

I identify four reasons for not concentrating on forced CEO departures alone. First, the timing of CEO turnover and board turbulence may reveal board control type, for whatever reason the departing CEO left the company. Thus, for this reason alone, all CEO changes belong to the data set. Second, public information on dismissals is likely to emerge only in a minority of cases, probably where the conﬂict is most acute, since both the ﬁrm and the departing CEO want to defend their reputational capital.

Gilson (1990) shows that directors in defaulting ﬁrms have trouble obtaining new directorships. This induces them to keep conﬂicts between CEO and board, or among board members, out of the public eye. Using only this subsample of turnovers is, consequently, likely to result in a seriously biased sample. To top it all, some studies judge a departure as forced if ﬁrm performance in advance has been weak. Then, there is no surprise in ﬁnding weak performance to be associated with CEO or director departure. Third, the often-used Parrino (1997) criteria3 for selecting forced departures include the explicit dismissals, but then assign a turnover to the forced category if the departing CEO is less than 60 years old. In practice, this means that nearly all departures of CEOs less than 60 years of age are forced. This comes close to the all CEO turnover measure. And even though the public announcement may be that the CEO or the director has found new employment or wishes to leave, the protection of reputational 3 Parrino classiﬁes turnovers as either forced or voluntary, ascertained from the Wall Street Journal (WSJ). The turnover is forced in two main cases: (1) If WSJ reported that the CEO was ﬁred, forced from the position, or departed due to unspeciﬁed policy differences;

(2) If the departing CEO is under the age of 60 and the WSJ announcement of the succession (2a) does not report the reason for the departure as involving death, poor health, or the acceptance of another position; or (2b) reports that the CEO is retiring, but does not announce the retirement at least six months before the succession. Searches in other parts of the business and trade press are undertaken for the second group in order to reclassify, if necessary.

## CHAPTER 4. BOARD CONTROL

capital makes all such announcements untrustworthy. Franks and Mayer (2001) move even one step further, when they deﬁne board turnover as the number of directors leaving the board during the year, for reasons other than death or retirement.Fourth, CEO turnovers in my sample show that departing and arriving CEOs have about the same age distribution, see ﬁgure 4.1.

The measure is continuous and always positive and is new to the literature. A value of zero will signify no replacements and no board size change. Table 4.1 shows that the average turbulence is 1.43. Looking ahead, table 4.2 reports that the percentage of new directors on a board is 26.1 on the average Norwegian board. Thus, the turnover of both CEO and directors appears to be high in my Norwegian data. I return to this point in section 4.6.

This paper employs two measures of ﬁrm performance; one is the stock return, a market measure, and one accounting measure, the return on assets (ROA). Hermalin and Weisbach (1998) argue that the board will have a greater faith in the accounting measure than in a market measure. The Bøhren and Strøm (2007) measure of independence, that is, tenure difference between the shareholder elected directors and the CEO, is employed.

4.5. DATA AND METHODS 119

4.5 Data and methods 4.5.1 Data My sample comprises all non-ﬁnancial ﬁrms listed on the Oslo Stock Exchange (OSE) at year-end at least twice over the period 1989 to 20024. To reduce censoring bias in the tenure measures, I start collecting director data in 1986. The ownership structure data covers every equity holding by every investor in every sample ﬁrm. The public securities register provided the ownership data, accounting and share price data is from the OSE’s data provider, and board data was collected manually from Kierulf’s Håndbok and a public electronic register from 1995.

**4.5.2 Methods**

Since the CEO turnover is a discrete event, taking two values, 1 for a new CEO and 0 for the continuation of the incumbent, and the board turbulence measure is continuous, different estimation methods are employed.

For the CEO turnover the probit model is the estimation vehicle, and I use the general methods of moments (GMM) for the board turbulence. Both relationships are estimated with panel data methods.

I ﬁrst set out the GMM method. Since the measure of board turbulence is continuous, I may then use ﬁxed effects panel data estimation (Woolridge, 2002). In a panel of ﬁrms, the speciﬁc ﬁrm’s heterogeneity will cause residuals to be dependent. Therefore data pooling will lead to biased estimates. Fixed effects estimation has the advantage of removing the dependency in residuals, since each ﬁrm’s overall average on a particular variable is subtracted from the given year’s observation. The transformed observations will be independent. With ﬁrm heterogeneity thus removed, the ﬁxed effects method makes it unnecessary to include a host of control variables in regressions. In ordinary least squares estimation on pooled data the objective for the inclusion of control variables is, of course, to remove ﬁrm heterogeneity. In the regressions that follow, I retain only the control variables ﬁrm size and risk.

4 The OSE had an aggregate market capitalization of 68 bill. USD equivalents by yearend 2002, ranking the OSE sixteenth among the twenty–two European stock exchanges for which comparable data is available. During my sample period from 1989 to 2002, the number of ﬁrms listed increased from 129 to 203, market capitalization grew by 8% per annum, and market liquidity, measured as transaction value over market value, increased from 52% in 1989 to 72% in 2002 (sources: www.ose.no and www.fibv.com).

## CHAPTER 4. BOARD CONTROL

The ﬁxed effects estimations are implemented with GMM, whose great advantage over other methods is that few assumptions are needed. For instance, OLS needs assumptions of homoskedasticity and absence of serial independence. The consistency of the GMM estimator follows only from the fact that it satisﬁes certain moment conditions (Davidson and MacKinnon, 1993, page 585). When these are satisﬁed, instruments are given naturally. In this paper, the instruments are constructed from the explanatory variables. Using Davidson and MacKinnon (1993, page 584), the expected value of the dependent variable yt given the information set Ωt is written E (yt |Ωt ) = Xt β t = 1,..., n (4.2) where Xt is a vector of explanatory variables, and β the vector of k corresponding coefﬁcients. Since (4.2) provides the conditional moment condition E (ut |Ωt ) = 0, it follows that for any vector W with elements wt ; t = 1,..., n such that wt ∈ Ωt, the unconditional moments E (wt (yt − Xt β)) =0. Now, the regressors Xt belong to the information set Ωt, and there are precisely k of them. The practical implication is that the k regressors dene k unconditional moment conditions. I have gone one step further and added transformations of the explanatory variables as instruments (see below). This means that I have more than the necessary k instruments for identiﬁcation.

The transformations are as follows. First, I use the raw (Amemiya and MaCurdy, 1986), the time-demeaned, and the squared time-demeaned explanatory variables. Furthermore, I include the average and standard deviation of ﬁrm-demeaned explanatory variables (Breusch et al., 1989). Finally, the panel data structure allows the inclusion of contemporary as well as lagged instruments. Since explanatory variables are lagged, I use lagged instruments.

Next, I want to estimate the probability that the CEO is new. For this purpose, the probit method, applied to panel data, is used. Early studies (Hermalin and Weisbach, 1988 and Kaplan and Minton, 1994) employ the logit model on pooled data. Panel data methods for the study of CEO and director turnover have only recently come into use, one example being Fich and Shivdasani (2006).

With panel data, the unobserved effects probit model is in general P (CEO Turnoverit ) = 1|(Explanatory variables)it, ci (4.3) = Φ ((Explanatory variables)it β + ci ) t = 1,..., T where P is probability, ci is unobserved ﬁrm i heterogeneity, and β is the

4.6. DESCRIPTIVE EVIDENCE 121 vector of coefﬁcients to the explanatory variables. The Φ symbol indicates the standard normal distribution. Two common assumptions are added to (4.3). The ﬁrst is that the CEO turnover variable is independently distributed across time t, conditional on the explanatory variables and unobserved ﬁrm heterogeneity. The second assumption is that unobserved ﬁrm heterogeneity is normally distributed with zero mean and a ﬁxed standard deviation ci ∼ Φ 0, σc. With these assumptions in hand, I estimate the relationship with maximum likelihood methods.

For panel data the probit models meet with the incidental parameters problem (Woolridge, 2002, page 484) when assuming a ﬁxed effects model and performing within transformations, leading to inconsistent estimates.

For this reason, the ﬁxed effects model is dropped, and instead I add 14 year and 19 industry dummies to control for ﬁrm heterogeneity as much as possible. The dummy coefﬁcients are not reported. Since unobserved ﬁrm heterogeneity cannot be removed, the coefﬁcient values will be biased. However, the coefﬁcients will show the correct direction of impact of each explanatory variable. Furthermore, since the bias is known, the slope coefﬁcients are observable at the average of the distribution, that is, when ci = 0 given the assumptions for (4.3). These so-called average parI report the most tial effects (APE) βc are estimated as βc = β/ 1 + σc important in the text.

**4.6 Descriptive evidence**

The objective in this section is to give some stylized facts about CEO turnover and board turbulence. First, I report some overall statistics, and then turn to the simultaneity of CEO, chairman and director changes around the CEO turnover event.

** Table 4.1 in section 4.**

3.2 shows the main characteristics of the various deﬁnitions of turnover. It is plain that overall CEO turnover is about 20 per cent. This is higher than in American data, where Fich and Shivdasani (2006) report a departure rate of 11.28 per cent. The contemporaneous departure of CEO and chairman takes place at an average rate of 11 per cent.

Identiﬁed forced CEO dismissals happen at an average rate of approximately 4 per cent. In Fich and Shivdasani (2006) the forced CEO dismissals are 18 per cent of all dismissals, that is, comparable to the proportion in our data. Thus, my proxy for forced dismissals implies a high simultaneous CEO and chairman turnover. As will become evident, this is repeated for board turnover as well.

## CHAPTER 4. BOARD CONTROL

Do CEO, chairman and director changes occur with different intensities over the period? Table 4.2 shows changes in each year of my sample.Clearly, the changes happen with different intensities over the period.

At the lowest, 10.3 per cent of CEOs are new (1995), while the highest is in 2001 at 28.3 per cent of CEOs. The result for the CEO conﬁrms the Huson et al. (2001) ﬁnding that CEO ﬁrings have been trending upwards during the sample period, although the inﬂux of new directors is at its lowest in the middle. The variation in changes affects the CEO, the chairman and the directors usually in the same year.

In the same table, ﬁrm performance, expressed as stock return and return on assets (ROA), is set out for each year. The expectation is that turnover is high in years of low ﬁrm performance, but low in years of high ﬁrm performance, as former studies show. Yet, it is hard to detect any pattern. These summary statistics show no such deﬁnite relationship.

The simultaneity of CEO and chairman departures is shown in ﬁgure 4.2.

** Figure 4.2**

It turns out that most of the chairman changes occur simultaneously with the change of new CEO. Furthermore, notice that changes are fairly symmetric around the year of CEO turnover, with slightly more chairman changes taking place before the CEO turnover.

** Figure 4.3 shows the director changes around the CEO turnover event.**

** Figure 4.3**

The same-year spike shows up for director turnover as it does for the chairman. New directors arrive either because they substitute old, or because the board size is enlarged. Again changes are distributed around the CEO turnover year, but this time is clearly higher before CEO turnover than after. The values at the tails of the distribution are less reliable, since fewer observations appear here.

The conclusion to this simple overview is that chairman and director turnover tend to occur simultaneously with CEO turnover, and when they do not, they occur before rather than after the new CEO is in place. This contradicts the CEO control hypothesis, and supports the joint control hypothesis. Thus, the board and the CEO seem to form a team.

4.7. ECONOMETRIC EVIDENCE 123

**4.7 Econometric evidence**

Is the simultaneity in CEO and board departures in ﬁgures 4.2 and 4.3 also present in a multivariate setting? This section tests for this and for the effects of other variables. Section 4.7.1 reports the results for CEO turnover.