CEO tenure and the performance-turnover relation

In our formulation, “ability” refers to the manager’s selection and implementation of strategies that create shareholder value. Consequently, “ability” is a unique time invariant characteristic of an employment relationship (see Holmström 1999 and Hermalin and Weisbach 1998).

All increases in likelihood reflect marginal increases in turnover probability over a base case predicted probability derived from average covariate values.

As described later, these results are robust to using other transformations of the tenure variable.

Comparing the prior firm performance of departing CEOs with their cohort counterparts who survive the quarter holds constant both their tenure and the duration over which we measure their performance. However, if a CEO survives because he is entrenched as a consequence of his power over the board and owners, the one-quarter survivor is a noisy measure of this outcome. Consequently, we also include CEOs who survive 4 years since they potentially better represent survival due to their power over firm owners that is unrelated to their performance.

We thank an anonymous referee for suggesting this empirical test.

Allgood and Farrell (2000) find a U-shaped pattern in the weight on performance measures for CEO retention decisions, across CEO experience levels. While their results are sensitive to the definition of CEO experience and the proxy for annual firm performance, the performance-turnover relation in their sample is declining in tenure, consistent with our results, but only in their sub-sample of new and mid-tenure CEOs. However, we do not find any evidence of an increase in the performance-turnover relation for long-tenured CEOs.

In other contemporaneous work, Lee et al. (2012) find that, when performance is poor, a positive relation exists between the likelihood of CEO turnover and absolute management forecast errors. They also find this relation is stronger in a sample of CEOs in the lowest quartile of CEO tenure. These results are consistent with our main hypothesis; however, the focus of our study is on the moderating effect of uncertainty about CEO ability on the general performance-turnover relation, while the focus of Lee et al. (2012) is whether management forecast accuracy is a signal of CEO ability.

See Murphy (1999) for a review of the literature establishing the use of performance measures in executive compensation and Brickley (2003) for a summary of work relating firm performance to CEO departures.

Research has shown the economic importance of missing consensus analyst forecasts through, for example, negative stock price reactions (Matsumoto 2002) and lower annual CEO performance bonuses (Matsunaga and Park 2001).

Recent survey evidence suggests that a manager’s human capital concerns are heightened when firm performance falls below certain benchmarks. For instance, Graham et al. (2005) state: “Repeatedly failing to meet earnings benchmarks can inhibit the upward or inter-industry mobility of the CFO or CEO because the manager is seen as an incompetent executive or a poor forecaster.”

Using a belief revision model of managerial ability, Gibbons and Murphy (1992) examine the role of periodic performance reports and career concerns in the design of explicit incentive contracts. A key insight from their study is that, as periodic performance reports mitigate the uncertainty about the CEO’s ability, the board must substitute explicit incentives to offset the reduction of implicit incentives arising from weaker CEO career concerns. Since our study is focused on examining the effect of uncertainty in beliefs on management turnover, we abstract away from specific compensation contract design features.

Our framework also implies that longer serving CEOs are perceived to have, on average, higher ability than the total population of CEOs, because they have revealed sufficiently good performance to preclude their dismissal. Consequently, tenure reflects entrenchment in a statistical sense.

Our sample period begins in 1996 because the governance data (discussed later) used to measure monitoring intensity is only available starting in 1996.

We refine our dataset in the following ways. First, the IRRC database contains errors in the recording of the annual meeting dates, which we investigate and correct if we can independently identify the correct date; otherwise, we omit observations with incorrect dates. Second, we omit 1,402 out of 152,637 IRRC observations that contain coding mismatches between our IRRC dataset and the IRRC dataset available from Wharton Research Data Services (WRDS). Third, for firms with two meeting dates in the same fiscal year, we match the meeting date and corresponding IRRC data with the closest fiscal date and corresponding fiscal data from Compustat and CRSP.

In untabulated tests, we show that all results in the empirical analysis below are robust to including CEO retirements, deaths, and resignations due to illness in the measurement of the dependent variable (i.e., using 1,404 turnovers rather than 782 turnovers for the dependent variable). However, if we replace the 782 “forced” turnovers with the 622 “unforced” turnovers in redefining the dependent variable, results are consistent in sign but (as expected) considerably weaker in significance because of the additional noise in the redefined dependent variable.

See specific definitions of each measure in the “Appendix”.

While it is possible to include four separate variables showing the level of performance over the previous four quarters, it is difficult to compare the performance effects across quarters. For example, in some observations the most recent quarter may contain the largest sales for the year (e.g., November to January quarter for retailers). Further, as a practical matter, the analysis becomes difficult to interpret with interactions on each of the quarterly performance variables. Instead, for expositional clarity, we aim to capture the time-series spirit of quarterly performance through our Miss variables, which reveal the number of times the firm missed a benchmark in each of the past four quarters. In any case, our results are qualitatively unaffected if we include the prior six quarters of performance information despite the fact that we lose statistical power because of dropping observations corresponding to managers who are replaced within six quarters of their appointment.

For example, the upper quartile of our sample CEOs tenure is greater than 9.76 years. Estimating analyst based performance for these managers would require obtaining I/B/E/S analyst data prior to the mid 1980s, which is unavailable for many firms or is of poor quality (Bruce and Bradshaw 2004).

All variable definitions are provided in the “Appendix”.

As a robustness check, we include the natural logarithm of assets as a control variable in the regressions, but the coefficient is statistically insignificant in all tests. Moreover, the coefficients on the variables of interest (performance and the interactions), as well as the statistical significance of the incremental marginal effects, are virtually unchanged. In other untabulated results, we also median-industry-adjust the Miss-Fcast and Miss-Earn variables and the control variables Rvol, Evol, and Btm and re-estimate Eq. (3). The coefficients on the industry-adjusted number of misses are similar in magnitude to the corresponding coefficients reported in Table 2 and are statistically significant.

While we view each measure as an alternative proxy for firm performance, in untabulated analysis we include for completeness all performance surprise variables in the same specification. The coefficient on Nret and the magnitude of the analyst forecast surprise variables remain negative and significant, after controlling for the other performance measures.

We compute these probabilities by setting the continuous control variables at their sample averages, a CEO less than 64 years old, and operating in the two-digit SIC code equal to 10.

Our restriction to CEOs beginning their appointment from 1995 to 2001 allows us to capture firm performance in each quarter of each CEO’s first four years of employment within our full sample period.

We do not report the first quarter differences because of differences in starting dates—some CEOs started their appointments early in the quarter, and others did not.

Quarter one is excluded because of differences in starting dates for CEO appointments. Quarter 16 is excluded because any departees defined in quarter 16 would in effect satisfy the definition of a four-year survivor.

Including multivariate controls for firm characteristics (e.g., firm size, book-to-market, and returns volatility), CEO characteristics (e.g., age, CEO ownership, CEO chair), and governance characteristics (e.g., board independence, board size, board ownership) reduce the number of positive and significant coefficients on four-quarter cumulative return on assets to 10 out of 12 quarterly regressions. After including controls for firm, CEO, and governance characteristics for the regressions that specify cumulative returns as the performance variable, the number of positive and significant coefficients remains at 10 out of 12 quarterly regressions. .

Data available from http://​www.​law.​harvard.​edu/​faculty/​bebchuk/​data/​E_​Index_​1990-2008.​zip.

An alternative argument is that more outside directorships imply that a board member is busier and therefore results in weaker monitoring (Fich and Shivdasani 2006, Fracassi and Tate 2012). To the extent that such an argument holds, the multiple directorship results in our study are the opposite of what we would expect (i.e., it would imply less experienced CEOs attract less monitoring). This stands in contrast to the other results in Table 8. Moreover, we are ultimately interested in the relation between CEO tenure and multiple directorships, which we expect and find to be negative. On the other hand, the raw correlation between CEO tenure and the busy board measure in, for example, Fich and Shivdasani (2006) is significantly positive, which is inconsistent with both our sample and samples in other recent work (Hwang and Kim 2009). A possible explanation for the difference is that the Fich and Shivdasani (2006) sample period ends in 1995 while ours begins in 1996.

A limitation of some of our monitoring intensity variables is that they may be time invariant and therefore do not plausibly capture a change in monitoring intensity. To address this concern, in untabulated regressions, we estimate the relation between each of our monitoring variables and three annual lags of the monitoring variable. All three lags are significantly positively associated with the current period monitoring except for Chair and %OutsideDir, which had two of three lags with significantly positive coefficients. This contrasts with E-index, which is based on charter provisions and is very unlikely to change from year to year—in that case, the previous year explained almost all of the variance in the current value of E-index; year 2 and year 3 lags were not statistically significant. In addition, we estimated the Table 8 regressions controlling for firm fixed effects, and the results in each regression hold except for %Indep, Bsize, and NumMeet, the coefficients on which become statistically insignificant.

In untabulated regressions, we test the robustness of the results reported here using logistic regressions with the dependent variable equal to one if the percentage of outside directors is above a 60% cutoff and zero otherwise. This measure was originally used by Weisbach Weisbach (1988). Our inference based on this logistic regression is the same as that based on the reported OLS regressions. We also use a cutoff of 50% outside directors or a cutoff of the top quartile (which represents 67% outside directors), and the inferences remain identical.

We also re-estimate the regressions using as the variable of interest the change in, instead of the level of, the natural logarithm of tenure. The idea here is in any given year, the change variable would be smaller for longer-serving CEOs and capture more of a lead-lag relation, i.e., a change in an independent variable leads to a change in a dependent variable. Untabulated results using the change variable are generally similar to those reported Table 8, except that the coefficient on the change in Lntenure is no longer statistically significant.

Specifically, we: (1) regress Lntenure on the monitoring intensity variables plus squared values of those variables (to capture non-linearities), (2) save the residual from this regression, which represents the component of Lntenure that is orthogonalized with respect to monitoring intensity, (3) redefine the New variable, using the orthogonalized tenure variable (4) interact the updated New variable with our performance variables and (5) observe that our results are slightly weaker but our overall inferences are unchanged.

Further, our results here are sensitive to the measurement of outside CEO hire, which lacks consistency in the literature (e.g., Karaevli 2007). Specifically, an inside hire might be an outside hire who is initially hired in a non-CEO executive role in expectation that the individual will eventually succeed the incumbent CEO. Moreover, the typical research approach for defining outside hires is time-invariant, e.g., a five-year CEO incumbent who was originally hired from the outside is much likely to exhibit characteristics very similar to an inside hire after five years. Future research will likely investigate the extent to whether differences in inside versus outside CEO appointments matter for performance evaluation purposes.