Short-termism in business: causes, mechanisms and consequences APPENDIX. Details of the econometric analysis

Transcription

1 Short-termism in business: causes, mechanisms and consequences APPENDIX Details of the econometric analysis

2 Table of Contents Abbreviations and definitions Introduction The data used in the EY analysis Econometric analysis of the short-term performance of companies Models explaining the company s performance Short term panel models of the market cap Short term panel models of the ROE Models explaining the investment activity Econometric analysis of the long-term performance of companies Cross-section models explaining the company s performance Long-term cross-section models of the market cap Long-term cross-section models of ROE Long-term cross-section models explaining the company s investment activity EY 2

3 Abbreviations and definitions OLS IV GMM FE RE SPIQ GICS LTIP Ordinary Least Squares method Instrumental Variables method Generalized Method of Moments Fixed Effects panel estimator Random Effects panel estimator Standard & Poor s Capital IQ Global Industry Classification Standard Long-Term Incentive Plan EY 3

4 1. Introduction The purpose of this appendix is to explain the details of the econometric methods used in the EY Poland Report Short termism in business: causes, mechanisms, consequences. First, we describe the data used in the empirical study, with special attention paid to the geographical and sectoral structure of the sample. Second, we discuss panel models used in the analysis, which allow us to study the short-term impact of CEO tenure and of investment activity on a company s performance. Short-termism leads to too much focus on short-term goals at the cost of long-term objectives. Therefore, investigating this issue requires the use of econometric methods that will allow us to distinguish between short- and long-run effects of the analysed variables. The panel data includes information regarding both horizons and, if we use proper transformations of data (i.e., the Within and Between Group transformations 1 ), we can make this time distinction explicit in our models. As Baldi H. Baltagi writes, applied studies using panel data find that the Between estimator (which is based on the cross-sectional component of the data) tends to give long-run estimates whereas the Within estimator (which is based on the time series component of the data 2 ) tends to give short-run estimates 3. In the spirit of this approach, in order to explain how market capitalization and ROE are affected in the short run, we use the system GMM (Blundell-Bond method), as well as standard fixed effects (FE) and random effects (RE) specifications. As regards the measurement of the long run impact of distinct variables on the company s performance, we use the cross-section models which we estimate using the ordinary least squares (OLS) method. The above selection of different econometric models allows us to investigate whether there are differences between the impact of the selected variables on a company s performance in the short vs the long run. 1 The Within Group transformation is the deduction of the unit-level means from the values of the panel variables, whereas the Between Group transformation is simply the calculation of these means in order to remove the time dimension from the panel. 2 Critically, this component still remains a panel (this footnote has been added by EY) 3 Baldi H. Baltagi, Econometric Analysis of Panel Data, Third edition, John Wiley & Sons, Ltd, EY 4

5 2. The data used in the EY analysis We have collected information on 1024 of the largest companies listed on European stocks over the period. We have selected those firms that were the largest, in terms of their market capitalization, as of the end of 1998, and continued to be listed over the entire sample period. Basing on this selection, we have built a dataset that includes: the tenures of CEOs that were in office over the years , market capitalization, share prices, capital expenditures, and other basic financial indicators of the analyzed companies. All the variables used in the analysis, including their respective sources, are outlined in Table 1. Note that simple multiplication of the number of firms and years gives us as many as observations in our sample data. However, depending on the model specification, there are more or less missing data in the case of some variables, which reduces the number of observations actually used. 4 Outliers are another reason for reducing the number of observations. Following a standard statistical and econometric practice, we remove them from our dataset. We remove a particular observation when: the stock variables have negative values (especially assets); debt/equity ratio is higher than 2000%; ROE is higher than 500% or lower than -500%; capex/total revenue ratio is higher than 1000%. Inclusion of such extreme observations would make many of the statistical results not applicable for the rest of the firms as it might bias the average estimates. In sum, we remove 131 observations. However, not all observations are used in all parts of our empirical analysis, making the number of observations effectively removed different across particular econometric models. For instance, in the panel models, the number of observations removed is very marginal and varies from 34 to 56. Firms within our sample are listed on stock exchanges in 24 countries. In particular, 301 companies are listed in the United Kingdom, 143 in France and 101 in Germany with these three countries accounting for nearly half of the sample. The geographical structure of the sample is outlined in Chart 1. 4 However, it should be noted that we do not need the values of a particular variable for a given firm for all years in the sample period, but for at least one year. For instance, if we observe the time in role variable in the Firm A for 16 years and in the Firm B for 14 years, there is no need to delete two observations in the Firm A, because the standard panel estimators make adequate adjustments in order to use all 30 observations available. However, if we consider the Firm C with no information on CEO tenure whatsoever, we have to remove the Firm C from any estimation that requires the use of the time in role variable. EY 5

6 Table 1. The variables used in the study and their sources. Variable Description Unit Source Market cap Market capitalization of a firm at the end of the calendar year. The value of this particular variable in 1998 has been the criterion of choosing the largest companies listed on the European stock markets. millions EUR Standard & Poor's Capital IQ Close price Share close price at the end of the year. This variable has also been used to calculate the average sectoral price indices in order to control for sectoral effects in the regressions for the market cap. EUR Standard & Poor's Capital IQ ROE Return on equity over the year. % of equity Standard & Poor's Capital IQ Debt/equity Debt to equity ratio at the end of the year. % of equity Standard & Poor's Capital IQ Time in role Tenure of incumbent CEO. When there was more than one CEO in a company in a given year, average of their tenures is used. Years BoardEx Share of LTIP in overall remuneration of the CEO Conditional compensation in the form of the Long Term Incentive Program (LTIP) granted in a given year, divided by the total CEO remuneration calculated as a sum of salary, bonus and LTIP. % of total remuneration EY calculations based on BoardEx Capex/revenue Capital expenditures (additions to fixed assets) divided by the total annual revenue. % of total revenue EY calculations based on Thomson One (Capex) and Thomson Reuters Eikon (Total Revenue) Capex/assets Capital expenditures (additions to fixed assets) divided by the total assets at the end of the previous year. The total assets are computed as a sum of total equity and total debt at the end of the previous year. % of total assets EY calculations based on Thomson One (Capex) and Thomson Reuters Eikon (Total Equity and Total Debt) Outsider Binary variable equal to one if the incumbent CEO was appointed from outside of the company, and zero when CEO was an insider successor. Binary variable EY calculations based on BoardEx Source: EY. EY 6

7 Chart 1. Regional structure of the sample (N=1024) 5. Eastern Europe 3% Southern Europe 12% Western Europe 37% Scandinavia 18% British Isles 30% Source: SPIQ, EY calculations. The firms within our sample have been divided according to the Global Industry Classification Standard (GICS) developed by MSCI and Standard & Poor s (S&P), which consists of 10 sectors, 24 industry groups (see Table 2), 67 industries and 156 sub-industries (not reported in this study). This taxonomy is widely used to construct sectoral sub-indices of listed companies and therefore we apply it for the purpose of our study. 5 British Isles include Great Britain and Ireland, Western Europe Netherlands, Austria, Belgium, France, Germany, Switzerland and Luxembourg, Scandinavia and Finland Denmark, Finland, Sweden, Norway, Southern Europe Greece, Portugal, Spain, Italy, Malta, and Eastern Europe and Balkans Hungary, Croatia, Czech Republic, Slovakia, Slovenia and Poland. EY 7

8 Table 2. Sectors in the sample according to the GICS, used for the description of the data. Code Sector Subcode Industry Group 10 Energy 1010 Energy 15 Materials 1510 Materials 2010 Capital Goods 20 Industrials 2020 Commercial & Professional Services 2030 Transportation 2510 Automobiles & Components 2520 Consumer Durables & Apparel 25 Consumer Discretionary 2530 Hotels Restaurants & Leisure 2540 Media 2550 Retailing 3010 Food & Staples Retailing 30 Consumer Staples 3020 Food, Beverage & Tobacco 3030 Household & Personal Products 35 Health Care 3510 Health Care Equipment & Services 3520 Pharmaceuticals & Biotechnology 4010 Banks 40 Financials 4020 Diversified Financials 4030 Insurance 4040 Real Estate 4510 Software & Services 45 Information Technology 4520 Technology Hardware & Equipment 4530 Semiconductors & Semiconductor Equipment 50 Telecommunication Services 5010 Telecommunication Services 55 Utilities 5510 Utilities Source: MSCI, Standard & Poor s. The structure of the sample according to GICS is outlined in Chart 2. EY 8

9 Chart 2. The sectoral structure of the sample of firms (N=1024). Materials 10% Telecommunication Services 1% Utilities 2% Consumer Discretionary 19% Information Technology 9% Consumer Staples 7% Energy 4% Industrials 28% Healthcare 6% Financials 14% Source: SPIQ, EY calculations. EY 9

10 3. Econometric analysis of the short-term performance of companies The primary purpose of the short-term econometric analysis is to quantify the impact of the time in role of a CEO and of the firm s investment activity on the short-term performance of a company. This performance is approximated by the changes in the company s market capitalisation and ROE. In this section, in order to achieve short-term estimates, we use the fixed effects (FE) panel estimation techniques and System General Method of Moments (GMM). We also consider the random effects (RE) panel estimation as an alternative (where applicable) method, which may give similar estimates to FE method (the similarity of estimates can be tested using the Hausman test). When justified, we additionally use a pooled OLS model. We estimate two kinds of panel models: models explaining the company s performance: the models of the market cap; the models of ROE; models explaining the company s investment activity: the models of the capex/total revenue ratio Models explaining the company s performance To start with, panel models explaining the performance of the i-th firm in the year t are based on the following equation: where: performance it = m i + gperformance it-1 + β 1(time in role) it + β 2outsider it + β 3(capex/total revenue) it + β 4(LTIP share) it-1 + β 4+s(sector performance) sit + ε it, (1) performance it stands for either the market cap (with necessary transformations) or ROE variable. Depending on the explained variable, we select different explanatory variables in particular econometric models. For instance, the dynamic model of the logarithm of the market cap variable is the only model with a lag of the explained variable on the right-hand-side of the corresponding econometric equation. In other cases the specification is static, i.e., g=0. Intercept m i is the firm level individual effect (indicating any kind of the characteristics of an individual firm that do not change over time), whose inclusion is crucial for assuring the comparability of different firms and drawing conclusions as regards all of them. In case of the FE approach, this is a deterministic dummy, and in RE approach, this is the part of the error term and therefore is treated as a random variable. (sector performance) sit is the control variable describing the situation in sector s (average close price in models of the market cap and average ROE in models of ROE). Importantly, this variable equals zero if the firm i does not belong to sector s. As a result, the sum of 10 sectoral variables in EY 10

11 equation (1) means that we consider only the impact of the sector that the firm i belongs to, while the impact of the remaining sectors on the company s performance equals zero. Other explanatory variables: (time in role) it, outsider it, (capex/total revenue) it and (LTIP share) it have been defined in Table 1, Finally, ε it is the idionsyncratic shock affecting the performance of a particular company in period t. Transformations and roles of explanatory variables are outlined in Table 3. Table 3. The description of explanatory variables in the models of a company s performance. Variable Transformation Role performanceit-1 It is logarithm of the first lag of the dependent variable and is considered only in the case of the dynamic panel model explaining the logarithm of the level of the market cap. This variable allows us to account for the persistence inherent in the market cap, which is a state variable. For instance, large firms in one year are likely to remain large in the following, which has little to do with the performance of a particular CEO. Inclusion of the first lag of the dependent variable in the model explaining the logarithm of the level of the market cap makes it possible to see the additional value generated due to the stability of management (or the investment activity of the firm). (time in role)it This variable remains untransformed in all the panel models. Shortened CEO tenure, and thus management instability, is one of the symptoms of short-termism. Therefore, measuring its impact on the performance of a company plays a pivotal role in our analysis. In particular, we want to investigate whether CEO tenure positively affects company s performance already in the short-term or maybe it is statistically significant in the long-term only (if at all). Since the time in role accumulates over time, only a fraction of this variable (i.e. up to one additional year) might be influenced by the explained variable, i.e. the current performance of a company, which, if favorable, can reduce the risk of shortening CEO tenure. Therefore, the endogeneity issue should not be a big problem here. Nevertheless, we (partly) address it through the use of some robust specifications (GMM). outsiderit (capex/total revenue)it This variable remains untransformed in all the panel models. This variable remains untransformed in all the panel models. It is a ratio of two flow variables measured over the entire year. We investigate whether it matters for the performance of a company whether the CEO was appointed from the outside or was an insider successor. The latter should have more knowledge of how the firm functions whereas the former may bring a fresh perspective to the company. We test empirically the relative importance of these factors for the performance of a firm. We investigate whether an increase in investment may deteriorate short-term financial performance of a company. If confirmed, it would support the hypothesis that executives excessively focused on short-term targets may be discouraged from increasing investment expenditure. We include this variable in the model of the market cap. Since the investment activity has no impact on the contemporaneous profits, it is not included in the models of ROE. EY 11

12 Variable Transformation Role (LTIP share)it-1 This variable remains untransformed in all the panel models. It is a ratio of flow variables measured over the entire fiscal year. We use the lagged value of this variable (we investigate also more lags) We investigate whether increasing the share of CEO remuneration that is conditional upon the future performance of a company positively affects the motivation of the executives and, consequently, the future performance of a company. It is therefore obvious that we have to apply the lagged LTIP share variable. One might claim that due to a few-year horizon of the conditional LTIP-remuneration, this motivational effect may not be visible already next year. We verify empirically whether this is the case. In the case of the dynamic panel model of the market cap, the sector performance variable is a logarithm of the average close price in the sector s (see Table 1). We use the average close price instead of the average market cap because the average close price constitutes a price index that can play the role of a deflator of the market cap. Moreover, by using the lagged variable, we avoid the problem of endogeneity (i.e. LTIP being granted as a result of the company s good performance). (sector performance)sit In the case of the non-dynamic panel model of the market cap, this variable is the first difference of logarithm of the average close price in the sector s. In the case of the model of ROE, it is the average ROE in the sector s. While doing these transformations, we deal with the problem of endogeneity (arising from the fact that the performance of a particular company might affect the performance of the whole sector) by using the arithmetic averages. This allows us to reduce the impact of big market players on the aggregate measure of the sectoral performance. Part of the performance of the company may result from the factors that are typical for the sector that this firm belongs to. In particular, demand or supply shocks affecting that sector would most likely impact on a given company irrespective of its executives decisions. We extract this effect by including sectoral variables in our models. Source: EY analysis. Crucially, the sectoral variable for the sector s is transformed in such a way that it impacts only on those firms that belong to that sector. Therefore, in spite of the fact that for each firm all 10 sectoral controls are used, only one of them (describing the situation in the sector corresponding to the firm i) has an impact on the performance of a given firm, while other sectoral variables are set to zero. As it has been stressed in the Report, market cap and ROE are different kind of variables, as the former is a stock variable, and the latter is a flow variable, which requires a separate econometric treatment Short term panel models of the market cap In the case of the market cap, the fundamental problem is the fact that this variable describes some accumulation over time, which means that the level itself says little about the impact of the tenure of a particular CEO or of the current investment activity. In particular, a new CEO (with zero tenure) might inherit a firm with a high market value in relation to other firms from the sample. We propose to resolve this issue through the use of: EY 12

13 the dynamic panel model; the panel model in first differences. In the dynamic panel model, the market cap 6 is explained, inter alia, through its own lag. In terms of equation (1) this means that g 0. As a result, owing to the ceteris paribus assumption, the impact of all other explanatory variables does not include the influence of the lagged performance variable. This means that the persistence in the market cap variable does not affect our estimates of the impact of the CEO tenure, company s investment activity and other variables on the performance of a firm. In the case of dynamic panel models with the fixed effects, the fundamental problem of endogeneity of the lagged dependent variable arises (the so called dynamic panel bias). In the case of simple OLS and FE estimators, inclusion of the lagged market cap as an explanatory variable generates the endogeneity problem by nature. There are several econometric methods to estimate the parameters of such dynamic panel models consistently. According to the recent literature, the Blundell and Bond system general method of moments (referred to as GMM in this Appendix) is relatively robust to problems that often arise in econometric analyses of the corporate finance datasets (which means that the GMM estimates are not severely biased as a result of some problems with these data, such as persistence in the dependent variables, endogenous explanatory variables, short sample period available) 7. The idea of the system GMM is to estimate the parameters of two econometric equations at the same time: untransformed and transformed one (which is an auxiliary equation). In the latter case, we use special mathematical transformation of the main equation (in levels) 8. The main issue is the proper selection of the instruments for both equations that are used in the estimation. The instruments are generated according to some automated rules using the transformations and levels of the variables from within the model in such a way that we preserve their explanatory power, but also get rid of their correlation with the error term (which is the source of endogeneity). Such an approach makes the pool of instruments potentially huge. The higher the number of instruments we use, the more precise become our estimates 9. However, the ultimate goal of the system GMM method is to remove the natural endogeneity of the lagged dependent variable and cope with the potential endogeneity of the remaining right-hand side variables in the equation (1). Notably, if we use too many instruments, we fail to remove the 6 We take the logarithm of this variable in order to reduce the problem of heteroskedasticity. 7 Mark J. Flannery, Kristine Watson Hankins, Estimating dynamic panel models in corporate finance, Journal of Corporate Finance 19 (2013) We apply this method using a popular Stata package xtabond2. 8 We use a special orthogonal deviations transformation instead of first differences in order to save the degrees of freedom, as suggested by Roodman (2009), op. cit. 9 We use the two-step GMM estimator to cope with the possible heteroskedasticity, but the estimates of standard errors in this method tend to be downward-biased. In order to account for this issue, we use the Windmeijer correction of the standard errors. EY 13

14 bias resulting from endogeneity as we over-fit the first stage regression 10. To avoid that, we should set the number of instruments that is not too high relative to the number of units in the panel. Reducing the maximum lag of the instruments 11 to 6 allows us to keep the overall number of instruments within the safe range, below the number of firms that enter the regression (265 instruments vs. 288 firms) 12. Description of the instrument sets both for the transformed and the levels equations is provided in Table 4. The number of lags of instruments used in the levels equations is typically lower than this is the case for the transformed equations. This is because once we include the instruments for the transformed equation, most of the instruments for the levels equation become mathematically redundant. There is a number of ways to validate particular choices as regards the GMM specification. Notably, the estimated coefficient for the first lag of the dependent variable should be somewhere between the estimates given by the ordinary least squares (OLS) and FE methods, as the former method produces upward bias, whereas the latter downward bias for this autoregressive coefficient. These estimators produce the limiting, plausible cases of the autoregressive coefficient, which allows us to conclude whether our choices as regards the specification of the GMM method are valid. In our case, the coefficient produced with the FE method equals 0.498, with the OLS method: 0.962, whereas with the system GMM we obtain 0.776, which is within the above plausible limits. Other criteria include the Hansen test of overidentifying restrictions, in which we do not reject the null hypothesis of the exogeneity of the instrument set (p-value equals ). The final test included in this analysis is the Arellano-Bond test for serial correlation (see also Table 4), which points at serial correlation of order 3 in the first differences equation used for the test, which means serial correlation of order 2 of ε it ( p-values are close to 0 up to order 3, and in the case of order 4 the p-value equals 0.091). We do not consider it as a fundamental problem, though, as we adjust the instrument set to account for this problem (see Table 4). Also, according to the latest findings as regards the corporate finance datasets, in spite of the fact that the serial correlation of order 2 is a breach of the assumptions underlying the use of the system GMM method, the consequences seem to be relatively limited We define the first stage regression as a regression of each of the endogenous variables on the full set of instruments and exogenous variables. The purpose of this regression is to remove the endogenous term, leaving only the linear combinations of exogenous variables for the second stage, main regression. Therefore, the over-fitting means that these linear combinations are not much different from the initial endogenous variables leaving us with the biased estimates of the main regression coefficients. 11 Here we talk about the lags of those instruments that are created using all but the strictly exogenous variables. In case of the strictly exogenous variables, we consider only the contemporaneous values. 12 Roodman (2009), op cit. 13 We use the significance level of Blundell-Bond system generalized method of moments appears to be the best choice in the presence of endogeneity and even (surprisingly) second order serial correlation if the dataset includes shorter panels, Flannery and Hankins (2013), op cit., p. 16. EY 14

15 Table 4. The choice of the instruments for the system GMM estimation. Basis for the instruments Endogeneity of the basis variable Lags of the levels, the transformed equation. Lags of the first differences, the levels equation. log(market cap)it-1 Fist lag of the market cap is the source of the dynamic panel bias. Due to serial correlation of order 2 (according to the Arellano-Bond test), we use the lags of order from 4 to 6 rather than of order from 2 to 6. Due to serial correlation of order 2, we use the lag of order 3. (time in role)it (capex/total revenue)it Time in role might be to some extent shaped by the log of the market cap (good performance of the company increasing the probability of extended tenure) we account for that potential source of endogeneity as well. Investment activity may, to some extent, be shaped by the market capitalization of a firm (e.g. a decline in the market cap may make executives reduce investment expenditure to achieve a short-term increase in the share price) - we account for that potential source of endogeneity. We use the lags of order from 2 to 6. We use the first lag only. (LTIP share)it-1 We treat the first lag of the LTIP share variable as the predetermined variable. This requires instrumentation as well. We use the lags of order from 1 to 6. We use only the contemporaneous value (zero lag). outsiderit (sector performance)st We consider the outsider variable as strictly exogenous. We consider the average share price strictly exogenous by construction. Both the outsider variable and the sectoral controls are considered strictly exogenous. This means that we use the contemporaneous values of these variables as the instruments for themselves (transformations in the transformed equation and levels in the levels equation) in a collapsed form. Source: EY analysis. Another way to account for the fact that the market cap is a stock variable is estimation of the model in first differences. We simply calculate the first differences of logarithms of the market cap and insert it as the performance variable into equation (1) together with the first-differences of logarithms of the sectoral controls. Such a simple solution allows us to capture the impact of time in role and capex/total revenue ratio on marginal changes in the market cap, without the need to resort to dynamic specification (denoted as GMM in Table 5). At first, we estimate these models using the standard fixed effects panel estimator (FE) and random effects panel estimator (RE). The Hausman test does not reject the null hypothesis stating that the RE model is consistent (with the p-value equal to 0.690), which is related to the fact that the parameters of the FE and RE models are relatively similar (especially in terms of the sectoral controls), with the RE estimates having lower standard errors. We report both FE and RE specifications in Table 5. In addition, we carry out an F test, which does not reject the null of individual effects in the FE specification being equal to zero (with the p-value equal to 0.580). This suggests that the pooled, OLS model is also appropriate for the model of the market cap in first differences and we present it in the Table 5 as well (OLS). We can also compare the OLS and RE model using the Breusch-Pagan test for random effects. This test rejects the null hypothesis which states that there are no random effects (i.e., both OLS and RE models are consistent, but OLS has lower standard errors), with the p-value equal to The estimates in the OLS and RE models turn out to be largely the same. EY 15

16 Another step, in the short-term FE models of market cap, is to test for heteroskedasticity. We carry out the modified Wald test for groupwise heteroskedasticity in the fixed effects regression model, which leads to the rejection of the null hypothesis of no heteroskedasticity (with the p-value equal to 0). This problem is typical of the panel models and we have accounted for that partly by introducing the log of the explained variable in the models of the market cap. Moreover, for the sake of the robustness check, we calculate the robust standard errors of the FE estimator as well (FE with robust standard errors). The results presented in Table 7 show that it does not affect our initial conclusions as regards the significance of the FE estimates. Note that the GMM method uses the logarithm of the level of the market cap as the performance variable, whereas the FE, FE with robust standard errors, RE and OLS specifications use the first differences of the logarithm of the market cap. The tests that we carried out do not suggest that the estimates of the coefficients in any of these models are inconsistent. The estimation results show that CEO tenure has a positive influence on the market cap of a company in the short-term, however it is not always statistically significant. In particular, it is statistically significant (under 5% significance level) in the RE and OLS models, in the GMM it is significant only under 10% significance level, while in the FE and FE with robust standard errors models it is not significant under any standard significance levels. The obtained results are much more robust across different models when it comes to the impact of investment activity on the market cap of a company. All the models indicate a statistically significant negative impact of an increase in investment activity on a short-term market valuation of a firm (see the negative coefficient for the capex/total revenue variable). In other words, in the short-term, shareholders tend to punish an increase in the firm s investment expenditure through the sale of shares in the company and a subsequent decline in the share price. In the case of the GMM and simple FE specifications, the coefficient for capex/total revenue ratio is similar and significant, whereas in the RE and OLS models, it is still significantly negative, but considerably smaller (in absolute terms). In addition, the estimation results show that the lag of the LTIP share variable is insignificant in all the specifications. The outsider variable is also insignificant in all the cases but the GMM, where its impact on the market cap is negative. This might suggest that an insider successor provides more beneficial experience to the company relative to the CEO coming from the outside. The estimation results for sectoral variables, not least their statistical significance, confirm the need of their inclusion in our models. EY 16

17 Table 5. Short-term panel models of the market cap (logarithm in case of GMM and first differences of logarithms in the case of FE, RE, and OLS). log(market cap it-1 ) time in role it capex/total revenue it outsider it LTIP share it-1 Consumer Discretionary Consumer Staples Energy Financials Healthcare Industrials Information Technology Materials Telecommunication Services Utilities GMM Source: EY analysis. Note: p-values in parentheses, * p < 0.10, ** p < 0.05, ***p < FE Autoregression coefficient FE (robust s.e.) 0.776*** (0.000) RE OLS Main variables of interest 0.011* ** 0.005** (0.091) (0.375) (0.392) (0.041) (0.012) *** *** *** *** *** (0.000) (0.000) (0.006) (0.001) (0.001) Individual control variables *** (0.003) (0.642) (0.616) (0.830) (0.808) * 0.087* (0.458) (0.191) (0.240) (0.071) (0.060) Corresponding sectoral controls 0.457*** 1.114*** 1.114*** 1.128*** 1.131*** (0.000) (0.000) (0.000) (0.000) (0.000) 0.402*** 0.914*** 0.914*** 0.930*** 0.930*** (0.000) (0.000) (0.000) (0.000) (0.000) 0.771*** 0.898*** 0.898*** 0.911*** 0.917*** (0.000) (0.000) (0.000) (0.000) (0.000) 0.343*** 0.975*** 0.975*** 0.989*** 0.968*** (0.000) (0.000) (0.000) (0.000) (0.000) 0.530*** 1.235*** 1.235*** 1.271*** 1.285*** (0.000) (0.000) (0.000) (0.000) (0.000) 0.360*** 0.822*** 0.822*** 0.812*** 0.809*** (0.000) (0.000) (0.000) (0.000) (0.000) 0.608*** 0.858*** 0.858*** 0.820*** 0.788*** (0.000) (0.000) (0.000) (0.000) (0.000) 0.462*** 1.080*** 1.080*** 1.086*** 1.086*** (0.000) (0.000) (0.000) (0.000) (0.000) 0.596*** 1.571*** 1.571*** 1.521*** 1.517*** (0.000) (0.001) (0.000) (0.001) (0.001) 0.662*** 0.560** 0.560*** 0.583** 0.592** (0.000) (0.022) (0.000) (0.012) (0.012) The intercept * (0.917) (0.624) (0.656) (0.135) (0.095) Additional model information Number of observations Number of firms Number of instruments R EY 17

18 3.1.2 Short term panel models of the ROE The other key performance variable that we analyse is ROE. We insert it as the performance variable in equation (1), set g to 0 and insert averages of ROE as the sectoral controls. Since ROE is a ratio and, in terms of its numerator, it is a flow variable, it does not require any transformations (neither logarithms nor differencing) or development of a dynamic specification of the model 15. However, we remove investment activity from explanatory variables, because we should not expect its effects to materialise so quickly (in the same financial year) to be reflected in the company s short-term profits. We can now use the FE and RE panel estimators to estimate the coefficients of the econometric equations for ROE (see Table 6). We again carry out the Hausman test, which does not reject the null hypothesis stating that the estimates in the RE model are consistent (with the p-value equal to 0.992). Therefore, both the RE and FE models produce consistent estimates. One may note that the estimates obtained in both models are similar. In addition, the modified Wald test for groupwise heteroskedasticity in the fixed effects regression model once again leads to the rejection of the null hypothesis of no heteroskedasticity (with the p-value close to 0). In order to address this issue, we calculate the robust standard errors using the heteroskedasticity-consistent White estimator. However, it does not affect our conclusions for the main variables of interest in the FE model, not least their statistical significance. Finally, the F test does reject the null of no individual effects (with the p-value close to 0), so we do not include the pooled, OLS model. Moreover, the Breusch-Pagan test for random effects rejects the null hypothesis of no random effects (with the p-value close to 0), which is yet another argument against the inclusion of the OLS model. According to all the short-term models of ROE, time in role leads to an increase in the profitability of a firm. It is worth to note that the estimated coefficient is statistically significant under 1% significance level across all the models. Moreover, the lagged LTIP share variable has a strong and positive impact on the company s profitability, which suggests that long term incentive plans motivate CEOs to achieve higher profits already in the short term. Finally, we find a negative impact of appointing an outsider on the company s ROE in the short term, though this effect is not always statistically significant. 15 In order to make sure that the lack of dynamic specification does not affect the results for ROE, we also included the lagged value of ROE in the FE and RE models. It did not change our conclusions, which means that the problem of serial correlation is not relevant in these models. EY 18

19 Table 6. Short term panel models of ROE. time in role it outsider it LTIP share it-1 Source: EY analysis. Note: p-values in parentheses, * p < 0.10, ** p < 0.05, ***p < FE FE (robust s.e.) RE Main variable of interest 0.005*** 0.005*** 0.005*** (0.001) (0.007) (0.000) * *** (0.058) (0.282) (0.009) 0.087*** 0.087*** 0.093*** (0.003) (0.012) (0.001) Corresponding sectoral controls 0.961*** 0.961*** 1.046*** Consumer Discretionary (0.000) (0.001) (0.000) Consumer Staples Energy Financials Healthcare Industrials Information Technology Materials Telecommunication Services Utilities Individual control variables *** 0.986*** (0.244) (0.003) (0.001) ** 0.635* (0.122) (0.018) (0.062) 0.918*** 0.918*** 0.840*** (0.000) (0.000) (0.000) (0.933) (0.936) (0.878) 0.869*** 0.869** 0.980*** (0.000) (0.015) (0.000) 0.764*** 0.764** 0.739*** (0.000) (0.027) (0.000) 0.956*** 0.956*** 0.978*** (0.002) (0.003) (0.000) 0.873*** 0.873** 0.839*** (0.003) (0.020) (0.001) ** 0.907** (0.175) (0.027) (0.018) The intercept (0.797) (0.838) (0.803) Additional model information Number of observations Number of firms R EY 19

20 3.2. Models explaining the investment activity In order to check whether short-termism through the shortened CEO tenure channel may lead to the neglect of investment activity, we estimate the models explaining the short-term investment activity of the i-th firm in the year t. In general, we specify these models basing on the following equation: (capex/total revenue) it = m i + β 1(time in role) it + β 2outsider it + β 3(LTIP share) it-1 + β 3+s(sector investments) sit + ε it, (2) with all the variables defined as in the short-term models of ROE. The transformations and roles of particular explanatory variables in the investment model framework are outlined in Table 7. Since the capex/total revenue is a flow variable (flows in both the numerator and denominator), we estimate the models of investment activity using a similar approach to that applied in the models of ROE. The results of our estimation are presented in Table 8. The Hausman test does reject the null hypothesis stating that the RE model is consistent (with the p-value close to 0), so we should treat FE as the main model here. The modified Wald test for groupwise heteroskedasticity in the fixed effects regression model leads to the rejection of the null hypothesis of no heteroskedasticity (pvalue is close to 0). Therefore, we present the robust FE estimates as well. The correction of the standard error weakens the statistical significance of the estimate for the time in role (it is no longer significant under 5% significance level, but remains statistically significant under 10% significance level). Finally, the F test does reject the null of no individual effects (with the p-value close to 0), so we do not include the pooled, OLS model. Also, the Breusch-Pagan test for random effects rejects the null hypothesis of no random effects (with the p-value close to 0), which is yet another argument against the inclusion of the OLS model. The main conclusion from the estimation results is that the time in role positively influences the company s propensity to invest. Therefore, an increased CEO tenure may help to counteract the neglect of investment resulting from too much focus on short-term goals (see section 3.1.1). As for individual control variables (outsider and LTIP share), the estimation results do not confirm their impact on the company s investment activity. EY 20

21 Table 7. The description of explanatory variables in the models of investment activity. Variable Transformation Role (time in role)it outsiderit (LTIP share)it-1 In relation to the corresponding variable described in Table 1, this variable remains untransformed in all the panel models. By construction, it is a state variable. This variable remains untransformed in all the panel models. It is a ratio of flow variables measured over the entire fiscal year. In order to avoid the problem of endogeneity, we use the lag of this variable (see Table 3). Shortening CEO tenure is one of the symptoms of short-termism. We investigate whether it may affect executives decisions on capital outlays, and thus impact on the company s investment activity. We investigate whether it matters for the company s investment activity if a new CEO is appointed from the outside. In particular, one might claim that an outsider may need more time than an insider successor to learn how the company functions, which may result in reduced investment activity of the firm. We investigate whether CEO remuneration conditional upon longer-term performance of a company influences current investment decisions and therefore helps to address the problem of the neglect of investment activity. (sector investments)sit Source: EY analysis. This variable stands for the average capex/total revenue in the sector s. We reduce the problem of endogeneity by using the arithmetic averages. This allows us to reduce the impact of big market players on the aggregate measure for sectoral investment activity (see Table 3). Crucially, the sectoral variable for the sector s is transformed in such a way that it impacts only those firms that belong to that sector (see Table 3). These variables are responsible for the control of capital intensity typical of a given sector. It is natural that some sectors invest more and some invest less. These sectoral controls also grasp the technological change. For instance, the entire sector might engage in implementation of some new technology. We extract this effect by including the sectoral variables in our models. EY 21

22 Table 8. Short term panel models of the investment activity. time in role it outsider it LTIP share it-1 Source: EY analysis. Note: p-values in parentheses, * p < 0.10, ** p < 0.05, ***p < FE FE (robust s.e.) RE Main variable of interest 0.002*** 0.002* 0.003*** (0.007) (0.081) (0.001) (0.249) (0.178) (0.115) (0.919) (0.939) (0.802) Corresponding sectoral controls *** Consumer Discretionary (0.239) (0.000) (0.484) Consumer Staples Energy Financials Healthcare Industrials Information Technology Materials Telecommunication Services Utilities Individual control variables (0.927) (0.710) (0.811) 0.688*** 0.688* 0.643*** (0.000) (0.071) (0.000) *** (0.629) (0.680) (0.000) (0.806) (0.583) (0.614) 0.917** 0.917*** 0.465* (0.012) (0.000) (0.071) ** (0.494) (0.018) (0.710) (0.935) (0.677) (0.164) * 0.736** (0.226) (0.067) (0.025) (0.912) (0.388) (0.152) The intercept (0.105) (0.114) (0.206) Additional model information Number of observations Number of firms R EY 22

23 4. Econometric analysis of the long-term performance of companies The primary purpose of the long-term econometric analysis is to quantify the impact of the length of CEO tenure and of the investment activity on the long-term performance of a company, approximated by the changes in the long-term growth of the market cap and the average level of ROE. In order to study the long-term consequences of short-termism, we apply cross-section econometric models because, according to the already cited book of Baltagi, cross-section models tend to give the long-term estimates. The idea of cross-section models based on transformations of the panel data (the Between Group transformation, see Section 3) is to calculate some kind of summary of each variable for the entire sample period, which means that in each case we consider only the firm dimension, without the time dimension. We estimate the following groups of cross-section models: the models explaining the company s performance: the models of the market cap; the models of ROE; the models explaining the investment activity: the models of the capex/total revenue ratio; the models of the capex/assets ratio. In this section all the models are estimated using the OLS method, with possible extensions to account for non-spherical error terms. Our long-term estimates are subject to much higher estimation errors, as they use around 700 observations, compared to the short-run estimates, which use more than 2000 observations Cross-section models explaining the company s performance The general form of the models explaining the company s performance is defined for the i-th firm without the time dimension: performance i = m + β 1(time in role) i + β 2(time in role) 2 i + β 3outsider i + β 4(capex ratio) i + β 5(sector performance) i + ε i, (3) where performance i stands for the long-run transformation of the performance variable from the panel models. In the case of the market cap, it is the logarithmic growth rate over the whole sample period, whereas in the case of the ROE, it is the sample average. Such an approach reflects the fact that the former is the state variable and the latter is the flow variable (in terms of the numerator of the ROE ratio). At first glance, we may seem to apply different approaches to modelling of the EY 23

24 market cap and ROE. However, in both cases we actually deal with roughly the same kind of transformation of the variables that we used earlier in the FE and RE panel models. This is because the sample logarithmic increment is in fact proportional to the average of the first differences of logarithms used in the panel models. 16 When it comes to the explanatory variables, (time in role) i and outsider i are the averages of the respective panel variables over the entire sample period available for the i-th firm, (capex ratio) i is the average of the capex/total revenue or the capex/assets variable, (sector performance) i is the respective sectoral performance variable for the sector that the i-th company belongs to (in the case of the cross-section models we use the sectoral control in the form of a single vector). In the case of models of the market cap we use the logarithmic growth rate of the average close price in the corresponding sector over the whole sample period, whereas in the case of ROE, we use the average sectoral ROE. Finally, we denote the error term with ε i and the model constant with m, (the latter with no economic interpretation in the cross-sectional framework). Note that in cross-section models of the company s performance we account for a potential nonlinear (quadratic) impact of the time in role on the performance of the firm. This is to investigate whether the marginal benefits from every additional year in the CEO s chair are decreasing. One may also notice that the (capex ratio) i in equation (3) is expressed as capex/total revenue (for the market cap model similarly to the short-term models) or as capex/assets (for the ROE model). We do not use the capex/total revenue in the latter case due to the positive relation between the total revenue and the profits of a company (i.e., the numerator in the explained variable). Using the capex/assets ratio instead allows us to solve the above problem. Finally, we do not include the LTIP share variable as it is clearly endogenous in the cross-sectional framework the average performance of a company affects the average LTIP share, because the latter is often granted as a reward for good results. 17 The transformations and roles of particular explanatory variables within the performance model framework are outlined in Table Strictly speaking, the econometric equivalence is not complete as we work with an unbalanced panel, but the differences are actually small. 17 As an alternative solution, instead of the LTIP share variable, we introduced its volatility (standard deviation to mean ratios) over the period. It has allowed us to cope with the problem of endogeneity of the remuneration. However, the estimation results have shown that this new variable is not statistically significant. EY 24