This chapter begins with justifying our chosen research method and presents the techniques for empirical data collection. Moreover, we provide a comprehensive explanation of the parametric and nonparametric statistical tests we used for data analysis.
Choice of method
In order to achieve the purpose of this thesis and to investigate how the structure of the CEO compensation relates to firm performance in the auto industry, we have decided to conduct quantitative analysis of publicly listed auto firms whom disclose CEO compensation information.
Aliaga and Gunderson (2000) define quantitative research as the process of collecting numerical data and analyzing it with the help of mathematically based methods. By getting numerical data or data that is possible to transform into useful statistics, the method is used to quantify the problem (Saunders, Lewis, & Thornhill, 2011). Moreover, Creswell (2007) claims that the quantitative approach is the best when the initial problem is to identify “factors that influence an outcome, or the utility of an intervention, or understanding the best predictors of outcomes” (Creswell, 2007, pp. 21-22). The qualitative method, on the other hand, is exploratory and is thus used to gain knowledge of opinions, causes or motivations that underlie the particular problem (Saunders, et al., 2011). Our research question, “What is the relationship between CEO compensation structure and company performance in the auto industry?” leads us to use the quantitative method. This method is appropriate since our main goal is not to attempt to find the underlying reasons or motivations for why the relationship is a certain way, instead we attempt to quantify the relationship between pay structure and performance. As such, our quantitative analysis do not explain the reasons behind the findings for how the CEO compensation structure relates to the market cap percentage change, but only measures the significance of the relationship between the tested variables. The theories underlying our theoretical framework, however, are incorporated in the analysis part, and with them we provide an interpretation of the obtained statistical results.
According to Creswell (2007) quantitative studies advance the relationship among particular variables and express them in the form of questions or hypotheses.
For each statistical test, we state null- and alternative hypotheses whose rejection or acceptance determines whether there is a relationship between the variables.
The techniques and procedures used for data collection and analysis is an essential part of the research process. Taking the quantitative approach into consideration, we based our paper mainly on secondary empirical data gathered from well-known and publicly accessible sources. The information is retrieved primarily online from respective company’s websites, annual reports, Securities and Exchange Commission (SEC) filings, published research papers and surveys. We gather information for the fiscal year 2013 which is published in annual reports the 31st of March 2014. We used the database Datastream, provided by Jönköping University, in order to retrieve market cap data. The compensation and market cap information can be seen in Appendix 2, Table 3.1. The table includes 33 firms and their respective CEO remuneration broken down into base salary, bonus, stocks and stock options, pensions, total compensation, as well as the market capitalization on the 28th of March 2013 and 31st of March 2014 separately, and the market cap change in percentages.
Sample derivation and size
The First Trust NASDAQ Global Auto ETF (CARZ) corresponds to the price and yield of the equity index called NASDAQ OMX Global Auto Index (SM) which is designed to track the performance of the largest and most liquid auto manufacturing companies (Yahoo Finance, 2015). It is the most traded auto-focused ETF (Yahoo Finance, 2015). Our sample contains shares from 36 companies. This ETF tracks auto firms that have a market cap of or above $500 million, which means most well-known auto companies. According to Datastream, these auto firms, excluded by the ETF, have a market cap above $500 million: SAIC Motor, Tata Motors, Suzuki Maruti, Fiat Chrysler Automobiles, Geely and Volvo. However, 9 out of the total 42 were excluded from the sample. This is because, even though publicly traded, they do not disclose CEO compensation information, neither in their financial reports nor in filings to accessible government institutions, making our task of finding such information impossible in due time. The result is a sample of 33 companies.
Our choice of evaluation with respect to executive remuneration is limited only to the CEO of each firm and not the whole top management. The compensation information for all of the top executives is included in many annual reports, however, usually in a lump sum, therefore, we abstain of taking them into consideration.
In order to gain more information on the pay-performance relationship and CEO compensation, we have used well-known databases such as Scopus, Google Scholar, Jönköping University Library, and Primo to retrieve relevant articles. Among the key-words used when searching were CEO compensation structure, pay-performance relationship, CEO remuneration, executive pay, executive compensation, and executive remuneration. All articles were published in English but for one Swedish exception.
In order to have comparable data and to provide valid analysis, we need to convert the numbers into a common currency. We chose US dollars ($) because it is an international currency (Tavlas, 1997; Blinder, 1996). Naturally, the numbers in some of the annual reports are published in currencies, such as Euro (EUR), Japanese yen (JPY), Indian rupees (INR) and Chinese yuan (CNY). We use the spot exchange rate for 31st of March 2014, the date of the official publication of the annual reports, in order to convert those currencies into a common one.
We limit our research to only one year – 2013. This was decided due to the time limitations of our thesis and the extensive analysis and calculations required. It is very time consuming to gather compensation information. It might be advantageous to extend it to more years, but we needed to narrow it. Moreover, the lack of data for some companies included in the CARZ ETF fund, negatively affected our sample size decreasing it to 33.
The firm performance measurement
We measure firm performance by the market cap change between the beginning and the end of fiscal year 2013. Generally, the market cap is a reflection of the business development of a particular company, as well as a reflection of accounting performance. If a firm has unexpectedly high profits the share price will rise and the market cap with it (Gitman, Juchau,
Flanagan, 2010). The market cap change, through the share price, is an indication of whether the markets believe the firm has increased profitability today, or in the future. As such, the market cap change for a year includes not only an evaluation of profitability for the year, but an evaluation of future profitability as well. In order to compare and evaluate if and how the CEO compensation structure relates to market cap for the year 2013, we selected 31st of March 2013 until the 31st of March 2014 and calculated the percentage change between the beginning and the end of the fiscal year. However, the US stock market was closed from the 29th of March 2013 to 31st of March 2013 due to official Easter holidays. Considering that a significant number of the companies in our sample are listed on the US stock exchange we shifted the beginning date to 28th of March 2013.
When doing statistical analysis, parametric tests are more sensitive and robust than their nonparametric equivalents. The most widely used parametric test for correlation is Pearson product-moment correlation coefficient, known as Pearson’s r (Hauke & Kossowski, 2011). Pearson requires fulfilment of certain assumptions for the data:
(Hauke & Kossowski, 2011)
Normality is required to perform parametric tests. Most of statistical procedures such as correlation, regression, t tests, and analysis of variance, are based on assumptions that the data is normally distributed (Ghasemi & Zahediasl, 2012; Altman & Bland, 1995). For large enough sample sizes, 40 and above, a deviation from normality would not cause significant problems, meaning that we can use parametric procedures even if the data is not normally distributed, and for hundreds of observations one can simply ignore the distribution of data (Ghasemi & Zahediasl, 2012).
Normality can be visually shown in histograms (Ghasemi & Zahediasl, 2012). Altman and Bland (1995) claim that samples from a normally distributed population might not seem normal graphically if the sample size is too small. To visually inspect the distribution is not very reliable unless accompanied with significance tests, normal Q-Q plots and boxplots.
Shapiro-Wilk test is a test of normality where a resulting p-value indicates normality (Shapiro Wilk, 1995). The null hypothesis that the data is normally distributed will be rejected when the p-value is below 0.05 to a confidence interval of 95% (Shapiro & Wilk, 1995). The visual representation of histograms, normal Q-Q plots and boxplots also hint graphically to whether the tested variables are normally distributed or not.
For normally distributed data to a 95% confidence interval, the values of skewness and kurtosis divided by the standard error should be within the range of -1.96 to +1.96 (Cramer, 1998; Cramer & Howitt; 2004; Doane & Seward, 2011). The skewness is a measure of symmetry or asymmetry of the distribution, whereas kurtosis shows whether there are the right amount of data allocated in the middle for it to have the bell-shape of normality .
Outliers can be illustrated with a box-and-whisker plot (Tukey, 1977). In SPSS, for example, it detects outliers by marking those values that do not fit within 1.5 times the interquartile range (Q3-Q1). Values that are not within the range of 3 times the interquartile range are considered extreme outliers. The outliers can also be presented in a stem-and-leaf diagram.
As stated earlier, parametric tests such as Pearson’s r, require more than normally distributed data (Lund Research Ltd, 2013). Assuming that the criteria for a normally distributed data is fulfilled one need to test for linearity.
The linear regression models the relationship between two variables, and shows whether the relationship is linear or not (Seber & Lee, 2012). Again, the p-value shows whether the regression is significant for a confidence interval. If the p-value is below 0.05 it indicates that the regression model can significantly predict the value which also means that the data is linear. The R value indicates how strongly the independent variable influences the dependent (Seber & Lee, 2012).
Pearson’s r measures the strength of linear correlation between two variables (Hauke & Kossowski, 2011). It requires several assumptions about the data in order to be performed:
Approximately normally distributed data
(Hauke & Kossowski, 2011)
If these assumptions are met, Pearson identifies if there is a positive or a negative correlation, and how strong it is (Kent State University, 2014). The r-value shows the relationship direction and the p-value denotes the significance. An r-coefficient sign of -1 would indicate a perfectly negative linear relationship (downward sloping), +1 would be perfectly positive linear relationship (upward sloping), whereas 0 shows no relationship (Hauke & Kossowski, 2011).
If the assumptions of normality, linearity and no outliers are not met, nonparametric tests can be used (Hollander, Wolfe & Chicken, 2013). Nonparametric tests, because they are less sensitive at measuring differences between samples, are less robust and give less accurate information compared to parametric tests. Nonparametric tests are seldom used on large samples (n > 100), since in data of that magnitude, we can assume the sampling distribution to be normal (Hollander et al., 2013). A nonparametric test of correlation is the Kruskal-Wallis test.
Table of Contents
1.2 Problem discussion
1.4 Research questions
2 Theoretical framework
2.1 CEO compensation components
2.2 Previous findings – pay-performance relationship
2.3 Agency Theory
2.4 Stewardship theory
2.5 Market index as a benchmark
3.1 Choice of method
3.2 Data collection
3.3 Data analysis
3.4 Research Quality
4 Empirical results
4.1 Aggregated compensation structure
4.2 Total compensation
4.3 Base salary
4.5 Stocks and Stock options
4.7 Parametric Testing
4.8 Nonparametric Testing
GET THE COMPLETE PROJECT
CEO Compensation Structure and Firm Performance Evidence from the auto industry