Contrastación de las Hipótesis Específicas

This section discusses the process of data analysis. This process includes preliminary analysis, multivariate analysis and robustness checks. Each of these phases is outlined below.

4.7.1 Preliminary Analysis

Preliminary analysis of the data includes descriptive statistics, univariate analysis and correlation matrix. Descriptive statistics describe the data in terms of central tendency and shape of distribution on a single variable in an organised form. Central tendency tests include the mean, median, standard deviation, minimum and maximum, while the shape of distribution includes the skewness and kurtosis tests. In particular, while skewness measures the symmetry of distribution, kurtosis tests the peakedness (flatness) of distribution. Both measures are used to test for the

95

normality of data distribution. According to Abdul Rashidah and Ali (2006); and Haniffa and Hudaib (2006), a normal distribution of the data requires that the standard of skewness and kurtosis are ± 1.96 and ± 2 respectively. In addition to descriptive statistics, univariate analysis is performed to test whether the mean value is significantly different from zero for firms that engage in high EM compared with those that engage in low EM.

The correlation between sample variables is made by pairwise correlation matrix to explain the degree of linear relationship between two variables. The correlation coefficient is in a range between -1 to +1, where ±1 correlation refers to a perfect linear relationship between variables. According to Gujarati (2003), a higher degree of correlation coefficient between independent variables may affect the results of regression analysis because of the multicollinearity problem, and the study recommends ±80% or above as the beginning of the multicollinearity, which affects the regression results.

4.7.2 Multivariate Analysis

4.7.2.1 Regression Analysis

Statistical multivariate data analysis methods can be classified, in general, under two broad categories: the parametric and non-parametric methods. However, the nature and characteristics of the data determine which method should be applied. Therefore, Gujarati (2003) suggests five fundamental assumptions to be tested before choosing the multivariate analysis model. These assumptions include the following:

96

1. Normality: This assumption requires that the data be normally distributed. 2. Linearity: This assumption suggests that the relationship between dependent

and independent variables should be linear.

3. Independence: Under this assumption, the error term of one observation should not be correlated with the error terms of other observations.

4. Heteroscedasticity: This assumption requires that the variance of the dependent variable be constant.

5. Multicollinearity: This assumption suggests that the collinearity among independent variables should be not exist.

The above assumptions are checked using several tests to determine which approach (i.e. the parametric and non-parametric methods) is more appropriate for the study data. First, Skewness and Kurtosis tests are applied to check for the normality. Second, the Quantile–Quantile (Q-Q plot) test is used to test for linearity. Third, the Variance Inflation Factor (VIF) test is used to test for independency and multicollinearity. Fourth, the Breusch-Pagan/Cook-Weisberg and White’s general tests are used to check for heteroscedasticity.

The results of Ordinary Least Squares (OLS) will generate inconsistent and biased, when these assumptions are violated (Gujarati 2003). However, several regression estimators, such as OLS with robust standard error, weighted least square regression (Generalised Least Squares (GLS)), and robust regression provide an alternative to OLS regression when the assumptions have been violated (Judge et al. 1985). In the presence of heteroscedasticity, and autocorrelation, either the least square estimator with robust standard error (Huber-White standard errors) or GLS

97

regressions are able to reweight the error variance and thus to correct heteroscedasticity and autocorrelation (Gujarati 2003).

In general, the present study finds that most of the OLS assumptions are not adequately fulfilled, even though several steps (e.g. data transformation) have been taken to conform to these assumptions. In this regard, Glass et al. (1972) indicate that mild violations of the OLS assumptions are robust and unaffected in many situations. Thus, pooled OLS regression is performed in the main analyses while additional tests using least square estimator with robust standard is used in the sensitivity analysis as an alternative estimator.

4.7.3 Further Analyses and Robustness Checks

The purpose of further and sensitive analyses is to ensure that the main results are robust to various measurements and estimators. While the aim of further analyses is to control for the potential effect of the direction of discretionary accruals and the type of industry in which a firm operates on the main findings, the robustness checks aim to ensure that the main results are robust to various measurements and estimators as well as to control for endogeneity.

98

4.8

Summary

This chapter provides the justifications for the study methods in accordance with the study objectives and research questions. In an attempt to provide evidence on the impact of CSR on EM in UK companies, the process of data analysis is performed using the pooled OLS regression models to examine the study hypotheses. The 515 (FTSE 350 Index) company-year observations are drawn for the period 2008-2010.

The next two chapters, chapter five and chapter six, will highlight and analyse EM and CSR practices in UK companies respectively. Chapter seven will then analyse and discuss the findings on the association between CSR and EM in UK companies.

99

5

Chapter Five