To be able to answer the research question for this thesis Does ESG score affect the financial performance of Nordic companies? a suitable model must be chosen. In this chapter, we will start by explaining the data set and then go through the model selection process, before ultimately discussing some concerns about the validity of the results. The procedure for choosing the right model will be repeated for all three segments. The results from the tests are presented in the results (7.1, Table 4).
5.1.1 Panel Data
The data set at disposal contains observations for 14 years and 139 companies and is structured as a panel data set. The panel data set is unbalanced due to the lack of data. The advantage of using a panel data set is that we can control for unobservable variables across firms and years given that we model it accurately (Stock & Watson, 2015). In this thesis, such variables could be increased focus on ESG over time, the importance of ESG in different companies or in different industries. It accounts for individual heterogeneity. A panel data set is rich on information and therefore allows us to investigate more complex problems than with pure cross-sectional or pure time-series data. It would require a long time-series to investigate how the variables move dynamically in a pure time-series model, which would induce a problem since ESG rating is updated yearly and goes back to 2002. With panel data, the number of observations will be higher and thereby increase the power of the test (Brooks, 2014).
5.1.2 Model Building
Due to the fact that we have a panel data set and wish to take advantage of its features, a model for panel data will be chosen. There are different types of models to be applied on panel data, where Fixed Effect Model, Random Effect Model and Pooled OLS are the three most common. Velte (2017) and Shih-Fang Lo and Her-Jiun Sheu (2007) investigate the same topic as in this thesis and are both applying the Fixed Effect model. Still, our choice of model will be based on a Poolability Test, a Breusch-Pagan Multiplier Test and a Hausman Test, which will determine which of the three models is the most suitable for our panel data set.
Another approach that could be used instead of the panel data methods chosen in this thesis is a portfolio analysis. This methodology is applied by Guenster et al.
(2011) who use the Fama & French (1993) methodology. A suggestion for further research is to use both the portfolio analysis and the panel data models to secure the robustness of the results.
5.1.2.1 Pooled model
A Pooled OLS would not take advantage of the benefits of the panel data set (Hill, Griffiths, & Lim, 2012). In a Pooled OLS the dependent variable is pooled together, both cross-sectional and time-series observations. The explanatory variables are stacked the same way. The Pooled OLS will be estimated using simple OLS. This method of handling a panel dataset is easy, and assumes that the average values of the variables and the relationship between them are constant across all entities (cross-sectionally) and over time (Brooks, 2014).
Put differently, the Pooled OLS use simple betas, meaning that they do not take into account the cross-sectional nor time-sectional characteristics. The Pooled OLS will be chosen if the data does not contain fixed effects or random effects. The regression equation for segment 1 with a Pooled OLS is:
π ππ΄;+ = π½-+ π½?πΈππΊ;++ π½@π &π·;++ π½Cπ΅ππ‘π;++ π½Eπ·πππ‘π ππ‘ππ;+ + π½Gπππ§π;++ π’;+
π€βπππ π = 1, β¦ ,139 πππ π‘ =2006, β¦ ,2018
5.1.2.2 Fixed effects
The Fixed Effect Model controls for unobserved heterogeneity that is constant in the time dimension. It assumes that there are omitted variables in the panel data that varies across entities, but not across time (Stock & Watson, 2015). To control for the variation across firms, the model has one intercept for each firm (πΌ;). The intercepts absorb the omitted effect that is constant over time, but the variation across time is still not accounted for. There is a difference within each firm that is not accounted for by the control variables but is captured by the intercepts for each entity. For segment 1 the regression equation with fixed effects is:
To test whether there are fixed effects in our panel data set, we use the F-Test for fixed effects, also referred to as the Poolability Test (Kunst, 2009). The null hypothesis states that individual effects do not exist, while the alternative states that there are individual effects. If the null hypothesis is rejected, a Pooled OLS cannot be used, and a Fixed Effect Model is preferred over Pooled model (Kunst, 2009).
Kunst (2009) states that it is necessary to check for random effects before deciding if the Fixed Effect Model is the right choice.
5.1.2.3 Random Effects Model
The Random Effect Model takes the individual effects into account and uses one intercept per entity (π;Y). The difference between a fixed and random effect model is that the random effect model assumes that the entities are randomly selected and that the individual effect is not fixed, but random (Hill et al., 2012). The Random Effect Model assumes that the random effects arise from a common intercept that is the same for all units over time, plus a random effect that is constant over time and measures the random deviation from the global intercept for each entity (π;) (Brooks, 2014). The regression equation for segment 1 with a random effect model is shown below.
π ππ΄;+ = π½-πΈππΊ;++ π½?π &π·;++ π½@π΅ππ‘π;++ π½Cπ·πππ‘π ππ‘ππ;+ + π½Eπππ§π;++ π’;++ π; + π;Y, π€βπππ π = 1, β¦ ,139 πππ π‘ =2006, β¦ ,2018
Our sample of data is not selected randomly, but still, there is a need to check whether there are random effects in the data. The Breusch-Pagan Lagrange Multiplier Test will test for random effects in the data set, thereby determine the need to check if a Random Effect Model is the best course of action. The null hypothesis is that individual-specific or time-specific error variance is zero, meaning that a Pooled OLS is preferred over a Random Effect Model (Park, 2011).
If there are both random and mixed effects in the data a Hausmann Test is suitable to determine which effect is the strongest, hence which model to choose. The Hausman Test will make us able to determine which is the best choice between the Fixed Effect Model and the Random Effect Model. The test examines whether the individual effects are uncorrelated with other regressors in the model. If the
individual effects are correlated the random effect model will violate a Gauss-Markov assumption and is therefore no longer Best Linear Unbiased Estimate (BLUE), this is because the individualβs effects are part of the error term in a Random Effect Model (Park, 2011). If the null hypothesis is rejected, the Fixed Effect Model is favoured.