The effect of the Norwegian gender quota on firm performance : the long run effects of the Norwegian gender quota for corporate board of directors on firm performance

(1)

1

The effect of the Norwegian gender quota on firm performance

The long run effects of the Norwegian gender quota for corporate board of directors on firm performance

The Norwegian gender quota of 2006 mandated companies to have at least 40% of women in their board of directors. Previous work focused mostly on short term effects on firm value and firm performance. In this thesis, I investigate the long term effects on firm performance measures by Return on Assets. I use a panel dataset of Norwegian ASA and AS firms from 2002 to 2015. Using a difference-in-differences regression method, the effects on firm performance are estimated for each year. Negative effects are found in the short run. Positive effects are found in the long run. These positive long term effects implicate governments should consider implementing gender quotas, as it increases the firm performance.

Master thesis

Name: Tzachi Sigi (V.S.) Fischer

Student number: 10373683

MSc- programme: Finance

Specialization: Corporate Finance

Thesis supervisor: Dhr. Simas Kučinskas Submission date: July 1st_{, 2018, Amsterdam}

(2)

2

Statement of Originality

This document is written by Tzachi Fischer who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document is original and that no other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

3

Acknowledgements

First of all, I would like to thank my supervisor, Simas Kučinskas, for his efforts regarding supervising my thesis. His comments were always very useful and helped me improving this thesis a lot. Furthermore, I would like to thank David and Rowena Reynolds from Australia, for discussing with them the topic beforehand, and later on for proofreading the whole paper. Thank you.

(4)

4

1. Introduction

In 2002, an interview with the Minister of Trade and Industry, Ansgar Gabrielsen was posted in the Verdens Gang, the biggest newspaper of Norway. There, he announced legislation to increase the percentage of women on boards to 40%. Apparently, he did this without consultation with other members of the government (Haugnes, 2008). It caused an uproar in the government. During the following government meeting, his action was criticized massively. Despite all the uproar, Gabrielsen was “convinced this diversity in boards would increase value creation” (Haugnes, 2008).

But was Gabrielsen right? Did the women quota indeed increase value creation and firm performance? A number of studies have been done on this specific Norwegian quota. Most research was executed around 2012, so only the short-term effects could be investigated. The results found were either negligible (Dale-Olsen, Schøne and Verner, 2013), or negative (Matsa and Miller, 2013; Ahern and Dittmar, 2012). The papers finding negative results specifically advice to investigate the long term effects. Until today, these long term effects have been ignored. In this thesis, the long term effects of this Norwegian women quota on firm performance is researched. For further implementation of quotas around the world, the short-term effects are important. But even more important might be the long-term effects. During the first years after the introduction, most companies have to intensify their search of female directors. Firms that might not comply with the quota in time, might hire unsuitable women. The effects observed after a few years are therefore not representative for the true effects for society.

The long term effects on firm performance are especially interesting, as more countries have introduced mandatory quotas since. Israel was the first country to implement mandatory female quotas for government owned companies (Izraeli, 2003). Norway followed in 2006. Since then, Italy, Spain, Germany, France, Belgium and more followed (Comi, Grasseni, Origo and Pagani, 2017). With quotas implemented in multiple countries, the question whether it enhances firm performance in the long run is even more important. Furthermore, these mandatory quotas made empirical research more reliable. Until the late ‘00s, research has been done of the effects of greater gender diversity, but endogeneity was always an issue. Did companies with high performance had many women because they could afford to? Or did companies have high performance because of these women? Due to the natural experiment

(6)

6 nature of the quota, difference-in-differences can be used. Therefore, it is assumed the effects of more women in boards can be investigated more reliable than previous research could.

One of the main arguments used in favor of gender-diverse boards is the assumption that a mix of females and males results in a mix of different backgrounds, opinions and creative ideas. The more diverse ‘input’ can be given in a board meeting, the higher the chance of choosing the best solution. A counterargument relates to this ‘diversity in opinions’. When a group has a lot of diversity, these disagreements can lead to endless discussions and arguments, which makes the board ineffective. It is a priori unclear whether the input of a lot of diverse opinions, or the homogeneous agreeable board prevails.

In this thesis, I use panel data of 2002 to 2015 of all ASA and AS companies in Norway. Using a matching algorithm and difference-in-differences (DID), I test the effect of the quota for each post year. For the short run, I find negative results on Return on Assets (ROA). For the long run, I find positive results for the years 2013 and 2014. The quota appears to lead to an increase in ROA of 7 percentage points. The short term results are in line with Matsa and Miller (2013) and Ahern and Dittmar (2012). The long term results are the added value of this thesis and therefore cannot be compared with current literature. A noticeable difference with current literature is that my sample includes all treated companies, whereas Matsa and Miller (2013) only include the listed treated firms. The most important addition is that I use an extended time period of 14 years, due to which I can investigate the long term effects.

In this thesis, I start with the literature review. Here, the details of the quota are explained and the empirical evidence between men and women is discussed. Also, current evidence on Norway’s quota is analyzed. Next, the methodology and data is described. The results and robustness checks follow. Finally, in the conclusion the main limitations and findings are discussed.

(7)

7

2. Literature review

In this literature review, first the specific details of the Norwegian women quota are given. Secondly, multiple arguments are given to support the assumption that a change in a company´s board of directors may affect firm performance. Next, a literature review of differences between males and females in corporate and non-corporate environments is given. Then, the current evidence of quotas in Norway and other countries is described. Finally, the added value of this thesis is explained.

2.1 Norway´s women quota

The public limited companies (PLCs) are in Norway registered as Allmennaksjeselskap (ASA), while the private limited companies (LTDs) are registered as Aksjeselskap (AS). In both types of legal forms, the board members are not personally liable for the commitments of the company. The shareholders are only liable for their share of the capital.

A limited company has to have a board of directors with at least one member. The minimum share capital is 30,000 Norwegian Kroner (NOK) (approximately EUR 3,000). For a public limited company, the requirements are stricter. The minimum required share capital increases to NOK 1 million (EUR 104 thousand), and the board of directors must consist of at least three members (“Running a private limited company”, 2017). Being registered as an ASA is a legal requirement in order to be listed on the Oslo Stock Exchange. All ASA companies together account for 73% of all private employment in Norway (Dale-Olsen et al., 2013). Therefore, the quota affected a big part of the Norwegian economy.

In September 2002, 6% of all board of directors seats of ASAs were assigned to women (Nergaard, 2003). In June 2003, the Norwegian government proposed a legislation that would require Norwegian public limited firms to have at least 40% of both sexes represented on corporate boards of directors. The proposed quota of 2003 applied on all state-owned companies, inter-municipal enterprises and public limited companies (PLC). This corresponds with 438 companies at that time. The AS companies were excluded from the quota because their shares are typically less dispersed: the majority of shares are owned by the same family. In many cases, the shareholders are at the same time members of the board (Reiersen and Sjåfjell, 2008).

(8)

8 The 2003 proposal was implemented on a voluntary basis only. After two years, the government would evaluate the rate of implementation. As the affected companies did not increase their share of women enough by July 2005, the government decided to implement a mandatory quota. In January 2006, the government completed the implementation of the mandatory quota. The non-compliers as of January 2006 were given two years to comply with the quota. Companies that had not met the quota by January 2008 were forced to cease to exist. In January 2008, a final warning was given to 77 firms. By April of that year, all firms complied with the quota (Ahern and Dittmar, 2012). In 2005, the average board of directors of a PLC consisted for 16% of women, while in 2008 this percentage was 41% (Nygaard, 2011). In order to comply with the quota, ASA-companies replaced about one third of their male directors with females. The total number of female directors increased from 165 in 2003 to 592 in 2008, while the number of male directors decreased from 1,516 in 2003 to 938 in 2008 (Bøhren and Staubo, 2014).

2.2 Theoretical framework

The main theories used to explain the evidence of performance effects of a gender quota are the agency theory and the resource dependency theory (Dalton, Hitt, Certo and Dalton, 2007; Zahra and Pearce, 1989). In addition, the efficient market hypothesis can be used.

Agency theory describes the Principal-Agent situation where the owners of the firm, the shareholders or the ‘Principal’, do not act at the same time as the manager of the firm. The CEO or the ‘Agent’ acts as the manager. However, his interests might not be aligned with those of the Principal. This is the Principal-Agent problem. The costs of not aligned interests are called agency costs.

The board of directors is an internal governance mechanism that is intended to ensure that the interests of shareholders and managers are closely aligned. Their main functions are monitoring and advising the management. Especially the monitoring is essential (Miller, 2002; Fama, 1980) and reduces the agency problems between the two parties (Carter, Simkins and Simpson, 2003). By monitoring, the directors can influence the management’ actions and in this way, the firm performance. But if the board of directors is not functioning or exercising their task of monitoring well enough, managers can make decisions or take actions which entrench themselves in the company in such way that it is costly for the firm to replace them

(9)

9 (Shleifer and Vishny, 1989). This is an example of agency costs. One of the reasons for introducing the board of directors was to reduce agency costs. Strong boards can thus prevent managers from entrenching themselves and thereby prevent the firm from losing value or deteriorate performance. Coles, Daniel and Naveen (2008) and Linck, Netter and Yang (2008) found that companies compose their boards based on firm specific characteristics, in such a way that they function optimally and optimize the costs and benefits of the board’s monitoring role (Nygaard, 2011).

Resource dependency theory (RDT) was developed by Pfeffer and Salancik in 1978. It

argues that the rate of success and constraints of a firm depends on the environmental organization/climate in which it operates. Before the quota, the average board of ASA companies consisted continuously of 5% women (Huse, 2016). This could mean that during the search for new board members, the nomination committee mainly proposed male directors1_{. After the quota, their scope widened and the search was no longer restricted to} predominantly men (Reiersen and Sjåfjell, 2008). The environmental conditions the firm is limited by, is largely determined by the board of directors (Hillman, Withers and Collins, 2009). It can be argued that this quota therefore enhanced the environmental climate of all treated firms in Norway. Sanders and Carpenter (1998) find that the board size is related to the environmental dependence. The more capable directors in the market, the better resources firms can have and thus the better firms perform. By enforcing the quota, the number of directors increases and thus firm performance can increase as a result.

According to the Classical Finance Theory, economic agents are fully rational in preferences and beliefs. This rationality makes the markets efficient. This is called the Efficient Market Hypothesis (EMH). Shiller (2003) concludes there has been a shift from generally assuming the EMH holds, towards a Behavioral Finance theory, which assumes irrationality among people. Irrational choices can be explained by this Behavioral Finance theory. If the EMH holds, the current board is the result of optimal choices by rational people. The current board should then be the optimal board. Forced changes, by for example a quota, should then

1_{To my knowledge, the fraction of females on the nomination proxy in Norway before 2006 has not been}

investigated. I have to assume the fraction of females was small, and therefore the number of female directors that could be appointed was around 5%. Nygaard (2011) mentions that the CEO can have great influence in the nomination process. Often, females were not part of the informal network the CEO picked his new candidates from. Therefore, according Nygaard (2011), it is likely that the proxy often displayed more men than women.

(10)

10 be value-reducing. As boards are chosen to optimize firm value, they are also chosen to optimize firm performance. Forced changes leading to value-reduction, leads therefore also to a decrease in firm performance. In other words, negative effects on firm performance are in support of the EMH, while positive effects of the quota are in support of the behavioral finance theory and invalidate the EMH.

2.3 Empirical evidence on difference between men and women

Are there any characteristics overrepresented among women? Characteristics assigned to females from experiments or questionnaires? And are these characteristics externally valid to assume females in an organization such as board of directors, will behave according to these characteristics? Or do females in groups behave differently than females individually?

Findings are that women are more risk averse than men in gambling settings (Eckel and Grossman, 2002; Schubert, Brown, Gysler and Brachinger, 1999) and also invest less than men (Charness and Gneezy, 2012). Wilson and Altanlar (2009) conclude that having female directors on the board reduces insolvency risk. Regarding CEOs, it is found that female CEOs tend to take on less leverage (Faccio, Marchica and Mura, 2016; Palvia, Vahamaa and Vahamaa, 2015). This could be related to the finding that males are often more overconfident than females (Huang and Kisgen, 2013).

According to Nielsen and Huse (2010), female directors are less hierarchical and more cooperative in organizations. Therefore more women on the board leads to lower levels of conflicts.

More specifically related to the behavior of women in their role as member of the board, a difference found between men and women is described by Adams and Ferreira (2009). They write that gender-diverse board are tougher monitors because their attendance rates to board meetings are higher than those of the male directors. If the agency theory holds, tougher monitors in the board lead to a better functioning board of directors, as their main function is the monitoring of the executive management. This should at its turn lead to better firm performance. However, Adams and Ferreira add in the same paper to this that at well-governed firms, tougher monitoring can harm the firm performance, due to over-monitoring. The optimal board is therefore company specific.

(11)

11 According to Van den Berghe and Levrau (2004), the most important element of a good board of directors is the quality of the board meetings. Females not only have a higher attendance rate to board meetings, they are also found to prepare better for board meetings than males (Huse and Solberg, 2006), and ask more questions (Terjesen, Sealy and Singh 2009). From this, it follows that more women in the board results in higher quality board meetings. However, it seems that women only become active when there are three women or more in the board (Konrad, Kramer and Erkut, 2008). This is also referred to as the critical mass. So, do women only deliver their valuable characteristics if they are not the only woman in the board? Rose (2007) argues that “board members with an unconventional background are socialised unconsciously adopting the ideas of the majority of conventional board members” (p. 404). In Norway, the average board size in 2017 was 8.4 members (Spencer Stuart, 2015). Combining this with the 40% quota, this corresponds with 3.36 women in each board of directors. As this is an average, some boards do not reach the critical mass of three women. It is therefore uncertain whether all positive characteristics of females come to expression in boards.

Smith, Smith and Verner (2006) find a positive link between female directors and firm performance. According to them, the moderator for the positive effect is the education level of the female directors. A key condition for a quota to be performance-enhancing is therefore the qualifications of the new female directors. Using data from Norway after the quota, Ahern and Dittmar (2012) found that new female directors had less CEO experience and were 8 years younger than the current male directors. Singh, Terjesen and Vinnicombe (2007) also show that newly appointed females have less experience as a member of board of directors of big companies. A possible explanation for this can be the shortage of experienced women in Norway. The percentage of females in ASA boards of directors fluctuated around 5% from 1990 till 2002. Therefore, the women were not able to gain experience as directors (Huse, 2016) This shortage of experienced women can lead to negative results on firm performance in the short term. As described later, multiple studies find negative results of the quota. In the long run, a couple years after the quota, these female board members have caught up with their lack of experience. If the lack of experience is a determinant of firm performance, then in the long run, the effect of the quota might be positive.

(12)

12 Studies have also been done which can imply employment consequences of appointing more females to the board. Women are better capable of delaying gratification (Bjorklund and Kipp, 1996), and therefore have a longer time horizon of reaping the rewards of investing (Silverman, 2003). Frederick (2005) and Castillo, Ferraro, Jordan and Petrie (2011) even found that smart women are more patient. In terms of employment numbers within the firms, this could mean that females are more reluctant to insist on employment reduction measures to improve firm earnings in times of bad performance. Matsa and Miller (2013) did find that affected firms (with more females in the board) undertook fewer workforce reductions. More women in the board can therefore lead to an excessive number of employees in the firm. This could lead to negative firm performance in the short term, as the cost of labor is too high. However, in the long run this could lead to better performance. The capital invested in HR remains in the firm. So in the long run the costs of employees is lower than that of companies which significantly reduced their workforce.

2.4 Empirical evidence on firm performance in other countries

Multiple studies have investigated the effect of female representation in boards on firm performance, either Tobin’s Q, Return on Assets (ROA) or announcement returns.

Campbell and Minquez-Vera (2009) found a positive cumulative abnormal return (CAR) after the announcement of new female board members. A positive CAR implies investors regard the appointment of females as enhancing for firm value. They researched appointments by quoted Spanish firms between 1989 to 2001. A Gender Equality Act by the Spanish government was implemented in 2007. According to this Act, boards had to ‘comply or explain’ to the principle that boards with few female directors should take initiatives to improve female representation. This Act was the reason to investigate possible effects on firm value.

Perryman, Fernando and Tripathy (2016) did research on female representation in top management teams worldwide. Their results indicated that firms with higher gender diversity in top management teams had higher a Tobin’s Q. It was also concluded that high gender diverse firms displayed less risk than firms with low gender diversity. Gender diversity is measured by the proportion of female executives in the firm. Risk was measured by both the Beta and the standard deviation of daily returns. For both measures, negative coefficients

(13)

13 were found. Control variables used are size of the board, age of the executives and dividend. Other control variables as mentioned in 3.4 Control variables are not used in the paper of Perryman et al. (2015).

Erhardt, Werbel and Shrader (2003) performed correlation and regression analysis on 112 large public US companies. They concluded that having more women on the board is associated with higher ROA. Because they only conclude there is an association, the lack of using Instrumental Variable (IV) regression is not a drawback. Conyon and He (2017) used data of 3,000 US firms and concluded women on boards improve firm performance. A drawback of their methodology is the lack of a natural experiment. To still account for endogeneity, they used IV endogenous quantile regressions. As instrument they use the percentage of women living in the state of incorporation. They show this instrument is correlated with the percentage of women on the board, so it satisfies the relevance criteria. But the exclusion criteria cannot be tested empirically. They only argue that their instrument is not correlated with the organizational performance. But if it is correlated with the percentage of women on the board, why would it not be correlated with the organizational performance? This is a weakness of their methodology.

Using data from France, Sabatier (2015) found that more females on the board is linked to higher firm performance and reduced corporate inefficiencies. She used IV regressions with variation in the pre-reform fraction of women on board as instrument. By performing a Hansen-Sargan overidentification test, she empirically proved the satisfaction of the exclusion restriction.

But not all results are clear cut and positive in terms of firm performance. Researchers described above, did not consider the possibility that the effect of more females can be company specific. This can be the reason why the empirical results on the added benefits of gender diversity are not clear cut. Abdullah, Ismail and Nachum (2016) found inconsistent results using Malaysian data and argued that the firms’ ownership characteristics are the main moderator for improved firm performance. This is in line with the conclusion of Dwyer, Richard and Chadwick (2003), who conclude that a supportive organizational environment is needed for positive results. The results could even be different for each country. From Jurkus, Park and Woodard (2011) it follows that results on agency costs are dependent on the competitiveness of the country. Ruigrok, Peck and Tacheva (2007) mention that the effects of

(14)

14 gender diversity on firm performance are not always externally valid/transmittable to another country, because the country-specific corporate governance systems are important. Therefore, a geographic focus on Norway specifically is needed.

2.5 Empirical evidence on firm performance in Norway

Because Norway was one of the first countries to implement gender quotas on such a large scale, multiple researchers investigated the effect on firm performance. The added value of such quota was the natural experiment nature of it. Researchers did now not need to account anymore for endogeneity by using, for example, IV instruments. Studies that researched the short-term effects found a large decline in Tobin’s Q and a significant drop in the share price of affected companies around the announcement of the law (Ahern and Dittmar, 2012). A significant reduction in profits was found by Matsa and Miller (2013). Dale-Olsen et al. (2013) concluded that the effect on the ROA was negligible. However, Nygaard (2011) concluded that for firms with low information asymmetry, the CAR after the quota was positive.

Matsa and Miller (2013) analyzed a panel dataset from 2003 to 2009. Among the affected Norwegian firms, they only kept the 159 listed non-finance-ASA companies (and dropped the non-listed ASA companies). Using inverse variance weighting, a matched control set of firms was created. They matched total assets, operating profits, labor costs and industry. The ROA (operating profits/assets) of these listed ASA firms were compared with matched samples of AS companies. In addition, the listed ASA firms were compared to public and private firms in other Scandinavian companies. For all sets of control groups, they found a significant drop in ROA for the treated firms. They further found that the affected firms, which had more females on the board due to the quota, undertook fewer workforce reductions than the control group. This might be in accordance with the findings of Silverman (2003) who concluded that females are able to delay the rewards of an investment longer than males. Because Silverman concluded this on the basis of an experiment, she notes that the outcome of this experiment may not be applicable to the real world. The finding that affected firms undertook fewer workforce reductions is the basis of the hypothesis that the negative effect on firm performance might be caused by the long-term focus of the new female directors. This negative effect on performance is proven by Matsa and Miller (2013). A drawback in their study is the lack of many control variables. Only the board size and the average number of

(15)

15 other board seats are used. They lack the use of other potential relevant control variables as described in section 3.4 Control variables.

Ahern and Dittmar (2012) used a panel dataset of 248 listed ASA companies from 2001 to 2009. They ran placebo effects using data from Danish, Finnish, Swedish and US companies. Using instrumental variables, they found that a 10% increase in female board membership leads to a drop of 0.19 in Tobin’s Q. Their instrument is the pre-quota variation in women on board, which is also used by Sabatier (2005). In this paper, Sabatier proved the satisfaction of the exclusion restriction. But why use IV regression with an instrument in the first place, if we have a natural experiment? Because the level of constraints put by the quota is dependent on the percentage of women on the board before the quota. The smaller this percentage was, the bigger the constraint. A drawback of their methodology is that they only used the control variables of board size and firm size.

Dale-Olsen et al. (2013) execute a difference-in-difference method on panel data from 2003 to 2007 of non-finance ASAs and AS companies. They argue that the years 2008 and 2009 are not suitable comparison years, due to the financial crisis. The data range is limited to 2003 to 2007. They concluded that the short-term effect of the quota on firm performance is negligible.

2.6 My research related to existing research

The empirical papers summarized above, investigated the short-term effects and advise specifically to investigate the long-term effects.

Dale-Olsen et al. (2013) wrote: “Since the reform was implemented quite recently, we are not able to investigate the long-run impact of the reform. (…) In the long run, the effects might be different. (…) Future research should look at potential long-term effects of the reform.” (p. 129). Matsa and Miller (2013) agree: “The long-term effects of greater gender diversity in corporate leadership present an important area for future research.” (p. 166).

Unfortunately, the financial crisis arose in the middle of my sample period 2002 to 2015. It is likely that this financial crisis is a shock to the performance of companies in Norway. This shock is unrelated to the quota. In other words, a drop in returns starting in 2008 should be attributed to the financial crisis rather than to the quota. Therefore, the results on the

(16)

16 short-term are not investigated intensively in this thesis. Rather, the results from 2012 and onwards are focused on.

Due to the findings that females are reluctant to fire employees, are more long-term oriented and that new female board members have less experience, the short-term effect of the quota is expected to be negative, in accordance with the EMH. This is indeed confirmed empirically (Dale-Olsen et al., 2013; Matsa and Miller, 2013; Ahern and Dittmar, 2012). To confirm the empirical negative results of previous literature, the short-term effects are tested. For this, the full sample period of 2002 to 2015 is used, but only the coefficient for 2007 is of importance.

𝐻𝑦𝑝𝑜𝑡ℎ𝑒𝑠𝑖𝑠 1: 𝐻0 = 𝑇ℎ𝑒 𝑠ℎ𝑜𝑟𝑡 − 𝑟𝑢𝑛 𝑒𝑓𝑓𝑒𝑐𝑡 𝑜𝑓 𝑡ℎ𝑒 𝑞𝑢𝑜𝑡𝑎 𝑖𝑠 𝑛𝑒𝑢𝑡𝑟𝑎𝑙 𝐻𝑦𝑝𝑜𝑡ℎ𝑒𝑠𝑖𝑠 1: 𝐻₁= 𝑇ℎ𝑒 𝑠ℎ𝑜𝑟𝑡 − 𝑟𝑢𝑛 𝑒𝑓𝑓𝑒𝑐𝑡 𝑜𝑓 𝑡ℎ𝑒 𝑞𝑢𝑜𝑡𝑎 𝑖𝑠 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒

However, after a couple of years, these new directors have gained experience and therefore have caught up with their delay. One argument for a drop in performance is therefore assumed to be cancelled. Also, females are more long-term oriented than males, prepare better and have higher attendance rates for board meetings and cause less conflicts in meetings.In the long-run a positive effect of the quota on firm performance is expected. The hypothesis is therefore:

𝐻𝑦𝑝𝑜𝑡ℎ𝑒𝑠𝑖𝑠 2: 𝐻₀ = 𝑇ℎ𝑒 𝑙𝑜𝑛𝑔 − 𝑟𝑢𝑛 𝑒𝑓𝑓𝑒𝑐𝑡 𝑜𝑓 𝑡ℎ𝑒 𝑞𝑢𝑜𝑡𝑎 𝑖𝑠 𝑛𝑒𝑢𝑡𝑟𝑎𝑙 𝐻𝑦𝑝𝑜𝑡ℎ𝑒𝑠𝑖𝑠 2: 𝐻1 = 𝑇ℎ𝑒 𝑙𝑜𝑛𝑔 − 𝑟𝑢𝑛 𝑒𝑓𝑓𝑒𝑐𝑡 𝑜𝑓 𝑡ℎ𝑒 𝑞𝑢𝑜𝑡𝑎 𝑖𝑠 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒

(17)

17

3. Methodology

In this methodology, first, the type of data is explained. Second, the econometric model used to estimate the effects of the quota is discussed. Third, explained is why the DID method shows a causal effect. Then, the control variables are clarified. Finally, the interpretation of the signs of the coefficients is explained.

3.1 Type of data

The treated group, the ASA firms in Norway, are compared with a control group, the AS firms in Norway. To answer the research question of this thesis, a proxy for firm performance has to be used. Firm performance can be measured by multiple accounting measures such as ROA, Return on Equity (ROE), Return on Capital (ROC), or profitability ratios such as net profit margin or asset turnover. Another used proxy for firm perfomance is Tobin’s Q. Tobin’s Q divides the market valuation of the company by its book value of the assets. An advantage of this proxy is the inclusion of the market based valuation of the company. Although this can be considered less objective due to the fact that investors’ valuation can be biased, following the Efficient Market Hypothesis, it can be argued that this market based valuation is a fair valuation. Ahern and Dittmar (2012) and Comi et al. (2017) used Tobin’s Q in their research, as they compared the treated and listed ASA firms to comparable listed firms in foreign countries. The dataset used in this thesis contains both listed and non listed firms. To compare listed (and non-listed) ASA firms with the (non-listed) AS control group, my research is limited to the usage of accounting measure based proxies only. The regression analyzes are therefore done using the ROA, where the Operating Income is used as the earnings measure. Regressions with dependent variables ROA using P/L (Profit or Loss) before tax, P/L after tax and Net income are done as robustness checks in the 6.4 Dependent variables section.

Ideally, for each firm of both the treated and control group, the percentage of women directors and size of the board has to be known. Due to no access to the Norwegian Statistics Bureau (SSB), this data cannot be accessed. The assumption has to be made that in the treated group, all companies experience a marked increase in their female directors at the time of the treatment, while in the control group, this increase does not take place. See the

(18)

18 4. Data section for more details and empirical evidence regarding this.

3.2 Econometric model

The sample period is 2002 to 2015. The same firms are observed for multiple years. Since the data is a panel dataset with longitudinal observations for the same firm, I can to control for firm level fixed effects and year fixed effects. By including firm fixed effects, I eliminate any confounding arising from effects that are constant over time within each firm. By including year fixed effects, I eliminate the effects caused by constant changes across all firms within each year. The quota implemented in 2006 is a natural experiment. Using a difference-in-differences model, the relationship between the quota and the firm performance is tested. A DID model tests the impact of an event by comparing the pre- to post event change in outcome for the treated relative to a control group. For a difference-in-differences, the parallel trend assumption has to be met. Although this cannot be proved, there must be enough evidence to reject the hypothesis that two type of firms follow a different trend before the quota. The following regression is tested:

𝑌_𝑖𝑗𝑡 = 𝛽₁∗ 𝑡𝑟𝑒𝑎𝑡𝑒𝑑 ∗ 𝑦𝑒𝑎𝑟_𝑡+ 𝛽₂∗ 𝑦𝑒𝑎𝑟_𝑡+ 𝜓_𝑖 ∗ 𝑦𝑒𝑎𝑟_𝑡∗ 𝑖𝑛𝑑𝑢𝑠𝑡𝑟𝑦_𝑗 + 𝛼_𝑗 + 𝜔_𝑡

Where the i is the industry, j is the company and t is the year. Y is the dependent variable: ROA using operating income as measure of earnings. 𝛹 allows for different linear time trends by industry group. This is followed from the method used by Matsa and Miller (2013). The 𝛼_𝑗 allows for the company fixed effects and the 𝜔𝑡 allows for the year fixed effects. Standard errors are adjusted for clustering at the firm level.

3.3 Causal effect

The differences-in-differences regression is good for testing the causal effect of a natural experiment only if clearly defined treated and control groups are present. Most probably, only the first years after the treatment is the causuality implied by a DID the strongest. For these first years after 2006, the DID gives strong results with high internal validity. However, it remains unclear how strong the DID predictor is for for example years 2014 or 2015.

During the sample period, no other major events have happened that could have affect the ROA, aside from the financial crisis starting in 2008 and the quota itself. Therefore, the results of the DID can be assumed to show causality. For the first couple of years, this causality

(19)

19 is the strongest. The weakness of the DID is that the evidence for causality decreases with the sample time period. Is the change in ROA seven years later on the account of the quota still? If no other major events happened in the sample period, then it can be ruled out that the change in performance is not due to the quota.

3.4 Control variables

Important contol variables when estimating the effect on ROA are the employment number, total assets and industry. Furthermore, the regional location, the board size (Van den Berghe and Levrau, 2004), board composition in terms of independent directors, the age diversity in the board, board activity and attendance rates, nationality, and blockholder ownership and family ownership and change in shareholder ownership can be important variables influencing the ROA. In this thesis, data on only the former three are used due to access restrictions. According to Yermack (1996), the board size is inversely related to firm value. According to Nguyen and Nielsen (2010) and Chou, Chung and Yin (2013), independent directors are valuable to a firm, although which percentage of the board should be independent is not investigated. There is empirical evidence that board activity has a positive effect on firm value (Brick and Chidambaran, 2010) and enhances firm performance (Chou, Chung and Yin, 2013). Ruigrok, Peck and Tacheva (2007) argue that the nationality diversity is an important aspect to consider. They write that the nationality of a director influences the added value of a director, because of associated characteristics as for example family contacts or national culture. Andres (2008) concludes that founding-familiy being still represented in the board is associated with enhanced firm performance. Finally, the higher education level of females is among the main reasons for enhanced firm performance (Smith, Smith and Verner, 2006).

Summarized, there are many board characteristics that are influencing either firm value or firm performance. During my sample period, the omitted control variables described above can change. To isolate the effects of the gender quota, I ideally have to control for all these important determinants of firm performance. The lack of inclusion of these control variables can be a reason why studies on gender equality often not find the same results. Probably, many studies do not control for all neccesary control variables.

(20)

20

3.5 Signs of coefficients

The main coefficient of interest is the 𝛽1. This coefficient estimates the change in ROA in year t relative to the years before the quota for treated firms, over the change in ROA for the control firms. If this coefficient is significant and positive, the effect of the quota is positive for that specific year t.

(21)

21

4. Data

In this section, described is how I gathered the data used in this thesis. Then, the data cleansing process is described. Finally, the variables are defined and the summary statistics are shown.

4.1 Data

For this thesis, data from 2002 up and until 2015 of all ASA and AS companies in Norway is used. For each company, the Bureau van Dijk company identifier, country, city, industry qualifier SIC/NACERev1.1/NACERev2/NAICS, accounting standard, legal form, number of employees, cost of employees and various accounting data such as total assets and Operating income is downloaded. The main source is the Bureau van Dijk Orbis (BvD) database, which is compiled of detailed information of more than 200 million, both public and private companies worldwide. Since all non-accounting data is only available for the last year, such as the BvD company identifier, accounting standard and industry classifier, using the online BvD database as of today contains not sufficient data. Next to that, this online database only contains accounting data for the last five years since the last known update. In order to gain access to the yearly information, for more than 16 years back, multiple archived versions of the BvD database are used. The BvD database as of December 2014 was not accessible and therefore the accounting information over 2014 is downloaded through the December 2015 edition. I assume that the non-accounting information as of 2015 also applies for 2014. The same method is applied for the years 2004, 2003 and 2002, using the 2005 database.

4.2 Data selection

Using the raw dataset as described above, the following procedure is followed to obtain the cleansed sample. All companies that did not have data for at least one year before and one year after 2006 are excluded. A company which became inactive in 2009 can still be part of the dataset for the years until 2009. This means that for each year, a different amount of observations can be observed. Observations with missing BvD company identifier, industry, total assets, legal form or year are removed. The remaining observations for these companies are not deleted. Companies with zero employees are removed, as well as all companies in the financial sector. Financial firms are removed because a law affecting only finance sector firms

(22)

22 was implemented in 2007. This caused financial firms to change their legal form from ASA to AS (Bøhren and Staubo, 2014). 221 companies changed their legal form during the sample period. According to Ahern and Dittmar (2012), these companies changed their legal form to avoid the law. For these companies, the costs of delisting and changing to a private limited liability company probably outweight the costs of meeting the quota. As these companies would be affected harder than average, these companies are removed. I acknowledge the potentiel selection bias caused by this decision. In the 6. Robustness checks section, these companies are not excluded. The results are robust to this. Furthermore, all AS-companies with total assets less than 6.7 million NOK2_{during 2005 or 2006 are removed. This is done to} avoid comparing the treated firms to very small (<1 million USD total assets) non-treated firms.

A comparable control group of AS companies is created using the teffects nnmatch program as described by Abadie, Drukker, Herr and Imbens (2004). This method of nnmatch (nearest neighbour match) enables me to compare the treated companies to similar non-treated companies. The five most comparable companies based on ROA_operatingincome, Laborcosts_assets and are estimated. In the next section, these three variables will be defined. For every treated firm, the nnmatch method finds five observations from the control group which have the closest values for the matching variables. The nnmatching method follows matching with replacement. It allows observations from the control group to be used as a match multiple times. This is the reason why there are 475 observations, but only 455unique firms are used in the regression described in the next section. Matching with replacement reduces biases, since it produces matches on higher quality than matching without replacement (Abadie and Imbens, 2002). The outcome of the teffects nnmatch is disregarded. This nnmatch is solely used to compile a matched control group. This matched control group is described in this thesis as the control group.

Robustness checks done with datasets cleansed using different methods are found in the 6. Robustness checksError! Reference source not found. section. The dependent variable used throughout this thesis is the ROA, where the earnings measure is the operating income. In the robustness checks, other dependent variables are used.

(23)

23

4.3 Variables & summary statistics

In this section the used variables for all regressions will be defined. The company ID is the Bureau van Dijk ID number, which is a unique identification number. It consists of the ISO nation code and the identification number used by the local tax authorities.

The accounting standard is the type of accounting practice the companies use and can either be local GAAP or IFRS. The legal form is the legal type of company. This can be an Aksjeselskap (AS), which is a Private limited liability company and the equivalent of an English Ltd. The legal form can also be an Allmennaksjeselskap (ASA), which is a Public limited company. The P/L before tax, P/L after tax, Net income and Operating income are four different measures for earnings. They are all used for generating different Return on Assets (ROA) ratios: ROA_beforetax, ROA_aftertax, ROA_netincome and ROA_operatingincome. The ROA is calculated by dividing the earnings by total assets. To compare the results of this paper with other papers, such as Dale-Olsen et al. (2013) and Matsa and Miller (2013), all four ratios are given and used in the 6. Robustness checks section. The Laborcosts_assets is the Labor costs divided by the total assets.

Table 1 describes the summary statistics for the treated, matched control and non-matched control groups in 2006, as this is the year of the quota. As can be seen in Table 1 below, the ROA ratios have big outliers. These ratios of ASA companies have been checked using their official annual reports. Because the annual reports of AS firms are not freely available, these outliers from the used Bureau van Dijk database cannot be checked with the official numbers. Therefore, the four ROA ratios and the Laborcosts_assets are winsorized on the 1% level, as is the standard within finance research. The debt is the long-term debt.

Until 2005, the NACERev1.1 industry classification code was standard. The standard classification changed in 2005 from NACERev1.1 to NACERev2. Some companies are also classified using the NAICS or SIC code classification. The companies that are only classified using the NACERev2, are converted to a NACERev1.1 classification. The NACERev2 code is dropped and the Rev1.1 is used. Next, the more than 800 different industries are merged into 11 different sections.

A POST and TREATED dummy are created. As mentioned in the literature review, all affected companies had to comply with the quota before January 1, 2008. However, the

(24)

24 highest percentage point increase in the share of women can already be seen in 2007 (see Figure 2). This is also the first year that the share of women in PLCs is higher than for LTDs (Dale-Olsen et al., 2013). Therefore, the year 2007 is regarded as the first year of the treatment. All the observation after 2006 receive a 1 for the POST dummy. All the ASA companies receive a 1 for the TREATED dummy, regardless of the year. The interaction variables postxtreated is defined as POST*TREATED, and is the main variables of interest.

Table 1 – Summary statistics of treated, matched control and all control firms, in 2006.

Variables N Mean Median Min Max SD

ASA - Treated Total assets 108 12,037.980 1,450.975 12.450 315,468.000 41,148.250 Number of employees 82 2,966.512 349 1 41,428.000 7,826.737 ROA_aftertax 108 2.907 6.201 -181.800 45.483 22.819 ROA_beforetax 108 4.399 7.301 -181.800 63.450 24.507 ROA_netincome 108 2.761 6.088 -181.800 45.483 22.929 ROA_operatingincome 108 3.832 6.696 -184.280 62.238 23.933 Laborcosts_assets 95 21.120 16.690 0.270 103.257 19.309 AS - Matched controls Total assets 475 463.751 35.911 6.819 32,969.000 2,182.374 Number of employees 390 39.321 11 1 741.000 90.677 ROA_aftertax 475 6.555 5.290 -68.184 65.180 13.922 ROA_beforetax 475 8.746 7.259 -94.702 80.702 17.408 ROA_netincome 475 6.753 5.358 -68.184 66.452 14.174 ROA_operatingincome 475 8.192 7.124 -95.193 74.193 15.687 Laborcosts_assets 475 21.118 16.696 0.269 103.548 19.218 AS – Non-matched controls Total assets 29,662 135.322 19.633 6.750 132,449.600 1,336.980 Number of employees 17,184 44.459 14 1 22,309.000 293.916 ROA_aftertax 29,662 7.415 4.872 -248.936 65.180 14.529 ROA_beforetax 29,662 9.703 6.382 -249.753 80.702 17.396 ROA_netincome 29,662 7.468 4.897 -248.276 66.452 14.588 ROA_operatingincome 29,662 8.568 6.167 -224.888 77.972 14.878 Laborcosts_assets 21,108 33.714 24.759 0.154 312.435 38.554

Table 1 shows the summary statistics of the cleansed samples: the treated firms, the matched control group and the non-matched controls group. As 2006 is the last year before the quota, this is the year the summary statistics are shown for. Total assets is displayed in NOK millions. Number of employees is in acutal amounts. The four different ROA measures are shown in percentages. The Laborcosts_assets is also shown in percentages. All ratios are winsorized on a 1% level.

(25)

25 In 2006, data is available of 108 firms that are affected by the quota. Of these 108 firms, data on the ROA ratios and the total assets is available for all firms. Only 95 firms have data on costs of labor and 82 have data on the number of employees employed. For the treated group, the mean amount of total assets is NOK 12,038 million (USD 1,784 million). The mean number of employees is 2,967. The mean ROA_operatingincome is 3,832%. The means for the ROAs using other income measures are between 2,761% and 4.399%. The mean Laborcosts_assets is 21,12%. The minimum values of the ROAs of -184% can be considered big, but are correct.

The control group of firms consists of 475 unique firms. The mean ROA_operatingincome for the treated and control group are with 3.8% and 8.2% respectively, not close to each other. However, there is also matched on the Laborcosts_assets and the industry.The figures for Laborcosts_assets figures as more comparable, with means of 21.1% and 21.1% for both groups. This is a sign of good comparability. It is worth noticing the mean and median of the control group and non-matched control group are more closely aligned than the mean and median of the treated group. Probably, this is due to some negative outliers for treated firms. As the control group consists of more firms, these outliers do not cause a low mean for the control group. In Table 2, the percentage of firms categorized in each industry is given. In terms of industry, the control group is highly comparable to the treated group. The most common industry group is Manufacturing and Real estate, renting and business activities. No companies are classified into the Financial intermediation group, because these companies were excluded during the data cleansing.

(26)

26 Table 2 - Distribution of groups amongst Industries

Group

Industry Treated Matched

control group

Non-matched

control group Total

Financial intermediation 0 0 0 0

Public sector and other community activities 0.23 0.80 3.71 3.64

Construction 1.96 2.43 7.03 6.92

Agriculture, hunting, forestry and fishing 2.66 2.84 2.92 2.92

Electricity, gas and water supply 3.21 3.31 1.33 1.37

Wholesale, Retail and Hotels 5.16 9.87 24.53 24.19

Mining and quarrying 14.48 9.01 1.25 1.45

Transport, storage and communication 16.82 12.53 6.16 6.32 Real estate, renting and business activities 20.03 22.31 38.69 38.32

Manufacturing 35.45 36.90 14.38 14.87

Total 100 100 100 100

Table 2 describes the distribution of companies among different industries, as classified by the NACE Rev 1.1 industry standard. The numbers are percentages. As can be seen, the Matched control group is in terms of industries better comparable to the treated group than the non-matched sample group is.

Figure 2 (Dale-Olsen et al., 2013) describes the share of women on boards, for the treated (PLC) and the control (Ltd) group, for each year from 2002 to 2009. Using this graph, the assumption that year 2007 is the first effective year of the treatment can be supported. Also, the assumption that the increase of women on boards in general only happened in boards of treated companies is supported by this graph. The percentage females in the boards for 2005 up and until 2017 is shown in Figure 3 (“Women in the minority, 2017).

(27)

27 Figure 2

“The share of women on boards of directors (PLC and LTD firms). The figure shows the average proportion of women board

members (in percent) for PLCs and LTDs for all firms and excludes the finance sector, based on firms at least surviving from 2003 to 2007.” (Dale-Olsen et al., 2013, p. 120).

(28)

28

5. Results

In Figure 1, the mean of the ROA is plotted against the years, for the treated and the control group. When eyeballing the pre-2006 means, the two groups appear to follow the same trend. Therefore, it is not possible to reject the parallel trend assumption.

Figure 1 describes the mean of the ROA for the treated firms and the firms in the control group. The yearly means are connected using lines. The vertical axis represent the mean ROA in percentages. The vertical line in year 2006 represents the year of the quota.

Multiple Differences-in-Differences regressions are done. The sample includes firm-year observations from 2002 up and until 2015. The standard errors are clustered at the company level in all four regressions. All four regressions include year fixed effects and firm fixed effects. Robust standard errors are reported in parentheses. Table 3 in the Appendix shows the details of all four regressions.

Three regressions (#2, #3, #4) are done with either no control variables or no allowance for different linear time trends per industry. Therefore, the results of these three regressions should be looked at skeptical. Because the controls are included, the number of

(29)

29 observations/IDs is higher in regressions #3 and #4. Not every observation has data on all control variables. In regression #1, the control variables are included and there is allowed for different linear time trends per industry. Table 4 shows the results of a difference-in-differences regression on a Post dummy and TreatedxPost interaction variables. The dependent variables is the ROA using operating income.

Table 4

Dependent variable

(1)

Independent variables ROA

Post -2.551 (-1.342) Treated x Post 2.636 (1.088) Constant 5.375*** (4.411) Observations 3,284 Number of ID 455 Adjusted R-squared 0.001 Firm FE Yes Year FE Yes Controls Yes

Linear time trends in industry Yes Robust t-statistics in

parentheses

*** p<0.01, ** p<0.05, * p<0.1

Table 4 describes the results of a DiD regression with the ROA as dependent variable. Only the Post and interaction variables Treated x Post are used as independent variables. Basic controls as well as firm fixed effects and year fixed effects are included. Also linear time trends per industry is allowed for. The robust t-statistics are in parentheses.

Figure 4 below visually illustrates the results of regression #1. The dependent variable is the ROA, with the operating income as measure of earnings. Coefficient estimates of the DID are displayed using the dots, with the 95% confidence interval displayed by the vertical bands. Each estimates represent the percentage point increase in ROA for treated firms over non-treated firms, relative to the years before the quota.

(30)

30

Figure 4 describes the results of the DiD with ROA as dependent variable. The dots represent the coefficient for that year. This coefficient is the estimated percentage points increase of ROA in that year, for treated firms over non-treated firms, relative to the years before 2007. The vertical bands represent the 95% confidence interval. Clustered standard errors adjusted at company level are used. Basic controls as well as firm fixed effects and year fixed effects are included.

Regression #1 regresses the treatedxpost for each post year on the ROA. Positive and statistically significant on a 5% level effects for treated firms are found for 2013 and 2014. I define years 2013 and 2014, respectively seven and eight years after the quota, as long term. This means that in the long run, the treated firms experienced a bigger increasse in ROA than the control firms, relative to the years before the quota (2002 untill 2006). For year 2014, the coefficient is 8.536. For year 2013, the coefficient of year 2013is 6.994 and has a t-statistic of 2.06. Therefore, the coefficient is statistically significant on a 5% level. The coefficient of 6.994 corresponds with a 6.994 percentage point in increase in ROA. This means that the difference between the ROA in 2013 relative to before the quota is 6.994 percentage points higher for treated firms than for the control group. As the mean amount of total assets for the treated firms is NOK 12,038 million in 2006, a 6.994 percentage point increase in the ROA corresponds with an increase of earnings of NOK 842 million (USD 125 million) in that year for each firm. The GDP in Norway was NOK 1,708 billion in Norway in 2006 (OECD, 2018). Considering there

(31)

31 are 108ASA firms in Norway in 2006 in my sample, this NOK 842 million earnings corresponds with a total increase in earnings of NOK 90.93 billion, or 5.3% of the GDP. The economic effect appears to be significant.

The coefficient for 2007 is statistically significant and negative. The DID estimate for year 2007 has the coefficient of -8.945. This means that the change in ROA_operatingincome over the years before 2007 to 2007 is 8.945 percentage points lower for treated firms than for the control firms. This implies a negative effect on the ROA on the short term. This is in line with the findings of Ahern and Dittmar (2012) and Matsa and Miller (2013).

𝐻𝑦𝑝𝑜𝑡ℎ𝑒𝑠𝑖𝑠 1: 𝐻0 = 𝑇ℎ𝑒 𝑠ℎ𝑜𝑟𝑡 − 𝑟𝑢𝑛 𝑒𝑓𝑓𝑒𝑐𝑡 𝑜𝑓 𝑡ℎ𝑒 𝑞𝑢𝑜𝑡𝑎 𝑖𝑠 𝑛𝑒𝑢𝑡𝑟𝑎𝑙 𝐻𝑦𝑝𝑜𝑡ℎ𝑒𝑠𝑖𝑠 1: 𝐻1= 𝑇ℎ𝑒 𝑠ℎ𝑜𝑟𝑡 − 𝑟𝑢𝑛 𝑒𝑓𝑓𝑒𝑐𝑡 𝑜𝑓 𝑡ℎ𝑒 𝑞𝑢𝑜𝑡𝑎 𝑖𝑠 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒

𝐻𝑦𝑝𝑜𝑡ℎ𝑒𝑠𝑖𝑠 2: 𝐻₀ = 𝑇ℎ𝑒 𝑙𝑜𝑛𝑔 − 𝑟𝑢𝑛 𝑒𝑓𝑓𝑒𝑐𝑡 𝑜𝑓 𝑡ℎ𝑒 𝑞𝑢𝑜𝑡𝑎 𝑖𝑠 𝑛𝑒𝑢𝑡𝑟𝑎𝑙 𝐻𝑦𝑝𝑜𝑡ℎ𝑒𝑠𝑖𝑠 2: 𝐻1= 𝑇ℎ𝑒 𝑙𝑜𝑛𝑔 − 𝑟𝑢𝑛 𝑒𝑓𝑓𝑒𝑐𝑡 𝑜𝑓 𝑡ℎ𝑒 𝑞𝑢𝑜𝑡𝑎 𝑖𝑠 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒

Based on the results as shown above in Figure 4, there is enough evidence to reject the H0 of

the first hypothesis (if and only if the short run effect is defined as the effect in 2007, that is, after one year after the quota). There is also enough evidence to reject the H0 of the second

hypothesis, where the long run effects are defined as seven and eight years after the quota. The research design is a DID model, based on a natural experiment. This natural experiment only affected one clearly defined treatment group. The percentage women in this treatment group rose according to the quota, while the percentage women in the control group remained stable. The results described above can be assigned to be direct effects of the quota, but the question remains how many years after the quota the changes in ROA are still caused by this quota.

For the years after 2008, the internal validity of the research design can be doubted. This is due to the financial crisis which started late 2008. The treated group in this thesis consists of only large, public firms, while the control group consists of non-listed and predominantly smaller firms. It could be argued that the difference in differences of ROA between my treated and control group is due to the fact that the treated companies are bigger (in terms of assets) and are therefore affected differently by the financial crisis than the smaller (in terms of assets) control group of firms. However, by running comparable regressions (for the effect on the short-term) using a control group compiled of similarly sized

(32)

32 listed Danish/Swedish firms, Matsa and Miller (2013) showed that the difference in differences of ROA between the treated and control group is not due to the fact that the control group consists of non-listed firms. Therefore, the internal validity violation by the financial crisis is not proved. One limitation could be the external validity. Many papers researching the gender diversity effect found different results. Not only the organizational culture in each firm is important for the increase in firm performance (Dwyer, Richard and Chadwick, 2003), but also the country-specific corporate governance systems are relevant (Ruigrok, Peck and Tacheva, 2007). Therefore, the results found in Norway, are not directly applicable to other countries.

Endogeneity due to measurement errors in the data are unlikely. Bureau van Dijk is known to deliver data of high quality3_{. As can be seen in Table 5 in the Appendix, the number} of ASA companies drops slowly but shows no big changes. This decrease of ASA firms is mostly caused by the exclusion of delisted firms in my sample. In 2006, the amount of companies in my sample is the biggest. Ribeiro, Menghinello and De Backer (2010) find indeed that the coverage of Orbis is the highest in 2006 and a few years before. The same holds for the amount of AS firms. This means that for every year, I can assume all active ASA and AS companies are included. However, the omitted variable bias can be a problem in this thesis. For both the short- and long-term effects of the quota, the lack of many control variables can cause endogeneity to arise. Control variables such as the board size and board meeting frequency are now included in the error term instead of controlled for. This could be solved by controlling for these omitted variables, but in this thesis, these are not available.

3_{According to Ribeiro, Menghinello and De Backer (2010), the quality of the balance sheet data of the Orbis}

database is high. Especially for bigger firms and when the data is compared with other data from the same country. The largest problem of the Orbis database arises when companies from different countries are compared. In this thesis, this drawback does not apply, as only Norwegian companies are used. Another drawback is the lack of panel data on small companies. As in my analysis, the biggest ASA firms are compared to firms with at least more than USD 1 million of assets, this weakness does not influence my results.

(33)

33

6. Robustness checks

Multiple robustness checks have been done to make sure the results of this thesis are robust to certain data preparation methods and sample selection biases. In most robustness checks, all results have the same statistical significance and sign as my main results. However, in some robustness checks, the short term negative effect is not significant at a 5% level, but at a 10% level. The Figures in this section should all be interpret the same as Figure 4 in the main results section. That is, the dots are the coefficient estimates, and the vertical bounds are the 95% confidence intervals. The dots represent the percentage points increase of ROA in that year, for treated firms over non-treated firms, relative to the years before 2007.

6.1 Oil and Finance companies

During the data cleansing process, I excluded financial firms, but kept oil companies. I showed a law affecting finance companies allowed them to change their legal form. As can be seen in Figure 5 on the next page, my results are robust to keeping the finance firms. The oil firms are now not dropped. However, one could argue that the performance of oil companies is less sensitive to the financial crisis than companies of other industries. My results are robust to excluding oil companies and finance companies (Figure 6) and excluding oil companies but keeping finance companies (Figure 7). When excluding oil firms, all the signs and statistical significance are equal to my main results, except for year effect in 2013. This is significant on the 10% level.

(34)

34

Figure 5 describes the results of the difference-in-differences regression with ROA as dependent variable. Financial firms and firms in the oil industry are included. The dots represent the coefficient for that year. This coefficient is the estimated points increase of ROA in that year, for treated firms over non-treated firms, relative to the years before 2007. The vertical bands represent the 95% confidence interval. Clustered standard errors adjusted at company level are used. Also basic control variables are included.

Figure 6 describes the results of the difference-in-differences regression with ROA as dependent variable. Financial firms and firms in the oil industry are excluded. The dots represent the coefficient for that year. This coefficient is the estimated points increase of ROA in that year, for treated firms over non-treated firms, relative to the years before 2007. The vertical bands represent the 95% confidence interval. Clustered standard errors adjusted at company level are used. Also basic control variables are included.

(35)

35

Figure 7 describes the results of the difference-in-differences regression with ROA as dependent variable. Companies in the oil industry are excluded. Finance firms are included. The dots represent the coefficient for that year. This coefficient is the estimated percentage points increase of ROA in that year, for treated firms over non-treated firms, relative to the years before 2007. The vertical bands represent the 95% confidence interval. Clustered standard errors adjusted at company level are used. Also basic control variables are included.

(36)

36

6.2 Listed firms

To show my results are in line with the findings of relevant literature, I investigated the short term effects. Because I do not use Tobins Q or compare my treated firms with listed firms in other countries, I was able to use both listed and non-listed treated firms in Norway. Matsa and Miller (2013) and Ahern and Dittmar (2012), however, only used the treated listed firms. They did this to be able to either calculate the Tobins Q, or to use comparable foreign listed firms as control group. Using non-listed and listed firms in this thesis, the negative short term effects are comfirmed. To show my results are robust to using only listed firms in the treatment group, I dropped the non-listed firms. See the results in Figure 8. All statistically significant results found in my main regression have the same signs and significances as the results when the non-listed treated firms are dropped. The 2007, 2013 and 2014 coefficients are -10.693 (SE of 4.258), 8.115 (SE of 3.807) and 10.029 (SE of 3.440) respectively, and are all significant on a 5% level. Therefore, my results are robust to excluding non-listed treated firms.

Figure 8 describes the results of the difference-in-differences regression with ROA as dependent variable. Of the treated firms, only the listed ones are used. The dots represent the coefficient for that year. This coefficient is the estimated percentage points increase of ROA in that year, for treated firms over non-treated firms, relative to the years before 2007. The vertical bands represent the 95% confidence interval. Clustered standard errors adjusted at company level are used. Also basic control variables are included.

(37)

37

6.3 Selection bias

In this thesis, as described in the 4.2 Data selection section, only firms that have observations before and after 2006 are kept. Firms which have for example observations from 2004 to 2009 are kept. But suppose a firm had very low ROA in years 2010 and 2011 and went bankrupt in the beginning of 2012. This firm is included in the regression for years 2010 and 2011, but not for 2012. Then, in 2012, it will look like the firm performance on average went up, relative to the pre-2007 years. This is then not due to a positive quota-related effect in 2012, but rather due to the missing bad performing firm for the year 2012. Therefore, the estimations are likely to suffer from sample selection bias. To avoid this sample selection/survivorship bias, a robustness check is run. Here, only firms that had observation throughout the whole sample period are included. Also, the firms that changed their legal form during the sample period are included. The results are visually shown in Figure 9.

All result are robust to including these firms. The effect of the quota is statistically significant and negative for year 2007. For the years 2012, 2013 and 2014, the effect is statistically significant and positive. So when only the firms with observations throughout the whole sample period are included, an additional positive effect of the quota for 2012 is found. Yet, it remains uncertain whether this method is the correct way of handling the selection bias. The quota may have been the reason why companies went bankrupt in the first place. By excluding these companies, the real effects of the quota may be underestimated. Due to this uncertaincy, I decide to regard the effect as shown in Figure 4 as the main results.

The effect of the Norwegian gender quota on firm performance : the long run effects of the Norwegian gender quota for corporate board of directors on firm performance