Do Dividends Signal Future Earnings in the Nordic Stock Markets?

We study the informational content of dividends on three Nordic civil law markets, i.e. on markets where the evidence for dividend catering has been weak. Using aggregate data on real earnings per share and payout ratios, long time series from 1969 to 2010, and methodologies which address problems of endogeneity, non-stationarity and autocorrelation (including a VECM approach), we find evidence on dividend signaling as well as the stickiness of dividends. However, we also find heterogeneity in the relationship between dividends and earnings on markets similar in many respects, suggesting that even small variations in the institutional surroundings may be important for the results.


Introduction
Recent studies of dividends in the U.S. have reported on both the "disapperance" of dividends (Fama and French 2001), as well as of the fact that aggregate real dividends paid by industrial firms in fact have increased over the past two decades, although there has been a substantial concentration in the number of dividend paying firms (DeAngelo, DeAngelo, and Skinner 2004).
Also theories competing with the classical dividend signalling hypothesis have gained ground, such as the catering theory of dividends developed by Baker and Wurgler (2004) and tested by e.g. Ferris, Narayanan, and Sabherwal (2009). At the same time, new methodologies used to test the information content of dividends (such as VAR methods, often applied on aggregate dividends) have started to give supporting evidence to the classical dividend signalling hypothesis, see e.g. the results of Arnott and Asness (2003), Gwilym, Seaton, Suddason, and Thomas (2006), Lee and Rui (2007), as well as Lee (2010a) and (2010b). The findings of Ferris et al. (2009) indicate that the legal regime and its accompanying set of investor protections play a large role in whether dividend catering takes place. i They find weaker catering among civil law firms, suggesting that the controlling shareholders in such firms are less interested in exploring transitory market misvaluations of their firms' stock due to dividend policy.
Most of the studies on dividend signalling on aggregate data, and applying methods that better take the time-series properties of the data into account, have been performed in common law countries.
Based on the findings of Ferris et al. (2009), we expect that dividend signalling may easier be detected in civil market countries, if the catering motives for dividends do not hold and produce behaviour which clouds the information content of dividends. We contribute to the literature on dividend signalling by offering new evidence from three civil law countries, the Nordic countries of Denmark, Norway, and Sweden. Interestingly, although close to each other in terms of general corporate governance levels (see e.g. Aggarwal, Erel, Stultz, and Williamsen 2009), the countries differ largely in terms of the average value given to dividends, as evidenced in the dividend premiums reported by Ferris et al (2009). ii Using aggregate data on real earnings per share and dividend payout ratios, together with a methodology which takes problems due to non-stationarity as well as autocorrelation into account, we find strong support for dividend signalling in Sweden, and weak support in Norway. Our results also support the assumption of sticky dividends.
Besides the question of whether dividends convey information of future earnings, we also contribute to the debate concerning the right way to test for the information content of dividends.
Our results indicate that, contrary to some arguments in the literature, neither the long length of the sample period, nor the use of aggregate data instead of individual data, reduces the chances of finding significant relationships supporting dividend signaling. The methodology seems more likely to be the determining factor. Our varying results for the three Nordic countries also suggest that even smaller variations in legal regimes, corporate governance, ownership structures and / or macroeconomic environments may be of importance for the results.
The paper is structured as follows. Section 2 presents a literature review and the hypotheses to be tested in this paper. In section 3, the testing procedure is presented. The data is described in section 4. In section 5, we report on the results. Finally, section 6 offers a summary and concluding remarks. Lintner (1956) was the first to recognize the information content of dividends, and the managerial reluctance to reduce dividends. According to Lintner (1956), the underlying reason behind the managerial reluctance to reduce dividends is that investors might interpret the action as a weakness in expected future earnings. Consequently, the decision to increase dividends would depend on the firm's expected earnings stability, since managers want to make sure that the firms earnings will not decline after a raise in dividends. Dividends would, therefore, in a sense lag behind expected earnings, and be characterized as sticky. The idea that dividends could convey information about expected future earnings has later formally been presented in a number of signaling models, see e.g. Bhattcharaya (1979), Miller and Rock (1985) and John and Williams (1985), where it is the informational asymmetry between managers and outside investors that provides room for signalling with dividends. iii Alternative hypotheses for why dividends might be related to firm value include various agency cost based explanations: dividends reduce free cash flow (Easterbrook 1984), and may reduce agency problems between inside and outside shareholders (La Porta, Lopez-de-Silanes, Shleifer, and Vishny 2000). Also clientele effects may produce a price reaction around dividend changes.

Literature Review
Lintner's observation of managerial reluctance to reduce dividends has empirically been studied on many markets, with varying results. Such studies may also be seen as preliminary studies of whether dividends might be able to carry information, since Kalay (1980) has shown that in order for dividends to convey information about future earnings, managers have to be reluctant to reduce dividends. iv Kalay's (1980) own empirical evidence on US data on managerial reluctance to reduce dividends remains inconclusive, while DeAngelo et al. (1992) report supporting evidence.
Supporting findings have also been reported for the UK, where Edward and Mayer's (1986) used a survey approach regarding dividend policies directed at finance directors. They concluded that dividends are likely to be reduced if a decline in earnings is persistent, but unlikely if transient.
Likewise, Marsh (1992) has shown that managers are reluctant to reduce dividends in the UK, and more recently also Brav et al. (2005) report, on the basis of the results from a survey to financial managers in the U.S, that 94 percent of the managers in dividend paying firms strongly or very strongly agree that dividend cuts are being avoided.
Also event study tests typically confirm that the price reactions to dividend cuts / omissions are typically greater than to dividend increases, and that dividends do have some information content, since stock prices react to dividend changes (see e.g. Aharony and Swary, 1980;Dielman and Oppenheimer, 1984;Healy and Palepu 1988;Fehrs et al., 1988;Asquith and Mullins 1983;Woolridge, 1983;Brickley, 1983;Kane et al., 1984;Kalay and Lowenstein, 1985;Ofer and Siegel, 1987;Dhillon and Johnson, 1994;and Christie, 1994;Abeyratna et al. 1996). v Michaely et al. (1995) have showed that dividend initiations lead to an increase in the stock price while omissions are followed by a stock price decline. Lee (1995), who investigated the relation between dividend shocks and stock price movements, has showed that stock prices do not only react to permanent changes but also to temporary changes in dividends. Finally, Amihud and Li (2006) have reported on the declining information content of dividend announcements, and found the effect to be a function of institutional holdings. However, most of the empirical evidence on stock price reactions to dividend announcements has been concentrated around developed markets (Mollah, 2007) like the U.S and the UK. Firm size has been shown to be a distinguishing factor regarding the relationship between dividend and security prices (Eddy andSeifert, 1988, andBajaj andVijh, 1995). Eddy and Seifert (1988) find that stock price reactions of small firms in comparison with big firms are greater when they experience an unexpected dividend increase. Bajaj and Vijh (1995) find that on average during dividend announcements, abnormal stock returns get larger the smaller the size of the firm. Zeghal (1983) has found empirical evidence showing that small firms are more transparent with their financial statements compared to large firms, and therefore the information content of dividends may depend on firm size.
Since there are, besides signalling, other potential explanations for why stock prices may react to dividend changes, a more direct test of whether dividends carry information about future dividends is to study the time-series relationship between dividends and future earnings /earnings growth.
While some researchers have found evidence of a significant positive relationship between dividends and earnings (see e.g. Healy and Palepu, 1988;Kao and Wu, 1994;and Brook et al., 1998;), many others have found only a weak or no relationship at all (Venkatesh, 1989;DeAngelo et al., 1996;Benartzi et al., 1997;Daniels et al., 1997;Baker et al., 2001;and Farsio et al., 2004).
The results from most more recent research on the information content of dividend does not offer results different from the earlier ones, since there still is no consensus in empirical results. Nissim and Ziv (2001), Lee and Rui (2007) and Hanlon et al. (2006) find that dividends convey information on future earnings, while Fukuda (2000) and Grullon et al. (2005) find weak or no evidence of a dividend signalling effect. However, recent research on managerial reluctance to reduce dividends supports Lintner's (1956) observation of managerial reluctance to reduce dividends (Baker et al., 2001 andLie, 2005).
In addition to the empirical evidence on US and UK data, some empirical studies have been conducted on Western European markets. Inspired by John and Williams (1985) with their dividend signaling model, based on different tax rates for capital gains and dividends, Amihud and Murgia (1997) have tested the dividend signaling model on German data. Their results do not contradict the US findings even though dividends are taxed at a preferable rate compared to the U.S. McDonald et al. (1975) have studied 75 French firms during a 7 year period in the 1960's and found support for the Lintner model. Dasilas et al. (2008) found evidence supporting the information content of dividends in their study of Greek listed firms. A handful of studies on the relationship between dividends and earnings have also been conducted on data beyond the US, the UK and Western Europe. Lee (2010a and Lee 2010b) has tested dividend signalling on the Singapore and Australian markets, and found evidence of a dividend signalling effect and support for the observation of managerial reluctance to reduce dividends by Lintner (1956).
Most empirical studies of the relationship between dividends and earnings have been conducted using weak testing methods, such as the simple ordinary least squares (OLS) methodologies (Lee, 2010b). A simple OLS methodology assumes that the variables employed follow a stochastic process, which means that variables are assumed to be stationary. However, problems occur when conducting a simple OLS methodology but the variables are non-stationary, that is, when the variables are at least an I(1) process. In other words, they are integrated of order one and I(0) after being differenced in order to make them stationary. When there is integration of order one, and the variables are not cointegrated, OLS regressions are likely to produce spurious results (Brinca, 2006). This implies that the t-statistics will be statistically significant, and the coefficient of determination will be high, indicating a meaningful relationship that is actually not there.
However, if the variables are differenced in order to make them stationary, the long-run relationship might be destroyed (Munnell, 1992).
Another problem is autocorrelation. When the model applied is misspecified, autocorrelation in the error term might arise. An example is when the model includes lagged dependent variables. Part of the Lintner model includes prior year's dividends, so if the Lintner model is being tested with a simple OLS regression, the model will include a lagged dependent variable which means that the simple OLS regression may produce biased estimates (Frankfurter and Wood, 2003).
A third problem that may occur when applying a simple OLS technique is the endogeneity bias (Mayston, 2005), which may e.g. occur in a multivariate regression when the independent variable is correlated with the error term of the regression model (Binstock et al. 2010). This specific issue can be dealt with by applying a Vector Auto-Regressive (VAR) framework, because all variables are treated as endogenous (Sims, 1980). So far, most of the empirical research on the informational content in dividends has also been conducted on US or UK data, while continental Europe has been overshadowed (Correia da Silva et al., 2004).
In summary, it is clear that the existing research on the topic is inconclusive, i.e. offers limited evidence on the information content of dividends. With respect to major methodological developments over time, there is also a lack of studies using the most recent tools and covering long enough time periods. The major objective of this paper is, therefore, to investigate if there is any informational content of dividends in the Nordic aggregate market. The Nordic countries are also civil law countries, which may imply that dividend catering is absent (Ferris et al., 2009), and thus offer a cleaner environment for the search of the information content of dividends. In addition, since managerial reluctance to reduce dividends has been shown to be a necessary condition for dividends to convey information (Kalay, 1980) about future earnings, sticky dividends are examined. To be more specific, the following research questions were addressed in the paper: RQ1: Do dividends convey information about future earnings? RQ2: Are dividends sticky?

Tests of stationarity and cointegration
The key relationship that we want to test is that between the time-series variables of real earnings per share (EPS) and the dividend payout ratio (DPR), in logarithmic form. First, we test them for non-stationarity and the order of integration by applying the Augmented Dickey-Fuller (ADF) procedure vi , which tests for unit root when the time series variables are an I(1) process, and the Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test, which is a Lagrange Multiplier (LM) test where the time series variables are an I(0) process. vii The ADF test is applied first on a model with both a drift and trend term, and then also for trend only, and finally with neither a drift nor a trend. If the variables (EPS and DPR) are non-stationary, they have to be differenced in order to make them stationary. After differencing the variables EPS and DPR once, the same testing procedure is conducted until stationarity is achieved. The KPSS test in turn starts by regressing the variable whose stationarity is to be tested on a constant and a trend, or just a constant, which tests for trend-stationarity in the former and level-stationarity in the latter viii . The null hypothesis is a stationary process. When the null hypothesis is rejected, the test indicates non-stationarity and the variables are differenced once and tested again until achieving a stationary process. The lag order is determined by the truncation parameter where q is frequently set to 4 and 12 (Schwert, 1989).
If both variables (e.g., EPS and DPR) are integrated of order one, that is an I(1) process and I (0) after being differenced once in order to make them stationary, the variables might be cointegrated.
Therefore we next apply the Johansen (1991)  itself. If there is a cointegration between the variables, the VAR model has to be transformed into a VECM, which is a restricted VAR created to deal with non-stationary variables xi . The VECM in this paper is in line with Johansen (1991) framework as: where represents the product of the α and β matrix with the dimensions ( ), and K is the set of variables and r is the number of cointegrating relationships such that . is often referred to as the long run parameter and the short run parameter, which captures the long in the former and short run dynamics in the latter of the variables. But, when there is one cointegrating relationship, can be written as: ( 3) Granger-Causality (Granger, 1969) is a widely used phenomenon in this context. When lagged values of a variable contribute to an improvement of the forecast of a variable after being controlled for lagged values of , it is said to be casual (Wooldridge, 2005 andLutkepohl, 2004), i.e., granger causes if is a useful predictor of . The Granger causality test has the ability of predicting each one of our variables statistical impact on their future value (Brooks, 2008), that is, telling us if one variable Granger -causes the other. Brooks (2008) also argues that the information of a positive or negative relation or the time horizon of when the impact will take place cannot be predicted by a Granger-causality test but Impulse responses, xii on the other hand, have the ability of doing so. The impulse response function for the paper is presented as: ( 4) where is the identity matrix [ ] and ∑ with and for .

Data
Our study is based on the aggregate monthly data for the Danish, Norwegian and Swedish markets for the period from 1969 to 2010. Morgan Stanley Capital International (MSCI) Indexes have been chosen due to their high representation of the market capitalization in the Nordic countries, and the availability of the longest historical data. We have collected monthly levels of the price index, price-earnings ratios (P/E-ratios), dividend yields and the consumer price index (CPI) from Statistics Sweden and Investment Strategist at Handelsbanken Capital Markets. We have derived earnings per share (EPS) and dividend payout ratio (DPR) from these data and the derivation procedure xiii for EPS and DPR is the same as Lee (2010b) and is also inspired by Arnott and Asness (2003) and Gwilym et al. (2006). A time series plot of EPS and DPR is presented in figures 1 and 2, respectively, to visualize the pattern of the data for these countries below.

INSERT FIGURES 1 AND 2 HERE
Non-stationarity of the EPS and DPR data are clearly visible in figures 1 and 2. However, a structural break is identified in all these three Nordic countries. This is due to the severe banking crisis in the early 1990's. The descriptive statistics are presented in Table 1.

INSERT TABLE 1 ABOUT HERE
The EPS series (as in Lee 2010b), is negative by construction, coming from the indexing and taking the natural logarithm of the resulting ratio, but it shows a slight upward trend. It is also notable that the Scandinavian countries pay a very high and rather stable dividend.

Results from stationarity tests
First, we test the EPS and DPR series for non-stationarity in their levels and first differences by applying the ADF test. The results are presented in Table 2.
INSERT Next, the variables are tested for cointegration by applying the Johansen xiv Maximum Eigenvalue and Trace tests including constant. If too many lags are used, the degrees of freedom may consume too much power from the test, while not using enough lags may limit the test from capturing the proper dynamics (Brinca, 2006). Therefore, the chosen lag order is 4 for both tests. In order to capture the short run effects of the relationship between EPS and DPR, the Johansen Maximum Eigenvalue and Trace tests are specified with a transitory procedure and the test results are reported in Table 4.

INSERT TABLE 4 ABOUT HERE
The result of the Trace test for cointegration in Table 4 indicates a cointegrating relation between EPS and DPR. The null hypothesis of a cointegrating relation being r=0 is rejected between the variables at 5 and 10 percent significance levels. But the null hypothesis of the cointegrating relation being r=1 between the two variables failed to be rejected at a 5 and 10 percent significance level, which indicates that EPS and DPR are cointegrated. The results of the Maximum Eigenvalue test in turn indicate a cointegrating relation between EPS and DPR. The null hypothesis of a cointegrating relation being r=0 can also be rejected at a 5 and 10 percent significance level, while the null hypothesis of a cointegrating relation being r=1 failed to be rejected at a 5 and 10 percent significance level, which also support that the EPS and DPR are cointegrated.
Nevertheless, including a deterministic term (in our case a constant) might produce biased estimates if it is formulated incorrectly. Therefore, in addition to the Johansen procedure, a cointegration test that accounts for the deterministic term is applied with the same lag order as in the Johansen procedure since it goes by the Johansen Trace test after the deterministic term is estimated and subtracted. The outcome of the test are reported in Table 5.

INSERT TABLE 5 ABOUT HERE
The results of the Saikkonen and Lutkepohl test also support a cointegrating relation between EPS and DPR, and the null hypothesis of a cointegrating relation between the variables being r=0 is rejected at a 5 and 10 percent significance level. The null hypothesis of the cointegrating relation being r=1, on the other hand, failed to be rejected at 5 and 10 percent significance level.
The cointegration results based on the three tests performed clearly indicate that EPS and DPR are cointegrated, which means that the model will be specified with a VECM approach in order to avoid biased estimates.

The VECM approach
Since the models applied in the paper deal with lag values, autocorrelation might cause biased and inconsistent estimates. In order to avoid this inefficiency, the residuals in the VECM are tested for autocorrelation by applying the Portmanteau test. The test results are reported in Table 6.
There is an ongoing debate in the literature concerning the use of aggregate versus individual firmlevel data in studies of the information content of dividends. Disappointing results on aggregate data include Farsio et al. (2004), who did not found any relationships between dividends and earnings. Daniels et al. (1997) have argued that the use of aggregate data may be the reason for not being able to find a causal relationship between dividends and future earnings in previous empirical research. They specifically argue that the information content of dividends cannot be tested without obtaining biased results on aggregate data. However, they only found a Grangercausal relationship between dividends and future earnings in 50 percent of their randomly picked individual firm-level U.S observations, which does not give any strong support for their argument of getting results very different from those when using aggregate data. Dittmar and Dittmar (2002) in turn argue that firm-level data is more likely to affect the analysis due to corporate manipulations and earnings smoothing. Also Lee and Rui (2007) argue that firm-level data may cause biased estimates, and that aggregate data would therefore be preferred. Their results from a study on US data suggest that dividends convey information about future earnings, since dividends are found to Granger-cause future earnings, results similar also to Lee, 2010a and Lee, 2010b.
Interestingly, the studies that have found a Granger-causal relationship between dividend and earnings on aggregate data are the ones using a VAR framework.
Our study provides some additional information to the debate on how to study the relationship between dividends and earnings. The results indicate that, contrary to some arguments in the literature, neither the long length of the sample period, nor the use of aggregate data instead of individual data, reduces the chances of finding significant relationships supporting dividend signaling when using a Granger-causality test. The methodology seems more likely to be the determining factor. Our varying results for the three Nordic countries (in many ways similar to each other) also suggest that even smaller variations in legal regimes, corporate governance, ownership structures and macroeconomic environments are important for the results. Australia and the US are both Anglo-Saxon economies, and common law countries, while Sweden is a civil law country and differs both from the Anglo-Saxon and the Continental European model, even though Sweden is moving towards a more Anglo-Saxon inspired economy (Crouch and Streeck, 1997). To answer our first research question, based on the Granger-causality test, dividends do convey information about future earnings in Sweden, less so in Norway, while no relationship was found for Denmark.

Impulse response functions
The impulse response analysis is performed in order to trace the responses of DPR (EPS) to a unit shock in EPS (DPR). The results are presented in Figures 3 and 4, respectively.  response of dividends to a unit shock in earnings shows that dividends experience a drop for seven months when earnings experience a shock increase and then gradually increases almost throughout the whole forecast period. xv The combination of an initial drop in DPS, followed by a subsequent

Summary and Conclusions
This study addresses the issue of the informational content of dividends in the aggregate markets for three Nordic countries, and makes several contributions to the literature.
First, we contribute to the literature on the determinants of dividend policy, and the information content of dividends. The dividend signaling theory has recently been challenged e.g. by the theory of dividend catering. Ferris et al (2009) find weaker evidence of catering for civil law firms, while most of the studies of the information content of dividends using more recent methods have been performed in common law countries such as the U.S., Singapore, and Australia. If dividend catering does produce behaviour that may cloud the information content of dividends, then civil law countries offer better opportunities for finding evidence in favour of dividend signalling. We report results for three civil law countries, Denmark, Norway, and Sweden, using a long time series from 1969 to 2010. We apply the Granger-causality test and the impulse response function.
These tests indicate that dividend payout conveys information about future earnings in Sweden, while some support of Granger-causality is also obtained for Norway. We also study Lintner's (1956)            . Impulse Response function of DPR to a unit shock in EPS, for Sweden i Dividend catering refers to the managerial behaviour of opportunistically modifying corporate payout policies when investor sentiment favors the payment of dividends. Investor sentiment for dividends is in turn often measured from the dividend premium, i.e. the difference between the market-to-book ratios between dividend payers and non-payers. ii While Sweden had, in a set of 23 countries, the highest median dividend premium (0.433), closely followed by Denmark (0.327), Norway had a negative one (-0.099), indicating that dividend paying firms were not valued above non-payers in Norway. iii The idea that dividends could signal information about future earnings is not at odds with the dividend irrelevance proposition of Miller and Modigliani (1961), since there are no informational asymmetries in that model, and dividends can alternatively to being paid out, be left within the firm (however, keeping the investment policy of the firm fixed). iv However, DeAngelo and DeAngelo (1990) have argued that there are several reasons for the reluctance to reduce dividends, besides that suggested by Lintner (1956). A reluctance to cut dividends might thus be a necessary but not sufficient condition for dividends to carry an informational content.