Research and development and firm risk
Matthijs S. Suurmeijer MSc Finance Master’s thesis
Faculty of Economics and Business, University of Groningen
Supervisor: Dr. P.P.M. Smid
June 21, 2013
Abstract:
By using a panel data set this thesis examines how research and development affects various measures of firm risk such as systematic and idiosyncratic risk. Panel least squares regression, with fixed effects and robust clustered standard errors, as well as quantile regression are employed. The findings are that research and development does not impact upon firm risk for the full unfiltered sample. After removing extreme observations the impact upon systematic risk and total risk are highly significant, but not with respect to idiosyncratic risk. The economic significance, however, is smaller. In addition, the impact is non-linear.
Keywords: research and development; firm risk; beta; systematic risk; idiosyncratic risk JEL classification: G12; G14; G32
2
1.
Introduction
Research and development (R&D) investments may lead to innovative products and services and thereby to a competitive advantage. For this reason, R&D is often linked to good financial performance. For example, in their meta-analysis on financial performance, Capon, Farley, and Hoenig (1990) find a strong relation between money spent on R&D and increased profitability. Similarly, R&D is linked to high stock returns. On the flip side of the coin, there is the impact of R&D on risk. Classic finance tells us that returns and risk are positively correlated such that R&D must increase risk. This paper empirically investigates this hypothesized effect of R&D on risk. The relevance of research on the consequences of R&D is reconfirmed by a study by Hirschey, Skiba, and Wintoki (2012) who find that from various types of investments including advertising and capital expenditures, R&D spending has grown the fastest during the last two decades.
This paper does not study both the return and risk consequences of R&D investments because reviewing the literature leads to the conclusion that a positive effect of R&D on returns is a stylized fact. For example, Chan, Martin, and Kensinger (1990) investigate how announcements of increased R&D expenditures affect share value using event study methodology. They find that stock markets react to these announcements with positive and significant abnormal returns, especially for high-tech firms. In addition, Chauvin and Hirschey (1993) find consistent evidence that R&D expenditures positively affect the market value of common equity. Eberhart, Maxwell, and Siddique (2004) examine the long-term impact of an increase in R&D expenditures. They find that for a period of five years after the announcement of the increase, firms experience significant and positive abnormal stock returns as well as operating performance. According to them, investors underreact to the managerial investment decision of changes in levels of R&D expenditures. In short, a positive impact of R&D on return has been identified by various authors so that the impact of R&D on risk is particularly interesting.
3
internal cost of capital for R&D intense firms. First, because of information asymmetry and the “lemon” problem investors will require a premium since it is difficult to distinguish between good and bad R&D intense projects. Second, there is the moral hazard problem as caused by separation of ownership and control where managers do not want to invest in risky R&D projects. With regard to the uncertainty, Kothari, Laguerre, and Leone (2002) indeed find that higher R&D investments are associated with greater uncertainty of future earnings. Also, analysts are more uncertain about the future earnings for firms with higher R&D expenditures relative to those with lower (Barth, Kasznik, and McNichols, 2001). These results are confirmed by Chambers, Jennings, and Thompson (2002). Chambers et al. argue that that the association between excess returns and levels of R&D investments (in contrast to changes in the levels) stems from a risk factor associated with R&D.
This study thus empirically addresses the question whether and how the R&D intensity of a firm affects its risk. More specifically, firm risk is measured using stock market based measures because shareholders are forward-looking. Therefore, stock markets are assumed to be efficient in at least semi-strong form (Fama, 1970). Berk, Green, and Naik (2004) develop a model in which they show that R&D impacts upon systematic risk as well as unsystematic risk. The impact on systematic risk results from uncertainty in ultimate cash flows while the impact on unsystematic risk results from technical risks. Therefore, more risk measures than just beta, as proxy for systematic risk, are considered. Consequently, the main question could be decomposed into various sub questions each relating to the effect of R&D intensity on a component of stock return risk. These components include total risk, systematic risk, and idiosyncratic risk. Finally, another risk measure which is of potential interest, downside deviation, is also part of this study.
This decomposition is also one of the reasons why this study is relevant and expands the current body of literature. While research has been performed on this topic, previous studies consider only one risk measure or can be critiqued on methodological issues. More clarity on the relation between R&D and risk is therefore of value. Apart from its academic relevance, this study could be of potential interest to practitioners and managers. While decisions on R&D are usually made from a product-market perspective, this finance-based perspective is important because if R&D impacts upon risk, it also affects access to capital markets and the cost of capital, and is relevant in performance measurement. In addition, this paper might be useful to investors who need insights on the effect of R&D on the total and idiosyncratic risk of a stock as well as how R&D contributes to systematic risk in a portfolio context.
4
distribution to distinguish between high- and low-risk firms. The sample employed in this study consists of 3,008 firms (31,033 observations) listed in the United States (U.S.), over a period from 1986 up to and including 2011.
The remainder of this paper is organized as follows. The next section, 2, will review relevant literature related to risk as well as review previous empirical research on the effects of R&D on risk and develop the hypotheses. Section 3 details the data collection as well as methodology, and describes the data. Subsequently, section 4 contains the results. Finally, section 5 concludes this study.
2.
Literature review
2.1.
Market determined and fundamental risk measures
Based on mean-variance portfolio theory by Markowitz (1952) on how a risk averse investor should select securities, the capital asset pricing model (CAPM) was independently developed by Sharpe (1964), Lintner (1965), and Mossin (1966). This model states that only systematic risk, or covariance with the market returns as proxied by beta, is relevant in explaining and determining the expected return on a security. Other sources of variance, idiosyncratic risk, can be diversified away in a portfolio. Therefore, investors are only rewarded for bearing systematic risk. The beta can be computed by dividing the covariance of the returns of a security with the returns on the market portfolio by the variance of the market returns.
Since, ultimately, the risk of a firm is determined by the fundamentals of the firm and market characteristics, many authors try to explain and forecast betas based on firm variables. A study on the link between firm variables and risk of stock returns is by Beaver, Kettler, and Scholes (1970). They assume that several accounting risk measures attempt to explain earnings uncertainty and thus can be a proxy for total risk (variability of stock returns). Furthermore, based on the assumption that total risk and beta are strongly positively correlated, the following variables are associated with beta according to Beaver et al. (1970).
1) Higher dividend payout lowers beta. Since firms are reluctant to cut dividends and want to stabilize their payout, firms with greater earnings uncertainty will have lower payout ratios. 2) High growth firms are expected to be more risky since competition will erode away excess
earnings.
3) Higher leverage leads to higher betas since debt financing increases the volatility of earnings. 4) Liquidity is negatively associated with risk since current assets are less volatile than fixed
assets. However, Beaver et al. (1970) predict only a weak association since they find the role of fixed assets more important in explaining risk.
5
7) Accounting beta measures the covariance of the earnings with the market, and should therefore be positively correlated with beta.
The study by Beaver et al. (1970) inspired other academics to investigate the link between fundamental firm characteristics and beta by considering new variables. For example, a study by Rosenberg and Marathe (1975) reviews 101 variables and their influence on firm risk so as to generate a model with greater predictive power.
As mentioned above, total risk can be decomposed in systematic risk and idiosyncratic risk. Over time the CAPM has been challenged because other risk factors that covariate with non-systematic risk, or idiosyncratic risk, have been identified (Rosenberg, 1974). Roll (1977) argues that this is because the stock market indexes are not fully efficient proxies for the market portfolio. These developments led to the Arbitrage Pricing Theory and models such as the Fama and French (1993) three-factor model. In addition, since the sixties there has been an increase in idiosyncratic volatility relative to systematic volatility (Campbell, Lettau, Malkiel, and Xu, 2001). This decreases the explanatory power of the market model and requires more stocks in a portfolio to achieve full diversification. Furthermore, it has been demonstrated that idiosyncratic risk is positively associated with returns (Merton, 1987; Fu, 2009). Various explanations for the increase in idiosyncratic risk have been put forward. One of these relates to growth options (Cao, Simin, and Zhao, 2010), interestingly R&D is often used as proxy for growth options. Other explanations are an increase in general economy-wide competition (Irvine and Pontiff, 2009), new listings by risker companies (Brown and Kapadia, 2007), and retail ownership (Brandt, Brav, Graham, and Kumar, 2010). Apart from the importance of idiosyncratic risk to under-diversified investors, the increase of idiosyncratic volatility combined with the positive link with returns and the fact that idiosyncratic volatility impacts upon the rating by analysts (Lui, Markov, and Tamayo, 2007) demonstrate that idiosyncratic volatility matters as a risk measure.
2.2.
Previous research and hypotheses
6
investigate the relation between R&D intensity and systematic risk. They however find statistically significant support for their hypothesis that R&D leads to lower systematic risk. They argue that R&D effort creates strategic differentiation and insulates a firm from market downturns. This finding is despite a positive correlation coefficient between R&D intensity and systematic risk. According to McAlister et al. (2007) this discrepancy is possible due to the fact the correlation matrix does not account for fixed cross-section effects and serial correlation of beta.
The finding by Campbell et al. (2001) that idiosyncratic risk has increased led to a study by Mazzucato and Tancioni (2008) whom argue that innovation could be related to this increase. They focus on five industries and use R&D intensity as a proxy for innovation but acknowledge that R&D is only the input and not the output, which for example could be a patent. For idiosyncratic risk they use a proxy namely the ratio of firm-level return variance over market return variance. Mazzucato and Tancioni (2008) find that R&D intensity significantly and positively impacts upon idiosyncratic risk. Their model specification however lacks a clear theoretical basis.1 Relatedly, Bartram, Brown, and Stulz (2012) research the question why U.S. stocks are more volatile than stocks from other countries. They conclude that the higher return volatility is mainly caused by higher idiosyncratic risk and that higher R&D intensity is one of the reasons why idiosyncratic risk is higher for U.S. stocks. Chen, Peng, and Wei (2012) examine the effect of R&D intensity on systematic risk as well as idiosyncratic risk. In addition to OLS they employ quantile regression to account for the non-normality of risk and heteroskedasticity. This technique also enables them to compare the impact of R&D on low-risk and high-risk firms. Chen et al. (2012) find that higher R&D intensity increases the beta as well as idiosyncratic risk, especially for firms with median to high risk. However, they do not perform additional analyses, include fixed effects, or use adjusted standard errors to proof the robustness of their general results. Table 1 presents a summary of these findings.
While the literature is not conclusive on the impact of R&D on risk in stock returns most of the studies indicate a positive relation. This corresponds with the findings and theoretical implications discussed in the introduction. Therefore, I formulate the following hypotheses in an attempt to answer the research question.
H1: the higher the R&D intensity of a firm, the higher is its systematic risk;
H2: the higher the R&D intensity of a firm, the higher is its total risk;
H3: the higher the R&D intensity of a firm, the higher is its idiosyncratic risk.
1
7
Table 1 Overview of the empirical studies on the effect of R&D on the risk of stock returns
Study Risk measure(s) Finding(s) Significance
Chan et al. (2001) Total risk R&D contributes to higher stock volatility. The impact of R&D is both statistically significant and economically important.
Xu and Zhang (2004) Total risk R&D intensity has, on average, a small but positive effect on the total risk of stock returns in Japan.
In the first, bubble-forming, sub period R&D is negatively related to risk. In the second, burst-of-bubble, period R&D is positively related to risk. In the third, post-bubble, period R&D is positively related to risk.
This result is not statistically significant.
Only for the third sub period the result is statistically significant, although the magnitude during period two is large
Ho et al. (2004) Beta Simulation and correlation analysis indicate that R&D is positively associated with beta through the intrinsic business risk component. R&D intense firms carry slightly less financial leverage.
The results are particularly strong for the manufacturing sector. For the non-manufacturing sector, the results are not robust.
McAlister et al. (2007) Beta Higher R&D intensity results in lower betas. This finding is despite a positive correlation coefficient, but according to McAlister et al. (2007) this is due to the incorporation of fixed effects and serial correlation in errors.
This result is statistically significant at 1%.
Mazzucato and Tancioni (2008) Idiosyncratic risk Using firms from five industries (agriculture, textiles, pharmaceutical, computers, and biotechnologies) a positive link between R&D and idiosyncratic risk is established.
The economic importance of R&D is higher for industries which are classified as low on innovation due to clearer differences between firms in terms of innovativeness.
At 5%, this finding is statistically significant.
Bartram et al. (2012) Idiosyncratic risk Stock volatility in the U.S. is greater than in other countries because idiosyncratic risk increases with R&D.
The effect is statistically significant and economically important. Chen et al. (2012) Beta, and
Idiosyncratic risk
OLS indicates that higher R&D intensity leads to higher betas and greater idiosyncratic risk.
Quantile regression indicates that higher R&D intensity is associated with higher betas and idiosyncratic risk. The higher the riskiness of the firm, the greater the effect of R&D.
These findings are statistically significant at 1%.
8
Although these three risk measures which follow directly from asset pricing literature are the main focus of this study, another measure of return risk is included, namely downside deviation. This risk measure considers negative returns only and therefore does not treat upside potential as risk. This measure thus relates more closely to the loss aversion which characterizes human nature.
H4: the higher the R&D intensity of a firm, the higher is its downside deviation.
3.
Data and methodology
3.1.
Sample selection and data collection
Following Ho et al. (2004), McAlister et al. (2007), and Chen et al. (2012) this study uses a sample which consists of firms listed in the U.S.2. The first step in selecting the sample was to obtain constituents of major stock indexes and stock exchanges at various points in time. The following constituent lists were collected from Thomson DataStream: S&P 500 at the end of September 1989; S&P 1500 at the end of December 1994, February 2002, and February 2013; NASDAQ Composite at the end of April 2004 and February 2013; and finally all constituents of the New York Stock Exchange (NYSE) at the end of February 2013. After removing duplicates based on the International Security Identification Number (ISIN) and removing entities for which the ISIN is not available, a list of 9,334 listed entities remains. Furthermore, based on the Industry Classification Benchmark (ICB) data, several sectors are filtered out of the sample, namely Equity Investment Instruments, Nonequity Investment Instruments, and Real Estate Investment Trusts. This procedure results in a base sample of 7,323 companies. In Section 3.3 attention will be paid to describing the firms which eventually end up being used in the sample.
The data used to test the hypotheses were collected from Thomson DataStream. First, static information such as industry classification, location, and date of incorporation were obtained. Second, total return indices were collected for the period from January 1981 until December 2011 as well as price index data on the S&P 500 index which serves as market index. Third, yearly time series data were obtained for the same period including accounting, financial, and market data. These data types are: market value of common equity (share price times number of shares), total assets, revenues, net income, R&D expenditures, cash dividends, market-to-book ratio, current ratio, total liabilities, and the number of shares.
3.2.
Methodology
Although model specifications as Rosenberg and Marathe (1975) use may improve the fit relative to the specification by Beaver et al. (1970), the high number of variables makes it harder to understand the model, also due to possibly counterintuitive or conflicting coefficients. Therefore, I
2
9
will stick to the simpler model by Beaver et al. (1970) as reviewed above. Consequently, the general model for the panel data set with five-year moving averages has the structure as described in equation (1):
(1)
Where the dependent variable is one of the risk measures: beta, total risk, systematic volatility, idiosyncratic risk, or downside deviation for firm i at time t. The is a constant, is the main variable of interest (R&D intensity) which like in McAlister et al. (2007) and Chen et al. (2012) is lagged because of an anticipated lagged effect due to the payoff properties of R&D. In addition, it functions as a basic method of ruling out reverse causality. The vector contains the control
variables as described in the literature review above, except for accounting beta since five observations are not sufficient for an efficient estimate. In addition, two other control variables are included. First, firm age, which is statistically significant in McAlister et al. (2007). Second, a measure of competitive intensity is included because Meng (2008) argues that in a highly competitive environment, higher R&D intensity will result in higher betas as well as greater total return volatility. Also, both McAlister et al. (2007) and Chen et al. (2012) identify competitive intensity as a significant predictor. The vector contains the coefficients for these various control variables x. The error term is assumed to be independent and identically distributed with a mean of zero and variance σ2.
After calculating logarithmic returns based on the total return indices (and price index for the market index) the various risk measures can be determined. For this purpose the market model, or the practical application of the CAPM, is assumed to be the return generating process (Blume, 1970). Thus, equation (2) describes returns.
(2)
Where is the return of firm i at time t, is the beta of firm i at time t, is the return
on the market portfolio (S&P 500) at time t, and is an error term.
10
Figure 1 Variable measurement and moving windows procedure
Window Variable(s) Measurement, time (t)
-5 -4 -3 -2 -1 0 1
1 (observation t=0) Risk measure Control variables Lagged R&D intensity 2 (observation t=1) Risk measure
Control variables Lagged R&D intensity etcetera
In addition to β-risk, other measures of risk are also used as dependent variables. Total risk (TR) for firm i at time t is calculated as the standard deviation of the 60 monthly stock returns. Systematic risk (SR), as opposed to beta which is a proxy for systematic risk, can be calculated as follows:
√ (3)
Where σ represents the standard deviation operator and thus the second item is the squared total risk (variance) of the market index. The reason for including systematic risk in addition to beta is that this results in three risk measures using equal units of measurement. Accordingly, idiosyncratic risk (IR) equals the standard deviation of the residuals from equation (2) and can be calculated as follows:
√ (4)
Thus, with idiosyncratic risk I mean idiosyncratic variance relative to the market model. Also, this is not a proxy for idiosyncratic risk as Mazzucato and Tancioni (2008) use, but a measure of idiosyncratic risk. To show the importance of using a real measure instead of a possibly flawed proxy, consider a hypothetical firm which returns exactly follow the market index but with a multiplier effect of four. This security would thus have a market model beta of four and the ratio of firm-level return variance to market variance, which is the proxy of Mazzucato and Tancioni (2008), would be considerably higher than one.3 However, the market explains all of this firm’s return variance and idiosyncratic risk would therefore actually be equal to zero. Finally, the last measure of risk, downside deviation, δ, can be calculated as follows (Sortino and Van der Meer, 1991):
3
The exact level of the ratio depends on the distribution of the returns. The higher the degree of uniformity of the distribution, the higher is the ratio.
11
Where is the minimal acceptable rate of return which equals zero such that all negative
returns are considered.
Following McAlister et al. (2007) and Chen et al. (2012), R&D intensity is measured as the five-year moving average4 of R&D expenditures to revenues:
∑
(6)
The accounting variables which are all measured using book values, and which are included as control variables just as McAlister et al. (2007)5 do are defined as follows:
∑ (7) (8) ∑ (9) ∑ (10) ∑ (11) √ ∑ ̅̅̅̅ (12)
Where is the earnings-price ratio for firm i at time t which is calculated as:
And where ̅̅̅̅ is the expected value of the earnings-price ratio for firm i in a certain window. The market value refers to the end of year market value of outstanding shares. This earnings-price ratio can therefore be interpreted as an accounting measure of return with net income in contrast to capital gains plus dividends as used in the market measure of return.
In addition to these accounting control variables the two other control variables, firm age and the Herfindahl-Hirschman index (HHI) as measure of competitive intensity, are defined as follows:
4
Five-year averages are used to reduce the impact of noise (Beaver et al., 1970).
5
12
( ) (13)
∑
∑
(14)
HHI is the sum of squared market shares for a certain sector k. Thus, all firms i at time t belonging to industry k have the same HHI. The 19 super sectors as identified by ICB are used for the classification of the firms. This measure of competitive intensity equals one in the case of one firm with a market share of 100% in sector k and approaches zero in the case of many firms with roughly equal market shares in sector k. A potential issue is that I assume that the sample equals the whole market. If this assumption does not hold the HHI measure might be inaccurate. Based on the results by McAlister et al. (2007), the expected sign of the variable firm age is negative, such that older firms are less risky. With regard to the HHI, a negative sign is expected, such that more competitive industries (a lower HHI) are associated with higher levels of risk. Returning to the accounting variables, the expected signs follow from the relations as described above in the literature review. Thus, dividend payout is expected to have a negative coefficient. A positive sign is expected for asset growth. Also for leverage a positive relation is to be expected. Next, liquidity is expected to have a negative sign. Firm size as measured by the average of the natural logarithm of book value of total assets is expected to be negatively related to risk. Naturally, the expected sign for the relation between stock return risk and earnings variability is a positive one.
In addition, I also test for the effects of R&D intensity in a model which follows the original specification by Beaver et al. (1970) more closely. Therefore, the variable accounting beta is added to the model which is calculated based on 10 years of data. While the market beta measures the co-variability of market returns, the accounting beta measures the co-co-variability of earnings. If covar represents the covariance operator and var the variance operator:
(15)
where ∑
13
While this study uses OLS to examine the impact of R&D on risk and the robustness of this relation, quantile regression is also employed following Chen et al. (2010). This regression technique put forward by Koenker and Basset (1978) minimizes the absolute deviations relative to a quintile in the empirical distribution through linear programming. Median regression is thus a special case of quantile regression. According to Koenker and Basset (1978) the estimator has comparable efficiency to least squares while being robust to extreme values of the dependent variable. Furthermore, quantile regression minimizes bias in case of a skewed distribution (Koenker and Hallock, 2001). Patton (2009) shows that that returns are often non-normally distributed and that there may be heterogeneity of risk and argues that quantile regression is an appropriate choice in these circumstances. Also, in those circumstances estimations focused on the conditional mean can be inefficient and unreliable (Barnes and Hughes, 2002).
3.3.
Data description
14
is important to note firms from the technology sector, which are characterized by relatively high R&D expenditures as well as risk, account for the majority of the sample which might result in biases.
Table 2 Description of firms in the sample
This table describes the 3,008 firms which make up the sample. Panel A shows to which industries (super sectors from the ICB data) these firms belong. Panel B categorizes the firms with respect to country. Next, Panel C displays the current status of the equity which can be active, dead, or suspended from trading. Finally, Panel D classifies the firms based on year of incorporation.
Panel A: Panel B:
Industry (Super sector) Frequency Country Frequency
Industrial Goods & Services 583 United States 2,939
Travel & Leisure 97 Canada 46
Basic Resources 67 Israel 16
Health Care 500 United Kingdom 1
Technology 672 Taiwan 1
Banks 0 Czech Republic 1
Chemicals 83 Singapore 1
Personal & Household Goods 196 Ireland 1
Retail 204 Switzerland 1
Insurance 9 Channel Islands 1
Oil & Gas 147 Total 3,008
Utilities 89
Financial Services 28 Panel D:
Telecommunications 42 Year of incorporation Frequency
Food & Beverage 90 until 1900 45
Construction & Materials 65 1901-1910 58
Media 94 1911-1920 69
Automobiles & Parts 32 1921-1930 122
Real Estate 10 1931-1940 48
Total 3,008 1941-1950 86
1951-1960 112
Panel C: 1961-1970 235
Current status Frequency 1971-1980 256
Active 2,111 1981-1990 836
Dead 894 1991-2000 939
Suspended 3 2001-2010 202
Total 3,008 Total 3,008
15
boomed in the nineties. Also, they find that firms from the Internet and high-tech sectors were primarily responsible for this boom. The sample is thus consistent with these two findings by Ritter and Welch (2002). For the unbalanced data sample of 31,033 observations, Table 3 shows the descriptive statistics.
Table 3 Descriptive statistics for the sample of 3,008 firms during 1986-2011
This table shows descriptive statistics on all independent and dependent variables in the main regression sample which contains 31,033 firm-year observations. The dependent variables are beta as in equation (2), systematic risk as in equation (3), total risk as the standard deviation of 60 monthly stock returns, idiosyncratic risk as in equation (4), and finally downside deviation as in equation (5). The independent variables are the R&D intensity ratio, the dividend payout ratio, the average logarithmic asset growth, the leverage ratio, the liquidity ratio, the log of total assets, the five-year variability of the earnings-price ratio, the log of firm age, and the Herfindahl-Hirschman index as measure of competitive intensity. The Jarque-Bera test for normality rejects the null hypothesis of a normal distribution for all variables.
Variable Mean Median Max. Min. St. dev. Skewness Kurtosis
Beta 1.1504 1.0415 6.213 -3.612 0.7503 1.126 5.840 Total risk 0.1410 0.1253 0.939 0.028 0.0694 1.294 6.241 Syst. risk 0.0541 0.0460 0.347 0.000 0.0379 1.462 6.405 Idiosync. risk 0.1264 0.1110 0.935 0.025 0.0660 1.364 6.895 Down. dev. 0.0952 0.0840 0.631 0.012 0.0521 1.500 7.164 R&D int. 1.9266 0.0206 3,755.000 0.000 47.4279 46.815 2,641.950 Div. payout 0.2175 0.0000 76.150 -102.261 2.4890 0.667 691.128 Growth 0.0735 0.0631 2.058 -1.117 0.1329 0.951 12.408 Leverage 0.4563 0.4559 20.205 -0.307 0.3055 29.330 1,755.610 Liquidity 3.1248 2.2300 121.356 0.108 3.6148 10.074 199.952
Total assets (log) 12.9113 12.8155 20.459 5.849 1.9921 0.222 2.733 Earnings var. 0.1191 0.0503 27.822 0.001 0.4882 35.094 1,689.646 Firm age (log) 3.2334 3.1781 5.247 0.000 0.8099 -0.253 3.248
HHI 0.0385 0.0283 0.293 0.014 0.0277 3.677 22.871
16
OLS regression equation and the median quantile regression equation interesting. While the great non-normality in the independent variable R&D intensity could be resolved by trimming the distribution from outliers, this procedure will only be done later in the study since in the end trimming is a data massage method. At first glance, The HHI does not seem to be wrongly measured because its mean is 0.039 (median is 0.028) ranging from 0.014 to 0.29. The antitrust division of the U.S. Department of Justice6 considers markets where the HHI is between 0.15 and 0.25 to be moderately concentrated and markets with a HHI higher than 0.25 to be highly concentrated.
Figure 2 shows how the measures of risk evolve over the sample period. In addition, it shows the one standard deviation boundaries. Due to the sampling procedure these graphs do not necessarily represent the total U.S. stock market. However, we can see that, possibly due to time effects but also as a result of the sampling procedure, the dependent variables are not stable over time. Of the measures of risk, beta (Panel A) as proxy for systematic risk is the most stable. While the pattern of systematic risk (Panel B) over time looks like that of beta, it is clear systematic risk itself is more volatile. Also, we see that total risk (Panel C), idiosyncratic risk (Panel D), and downside deviation (Panel E) follow each other closely. Though, around the financial crisis of 2007/2008 the increase of downside deviation is naturally steeper than for the other two measures. The increase of total return volatility, leading up to the peak in the sample, however takes place before inclusion of many NASDAQ listed firms. Table 4 presents the correlation matrix for all variables for the sample as described above. As row 6 of Table 4 indicates, R&D intensity is positively and significantly correlated with the various risk measures which corresponds with the findings in previous literature. Furthermore, the correlations between R&D intensity and systematic risk as well as its proxy, beta, are smaller than the correlations with the other risk measures suggesting that the impact of R&D intensity on risk runs mainly through idiosyncratic risk which is consistent with Bartram et al. (2012). Among the correlation coefficients between the explanatory variables, the highest correlation of -0.317 occurs between leverage and liquidity. Based upon the basic method of looking at the correlations, multicollinearity issues, while not impacting upon coefficient estimates, are not expected to hamper drawing inferences or introduce instability in the estimates as a consequence of adjusting the model specification (Brooks, 2008). The correlation coefficient between systematic risk and beta is, as expected, high at 0.904 whereas the correlations of these with idiosyncratic risk are 0.402 or lower. These coefficients provide opportunity for a comparison between the impacts upon systematic risk and idiosyncratic risk. However, due to the high correlation between total risk and idiosyncratic risk the impacts upon these two variables are expected to be the comparable. Furthermore, the high correlation between downside deviation and total risk as well as idiosyncratic risk, which is not surprising given the method of measurement, raises the question whether the investigation of the impact upon downside deviation adds value.
6
17
Figure 2 Means of the measures of risk during the sample period 1986-2011
For the various risk measures, these graphs plot the means over time, for 31,033 observations. Panel A does this for beta, which is measured as in equation (2). Panel B does this for systematic risk, which is measured as in equation (3). Panel C does this for total risk, which is measured as the standard deviation of 60 monthly stock returns. Panel D does this for idiosyncratic risk, which is mesaured as in equation (4). And, finally, Panel E does this for downside deviation, which is measured as in equation (5). The dotted lines represent the plus and minus one standard deviation boundaries. The vertical axes represent the means, the horizontal axes the years.
Panel A: Beta Panel B: Systematic risk
Panel C: Total risk Panel D: Idiosyncratic risk
Panel E: Downside deviation
18
Table 4 Matrix of correlations for the sample of 3,008 firms during 1986-2011
This matrix contains pairwise correlation coefficients for all independent and dependent variables for the sample of 31,033 observations. The dependent variables are beta as in equation (2), systematic risk as in equation (3), total risk as the standard deviation of 60 monthly stock returns, idiosyncratic risk as in equation (4), and finally downside deviation as in equation (5). The independent variables are the R&D intensity ratio, the dividend payout ratio, the average logarithmic asset growth, the leverage ratio, the liquidity ratio, the log of total assets, the five-year variability of the earnings-price ratio, the log of firm age, and the Herfindahl-Hirschman index as measure of competitive intensity.
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (1) Beta 1 (2) Total risk 0.523 1 (3) Syst. risk 0.904 0.590 1 (4) Idiosync. risk 0.361 0.974 0.402 1 (5) Down. dev. 0.543 0.898 0.622 0.848 1 (6) R&D int. 0.031 0.068 0.030 0.071 0.06 1 (7) Div. payout -0.047 -0.067 -0.04 -0.066 -0.059 -0.004 1 (8) Growth -0.018 -0.030 -0.059 -0.016 -0.065 -0.024 -0.018 1 (9) Leverage -0.076 -0.072 -0.064 -0.071 -0.051 -0.009 0.025 0.025 1 (10) Liquidity 0.075 0.186 0.072 0.195 0.157 0.061 -0.029 0.009 -0.317 1
(11) Total assets (log) -0.026 -0.505 -0.028 -0.568 -0.367 -0.050 0.051 -0.001 0.253 -0.289 1
(12) Earnings var. 0.077 0.252 0.101 0.256 0.225 0.015 -0.008 0.054 0.147 0.028 -0.105 1
19
4.
Results
4.1.
Preliminary least squares regression results
The results related to the OLS estimation of the main equation (1) for the various risk measures are displayed in Table 5. Petersen (2009) shows that for 42% of the recently published finance studies that use panel regression, the standard errors are not adjusted for possible dependence in the residuals. While OLS standard errors are unbiased in the case of a well behaving error term (identically and independently distributed), they may be biased if this is not the case. Therefore, the equation (1) is estimated using White’s robust standard errors clustered at the cross-section such that heteroskedasticity and serial correlation within cross-sections are accounted for. The fact that this adjustment leads to considerably different standard errors, also later on when introducing fixed effects into the model specification, can be interpreted as evidence that there indeed is clustering at the firm level.
20
Table 5 Pooled least squares regression results for the impact of R&D on risk
This table details all the regression coefficients from the panel least squares regression estimations for the various risk measures. Below the coefficients, between parentheses, are the t-ratios. The equations have been estimated using robust standard errors (with adjustment of the degrees of freedom) clustered at the cross-section to allow for heteroskedasticity and serial correlation within a cross section.
The dependent variables are beta as in equation (2), systematic risk as in equation (3), total risk as the standard deviation of 60 monthly stock returns, idiosyncratic risk as in equation (4), and finally downside deviation as in equation (5). The independent variables are the R&D intensity ratio, the dividend payout ratio, the average logarithmic asset growth, the leverage ratio, the liquidity ratio, the log of total assets, the five-year variability of the earnings-price ratio, the log of firm age, and the Herfindahl-Hirschman index as measure of competitive intensity.
Variable Beta Syst. risk Total risk Idiosync. risk Down. dev.
Constant 1.4066 0.0665 0.3876 0.3885 0.2336
(17.499)*** (16.903)*** (48.486)*** (52.057)*** (41.749)***
Lagged R&D int. 0.0004 0.0000 0.0001 0.0000 0.0000
(3.92)*** (2.818)*** (2.37)** (2.324)** (1.758)* Div. payout -0.0125 -0.0005 -0.0010 -0.0008 -0.0007 (-3.033)*** (-2.821)*** (-3.153)*** (-3.247)*** (-3.059)*** Growth -0.2303 -0.0237 -0.0345 -0.0253 -0.0395 (-3.6)*** (-7.345)*** (-6.403)*** (-5.114)*** (-9.538)*** Leverage -0.1707 -0.0074 0.0110 0.0142 0.0065 (-1.714)* (-1.748)* (2.564)** (2.564)** (2.454)** Liquidity 0.0096 0.0005 0.0008 0.0007 0.0008 (2.228)** (2.337)** (2.617)*** (2.377)** (3.114)***
Total assets (log) 0.0252 0.0011 -0.0142 -0.0159 -0.0070
(3.903)*** (3.803)*** (-28.151)*** (-32.936)*** (-19.422)***
Earnings var. 0.1260 0.0084 0.0267 0.0244 0.0190
(2.369)** (2.486)** (2.915)*** (2.925)*** (3.069)***
Firm age (log) -0.1657 -0.0081 -0.0202 -0.0183 -0.0151
(-12.945)*** (-13.044)*** (-18.072)*** (-18.156)*** (-17.857)*** HHI 0.1629 0.0408 -0.1496 -0.1874 -0.0978 (0.595) (2.965)*** (-5.78)*** (-7.609)*** (-5.337)*** No. observations 31,033 31,033 31,033 31,033 31,033 F-statistic 174.650*** 190.770*** 1916.982*** 2516.149*** 1059.539*** Adjusted R-squared 4.8% 5.2% 35.7% 42.2% 23.5% *** statistically significant at 1%, ** at 5%, * at 10%
21
reason for this poor fit is that stock returns are non-linear functions of the returns on these R&D projects (Da et al., 2012). Consequently, having a sample which is possibly biased toward high-tech industries might influence the efficiency of the estimations with regard to measures of systematic risk. Therefore, this could be a potential explanation for counterintuitive signs of coefficients for regressions for beta and systematic risk, such as seen for leverage.
To further examine the relation between leverage and measures of systematic risk, an OLS regression with beta as dependent variable, containing only a constant and average leverage as well as fixed firm and time effects is estimated. For the unrestricted sample (7,323 firms) of 73,600 observations, the coefficient of leverage is also negative. Thus, a sample bias as a consequence of sample restrictions, mainly related to R&D data availability, does not seem to drive this confusing finding. Furthermore, removing extreme observations by trimming the top and bottom 1% of the leverage data series for this regression so as to reduce the potential impact of outliers does not result in a positive coefficient. Apart from these reasons, another explanation could be that there are endogeneity issues where the determination of beta by the market and the decision on capital structure by the firm occur simultaneously and interact with each other (Molina, 2005). As a simple solution, lagging average leverage up to five years such that there is no overlap of measurement time frame, for either regression equation (1) or the OLS regression solely including leverage does not alter the findings on the sign of leverage. The negative impact of leverage on beta as well as systematic risk thus seems to be robust.
22
Finally, for downside deviation these numbers are 0.000004, 0.0037%, and 0.0068% respectively. Thus, while the impacts are statistically significant, from an economic standpoint these impacts of R&D upon risk are negligible. These small coefficients probably occur because of the non-normal distribution of R&D intensity, mainly caused by the extreme values in the higher tail.
Potential multicollinearity issues will be explored now since, while not alarmingly high, there are many significant correlations among the explanatory variables. Table 6 shows all variance inflation factors (VIFs) for all the explanatory variables for the various risk measures. As is clear, rules of thumb, although these are not undisputed, such as a maximum VIF per variable of ten or 1.9 as mean of all VIFs, are not breached.
Table 6 Variance inflation factors (VIFs) for the independent variables for all risk measures
The dependent variables are beta as in equation (2), systematic risk as in equation (3), total risk as the standard deviation of 60 monthly stock returns, idiosyncratic risk as in equation (4), and finally downside deviation as in equation (5). The independent variables are the R&D intensity ratio, the dividend payout ratio, the average logarithmic asset growth, the leverage ratio, the liquidity ratio, the log of total assets, the five-year variability of the earnings-price ratio, the log of firm age, and the Herfindahl-Hirschman index as measure of competitive intensity. As can be seen, rules of thumb are not breaches. These rules of thumb are a maximum VIF per variable of 10, or 1.9 as average over all variables.
Variable Beta Syst. risk Total risk Idiosync. risk Down. dev.
Lagged R&D int. 1.0520 1.0331 1.0138 1.0123 1.0161
Div. payout 1.0099 1.0095 1.0115 1.0092 1.0107
Growth 1.0911 1.0852 1.1048 1.1059 1.0865
Leverage 2.1731 1.9024 1.3976 1.6929 1.3088
Liquidity 1.6307 1.5376 1.4049 1.5166 1.3743
Total assets (log) 1.7527 1.7143 1.5235 1.5825 1.5261
Earnings var. 1.1062 1.1531 1.3050 1.2812 1.2504
Firm age (log) 1.2772 1.2963 1.3176 1.2599 1.3543
HHI 1.0299 1.0267 1.0277 1.0304 1.0204
Average VIF 1.3470 1.3065 1.2341 1.2768 1.2164
As mentioned above the accounting beta measure as specified by Beaver et al. (1970) is also part of the study. The in-sample correlation between beta and accounting beta is -0.0143 and thus very small and negative. In the regression with dependent variable beta it has a positive sign but is not statistically significant.7 In all other regression it is not significant either.8 Thus, as expected, the measurement of the accounting beta seems to be inaccurate. To achieve a better measurement, a market portfolio for the earnings which is more stable would probably need to be defined, for example all S&P 500 firms each year. In addition, I suspect this measure will be more accurate for more mature firms in contrast to smaller firms from the NYSE and NASDAQ as also included in the unadjusted sample. With regard to the earnings variability, using ten years of data instead of five years hardly
7
These results are not tabulated but are available upon request, as is applicable for other unreported results.
8
23
impacts upon the coefficient. As a consequence, I will return to using the main regression specification as in equation (1) again.
Figure 3 Quantile regression coefficients for the impact of R&D on risk
These graphs show the coefficients of the lagged R&D intensity variable in the quantile regression models for the various risk measures. Panel A does this for beta, which is measured as in equation (2). Panel B does this for systematic risk, which is measured as in equation (3). Panel C does this for total risk, which is measured as the standard deviation of 60 monthly stock returns. Panel D does this for idiosyncratic risk, which is mesaured as in equation (4). And, finally, Panel E does this for downside deviation, which is measured as in equation (5). The dotted lines represent the 95% confidence interval. The vertical axes represent the coefficients, the horizontal axes the quintiles of the distribution of the dependent variable.
Panel A: Beta Panel B: Systematic risk
Panel C: Total risk Panel D: Idiosyncratic risk
Panel E: Downside deviation
24
4.2.
Quantile regression results
Chen et al. (2012) use quantile regression since this technique is more robust to non-normality of the dependent variable. In addition, it allows for examining the impact of the variables at different points in the distribution of the dependent variable so as to explore heterogeneity of risk (Patton, 2009). All the tables detailing the results for these estimations can be found in the Appendix. Some observations related to the control variables are that growth impacts positively upon risk for firms with low or median betas (low risk for other risk measures), which is the expected relation. For higher risk firms this relation is negative, which is the sign in the OLS regression. This observation reconfirms the remark that the sample seems to contain many stable growth companies. With regard to the leverage variable, which lowers beta according to the OLS regression, we see that the coefficient is also negative for all quintiles. However, the impact is greater for low risk firms, which is in accordance with the proposition that low risk firms can have more debt in their optimal capital structure. Also, we notice that leverage, increasingly, increases the total risk and idiosyncratic risk of a firm. Furthermore, the earnings variability variable shows an impact of greater magnitude for higher levels of firm risk for all dependent variables. Lastly, following the OLS regression results competitive intensity affects, unexpectedly, both systematic risk measures positively. However, the quantile regressions show that the higher the risk of the firm the smaller the coefficient. In addition, competitive intensity has the expected negative sign in the equations with total risk and idiosyncratic as dependent variables. The coefficients become of larger size for higher risk firms as economic intuition predicts.
Chen et al. (2012) find that for higher risk firms, the impact of R&D intensity on firm risk measures becomes larger. Figure 3 graphs the coefficients at the various quintiles for the R&D intensity variable with respect to the various risk measures. As Panel A shows for beta, the consistently increasing coefficients as identified by Chen et al. (2012) are not visible in this study. In contrast, the coefficient peaks around the median. In addition, both at the left and right tails the coefficients are significantly different from each other. With regard to systematic risk (Panel B) we see a rather flat line for the coefficient with a wide 95% confidence interval at the 0.4 quintile. The other risk measures show an, on average, increasing impact of R&D intensity on risk for higher levels of firm risk. However, there is considerable uncertainty in the estimations as indicated by the 95% confidence intervals as well as the humps in the lines. In addition, in resemblance of the results from the OLS analysis we see small coefficients indicating that while quantile regression is robust to non-normality of the dependent variable it is of little use to this paper since here the biggest issue is the non-normality of the independent variable R&D intensity. Consequently, after trimming9 the top 2.5% observations of the raw, as opposed to in-sample, R&D intensity data series, removing observations with negative revenues, and subsequently re-estimating the equations the results change (see Figure 4).
9
25
Figure 4 Quantile regression coefficients for the impact of R&D on risk after deletion of extreme values
These graphs show the coefficients of the lagged R&D intensity variable in the quantile regression models for the various risk measures. Panel A does this for beta, which is measured as in equation (2). Panel B does this for systematic risk, which is measured as in equation (3). Panel C does this for total risk, which is measured as the standard deviation of 60 monthly stock returns. Panel D does this for idiosyncratic risk, which is mesaured as in equation (4). And, finally, Panel E does this for downside deviation, which is measured as in equation (5). The dotted lines represent the 95% confidence interval. The vertical axes represent the coefficients, the horizontal axes the quintiles of the distribution of the dependent variable. These are for the adjusted sample in which 1.5% of the observations relative to the other figure have been deleted. These deleted observations consist of some with negative revenues, and for the largest part extreme (high) records for the variable R&D intensity.
Panel A: Beta Panel B: Systematic risk
Panel C: Total risk Panel D: Idiosyncratic risk
Panel E: Downside deviation
First, the coefficients are of bigger size and the confidence intervals narrow. Second, the patterns observed by Chen et al. (2012) show up for all risk measures, namely an increasing impact of R&D intensity on the dependent variables for higher risk levels. The sample restriction leads to 30,560
26
observations instead of 31,033, or a decrease of the number of observations of 1.5%. The impact of this extreme value removal procedure on least squares regression will be discussed in the next section. In addition, the next section will discuss another issue. There is the potential problem that R&D is not significant in the first place because there is an omitted variable which could potentially be captured by firm fixed effects. Quantile regression however does not provide this opportunity. In addition, quantile regression does not allow for the use of standard errors which are robust to serial correlation within cross-sections.
4.3.
Further least squares regression results
To check the robustness of the results from the least squares regression discussed above, firm fixed effects have been added to the model specification to control for potential unobserved heterogeneity problems, or more specifically omitted variable bias. While McAlister et al. (2007) use fixed effects in their model; Chen et al. (2012) do not include them even though their model includes less control variables such that we cannot be assured that their OLS results are robust to the omitted variable bias. Table 7 shows the regression results with firm-level fixed effects such that the constants are allowed to differ between cross-sections. Tests of significance indicate that fixed effects are indeed present for all risk measures. Hausman tests for correlated random effects strongly reject the hypothesis that the effects are uncorrelated with a p-value of 0.0000 and thus lead to the acceptance of fixed effects in the various models. Furthermore, Hirschey et al. (2012) report that R&D spending is clustered within certain industries. Adding fixed effects makes potential industry effects superfluous since these cross-section effects can capture both industry-specific and firm-specific variation.
We now see that liquidity has its expected sign in all but one of the cases. However, the HHI coefficient in the beta specification now takes a larger value, still with a counterintuitive sign. With regard to the research question of this study, Table 7 shows that in the fixed effects specification R&D intensity does not significantly affect any of the firm risk measures. In addition, the coefficients are still of small size. The question therefore arises, whether the R&D intensity variable would have been significant in Chen et al. (2012) their OLS regression results if they would have included firm fixed effects. In addition, the result by McAlister et al. (2007) that more R&D intense firms have lower betas is still puzzling because, despite the possible impact of extremely high R&D intensity values in this study, and irrespective of the small coefficients and/or the lack of their significance, the sign of the coefficient of the R&D intensity variable has always been positive so far. In addition, the fact that an omitted variable as captured by firm fixed effects captures some variance previously attributed to R&D questions the relevance of the results from the previous section on quantile regression.
27
impact of the trimming procedure on the R&D intensity variable transparent, Figure 5 graphs the mean of the R&D intensity variable over time. 10 As is clear, trimming 1.5% of the number of observations results in a much more stable measure as it undoes the dramatic pattern as described in Section 3.3.
Table 7 Panel fixed effects regression results for the impact of R&D on risk
This table details all the regression coefficients from the panel least squares regression estimations with cross-section fixed effects for the various risk measures. Below the coefficients, between parentheses, are the t-ratios. The equations have been estimated using robust standard errors (with adjustment of the degrees of freedom) clustered at the cross-section to allow for heteroskedasticity and serial correlation within a cross section. The Hausman test for correlated random effects indicates, for all dependent variables, that the effects are of a fixed nature.
The dependent variables are beta as in equation (2), systematic risk as in equation (3), total risk as the standard deviation of 60 monthly stock returns, idiosyncratic risk as in equation (4), and finally downside deviation as in equation (5). The independent variables are the R&D intensity ratio, the dividend payout ratio, the average logarithmic asset growth, the leverage ratio, the liquidity ratio, the log of total assets, the five-year variability of the earnings-price ratio, the log of firm age, and the Herfindahl-Hirschman index as measure of competitive intensity.
Variable Beta Syst. risk Total risk Idiosync. risk Down. dev.
Constant 0.6708 -0.0271 0.2450 0.2766 0.1104
(3.843)*** (-2.917)*** (19.389)*** (23.051)*** (10.757)***
Lagged R&D int. 0.0002 0.0000 0.0000 0.0000 0.0000
(1.378) (1.158) (0.987) (0.88) (1.238) Div. payout -0.0056 -0.0002 -0.0004 -0.0003 -0.0003 (-2.366)** (-1.89)* (-2.112)** (-2.129)** (-1.823)* Growth -0.1999 -0.0255 -0.0297 -0.0184 -0.0381 (-3.129)*** (-7.363)*** (-5.344)*** (-3.661)*** (-8.296)*** Leverage -0.0831 -0.0033 0.0141 0.0163 0.0085 (-2.364)** (-3.088)*** (1.529) (1.712)* (1.474) Liquidity -0.0005 0.0001 -0.0009 -0.0011 -0.0005 (-0.122) (0.752) (-2.763)*** (-3.424)*** (-1.738)*
Total assets (log) 0.0537 0.0052 -0.0054 -0.0080 -0.0001
(3.557)*** (6.292)*** (-4.979)*** (-7.989)*** (-0.139)
Earnings var. 0.0721 0.0054 0.0229 0.0213 0.0158
(1.3) (1.422) (2.099)** (2.125)** (2.23)**
Firm age (log) -0.0767 0.0018 -0.0102 -0.0125 -0.0046
(-2.316)** (0.999) (-3.925)*** (-5.161)*** (-2.069)** HHI 2.1047 0.2823 -0.1352 -0.2952 -0.0018 (5.377)*** (12.879)*** (-5.334)*** (-13.04)*** (-0.085) No. observations 31,033 31,033 31,033 31,033 31,033 F-statistic 11.678*** 10.183*** 28.985*** 33.896*** 16.921*** Adjusted R-squared 50.9% 47.2% 73.1% 76.2% 60.7%
p-value Hausman test 0.000 0.000 0.000 0.000 0.000
Cross-section effects Fixed Fixed Fixed Fixed Fixed
*** statistically significant at 1%, ** at 5%, * at 10%
10
28
Figure 5 Comparison between the R&D intensity variable and the trimmed R&D intensity variable
This figure shows the evolution of the R&D intensity variables over time. The dotted line does this for the R&D intensity variable of 31,033 observations; this pattern was described in Section 3.3. The solid line does this for the R&D intensity variable where extremely high observations (as well as some observations with negative revenues) are trimmed, leading to a total of 30,560 observations.
Table 8 contains the results on the panel least squares regression for the sample adjusted for extreme values of the variable R&D intensity. For this restricted sample, the coefficients for this variable are larger and in all cases different from zero with four decimal places. Furthermore, R&D intensity significantly increases beta at a significance level of 1%, increases total risk at a significance level of 5%, and is positively associated with the other risk measures at a weak significance level of 10%. The economic significance of these variables will be investigated in a similar manner as in Section 4.1. Thus, the economic impact upon risk of spending 10% of your revenues on R&D (or, 10% more than a comparable firm) will be presented as well as related to the means and standard deviations of the various risk measures. First, for beta the impact of this 10% is 0.009304, or 0.8128% of the mean which is 1.2442% of the standard deviation, ceteris paribus. Second, with regard to systematic risk these numbers are, respectively, 0.000254, 0.4725%, and 0.6719%. For total risk these are 0.000645, 0.4632%, and 0.9507%. These effects are 0.000537, 0.4311%, and 0.8358% for idiosyncratic risk. Last, with regard to downside deviation, 0.000432, 0.4586%, and 0.8439% are the numbers respectively. Thus, the impact upon beta is the largest. However, consider the average firm of this sample which has a market beta of 1.14 and imagine a situation in which this firm decides to adapt its strategy by spending 10% more of its revenues on R&D activities so as to innovate. As a result, the firm will have a beta of 1.15. For such a major business decision the financial consequences are rather small. These results, however, reconfirm that the finding of McAlister et al. (1970) that R&D lowers beta seems fragile. In addition, it supports the finding by Chen et al. (2012) and Ho et al. (2004) that the higher the R&D intensity of the firm the higher its beta (or total risk). However, these results from Table 8 seem more robust due to precluding an omitted variable bias and controlling for heteroskedasticity and serial correlation within a cross-section.
29
Table 8 Panel fixed effects regression results for the impact of R&D on risk after deletion of extreme values
This table details all the regression coefficients from the panel least squares regression estimations with cross-section fixed effects for the various risk measures. Below the coefficients, between parentheses, are the t-ratios. The equations have been estimated using robust standard errors (with adjustment of the degrees of freedom) clustered at the cross-section to allow for heteroskedasticity and serial correlation within a cross-section. The Hausman test for correlated random effects indicates, for all dependent variables, that the effects are of a fixed nature. These are for the adjusted sample in which 1.5% of the observations relative to the other tables have been deleted. These deleted observations consist of some with negative revenues, and for the largest part extreme (high) records for the variable R&D intensity.
The dependent variables are beta as in equation (2), systematic risk as in equation (3), total risk as the standard deviation of 60 monthly stock returns, idiosyncratic risk as in equation (4), and finally downside deviation as in equation (5). The independent variables are the R&D intensity ratio, the dividend payout ratio, the average logarithmic asset growth, the leverage ratio, the liquidity ratio, the log of total assets, the five-year variability of the earnings-price ratio, the log of firm age, and the Herfindahl-Hirschman index as measure of competitive intensity.
Variable Beta Syst. risk Total risk Idiosync. risk Down. dev.
Constant 0.5755 -0.0315 0.2262 0.2589 0.0976
(3.246)*** (-3.339)*** (19.131)*** (23.133)*** (9.722)***
Lagged R&D int. 0.0930 0.0025 0.0065 0.0054 0.0043
(2.61)*** (1.851)* (2.253)** (1.95)* (1.712)* Div. payout -0.0056 -0.0002 -0.0004 -0.0003 -0.0003 (-2.378)** (-1.906)* (-2.135)** (-2.147)** (-1.839)* Growth -0.1896 -0.0259 -0.0270 -0.0152 -0.0348 (-2.934)*** (-7.342)*** (-5.04)*** (-3.15)*** (-7.485)*** Leverage -0.0604 -0.0032 0.0127 0.0151 0.0074 (-2.162)** (-2.849)*** (1.445) (1.625) (1.352) Liquidity 0.0003 0.0003 -0.0010 -0.0012 -0.0004 (0.058) (1.091) (-2.765)*** (-3.683)*** (-1.187)
Total assets (log) 0.0578 0.0055 -0.0041 -0.0068 0.0008
(3.781)*** (6.575)*** (-3.986)*** (-7.206)*** (0.883)
Earnings var. 0.1713 0.0136 0.0490 0.0455 0.0329
(3.272)*** (4.986)*** (9.671)*** (9.689)*** (8.379)***
Firm age (log) -0.0776 0.0014 -0.0112 -0.0134 -0.0054
(-2.344)** (0.761) (-4.534)*** (-5.793)*** (-2.492)** HHI 2.1511 0.2834 -0.1243 -0.2846 0.0041 (5.516)*** (12.993)*** (-4.995)*** (-12.831)*** (0.191) No. observations 30,560 30,560 30,560 30,560 30,560 F-statistic 11.818*** 10.281*** 29.584*** 34.715*** 17.071*** Adjusted R-squared 51.3% 47.4% 73.6% 76.6% 61.0%
p-value Hausman test 0.000 0.000 0.000 0.000 0.000
Cross-section effects Fixed Fixed Fixed Fixed Fixed
*** statistically significant at 1%, ** at 5%, * at 10%
30
addition, they are correctly specified and indeed of fixed nature according the Hausman test. With regard to the significance, signs, and magnitudes of the coefficients of R&D intensity nothing changes. For the beta regression the constant becomes closer to one. Also, the coefficient of the competitive intensity variable is now negative but still not significant. These two changes make the economic interpretation of the equation easier. For the other risk measures there are no noteworthy differences in results. For the adjusted sample in which extreme values are trimmed we now see that the impact of R&D, in the presence of time effects, on the measures systematic risk and total risk are statistically significant at 1% and the impact on downside deviation at 5%. This is primarily caused by lower standard errors. With regard to idiosyncratic risk there are no changes. Thus, while the positive impact on risk is statistically significant at a weak 10% for idiosyncratic risk, and at stronger levels for the other measures of risk, the economic significance of the effects on stock return risk is rather low.
4.4.
Additional analyses
Since R&D is often used as proxy for growth opportunities (Hirschey et al., 2012), it is important to check whether the impact of R&D intensity is really due to investments so as to innovate or whether it proxies growth opportunities. Another variable which is often used to proxy growth opportunities is the market-to-book ratio (Fama and French, 1993). Although the growth of assets is already included in the model, the market-to-book ratio measures a different concept. The growth of assets measures historical growth of an accounting measure during the last five years. On the other hand, the market-to-book ratio includes the market-based valuation by shareholders who are forward looking and thus this variable measures perceived growth opportunities. For the adjusted sample, the inclusion of the market-to-book ratio in the model does not alter the results on the impact of R&D intensity of risk. All coefficients hardly change and for none of the risk measures there is a noticeable shift in statistical significance for any of the explanatory variables. With regard to the market-to-book ratio there is, as expected, a positive impact upon risk for all risk measures, however only statistically significant at 5% for total risk and idiosyncratic risk. Thus, growth opportunities do not seem to be responsible for the results related to the R&D intensity of a firm.
