Religion, Gambling Attitudes and Corporate Innovation

Transcription

1 Religion, Gambling Attitudes and Corporate Innovation Binay Kumar Adhikari and Anup Agrawal* Current draft: November 2013 Comments welcome *Both authors: University of Alabama, Culverhouse College of Business, Tuscaloosa, AL Adhikari: (205) ; Agrawal: (205) We thank David Cicero, Jack He, Paul Hsu, Chuck Knoeber, Jim Ligon, Dave Mauer, Shawn Mobbs, Harris Schlesinger, Tony Via and seminar participants at the University of Alabama for helpful comments and suggestions. We acknowledge financial support from a summer research grant from the Culverhouse College of Business (Adhikari) and William A. Powell, Jr. Chair in Finance and Banking (Agrawal).

2 Religion, Gambling Attitudes and Corporate Innovation Abstract We examine the effect of local gambling preferences on corporate innovation. Using a county s Catholics to Protestants ratio (CPRatio) as a proxy for local gambling preferences, we find that firms headquartered in areas with greater tastes for gambling tend to be more innovative, i.e. they obtain more and better quality patents. They do so partly by spending more on R&D, which makes their stock more lottery-like, a feature desired by investors who like gambling. R&D spending is more sensitive to local gambling preferences if local investors are economically more important to a firm. Moreover, CEOs hold more stock options in higher CPRatio areas, and their option incentives positively predict firms R&D spending. Consistent with the notion that part of R&D expenditure is motivated by local gambling preferences, firms in more gambling-tolerant areas tend to be less efficient in producing patents and citations for a given R&D expenditure; they also hold more cash, partly to finance innovation. Finally, firms in areas with greater gambling preferences appear to be more adept at translating industry growth opportunities into firm value. Keywords: Religion, Gambling Attitudes, Corporate Innovation, R&D, Patents JEL Codes: G30, L26

3 Religion, Gambling Attitudes and Corporate Innovation 1. Introduction Do local culture and beliefs affect real economic activity? While it seems natural for firms to respond to the preferences of local residents, many of whom are connected to the firm as investors, managers, employees, customers or suppliers, not much research has been done in financial economics to identify the exact mechanisms through which firms react to the opportunities and challenges posed by local culture. 1 We address this question by focusing on how a local cultural trait, namely gambling preferences, affects corporate innovation, which is a key driver of real economic activity. The effect of investors gambling preferences on financial markets has drawn substantial interest in recent times. Influential works by Friedman and Savage (1948), Kahneman and Tversky (1979), Tversky and Kahneman (1992), and Barberis and Huang (2008) have uncovered aspects of human nature involved in transforming objective probabilities that lead them to develop a preference for skewness. A common theme in these papers is that investors tend to overweight a small probability of very high returns and underweight large probabilities of smaller losses, and exhibit a preference for lottery-like or positively skewed return distributions. Empirically, Kumar (2009) finds that a majority of individual investors prefer stocks that possess lottery-like features such as low price, high idiosyncratic volatility and high idiosyncratic skewness. Moreover, rather intriguingly, Kumar, Page and Spalt (2011, hereafter KPS) find that even institutional investors, which generally seek safer investments, tend to overweight lottery-type stocks in their portfolios if they are located in gambling-tolerant areas. All these findings support the notion that, either because of a belief in their luck or overconfidence in their abilities, some people have deep-rooted tendencies to gamble. While existing papers clearly establish the effect of investors attitudes towards gambling on financial markets, surprisingly little research examines how gambling attitudes affect corporate policies. This is an important issue. Given prior evidence of a local bias in individual and institutional stock ownership (see, e.g., Huberman (2001), Grinblatt and Keloharju (2001), Ivković 1 Hilary and Hui (2009), discussed later, is an important exception. 1

4 and Weisbenner (2005), and Massa and Simonov (2006)), firms have incentives to respond to the preferences of local investors when making financial decisions. Consistent with this idea, Becker, Ivkoviĉ and Weisbenner (2011) find that firms respond to the demand for dividends from older local investors. We extend this line of research and examine the relation between the gambling preferences of the local community and a corporate policy directly affected by such preferences, namely a firm s innovation activity. Our premise is simple and intuitive. Both retail and institutional investors tend to overinvest in local stocks. Moreover, in areas which are culturally more tolerant to gambling, investors are more likely to hold lottery-type stocks and are even willing to pay a premium for the lottery factor (see Kumar (2009), KPS (2011)). These two tendencies together should incentivize firms in gamblingtolerant areas to introduce lottery-like characteristics in their stock to cater to local investors. In addition, managers in these areas may also have a preference for gambling, which can lead them to hold more stock options in their firms, 2 giving them an added incentive to make their stock lotterylike. In this paper, we identify one potential mechanism that firms employ to accomplish this objective, namely higher levels of research and development (R&D) expenditure, which contributes to both the lottery factor of a stock and higher innovation in terms of patents and citations. We conjecture a positive relation between gambling preferences and innovation. Many attempts to come up with new products, services and methods can be viewed as gambles because these endeavors often promise relatively small probabilities of large success and large probabilities of failure. While successful innovations generate fame and fortune for a few, the vast majority of innovative ventures fail. 3 While gambling preferences may not always spur innovation, innovators and gamblers share some common characteristics. For instance, gambling preference represents behavioral attributes such as tolerance of risk and failure (see, e.g., Tian and Wang (2013)), a focus on maximum possible reward (see, e.g., Bali, Cakici and Whitelaw (2011)), overconfidence in one s ability and luck, and a need for fantasizing (see, e.g., Burns, Gillett, Rubinstein, and Gentry 1990)). These attributes contribute positively to innovation. So, since innovative activities have lottery-like features, they should attract individuals and firms with a taste for gambling. 2 KPS (2011) find that broad-based employee stock option plans are more popular among firms in gambling-tolerant areas. 3 Even when R&D projects result in patents, few patents have substantial commercial value. Business Week (November 9, 2005) quotes Richard Maulsby, the Public Affairs Director of the U.S. Patent and Trademark Office as saying, There are around 1.5 million patents in effect and in force in this country, and of those, maybe 3,000 are commercially viable. 2

5 We further conjecture that the effect of investors gambling preferences on innovation should be larger in firms in innovative industries for two reasons. First, firms in innovative industries likely have greater flexibility and resources (e.g., a larger R&D budget) to engage in innovative activities, so they should have a greater ability to cater to their investors gambling preferences. Second, such firms likely have more growth opportunities and a history of past success in innovation. So, gambling-tolerant investors are more likely to be attracted to such firms, hoping that past success will be repeated. The focus of this paper is on how innovation outcomes are affected by the gambling attitudes of local residents who likely invest in local companies. Gambling attitudes of firm managers can amplify this effect for two reasons. First, innovative endeavors typically have embedded real options, so they are harder to value than usual stand-alone capital expenditure decisions. This provides more room for managers to exercise their discretion and act on their instincts. Second, the success of innovative endeavors requires an extended timeframe to judge, which further appeals to managers overconfidence and illusion of control. Indeed, the existing literature on innovation suggests that most successful innovators have overconfidence in their abilities (Galasso and Simcoe (2011), and Hirshleifer, Low and Teoh (2012, henceforth HLT), a characteristic shared by many gamblers (see, e.g., Goodie (2005), and Lakey, Rose, Campbell and Goodie (2008)). To test these conjectures, we make use of a large dataset of US public companies and investigate their research and development (R&D) expenditures as innovative inputs and patents and citations as innovative outputs. Since direct measures of people s gambling preferences are hard to find in workable detail for an extended period, we follow the previous literature and examine the heterogeneity in gambling preferences of firms local communities induced by their religious beliefs. Specifically, we follow Kumar (2009), KPS (2011), and Schneider and Spalt (2013) and consider the ratio of Catholics to Protestants (CPRatio) in the county of a firm s headquarters as a measure of local gambling preferences. There is ample justification for choosing this variable as a proxy for local gambling preferences. First, Protestant philosophy strongly condemns any kind of gambling activity, whereas Catholic philosophy is somewhat tolerant of gambling. 4 In fact, Catholic churches even use 4 The following excerpt from the online version of 1913 Catholic Encyclopedia ( conveys the view of the Catholic Church on gambling: In its moral aspect, although gambling usually has a bad meaning, yet we may apply to it what was said about betting. On certain conditions, and apart from excess or scandal, it is not sinful to stake money on the issue of a game of chance any more than it is sinful to insure one's property against risk, or deal in futures on the produce market. As I may make a free gift 3

6 bingo and lotto for their own fundraising. Consistent with this notion, Kumar (2009) finds that individual investors in predominantly Catholic (Protestant) locations invest more (less) in lotterytype stocks. Moreover, KPS (2011) find that even institutional investors, who generally avoid high risk stocks, hold disproportionately higher levels of lottery-type stocks if they are located in counties with relatively greater concentrations of Catholics. In a corporate finance setting, Schneider and Spalt (2013) find that firms in high CPRatio areas tend to acquire firms whose stocks have lotterylike features such as higher idiosyncratic volatility and idiosyncratic skewness. Second, following previous studies, we assume a local contagion effect, i.e., the dominant religion in an area affects local culture and systematically influences the behavior of local residents even if they do not personally adhere to that belief system. 5 We find that firms headquartered in counties with more Catholics than Protestants innovate more in terms of the number and quality of patents. Figure 1 gives a snapshot of a part of our main findings. We sort firms into quintiles based on their CPRatios, where the sorts are made in every year of available data on religion. Panels A through E depict the means of (eventually granted) patent application counts, citations per patent, technology-adjusted citations, R&D expenditure to total assets ratio, and cash plus marketable securities to total assets ratio, respectively over the CPRatio quintiles. Panels A, B and C show a clear upward trend in the mean of patent counts, citations per patent and technology-adjusted citations, respectively, for each quintile. Despite the lack of monotonicity, the evidence is clear that firms headquartered in counties with a high CPRatio tend to have higher levels of patents and patent citations. Panel D shows a monotonically increasing average R&D to assets ratio over the CPRatio quintiles, which suggests that greater R&D expenditure leads to more innovation by firms located in high CPRatio counties. All of these suggestive depictions are later confirmed by more rigorous analysis. In all of our regressions with patent counts, citations per patent and technology-adjusted citations as dependent variables, we include the average of the first and second lags of R&D expenditures as an explanatory variable. This lag choice is consistent with prior evidence (see, e.g., Pakes and Griliches (1980)) that the average gestation period between R&D spending and patent of my own property to another if I choose, so I may agree with another to hand over to him a sum of money if the issue of a game of cards is other than I expect, while he agrees to do the same in my favour in the contrary event. 5 Our results are similar when we use an alternate measure of local gambling preferences based on state laws on gambling (see section 3.4 on robustness checks below). 4

7 applications is between one and two years. We find that there is an incremental impact of local gambling attitudes on patents and citations in the sense that our gambling proxy obtains a positive and significant coefficient in predicting patents and citations even after controlling for the level of R&D expenditures. One possible interpretation of this effect is that firms in gambling prone areas invest relatively more in the R component in R&D which is riskier but more productive than the D component. 6 If we remove R&D expenses from the patents and citations regressions, the marginal effect of gambling attitude on innovation increases slightly with somewhat greater statistical significance. All these results are robust to several alternate specifications and are not driven by firms in just a few industries or locations. Moreover, they are supported by an analysis of firms that relocate their corporate headquarters. Exploring further, we uncover evidence suggesting that part of the incentive for these firms to spend more on R&D is to make their stock more lottery-like. In particular, we find that R&D expenditures positively and highly significantly predict both idiosyncratic volatility and idiosyncratic skewness of a firm s stock returns, both of which have been shown to be desired by investors with gambling preferences (e.g., Kumar (2009), and KPS (2011)). Digging deeper, we find that the sensitivity of the level of R&D expenditures to local gambling preferences is higher among firms for which local investors are likely to be economically more important. While we focus on the gambling preferences of local investors because our sample consists of all CRSP-Compustat firm-years with available data rather than the much smaller Execucomp subsample for which we also have data on CEO pay, the gambling preferences of local managers also appear to matter. In our Execucomp subsample, we find that in high CPRatio areas, CEOs hold more stock options in their firms, and incentives from their option holdings positively predict firms investments in innovation, in addition to an effect of local investor preferences. Moreover, the productivity of R&D expenditures in producing patents and citations is lower for firms in counties with a higher CPRatio. This finding is consistent with our conjecture that firms in high CPRatio areas overinvest in R&D to cater to local gambling preferences. As additional support for our main findings, we find that firms in higher CPRatio areas hold more cash, partly to fund R&D. Accordingly, Panel E of Figure 1 shows a monotonically increasing cash to assets ratio over the quintiles of CPRatio, a pattern confirmed by more rigorous analysis. Finally, firms in more 6 HLT (2012) offer a similar interpretation of the effect of CEO overconfidence on innovation. 5

8 gambling-tolerant areas appear to be more adept at transforming industry growth opportunities into firm value as measured by Tobin s Q. Our work contributes to the literatures on corporate innovation, religion and finance, and the influence of local demographics on financial decisions. First, the paper makes a unique contribution to the growing literature on finance and innovation. Most existing papers on firm innovation appeal to rationality and are based on factors such as a firm s industry, economic environment and managerial incentives. For example, in a formal model, Manso (2011) shows that managerial incentives that tolerate early failure and reward long-term success lead to higher innovation. Tian and Wang (2011) support this notion empirically by showing that IPO firms backed by more failure-tolerant venture capitalists tend to be more innovative. He and Tian (2013) find that firms followed by more analysts tend to be less innovative, consistent with the idea that analysts exert too much pressure on managers to meet short-term goals and encourage managerial myopia. Other recent papers examine the effects of product market competition (Aghion, et al. (2005)), financing risk (Nanda and Rhodes-Kropf (2011)), corporate governance (Atanassov (2013)), stock market liquidity (Fang, Tian, and Tice (2013)), and financial development (Hsu, Tian and Xu (2013)) on innovative activities. 7 However, to our knowledge, the only paper that identifies a behavioral determinant of innovation is HLT (2012), who find a positive influence of CEO overconfidence on innovation. We extend this literature by examining the role of local residents gambling preferences in corporate innovation. Second, this work builds on the growing literature on religion-induced behavioral differences and their impact on firms and financial markets. Our paper is closely related to Hilary and Hui (2009), who find that firms headquartered in counties with higher proportions of religious residents take significantly lower risk as evidenced by lower investments and R&D expenditures, and achieve lower growth. 8 We build on Hilary and Hui (2009) s work in at least two ways. First, we extend their analysis to patents and citations. While R&D expenditures are inputs in the innovative process and are subject to reporting biases, patents and citations are innovative outputs, less prone to reporting biases, and hence are likely to be more important for firm value. Hall (1999) reviews a large literature 7 Other research examines the effects of innovation on mergers (Bena and Li (2013)) and CEO pay (Balkin, Markman and Gomez-Mejia (2000)). 8 Other studies find that firms in more religious counties experience low stock price crash risk (see Callen and Fang (2013)) and avoid unethical behavior (see Grullon, Kanatas and Weston (2010)). 6

9 on innovation and firm value and concludes that patent measurements contain information about firm value beyond that conveyed by R&D expenditures. But R&D expenditures are obviously an essential innovation input and we extensively analyze them as such. Another way we differ from Hilary and Hui is that while they use the level of a community s religiosity as a measure of its risk aversion, we use the ratio of Catholics to Protestants as a proxy for a community s gambling preferences. Gambling preference is a more appropriate metric for our analysis because it is more strongly related to innovative endeavors, which tend to have highly skewed payoff distributions. Our paper also draws extensively from the work of Kumar (2009), who finds that individual investors prefer lottery-like stocks, especially if they live in a Catholic county, and KPS (2011), who find that even institutional investors in predominantly Catholic communities tend to overweight lottery-type stocks in their portfolios. The rationale is that Catholic culture is much more tolerant of gambling than Protestant cultures, so investors who live in predominantly Catholic communities are influenced by the dominant culture and prefer stocks that offer gambling-like return distributions. We integrate the ideas from these two strands of work on the impact of religiosity on corporate decisions and the influence of gambling preferences on financial markets. We examine the implications of gambling preferences on corporate policies relating to innovation, an arena where such preferences can be key. Finally, we contribute to a new and growing area of research that examines the effect of local demographics on firm policies. Corporate decisions can be influenced by local demographics in several ways. First, given a local bias in investment decisions, firms have incentives to adopt policies that cater to the preferences of local investors. Second, for many firms, local residents constitute a significant proportion of their workforce and even the top management team (see, e.g., Yonker (2012)). The environment managers grew up in is likely to affect their corporate decisions. We build on the work of Becker, Ivković and Weisbenner (2011) on the effect of local elderly population on corporate dividend policies, and Cohen, Gurun and Malloy s (2012) work on the effect of local diaspora on firms foreign trade. The paper proceeds as follows. Section 2 describes our data and presents summary statistics. Section 3 presents our analysis of patents, citations and R&D expenditures, and conducts some robustness checks. Section 4 discusses a potential incentive for firms in gambling-tolerant areas to be innovative. Section 5 presents tests of secondary implications related to our main conjectures. Section 6 investigates whether firms in gambling-tolerant areas are more adept at translating 7

10 potential growth opportunities into firm value. Section 7 discusses some issues about the interpretation of our results and concludes. 2. Data and Summary Statistics Our sample period extends from 1980 to For our analysis, we combine data from several different sources as described below. Financials and headquarters locations Data on company financials and stock returns come from Compustat and CRSP databases. Data on firm headquarter location come from the CRSP-Compustat merged file. This file has numerous missing values in the county field, which we hand-collect from several other sources, such as the US Census Bureau s website that finds county name from city and state names, and via internet searches. We exclude companies in the financial (2-digit SIC codes 60 to 69) and public utility (2-digit SIC code 49) industries because they are highly regulated. Most of our analyses use an unbalanced panel of 30,878 firm-years that contain 3,926 unique firms during our 27-year sample period across 566 US counties. The numbers of observations vary somewhat across the tables based on data availability. Religiosity and demographics We obtain county-level religion data from the Churches and Church Membership files of the American Religion Data Archive (ARDA) website, which contains decennial data on county-level religion statistics on 133 Judeo-Christian bodies. For our analysis, we use the datasets for 1980, 1990, 2000, and Following the previous literature (Alesina and La Ferrara (2000), Hilary and Hui (2009), KPS (2011)), we linearly interpolate the religion data to obtain estimates for the intermediate years. We collect most county-level information from the U.S. Census Bureau, which has data on several demographic and economic characteristics (age, sex, race, education, income and the proportion of married couples) for each US county. Again, when data are available only on decennial 8

11 basis, we linearly interpolate to obtain values for the intermediate years. Following Hilary and Hui (2009) and KPS (2011), we use these variables as county-level control variables in the regressions. 9 Innovation: Patents and Citations Our main variables of interest include the number of patents applied for in a given year that were eventually granted, the number of citations per patent and technology-class adjusted citations. Our main source of data for patents and citations is the 2006 edition of the NBER patent database. However, patent applications are recorded in the NBER database only when they are granted. Therefore, in the NBER data, patent application numbers drastically decrease in the latter part of the sample as many of the patent applications were not approved by the end of To obtain more complete data on patent applications until 2006, we supplement the NBER data with another database put together by Kogan, Papanikolaou, Seru and Stoffman (2012) from Noah Stoffman s website. 10 The later database covers patent application data until 2010, so it largely solves the issue of a mechanical decline in the number of patent applications toward the end of our sample period. We merge the patent data with firms financial data from CRSP/Compustat and county-level religion and demographic data from ARDA and Census Bureau. Following the previous literature (see e.g., HLT (2012), and He and Tian (2013)), we construct three measures of a firm s innovation outcomes. The first measure, which represents the quantity of innovation, is the number of patents applied for in a given year that are eventually granted. The second measure is the number of citations per patent, a measure of innovation quality, which is calculated as the total number of citations received during our sample period on all patents filed (and eventually received) by a firm in a given year, scaled by the number of the patents filed (and eventually received) by the firm during the year. To correct for the truncation problem in citation counts (as older patents are more likely to receive more citations), we follow the previous literature and multiply the raw citation count by the weighting index provided by Hall, Jaffe and Trajtenberg (2001, 2005) to arrive at the adjusted citation count. Our third measure of innovation outcome is the sum of the number of technology class-adjusted citations received during our sample period on all patents filed (and eventually received) by a firm in a given year. The adjustment is done 9 We follow KPS (2011) who find that the education measure and income are highly correlated (correlation =.82) and do not include income in our regressions

12 by scaling each patent s citation count by the average citation count of all patents filed (and eventually granted) in the same technology class in a given year. This alternative measure of innovation quality takes into account the non-uniform propensity for patents in different technology classes to cite other patents. This measure is also adjusted for the truncation problem in citations. Due to the right-skewness of patent and citation proxies, in our main regression analyses, we follow the previous literature and use the natural logarithm of one plus the number of patents applied in a given year (LnPatent), log of one plus the number of truncation-adjusted citations per patent (LnCitePerPat), and the log of one plus the number of technology-class adjusted citations (LnTechAdjCites). Following prior work, we set these variables to zero for firm-years without data available in the NBER database. Measuring gambling preferences We follow KPS (2011) and consider the ratio of Catholic adherents to Protestant adherents (CPRatio) in a county as a proxy for the gambling attitudes of its residents. Previous literature finds that among religious groups, Catholics and Jews are more active participants in lotteries compared to Protestants and Mormons (see, e.g., Tec (1964), and Grichting (1986)). KPS (2011) find that states with higher concentrations of Catholics relative to Protestants are more likely to legalize state lotteries and adopt them earlier. After controlling for several demographic factors, they find that the concentration of Catholics relative to Protestants in a county positively and significantly predicts both the existence of state lottery and per capita lottery sales in a given year. Accordingly, our regressions employ the natural logarithm of one plus the Catholic-to-Protestant ratio (LnCPRatio) as the main explanatory variable of interest. This log measure parallels our patents and citations variables, and is less skewed than the raw CPRatio variable. 11 Following the previous literature on innovation, all our regressions include industry and year fixed effects and control for a number of firm characteristics that are related to innovation input and output of a firm. For instance, in regressions of patent count, citations per patent and technologyadjusted citations, we control for firm size (sales), past investment in R&D, profitability (ROA), capital intensity (the ratio of net property plant and equipment to the number of employees), leverage, growth opportunities (Tobin s Q), financial constrains (Kaplan and Zingales KZ Index), 11 As shown in robustness checks in section 3.4 below, our results are essentially unchanged if we replace LnCPRatio by CPRatio. The two variables are almost perfectly correlated (Pearson correlation = 0.96). 10

13 and industry concentration (Herfindahl-Hirschman index based on sales). Moreover, we also control for analyst following and institutional ownership which the recent literature (e.g., He and Tian (2013), and Aghion, Van Reenen and Zingales (2013)) finds to be important for innovation. All firm-specific control variables are lagged by one year in all the regressions. The only exception is that in the regressions of patents and citations as dependent variables, we include the average of the first and second lags of R&D to assets ratio as a control. Our choice of this lag structure is based on prior evidence that the average lead time between investment in R&D and patent applications is between one and two years. 12 In addition to the firm-specific variables, we also control for a number of contemporaneous county-level variables in all our regressions, such as county population, age structure, education (fraction of college graduates), fractions of female and minority population, fraction of married households and religious adherents per thousand. The Appendix defines all the variables used in our regressions. Table 1 presents summary statistics of all the variables used in this study. Panel A shows summary statistics of county-level religion and demographic variables for all US counties in our sample. Even though there are more than 3,000 counties in the US, only about 566 of them have or ever had a publicly-traded company located there during our sample period. The typical (median) county in our sample has (per 1,000 people) 161 Catholic, 260 Protestant and 496 total adherents. The typical county is a metropolitan area (rural-urban continuum <=3), and has a population of about 164 thousand and has about 24% college graduates among the population aged 25 years and higher. The mean (median) ratio of Catholics to Protestants is 1.12 (0.62) whereas the mean (median) of one plus the ratio of Catholics to Protestants (LnCPRatio) is 0.61 (0.48). LnCPRatio is obviously much less skewed than CPRatio. Both the mean and median of the age group indicator variable is about 8, which denotes 35 to 40 year olds. Compared to the typical US county (untabulated), the counties in our sample are more highly populated and more educated, and have higher per capita incomes. These counties are slightly less religious and have higher ratios of Catholic to Protestant populations. Our summary statistics are comparable to KPS (2011), though not the same because of slight differences in variable definitions and our larger sample size. Since our sample is not restricted to counties where institutional investors are located, it includes counties with lower populations, higher number of 12 Pakes and Griliches (1980) find this lag to be 1.6 years on average. Data from Rapoport (1971) and Wagner (1968) also suggest this lag to be between one and two years for most industries. Our results do not change if we replace the average with the first lag or the second lag. 11

14 religious adherents (both Catholic and Protestant) and fewer college graduates than in the KPS (2011) sample. Panel B of Table 1 shows summary statistics of our main variables of interest at the firmyear level, which we use later to infer the economic significance of our regression estimates. The mean and median values of LnCPRatio and CPRatio at the firm-year level differ from those at the county-level due to a non-uniform distribution of firm-years within counties. On average, a firm applied for 9.5 patents per year which were eventually granted and each granted patent received an average of 5.1 citations. Similarly, a firm received an average of 7.9 citations adjusted for technologyclass. A firm spent an average of 4.4% of its assets on R&D. Patent applications, citations per patent, and technology class-adjusted citations have median values of zero. The average annual idiosyncratic standard deviation of stock returns is and idiosyncratic skewness (i.e., the skewness of the residuals from a regression of daily stock returns on market returns and squared market returns) is Following KPS (2011) s definition, we identify about 27.2% of stocks as lottery stocks (i.e., stocks with above-median idiosyncratic volatility and above-median idiosyncratic skewness). The distributions of other variables are comparable to those in the prior literature. 3. Analysis and Discussion We start by presenting simple correlations among our key variables of interest by way of a smell-test. We then proceed to tests of our main conjectures, followed by robustness checks. 3.1 Correlations Table 2 shows Pearson correlation coefficients among our main dependent variables and some key explanatory variables, where all non-italicized coefficients are statistically significant at the 1% level. These correlations are consistent with our story and with the more rigorous analysis that follows. For instance, LnCPRatio is positively correlated with LnPatent, LnCitePerPat and LnTechAdjCites consistent with our conjecture that firms in counties with more Catholics relative to Protestants innovate more by applying for more patents that are eventually granted and receiving more citations on these patents. Moreover, LnCPRatio is highly positively correlated with (R&D/Assets) 12 consistent with the idea that firms located in more gambling-tolerant counties invest larger fractions of their assets in R&D. Not surprisingly, R&D expenditure is positively correlated with LnPatent, LnCitePerPat and LnTechAdjCites, suggesting that a higher level of R&D investment is 12

15 an essential requirement for innovation. Furthermore, R&D expenditure is positively correlated with both idiosyncratic volatility and idiosyncratic skewness of a stock, consistent with our conjecture that R&D expenditure contributes positively to firm-specific volatility and firm-specific skewness, both of which are lottery factors for a stock. Finally, cash holding is positively correlated with LnCPRatio and (R&D/Assets) 12, consistent with our conjecture that firms in high CPRatio counties hold more cash partly to finance R&D. All of the county-level variables are highly correlated with each other and with our main explanatory variable of interest, LnCPRatio. Interestingly, the correlations suggest that Catholics are more concentrated in more religious counties. Counties with higher concentrations of Catholics relative to Protestants tend to be more populated, have older populations and be more urban. They also have higher proportions of college graduates and females, and lower proportions of minority populations and households with married couples. 3.2 Patents, Citations & R&D To formally examine how local gambling preferences affect firms innovation outcomes, we estimate the following regression model: LnPatent i,j,k,t or LnCitePerPat i,j,k,t or LnTechAdjCites i,j,k,t = α + βlncpratio k,t + γfirmlevelcontrols i,t + δcountylevelcontrols k,t + Year t + Industry j + ε i,j,k,t where i, j, k and t are indices of firm, industry (2-digit SIC code), county and year. The dependent variables are innovation outcomes. LnPatent is the natural logarithm of one plus the number of patents applied for and eventually granted by firm i in year t. Similarly, LnCitePerPat is the natural logarithm of one plus adjusted citations per patent, and LnTechAdjCites is the natural logarithm of technology class-adjusted citations. Our main explanatory variable of interest LnCPRatio, is the natural logarithm of one plus the Catholic to Protestant ratio in year t in county k of the firm s headquarters. FirmLevelControls is a vector that includes several time-varying firm characteristics found by the prior literature to affect a firm s innovation process (see, e.g., He and Tian (2013), and HLT (2012)) which are discussed in section 2. CountyLevelControls includes county-level variables that have been identified by prior studies to affect individual and firm decisions in the county, namely gender, racial and age compositions, population, education, household characteristics, and overall 13

16 religiosity (see, e.g., KPS (2011) and Schneider and Spalt (2013)). All of our regression models have standard errors robust to heteroscedasticity and within-firm correlation. 13 Panel A of Table 3 presents our first set of regressions of patents and citations using the full sample. In untabulated results, for each of the three dependent variables, namely patent count, citations per patent and technology-adjusted citations, we begin by estimating a parsimonious model that only includes our main explanatory variable, LnCPRatio, with industry and year fixed effects. The regression of LnPatent obtains a coefficient of on LnCPRatio, which is statistically significant at the 5% level. This finding suggests a positive relation between the ratio of Catholics to Protestants in a country and the number of eventually granted patents filed by a firm after controlling for any secular trend in innovation, industry effects and correcting the standard errors for multiple observations for a firm. Next, as shown in column 1 of Panel A, we add a number of firm-specific control variables found by previous studies to affect a firm s innovative output. Most of these firm-specific variables turn out significant in predicting LnPatent, but the coefficient on LnCPRatio remains essentially unchanged and becomes statistically significant at the 1% level. This result suggests that the gambling attitude of a community predicts the patenting activities of a firm beyond that explained by firm-specific factors. Finally, in column 2, we also add the county-specific variables found to be important by previous studies. The estimated coefficient on LnCPRatio in this model, which includes the full set of controls, increases from the estimate in the model without county-level control variables, with a noticeable improvement in statistical significance. This result indicates that a firm located in a county with a higher concentration of Catholics tends to apply for (and receives) a larger number of patents, and this relation stays intact after controlling for firm characteristics and county-level demographic variables. In terms of economic significance, the point estimate of suggests that moving from a county at the 25 th percentile to the 75 th percentile of LnCPRatio increases LnPatent by (= 0.144*( ), see Table 1, Panel B). This represents an increase of about 17% compared to the unconditional mean value of for LnPatent. Next, we estimate a similar set of regressions of the effect of LnCPRatio on citations per patent (LnCitePerPat). Similar to the patent counts regression, the untabulated parsimonious model obtains a positive and statistically significant coefficient (at the 5% level) of on LnCPRatio in predicting LnCitePerPat, consistent with our hypothesis. As shown in column 3 of Table 3 Panel A, 13 In robustness checks, we cluster the standard errors by the county of a firm s headquarters locations and find similar results. 14

17 this estimate slightly decreases in magnitude, but turns significant at the 1% level once we add firm specific-variables control variables to the model, underscoring the importance of firm-specific factors for innovation. In column 4, when we also control for county-specific variables, the estimated coefficient on LnCPRatio increases and now has a larger t-statistic. The point estimate of on LnCPRatio obtained from the model with the full set of controls suggests that going from a county at the 25 th percentile to one at the 75 th percentile of LnCPRatio increases LnCitePerPat by 0.108, which is an increase of 13% over its unconditional mean of Finally, we examine the effect of LnCPRatio on technology class-adjusted citations (LnTechAdjCites). Consistent with our conjecture, in the untabulated parsimonious model with only LnCPRatio, year and industry dummies as explanatory variables, LnCPRatio obtains a coefficient estimate of with a t-statistic of When we add the firm-specific variables to the regression, as shown in column 5, this estimate slightly increases in magnitude and becomes statistically significant at the 1% level. As shown in column 6, this estimate further increases in magnitude with even stronger statistical significance in the model with the full set of controls. In economic terms, the coefficient estimate of obtained from this full model suggests that moving from a county at the 25 th percentile to the 75 th percentile of LnCPRatio increases LnTechAdjCites by 0.121, which is an increase of about 21% over its unconditional mean of Throughout this analysis, most control variables obtain their expected signs. For instance, larger and older firms and firms with higher investment opportunities receive more patents and citations. Firms with higher levels of financial leverage obtain fewer patents and citations possibly because the need for leverage represents financing constraints given that innovation, as with other growth opportunities, is difficult to finance with debt (see Myers (1977) and Myers and Majluf (1984)). Not surprisingly, past R&D expenditure contributes to patents and citations in a very significant and positive way. Among the county-level variables, the proportion of college graduates (College Grads) in the county turns out to be the most significant, suggesting that higher education facilitates innovation. We now proceed to test our next conjecture that the effect of local gambling tendency on innovation should be larger in firms which are in innovative industries. We follow HLT (2012) and perform our analysis on samples split into innovative and non-innovative industries. We consider an 15

18 industry, defined by its 4-digit SIC code for this purpose, to be innovative if its citations per patent exceed the median value across all industries in a given year. 14 Panel B of Table 3 shows the estimates obtained from the regression with the full set of controls on samples of innovative and non-innovative industries separately. For brevity, we only show the estimated coefficients on our main explanatory variables of interest, LnCPRatio and R&D/Assets. Among innovative industries, the regression of (eventually granted) patent applications obtains a coefficient of on LnCPRatio which is statistically significant at the 1% level. In the sample of non-innovative industries, LnCPRatio obtains a much smaller coefficient of 0.094, which is significant at the 5% level. A Chow test rejects the null hypothesis of equality of these two coefficients (p= 0.052). These results support our hypothesis and suggest that the effect of LnCPRatio on a firm s patent output is about twice as big in innovative industries as in noninnovative industries. The greater influence of gambling attitudes in innovative industries is even more pronounced on citations per patent as an innovative output. As shown in columns 3 and 4 of Panel B, the coefficient estimate of LnCPRatio is about ten times as large in innovative industries (and statistically significant at the 1% level) as it is in non-innovative industries. Again, these coefficients on LnCPRatio in innovative and non-innovative industries are statistically different from each other (p <.01). The conclusion is similar with LnTechAdjCites as an innovative output. Among innovative industries, LnCPRatio obtains a statistically significant coefficient of 0.222, which is almost five times as large as that in non-innovative industries. Apart from being substantially larger in economic terms, the coefficient estimate on LnCPRatio in innovative industries is also statistically different from that in non-innovative industries (p <.01). Also notable is the differential contributions of R&D expenditures in generating patents and citations among firms in innovative and non-innovative industries. In patent count regressions, the coefficients of and on R&D/Assets among the innovative and the non-innovative firms, respectively, suggest that the impact of R&D expenditures in generating patents is about 31% larger in innovative industries than in non-innovative industries. Calculated similarly, the effect of R&D investment is 58% (143%) higher for citations per patent (technology-adjusted citations) in innovative industries than in non-innovative industries. The importance of R&D expenditures in 14 We define industry using 4-digit SIC codes here because they provide a sharper contrast in innovative activities across industries than 2-digit SIC codes. 16

19 generating patents and citations and its greater influence in innovative industries serve as the basis for our next analysis. 3.3 Channel for Innovation: R&D Expenditures Next, we attempt to identify the channel through which firms in higher CPRatio counties achieve more innovation. Given the results in Table 3 suggesting that R&D expenditures are an essential input for generating patents and citations, we examine whether firms in counties with higher CPRatios invest more in R&D. Table 4 shows the estimated coefficients from OLS regressions of R&D expenditures. All standard errors are corrected for heteroscedasticity and within-firm correlations. Panel A shows the results of the analysis on the full sample. As before, we start with an untabulated parsimonious model where the only explanatory variables included are our main variable of interest, LnCPRatio, along with year and industry dummies. Consistent with our conjecture, this model yields a positive and significant coefficient (at the 1% level) of on LnCPRatio, suggesting that gambling preferences of a county s residents positively predict R&D expenditures of firms headquartered in the county. Next, as shown in column 1, we control for several firm-specific variables measuring firm size, leverage, investment opportunities, past performance and capital intensity that are found by the previous literature to predict a firm s R&D expenditures (see, e.g., Coles, Daniel, and Naveen (2006), and HLT (2012)). 15 In addition, motivated by the recent literature on innovation (see, e.g., He and Tian (2013), and Aghion, Van Reenen and Zingales (2013)), we add controls for institutional ownership and analyst coverage. As shown in column 1, when we introduce firm-specific variables to the regression, the point estimate on LnCPRatio decreases, but remains statistically significant at the 1% level. This finding suggests that the relation between LnCPRatio is indeed influenced by firm characteristics, but the marginal influence of gambling attitude still remains significant. Finally, in column 2 we add several county-specific variables that are also likely to influence innovation decisions of firms located there. The estimated coefficient on LnCPRatio in this regression with the full set of control variables is slightly larger than that obtained from the model without county-level controls, and is statistically significant at the 1% level. So, even after controlling for firm-specific as well as country-level variables, the local gambling culture significantly predicts R&D expenditures of a firm. In economic terms, the point estimate of on LnCPRatio suggests that moving from a 15 We do not have CEO compensation-related variables, since our sample is not limited to S&P 1500 firms for which these variables are available in Execucomp after

20 county in the 25 th percentile LnCPRatio to a county in the 75 th percentile increases R&D/Assets by.0063 (=0.007*( ). This estimate is economically significant because it represents a 14% increase over the unconditional mean of for R&D/Assets. Overall, the evidence leads to the conclusion that higher levels of innovation outputs of firms in high CPRatio areas are at least partially driven by higher levels of R&D expenditures made by these firms. In Panel B, we repeat this analysis in sub-samples partitioned by industry innovativeness. We employ the model with the full set of control variables, but for brevity we only tabulate the coefficient on LnCPRatio. As before, we expect that the influence of gambling attitudes on R&D expenditures should be higher in firms in innovative industries because managers likely have a greater opportunity to engage in innovative endeavors in these firms. Consistent with this conjecture, Panel B of Table 4 shows that for firms in innovative industries, LnCPRatio has a coefficient estimate of which is statistically significant at the 1% level. In non-innovative industries, the coefficient estimate on LnCPRatio is a much lower 0.004, significant at the 10% level. Apart from being almost three times as large in economic terms, this coefficient in the innovative industries subsample is also statistically different from that in the subsample of non-innovative industries (p=.096). This finding corroborates our analysis in Panel B of Table 3, where we find that the influence of LnCPRatio on innovation output is significantly higher in firms in innovative industries. Control variables take the expected signs. Firms with higher investment opportunities (Tobin Q), low sales and low past stock performance spend more on R&D. Highly levered firms, and firms with higher institutional ownership spend less on R&D, both consistent with the findings of HLT (2012). Interestingly, we find a positive relation between analyst coverage and R&D expenditures, which is consistent with He and Tian (2013) s results in their model without firm fixed effects. Unlike in the previous literature, the coefficient of the excess cash variable is statistically insignificant, which we find to be the result of including additional control variables, such as analyst coverage and institutional ownership. Among county-level variables, the percentage of college graduates in a county positively predicts R&D expenditures of local firms, supporting the view that higher education fosters innovation. Consistent with Hilary and Hui (2009), the natural logarithm of the number of religious adherents per 1000 people is negatively related to R&D expenditures. 18