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

21 We find that firms in more gambling-tolerant counties spend more on R&D, which in turn allows them to generate higher innovation output as measured by patents and citations, especially in innovative industries. Taken together, these findings imply that gambling preferences of local residents lead firms to invest more in R&D which generates more patents and citations. Thus, R&D expenditure appears to be an important channel though which firms in gambling-tolerant counties achieve higher innovation. This conclusion leads us to probe deeper into the incentives of these firms to invest more in R&D, which we do in section 4 below. But before that, we examine the robustness of our results so far. 3.4 Robustness We examine the robustness of our main results in section and provide event study evidence from relocations of corporate headquarters in section Robustness of Main Results In this section, we present and discuss a number of robustness tests of our main results on patents, citations and R&D expenditures. Broadly, we examine if our results are robust to other plausible specifications, and if they are driven by certain industries or locations. Table 5 summarizes these results and reports the coefficient estimates and t-statistics on our main explanatory variable of interest, followed by the number of observations in different specifications. We separately report the results of regressions on our full sample and on the sub-sample of innovative industries. The four sets of six columns each present the results of regressions of patent counts, citations per patent, technology-class adjusted citations and R&D expenditures, respectively. We start by examining an alternate measure of local gambling preferences. We gather information on whether each of six types of gambling (charitable, pari-mutuel, lottery, commercial, Indian and racetrack) is legal in each of the 50 states 16. We omit charitable gambling, since its purpose is fund-raising for charities rather than gambling for personal gain. Using the remaining five types of gambling, we construct an index of gambling culture in a state by counting the number of types of gambling allowed in the state. This index takes a value of 0 to 5 for a state. This is a timeinvariant measure of a state s gambling culture. This gambling index has a highly significant Pearson 16 See Wikipedia article Gambling in the United States ( 19

22 correlation coefficient of 0.43 with LnCPRatio, consistent with our interpretation of LnCPRatio as a measure of local gambling preferences. We then replace LnCPRatio by this gambling index in the regressions of Tables 3 and 4. Row (1) of Table 5 shows that this index positively and significantly explains LnPatent, LnTechAdjCites and R&D/Assets; its coefficient is positive but statistically insignificant in regressions of LnCitePerPat. Moreover, these effects are consistently stronger in the subsample of innovative industries. We don t use this gambling index in our main analysis because it has limited variation (0 to 5), is available only at the state-level and is time-invariant, while CPRatio is a continuous variable, is measured at the finer country-level and varies over time. Next, in row (2), our conclusions do not change if we use CPRatio instead of LnCPRatio. Consistent with the results of our baseline specification, CPRatio positively and significantly predicts the number of patents, citations per patent, adjusted citations and R&D expenditures. In each case, the coefficient estimate is larger for the sub-sample of firms in innovative industries. There are many firm-years with zero patents, zero citations and zero R&D expenditures. So, one concern is whether our results are driven by a jump from zero patents (or citations or R&D) to a positive number. We deal with this issue in two ways. First, for each of the four dependent variables, we estimate a Tobit model, which explicitly accounts for a jump in the distribution at zero. In row (3), these results are qualitatively similar to those under the OLS and are statistically significant. Second, we repeat the regression analysis by excluding all firm-years with zero patents. By doing this we essentially estimate the influence of LnCPRatio among firm-years with at least one patent, thus eliminating the impact of the jump from zero to positive values. In row (4), despite losing about two-thirds of our observations, our results remain quantitatively similar and statistically significant. We next attempt to control for unobserved firm heterogeneity. Ideally, this is done by including firm fixed effects in the regressions. However, our data does not allow us to control for firm fixed effects because the religious composition of a county remains fairly stable over time. Consequently our main explanatory variable of interest, LnCPRatio, does not have enough timeseries variation within a firm. So, we use the random effects (RE) model which, although it relies on stronger assumptions about the error correlation structure, can estimate the effects of time invariant explanatory variables, while controlling for unobserved heterogeneity. We still correct the standard 20

23 errors for repeated observations and heteroscedasticity. In row (5), the estimated coefficient on LnCPRatio in the patent count regression under the RE model turns insignificant in the full sample, but stays significant in the sample of innovative industries. The estimates from the regressions of citations per patent, technology-adjusted citations, and R&D expenditures stay significant both in the full sample and the innovative industry sub-sample. The point estimates for the innovative industry subsample are larger than for the full sample in three of the regressions; they are equal in the R&D expenditure regression. Next, to mitigate the issue of causality and omitted variables bias, we follow Hilary and Hui (2009) and KPS (2011) and estimate a two-stage least squares (2SLS) model by using the three-year lagged value of LnCPRatio as an instrument for LnCPRatio. In row (6), the results from this model are very similar to those from OLS. We next perform several tests to check whether our results are driven by a large number of patents generated by a few firms in a few locations or industries. To this end, we identify the five industries (defined by 2-digit SIC codes) that generate the largest average number of patents during our sample period, namely transportation equipment (SIC2 = 37), chemicals and allied products (SIC2 = 28), petroleum refining and related industries (SIC2 = 29), paper and allied products (SIC2 = 26), and electronic and other electrical equipment manufacturers (SIC2 = 36). Despite a loss of more than 25 percent of our sample when we drop them, the results shown in row (7) remain significant and are in agreement with our main findings. Our results (untabulated) are similar when we omit high-tech industries, as defined by Loughran and Ritter (2004). Dropping these industries from the sample also addresses a potential endogeneity in our results that gambling-tolerant investors may invest more in firms in high-tech industries, which, on average, tend to invest more in R&D and obtain more patents and citations. In the next robustness test, we exclude the five U.S. states with the highest concentration of Catholics relative to Protestants during our sample. These states comprise of Massachusetts, New Jersey, New York, New Hampshire and Rhode Island. In row (8), our results from each of the four regressions for patents, citations and R&D remain essentially unchanged. This finding implies that the observed relations are not driven by firms in a few states with a large number of Catholic adherents. The same is true in row (9), when we exclude the five states with the lowest CPRatio (Tennessee, Arkansas, North Carolina, South Carolina, and Alabama). 21

24 Finally, to address the possibility that the decisions of different firms in a county might be correlated, we repeat our regression analysis by clustering the standard errors at the county-level instead of firm-level. This analysis poses a big hurdle to statistical significance because clustering at county-level effectively collapses the sample size to 516 independent county-level observations, thus significantly increasing the standard errors. Despite this, while the magnitudes of the t-statistics drop somewhat in row (10), most of our results remain statistically significant as in our baseline tests Evidence from Corporate Headquarters Relocations We next examine whether firms that move their corporate headquarters to a higher CPRatio area spend more on R&D following the move. We use the SEC Analytics Suite to identify changes in the state of our sample firms headquarters reported in their 10K filings for each year from 1994 to In cases where this database reports a move by a firm in one year followed by a move back the following year, we used Factiva news search and other Internet sources to verify that the firm s headquarters actually moved to a new location. We then average the CPRatios across all the counties in a state to create a state-level CPRatio each year. For each of the 194 firms in our sample that moved their headquarters state over this period, we calculate (LnCPRatioState) and (R&D/Assets), which measure the change in these variables for the year after the move compared to the year before the move. We exclude the year the firm appears in a new state because the relocation takes place at different times during the year for different firms and to allow time for a change in corporate policies. We then estimate the following regression, which has one observation for each mover firm, and use standard errors corrected for heteroskedasticity: (R&D/Assets) = b o + b 1 (LnCPRatioState)+ 2-digit SIC industry dummies + year dummies, Despite a small sample size, the estimated coefficient of (LnCPRatioState) in this regression is significantly positive, with a t-statistic of This finding suggests that firms increase their R&D spending in the year following a move to a state with a higher preference for gambling, and support our main findings based on panel regressions. We do not tabulate this result for brevity. 4. Incentive for Higher R&D Our results so far are consistent with the hypothesis that local gambling attitudes motivate firms to spend more on R&D, which consequently leads to higher levels of innovation in terms of patents and citations. In this section, we discuss one potential incentive that leads firms in higher 22

25 CPRatio areas to spend more on R&D. Specifically, we examine whether R&D investment makes a stock more lottery-like, which prior studies have found to be a desirable feature for individual and institutional investors inclined to gamble. Kumar (2009) argues that stocks with high idiosyncratic volatility, high idiosyncratic skewness and low price tend to have lottery-like characteristics. Investors with a preference for lotteries find such stocks attractive, even if they offer negative expected returns, because these investors expect the extreme positive return events of the past to be repeated in the future. Kumar (2009) finds that individual investors in Catholic- (Protestant-) dominated areas favor lottery-like stocks more (less) than the rest of the sample. KPS (2011) find that even institutional investors, who generally avoid highly risky stocks, hold higher proportions of lottery-type stocks if they are located in counties with higher concentrations of Catholics. We begin our analysis by identifying stocks that are more likely to be viewed as lottery-like. Motivated by Kumar (2009), we classify a stock as lottery-type if its returns exhibit above-median idiosyncratic volatility and above-median idiosyncratic skewness in a given year. 17 Since there is no reason to expect R&D expenditures to influence stock price levels per se, we follow KPS (2011) and do not consider price per share while constructing the lottery stock set. 18 But we do control for stock price levels in the regressions. Idiosyncratic volatility is the standard deviation of the residuals obtained by regressing daily returns on a stock on Fama and French (1993) and Carhart s (1997) four factors over a year. Idiosyncratic skewness is calculated as the skewness of the residuals obtained by regressing daily returns on a stock on excess market return and excess market return squared (see, Harvey and Siddique (2000), and Kumar (2009)). We then estimate a logit model of a dummy variable indicating whether a stock is lottery-type, using lagged R&D/Assets as the main explanatory variable. Following Kumar (2009, Table II), we control for firm-specific variables (such as firm size, firm age and market-to-book ratio), variables related to asset pricing (such as market beta, SMB and HML) and market microstructure variables (turnover and illiquidity). We also include all county-level control variables along with LnCPRatio and LnPatent. In addition, the models include year fixed effects and either industry or firm fixed effects. Table 6 presents the results of various analyses of stocks lottery-like features. For brevity, county-level control variables are not tabulated. Column 1 shows the results of a logit regression that 17 High volatility can lead investors to believe that extremely high past returns are likely to be repeated. 18 Our conclusions are not affected by this choice. 23

26 predicts the probability of a stock to be lottery-like. In this regression, the coefficient on lagged R&D/Assets is positive and highly significant after controlling for county demographics and a host of variables explaining the lottery features of a stock. This result suggests that R&D expenditure is an important factor that gives lottery-like features to a stock by increasing its idiosyncratic volatility and idiosyncratic skewness. A potential concern with this analysis is that the positive relation between R&D expenditure, idiosyncratic volatility and idiosyncratic skewness may be driven by firm inertia, as stock return distribution and firm policies both tend to persist over time. Therefore, to better capture the dynamic role of R&D expenditures in shaping the distributions of stock returns, we re-estimate logit model in column 2, after controlling for firm fixed effects. This fixed effect logit model examines the determinants of within-firm variability in the lottery-likeness of a stock. By definition, this regression only includes those firms where the stock switched from being non-lottery type to being lottery type, or vice-versa. The model yields a positive coefficient of on lagged R&D/Assets which is significant at the 1% level. This finding suggests that R&D expenditures significantly and positively predict the switching of a non-lottery type stock to lottery-type or vice-versa, and confirms that the relation between R&D expenditure and lottery-likeness of a stock is dynamic in nature. Next, we analyze the effect of R&D expenditures on each of the two lottery factors, a stock s idiosyncratic volatility and idiosyncratic skewness, separately. Column 3 shows the results of a regression of idiosyncratic volatility where lagged R&D/Assets is the main explanatory variable of interest. The control variables are the same as before. As expected, the coefficient estimate on R&D/Assets is positive and statistically significant at a very high level even after controlling for firmspecific and county-level variables. This result suggests that R&D expenditures significantly increase the idiosyncratic volatility of a stock to make them more lottery-like. To better capture the dynamics of R&D expenditures and idiosyncratic volatility, we repeat this regression after controlling for firm fixed effects. Since our main explanatory variable of interest here is R&D expenditure, which, unlike LnCPRatio, has meaningful time-series variation, the firm fixed-effects model is appropriate for this analysis. As shown in column 4, while both the point estimate and t-value decrease, R&D expenditure still obtains a positive coefficient, which is significant at the 1% level in predicting idiosyncratic volatility. This finding confirms the dynamic role of R&D expenditure in influencing idiosyncratic volatility. 24

27 Columns 5 and 6 of Table 6 show the results of regressions of idiosyncratic skewness. Column 5 shows the estimate of the regression model without firm fixed effects, which yields a positive and highly significant coefficient on lagged R&D/Assets in explaining idiosyncratic skewness. Column 6 presents the same regression with firm fixed effects. While the point estimate of the coefficient on R&D/Assets is slightly smaller than before, it is still significant at the 1% level, suggesting a dynamic relation between R&D expenditures and idiosyncratic skewness. Following Kumar (2009), we also experiment with adding the cotemporaneous idiosyncratic volatility measure as an additional explanatory variable in the idiosyncratic skewness regressions because volatility and skewness might be simultaneously determined. In untabulated results, we find that while the signs on some of our control variables change, the coefficient on R&D/Assets remains positive and significant at the 1% level in both models, either with or without firm fixed effects. This finding does not support the idea that the observed increase in a stock s idiosyncratic skewness because of higher R&D expenditures is an artifact of a simultaneous increase in volatility. The signs on control variables are largely consistent with Kumar (2009) in explaining a stock s lottery-likeness. The results presented in columns 1 and 2 suggest that stocks with lottery-like features belong mostly to younger firms with significantly smaller market capitalization, lower institutional ownership, higher illiquidity (i.e., lower liquidity) and higher trading volume (i.e., turnover). Moreover, lottery stocks earn lower returns and are less likely to be dividend payers. The negative point estimate on price per share suggests that stocks classified as lottery stocks based on idiosyncratic volatility and idiosyncratic skewness also tend to have lower prices, which is the third criterion Kumar (2009) uses to identify a lottery stock. Some variables take the opposite signs while predicting idiosyncratic volatility and idiosyncratic skewness. For instance, the market to book ratio (Tobin s Q) has a positive effect on volatility but a negative effect on skewness. Similarly, firm age seems to have a negative effect on volatility but a positive effect on skewness. There is also some evidence that the number of eventually granted patent applications is positively related to volatility, but negatively to skewness. While the relation between R&D expenditures and firm risk has been documented by prior research (see, e.g., Chan, Lakonishok, and Sougiannis (2001)), it was not in the context of corporate innovation, so it does not include the set of control variables relevant here. But, to our knowledge, we are the first to document the positive relation between R&D expenditures and idiosyncratic 25

28 skewness. This positive relation makes sense because R&D investments are high-risk endeavors that have positively skewed payoff distributions. While many of these projects fail, a few can yield ground-breaking inventions with very large returns. To summarize this section, we identify one motivation for firms in high CPRatio areas to invest more in R&D, which eventually results in more patents and more citations. We find that R&D expenditures make the stock return distribution more lottery-like by increasing both idiosyncratic risk and idiosyncratic skewness, features desired by individuals who prefer gambling. So it appears that part of the reason firms invest in R&D is to introduce lottery-like features in their stock to attract local investors with a preference for gambling. 5. Supporting evidence To improve identification, in this section we examine four subsidiary implications of our main conjectures. We basically present four pieces of evidence that support the notion that higher R&D expenditures made by firms located in gambling-tolerant areas are partly due to local residents preference for gambling. We find that: (1) The influence of local gambling preferences on R&D expenditures is driven by firms for which local investors are more important, (2) The gambling preferences of both local managers and local investors appear to motivate firms to invest more in innovation activity, (3) Firms in high CPRatio areas tend to be relatively inefficient in generating patents from R&D dollars, and (4) Firms in higher CPRatio areas hold more cash, especially to invest in R&D. 5.1 The role of local investors We find in section 3.3 that firms located in counties with high CPRatios invest more in R&D. We argue that firms do so partly to cater to the lottery preferences of local investors, and our findings in section 4 support this argument. This argument also implies that the relation between local gambling preferences and R&D expenditures should be more pronounced in firms for which local investors are economically more important. To test this conjecture, we examine four indicators of such firms. First, Becker, Ivković and Weisbenner (2011) argue that small firms are more likely to be reliant on the local investor base. We create an indicator variable, SmallFirm, which equals one if the market value of the firm is less than the median market value of all firms in a given year, and zero otherwise. Second, we identify firms located in counties with fewer investment opportunities 26

29 relative to local investment demand. Following Hong, Kubik and Stein (2008), we create a variable BE/PI which equals the aggregate book equity of all the public companies headquartered in a county in a given year divided by the aggregate personal income of all the residents in the county during the year. 19 The idea is that in the presence of local bias, a firm is likely to get a larger share of local investment if it has to compete with fewer other local firms, which they call the only-game-in-town effect. This effect makes local investors economically more important to firms. Specifically, we create a variable, Low BE/PI, which equals one if the BE/PI associated with a firm is below the median BE/PI of all sample firms in a given year, and zero otherwise. Furthermore, Hong, Kubik and Stein find that the importance of BE/PI is greater in firms which are less visible outside their home regions and consider the number of shareholders in a firm as a measure of visibility. Accordingly, we create a third variable, Low BE/PI Visible, an indicator variable that equals one if the firm has below-median number of shareholders and below median BE/PI in a given year, and zero otherwise. Finally, given that individual investors are more likely to be prone to behavioral biases such as gambling preferences (see, e.g., Bailey, Kumar and Ng (2011)) than institutional investors, local gambling preferences should matter more for firms with larger fractions of individual investors. Accordingly, we create a fourth variable, High Indiv Invest, an indicator variable that equals one if a firm has above-median fraction of individual investors (i.e. below-median fraction of institutional investors) in a year, and zero otherwise. Based on these four indicators of the importance of local investors to a firm, we examine whether the influence of local gambling preferences on innovative endeavors (R&D expenditures) is driven by firms for which local investors are likely to be more important. Accordingly, we reestimate our regression of R&D expenditures as in column 2 of Table 4, Panel A, after adding an indicator variable for the importance of local investors (i.e., SmallFirm or Low BE/PI or Low BE/PI Visible or High Indiv Invest) and its interaction with LnCPRatio. A positive and significant coefficient on these interaction terms would indicate that R&D expenditure is more sensitive to local investors preference for gambling if the local investor base is more important for the firm. Table 7 presents these results. Column 1 shows the results of the regression using the first indicator variable SmallFirm and its interaction with LnCPRatio. The results support our hypothesis that the influence of local gambling attitudes on innovative input, i.e., R&D spending, should be more pronounced in smaller firms. The point estimate of this interaction term, which measures the 19 We proxy a county s per capita income by its state s per capita income and multiply it by the county s population. 27

30 effect of gambling preferences in small firms, is and is statistically significant at the 1% level, whereas it is small and insignificant among large firms with a point estimate of (t-value = 0.98). The next indicator variable, Low BE/PI and its interaction with LnCPRatio are introduced in column 2. Once again, we find a more significant effect of LnCPRatio on R&D spending among firms in Low BE/PI counties with a point estimate of (t = 1.84), compared to firms in high BE/PI counties which have a point estimate of (t = 1.25). Next, we consider both low BE/PI and low visibility together (Low BE/PI Visible), which captures the importance of local investors more precisely. As shown in column 3 of Table 7, the effect of gambling preferences on R&D activities is more pronounced in firms which are located in low BE/PI counties and are also less visible. The point estimate on the interaction of Low BE/PI Visible and LnCPRatio is which is significant at the 5% level. On the other hand, the point estimate on the rest of the sample is a much smaller and is barely significant at the 10% level. Finally, column 4 shows the effect of a firm s reliance on individual investors on the relation between local gambling preferences and R&D. The interaction of High Indiv Invest and LnCPRatio obtains a positive coefficient of 0.01 which is significant at the 1% level. This result suggests that firms invest more in R&D to cater to local gambling preferences especially when larger fractions of their investors are individuals, who are more susceptible to behavioral biases. To recap, we find that the effect of gambling preferences on innovative input is driven by firms for which the local investor base is more important. In particular, smaller firms, firms in counties with fewer investment opportunities compared to demand (especially if they are less visible), and firms with more individual investors tend to be more influenced by local investors gambling preferences. Hence, they spend more on R&D. 5.2 The role of local managers We find in section 3.3 that firms located in high CPRatio areas invest more in R&D. Our findings in section 4 and 5.1 suggest that firms do so partly to cater to the lottery preferences of local investors. We next examine whether managers of firms in high CPRatio areas also display a preference for lotteries and if so, whether such preferences play a role in firms decisions to invest more in R&D. We measure managers lottery preferences based on their stock option holdings in their firms. Consistent with the fact that stock options have lottery-type characteristics, such as a 28

31 positively skewed payoff distribution, KPS (2011) find that broad-based employee stock option plans are more popular with firms in gambling-tolerant (i.e., high CPRatio) areas. For the subsample of our firm-years on Execucomp, 20 we start by examining whether stock option holdings of CEOs of firms in high CPRatio areas generate higher pay-performance sensitivity (PPS option ) and higher vega (Vega option ). We follow Core and Guay s (2002) approach to compute CEOs PPS option and Vega option, which measure the increase in the value of a CEO s option holdings in response to a 1% increase in stock price and a 1% increase in stock volatility, respectively. We then estimate regressions of PPS option and Vega option, where the main explanatory variable of interest is LnCPRatio. The control variables are based on Mobbs (2013). In addition, we include all the control variables in our R&D regression in Table 4, column (2) because a firm s R&D intensity and its CEO s option incentives are likely to share common determinants. We find that LnCPRatio significantly and positively predicts both PPS and Vega of CEO s option holdings, implying that CEOs of firms in more gambling-tolerant areas derive significant incentives from their stock option holdings. We then add either PPS or Vega or both as explanatory variables in our R&D regression in Table 4, column (2). The coefficient estimates of both variables are significantly positive, suggesting that incentives from CEO s option holdings positively predict firms investments in innovation. But even after controlling for this effect, the coefficient estimate of LnCPRatio remains significantly positive, and in fact becomes slightly larger (.009 instead of.007). These findings suggest that the gambling preferences of both local managers and local investors lead firms to invest more in innovation. These results are not tabulated for brevity, but are available upon request. 5.3 Gambling motivation and R&D efficiency If R&D expenditures made by firms in counties with higher CPRatios are partly driven by gambling motives, what can we conjecture about the relation between innovation input, i.e., R&D expenditures, and innovation output, i.e., patents and citations generated by these firms? Suppose that a part of R&D expenditures made by firms in gambling-tolerant areas is not based on reasonable expectations of success, but instead are risky long shots to cater to local investors lottery demands. Since a few of these gambles would turn out to be viable patents, but most would not, we should find that firms in high CPRatio areas have a lower marginal productivity of R&D 20 Our Execucomp subsample, which consists of S&P 1500 firms over the years that have all the data we need, is about one-third the size of our total sample, which consists of CRSP-Compustat firms over with all the needed data. This is why we use the full CRSP-Compustat sample for most of our analysis. 29

32 expenditures in generating patents and citations even if the total number of patents and citations obtained by these firms are higher because of higher R&D spending. To investigate this issue, we conduct two sets of analyses. First, we simply extend the model in column 2 of Table 3, Panel A and add the interaction of LnCPRatio and (R&D/Assets) 12. This model is analogous to HLT s (2012, Table XI) model of innovative efficiency, except that we add the interaction term to explicitly account for the moderating effect of gambling preferences on the effectiveness of R&D intensity in generating patents and citations. In Panel A of Table 8, the interaction terms obtain negative coefficients in explaining each of our three innovation output variables and, in two cases, are statistically significant at the 5% level. In each case, the coefficients on the main effects of LnCPRatio and (R&D/Assets) 12 remain positive and highly significant. These results suggest that, while firms in gambling-tolerant areas generate more patents and citations, on average they are less efficient in utilizing R&D intensity in doing so; consistent with the idea that part of the R&D expenditure is driven by gambling motives. Second, we develop a regression model which more formally estimates the productivity of R&D expenditure in generating innovation outputs. To this end, we modify the regressions of innovation outputs shown in Table 3, Panel A by recognizing two important issues. First, to estimate innovation output (i.e., patents and citations) per dollar of innovation input (i.e., R&D dollars), we need the unscaled R&D expenditures as an explanatory variable, and, in a parallel way, the raw number of successful patent applications, citations per patent and technology-adjusted citations as dependent variables. Second, the relation between R&D spending and patents or citations are likely to vary significantly across industries. For example, it might take several hundred million dollars in R&D expenditures to develop a viable patent in a drug company, while a toy company might generate a patent with a few million dollars in R&D. Consequently, our empirical model here has the raw values of each of the innovation output variables as the dependent variable, and the dollar value of R&D and its interaction with LnCPRatio and with each of the 2-digit SIC code industry dummy variables as explanatory variables. As with our main analysis, the regressions use the average of the first and second lags of R&D (i.e., R&D 12 ). The interaction of R&D expenditure with LnCPRatio estimates the effect of gambling attitudes on innovation efficiency, whereas the interactions with each industry dummy control for an industry effect on the relation between R&D spending and patents or citations. 30

33 The three columns in Panel B of Table 8 show the results of regressions of patent applications, citations per patent and technology-adjusted citations, respectively. The control variables are the same as in column 2 of Table 3, Panel A. For brevity, we only tabulate the coefficient estimates on R&D 12, LnCPRatio and their interaction. Consistent with our conjecture, we obtain a negative and statistically significant coefficient on the interaction of R&D 12 and LnCPRatio in explaining each of the three innovation outputs. These results imply that an extra dollar spent on R&D by firms in high CPRatio counties produces fewer patents, citations per patents and adjusted citations after controlling for firm-specific and county-level variables and for industry effects in patent productivity. As expected, the coefficients on the main effects of LnCPRatio and R&D 12 remain positive and significant. 5.4 Financing innovation: CPRatio, Cash and R&D The analysis in this sub-section is motivated by recent evidence that corporate cash holdings have become increasingly important as a strategic resource for innovation and R&D, and that the temporal increase in cash holdings is driven almost solely by R&D intensive firms (see Schroth and Szalay (2010), and Lyandres and Palazzo (2012)). These findings are an extension of the evidence in Opler, Pinkowitz, Stulz, and Williamson (1999), and Bates, Kahle and Stulz (2009) that R&D intensive firms tend to hold more cash as the cost of financial distress is higher for these firms. We test an implication of these findings for our results: if firms in higher CPRatio areas invest more in R&D, they should hold more cash and the levels of their cash holdings should be partially explained by higher R&D expenditures. To this end, we build on the previous literature that predicts the level of cash holdings (cash and marketable securities divided by total assets) of a firm, and add our gambling proxy variable, LnCPRatio, and its interaction with R&D/Assets ratio to the regression. If firms located in areas with higher concentrations of Catholics hold more cash, we expect to see a positive and significant coefficient on LnCPRatio. More importantly, if these firms hold more cash to invest in R&D, we expect a positive and significant coefficient on the interaction of R&D/Assets and LnCPRatio. Table 9 shows the results of regressions of cash holdings where LnCPRatio and its interaction with R&D/Assets are the main explanatory variables of interest. We use firm-level control variables found by Opler et. al. (1999) and Bates, Kahle and Stulz (2009) to significantly predict a firm s cash holdings. Consistent with our previous analysis, all firm-level control variables 31

34 are lagged by one year. We also include the full set of contemporaneous county-level control variables and year and industry fixed effects. Column 1 shows the coefficient estimates of LnCPRatio, R&D/Assets, Industry cash flow volatility (Industry Sigma), Acquisitions and the interactions of the former two variables. The coefficient estimates of the other control variables are not reported for brevity. As expected, LnCPRatio obtains a positive and highly significant coefficient in explaining level of a firm s cash holdings. Likewise, R&D/Assets obtains a positive and significant coefficient consistent with the previous literature. More importantly, the point estimate of the interaction between R&D/Assets and LnCPRatio is also positive and statistically significant at the 1% level. This result supports our conjecture that higher levels of cash holdings by firms in high CPRatio areas are partly driven by R&D intensity of these firms. A potential concern, however, is that firms in gambling-tolerant areas may have other unique characteristics that affect their cash holdings. So, the interaction of R&D and LnCPRatio might be tempered by other motivations for cash holdings that are omitted from the regression. The prior literature suggests two more determinants of cash holdings for such firms. First, more R&D intensive firms are also likely to face greater industry cash flow volatility (Industry Sigma), so the higher cash holding of firms in higher CPRatio areas might be driven by the precautionary motive of cash rather than by a strategic R&D motive. Second, recent evidence suggests that higher risk tolerance and higher gambling preferences of managers lead them to spend more on acquisitions (see, e.g., Gaham, Harvey and Puri (2013), and Schneider and Spalt (2013)), which can reduce the cash holdings of firms in more gambling-tolerant areas. To isolate the effect of the R&D motive from these other potential determinants of cash holdings in high CPRatio areas, we add the interactions of LnCPRatio with two other variables: Industry Sigma and Acquisitions to the regression in column 1. These interaction terms control for the precautionary and acquisition motives for cash holdings in high CPRatio areas, and make the estimate on LnCPRatio*R&D/Assets a cleaner measure of the incremental cash holding for R&D purposes by firms in gambling-tolerant areas. Column 2 of Table 9 shows the estimates of this regression. As expected, LnCPRatio*Industry Sigma indeed obtains a positive and significant coefficient, suggesting the importance of the precautionary motive for holding extra cash by firms in high CPRatio areas. Similarly, the interaction of Acquisitions with LnCPRatio obtains a negative and significant coefficient consistent with prior findings that managerial risk tolerance and gambling attitudes lead firms to spend more on M&A activities, leaving them with less cash. Nevertheless, the 32

35 coefficient estimate on LnCPRatio*R&D/Assets remains positive and statistically significant at the 5% level. It is also noteworthy that once these two interaction terms are introduced, the main effect of LnCPRatio on cash holdings turns insignificant. This finding suggests that all the extra dollars of cash held by firms in high CPRatio areas can be attributed to these three reasons. Moreover, we find that the R&D motive for higher cash holdings by firms in high CPRatio areas is significant in innovative industries (see column 3), but not in non-innovative industries (see column 4). This result further strengthens our evidence that greater cash holdings by firms in more gambling-tolerant areas are at least partly for funding innovation. 6. Firm Valuation We find in sections 3 and 5.3 that firms in more gambling-tolerant areas invest more in R&D, produce more and higher quality innovation output, but are less efficient at generating innovations per unit of R&D spending. While our empirical results so far do not do suggest an unambiguous direct effect of gambling preferences on overall firm valuation, we ask a more specific question: Does firm value vary more with local gambling preferences when innovation is more important? More generally, are firms in areas with a greater preference for gambling more adept at transforming their growth opportunities into firm value? To address this issue, we follow HLT s (2012) framework and estimate the following regression: Tobin s Q = b o + b 1 IndGrowthOpp + b 2 LnCPRatio + b 3 LnCPRatio*IndGrowthOpp + firm-level controls + county-level controls + 2-digit SIC industry dummies + year dummies, where Tobin s Q is the ratio of market value of assets to book value of assets and IndGrowthOpp is an exogenous, industry-level measure of a firm s growth opportunities. We use two alternate measures of IndGrowthOpp. First, we use industry innovativeness (Innovative Ind) which equals one for 4-digit SIC industries whose citations per patent exceed the median for all industries in a given year; it equals zero for other industries. Hall, Jaffe and Trajtenberg s (2005) findings suggest that firms in innovative industries have greater potential for value creation through innovation. So, Innovative Ind is a cross-sectional and time-varying measure of the importance of innovation for firm value. Our second, broader measure of industry growth opportunities is peer firms average price-toearnings ratio (PeerPE), which we compute as the natural log of total market value of equity divided 33

36 by total earnings of all other firms in a firm s 2-digit SIC industry. As HLT (2012) point out, PE is a noisy measure of growth opportunities because it is also influenced by discount rate changes, biasing our tests against finding a significant result. Firm-level control variables are the same as in HLT (2012). In addition, we control for all the county-level variables used in our earlier analyses. If firms in areas with a greater preference for gambling are more adept at transforming industry growth opportunities into firm value, we expect b 3 to be positive. However, we do not have a prior on the sign on b 2. Our finding in section 3.2 that investors in more gambling-tolerant areas produce more and better quality innovation output implies that b 2 should be positive. But our finding in section 5.3 that firms in high CPRatio areas are less efficient at generating innovations per unit of R&D spending implies that b 2 should be negative. Table 10 shows estimates of four variants of this regression. For brevity, we do not report estimates of the county-level control variables and the intercept. Models (1) and (3) include alternate measures of industry growth opportunities, without firm-level controls. The coefficient estimates of both measures are positive and highly significant, consistent with the idea that firms in industries with higher growth opportunities command higher valuations. Models (2) and (4) include LnCPRatio, a measure of industry growth opportunities, their interaction, and firm-level controls. The coefficient estimate of LnCPRatio is ambiguous: it is positive in model (2) and negative in model (4). But our main interest here is in the interaction term, LnCPRatio*Innovative Ind or LnCPRatio*PeerPE, which is positive and highly significant in both models. This result suggests that firms in areas with greater preference for gambling are more successful at converting their potential growth opportunities into realized firm value Conclusion Innovation leads not only to higher firm values (see, e.g., Hall, Jaffe and Trajtenberg (2005), Sorescu and Spanjol (2008), and Hirshleifer, Hsu and Li (2013)) but is also an engine of economic growth (see Solow (1957), and Kogan, Papanikolaou, Seru and Stoffman (2012)). Consequently, both academic research and public policy discussion has focused on identifying factors that lead to more and better innovation. While most prior research investigates rational factors related to firms and financial markets as contributors to innovation, we examine a behavioral determinant of innovation. 21 This finding parallels HLT s (2012) finding that firms with overconfident CEOs are more adept at translating their potential growth opportunities into firm value. 34

37 We find that religion-induced gambling preferences play a role in promoting innovative activities. In particular, firms located in US counties with higher concentrations of Catholics relative to Protestants produce more patents and generate more citations. Higher numbers of patents and citations are primarily the result of higher levels of R&D expenditures made by these firms. These results are robust to several alternative empirical 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. Supporting a gambling motive for higher R&D expenditures, we find that R&D spending makes stocks more lottery-like by increasing both their idiosyncratic volatility and idiosyncratic skewness, features of stocks found by Kumar (2009) to be desired by investors with a preference for gambling. In addition, we empirically confirm four auxiliary implications of our hypothesis. First, R&D spending shows greater sensitivity to local gambling preferences in firms that are more reliant on local investors. Second, 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, 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. Third, firms in high CPRatio areas tend to be less efficient in generating patents for a given level of R&D expenditures. This finding supports the idea that these firms overinvest in R&D to cater to local gambling preferences. Fourth, firms in more gambling-tolerant areas tend to hold more cash, partly to invest in R&D. Finally, firms in more gambling-tolerant areas appear to be more adept at converting industry growth opportunities into firm value as measured by Tobin s Q. Our findings suggest that innovative endeavors, like many other financial decisions, are partly a product of human behavior. Our findings are congruent with the rich literature on investor behavior, particularly regarding their preference for local stocks, gambling and skewness. However two caveats are worth noting. First, despite a battery of analyses we present showing a strong positive relation between local gambling preferences and innovation activities, our results are mostly silent on the direction of causality. There are two plausible explanations of this relation. The first is the causal explanation that local gambling preferences cause firms to pursue more innovative endeavors. The second is the endogenous matching explanation that firms which need more 35

38 innovation to succeed decide to locate in areas where local residents have a taste for gambling. Although our main empirical findings allow for both possibilities, we are somewhat skeptical of the endogenous matching story for at least two reasons. First, our data spans many years after a firm picks its headquarters location. Second, a firm s location decision is based on many other important factors apart from innovation motives, such as production cost, and proximity to suppliers, customers and workforce etc. Therefore, it seems less likely that our results are driven by a firm s decades-old decision to locate its headquarters in a particular area, especially given that the decision was likely based on many other factors unrelated to innovation. The second caveat is regarding the normative interpretation of our results. We examine one consequence of gambling preferences, namely firm innovation. However, we recognize that gambling has many other consequences on individuals and society, and do not argue that gambling is necessarily optimal for a society in an overall sense. But nurturing some aspects of gambling preference such as a tolerance for failure, a focus on the highest possible reward and perhaps the ability to endure and enjoy some risk might be beneficial for innovation, which is crucial for economic growth. 36

39 References Aghion, Philippe, John Van Reenen, and Luigi Zingales, 2013, Innovation and institutional ownership, American Economic Review 103, Aghion, Philippe, Nick Bloom, Richard Blundell, Rachel Griffith, Peter Howitt, 2005, Competition and innovation: An inverted-u relationship, Quarterly Journal of Economics 120, Alesina, Alberto, and Eliana La Ferrara, 2000, Participation in heterogeneous communities, Quarterly Journal of Economics 115, Atanassov, Julian, 2013, Do hostile takeovers stifle innovation? Evidence from antitakeover legislation and corporate patenting, Journal of Finance 68, Bailey, Warren, Alok Kumar, and David Ng, 2011, Behavioral Biases of Mutual Fund Investors, Journal of Financial Economics 102, Bali, Tarun G., Nusret Cakici, and Robert F. Whitelaw, 2011, Maxing-out: Stocks as lotteries and the cross section of expected returns, Journal of Financial Economics 99, Balkin, David B., Gideon D. Markman, and Luis R. Gomez-Mejia, 2000, Is CEO pay in hightechnology firms related to innovation?, Academy of Management Journal 43, Barberis, Nicholas, and Ming Huang, 2008, Stocks as lotteries: The implications of probability weighting for security prices, American Economic Review 98, Bates, Thomas W., Kathleen M. Kahle, and René M. Stulz, 2009, Why do U.S. firms hold so much more cash than they used to?, Journal of Finance 64, Becker, Bo, Zoran Ivković, and Scott Weisbenner, 2011, Local dividend clienteles, Journal of Finance 66, Bena, Jan and Kai Li, 2013, Corporate innovations and mergers and acquisitions, Journal of Finance, forthcoming. Burns, Alvin C., Peter L. Gillett, Marc Rubinstein, and James W. Gentry, 1990, An exploratory study of lottery playing, gambling addiction and links to compulsive consumption, Advances in Consumer Research, 17,

40 Callen, Jeffrey L., and Xiaohua Fang, 2013, Religion and stock price crash risk, Journal of Financial and Quantitative Analysis, forthcoming. Carhart, Mark M., 1997, On persistence in mutual fund performance, Journal of Finance 52, Chan, Louis K. C., Josef Lakonishok, and Theodore Sougiannis, 2001, The stock market valuation of research and development expenditures, Journal of Finance 56, Cohen, Lauren, Umit G. Gurun, and Christopher J. Malloy, 2013, Resident networks and firm trade, NBER Working Paper Coles, Jeffrey, Naveen Daniel, and Lalitha Naveen, 2006, Managerial incentives and risk-taking, Journal of Financial Economics 79, Core, John E. and Wayne Guay, 2002, Estimating the value of employee stock option portfolios and their sensitivities to price and volatility, Journal of Accounting Research 40, Fama, Eugene F., and Kenneth R. French, 1993, Common risk factors in returns on stocks and bonds, Journal of Financial Economics 33, Fang, Vivian W., Xuan Tian, and Sheri Tice, 2013, Does stock liquidity enhance or impede firm innovation?, Journal of Finance, forthcoming. Friedman, Milton, and Leonard J. Savage, 1948, The utility analysis of choices involving risk, Journal of Political Economy 56, Galasso, Alberto, and Timothy Simcoe, 2011, CEO overconfidence and innovation, Management Science 57, Goodie, Adam S., 2005, The role of perceived control and overconfidence in pathological gambling, Journal of Gambling Studies 21, Graham, John, Campbell Harvey, and Manju Puri, 2013, Managerial attitudes and corporate actions, Journal of Financial Economics 109, Grichting, W. L., 1986, The impact of religion on gambling in Australia, Australian Journal of Psychology 38,

41 Grinblatt, Mark, and Matti Keloharju, 2001, How distance, language, and culture influence stockholdings and trades, Journal of Finance 56, Grullon, Gustavo, George Kanatas, and James P. Weston, 2010, Religion and corporate (mis)behavior, Working paper, Rice University. Hall, Bronwyn H., 1999, Innovation and market value, NBER Working Paper Hall, Bronwyn H., Adam Jaffe, and Manuel Trajtenberg, The NBER patent citations data file: Lessons, insights and methodological tools, NBER Working Paper Hall, Bronwyn H., Adam Jaffe, and Manuel Trajtenberg, 2005, Market value and patent citations, Rand Journal of Economics 36, Harvey, Campbell R., and Akhtar Siddique, 2000, Conditional skewness in asset pricing tests, Journal of Finance 55, He, Jie (Jack), and Xuan Tian, 2013, The dark side of analyst coverage: The case of innovation, Journal of Financial Economics 109, Hilary, Gilles, Kai Wai Hui, 2009, Does religion matter in corporate decision making in America?, Journal of Financial Economics 93, Hirshleifer, David, Angie Low, and Siew Hong Teoh, 2012, Are overconfident CEOs better innovators?, Journal of Finance 67, Hirshleifer, David, Po-Hsuan Hsu, Dongmei Li, Innovative efficiency and stock returns, Journal of Financial Economics 107, Hong, Harrison, Jeffrey Kubik, and Jeremy Stein, 2008, The only game in town: Stock-price consequences of local bias, Journal of Financial Economics 90, Hsu, Po-Hsuan, Xuan Tian, and Yan Xu, 2013, Financial development and innovation: Crosscountry evidence, Journal of Financial Economics, forthcoming. Huberman, Gur, 2001, Familiarity breeds investment, Review of Financial Studies 14,

42 Ivković, Zoran, and Scott Weisbenner, 2005, Local does as local is: Information content of the geography of individual investors common stock investments, Journal of Finance 60, Kahneman, Daniel, and Amos Tversky, 1979, Prospect theory: An analysis of decision under risk, Econometrica 47, Kogan, Leonid, Dimitris Papanikolaou, Amit Seru, and Noah Stoffman, 2012, Technological innovation, resource allocation and growth, NBER Working Paper Kumar, Alok, 2009, Who gambles in the stock market?, Journal of Finance 64, Kumar, Alok, Jeremy Page, and Oliver Spalt, 2011, Religious beliefs, gambling attitudes, and financial market outcomes, Journal of Financial Economics 102, Lakey, Chad E., Paul Rose, W. Keith Campbell, and Adam S. Goodie, 2008, Probing the link between narcissism and gambling: The mediating role of judgment and decision-making biases, Journal of Behavioral Decision Making 21, Loughran, Tim and Jay Ritter, 2004, Why has IPO underpricing changed over time?, Financial Management 33, Lyandres, Evgeny, and Berardino Palazzo, 2012, Strategic cash holdings and R&D competition: Theory and evidence, Working Paper, Boston University. Massa, Massimo, and Andrei Simonov, 2006, Hedging, familiarity and portfolio choice, Review of Financial Studies 19, Mobbs, Shawn, 2013, CEO s Under Fire: The Effects of Directors on CEO Compensation and Turnover, Journal of Financial and Quantitative Analysis 48, Myers, Stuart, 1977, Determinants of corporate borrowing, Journal of Financial Economics 5, Myers, Stuart, and Nicholas Majluf, 1984, Corporate financing and investment decisions when firms have information that investors do not have, Journal of Financial Economics 13, Nanda, Ramana, and Matthew Rhodes-Kropf, 2011, Financing risk and innovation, Harvard Business School Working Paper

43 Opler, Tim, Lee Pinkowitz, René M. Stulz, and Rohan Williamson, 1999, The determinants and implications of corporate cash holdings, Journal of Financial Economics 52, Pakes, Ariel, and Zvi Griliches, 1980, Patents and R&D at the firm level: A First Look, in Z. Griliches, ed.: R&D, Patents, and Productivity, (University of Chicago Press, Chicago, IL). Rapoport, John, 1971, The anatomy of product-innovation process: Cost and time, in E. Mansfield, ed.: Research and innovation in the modern corporation, (Norton, New York). Schneider, Christoph, and Oliver Spalt, 2013, Acquisitions as lotteries: do managerial gambling attitudes influence takeover decisions?, Working Paper, University of Mannheim and Tilburg University. Schroth, Enrique, and Dezsö Szalay, 2010, Cash breeds success: The role of financing constraints in patent races, Review of Finance 14, Solow, Robert M., 1957, Technical Change and the Aggregate Production Function, Review of Economics and Statistics 39, Sorescu, Alina B., and Jelena Spanjol, 2008, Innovation's effect on firm value and risk: Insights from consumer packaged goods. Journal of Marketing 72, Tec, Nechama, 1964, Gambling in Sweden (Bedminster Press, Totowa, NJ). Tian, Xuan, and Tracy Yue Wang, 2013, Tolerance for failure and corporate innovation, Review of Financial Studies, forthcoming. Tversky, Amos, and Daniel Kahneman, 1992, Advances in prospect theory: Cumulative representation of uncertainty, Journal of Risk and Uncertainty 5, Wagner, Leonore U., 1968, Problems in estimating research and development investment and stock, in Proceedings of the Business and Economic Statistics Section (American Statistical Association, Washington, DC). Yonker, Scott E., 2012, Geography and the market for CEOs, Working Paper, Indiana University Bloomington. 41

44 Appendix Variable definitions Variable Name Patent Applications LnPatent Citations Per Patent LnCitePerPat Tech-Adjusted Citations LnTechAdjCites LnCPRatio R&D 12 R&D/Assets (R&D/Assets) 12 LnSales LnFirmAge Idio. Volatility Idio. Skewness Description The number of (eventually granted) patents applied for during a year. Replaced by zero if missing. Natural logarithm of one plus Patent Applications. The total number of citations received during the 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. The number of citations is adjusted by the weighting index of Hall, Jaffe and Trajtenberg (2001, 2005). Replaced by zero if citation counts are missing. Natural logarithm of one plus Citations Per Patent Total number of technology class-adjusted citations received during the sample period on all patents filed (and eventually received) by a firm during the year. Each patent s adjusted citation count is divided by the average adjusted citation count for all patents filed (and eventually granted) in the same technology class during the year by all the sample firms. A patent s adjusted citation count is its number of citations, adjusted by the weighting index of Hall, Jaffe and Trajtenberg (2001, 2005). Replaced by zero if citation counts are missing. Natural logarithm of one plus Tech-Adjusted Citations. Natural logarithm of one plus the number of Catholic adherents to the number of Protestant adherents in a county in a given year. The average of the first and the second lags of R&D expenditures (Compustat data item: XRD). Replaced by zero if missing. R&D expenditure (Compustat data item: XRD) divided by total assets (Compustat data item: AT). Replaced by zero if missing. The average of the first and the second lags of R&D/Assets Natural logarithm of sales (Compustat: SALE). Natural logarithm of firm age, approximated by current fiscal year minus the year the firm first appears in Compustat. Idiosyncratic volatility, computed as the standard deviation of the residuals obtained by regressing daily returns on a stock on Fama and French (1993) and Carhart s (1997) four factors over a year. Idiosyncratic skewness, computed as the skewness of residuals obtained by regressing daily returns on a stock on excess market return and excess market return squared over a year.

45 Lottery Stock ROA Ln(PPE/Emp) An indicator variable that equals one if a stock s returns exhibit above-median idiosyncratic volatility and above-median idiosyncratic skewness in a given year, and zero otherwise. Operating income before depreciation divided by total assets (Compustat: OIBDP/AT). Natural logarithm of net property, plant and equipment divided by the number of employees (Compustat: PPENT/EMP). Book Leverage Long-term plus short-term debt divided by total assets [Compustat: (DLC + DLTT)/AT]. Capex Tobin's Q KZ Index HHI Inst. Own LnAnalysts LnPopulation Younger Rural Urban Continuum Capital expenditure divided by total assets (Compustat: CAPX/AT) Ratio of market value of assets to book value of assets = Market value of equity plus book value of assets minus book value of equity minus accumulated deferred taxes, all divided by the book value of assets [Compustat: (AT - CEQ + CSHO*PRCC_F - TXDITC)/AT]. Kaplan and Zingales index for a given year calculated as *KZ_CashFlowToCapital *KZ_Q *KZ_Leverage *KZ_Div *KZ_CashToCapital Where, KZ_CashFlowToCapital = (IB + DP))/lag(PPENT) from Compustat; KZ_Q = (AT + PRCC_F*CSHO - CEQ - TXDB)/AT from Compustat; KZ_Leverage = (DLTT+ DLC)/( DLTT + DLC+ SEQ) from Compustat; KZ_Div = (DVC + DVP)/ lag(ppent) from Compustat; KZ_CashToCapital = CHE/ lag(ppent) from Compustat. Herfindahl-Hirschman index based on sales for a given SIC industry code measured at the end of a fiscal year. Percentage of a firm s stock held by institutional investors at the end of a fiscal year, reported in SEC Form 13F. Natural logarithm of one plus the number of analysts following a firm in a given year. Natural logarithm of a county s population in a given year. An indicator variable that equals one if the median age of the people in a firmcounty is less than the median age for all the firm-counties in a given year, and equals zero otherwise. A classification scheme that distinguishes metropolitan (metro) counties by the population size of their metro area, and nonmetropolitan (nonmetro) counties by the degree of urbanization and adjacency to a metro area or areas. Scaled from 1 to 9, where a higher number means more rural (1 to 3 refer to metro and 4 to 9 refer to nonmetro areas).

46 College Grads Minority Population Married Household Male to Female Ratio Percentage of 25 years or older residents in a county who have at least a Bachelor s degree. Percentage of non-white population in a county. Percentage of population living in a household of married couples in a county. Ratio of male population to female population in a county. LnAdherentsPer1000 Natural logarithm of the number of religious adherents in a county per 1000 residents. Excess Cash LnMarketCap Amihud Illiquidity Dividend Payer Turnover Stock Return Price Per Share Market Beta SMB Beta HML Beta SmallFirm Low BE/PI Residual from a regression of a firm s cash to assets ratio on control variables as in Opler et. al. (1999) and Bates, Kahle and Stulz (2009). Natural logarithm of the market capitalization of a firm calculated at the end of the fiscal year calculated as price per share multiplied by the number of common shares outstanding (Compustat: PRCC_F*CSHO). Absolute daily returns per unit of trading volume, averaged over the number of trading days in a year. An indicator variable that equals one if the firm pays a cash dividend in a given year, and equals zero otherwise. Average monthly shares traded divided by the number of shares outstanding during a year. Holding period stock return for a year. Price per share at the end of a fiscal year. Loadings on market risk premium estimated by a factor model using the prior sixty monthly returns. Loadings on the SMB factor estimated by a four factor model using the prior sixty monthly returns. Loadings on the HML factor estimated by a four factor model using the prior sixty monthly returns. An indicator variable that equals one if a firm in a given year has a smaller market capitalization than the median of all firms, and zero otherwise. An indicator variable that equals one if BE/PI associated with a firm-year is below its median, and equals zero otherwise. BE/PI is calculated as the ratio of the aggregate book value of equity of all publicly-traded firms headquartered in a county divided by the aggregate personal income (= per capita income * number of residents) of the county. A county s per capita income is approximated by its state s per capita income.

47 Low BE/PI Visible High Indiv Invest Cash Holding Industry Sigma Acquisitions Innovative Ind PeerPE An indicator variable that equals one if a firm has below median number of shareholders and below median BE/PI in a given year, and zero otherwise. An indicator variable that equals one if a firm has above-median fraction of individual investors (i.e. below-median fraction of institutional investors) in a given year, and zero otherwise. Cash and marketable securities divided by total book assets (Compustat: CHE/AT). The mean of the standard deviations of cash flow/assets ratio over the 10 year period of the industry, as defined by the two-digit SIC code. Dollars spent on acquisitions divided by the book value of assets (Compustat: AQC/AT). An indicator variable that equals one for 4-digit SIC industries whose citations per patent exceed the median for all industries in a given year; it equals zero for other industries. Natural logarithm of total market value of equity divided by total earnings of all other firms in a firm s industry, as defined by the two-digit SIC code. # Business Segments Number of business segments from Compustat segment file.

48 Table 1 Summary statistics The table reports the summary statistics of our key variables of interest. Panel A shows county-related variables at the county-level for the latest year that a county appears in our sample. Panel B shows variables at the firm-year level. The sample consists of U.S. public companies on CRSP and Compustat, excluding financial firms (2-digit SIC codes 60 to 69) and utilities (2-digit SIC code 49), from 1980 to All the variables used in the regression analyses are defined in the Appendix. Panel A: County level summary statistics Mean Std. Dev. 25th percentile Median 75th Percentile N All Adherents per Catholics per Protestants per CPRatio LnCPRatio College Grads (%) Deflated Income ($ 000) Male to Female Ratio Married Household Fraction Minority Population Age Group Rural Urban Continuum Total Population ( 000)

49 Panel B: Firm level summary statistics Mean Std. Dev. 25th percentile Median 75th Percentile N CPRatio LnCPRatio Patent Applications LnPatent CitePerPat LnCitePerPat Tech-Adjusted Citations LnTechAdjCites R&D R&D/Assets Lottery Stock IdioVolatility Idio. Skewness Total Assets ($ millions) Sales ($millions) ROA Firm Age PPE/Emp Book Leverage Capex KZ Index HHI Inst. Own Number of Analysts Tobin's Q Market Cap ($millions) Amihud Illiquidity Dividend Payer Turnover Stock Return Price Per Share Market Beta SMB Beta HML Beta Cash Holding Industry Sigma Acquisitions

50 Table 2 Correlation Matrix The table reports the Pearson correlation coefficients among select variables of interest. The sample consists of U.S. public companies on CRSP and Compustat, excluding financial firms (2-digit SIC codes 60 to 69) and utilities (2-digit SIC code 49), from 1980 to All the variables used in the regression analyses are defined in the Appendix. All continuous variables are winsorized at the 1% level in both tails. All non-italicized correlations are statistically significant at the 1% level. (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) (17) (18) CPRatio (1) 1 LnCPRatio (2) LnPatent (3) LnCitePerPat (4) LnTechAdjCites (5) (R&D/Assets) 12 (6) Lottery Stock (7) Idio. Volatility (8) Idio. Skewness (9) Cash Holding (10) LnPopulation (11) Younger (12) Rural Urban Continuum (13) College Grads (14) Minority Population (15) Married Household (16) Male to Female Ratio (17) LnAdherentsPer1000 (18)

51 Table 3 Innovation outcomes and gambling preference The table reports estimates of regressions of innovation outcomes (patents, citations per patent and technology-adjusted citations) on the local gambling preference proxy, LnCPRatio. All the variables are defined in the Appendix. All firm-level independent variables are lagged by one year except for R&D/Assets, which is the average of the first and second lags. County-level control variables are contemporaneous. Panel A shows the results of regressions on the full sample and panel B shows the results on sub-samples partitioned by industry innovativeness, where a 4-digit SIC code industry is defined as innovative if its average citations per patent exceed the median of all industries in a given year. All continuous variables are winsorized at 1% in both tails. All regressions include year and industry dummies where industry is defined based on 2- digit SIC codes. Intercepts are not reported. Standard errors are corrected for heteroscedasticity and are clustered at the firm level, and t-statistics are in parentheses. *, **, and *** indicate statistical significance at the 10%, 5% and 1% levels, respectively. Panel A: Patents and citations per patent full sample (1) (2) (3) (4) (5) (6) Dependent variable LnPatent LnPatent LnCitePerPat LnCitePerPat LnTechAdjCites LnTechAdjCites LnCPRatio 0.101*** 0.144*** 0.079*** 0.120*** 0.090*** 0.135*** (2.86) (3.28) (2.76) (3.23) (2.75) (3.32) (R&D/Assets) *** 1.723*** 2.217*** 2.111*** 1.413*** 1.320*** (8.79) (8.20) (10.55) (10.10) (7.21) (6.76) LnSales 0.273*** 0.275*** 0.144*** 0.145*** 0.241*** 0.243*** (14.46) (14.55) (12.53) (12.77) (13.67) (13.72) LnFirmAge 0.186*** 0.192*** 0.077*** 0.084*** 0.148*** 0.154*** (6.76) (6.94) (3.34) (3.64) (6.02) (6.24) ROA *** *** *** *** (-3.33) (-3.57) (-0.89) (-1.10) (-3.14) (-3.40) Ln(PPE/Emp) 0.104*** 0.103*** 0.062*** 0.062*** 0.073*** 0.072*** (5.53) (5.49) (4.48) (4.43) (4.11) (4.06) Book Leverage *** *** *** *** *** *** (-4.13) (-4.17) (-3.30) (-3.32) (-4.44) (-4.42) Capex ** 0.445** (0.33) (0.28) (0.62) (0.61) (2.44) (2.39) Tobin's Q 0.061*** 0.060*** 0.041*** 0.040*** 0.056*** 0.056*** (6.58) (6.54) (4.67) (4.62) (6.16) (6.14) KZ Index *** *** ** ** *** *** (-4.03) (-3.96) (-2.09) (-2.11) (-4.15) (-4.12) HHI 0.162* 0.159* (1.79) (1.77) (0.35) (0.37) (1.52) (1.51) Inst. Own *** *** *** *** (-3.63) (-3.66) (1.55) (1.59) (-4.38) (-4.40) LnAnalysts 0.166*** 0.163*** 0.131*** 0.127*** 0.155*** 0.153*** (8.20) (8.12) (7.16) (7.01) (8.22) (8.14) LnPopulation (-1.27) (-0.90) (-1.51) Younger (0.54) (1.15) (1.05)

52 Rural Urban Continuum (1.00) (1.26) (0.35) College Grads 0.008*** 0.006*** 0.006** (2.98) (2.81) (2.29) Minority Population (-0.50) (0.03) (-0.87) Married Household * (0.90) (1.85) (0.58) Male to Female Ratio (-0.91) (-1.20) (-0.82) LnAdherentsPer (-1.22) (-1.51) (-1.07) Year Fixed Effects Yes Yes Yes Yes Yes Yes Industry Fixed Effects Yes Yes Yes Yes Yes Yes Observations Adjusted R Panel B: Effect of Industry Innovativeness Dependent variable LnPatent LnCitePerPat LnTechAdjCites (1) (2) (3) (4) (5) (6) Non- Non- Non- Innovative Innovative Innovative Innovative Innovative Innovative Ind. Ind. Ind. Ind. Ind. Ind. LnCPRatio 0.195*** 0.094** 0.229*** *** (3.48) (2.14) (3.81) (0.71) (3.73) (1.32) (R&D/Assets) *** 1.447*** 2.310*** 1.459*** 1.705*** 0.701*** (7.67) (5.86) (7.46) (7.11) (6.61) (3.70) Year Fixed Effects Yes Yes Yes Yes Yes Yes Industry Fixed Effects Yes Yes Yes Yes Yes Yes Observations Adjusted R

53 Table 4 Innovative input and gambling preference The table reports the results of regressions of an innovation input (R&D expenditures scaled by total assets) on the local gambling preference proxy, LnCPRatio. All the variables are defined in the Appendix. All firmlevel independent variables are lagged by one year. County-level control variables are contemporaneous. Panel A shows the results of regressions on the full sample and panel B shows the results on sub-samples partitioned by industry innovativeness, where a 4-digit SIC code industry is defined as innovative if its average citations per patent exceed the median of all industries in a given year. All continuous variables are winsorized at the 1% level in both tails. All regressions include year and industry dummies, where industry is defined based on 2-digit SIC codes. Intercepts are not reported. Standard errors are corrected for heteroscedasticity and are clustered at the firm level, and t-statistics are in parentheses. *, **, and *** indicate statistical significance at the 10%, 5% and 1% levels, respectively. Panel A: R&D expenditure full sample (1) (2) Dependent variable R&D/Assets R&D/Assets LnCPRatio 0.006*** 0.007*** (3.08) (2.69) Book Leverage *** *** (-4.31) (-4.53) Tobin's Q 0.014*** 0.014*** (11.90) (11.85) Stock Return *** *** (-15.50) (-15.11) LnFirmAge ** (-1.98) (-0.95) LnSales *** *** (-15.13) (-14.87) Sales Growth (-0.38) (-0.54) Ln(PPE/Emp) *** ** (-2.60) (-2.45) Excess Cash (1.54) (1.24) Capex ** * (-2.16) (-1.78) LnAnalysts 0.020*** 0.019*** (15.15) (14.20) Inst. Own *** *** (-3.41) (-3.34) LnPopulation (-0.14) Younger (-0.82)

54 Rural Urban Continuum (-0.56) College Grads 0.001*** (7.09) Minority Population (1.17) Married Household (1.52) Male to Female Ratio * (-1.76) LnAdherentsPer *** (-5.89) Year Fixed Effects Yes Yes Industry Fixed Effects Yes Yes Observations Adjusted R Panel B: Effect of industry innovativeness Dependent variable R&D/Assets (1) (2) Non- Innovative Innovative Ind. Ind. LnCPRatio 0.011*** 0.004* (2.60) (1.72) Year Fixed Effects Yes Yes Industry Fixed Effects Yes Yes Observations Adjusted R

55 Table 5 Robustness tests This table reports the results of several robustness tests performed on the regressions of LnPatent, LnCitePerPat, LnTechAdjCites and R&D/Assets. The main specification is the regression on the full sample with the complete set of controls, as presented in columns 2, 4 and 6 of table 3, Panel A and column 2 of Table 4, Panel A for LnPatent, LnCitePerPat, LnTechAdjCites and R&D/Assets, respectively. The four sets of six columns each are for the regressions of LnPatent, LnCitePerPat, LnTechAdjCites and R&D/Assets, respectively. The first three columns in each set are for the full sample and the next three columns are for the sub-sample of innovative industries. A 4-digit SIC code industry is defined as innovative if its average citations per patent exceed the median of all industries in a given year. All continuous variables are winsorized at 1% in both tails. All regressions include year and industry dummies, where industry is defined based on 2-digit SIC codes. Standard errors are corrected for heteroscedasticity and are clustered at the firm-level, except in test 9, where they are clustered at the county-level. Dependent Variable LnPatent LnCitePerPat Full Sample Innovative Ind. Full Sample Innovative Ind. t- t- Coef. t-stat N Coef. t-stat N Coef. stat N Coef. stat N Main Specification Alternative Specifications 1) Gambling index instead of LnCPRatio ) CPRatio instead of LnCPRatio ) Tobit instead of OLS ) Sample with non-zero patents only ) Random effects, het. robust, firm cluster ) 2SLS - Instrument: 3 lags of LnCPRatio Geography and Industry Related 7) Exclude 5 industries with most patents ) Exclude 5 most Catholic States ) Exclude 5 least Catholic States ) Cluster by county FIPS

56 Table 5 (cont.) LnTechAdjCites R&D/Assets Full Sample Innovative Ind. Full Sample Innovative Ind. t- t- Coef. t-stat N Coef. t-stat N Coef. stat N Coef. stat N Main Specification Alternative Specifications 1) Gambling index instead of LnCPRatio ) CPRatio instead of LnCPRatio ) Tobit Instead of OLS ) Sample with non-zero patents only ) Random effects, het. robust, firm cluster ) 2SLS - Instrument: 3 lags of LnCPRatio Geography and Industry Related 7) Exclude 5 industries with most patents ) Exclude 5 most Catholic States ) Exclude 5 least Catholic States ) Cluster by county FIPS

57 Table 6 Motivation for higher R&D expenditures This table reports the results of regressions where the dependent variables are 1) an indicator variable for a lottery stock, 2) Idiosyncratic volatility of a stock in a given year, and 3) Idiosyncratic skewness of a stock in a given year. All the variables are defined in the Appendix. All independent variables at the firm-level are lagged by one year. County-level control variables (not reported for brevity) are contemporaneous. Models 1 and 2 are logit and fixed effects logit models where the dependent variable is a dummy variable for a lottery stock. Models 3 and 4 are OLS and firm fixed effects panel regressions with idiosyncratic volatility as the dependent variable. Models 5 and 6 are OLS and firm fixed effects panel regressions with idiosyncratic skewness as the dependent variable. All continuous variables are winsorized at 1% in both tails. All regressions include year dummies, and all regressions except those with firm fixed effects include industry dummies. Industry is defined by 2-digit SIC codes. Intercepts are not reported. Standard errors are corrected for heteroscedasticity and are clustered at the firm level, and t-statistics are in parentheses. *, **, and *** indicate statistical significance at the 10%, 5% and 1% levels, respectively. (1) Lottery Stock (2) Lottery Stock (3) Idio. Volatility (4) Idio. Volatility (5) Idio. Skewness (6) Idio. Skewness Dependent variable R&D/Assets 1.137*** 1.090*** 0.019*** 0.010*** 0.813*** 0.733*** (4.76) (2.78) (8.83) (3.64) (6.66) (3.68) LnPatent *** 0.000** * (-0.56) (-0.25) (6.19) (1.99) (1.39) (-1.66) LnCPRatio (-0.78) (-1.31) (-0.80) (-0.89) (1.12) (1.37) LnFirmAge ** *** *** *** 0.109** (-2.19) (-2.80) (-7.91) (0.09) (6.64) (2.57) LnMarketCap *** *** *** *** *** *** (-8.37) (-15.62) (-15.79) (-18.75) (-9.86) (-17.04) Tobin s Q 0.113*** 0.115*** 0.001*** 0.001*** ** (7.00) (5.81) (10.41) (9.76) (-2.01) (-0.04) Amihud Illiquidity 0.016*** 0.008*** 0.001*** 0.000*** 0.003*** (5.89) (3.24) (25.12) (15.09) (3.45) (-0.75) Dividend Payer *** *** *** *** (-12.44) (-3.67) (-13.33) (-4.18) (-0.81) (-0.05) Turnover 0.204*** 0.133*** 0.002*** 0.001*** 0.019*** (12.62) (6.41) (18.47) (5.97) (2.65) (1.11) Stock Return *** *** *** *** *** *** (-11.28) (-7.91) (-25.18) (-19.46) (-14.21) (-7.16) Inst. Own *** *** *** *** *** *** (-9.36) (-6.00) (-19.31) (-3.65) (-10.95) (-5.34) LnAnalysts ** ** (-1.60) (-0.21) (0.38) (1.97) (-2.24) (1.19) Price Per Share *** *** *** 0.000*** (-14.66) (-7.17) (-6.97) (5.29) (0.49) (-1.09) Market Beta 0.082*** *** * (3.73) (-0.05) (3.48) (-0.33) (-1.12) (-1.74) SMB Beta 0.071*** ** *** (4.70) (-0.39) (2.26) (-1.17) (-0.71) (-2.66) HML Beta *** ***

58 1.137*** 1.090*** 0.019*** 0.010*** 0.813*** 0.733*** Year Fixed Effects Yes Yes Yes Yes Yes Yes Industry Fixed Effects Yes Yes Yes Firm Fixed Effects Yes Yes Yes Observations Adjusted (Pseudo) R 2 (0.335) (0.159)

59 Table 7 Role of local investors This table reports the analysis of the effect of our local gambling preference proxy, LnCPRatio on an innovation input (R&D expenditure scaled by total assets). All the variables are defined in the Appendix. SmallFirm equals one if the market value of a firm is below the median of all firms in a given year. Low BE/PI equals one if the BE/PI associated with a firm is below the median BE/PI of all sample firms in a given year, and zero otherwise. BE/PI equals the total book equity of all public companies headquartered in a county during a year divided by the aggregate personal income of all residents of the county during the year. Low BE/PI Visible is an indicator variable that equals one if a firm has below-median number of shareholders and below-median BE/PI. High Indiv Invest is an indicator variable that equals one if a firm has above-median fraction of individual investors (i.e. below-median fraction of institutional investors) in a given year, and zero otherwise. Control variables are the same as in column 2 of Table 4, Panel A. All the independent variables at the firm-level are lagged by one year. County-level control variables are contemporaneous. All regressions include year and industry dummies, where industry is defined based on 2-digit SIC codes. All continuous variables are winsorized at 1% in both tails. Intercepts are not reported. Standard errors are corrected for heteroscedasticity and clustered at the firm level and t-statistics are in parentheses. *, **, and *** indicate statistical significance at 10%, 5% and 1% levels, respectively. (1) (2) (3) (4) Dependent variable R&D/Assets R&D/Assets R&D/Assets R&D/Assets LnCPRatio * (0.98) (1.25) (1.72) (0.86) LnCPRatio*SmallFirm 0.009*** (2.76) Small Firm * (-1.78) LnCPRatio*Low BE/PI 0.007* (1.84) Low BE/PI ** (-2.09) LnCPRatio*Low BE/PI Visible 0.009** (2.15) Low BE/PI Visible *** (-2.76) LnCPRatio*High Indiv Invest 0.010*** (3.01) High Indiv Invest * (-1.75) Year Fixed Effects Yes Yes Yes Yes Industry Fixed Effects Yes Yes Yes Yes Observations Adjusted R

60 Table 8 Innovative efficiency Panel A of the table reports the results of regressions of LnPatent, LnCitePerPat and LnTechAdjCites on LnCPRatio, (R&D/Assets) 12 and LnCPRatio*(R&D/Assets) 12 and other control variables. Panel B reports the results of regressions of Patent Applications, Citations Per Patent and Tech-Adjusted Citations on LnCPRatio, R&D 12 and LnCPRatio* R&D 12 and other control variables. Control variables are the same as in column 2 of Table 3, Panel A, but are not reported for brevity. All independent variables at the firm-level are lagged by one year except for R&D which is the average of the first and second lags. County-level control variables are contemporaneous. All regressions include year and industry dummies, where industry is defined based on 2- digit SIC codes. All regressions in Panel B include the interactions of R&D 12 with each industry dummy. All continuous variables are winsorized at 1% in both tails. Intercepts are not reported. Standard errors are corrected for heteroscedasticity and clustered at the firm level, and t-statistics are in parentheses. *, **, and *** indicate statistical significance at the 10%, 5% and 1% levels, respectively. Panel A: R&D Intensity and Patents and Citations (1) (2) (3) LnPatent LnCitePerPat LnTechAdjCites LnCPRatio 0.169*** 0.130*** 0.158*** (3.62) (3.35) (3.69) LnCPRatio*(R&D/Assets) ** ** (-2.02) (-0.80) (-2.15) (R&D/Assets) *** 2.372*** 1.895*** (6.21) (5.86) (5.63) Year Fixed Effects Yes Yes Yes Industry Fixed Effects Yes Yes Yes Observations Adjusted R Panel B: R&D Dollars and Patents and Citations (1) (2) (3) Patent Applications Citations Per Patent Tech- Adjusted Citations LnCPRatio 2.565*** 1.232*** 2.368** (2.68) (3.77) (2.47) LnCPRatio*R&D ** ** ** (-2.54) (-2.14) (-2.18) R&D *** 1.531** 1.537** (2.85) (2.30) (2.41) (R&D 12 *Industry dummies) Yes Yes Yes Year Fixed Effects Yes Yes Yes Industry Fixed Effects Yes Yes Yes Observations Adjusted R

61 Table 9 Cash holding for innovation This table reports the results of regressions of cash holding (cash and marketable securities divided by total assets) on our local gambling preference proxy, LnCPRatio and related interactions. All the variables are defined in the Appendix. Control variables include firm-specific variables from Opler et. al. (1999) and Bates, Kahle and Stulz (2009), and all our county-level controls. For brevity, we only report coefficient estimates for LnCPRatio, R&D/Assets, Industry Sigma, Acquisitions and the interactions of LnCPRatio with the latter three variables. Industry Sigma is the average volatility of cash flow for the industry defined by 2-digit SIC code in a given year. Acquisitions are dollars spent on acquisitions divided by total assets. All the independent variables at the firm-level are lagged by one year. County-level control variables are contemporaneous. Columns 1 and 2 show the results of regressions on the full sample and columns 3 and 4 show the results on sub-samples partitioned by industry innovativeness, where a 4-digit SIC code industry is defined as innovative if its average citations per patent exceed the median of all industries in a given year. All continuous variables are winsorized at 1% in both tails. All regressions include year and industry dummies where industry is defined based on 2- digit SIC codes. Intercepts are not reported. Standard errors are corrected for heteroscedasticity and are clustered at the firm level, and t-statistics are in parentheses. *, **, and *** indicate statistical significance at the 10%, 5% and 1% levels, respectively. (1) (2) (3) (4) Cash Cash Holding Holding (Full (Innovative sample) industries) Cash Holding (Full sample) Cash Holding (Noninnovative industries) LnCPRatio 0.016*** (3.56) (-0.11) (0.81) (-0.42) LnCPRatio*R&D/Assets 0.151*** 0.103** 0.105* (3.55) (2.35) (1.95) (1.21) LnCPRatio*Industry Sigma 0.239*** 0.185** 0.251*** (3.94) (2.27) (3.57) LnCPRatio*Acquisitions ** *** (-2.37) (-0.28) (-3.19) R&D/Assets 0.370*** 0.420*** 0.274*** 0.577*** (6.46) (7.29) (4.26) (6.36) Industry Sigma 0.256*** (6.80) (0.36) (-0.15) (0.37) Acquisitions *** *** *** *** (-13.08) (-4.98) (-4.39) (-3.17) Year Fixed Effects Yes Yes Yes Yes Industry Fixed Effects Yes Yes Yes Yes Observations Adjusted R

62 Table 10 CPRatio and firm valuation This table reports the results of regressions of Tobin s Q on LnCPRatio and two alternate measures of industry growth opportunities, namely industry innovativeness (Innovative Ind) or peer firm PE ratio (PeerPE), and their interactions. Innovative Ind equals one for 4-digit SIC industries whose citations per patent exceed the median for all industries in a given year; it equals zero for other industries. PeerPE is the natural log of total market value of equity divided by total earnings of all other firms in a firm s 2-digit SIC industry. All other variables are defined in the Appendix. All firm-level independent variables are lagged by one year. All models include contemporaneous county-level control variables (not reported for brevity). All continuous variables are winsorized at 1% in both tails. All regressions include year and 2-digit SIC industry dummies. Intercepts are not reported. Standard errors are corrected for heteroscedasticity and are clustered at the firm level, and t- statistics are in parentheses. *, **, and *** indicate statistical significance at the 10%, 5% and 1% levels, respectively. (1) (2) (3) (4) Dependent variable Tobin's Q Tobin's Q Tobin's Q Tobin's Q Innovative Ind 0.212*** (7.47) (0.63) PeerPE 0.056*** *** (4.12) (-2.60) LnCPRatio*Innovative Ind 0.147*** (3.13) LnCPRatio*PeerPE 0.116*** (4.00) LnCPRatio 0.103** ** (2.51) (-2.19) Ln(PPE/Emp) ** ** (-2.18) (-2.08) LnSales *** * (-2.64) (-1.96) Stock Return 6.825*** 7.136*** (26.85) (25.19) ROA *** *** (-11.88) (-11.79) # Business Segments *** *** (-8.66) (-8.69) Year Fixed Effects Yes Yes Yes Yes Industry Fixed Effects Yes Yes Yes Yes Observations Adjusted R

63 Figure 1: Catholics to Protestants ratio (CPRatio) and innovative activities. The figure shows the number of patents, citations per patent, technologyadjusted citations, R&D expenditure to total assets ratio, and cash to total assets ratio averaged within each quintile of Catholics to Protestants ratio in the county of firm headquarters. For each year of available data on CPRatio during our sample period (1980, 1990, and 2000), we sort firms into CPRatio quintiles, where quintile 1 is the lowest. Heights of the bars represent mean values within each quintile. Panel A shows the average number of patents applied for in a given year that are eventually granted, Panel B shows the average citations per patent, Panel C shows the average technologyadjusted citations, Panel D shows the average research and development expenditure normalized by total book assets, and Panel E depicts the average cash and marketable securities normalized by total assets.