Why Understanding Smoking Bans is Important for Estimating Their Effects: California’s Restaurant Smoking Bans and Restaurant Sales Robert K. Fleck Department of Agricultural Economics and Economics Montana State University Bozeman, MT 59717 phone: (406) 994-5603 e-mail: rfleck@montana.edu F. Andrew Hanssen Department of Agricultural Economics and Economics Montana State University Bozeman, MT 59717 phone: (406) 994-5616 e-mail: ahanssen@montana.edu January 15, 2007 Abstract: A large literature has sought to determine whether smoking bans help or hinder restaurants. Much of the literature improperly specifies its econometric equations, and thus mistakenly infers causality. Examining the relationship between restaurant smoking bans and restaurant revenues in 267 California communities, we reach two main conclusions. First, California’s municipal restaurant smoking bans are endogenous in a critical way – restaurant sales growth (or something correlated with restaurant sales growth) appears to cause restaurant bans, not vice versa. Consequently, failure to control properly for trends can produce spurious “evidence” of causation. Second, ban heterogeneity (e.g., state vs. local) can be exploited to sort out – or rule out – causal effects. In other words, pooling data and treating smoking bans implemented at different levels as homogenous (as many studies do) ignores an important source of information, and is likely to lead to erroneous conclusions. Our analysis holds lessons for the many studies that have examined the arguably more important question of how smoking bans affect smoking rates. We thank Joe Fitz and the California State Board of Equalization for sending us taxable sales data. We are also grateful to Len Casey and the American Nonsmokers’ Rights Foundation for generously sharing their data on smoking bans. For helpful comments, we thank Rosalie Pacula, Wally Thurman, and participants at the 2006 Western Economic Association Conference. The Department of Agricultural Economics and Economics at Montana State University provided the financial support for this research. I. Introduction A voluminous literature has sought to quantify the effect of smoking bans, typically on the quantity of smoking, but also on the performance of affected businesses.1 Various time periods have been explored, types of bans 1For reviews of studies on smoking bans and smoking behavior, see Chaloupka and Warner (1999, 37-38), Evans et al. (1999, 731-2), and Tauras (2006, 334). For a review of studies on restaurant bans and restaurant performance, see Scollo and Lal (2005). considered, locales examined, and specifications employed. Nonetheless, the majority of studies share two common failings. First, smoking bans of a given type (e.g., a restaurant ban) are treated as homogenous, when in fact, such bans have been enacted at various jurisdictional levels (municipality, county, or state), at various times, in various orders, and in various combinations (with other types of smoking bans, and with similar bans at other jurisdictional levels). Second, empirical specifications do not adequately control for preexisting trends in the dependent variable. As a result, many studies arrive at fundamentally wrong conclusions. We illustrate these points in an investigation of how California’s restaurant smoking bans have affected restaurant sales. We analyze the effect of smoking bans on restaurant sales for two reasons. First, the issue plays a role in the policy debate – more than 150 studies (many published in refereed journals and others in consulting reports undertaken for municipalities) have explored whether smoking bans help or hinder restaurants (see Scollo and Lal 2005). Yet few of these studies properly account for the effects of pre-existing trends (or changes in trends), for the closely related question of ban endogeneity, or for ban heterogeneity. Second, restaurant sales data are very rich: wellmeasured, available by municipality, by quarter, and over an extended period. Therefore, we can identify effects that would be obscured by cruder data (e.g., periodic survey responses), enabling us to draw lessons applicable to studies of the effect of other types of smoking bans (e.g., workplace bans) in other dimensions (e.g., smoking prevalence). We investigate 267 California municipalities for which the California State Board of Equalization provides quarterly taxable restaurant sales.2 We begin by employing “naive” specifications (which treat bans as homogenous and control simplistically for trends) representative of those typically estimated in the literature. The results of our naive specifications indicate that smoking bans have a large negative effect on restaurant revenues. However, when we examine municipal and state bans separately, we find a positive association between municipal bans and restaurant sales, and a negative association between the state ban and restaurant sales. Finally, when we control properly for preexisting trends in the dependent variable, we find no substantial evidence of any causal relationship running from either municipal bans or the state ban to restaurant sales. 2The Board of Equalization currently reports such data for the 272 cities with the highest total taxable sales. Our sample size is smaller because a full set of data (including taxable sales for restaurants, as well as the Census variables we use) is available for only 267 California cities. 2 Thus, we reach two main conclusions. The first is that California’s municipal restaurant smoking bans are endogenous in critical way – restaurant sales growth (or something correlated with restaurant sales growth) appears to cause restaurant bans, not vice versa. The second is that because communities are heterogenous, one must analyze the effect of a ban taking into account the level at which the ban was enacted. Pooling data and treating “smoking bans” as homogenous (as many studies do) is inappropriate – cities that impose bans on themselves are likely to differ systematically from cities that have bans forced on them by a higher jurisdiction (county or state). Relatedly, analyzing and/or instrumenting for smoking bans using state-level data (again, as many studies do) may fail to capture the relevant variation when the bans have force at a lower level of jurisdiction (e.g., a municipal ban). Our analysis, although focused on restaurant bans and restaurant sales, has implications for the more extensive literature on how smoking bans influence smoking behavior. Most importantly, it demonstrates the crucial nature of identifying and controlling for trends in the dependent variable (whether restaurant revenue, as in our data set, or smoking rates). Although a number of smoking ban studies explicitly acknowledge (and attempt to deal with) the possibility of endogeneity (e.g., Warner 1981; Chaloupka and Saffer 1992; Evans et al. 1999; Tauras 2006), their equations may be nonetheless improperly specified in the absence of appropriate controls for differentials in trends across the relevant units of observation. Furthermore, we find that demographic variables which might plausibly affect demand for restaurant meals and inspire smoking bans (e.g., income, education), do not explain the differential in restaurant sales trends, suggesting that more research into the causes of smoking bans is necessary to identify their true effects. More broadly, our conclusions are related to the general point that treating policy changes as natural experiments when using panel data can produce misleading results (e.g., Besley and Case 2000; Heckman and Vytlacil 2001; Bertrand et al 2004).3 In particular, our analysis demonstrates that in a panel setting, simple methods of addressing serial correlation (e.g., first order autocorrelation corrections, linear trends, time period dummies) are insufficient to identify the true effects of smoking bans. 3Also see Heckman’s (2000) more general discussion of the difficulties involved in efforts to infer causality. Heckman, Flyer, and Loughlin (2006) review the public health literature on the effects of cigarette advertising on youth smoking, and explain why the methods used in that literature fail to identify causal effects. 3 II. The Debate Over the Effects of Smoking Bans on Restaurants Many studies have investigated the influence of smoking bans on restaurants. The most complete overview of this literature is provided by Scollo and Lal (2005), who review 151 studies, about forty of which were published in peer-reviewed journals.4 The reviewed studies analyze a variety of “performance” measures: taxable sales (by far the most common), employment levels, number of establishments, number of permit applications, number of bankruptcies, unemployment insurance claims, self-reported patron intentions, and proprietor predictions. While the evidence is somewhat mixed, the majority of studies find bans to have either no effect or a positive effect on restaurant performance. Indeed, some notable studies find that smoking bans improve restaurant performance measures substantially.5 That an improvement in restaurant performance should follow a smoking ban may appear, at first blush, puzzling – why would forcing restaurants to do something that they could do voluntarily increase their revenues and/or reduce their costs? In fact, few of these studies offer any economic theory to support their empirical analyses, but rather simply assert that more non-smoking customers (or fewer smoking customers) will patronize a non-smoking restaurant, or that a non-smoking restaurant’s expenses will be lower, ceteris paribus.6 While either outcome is possible, neither requires coordinated action, so the contribution of a law is not clear. Nonetheless, a positive relationship between restaurant performance and smoking bans may indeed be consistent with rational profit maximizing behavior by restaurant owners. For example, there may be a prisoners’ dilemma-style problem among restaurants that renders unilateral action unprofitable but coordinated action profitable.7 4The majority of the peer-reviewed articles appeared in medical and health journals, such as the Journal of the American Medical Association, or the American Journal of Public Health. 5See, e.g., Glantz (2000), Alamar and Glantz (2004), and Cowling and Bond (2005). Glantz (2000) and Cowling and Bond (2005) focus on California. 6The restaurant will need to fumigate less frequently, for example, and will have fewer burns in carpets or curtains to mend (e.g., Alamar and Glantz 2004, 520-521). 7For example, if smokers have a very inelastic demand for dining out, but a strong willingness to travel within a city to smoke while they eat, a unilateral smoking ban may lose a restaurant more in smokers than it gains in nonsmokers, yet a coordinated ban may increase the number of nonsmoker patrons without substantially reducing the number of smoker patrons. Cowling and Bond (2005, 1280) describe a similar idea in the context of a free-rider problem. 4 Alternatively, rather than smoking bans increasing revenue, the direction of causality may run from revenues to bans.8 Or perhaps a third factor explains both restaurant sales growth and smoking bans. In short, the effect of restaurant smoking bans on restaurant performance remains an open – and empirical – question. We will analyze the effect of California’s comprehensive restaurant smoking bans on restaurant sales.9 California is an excellent candidate for empirical analysis because it has been a pioneer in the establishment of comprehensive private-place smoking bans, and because its pattern of passage of smoking bans provides econometrically useful variation in ban coverage and jurisdiction.10 Because California enacted smoking bans at several jurisdictional levels (state, county, municipality), a given California community may be covered by more than one ban. For example, 16 California cities put municipal bans in place before California’s statewide ban took effect in 1995, and another 19 did so afterwards (see on-line appendix B).11 Another reason to engage in econometric investigation is that casual perusal of the raw data (see on-line appendix B) reveals that some of the cities passing bans in the 1990s grew extremely rapidly in wealth and/or population during the 1990s – for example, Mountain View, Santa Clara, and San Jose during the Silicon Valley boom. Properly accounting for these and other factors that influence restaurant revenue growth is essential to determining a ban’s effects. III. Data Our data set is an unbalanced quarterly panel covering 267 California cities for the years 1980-2004. We 8For instance, if restaurants are a zero profit industry, restaurant owners will have less reason to lobby against a ban when market demand (and hence aggregate restaurant sales) is increasing – entry allows no more than a normal rate of return regardless of a ban. 9Our empirical analysis includes only comprehensive bans. We thus do not count laws that merely require separate smoking sections, or that have exemptions based on the size of the firm. Furthermore, a comprehensive restaurant ban need not imply a ban for bars that serve food. 10The first seventeen comprehensive workplace smoking bans, the first sixteen comprehensive restaurant smoking bans, and the first seven comprehensive bar smoking bans passed in the United States were enacted by California communities. As late as January 1, 2000, 84 of the nation’s 110 workplace bans, 32 of the nation’s 62 restaurant bans, and 20 of the nation’s 43 bar bans were in effect in California. An implication is that results of national studies of smoking bans during the 1990s will be heavily weighted towards California. The ANRF is the source for our data on comprehensive private-place smoking bans – see Section III. 11The online appendices referred to in this article may be found at . . . 5 include an observation in our econometric analysis if and only if all variables are available for that city and quarter. See on-line Appendix A for a list of data sources, variable definitions, and descriptive statistics. Our dependent variable will be Restaurant Revenue, the log of real taxable sales at eating and drinking establishments. The California State Board of Equalization publishes taxable sales data quarterly, and we were able to obtain quarterly eating and drinking establishment data for the years 1980-2004. Because the Board of Equalization reports data for cities with substantial taxable sales, the sample includes all of California’s large cities (i.e., none with a population above 125,000 in 1990 is omitted), but also includes some small-population cities with large amounts of taxable sales (business districts, industrial parks).12 Our data on smoking bans come from the ANRF’s (2006b) database on “100% Smokefree Ordinances” for cities, counties, and states. The organization classifies bans as “100% Smokefree” if they are “comprehensive”; i.e., the bans prohibit smoking completely rather than (for instance) simply requiring a nonsmoking section. On-line appendix B lists the California municipalities (and counties) classified by the ANRF as having passed comprehensive restaurant smoking bans, and the date the bans are said to have taken effect.13 The appendix also lists the date California passed a statewide restaurant smoking ban. For each municipality for each quarter, we define the variable City Ban to equal to 1 if a comprehensive municipal restaurant smoking ban is in effect, and 0 otherwise (and between 0 and 1 for the quarter in which the ban first took effect).14 We define the variable State Ban to equal 1 for all municipalities when the state ban is in effect (i.e., from the first quarter of 1995 onwards) and 0 otherwise. We define the variable Both Bans to equal 1 if a municipality’s restaurants are subject to both a city and state ban (i.e., it takes on the value of 0 for all communities prior to 1995, and 1 after 1995 for communities that also have municipal bans in place). Finally, we define the variable Any Ban to equal 1 if a municipality had a local ban in effect (City Ban = 1) and/or California’s state ban was in effect (State Ban = 1). Although the ANRF provides the best data available, there are two important caveats the reader should keep in 12We find that the results we present are not sensitive to the exclusion of these smaller cities; see Section IV. 13The county bans apply only to unincorporated sections of the relevant counties. 14For example, if a ban was in place for half of its initial quarter, the value of City Ban is set equal to one-half. 6 mind. First, because we rely upon the ANRF classifications, our estimated effects capture differences between cities coded by the ANRF as having “100% Smokefree” ordinances and other cities, with “other” including anything the ANRF does not classify as “100% Smokefree” (which could be a weak ban or no ban at all).15 As a result, there may be unobserved strength-of-ban differences among cities classified as having “100% Smokefree” bans, as well as among cities classified as not having “100% Smokefree” bans.16 Second, there are policies related to smoking that we cannot observe, most notably the degree to which smoking bans are actually enforced. That said, the main finding of this paper – namely that failing to control properly for trends can generate spurious results in line with those reported in the previous literature – clearly matters regardless of the degree to which the ANRF data actually measure the relevant variation in antismoking policy with perfect accuracy. Because restaurant revenues are likely to be influenced by an enormous variety of city-specific factors, and because all of the comprehensive restaurant smoking bans came into effect partway through our sample period, we include city fixed effects.17 To control for general trends in the economy (and in dining out), as well as year-to-year and seasonal fluctuations, we employ either (i) a linear trend with quarterly seasonal dummies or (ii) a full set of time dummies for the 100 quarters in our sample. Even with city fixed effects and quarterly time period dummies included, there remains the need to control for trends and fluctuations that are not common to the entire set of cities in the sample. For this reason, we use the variable 15Furthermore, the nature of collecting smoking ban data from city governments (as the ANRF does) introduces the potential for mis-classifications. Indeed, as Rosalie Pacula pointed out to us, this is apparent from the fact that the ANRF has revised its classifications over time as it has obtained more accurate information on the nature of the bans in place. Hence, the estimated effects of the smoking ban variables we use are most accurately interpreted as reflecting a comparison between cities reported as having “100% Smokefree” ordinances and all other cities in the data set. 16An econometric analysis of smoking bans for the entire U.S. would require careful attention to the fact that some weak state-level smoking bans preempt stronger local bans and, hence, actually reduce the restrictions on smoking in some localities (e.g., American Medical Association 2003; Centers for Disease Control 2006). For California, the state-level law did have some preemptive clauses for 1995-1997, according to the Centers for Disease Control’s (2006) interpretation of the law. Recall, however, that the state-level ban in place during those years is a “100% Smokefree Ordinance” (i.e., a strong ban) as coded by the ANRF (2006b). This suggests that, if the preemption of local laws had any effect, the effect likely would have been minor. In principle, we could test this empirically if we had data on the strength of local “100% Smokefree” ordinances passed before the state ban, but in practice we cannot (given the dichotomous coding of smoking bans). 17The literature on the determinants of state-level bans suggests a variety of potential causal factors, including cigarette consumption, income, political leanings, and tobacco production (e.g., Hersch, Del Rossi, and Viscusi 2004; Gallet, Hoover, and Lee 2006). Also see Shipan and Volden (forthcoming), who focus on the diffusion of antismoking policies from cities to states (i.e., the effect that the passage of city-level bans in a state has on the likelihood of that state passing a state-level ban). 7 Other Taxable Sales, defined as taxable sales from all sources other than eating and drinking establishments.18 Because Other Taxable Sales is available quarterly, the variable allows us to control for fluctuations and trends in economic growth specific to each city in the data set. We also use a group of city-specific trend controls consisting of demographic variables interacted with a linear time trend. The demographic variables are population, education (the percent of the population that has a high school degree or higher, the percent of the population that has a bachelor’s degree or higher), median household income, median home value, and the median residential rent paid.19 We include education because human capital may be related to both growth in restaurant revenue (highly educated people had more rapid increases in income over the sample period) and to smoking bans (highly educated people smoke less and may be more likely to lobby to prevent others from smoking). Median household income serves a similar purpose. We include home values to control for wealth-related factors not captured by education and income; home values may also proxy for opportunity costs related to the use of restaurant property. Median residential rent serves a similar purpose. We employ specifications including 1990 Census measures only, and including 2000 Census measures as well (thus allowing us to examine how trends in restaurant growth vary with inter-temporal changes in the Census variables).20 IV. Results Naive Specifications The objective of our empirical analysis is to illustrate the manner in which many studies of the effect of smoking bans have gone wrong. We will therefore begin by estimating what we refer to as “naive” specifications – 18In robustness tests, we controlled for seven separate components of Other Taxable Sales, by adding six variables (each a category of Other Table Sales) to a regression that already included Other Taxable Sales. The six categories were: apparel stores; food stores; building materials and farm implements; auto dealers and auto supplies (excludes service stations); service stations; non-restaurant retail stores. Data for each of these six categories were available for most of the cities in the California State Board of Equalization’s data set, though adding the six categories forced us to drop 1458 observations because of missing data. (We used the same data sources and units of measurement as we do for Other Taxable Sales.) 19All demographic variables are in logged form (as is the dependent variable). 20Other potentially useful variables are not available at the city level. For example, the prevalence of smoking among the population may influence the likelihood of adopting a ban and the effect of a ban on restaurants. Also, the effects of a ban on firmlevel revenue may differ between restaurants of different types – perhaps with positive effects for some and negative effects for others. But data on smoking prevalence and revenue by restaurant type (e.g., take-out versus full dining room with liquor) are available only at more highly aggregated levels. 8 naive in the sense that municipal and state bans are assumed identical in their effects (and can thus be pooled).21 These “naive” regressions provide a baseline for comparison with the previous literature, and with the specifications we will estimate subsequently. The first two columns of Table 1 present the results of estimating our naive specifications, using data for the entire 1980-2004 time period. Both regressions include city fixed effects, a linear time trend, and seasonal dummies, and correct for first-order autocorrelation; Regression 2 also includes Other Taxable Sales. The variable of interest is Any Ban, which equals 1 if the municipality was covered by either a state or local ban during that quarter. In both regressions, the coefficient on Any Ban is negative, statistically significant, and reasonably large (implying a roughly 4 percent reduction in restaurant revenues). These results are consistent with the hypothesis that restaurant smoking bans reduce restaurant sales. Is the coefficient on Any Ban picking up something other than just a smoking ban effect? In order to understand better what the coefficient on Any Ban is capturing, we will replace it with separate variables for state and municipal bans: City Ban (equal to 1 if the municipality has a city ban in effect), State Ban (equal to 1 during the state ban period), and Both Bans (equal to 1 if restaurants in the municipality are subject to both a city ban and the state ban). The results are shown in Regressions 3 and 4 of Table 1. The coefficient on State Ban is negative and statistically significant in both regressions, and of roughly the same magnitude as the Any Ban coefficient in Regressions 1 and 2. The coefficients on City Ban are negative, but less than one third the magnitude of those for State Ban and far from statistical significance. Finally, the coefficients on Both Bans are positive, and of magnitudes that offset the combined estimated effects of City Ban and State Ban.22 These coefficients would imply that the state-level ban alone (i.e., not in conjunction with a city-level ban) has a substantial negative effect on restaurant revenues, while having both a city ban and the state ban is roughly equivalent to having no ban at all. Clearly, this suggest the 21In fact, there are good reasons to expect that local bans and state bans will differ systematically in terms of their consequences. Notably, cities that choose to adopt any given policy may do so because those particular cities expect large benefits and/or small costs resulting from that policy. By contrast, a state-level policy is imposed on all cities in the state – not just those that would choose on their own to impose the policy. 22Recall that one should sum the coefficients for City Ban, State Ban, and Both Bans to find the estimated effect of having a city-level ban in place in the years after the state ban took effect. The estimated effect of having both bans in place is -0.00166 Regression 3 and 0.00201 in Regression 4. 9 possibility of specification error. In brief, it appears from Table 1’s results (if correct – a question we will address shortly) that the local and state-level smoking bans have different relationships to restaurant revenues, and that the negative coefficient on Any Ban in Regressions 1-2 reflects primarily the influence of the state ban on restaurants in municipalities that did not pass city bans.23 Therefore, our next step is to look at municipal and state bans independently. Municipal Bans We begin with municipal bans. We will start with specifications identical to those estimated in Regressions 1-2 of Table 1, with the difference that our variable of interest will now be City Ban (municipal bans) instead of Any Ban. Because our principal goal at this point is to test whether cities with bans are inherently different from cities without bans, we will estimate specifications that ignore the passage of the state ban. Regressions 1-2 of Table 2 shows the results. Interestingly, the sign of the City Ban coefficient has flipped from negative (in Table 1) to positive, with point estimates implying roughly 3 percent higher revenue (t=2.22) for the specification that includes Other Taxable Sales. Why does the City Ban coefficient change from negative to positive when the state-level ban is ignored? A possible reason is that the linear time trend (included in all the equations discussed so far) inadequately captures the difference between early and late years in the sample; this is a critical concern because controlling for the period of the state-level ban is econometrically identical to controlling for the last ten years of the sample. To account better for trends and fluctuations common across all cities in the sample, we estimate in Regressions 3-4 of Table 2 the same specifications as in Regressions 1-2, but replace the linear time trend with a full set of time period dummy variables.24 23By comparing Regression 1 to Regression 2, and by comparing Regression 3 to Regression 4, one can see that although Other Taxable Sales (coefficients of 0.39 with t>73) adds substantially to the explanatory power of Regressions 2 and 4, the coefficients on the smoking ban variables do not change dramatically with the addition of Other Taxable Sales. Furthermore, if we include the six additional controls for categories of taxable sales (described in Section III), we reach the same conclusion. With the additional controls added to Regressions 2 and 4, the coefficients are -.0275 (t=5.59) for Any Ban, -.0192 (t=0.97) for City Ban, -.0285 (t=5.74) for State Ban, and .0478 (t=2.46) for Both Bans. Thus, if the smoking ban coefficients in Table 1 are the result of an omitted variable bias, the omitted variable appears not to be highly correlated with various types of non-restaurant taxable sales. 24In other words, each of the 100 quarters in our data set has its own dummy variable. Note that with time period dummies included, the coefficient on City Ban measures the relative (to cities without bans) effect of the ban on revenues. For example, if restaurant revenues in a city are unchanged following passage of its municipal ban, but fall in other cities that do not have municipal bans, the coefficient on City Ban will be positive. 10 As Regressions 3 and 4 show, the coefficients on City Ban are larger when time period dummies are included and, regardless of whether we control for Other Taxable Sales, statistically significant.25 In short, controlling for crosssectional differences and fluctuations over time, municipal bans appear to be associated with between 4 and 5 percent higher restaurant revenues (relative to restaurants in non-ban cities). These specifications thus yield results similar to those of a number of previous studies that (as discussed in Section II) find that smoking bans have a positive and statistically significant effect on restaurant sales. Of course, correlation is not causation. To investigate the question of causation more thoroughly, we reestimate Regressions 3-4 separately for 1980-1994 (before the statewide ban) and 1995-2004 (after the statewide ban). The results are shown in Table 3. When the specifications are estimated on 1980-1994 data (Regressions 1-2), the coefficient on City Ban is negative and statistically insignificant. By contrast, when the specifications are estimated on 1995-2004 data (Regressions 3-4), the coefficient on City Ban is positive, nontrivial in magnitude (implying roughly 2 to 4 percent higher revenue) and statistically significant (t=2.51) in Regression 4 (i.e., when controlling for Other Taxable Sales). In short, the positive association between municipal restaurant smoking bans and restaurant sales apparent in the previous regressions (Table 2) appears to depend on including the period following the state ban. The fact that the municipal bans are associated with higher restaurant revenues in the later period (but not in the earlier period) may reflect the influence of the state ban; however, the results also suggest the possibility that restaurant sales may simply be trending differently in cities that passed municipal restaurant smoking bans than in other cities. In order to investigate this possibility, we define two new variables. The first is City Ban*Time, which is the dichotomous variable City Ban multiplied by a linear time trend (which runs from 1 to 100 over the sample period). City Ban*Time will capture trend differences between ban and non-ban cities at times when a municipal ban is in effect. The second is Ever Ban*Time, which is the product of a dummy indicating whether a city ever put into effect a restaurant smoking ban, multiplied by the linear time trend.26 Ever Ban*Time will capture long-term differences between ban and non-ban 25Again, this conclusion is robust with respect to including additional retail sales measures. Adding the six additional variables to Regression 2 yields a coefficient on City Ban of .0185 (t=1.69). Adding the six variables to Regression 4 yields a coefficient on City Ban of .0389 (t=3.75). 26Ever Ban is thus equal to 1 for ban-passing municipalities over the entire sample period (i.e., before and after their municipal bans were put into place). 11 cities – differences that are independent of whether or not the ban is in effect. The results of including these two additional variables are shown in Table 4. Regression 1 of Table 4 is estimated for the pre-statewide ban period (1980-1994). The coefficient on Ever Ban*Time is positive, statistically significant (t = 2.22), and non-trivial in magnitude (.33 percent greater revenue with each passing year, relative to other cities).27 In other words, restaurant sales growth was more rapid in cities that passed bans whether or not the ban was in effect. This is consistent with there being systematic pre-ban differences between ban-passing cities and non-banpassing cities. What about the effect of actually having a ban in place? That effect is captured by two variables: City Ban and City Ban*Time. The two variables have coefficients with offsetting signs (i.e., one positive, one negative); when the coefficients are interpreted jointly, the magnitude of the estimated effect is nontrivial (a roughly 3% decrease in revenue), but the coefficients on the two variables are far from statistical significance (individually and jointly).28 In short, the equation provides at most very weak evidence that city bans affect restaurant revenue. We next examine whether the pre-ban trend differences can be accounted for by a basic set of demographic variables plausibly related to restaurant revenue and smoking bans. Regression 2 includes six new variables, each an interaction of the linear time trend with a 1990 Census variable (population, high school degrees, bachelor’s degrees, income, home value, and rent).29 Regression 3 adds six more variables, each an interaction of the linear time trend with a 2000 Census variable. Thus, Regression 2 allows the growth in restaurant revenue to vary between cities in relation to the levels of demographic variables, and Regression 3 allows the growth in restaurant revenue to vary between cities in relation to levels and changes in the demographic variables over the decade of principal interest. Examining the coefficient estimates from Regressions 2 and 3 of Table 4, it can be seen that the levels and changes in the demographic variables do not explain the differences in trends between cities that passed bans and cities 27Over the period of the panel (100 quarters) an exogenous difference in trends of this magnitude could lead to a substantial overestimation of the effect of smoking bans if a naive estimating framework were employed. 28The joint effect ranges from -0.034 in 1990 (when the first city ban took effect) to -0.031 in 2005 (the end of our sample). In an F-test of joint significance, p=.48. 29See Section III and on-line Appendix A for descriptions of these variables. 12 that did not. In fact, adding the full set of demographic controls (Regression 3) more than doubles the size of the estimated coefficient on Ever Ban*Time (with t=5.26 on the coefficient in Regression 3). In sum, whatever caused the differences in trends does not appear to be highly correlated with any of a quite general set of demographic variables. And having a smoking ban in effect does not appear to have any statistically significant association with restaurant sales.30 Regressions 4-6 in Table 4 repeat the specifications used in Regressions 1-3, but use data from 1995-2004 to estimate the equations. As with the earlier period, the results show a substantial and statistically significant greater growth rate in ban-passing cities, whether or not the bans were in effect.31 Once again, this suggests a systematic difference in some exogenous influence(s) on restaurant revenue – influences that studies employing naive specifications would interpret as the causal effect of smoking bans on restaurant revenue.32 As robustness test, we examined the sensitivity of our results the exclusion of cities with small populations. This is potentially important because the sample consists of municipal data assembled by the California State Board of Equalization on the basis of the size of the taxable sales, not the size of the municipality. Thus, our sample includes some small cities that are potentially unrepresentative.33 When we re-estimated our equations excluding cities with fewer than 20,000 people in 1990, the results were very similar to those shown in Tables 3-6. 30As an additional test of whether cities that will pass bans in the future differ from cities that will not pass bans in the future, we re-estimated Regression 3 allowing the coefficient on Ever Ban*Time to differ between early and late passers of city bans (i.e., cities that passed bans prior to the passage of the state ban and cities that passed bans after the passage of the state ban). The results are quite striking: The coefficient for early passers is 0.00094 (t=1.71) and the coefficient for late passers is 0.00260 (t=5.69). Thus, the cities that passed bans after the passage of the state ban had rapid restaurant revenue growth prior to the passage of the state ban. As for the evidence that having a city ban in effect matters, it remains as weak as in Regression 3. 31As for Regression 3, we re-estimated Regression 6 allowing the coefficient on Ever Ban*Time to differ between early and late passers of city bans. Again the results are quite striking: The coefficient for early passers is 0.00318 (t=3.89) and the coefficient for late passers is 0.00282 (t=3.47). Thus, early passers had rapid restaurant revenue growth even after the state ban was in effect. And so did late-passers, whether or not the city bans had yet been passed. The evidence that actually having a city ban in effect matters remains as weak as in Regression 6. 32As a robustness test, we expanded the set of controls in Regressions 3 and 6 by adding the six additional controls for categories of taxable sales (described in Section III). The results were similar, although the adding the additional controls caused the estimated coefficients on Ever Ban*Time to decline somewhat in magnitude: 0.00142 (t=4.25) for the 1980-1994 period; 0.00174 (t=3.28) for the 1995-2004 period. 33For example, the City of Industry had only 580 people in 1990, and 777 in 2000, yet averaged over $19 million dollars per quarter in taxable sales at eating and dining places between 1980 and 2004 (see on-line Appendix A for descriptive statistics for population and other variables). 13 The State Ban In sum, there is little evidence of a causal relationship running from municipal restaurant smoking bans to restaurant revenues. How about the state ban? To investigate, we limit our sample to cities that never passed municipal bans (we have already established that ban-passing cities are systematically different from other cities). We include fixed effects, a linear trend, seasonal dummies, and controls for other taxable sales. Our variable of interest is State Ban, which takes on the value of 1 for all cities from the first quarter of 1995 onwards. The results are shown in column 1 of Table 5. The coefficient on State Ban is negative, statistically significant (t=6.78), and substantial (implying that the state ban is associated with roughly a 4 percent reduction in restaurant sales). This is consistent with the hypothesis that restaurant revenues in cities without a self-imposed ban were adversely affected by California’s statewide ban. But is the relationship causal? We will again investigate the role of pre-ban trends. Because the state ban affects all municipalities at the same time, we will look directly at coefficients on time period dummies. We define sixteen year dummies, running from 1989 (the year before California’s first municipal smoking ban) through 2004 (the end of the sample period). If the negative coefficient on State Ban truly reflects the effect of California’s state restaurant smoking ban on restaurant revenues, we should find evidence of a break – a deviation from the pre-state-ban trend – and that break should occur at the time the state ban took effect. Regression 2 shows the results. Although the ban year (1995) is indeed associated with a negative deviation from the pre-1989 restaurant revenues trend, so are the preceding six years. And the deviations increase nearly monotonically from 1989 right through 2004. Thus, the coefficient on State Ban in Regression 1 provides (at best) very dubious evidence of causality. That said, the dubious nature of the evidence only becomes apparent when one examines the pre-ban deviation from the trend.34 What are the general lessons here? First, Tables 1-4 taken as a whole show that municipal ban-passing cities differ systematically (i.e., even when the ban is not in effect) from other cities in ways that regressions with city fixed 34Repeating our earlier robustness tests, we added the six additional controls for categories of taxable sales, and we examined the sensitivity of our results the exclusion of cities with fewer than 20,000 people (1990 Census). In both tests, the results were similar to those shown in Table 5. 14 effects, time period dummies, and panel controls for other types of retail sales (and even demographic variables) will not fully capture. Second, the recognition that bans differ in type is essential for efforts to understand the ways in which systematic differences across units of observation can produce misleading results. Our analysis illustrates the importance of both points – we find that differences in trends between ban passers and non-ban passers, along with changes in trends over time, can generate spurious evidence that city-level bans increase restaurant revenue (Tables 2 and 3) and state-level bans decrease restaurant revenue (Table 5). V. Discussion Implications for Studies on the Effect of Smoking Bans on Restaurants One of the best of the many studies relating restaurant bans to restaurant revenues is by Bartosch and Pope (2002), who examine the effect of municipal bans on taxable restaurant meal receipts in Massachusetts, using a sample running from 1992 to 1998.35 Bartosch and Pope employ specifications containing various controls, including fixed effects and time trends. Their variable of interest is a linear trend for the months following the implementation of a smoking ban. They find the coefficient on that variable to be statistically insignificant, and conclude that smoking bans therefore did not affect restaurant sales. The results of our analysis suggest a potential problem with Bartosch and Pope’s specification: It fails to account for pre-ban trends in restaurant sales. What if restaurant sales in ban-passing cities were trending differently than in non-ban cities before the bans were passed? For example, if the ban-passing cities were experiencing faster growth in restaurant revenues before the bans went into effect (as we find for California), and the result of the bans was to cause revenue growth to fall to the level of that of non-ban cities, Bartosch and Pope’s specification would pick up the same statistically insignificant coefficient on the post-ban time trend it does. Yet the true effect of restaurant bans on sales growth would be negative. In short, although Bartosch and Pope’s general conclusion may be correct, their analysis is insufficient to indicate whether or not Massachusetts’ restaurant smoking bans affect restaurant revenues. Alamar and Glantz (2004) develop an interesting alternative approach, in which they examine whether 35Some Massachusetts municipalities enacted comprehensive restaurant smoking bans, some enacted laws simply requiring separately ventilated smoking rooms, and some enacted no private place antismoking laws at all. 15 restaurant smoking bans affect the market value of restaurants (when restaurants are put up for sale). Alamar and Glantz find that restaurant smoking bans are positively associated with the market value of restaurants, controlling for revenue, and conclude that smoking bans therefore make restaurants more valuable. Yet again, the potential problem is a lack of attention to preexisting trends, this time in property values. If communities that saw faster growth in property values were also more likely to pass restaurant smoking bans, Alamar and Glantz’s specification would produce the positive association between bans and restaurant market values that they find, but there need not be anything causal in the relationship. Particular concern is warranted because the time period covered, combined with the geographical concentration of smoking bans, renders the need to control for trends in property values acute. Between 1991 and mid-2002 (the period of the Alamar and Glantz study), the vast majority of local comprehensive restaurant smoking bans (99 out of 118) were implemented in two states, California and Massachusetts; furthermore, as discussed, California passed a state ban in 1995 which prohibited smoking at restaurants in every California community. Both California and Massachusetts saw enormous run-ups in property values over the same period. Because Alamar and Glantz do not control for trends in property values, it is impossible to tell whether they are picking up the effect of smoking bans, or simply the fact that communities that enacted restaurant smoking bans experienced more rapid growth in (or even just higher levels of) property values than did communities that did not enact bans. Implications for Studies on the Effect of Smoking Bans on Smoking A number of studies have investigated the effect of smoking bans on smoking behavior.36 Our analysis suggests that the results of such studies should be treated with caution. To illustrate, consider Evans et al. (1999), perhaps the best extant study on the effect of workplace bans on smoking. Evans et al. analyze a longitudinal data set of responses to 1991 and 1993 NHIS surveys, and conclude that smoking rates are lower in firms that ban smoking (taking self-selection of workers into account). However, because Evans et al. use longitudinal data, it impossible for them to establish whether or not firms with workplace smoking bans were experiencing a differential decline in 36The idea is that a ban raises the cost of smoking (by making smoking less convenient) and thus reduces the quantity of smoking. For overviews of the literature, see Chaloupka and Warner (1999, 37-38), Evans et al. (1999, 731-2), and Tauras (2006, 334). Also see, e.g., Saffer and Chaloupka (2000) on the effects of advertising bans. 16 smoking before the bans were implemented. Furthermore, Evans et al. note that in their sample, there was a large increase between 1991 and 1993 in the fraction of workers employed by firms banning smoking in all work areas (from 61.7 percent in 1991 to 73.4 percent in 1993). Over the same period, the number of U.S. communities with comprehensive workplace smoking bans increased from six to 34, and 26 of the 28 new bans were enacted in California. Therefore, Evans et al.’s results may be heavily influenced by California in a manner that their simple inclusion of state cigarette taxes (the study’s only location-based measure) will not capture. In short, although Evans et al.’s conclusions may be correct, additional tests are needed before one can confidently infer causality from their results.37 What Causes a Restaurant Smoking Ban? The evidence from our analysis is consistent with increases in restaurant sales leading to restaurant smoking bans, not vice versa. But what is the causal mechanism? There are two principal possibilities: Restaurant sales growth affects smoking bans directly (i.e., is causal), or a third factor influences both smoking bans and restaurant sales growth. While a thorough investigation of this issue is beyond the scope of our analysis, we can speculate. First, it is possible that restaurant sales growth causes restaurant smoking bans directly, by reducing the incentive of restaurant owners to lobby against bans. Restaurant owners and restaurant associations have typically been the most vocal opponents of restaurant smoking bans.38 If restaurateuring is a zero profit industry with a perfectly elastic supply of inputs, fast revenue growth – i.e., rapid increases in market demand – simply leads to entry. Even if a smoking ban reduces restaurant sales growth somewhat as compared to no ban, an incumbent restaurant owner in a rapid growth area will be affected (if at all) only in the short run (i.e., during the time it takes the rate of entry to adjust). If the ban leaves restaurant rents intact, the only long run effect of a ban would be less entry than would have otherwise 37Also note that similar econometric problems may arise when estimating the effects of advertising bans. If, for example, cigarette advertising bans tend to be adopted by countries that already have rapidly decreasing cigarette consumption, then regressions with year dummies and country fixed effects (e.g., Saffer and Chaloupka 2000) may yield erroneous evidence that advertising bans reduce cigarette consumption. Naturally, the bias would be in the other direction if cigarette advertising bans tend to be adopted by countries with rapidly increasing cigarette consumption. Whether such a bias exists is an important question for future research. 38See, for example, the discussion and citations in Bartosch and Pope (2002, ii38). smoke.org). 17 See also the ANRF website (www.no- occurred. By contrast, if total restaurant revenues are growing very slowly (or are stagnant or declining) and entry requires incurring a sunk cost, a restaurant smoking ban that reduces market demand to the point where prices fall will be opposed by incumbent firms. Alternatively, rather than there being a direct causal relationship between bans and revenue, an omitted factor may explain both. Human capital is a plausible candidate – as we suggest in Section III, human capital may affect both the likelihood of passing a ban (if wealthy, highly educated people are disproportionately opposed to other people smoking and more effective at lobbying) and the frequency of dining out. We found trend differences between ban and non-ban cities to be robust to the inclusion of obvious demographic proxies for human capital (e.g., education, income), but a third factor, imperfectly correlated with these measures, may cause smoking bans. For example, restaurant smoking bans may be more likely in communities with residents who are health conscious or more engaged in outdoor sports; alternatively, political variables, such as voter activism or the aggressiveness of antismoking groups may influence the passage of restaurant smoking bans (see Chaloupka and Saffer 1992; Hersch, Del Rossi, and Viscusi 2004; Shipan and Volden forthcoming). Such factors are likely related to education and income, but a key component of the relationship may not be captured by the Census variables we employ.39 VI. Conclusion In this paper, we have investigated the relationship between comprehensive restaurant smoking bans and restaurant revenues, examining 267 California cities over a twenty-five year period. Our findings provide evidence that localities passing bans differ exogenously from localities that do not, and that they differ in a manner that confounds naive econometric attempts to estimate the causal effects of smoking bans. For California, these exogenous differences remain even after controlling for city fixed effects, time dummies, and demographic factors. If in this respect California is representative of the country as a whole, then studies based on data sets from other states will have similar problems. And, of course, if California and other states that led the way in passing bans – notably Utah and Massachusetts – are not representative (especially in terms of rapid growth in the restaurant industry during the 1990s), then inference using 39For example, health conscious people may tend to be wealthy, but wealthy people as a whole may not be health conscious. 18 state-level data will face obstacles similar to those we demonstrate for city-level data. In this light, properly estimating the effects of smoking bans (on smoking or on restaurant revenues) requires a more thorough understanding of why some jurisdictions pass bans while others do not. Bibliography Alamar, Benjamin C., and Stanton A. Glantz. 2004. “Smoke-Free Ordinances Increase Restaurant Profit and Value”, Contemporary Economic Policy, 22: 520-525 American Medical Association. 2003. “Preemption: Taking the Local Out of Tobacco Control.” (http://www.smokelessstates.org/downloads/2003_Preemption.pdf) American Nonsmokers’ Rights Foundation. 2006a. “Percent of U.S. State Populations Covered by Local or State 100% Smokefree Air Laws.” April 17, 2006. http://www.no-smoke.org/pdf/percentstatepops.pdf. American Nonsmokers Rights Federation (2006b). “United States Population Protected by 100% Smokefree Workplace and/or Restaurant and/or Bar Laws.” April 17, 2006. http://www.no-smoke.org/pdf/EffectivePopulationList.pdf Bartosch, W.J., and G.C. Pope. 2002. “Economic Effect of Restaurant Smoking Restrictions on Restaurant Business in Massachusetts, 1992 to 1998", Tobacco Control, 11 (Supplement II): ii38-ii42 Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan. 2004. “How Much Should We Trust Differences-inDifferences Estimates?” Quarterly Journal of Economics, 119: 249-275 Besley, Timothy, and Anne Case. 2000. “Unnatural Experiments? Estimating the Incidence of Endogenous Policies.” Economic Journal 110: F672-F694. California State Board of Equalization. 1980-2004. Taxable Sales in California. Centers for Disease Control. 2006. “State Comparison Report.” (http://apps.nccd.cdc.gov/statesystem/) Chaloupka, Frank J., and Kenneth E. Warner. 1999. “The Economics of Smoking”, NBER Working Paper 7047 Chaloupka, Frank J., and Henry Saffer. 1992. “Clean Indoor Air Laws and the Demand for Cigarettes”, Contemporary Policy Issues, 10: 72-83 Cowling, David W., and Philip Bond. 2005. “Smoke-Free Laws and Bar Revenues in California – the Last Call”, Health Economics 14: 1273-1281 Economic Sciences Corporation. 2006. California Database. Evans, William N., Matthew C. Farrelly, and Edward Montgomery. 1999. “Do Workplace Smoking Bans Reduce Smoking?”, American Economic Review, 89: 728-747 Gallet, Craig A., Gary A. Hoover, and Junsoo Lee. 2006. “Putting Out Fires: An Examination of the Determinants of 19 State Clean Indoor-Air Laws.” Southern Economic Journal 73:112-124. Glantz, Stanton A. 2000. “Effect of Smokefree Bar Law on Bar Revenues in California.” Tobacco Control (letters) 9:111-112. Heckman, James J. 2000. “Causal Parameters and Policy Analysis in Economics: Retrospective.” Quarterly Journal of Economics 115: 45-97. A Twentieth Century Heckman, James J, Fredrick Flyer, and Colleen Loughlin. 2006. “An Assessment of Causal Inference in Smoking Initiation Research and a Framework for Future Research.” Working paper, University of Chicago. Heckman, James J., and Edward Vytlacil. 2001. “Policy-Relevant Treatment Effects.” American Economic Review (Papers and Proceedings) 91: 107-111. Hersch, Joni, Alison F. Del Rossi, and W. Kip Viscusi. 2004. “Voter Preferences and State Regulation of Smoking.” Economic Inquiry 42:455-468. Saffer, Henry, and Frank Chaloupka. 2000. “The Effect of Tobacco Advertising Bans on Tobacco Consumption.” Journal of Health Economics 19:1117-1137. Scollo, Michelle, and Anita Lal. 2005. “Summary of Studies Assessing the Economic Impact of Smoke-Free Policies in the Hospitality Industry – includes studies produced to July 2005", VicHealth Centre for Tobacco Control, Melbourne, Australia Shipan, Charles R., and Craig Volden. Forthcoming. “Bottom-Up Federalism: The Diffusion of Antismoking Policies from U.S. Cities to States.” American Journal of Political Science. State of California Division of Labor Statistics & Research. 2006. Consumer Price Index Historical Data Series. http://www.dir.ca.gov/DLSR/CPI/CPIHistDataSeries.xls (accessed May 25, 2006) Tauras, John A. 2006. “Smoke-Free Air Laws, Cigarette Prices, and Adult Cigarette Demand”, Economic Inquiry, 44: 333-342 United States Census Bureau. 2006. Data Sets. http://factfinder.census.gov/servlet/DatasetMainPageServlet (accessed May 17, 2006) U.S. Department of Health and Human Services. 1986. Smoking and Health: A National Status Report, Public Health Service, Centers for Disease Control, Washington, D.C. Warner, Kenneth E. 1981. “State Legislation on Smoking and Health: A Comparison of Two Policies”, Policy Sciences, 13: 139-152 20 Table 1: Naive Specifications by Type of Ban, 1980-2004 Any Ban 1 2 3 4 Restaurant Restaurant Restaurant Restaurant Revenue Revenue Revenue Revenue 1980-2004 1980-2004 1980-2004 1980-2004 -0.0459488 - - -0.011898 0.0375595 (7.44) (7.05) City Ban 0.0132173 (0.54) (0.55) - - 0.0476426 0.0391737 (7.64) (7.29) 0.0591985 0.0530823 (2.53) (2.53) State Ban Both Bans Other Taxable Sales Time 0.3930517 0.3929712 (73.86) (73.90) 0.0063177 0.0037068 0.0062685 0.0036693 (39.00) (31.10) (38.48) (30.62) City Fixed Effects included included included included Quarterly Seasonal included included included included Correction for yes yes yes yes Autocorrelation ρ = 0.83 ρ = 0.74 ρ = 0.83 ρ = 0.74 R2 (within) 0.1156 0.3238 0.1165 0.3253 obs 24193 24193 24193 24193 Dummies 21 cities 267 267 267 267 quarters 99 99 99 99 t statistics in parentheses. Dependent variable and independent variables (except smoking ban variables, Time, and dummies) measured in logs. 22 Table 2: City Ban Results, 1980-2004 City Ban 1 2 3 4 Restaurant Restaurant Restaurant Restaurant Revenue Revenue Revenue Revenue 1980-2004 1980-2004 1980-2004 1980-2004 0.0278366 0.0283496 0.0485734 0.044877 (1.64) (2.22) (3.12) (3.71) Other Taxable Sales Time 0.3927224 0.3259433 (73.43) (53.6) 0.0055748 0.0030987 (41.18) (34.57) City Fixed Effects included included Quarterly Seasonal Dummies included included Time Period Dummies Correction for Autocorrelation included included included included yes yes yes yes ρ = 0.83 ρ = 0.75 ρ = 0.82 ρ = 0.74 R2 (within) 0.1129 0.317 0.2227 0.3705 obs 24193 24193 24193 24193 cities 267 267 267 267 quarters 99 99 99 99 t statistics in parentheses.. Dependent variable and independent variables (except smoking ban variable, Time, and dummies) measured in logs. Table 3: City Ban Results by Time Period, 1980-1994, 1995-2004 City Ban Other Taxable Sales 1 2 3 4 Restaurant Restaurant Restaurant Restaurant Revenue Revenue Revenue Revenue 1980-1994 1980-1994 1995-2004 1995-2004 -0.0003415 -0.0169005 0.0214885 0.0356972 (0.01) (0.74) (1.28) (2.51) 0.3071133 0.4372404 23 (38.44) (55.14) City Fixed Effects included included included included Time Period Dummies included included included included yes yes yes yes ρ = 0.67 ρ = 0.58 ρ = 0.70 ρ = 0.64 R2 (within) 0.2778 0.3912 0.2551 0.4063 obs Correction for Autocorrelation 13782 13782 10156 10156 cities 255 255 267 267 quarters 59 59 39 39 t-statistics in parentheses. Dependent variable and independent variables (except smoking ban variable and dummies) measured in logs. Table 4: Differences in Trends Between Ban-Passing and Other Cities 1 2 3 4 5 6 Restaurant Revenue Restaurant Restaurant Restaurant Restaurant Restaurant Revenue 1995-2004 1980-1994 City Ban City Ban*Time Ever Ban*Time Other Taxable Sales Population 1990*Time High School 1990*Time College 1990*Time Income 1990*Time Rent 1990*Time Home Value 1990*Time Revenue Revenue Revenue Revenue 1980-1994 1980-1994 1995-2004 1995-2004 -0.0367351 -0.0088357 -0.0217451 0.0129353 0.002689 0.0181047 (0.72) (0.17) (0.43) (0.36) (0.08) (0.51) 0.0000554 0.0001431 0.0000949 -.0000762 0.0000317 -0.0000609 (0.07) (0.17) (0.12) (0.18) (0.08) (0.15) 0.0008334 0.0013693 0.0019311 0.0024648 0.0024137 0.0030000 (2.22) (3.75) (5.26) (3.77) (3.72) (4.65) 0.3061662 0.2918734 0.2773008 0.4392894 0.3706293 0.3555628 (38.27) (35.83) (33.80) (55.48) (41.58) (39.35) -0.0010565 -0.0077537 0.0000188 -0.0031905 (8.93) (10.22) (0.12) (5.42) -0.0005624 -0.0081403 0.0019615 0.0092538 (0.49) (3.21) (1.54) (2.71) -0.0010702 0.0001023 -0.0018802 -0.0086664 (2.17) (0.11) (4.48) (6.91) 0.006443 0.0025431 0.0031126 -0.0007902 (7.13) (1.25) (2.73) (0.31) 0.0005718 0.0032002 0.00079 0.0015791 (0.42) (1.61) (0.44) (0.63) -0.0027115 0.0015439 -0.0016259 0.0037957 24 (4.65) (1.89) Population 2000*Time High School 2000*Time College 2000*Time Income 2000*Time Rent 2000*Time Home Value 2000*Time (2.35) (3.55) 0.0067531 0.0033159 (9.09) (5.69) 0.0078655 -0.0087546 (2.71) (2.57) -0.0003433 0.0076886 (0.32) (6.44) 0.0026646 0.0070306 (1.33) (2.73) -0.0011769 -0.0053449 (0.62) (2.16) -0.0042849 -0.0055881 (5.48) (5.48) City Fixed Effects included included included included included included Time Period Dummies included included included included included included Correction for Autocorrelation yes yes yes yes yes yes ρ = 0.58 ρ = 0.60 ρ = 0.54 ρ = 0.64 ρ = 0.63 ρ = 0.61 R2 (within) 0.3915 0.41 0.4271 0.4097 0.4302 0.4547 obs 13782 13782 13782 10156 10156 10156 cities 255 255 255 267 267 267 quarters 59 59 59 39 39 39 t statistics in parentheses. Dependent variable and independent variables (except smoking ban variables, Time, and dummies) measured in logs. Table 5: Effects of the State-Level Ban in Cities Without Municipal Bans State Ban 1 2 Restaurant Revenue Restaurant Revenue 1980-2004 1980-2004 -0.0378087 (6.78) Other Taxable Sales Dummy 1989 0.4104262 0.4012202 (73.82) (71.79) -0.0354792 (5.53) Dummy 1990 -0.0769534 (9.02) Dummy 1991 -0.0899179 (8.89) Dummy 1992 -0.123288 25 (10.74) Dummy 1993 -0.1531029 (12.01) Dummy 1994 -0.1903584 (13.64) Dummy 1995 -0.2192906 (14.44) Dummy 1996 -0.2404854 (14.67) Dummy 1997 -0.261195 (14.86) Dummy 1998 -0.2632027 (13.99) Dummy 1999 -0.2735062 (13.67) Dummy 2000 -0.2952737 (13.92) Dummy 2001 -0.3019641 (13.43) Dummy 2002 -0.293944 (12.37) Dummy 2003 -0.3019628 (12.07) Dummy 2004 -0.3122987 (11.88) Time 0.0034722 0.0072216 (28.04) (21.21) City Fixed Effects included included Quarterly Seasonal Dummies included included Correction for Autocorrelation yes yes ρ = 0.74 ρ = 0.72 R2 (within) 0.3348 0.3557 obs 21420 21420 cities 238 238 quarters 99 99 26