The effects of recreational cannabis access on labor markets: evidence from Colorado

In 2014, the first recreational cannabis dispensaries opened in Colorado. Today, recreational cannabis dispensaries outnumber Starbucks and McDonald's locations combined (Colorado Department of Revenue [CDOR], 2019). As the number of recreational dispensaries grew between 2014 and 2019, so did the value of their sales, dwarfing those of the existing medical cannabis market. When comparing recreational and medical cannabis sales during this time period (Figure 1A), medical cannabis sales stayed around $35 million, while recreational cannabis sales increased 10-fold, from $10 million to $110 million.

These are retail (recreational) and medical cannabis sales, as reported to the Colorado Department of Revenue.

Figure 1

As in most other legal recreational cannabis markets, legalization in Colorado occurred at the state level, but individual counties and municipalities can choose to ban cannabis businesses, from growing facilities through retail dispensaries. Of the 64 counties in our sample, only 37 experienced entry of dispensaries between 2014 and the end of our sampling period in 2018. Many local areas continue to ban cannabis, including El Paso County, in which Colorado Springs, the second largest city in the state, is located.

https://www.visitcos.com/travel-tools/local-services-information/recreational-marijuana/. Accessed 07/16/21.

This paper is the first to analyze the effects of recreational dispensary access on labor market outcomes beyond workers’ compensation benefits, to use a county-level approach to explore the effect, and to identify demand-side effects, rather than supply-side effects. Using monthly data from the Local Area Unemployment Statistics (LAUS) and quarterly data from the Quarterly Census of Employment and Wages (QWEC) for the period 2011–2018, we test for changes in the unemployment rate, the number of people unemployed, the total labor force, the number of employees (overall and by major industry), and wages (overall and by major industry) in Colorado counties in response to the entry of recreational cannabis dispensaries. We investigate the effects on labor market outcomes by exploiting quasi-random variation in the timing of the entry of recreational cannabis dispensaries at the county level. With a difference-in-differences (DID) research design, we compare labor market outcomes within and across counties and across industries, before and after counties began selling recreational cannabis, controlling for the county, month, and year fixed effects and the number of medical cannabis patients.

We find that the opening of recreational cannabis dispensaries is associated with a decrease in the number of unemployed people and the unemployment rate, with no significant effect on the total labor force. We further find an increase in the total number of employees overall, with effects being the strongest in the manufacturing sector. For wages, we find no significant effect on average weekly wages overall or by industry. Our results suggest that labor demand effects dominate any impact on labor supply. Because of potential spillovers across counties within a state and possible endogeneity in dispensary entry, we corroborate the robustness of our results in multiple ways, including capitalizing on institutional factors limiting entry, attempting to predict pretreatment using pre-trends, assessing the existence and magnitude of county-level spillovers using a nearest-neighbor, distance-based approach, and using the generalized synthetic control (GSC) method presented by Xu (2017).

That the creation of a newly legal industry affects local labor markets, either directly or indirectly, is supported by the broader literature on new industry entry. Benefits associated with the arrival of new casinos to local areas include doubling of earnings in the local gambling industry and indirect spillover effect on employment growth in closely related local industries (Cotti, 2008; Humphreys and Marchand, 2013). With respect to abrupt increases in the size of legal industries, a study found that the opening of an ethanol plant in a county increases employment opportunities, labor earnings, and demand for land and housing, with spillover effects on other industries (Low and Isserman, 2009), while Black et al. (2005) found that spill-over effects from coal booms increase employment but not wages in the construction and services sectors. The empirical evidence from these two industries supports that positive benefits also arise in the cannabis industry, with a direct increase in labor demand within the cannabis industry and spillovers across other industrial sectors.

The fact that demand-side effects seem to dominate labor supply-side effects matches a literature that finds ambiguous and limited effects on health and productivity. Older studies using survey data have typically identified negative associations between cannabis use and health-and-productivity-related outcomes (Van Ours, 2007; Hanson et al., 2010; Van Ours et al., 2013; Volkow et al., 2014; Williams and Van Ours, 2020; Winward et al., 2014; Kaestner, 1994; Register and Williams, 1992; Irons et al., 2014), or no association (MacDonald and Pudney, 2000). More recent studies of medical cannabis legalization indicate potential positive health and productivity effects, often through substitution away from more harmful substances (Ullman, 2017; Nicholas and Maclean, 2019; Bonn-Miller et al., 2007; Vigil et al., 2017; Doremus et al., 2019; Bradford et al., 2018; Bradford and Bradford, 2018; Anderson et al., 2018, 2014; Chu, 2015; Baggio et al., 2018; Anderson et al., 2013; Choi et al., 2019).

The limited literature on how medical marijuana laws (MMLs) affect labor supply shows mixed results as well. On the one hand, medical cannabis dispensaries’ entry may have potential positive employment and earnings effects, at least in older populations (Nicholas and Maclean, 2019), and for more extreme outcomes, such as workplace deaths (Anderson et al., 2014), workers’ compensation claims (Ghimire and Maclean, 2020), and job absences (Ullman, 2017). On the other hand, Sabia and Nguyen (2018) find a small decrease in wages among males aged 20–39 years. Older studies, using survey data on cannabis use rather than the introduction of legal medical cannabis, found that cannabis use reduced wages (Van Ours, 2007). These effects were the strongest among men (Kaestner, 1994) and, within men, among younger men (Register and Williams, 1992), although Kaestner (1994) found some evidence of positive effects for women. How much these studies relate to our study depends on the extent to which medical and recreational uses are similar. Medical legalization and associated dispensary access affect a significantly smaller, but potentially sicker, population than does recreational access. Medical legalization limits use to those with severely debilitating symptoms, typically for a limited range of conditions. Recreational access facilitates use not just for recreational purposes, but for medical use as well among those with conditions not previously approved for medical cannabis authorization, whose conditions were insufficiently severe to qualify, or with severe approved conditions but who were unwilling to join a registry.

The literature on the effects of recreational legalization on labor-related outcomes is even more nascent. The possibility that cannabis legalization may improve labor supply is supported by recent work finding reductions in disability claims (Maclean et al., 2021) and workers’ compensation benefits (Abouk et al., 2021) following recreational cannabis legalization. This reduction in worker injuries and injury severity could translate into positive effects on labor market outcomes more generally, but our results suggest that such effects are not large enough to translate into measurable changes in employment or wages. A recent working paper focuses on the demand-side effects in agriculture and retail sectors following recreational legalization in Washington and Colorado and, similar to this study, generally finds increased employment with limited wage effects (Jiang and Miller, 2021).

In summation, this paper explores the effect of entry of recreational cannabis dispensaries on local labor markets, using Colorado county-level data on employment and wages. We begin by providing a brief summary of cannabis legalization and detailing the potential effects of cannabis on labor supply and labor demand. The paper then proceeds to present our empirical analysis before discussing our results and concluding.

We assembled data for all 64 counties in Colorado in terms of labor market outcomes, recreational access, and county characteristics from various federal and state-level agencies. Table A1 in the Appendix summarizes a timeline for the period covered by each variable type we use and its data sources.

In order to estimate the impact of recreational dispensary entry on our variables of interest, we use a DID model with fixed effects. The variation in the location and timing of dispensary entry across counties in Colorado is exploited to accurately identify local labor market spillovers. We analyze the unemployment rate, the size of the labor force, the number of unemployed, and the number of employees, as well as the average weekly wages overall and by industry sector. All the dependent variables except the unemployment rate have been transformed using the natural logarithm to reduce the impact of outliers.

We estimate the following two specifications of our model: (1) ln(Ycmy)=α0+βRcmy+θXcmy+δy+δm+γc+ccmy {ln}\left( {{Y_{cmy}}} \right) = {\alpha _0} + \beta {R_{cmy}} + \theta {X_{cmy}} + {\delta _y} + {\delta _m} + {\gamma _c} + {c_{cmy}} (2) ln(Ycmy)=α1+λ1salescmylow+λ2salescmyhigh+θXcmy+δy+δm+γc+ccmy {ln}\left( {{Y_{cmy}}} \right) = {\alpha _1} + {\lambda _1}sales_{cmy}^{low} + {\lambda _2}sales_{cmy}^{high} + \theta {X_{cmy}} + {\delta _y} + {\delta _m} + {\gamma _c} + {c_{cmy}} Y_cmy is our main outcome variable for county c in month/quarter m and year y. R_cmy in Eq. (1) and salescmyi sales_{cmy}^i in Eq. (2) are the two types of treatment that we consider. R_cmy is a dummy variable indicating any recreational cannabis sales, whereas salescmylow sales_{cmy}^{low} indicates when $0 < sales ≤ $500,000 and salescmyhigh sales_{cmy}^{high} when sales > $500,000. β reports the average effect with the commencement of the sale of recreational cannabis, and λ₁ and λ₂ report the average effects of lower and higher levels of recreational cannabis sales relative to no sales. Thus, β in Eq. (1), and λ₁ and λ₂ in Eq. (2) are our primary coefficients of interest, summarizing the effect of access a recreational cannabis dispensary. X_cmy controls for the natural log of the number of medical cannabis patients. δ_y, δ_m, and γ_c are the fixed effects at the year, month/quarter, and county levels, which capture the year and month/quarter effects that are common across counties and the time-invariant county-level factors.

County-level fixed effects enable us to control for population density and demographics (e.g., age and socioeconomic status), which do not change over time during our sample period from 2011 to 2018. Month fixed effects account for seasonality, which affects all counties and differs only in levels, not in terms of which months are peak tourism months. For example, even though some counties always have more tourism than others (captured by the county fixed effects), most Colorado counties have spikes in tourism in the summer and ski seasons. Robust standard errors, c_cmy, are clustered at the county (treatment) level to correct for arbitrary correlation among observations in a given county.

The opening of dispensaries may be endogenous, although ideological factors probably play a role in outright bans (Bradford and Bradford, 2017), and initial licenses were restricted to existing medical licensees, potentially limiting the choice of entry location. It is plausible that dispensaries open earlier in counties with favorable economic conditions and where expected demand for recreational cannabis is high. Therefore, a dispensary opening may not be a random event. Although we use county, month, and year fixed effects, they may not fully absorb the unobservable differences between treated and control counties leading to the opening of a dispensary. We attempt to address this potential endogeneity by using event studies to assess pre-trends, attempting to predict entry with preexisting trends in our outcome variables, capitalizing on the institutional factor that only existing medical dispensaries were allowed to sell initially, and using a GSC method.

Our identification is based on both the timing and location of new dispensaries; thus, we use an “event study” research design to ensure that preimplementation factors do not drive the heterogeneity in the timing of implementation and to assess effects that change over time (Autor, 2003; Lovenheim, 2009; Lee and Mas, 2012). We estimate Eq. (3) for our event study model.

More specifically, we regress our outcomes, Y_cmy, on a series of “event–time” dummies Z^k_cmy.

Z^k_cmy equals one when county c is k months/quarters from the commencement of recreational cannabis sale in month/quarter m and year y. These indicator variables are set to zero for counties that never sell recreational cannabis during our given sample period from 2011 to 2018. Periods >40 months (10 quarters) before or after dispensary entry are collapsed into the first and last periods.

Following Borusyak and Jaravel (2017) and to abstract away from seasonality around the time of dispensary entry, the reference period is a half year prior to dispensary entry.

For our event study regressions, we normalize τ₋₆ = 0 for monthly data and τ₋₂ = 0 for quarterly data.

An additional reason for using the event study approach is that the DID estimate, i.e., the β coefficient in Eq. (1), differentially weights the effects close to the initial timing of the treatment from the effects of later treatment (Borusyak and Jaravel, 2017; Athey and Imbens, 2021; Goodman-Bacon, 2021; De Chaisemartin and d’Haultfoeuille, 2020). One way to address this concern is to estimate the dynamic treatment effect over the post-treatment period and take the average or weighted average of the coefficients (Borusyak and Jaravel, 2017). Therefore, we also run our event study including only the τ_k variables for the post-dispensary entry period, but otherwise following Eq. (3). We calculate the simple average and the weighted average with the weighted average coefficient calculated as k=042 τk*sk,∑k=042 \mathop {k = 0}\limits^{42} \,{\tau _k}*{s_k},\sum\limits_{k = 0}^{42} {} where s_k is the share of observations in the post-dispensary entry period after the counties started selling recreational cannabis.

Standard errors are calculated using the delta method.

Along with the event studies, we test for endogeneity by regressing the binary variable for dispensary entry in county c at month/quarter m and year y on preexisting trends in the labor market outcome variables used in our study. We create the preexisting trend variable by calculating the change in the variable's value relative to that 1 year and 6 months prior to entry.

(4) Rcmy=α0+βY changecmy+δy+δm+γc+Ecmy {R_{cmy}} = {\alpha _0} + \beta Y\,chang{e_{cmy}} + {\delta _y} + {\delta _m} + {\gamma _c} + {E_{cmy}} where Y change_cmy = Y_t − Y_{t−12 months} or Y_t − Y_{t−6 months}

We conduct additional robustness checks on our main specification, including controlling for population, county-specific seasonality (county–month fixed effects), and county–year trends; and omitting counties with the highest recreational sales overall (Denver) and per capita (Las Animas), both potential indicators of significant sales to out-of-state customers and potentially much larger demand-side effects than those experienced by counties serving primarily local consumers. Further robustness checks include comparing early entrants

Early entrants are defined as dispensaries opening by June 2014, when only previously licensed medical dispensaries were allowed to sell.

The benefits of dispensary opening may spill over to nearby counties, especially to nondispensary counties. For example, dispensary opening in treated counties may draw workers out of nondispensary counties so that the employment level declines in nondispensary counties, thereby, creating negative spillovers. Similarly, neighboring counties may benefit from positive spillovers related to improved supply chains. Failure to capture such spillovers may lead to biased estimates of the average treatment effect. Therefore, to remove the bias arising from average spillovers to both the nearby dispensary as well as nondispensary counties, following Butts (2021), we create an indicator variable border, (5) ln(Ycmy)=α0+βRcmy+λRcmy*bordercmy +θXcmy+δy+δm+γc+ccmy {ln}\left( {{Y_{cmy}}} \right) = {\alpha _0} + \beta {R_{cmy}} + \lambda {R_{cmy}}*borde{r_{cmy\,}} + \theta {X_{cmy}} + {\delta _y} + {\delta _m} + {\gamma _c} + {c_{cmy}} which equals one if a county c is within 50 km (approximately 30 miles)

Based on the proposal by Jeffrey Beall (2021), the maximum average commute is 29.2 minutes in Colorado. If people travel by freeway at 65 m/hour, then the maximum distance they travel is within 30 miles. We also check and confirm our results with 60 km (37.82 miles) and 70 km (43.5 miles).

In addition to corroborating the validity of our main results, an absence of spillovers would support the use of a county-level GSC method as better approximating our main analyses than a state-level synthetic control approach.

We further address how endogeneity may be affecting our results, using a GSC method to corroborate our results. The aim of this additional specification is to address potential endogeneity due to the violations of the parallel trends assumption, as observed from Figure 2, and improve the credibility of our estimates by utilizing the GSC methodology suggested by Xu (2017). This method is used to minimize the bias by creating an appropriate comparison group, especially in cases when the parallel trends assumption does not hold and data have untreated units serving as control units. The GSC method combines the synthetic control method proposed by Abadie et al. (2010) with linear fixed effects models and allows the treatment to be correlated with unit and time unobservables. This method is appropriate for our case as we have multiple treated units with heterogeneous treatment periods. It uses a parametric bootstrap procedure that provides valid inference. A cross-validation scheme selects the number of factors in the interactive fixed effect (IFE) model automatically and thus reduces the risk of overfitting. Similar to the synthetic control method (SCM), GSC uses the pretreatment outcome data for the treated units to choose weights for the control units and correlations between the control and treatment groups to predict the counterfactual. As units move from the control group to the treatment group over the study period, with some units untreated at the end of the period, GSC addresses the breakdown of the units into the treatment and control groups.

Figure 2

The GSC model from Xu (2017) has the following functional form: (6) yct=δctRct+γcft+μct {y_{ct}} = {\delta _{ct}}{R_{ct}} + {\boldsymbol {\gamma }_c}{\boldsymbol {f}_t} + {\mu _{ct}} where the treatment dummy indicator R_ct lets different counties experience the treatment in different years, δ_ct is the heterogeneous treatment effect on county c at time t; γ_c is the vector of unobserved common factors; and f_t denotes unknown factor loadings.

The time-varying coefficients are the (latent) factors, while the unit-specific intercepts are factor loadings in the paper by Xu (2017).

Therefore, an estimate for the treatment effect on the treated group is given by Δ^ct=yct−y^O {\hat \Delta _{ct}} = {y_{ct}} - {\hat y^{\rm{O}}} , for t > T_O.

Because the optimal number of included factors in the vector γ_c is unknown, GSC implements a leave-one-out cross-validating procedure before calculating the causal effect [for more details, see Xu (2017)]. As discussed by Xu (2017), the GSC estimation framework has one key limitation, namely, the “incidental parameters” problem if the number of pretreatment periods is small or if there are too few control units. In our case, such concerns are minimized as almost half our counties are never treated and we have relatively long pretreatment periods.

This method combines the synthetic control method with IFE models and is less sensitive to small sample size than the traditional synthetic control method. It is being used by a rapidly growing literature using panel data models with U.S. county-level treatment (Cook, 2021; Cengiz and Tekgüç, 2018; Nguyen, 2020), or with treatment measured in similar smaller geographic areas in other countries (Skorobogatov, 2021; Mäkelä, 2017). As the GSC eliminates confounding factors that violate the identifying parallel trends assumption in event study designs, particularly for unemployment rate in our case, our use of the GSC provides causally interpretable results.

We first report the results from event study models to assess pre-trends, then turn to DID regressions, and finally, we present our GSC model results that estimate the average treatment effect on the treated (ATT) of recreational cannabis dispensary entry.

5.1

Event study

The event study allows us to assess whether trends in labor market outcomes determine or even simply precede dispensary entry rather than dispensary entry affecting labor market outcomes. In other words, we expect that the coefficient τ_k for the pretreatment period, −40 ≤ k ≤ 0 for monthly data and −10 ≤ k ≤ 0 for quarterly data, would be clustered around zero in the absence of pre-trends. Figures 2–4 plot the estimated τ_k coefficients from Eq. (3).

For Figures 2–4, we generally observe no pretreatment trend in the coefficients except for the unemployment rate. For the natural logs of the total labor force, unemployed, all industry employees, and employees in manufacturing, construction, service-providing industries, and natural resource and mining, and for the average wages overall and by industry sector, our assumption of no pre-trends holds. Because of its usefulness as a policy benchmark, we continue to include the unemployment rate as an outcome. Even for the unemployment rate, we still observe a significant change. From Figure 2, we observe that the unemployment rate is higher in the preimplementation period, whereas it declines post-dispensary entry in recreational cannabis-selling counties. Similarly, the number of unemployed appears to be declining post-dispensary entry, while no obvious changes in labor force exist over time. The declines in unemployment and the number unemployed begin in anticipation of the market opening, about 6 months prior to dispensary entry. In order to stock stores on the day dispensaries open, growing and processing must begin in advance.

Growing cannabis can take 3–8 months, with additional processing time that varies by product type. https://www.leafly.com/learn/growing/marijuana-growth-stages. Accessed 07/16/2021.

From the event study in Figure 3, we observe that the number of employees for all industries combined has a sustained and significant increase after dispensary entry for treated counties. The increase in employees overall appears to be driven by the manufacturing sector, which experiences a sustained increase about 10 months post-dispensary entry for treated counties. These results also show that the sector arguably least likely to be affected by recreational cannabis dispensary entry, namely, Natural Resources and Mining, shows no trend pre- or post-dispensary entry, as expected.

Figure 3

Figure 4 shows some evidence of an increasing trend in wages, but the differences in the coefficients are not generally statistically significant.

Figure 4

5.3

GSC method

Results from the SCM

We also conduct a falsification test by assigning the treatment to 1 year post-dispensary opening. We find from Figures A10–A12 that our results still hold and the gap between the treated counties and counterfactual starts to increase exactly from “0” (1 year post-dispensary opening).

are shown in Figures 5–7. We observe that, prior to opening a dispensary, the treated and the control counties follow the same trajectory. However, after dispensary opening, the trajectory for our counterfactual counties starts to diverge for unemployment rate and manufacturing sector employees. The gaps become significant approximately 1 year after the recreational cannabis sale started. These effects are similar to the effects we observed from our event study graphs. Moreover, from Tables A37–A39 in the Appendix, we observe that the ATT for the unemployment rate is 0.566, which is close to our DID coefficients from Table 3. Although we do not find any significant effect for all industry employees, we find manufacturing sector employees in dispensary counties to increase by 13.7%. In addition to the event studies and DID regressions, these results provide evidence of a positive effect of a dispensary opening on the unemployment rate and manufacturing sector employees.

Figure 5

Figure 6

Figure 7

Table 8

The effect of recreational dispensary entry and sales on wages – regression analysis

	All	Cons	Manu	NR	Service
	(1)	(2)	(3)	(4)	(5)
Panel A: amount of sales
Recreational sale	0.002 (0.0114)	0.018 (0.0269)	0.007 (0.0229)	0.041 (0.0326)	0.002 (0.0136)
Recreational sale=1 × <50km not-sellingmy=1	−0.004 (0.0114)	−0.035 (0.0363)	0.014 (0.0385)	−0.051* (0.0227)	0.006 (0.0140)
Ln(number of medical patients)	0.004 (0.0113)	−0.084+ (0.0459)	−0.035 (0.0271)	0.066+ (0.0340)	0.002 (0.0173)
R2	0.937	0.778	0.932	0.910	0.925
Observations	2,048	1,536	1,312	1,536	2,048
Quarter FE
Year FE
County FE

Notes: Table reports the τ_k coefficients for k ≥0 from the event study regressions. Each column represents a separate regression. Below the table, we report the average and weighted average of the τ_k coefficients from the event study, including only the τ_k coefficients for the periods post-dispensary entry. Standard errors are clustered at the county level and reported in parentheses. The industry subsectors are Construction (Cons), Manufacturing (Manu), Natural Resources and Mining (NR), and Service.

+

p < 0.1,

*

p < 0.05,

**

p < 0.01.

FE, fixed effect.

While prior studies have focused on the effects of self-reported cannabis use or state-level medical cannabis legalization on labor outcomes, this study contributes to the literature by exploring the relationship between recreational cannabis dispensary access and labor market outcomes, distinguishing between supply- and demand-side effects. In addition, rather than focus on state-level differences, which a variety of unobservable factors may mediate, we exploit county-level variation in the timing of commencement of sale in Colorado, lending potentially greater internal validity to our results. Furthermore, we assess changes in labor market outcomes across industries, as this has not yet been explored in the literature on illegal, medical, or recreational cannabis legalization and use.

Using a DID estimation strategy and event studies, corroborated by a GSC model, we find consistent evidence of a decrease in unemployment and increases in the number of employees in manufacturing. A large decrease in the unemployment rate suffers from pre-trends in our event studies, but this is supported by the results of the GSC method. Our results showing a decrease in the number of unemployed and increases both in the number employed overall and in the service sector are only weakly supported by the synthetic control method results and, in the case of the overall effects, may be somewhat transient. We did not find any effect on the size of the labor force or on average wages overall or by industry sector.

The decrease in unemployment without a change in the size of the labor force suggests a tightening of the labor market driven by demand-side effects rather than by a decrease in labor supply. The lack of an effect on wages most likely arises from labor supply continuing to exceed labor demand in excess of frictional unemployment. These results match those found by Sabia and Nguyen (2018) in that we find no overall effect on wages. The lack of an effect on wages also means the tighter labor market is unlikely to negatively affect firms in unrelated industries (Black et al., 2005). Unlike Nicholas and Maclean (2019), we do not find an effect on the size of the labor force. However, this could be explained by their smaller affected population (elderly adults), a likely higher prevalence of medical conditions potentially treatable by cannabis, and a different treatment variable, namely, state-level medical cannabis access, rather than county-level recreational dispensary access in Colorado.

The predominance of the labor demand-side effect could be explained by the limited effects of cannabis on health and labor productivity, but it could also be influenced by the amount of cannabis-related tourism and an illicit market that existed prior to legalization. Tourists do not fill jobs, but they increase demand for primarily recreational, nonmedical cannabis, which – in turn – increases labor demand from recreational cannabis suppliers. Cannabis use by tourists has no effect on the health and productivity of the local labor supply. With respect to the illicit market, if individuals legally purchasing cannabis after legalization previously purchased on the illicit market and did not substantially change their consumption habits, then one would expect little effect on labor supply. As reported by Hunt and Pacula (2017), respondents from the Colorado survey, who had previously bought their cannabis from a “dealer” or “friend”, reported purchasing cannabis from either a medical or recreational dispensary after legalization. Potentially reconciling our results with the increased labor supply documented by Nicholas and Maclean (2019), older medical cannabis users may be less likely to be tourists and may be less likely to previously have bought cannabis on the illicit market than those purchasing cannabis from legal recreational dispensaries.

Our findings that cannabis dispensaries can trigger employment growth are in keeping with employment effects from the legalization of another previously illegal activity, namely, gambling. Opening a casino connects residents to a potentially addictive entertainment option while, at the same time, attracting even distant tourists, leading to spillovers into the service sector through effects on leisure and hospitality markets (Cotti, 2008; Reece, 2010; Humphreys and Marchand, 2013; Garrett et al., 2004).

Our industry-level results indicate that the increase in labor demand arose primarily from changes in demand for manufacturing sector employees. Manufacturing effects seem to primarily arise with entry, which could be explained by labor being a relatively a fixed cost in manufacturing. The fact that the magnitude of the effect on employees is larger than the apparent effect on the unemployment rate suggests that self-employed individuals are becoming employed by others. Considering that employees tend to experience greater financial stability, this could be an additional benefit from the labor demand shock induced by recreational cannabis dispensary entry.

This study does have limitations with respect to cross-border travel, substitution between medical and recreational cannabis, and the generalizability of results for future recreational legalization. Although we attempted to assess the magnitude of spillovers by controlling for selling in nearby counties, it is possible that we may still not have completely accounted for all spillovers, somewhat influenced by the manner in which our data were collected. On the demand side, we expect that increased employment occurs through the cannabis industry itself and spillovers to related industries that are geographically proximate. However, for the supply side, the potential for cross-county border travel to contaminate our results is much higher. Firms are more localized to a specific county than are workers or the labor force more generally, and cannabis users seem likely to cross borders to purchase cannabis, perhaps affecting labor markets in which dispensaries are not located. The different data sources we use could also lead to differing precision levels with respect to demand versus supply. Our data on employees are derived from payroll information reported by firms in a county, but our data on unemployment and labor force participation rely on household reporting, meaning some affected households may not be showing up as treated because they live in an untreated county but work in a county in which a dispensary entered.

The medical market for cannabis in Colorado was quite mature at the time of recreational cannabis legalization, and graphing sales over time for medical cannabis shows little change with recreational legalization. However, we did find that, while increases in the number of medical patients may augment the effect of recreational dispensary entry on the unemployment rate, an increase in the number of medical patients is associated with a reversal of some of the increase in the number of manufacturing sector employees associated with recreational dispensary entry. The magnitude of the effect from increases in the number of medical patients is much smaller than for recreational dispensary entry, but the opposite sign associated with the number of manufacturing sector employees suggests that markets in which the number of medical patients is growing may experience fewer benefits from dispensary entry, perhaps through market saturation or heterogeneity in the quality of dispensary entrants. Future studies should be conducted to understand the level of substitution between recreational and medical cannabis and how differences in the preexisting structure of the market into which a dispensary enters could lead to different labor market effects from entry.

Lastly, while this study finds that the introduction of a legal recreational cannabis market has a statistically significant impact on the number of employees in counties with dispensaries, this evidence should be extrapolated carefully because the industry's long-term evolution is uncertain and substantial heterogeneity exists in how states are choosing to legalize cannabis (Pacula et al., 2015). If the past is an indicator, states will continue to enact heterogeneous state-level policies with respect to cannabis, even in the presence of federal legalization, meaning that the effects of state-level recreational cannabis laws likely will differ from those experienced by Colorado in line with differences in the underlying laws. The other factor limiting the generalizability of Colorado's experience is that Colorado likely benefited from being one of the first states to legalize recreational cannabis. Even by the end of our sample period in 2018, as we observe from Figure A1 in the Appendix, Colorado was the only state among its neighbors with legal recreational cannabis sales. It is also closer to the Midwest and East Coast than Washington, California, and Oregon, making it the closest cannabis tourism destination for a large portion of the U.S. population. However, by April 2020, Alaska, Massachusetts, Nevada, Michigan, and Illinois also had legal recreational cannabis dispensaries, potentially reducing the labor market benefits of cannabis legalization even within Colorado.

In June 2019, Colorado Governor Jared Polis stated concerning the cannabis industry, “It's going very well.... It's creating tens of thousands of jobs, tax revenue for the state, filling up buildings for landlords, and reducing crime...” (Rosenbaum, 2019). Overall, the findings in this paper provide evidence that recreational cannabis dispensaries improved county-level labor market conditions in Colorado. Although some studies in the literature show adverse effects on labor markets, the equilibrium outcomes that we observe suggest that labor demand-side effects have a substantial impact on employment through the creation of jobs in the cannabis industry and associated spillover effects to other industries.

These positive effects on labor markets should be weighed against potential negative effects on public health. However, even the existence of a net negative public health effect from cannabis legalization is not consistently supported by the literature. Van Ours (2020) states that legalizing cannabis will lead to complex outcomes regarding public health but also will increase knowledge of the therapeutic potential of cannabis for mental health. Medical cannabis legalization has been shown to increase consumption (Martins et al., 2016), and excessive use of cannabis may lead to adverse health issues (Van Ours, 2007; Van Ours and Williams 2011, 2012; Van Ours et al., 2013; Volkow et al., 2014; Irons et al., 2014). It seems likely that recreational cannabis access increases use as well. At the same time, even these documented negative effects from use need to be calibrated by the substitution or harm reduction potential of cannabis, which has been shown to decrease the use of more harmful substances, such as alcohol (Baggio et al., 2018; Mark Anderson et al., 2013) and opioids (Vigil et al., 2017; Bradford et al., 2018; Stith et al., 2019; Carrieri et al., 2020), with Carrieri et al. (2020) documenting this effect even for cannabis containing low levels of THC.

Although this study focuses solely on Colorado and is limited by the data available, it represents an important step in understanding the impact of recreational cannabis laws on labor markets. Our results suggest that policymakers considering recreational access to cannabis should consider employment effects as a potential outcome from recreational cannabis legalization but should also account for significant differences for counties with and without recreational dispensaries. As more states are likely to consider legalizing recreational cannabis, understanding how local effects vary with greater access at the regional and national levels is an open question for future research.