Criminal Justice and Sexually Predators

THE POT RUSH: IS LEGALIZED MARIJUANA A POSITIVE LOCAL AMENITY?

DIEGO ZAMBIASI and STEVEN STILLMAN∗

This paper examines the amenity value of legalized marijuana by analyzing the impact of marijuana legalization on migration to Colorado. Colorado is the pioneering state in this area having legalized medical marijuana in 2000 and recreational mari- juana in 2012. We test whether potential migrants to Colorado view legalized marijuana as a positive or negative local amenity. We use the synthetic control methodology to examine in- and out-migration to/from Colorado versus migration to/from counterfac- tual versions of Colorado that have not legalized marijuana. We find strong evidence that potential migrants view legalized marijuana as a positive amenity with in-migration sig- nificantly higher in Colorado compared with synthetic-Colorado after the writing of the Ogden memo in 2009 that effectively allowed state laws already in place to be activated, and additionally after marijuana was legalized in 2013 for recreational use. When we employ permutation methods to assess the statistical likelihood of our results given our sample, we find that Colorado is a clear and significant outlier. We find no evidence for changes in out-migration from Colorado suggesting that marijuana legalization did not change the equilibrium for individuals already living in the state. (JEL I18, R23, K42, C22)

I. INTRODUCTION

In the past decade, a number of U.S. states have legalized marijuana either for medical or recreational proposes. A large literature has examined the impact of these law changes on the consumption of marijuana (Cerdá et al. 2017; Jacobi and Sovinsky 2016; Pacula 2010) and other related vices such as alcohol and tobacco (DiNardo and Lemieux 2001), and on a wide range of health outcomes including mental health (Henquet et al. 2005; Zammit et al. 2002), suicide (Anderson, Rees, and Sabia 2012), car accidents (Anderson, Hansen, and

∗This paper began life as part of Zambiasi’s Master’s Thesis in Economics and Management of the Public Sector at the Free University of Bozen-Bolzano. We would like to thank Mirco Tonin for helpful comments on the thesis, Javier Gardeazabal for insightful discussions about the synthetic control methodology and two referees for excellent sugges- tions on how to improve the paper. Zambiasi: PhD Student, Department of Economics, Uni-

versity College Dublin, Dublin 4, Ireland. Phone +39 348 069 3476, Fax +353 1 716 8188, E-mail [email protected]

Stillman: Professor, Faculty of Economics and Manage- ment, Free University of Bozen-Bolzano, Bozen- Bolzano, 39100, Italy. Phone +39 047 101 3132, Fax +39 0471 011009, E-mail [email protected]

Rees 2013; Hansen, Miller, and Weber 2019), and crime (Chu and Townsend 2019; Dragone et al. 2017). One so far unexamined question has been whether individuals, on average, view the legalization of marijuana as a positive or negative amenity and whether this has led to changes in individuals’ migration decisions and the spatial equilibrium of the U.S. population.

There is a large literature examining how household migration decisions help shape human geography (e.g., Blanchard and Katz 1992; Glaeser, Ponzetto, and Tobio 2014). The typical starting point is the neoclassical model of spatial equilibrium pioneered by Rosen (1974) and Roback (1982). This model assumes that in equilibrium there are no utility gains derived from moving locations. Hence, any change in amenities (or in how they are valued) will lead to population movement as any location offering extra-normal utility will experience in-migration until wages fall or rents rise sufficiently to elim- inate the utility differential. While the value of amenities is crucial for determining the spatial

ABBREVIATIONS

ACS: American Community Survey MSPE: Mean-Square Prediction Error

667

Economic Inquiry (ISSN 0095-2583) Vol. 58, No. 2, April 2020, 667 – 679

doi:10.1111/ecin.12832 Online Early publication August 14, 2019 © 2019 Western Economic Association International

668 ECONOMIC INQUIRY

allocation of the population, it is typically dif- ficult to estimate as there are many unobserved factors that covary with both any considered amenity, and house prices and wages.1 For this reason, most papers in this literature rely on natural experiments that exogenously change the value of amenities without directly affecting wages and rents (e.g., Albouy et al. 2014; Chay and Greenstone 2005).2

In this paper, we contribute to the subset of this literature that focuses on differences in local laws as a type of amenity (Malani 2008).3 Specifically, we examine the amenity value of legalized mari- juana by analyzing the impact of marijuana legal- ization on migration to Colorado. Colorado is the pioneering state in this area having legalized medical marijuana in 2000 and recreational mari- juana in 2012. We test whether potential migrants to Colorado view legalized marijuana as a posi- tive or negative local amenity.

We do this by using the synthetic control methodology, as developed in Abadie and Gardeazabal (2003) to examine in- and out- migration to/from Colorado versus migration to/from counterfactual versions of Colorado that have not legalized marijuana. This method chooses the optimal combination of donor units (in our case, states that have not legalized mar- ijuana as of 2017) that match both migration patterns and other chosen covariates in Colorado prior to marijuana legalization. We can then directly compare outcomes in Colorado after marijuana legalization to those in the optimal counterfactual version of Colorado. Unlike a traditional difference-in-differences approach, here the “control” group is chosen by an algo- rithm as opposed to potentially arbitrarily by the researcher and it is not necessary to have parallel trends in counterfactual outcomes between the treated unit (i.e., Colorado) and the defined control group.4

1. For example, people living on smaller islands typically live close to the coast but also have a more constrained housing supply and less jobs from which to choose.

2. A recent exception is Albouy and Stuart (2014) which directly estimates a parametrized version of the Rosen- Roback general equilibrium model. There are also a large number of descriptive papers looking at the relationship between local amenities and migrant settlement decisions, for example, Rodríguez-Pose and Ketterer (2012).

3. For example, Hellinger and Encinosa (2003) examine how laws that limit malpractice awards affect the location decisions of physicians while Young et al. (2016) examine how state tax laws on millionaires affect their migration decisions.

4. The synthetic control methodology is ideal for examining the impact of law changes in a single geographical

We find strong evidence that potential migrants view legalized marijuana as a positive amenity with in-migration significantly higher in Colorado compared with synthetic-Colorado after the writing of the Ogden memo in October 2009 that effectively allowed state laws already in place to be activated, and additionally after marijuana was legalized in 2013 for recreational use. From 2005 to 2009, on average, 187,600 people migrated to Colorado in each year. Between 2010 and 2013, in-migration increased by 21,372 people per year (a 11.4% increase) in Colorado compared with synthetic-Colorado. After full legalization in 2013, in-migration further increased by 14,087 people per year (an additional 7.5% increase).

When we employ permutation methods to assess the statistical likelihood of our results given our sample, we find that Colorado is a clear and significant outlier.5 We find no evidence for significant changes in out-migration from Col- orado relative to synthetic-Colorado suggesting that marijuana legalization did not change the equilibrium for individuals already living in the state. In total, 156,406 more people moved to Colorado than predicted based on migration into synthetic-Colorado after the signing of the Ogden memo. Given that we find no impact on out- migration, this implies that marijuana legaliza- tion increased Colorado’s population by 3.2% as of 2015.

Our findings are consistent with evidence that around 60% of Americans support marijuana legalization (Pew Research Center 2018) and anecdotal evidence that states that have legalized marijuana have attracted migrants (The Hemp Connoisseur 2014). They are also consistent with most previous research that has in general found positive effects of legalization, such as reductions in youth suicide, traffic accidents and crime (see above for citations). Furthermore, as legal mari- juana is taxed and raises significant revenues for states (Business Insider 2018), our findings are also consistent with individuals being attracted to states with increasing revenues.

Given that Colorado was a first-mover state, it is unclear whether later-mover states will attract population flows to the same extent, but overall

unit (e.g., state) which can be seen by the recent explosion of papers using this approach (e.g., Abadie, Diamond, and Hain- mueller 2010; Bohn, Lofstrom, and Raphael 2014; Jones and Marinescu 2018)

5. Our results are robust to how the synthetic weights are chosen and to further controlling for state and time fixed effects in a weighted difference-in-differences framework.

ZAMBIASI & STILLMAN: POT RUSH 669

our findings suggest that the widespread move- ment to legalize marijuana in the United States has had consequences for the spatial equilib- rium of the U.S. population. Depending on the composition of the individuals that are attracted to states that are earlier legalizers of marijuana, there could also be flow-on consequences for marijuana consumption, direct and indirect tax collection and the social welfare system. In line with the Rosen and Roback model, there should also be either a decline in wages and/or an increase in housing costs in Colorado as a result of the increased in-migration induced by mari- juana legalization.

A recent complimentary paper by Cheng, Mayer, and Mayer (2018) examines the impact of marijuana legalization in Colorado on house prices by comparing municipalities that have allowed retail sales versus those that have not in a difference-in-differences framework. Consis- tent with our findings and the Rosen and Roback model, they find that legalization leads to an average 6% increase in housing values which is caused by an increase in housing demand. They find that this increase does not occur in a munic- ipality until after marijuana is made available for retail sales. On the other hand, we find large increases in migration even during the period where marijuana is only available via medical dispensaries. Taken together, this suggests that, at least initially, easier access as opposed to the generation of new jobs and local tax revenues, was the main driver of migration inflows to Colorado. While after legalization, it is quite likely that both channels were important with newly arriving migrants drawn to municipalities that legalized retail sales which led to both more local job opportunities and tax revenues.

Our paper proceeds as follows: in the next section, we provide background on marijuana legalization in general and the situation in Col- orado. In Section III, we discuss the synthetic control method and the data we use in our analysis. In Section IV, we present our results and robustness checks. We then conclude in Section V.

II. BACKGROUND

A. Marijuana Policy in Colorado

On November 7, 2000, Amendment 20 (Medical Marijuana Law) was approved via bal- lot initiative by Colorado’s constituents. Starting January 1, 2011, individuals with debilitating

medical conditions could gain access to mari- juana by getting a certificate from a physician certifying their illness. Both patients and pri- mary caregivers were required to register with the Colorado Department of Public Health and Environment and were allowed to possess up to two ounces (57 g) of marijuana and cultivate up to six marijuana plants, with only three of them being mature at the same time. However, these activities remained illegal under the federal Controlled Substances Act of 1970, which clas- sifies marijuana as a Schedule 1 drug, and hence marijuana was only available from plants grown in noncommercial, home settings and the number of medical users remained relatively small (Eddy 2010).

A large change in access to medical mari- juana occurred October 19, 2009 when Attorney General Eric Holder announced formal guide- lines for federal prosecutors in states that had enacted laws authorizing the use of marijuana for medical purposes. In what is widely known at the “Ogden memo,” the Justice Department clarified that in states like Colorado, “the federal government would not prosecute anyone operat- ing in clear and unambiguous compliance with the states’ marijuana laws” (Miron 2014). Col- orado state law was quickly changed to permit commercial production and distribution of med- ical marijuana. The number of registrants grew rapidly from 4,819 in December 2008 to 115,467 in December 2014 (Ghosh et al. 2015).

On November 6, 2012, Colorado’s con- stituents approved Amendment 64 via ballot initiative by a slight margin, legalizing mari- juana use for recreational purposes. This new amendment authorized any individual, age 21 or over, to purchase and possess up to one ounce (28 g) of marijuana and cultivate six marijuana plants. For non-Colorado residents, a one-quarter of an ounce (7 g) purchasing limit was placed aimed at limiting the transportation of large quantities of marijuana to other states (Miron 2014). Under this new amendment, medical mar- ijuana provisions remained unaffected. The first dispensaries selling marijuana for recreational use opened in 2014 so up until that point there was little direct impact of the amendment except that marijuana could now be grown at home for purely recreational usage.

B. Marijuana Policy in the Rest of the United States

States can be divided into three main cate- gories in terms of how they regulate the selling

670 ECONOMIC INQUIRY

TABLE 1 State Weights for Synthetic Colorado

State In-Migration Out-Migration

Alabama 0 0 Florida 0 0.043 Idaho 0 0.004 Iowa 0 0 Kansas 0 0 Mississippi 0 0 Missouri 0 0 Nebraska 0 0 North Dakota 0 0.008 Oklahoma 0 0 Pennsylvania 0 0 South Carolina 0 0 South Dakota 0 0.004 Tennessee 0 0 Texas 0 0 Utah 0.436 0.001 Virginia 0.564 0.301 West Virginia 0 0 Wisconsin 0 0.628 Wyoming 0 0.011

Note: This table reports all states in the donor poll and the weights chosen using the method described in Section III to construct a synthetic control for Colorado for in-migration and out-migration outcomes.

of marijuana: (a) states in which all sales are illegal, (b) states in which selling is legal only for medical purposes, and (c) states in which selling is also legal for recreational purposes.6

Laws regarding possession are more nuanced and have greater variation across states. In order to assess the impact of marijuana legalization in Colorado using the synthetic control method, a donor pool of states must be used to create a synthetic-Colorado, which “looks” like Colorado in terms of prelegalization migration levels and key covariates, but in which marijuana is not legal. Hence, we only include in our donor pool the 20 states where in 2017 selling marijuana is still illegal for all purposes (see Table 1 for the list of donor pool states). This ensures that no state in synthetic-Colorado is directly affected by mari- juana legalization.

III. METHODOLOGY AND DATA

A. The Synthetic Control Method

We rely on the synthetic control method, as developed in Abadie and Gardeazabal (2003), to examine the impact of marijuana legalization

6. Legality of cannabis by U.S. jurisdiction, Retrieved January 24, 2017 from https://en.wikipedia.org/wiki/ Legal- ity_of_cannabis_by_U.S._jurisdiction#By_state.

on migration to/from Colorado. This approach compares outcomes for Colorado to those for a synthetic-Colorado created as a weighted combi- nation of other U.S. states chosen to best repro- duce the characteristics of Colorado before mar- ijuana legalization. Let J be the number of donor states (the 20 U.S. states in the donor pool), W is a (J × 1) vector of nonnegative weights such that w1 + · · ·wj = 1 and wj ≥ 0 ( j = 1, 2, … , J) where wj( j = 1, … , J) is the weight of state j in the synthetic-Colorado, X1 is a (K × 1) vec- tor of pretreatment average values of K migra- tion predictors for Colorado before legalization of marijuana, and X0 is a (K × J) matrix with all the migration predictors for all possible control states J.7

The synthetic control method chooses W*, which is the vector of weights that determines the best control region for Colorado by min- imizing (X1 − X0W)

′ V(X1 − X0W) subject to

w1 + · · ·wj = 1 and wj ≥ 0 ( j = 1, 2, … , J) where V is a diagonal matrix with nonnegative compo- nents and values on the diagonal represent the relative importance of each migration predictor. The optimal V and W are chosen simultaneously according to a nested optimization algorithm that minimizes the mean-square prediction error (MSPE) in the pretreatment period, as described in Abadie and Gardeazabal (2003). We imple- ment this using R package “synth” with the default option that uses both the Nelder – Mead and BFGS algorithms, and then picks the solution with the lowest MSPE.

Once the optimal weights minimizing the dif- ference in outcomes for the pretreatment period have been established, the effect of marijuana legalization can be calculated by examining the differences in outcomes after legalization between Colorado and synthetic-Colorado. More formally, (1) 𝔪t = Y

I it − Y

N it ,

where 𝔪 is the policy effect for Colorado (state i) at time t calcuated as the difference in the outcome for Colorado, Y I

it , and the generated

outcome for synthetic-Colorado,

(2) YNit = J∑

j = 1 j ≠ i

wjY N jt ,

7. The vector X can include predictors that are both time- variant (in which case pretreatment averages are matched) and time-invariant during the pretreatment period. The latter case are called special predictors and are treated equivalently to the time-varying predictors (Abadie, Diamond, and Hainmueller 2010).

ZAMBIASI & STILLMAN: POT RUSH 671

which is the combination of observed outcomes for the states ( j) in the donor pool, using the optimal weights calculated with the synthetic control method.8

Although the synthetic control method always generates the “best” match in the pretreatment period, the constraint that the weights must be nonnegative means that it is not guaranteed to fit pretreatment trends well. This depends on whether or not X1 lies within the convex hull of the Xo vector of the control states.

9 Following Abadie, Diamond, and Hainmueller (2010), we judge the quality of the match by estimating the MSPE for preintervention outcomes, that is, the square root of (YIit − Y

N it )

2 in the pretreatment period, for our main estimate and for each of our placebo estimates described below. We then rank the fit across all placebos.

To calculate the significance of our estimates, we implement a permutation method suggested by Abadie, Diamond, and Hainmueller (2015), comparing our synthetic control estimate to a distribution of placebo estimates. Formally, we implement the above synthetic control procedure for all 20 states in our donor pool and repeat this exercise as if the treatment year occurred in each year of the posttreatment time period (from 2010 to 2015) for a total of 120 placebo experiments. We define m̂jt as the estimate for state j with placebo treatment year t. We then conduct a two- tailed test of the null hypothesis of no effect in Colorado by comparing the true treatment effect to the empirical distribution of placebo estimates. Specifically, our placebo “p value” is defined as the proportion of estimates where the treatment effect in Colorado is smaller in absolute value than the placebo estimate.

We can also calculate confidence intervals by inverting the permutation test. Specifically, as suggested in Firpo and Possebom (2017), we rerun the permutation method where the

8. Abadie, Diamond, and Hainmueller (2010) provide more details about the synthetic control methodology. We use the “Synth” package in R to implement the approach on our data (Abadie, Diamond, and Hainmueller 2011). The policy effect can also be estimated in a more traditional difference-in-differences model where state and time fixed effects are included and the synthetic control weights are used to generate the correct control group. This has the advantage that aggregate macroeconomic effects are controlled for via the time fixed effects but with the additional assumption that there must be parallel trends in counterfactual outcomes between the treatment (i.e., Colorado) and control group (i.e., synthetic-Colorado).

9. In other words, if the treated state is an outlier in terms of the outcome in the pretreatment period, it will not be possible to create a good match from a pool of other states.

posttreatment outcome for Colorado is trans- formed by a particular 𝛼*. The 90% confidence interval is then defined at the set of 𝛼* where we cannot reject the null hypothesis of Colorado being significantly different than more than 90% of placebo treatments.

B. Data

Our main data source is the American Com- munity Survey (ACS). Our outcome variables are calculated using the State-to-State migra- tion tables generated using the ACS starting in 2005 and available in the U.S. Census Bureau’s online database. These tables indicate how many individuals changed residence within the United States during the previous year, and specify the origin and destination state. We aggregated the data to calculate both the total inflow into and the total outflow out of every state in every year from 2005 to 2015. These are the two dependent vari- ables in our analysis.10

We consider a number of predictors of migra- tion levels. Also from the ACS, we control at the state-level for the average total population, age-distribution (percent aged below 15, 15 – 34, 35 – 64, and 65 and above), percent with bachelor degrees and higher, median income, labor force participation rate, unemployment rate, and occu- pational composition (at the one-digit level) in the pretreatment period. We also control for per- capita state level tax revenues in the pretreatment period from the Annual Survey of State Gov- ernment Tax Collections11; and average monthly housing costs in the pretreatment period using the price of a vacant 2-bedroom rental unit at the 45th percentile of the Metropolitan Statistical Area distribution as measured in the Department of Housing and Urban Development Fair Market Rent series (Saiz 2007).12 Finally, we include as non-time-varying predictors; the average temper- ature in each state over the period 1971 – 2000;

10. We examined migration in levels, logs and as a per- centage of the previous year’s population. We get the best fit in terms of RMSE in the pretreatment period when we focus on migration levels so present those results in the paper. Our findings are qualitatively similar if we instead examine logs or rates. This is also the most common way used in the literature to measure internal migration (Molloy, Smith, and Wozniak 2011).

11. Taxes are defined as all compulsory contributions exacted by a government for public purposes, except employer and employee assessments for retirement and social insurance purposes.

12. We generate state level data by averaging all MSAs belonging to the same state ignoring those that belong to multiple states.

672 ECONOMIC INQUIRY

the average percent of votes for Republican can- didates in the 2004 and 2008 presidential elec- tions; and the average in- or out-migration rate from 2000 to 2005 for each state.13

As we only have 20 states in our donor pool, we can only control for key covariates that are potentially related to trends in migration. As a robustness check, we also do the synthetic control match without including covariates, and as will be shown below, get similar results.

IV. RESULTS

A. Synthetic Control Weights and Match Quality

As discussed above, we have a donor pool of 20 states in which the sale of marijuana is fully illegal as of 2017 for which to generate a synthetic-Colorado. As marijuana was not widely available in Colorado (or any states) until the writing of the Ogden memo in October 2009, we treat 2010 – 2015 as the treatment period in our analysis and 2005 – 2009 as the pretreatment period.14 Hence, our synthetic control weights will be calculated to create a synthetic-Colorado that best matches in- and out-migration levels in the 2005 – 2009 period as well as the covariates discussed in the previous section.

Table 1 presents the results from this exercise done separately for in-migration and out- migration. For in-migration, synthetic-Colorado is a composition of only two states from the donor pool, Utah with 43.6% and Virginia with 56.4% weight.15 Virginia is a larger state with a similar striking urban/rural divide as in Colorado, while Utah in a neighboring state that is similar in terms of weather and occupational composition. Eight states make up the donor pool for out-migration with Virginia (30.1%) and Wisconsin (62.8%) having the largest weights. Wisconsin is also a state with a large urban/rural divide.

Table 2 shows how the two weighted synthetic-Colorados look in terms of match- ing the real Colorado in the pretreatment period

13. Prior to 2005, only five-yearly in- and out-migration data is available at the state level. Hence, we cannot extend the standard synthetic control pretreatment period earlier in the time.

14. Zambiasi (2017) explores other definitions of the treatment period and discusses why this one makes the most sense. Examining the raw data on migration into Colorado, one also observes little change until 2010 (see Figure 1).

15. In the working paper version of the paper with a reduced set of control variables, 13 states contributed to gen- erating synthetic Colorado (with Virginia and Pennsylvania as the most important states), however the main results were qualitatively similar.

in terms of outcomes and covariates. Most importantly, there is a nearly perfect match in terms of the two outcomes variables even though the donor pool candidates have, in general, less migration than Colorado. Synthetic-Colorado is generally a bit larger in population, colder and more likely to vote for Democratic presidential candidates than real Colorado both otherwise looks pretty similar. The MSPE shows that we do an excellent job fitting the pretreatment in- migration and out-migration trends for Colorado and that the fit is in the middle of the distribution in terms of how well we can fit pretreatment outcomes for all the placebo interventions.

B. Main Results

Figures 1 (in-migration) and 2 (out-migration) present our main results. In the top panel of each figure, we graph the outcome for Colorado (the solid line labeled treated) and for synthetic- Colorado (the gray line labeled synthetic) cre- ated as described above. In the bottom panel of each figure, we show the difference over the whole sample period between real Colorado and synthetic-Colorado (the heavy black line) and then the difference between each placebo and synthetic treatment for 108 of 120 possible state/year combinations (the lighter lines).16 The vertical dashed lines indicate 2010, the year that medical marijuana became widely available and 2014, the year that recreational marijuana became widely available.

Examining the pretreatment period in both Figures 1 and 2 gives another indication of the quality of our synthetic control estimates. In the 5 years prior to medical marijuana being readily available, synthetic-Colorado and real Colorado have nearly identical in- and out-migration lev- els in all years and the general trends in the two “states” are similar. While this is somewhat by design, recall that the synthetic control estima- tor attempts to match average outcomes in the pretreatment period, hence there is no guarantee that outcomes in the real and synthetic states will follow the same trend or be similar in all pre- treatment years. Hence, this finding is a strong

16. As recommended by Abadie, Diamond, and Hain- mueller (2010), we drop states/years from the bottom panel where the MSPE is more than five times that for the actual treatment indicating that we do a very poor job of matching for these states in the pretreatment period. In practice, this drops all estimates for Florida and Texas (six for each state) and no other estimates. Each point on each lighter line shows the results from a placebo experiment hence there is one line per state.

ZAMBIASI & STILLMAN: POT RUSH 673

TABLE 2 Pretreatment Covariate Balance

Colorado All Donor Pool States

In-Migration Match

Out-Migration Match

In-migration (100,000s) 1.88 1.50 1.87 Out-migration (100,000s) 1.62 1.32 1.62 Population (100,000s) 48.4 55.5 55.0 66.4 Median income (1,000s) 65.1 52.7 66.0 59.8 House rent per month 725 565 673 631 Per-capita tax revenues (1,000s) 3.34 4.38 4.27 4.54 % age 0 – 14 0.207 0.203 0.226 0.193 % age 15 – 34 0.287 0.275 0.299 0.269 % age 35 – 64 0.359 0.341 0.327 0.359 % age 65+ 0.147 0.180 0.147 0.177 % bachelors degree or higher 0.328 0.230 0.290 0.258 % in labor force 0.703 0.658 0.686 0.685 % employed 0.933 0.933 0.935 0.935 % unemployed 0.058 0.059 0.012 0.006 % armed forces 0.007 0.005 0.049 0.056 % management and professional Occs 0.351 0.303 0.345 0.323 % service occupations 0.150 0.152 0.140 0.147 % sales and office occupations 0.237 0.237 0.243 0.231 % resources and construction Occs 0.106 0.108 0.099 0.094 % production and transportation Occs 0.088 0.132 0.107 0.140 Average temperature celsius 1971 – 2000 7.30 11.83 11.23 8.83 % voted Republican in 2004/2008 elections 0.482 0.568 0.574 0.477 In-migration 2000 – 2005 (100,000s) 6.438 4.413 5.688 Out-migration 2000 – 2005 (100,000s) 4.812 3.892 4.886 Preperiod mean-squared predication error 0.011 0.001 Preperiod MSPE percentile 61st 39th

Notes: Table reports average value of variables during the pretreatment period for Colorado, all donor states and the synthetic control constructed using the method in Section III. Columns (3) and (4) differ in the outcome matched on. The mean squared prediction error (MSPE) is calculated using all pretreatment data and the percentile is based on a comparison among all placebo estimates.

indicator that the synthetic control approach is applicable for our research question.

Comparing real Colorado to synthetic- Colorado in Figure 1 reveals that, once medical marijuana became widely available after the writing of the Ogden memo in October 2009, in-migration to Colorado increased by a large amount relative to in-migration to synthetic- Colorado. Large differences between the two are already apparent in 2011 and appear to get larger with each subsequent year. The lower panel shows that the difference found between real Colorado and synthetic-Colorado is larger than the difference found for all other placebo treatments. While we will present the formal statistics later in the paper, this clearly shows that in-migration to Colorado is a clear and sig- nificant outlier starting in 2011. We know of no other Colorado-specific changes that could have attracted additional migrants starting at exactly that point in time. On the other hand, it is clear in Figure 2 that increased availability of medical marijuana had no impact on out-migration from Colorado as out-migration is nearly identical

for real Colorado and synthetic-Colorado in the entire posttreatment period.

In Table 3, we present the treatment effects illustrated in Figures 1 and 2 along with the p value and confidence interval calculated for each estimate using the permutation method described in Section III. We calculate treatment effects separately for each year after the Ogden memo was written and for two combined time periods, 2010 – 2013 when medical marijuana was widely available and 2014 – 2015 when recreational mar- ijuana was widely available.

Panel A directly matches to the effects illus- trated in Figure 1 for out-migration. Starting in 2011, there is a statistically significant increase in migration to Colorado. Between 2011 and 2013, 18,900 – 31,200 extra people migrated to Colorado each year because of access to medical marijuana. In 2014 and 2015, an extra 33,600 – 37,300 people migrated to Colorado each year because of access to recreational marijuana. In the 2005 – 2009 period, Colorado averaged 188,000 migrants per year hence mar- ijuana legalization increased in-migration by

674 ECONOMIC INQUIRY

FIGURE 1 Impact of Marijuana Legalization on In-Migration to Colorado

Note: The top panel plots the synthetic control estimates of in-migration in 100,000s for Colorado from 2005 to 2015. The solid line plots the actual in-migration level in Colorado, while the dotted line plots the synthetic control estimate. The vertical dashed lines indicate 2010, the year that medical marijuana became widely available and 2014, the year that recreational marijuana became widely available. The bottom panels plot the results of a permutation test of the significance of the difference between Colorado and synthetic Colorado. The solid dark line plots the difference for Colorado using the true introduction of the treatment in 2010. The lighter lines plot the difference using other states and treatment years. We drop states/years where the RMSE is more than five times that for the actual treatment indicating that we do a very poor job of matching for these states in the pretreatment period.

between 10.1% and 19.9% in each year since 2011. In total, 156,406 more people moved to Colorado than predicted based on migration into synthetic-Colorado after the signing of the Ogden memo. Given that we find no impact on out-migration, this implies that marijuana legalization increased Colorado’s population by 3.2% as of 2015.

In Panel B, we present alternative esti- mates derived from a weighted difference-in- differences framework. Only data on the outcome variable is used here but both state and year fixed effects are included in the model along with interactions between posttreatment indicator and an indicator for Colorado. The data for each state are then weighted by the derived synthetic

ZAMBIASI & STILLMAN: POT RUSH 675

FIGURE 2 Impact of Marijuana Legalization on Out-Migration from Colorado

Note: The top panel plots the synthetic control estimates of out-migration in 100,000s for Colorado from 2005 to 2015. The solid line plots the actual out-migration level in Colorado, while the dotted line plots the synthetic control estimate. The vertical dashed lines indicate 2010, the year that medical marijuana became widely available and 2014, the year that recreational marijuana became widely available. The bottom panels plot the results of a permutation test of the significance of the difference between Colorado and synthetic Colorado. The solid dark line plots the difference for Colorado using the true introduction of the treatment in 2010. The lighter lines plot the difference using other states and treatment years. We drop states/years where the RMSE is more than five times that for the actual treatment indicating that we do a very poor job of matching for these states in the pretreatment period.

control weight for that state. This approach has the additional advantage that any macroeconomic effects on migration in all states will now be con- trolled for, but requires the additional assumption that macroeconomic effects are the same for all states.17 Consistent with the graphical evidence

17. This is equivalent to the typically discussed paral- lel trends assumption required for unbiased estimation of

that Colorado and synthetic-Colorado followed the same in-migration trends in the pretreatment period, these results are nearly identical. Overall, we find strong support for our main results above.

In Panels C and D, we present similar estimates for out-migration. Confirming the

difference-in-differences models. The standard synthetic con- trol estimator does not require this.

676 ECONOMIC INQUIRY

TABLE 3 Impact of Marijuana Legalization on Migration to/from Colorado

Year 2010 2011 2012 2013 2014 2015 2010 – 2013 2014 – 2015

Panel A: Impact on in-migration (100,000s), Colorado compared with synthetic Colorado Treatment effect 0.117 0.237 0.312 0.189 0.373 0.336 0.214 0.355 p value 0.380 0.046 0.009 0.074 0.000 0.000 0.055 0.000 95% CI [0.013,

0.124] [0.026, 0.249]

[0.039, 0.374]

[0.053, 0.498]

[0.066, 0.623]

[0.079, 0.748]

[0.027, 0.249]

[0.073, 0.686]

Panel B: Impact on in-migration (100,000s), weighted difference-in-differences estimates with state and year fixed effects Treatment effect 0.115 0.235 0.310 0.188 0.371 0.334 0.212 0.353 p value 0.111 0.014 0.005 0.029 0.003 0.004 0.001 0.000

Panel C: Impact on out-migration (100,000s), Colorado compared with synthetic Colorado Treatment effect −0.162 0.084 0.119 0.089 0.050 0.159 0.032 0.104 p value 0.130 0.343 0.269 0.324 0.556 0.130 0.685 0.287 95% CI [−0.012,

0.055] [−0.025, 0.110]

[−0.038, 0.165]

[−0.051, 0.220]

[−0.064, 0.275]

[−0.077, 0.330]

[−0.045, 0.138]

[−0.071, 0.303]

Panel D: Impact on out-migration (100,000s), weighted difference-in-differences estimates with state and year fixed effects Treatment effect −0.162 0.084 0.119 0.090 0.050 0.159 0.033 0.105 p value 0.018 0.113 0.046 0.098 0.297 0.019 0.599 0.197 Placebos 108 108 108 108 108 108 108 108

Notes: This table presents estimates of effect of access to legalized marijuana on migration to/from Colorado using the synthetic control method outlined in Section III. In Panels A and C, the treatment effect calculated by comparing Colorado to synthetic Colorado is shown for each posttreatment year and for the average of 2010 – 2013 and 2014 – 2015. In Panels B and D, treatment effects are from weighted difference-in- differences models that also control for state and year fixed effects. All p values and confidence intervals are constructed using the permutation test described in Section III.

graphical analysis in Figure 2, we find no evi- dence that legalization of marijuana has had an overall impact on out-migration from Colorado following the Ogden memo.18 Our estimates are precise enough to rule out effects half as large as those found for in-migration. This suggests that marijuana legalization had little impact on the equilibrium for individuals already living in the state.

C. Robustness Checks

Our final analysis examines the robustness of our findings to two choices made during the pro- cess of generating the synthetic control version of Colorado. First, as the correct choice of covari- ates is unclear and we only have 19 degrees of freedom to work with, we examine whether our results are robust to excluding covariates entirely and just matching on the outcome variable in the pretreatment period (in other words, the aver- age pretreatment outcome for each state is the only covariate used in the matching algorithm). Second, as pointed out in Kaul et al. (2015), if one believes that the covariates are particularly important for explaining the outcome variable then one is better off not controlling for the out- come in the entire pretreatment period. In fact,

18. Our difference-in-differences estimates show a significant negative impact on out-migration in 2010 and similar size positive impacts in 2012 and 2015. These effects are less than half the size of the impacts on in-migration in those years.

it is optimal just to control for the outcome in a short window prior to treatment. Because we already have a short pretreatment period, we examine whether our results are robust to only including migration in 2009 along with covari- ates when generating the synthetic control ver- sion of Colorado.

Table 4 presents the results from this exercise. In both cases, our main findings are qualitatively unaffected. When we do not use covariates when deriving synthetic-Colorado, we find a slightly smaller significant impact of marijuana legaliza- tion on in-migration, but cannot reject that the impacts are the same as in our main specification. On the other hand, when we only use information only for 2009 when deriving synthetic-Colorado, our estimated impacts of in-migration are nearly identical. Overall, our main results are extremely robust to our choice of covariates in the synthetic control model.19

V. CONCLUSION

In this paper, we examine the amenity value of legalized marijuana by analyzing the impact of marijuana legalization on migration to Col- orado. Colorado is the pioneering state in this area having legalized medical marijuana in 2000 and recreational marijuana in 2012. We test whether

19. We also find nearly identical results using a reduced set of covariates in the working paper version of the paper.

ZAMBIASI & STILLMAN: POT RUSH 677

TABLE 4 Robustness Checks

Year 2010 2011 2012 2013 2014 2015 2010 – 2013 2014 – 2015

Panel A: Impact on in-migration (100,000s), no covariates used when deriving weights Treatment effect 0.144 0.235 0.252 0.206 0.300 0.321 0.210 0.311 p value 0.139 0.000 0.000 0.019 0.000 0.000 0.019 0.000

Panel B: Impact on in-migration (100,000s), only 2009 outcome used when deriving weights Treatment effect 0.132 0.250 0.318 0.224 0.370 0.338 0.231 0.354 p value 0.287 0.120 0.065 0.111 0.028 0.037 0.046 0.028

Panel C: Impact on out-migration (100,000s), no covariates used when deriving weights Treatment effect −0.114 0.068 0.060 0.095 0.115 0.169 0.027 0.142 p value 0.176 0.370 0.417 0.231 0.167 0.074 0.648 0.130

Panel D: Impact on out-migration (100,000s), only 2009 outcome used when deriving weights Treatment effect −0.129 0.052 0.042 0.076 0.096 0.149 0.010 0.123 p value 0.185 0.509 0.611 0.352 0.287 0.176 0.889 0.204 Placebos 108 108 108 108 108 108 108 108

Notes: This table presents estimates of effect of access to legalized marijuana on migration to/from Colorado using the synthetic control method outlined in Section III. More restricted sets of covariates are used when generating the synthetic control weights in these specifications. All p values are constructed using the permutation test described in Section III.

potential migrants to Colorado view legalized marijuana as a positive or negative local amenity. We do this by using the synthetic control method- ology to examine in- and out-migration to/from Colorado versus migration to/from counterfac- tual versions of Colorado that have not legal- ized marijuana.

We find strong evidence that potential migrants view legalized marijuana as a positive amenity with in-migration significantly higher in Colorado compared with synthetic-Colorado after the writing of the Ogden memo in October 2009 that effectively allowed state laws already in place to be activated, and additionally after marijuana was legalized in 2013 for recreational use. When we employ permutation methods to assess the statistical likelihood of our results given our sample, we find that Colorado is a clear and significant outlier. We find no evidence for significant changes in out-migration from Colorado relative to synthetic-Colorado suggest- ing that marijuana legalization did not change the equilibrium for individuals already living in the state. In total, 156,406 more people moved to Colorado than predicted based on migration into synthetic-Colorado after the signing of the Ogden memo. This implies that marijuana legalization increased Colorado’s population by 3.2% by 2015.

Given that Colorado was a first-mover state, it is unclear whether later-mover states will attract population flows to the same extent, but overall our findings suggest that the widespread move- ment to legalize marijuana in the United States

has had consequences for the spatial equilib- rium of the U.S. population. Depending on the composition of the individuals that are attracted to states that are earlier legalizers of marijuana, there could also be flow-on consequences for marijuana consumption, direct and indirect tax collection and the social welfare system. In line with the Rosen and Roback model, there should also be either a decline in wages and/or an increase in housing costs in Colorado as a result of the increased in-migration induced by mari- juana legalization.

A recent paper by Cheng, Mayer, and Mayer (2018) examines changes in house prices in Col- orado after legalization comparing municipali- ties that have allowed retail sales versus those that have not in a difference-in-differences frame- work. They find that house prices increase in municipalities only after marijuana is made avail- able for retail sale. Taken together with our results, this suggests that, at least initially, easier access as opposed to the generation of new jobs and local tax revenues, was the main driver of migration inflows to Colorado. While after legal- ization, it is quite likely that both channels were important with newly arriving migrants drawn to municipalities that legalized retail sales which led to both more local job opportunities and tax revenues. In order to estimate the relative impor- tance of each channel, it would be necessary to estimate a structural model that includes, at a minimum, information on the collection and dis- tribution of tax revenues, local job opportunities, and the demand and supply for housing. This is a very interesting avenue for future research.

678 ECONOMIC INQUIRY

As the majority of previous research (see cita- tions in the introduction) finds little evidence for negative impacts of marijuana legalization on health (including mental health), accidents or crime, which are thought to be the three areas with the largest potential for negative spillovers, and local and state tax revenues have increased, it is highly likely that marijuana legalization has had a positive effect on local and state resources in Colorado. Beyond this, a large industry, espe- cially in tourism, has sprung up to take advantage of the change in legal status which has also likely contributed to increased employment and tax rev- enues (Kang, O’Leary, and Miller 2016). Again, given all of the potential indirect and spillover effects from legalization, a general equilibrium model is needed to rigorously answer the ques- tion of the extent to which Colorado his benefited from legalization (and if this changed once other states also legalized).

REFERENCES

Abadie, A., A. Diamond, and J. Hainmueller. “Synthetic Con- trol Methods for Comparative Case Studies: Estimat- ing the Effect of California’s Tobacco Control Pro- gram.” Journal of the American Statistical Association, 105(490), 2010, 493 – 505.

. “Synth: An R Package for Synthetic Control Meth- ods in Comparative Case Studies.” Journal of Statistical Software, 42(13), 2011, 1 – 17.

. “Comparative Politics and the Synthetic Control Method.” American Journal of Political Science, 59(2), 2015, 495 – 510.

Abadie, A., and J. Gardeazabal. “The Economic Costs of Con- flict: A Case – Control Study for the Basque Country.” The American Economic Review, 93(1), 2003, 113 – 32.

Albouy, D., W. Graf, R. Kellogg, and H. Wolff. “Climate Amenities, Climate Change, and American Quality of Life.” Journal of the Association of Environmental and Resource Economists, 3(1), 2014, 205 – 46.

Albouy, D., and B. Stuart. “Urban Population and Amenities: The Neoclassical Model of Location.” NBER Working Paper No. 19919, 2014.

Anderson, D., B. Hansen, and D. I. Rees. “Medical Mari- juana Laws, Traffic Fatalities, and Alcohol Consump- tion.” Journal of Law and Economics, 56(2), 2013, 333 – 69.

Anderson, D., D. Rees, and J. Sabia. “High on Life? Medical Marijuana Laws and Suicide.” IZA Discussion Papers No. 6280, 2012.

Blanchard, O. J., and L. F. Katz. “Regional Evolutions.” Brookings Papers on Economic Activity, 1, 1992, 1 – 75.

Bohn, S., M. Lofstrom, and S. Raphael. “Did the 2007 Legal Arizona Workers Act Reduce the state’s Unau- thorized Immigrant Population?” Review of Economics and Statistics, 96(2), 2014, 258 – 69.

Business Insider. “Marijuana Legalization Could Inject over $130 Billion into US Tax Coffers by 2025 — If the Trump Administration Stays Hands-off.” 2018. Accessed February 25, 2018. http://www .businessinsider.com/cannabis-to-add-a-million-jobs- 132-billion-tax-revenue-to-us-by-2025-2018-1?IR=T

Cerdá, M., M. Wall, T. Feng, K. M. Keyes, A. Sarvet, J. Schu- lenberg, P. M. O’Malley, R. L. Pacula, S. Galea, and D. S. Hasin. “Association of State Recreational Mar- ijuana Laws with Adolescent Marijuana Use.” JAMA Pediatrics, 171(2), 2017, 142 – 9.

Chay, K. Y., and M. Greenstone. “Does Air Quality Mat- ter? Evidence from the Housing Market: EBSCOhost.” Journal of Political Economy, 113(2), 2005, 376 – 424.

Cheng, C., W. J. Mayer, and Y. Mayer. “The Effect of Legalizing Retail Marijuana on Housing Values: Evi- dence from Colorado.” Economic Inquiry, 56(3), 2018, 1585 – 601.

Chu, L., and W. Townsend. “Joint Culpability: The Effects of Medical Marijuana Laws on Crime.” Journal of Eco- nomic Behavior & Organization, 159(March), 2019, 502 – 25.

DiNardo, J., and T. Lemieux. “Alcohol, Marijuana, and Amer- ican Youth: The Unintended Consequences of Govern- ment Regulation.” Journal of Health Economics, 20(6), 2001, 991 – 1010.

Dragone, D., G. Prarolo, P. Vanin, and G. Zanella. “Crime and the Legalization of Recreational Marijuana.” IZA Discussion Papers No. 10522, 2017.

Eddy, M. “Medical Marijuana: Review and Analysis of Fed- eral and State Policies.” Congressional Research Ser- vice, 2010, 7 – 5700.

Firpo, S., and V. Possebom. “Synthetic Control Method: Infer- ence, Sensitivity Analysis and Confidence Sets.” 2017. Accessed February 27, 2018. https://www.sites.google .com/site/vitorapossebom/research

Ghosh, T. S., M. Van Dyke, A. Maffey, E. Whitley, D. Erpeld- ing, and L. Wolk. “Medical Marijuana’s Public Health Lessons — Implications for Retail Marijuana in Col- orado.” New England Journal of Medicine, 372(11), 2015, 991 – 3.

Glaeser, E. L., G. A. M. Ponzetto, and K. Tobio. “Cities, Skills and Regional Change.” Regional Studies, 48(1), 2014, 7 – 43.

Hansen, B., K. Miller, and C. Weber. “Early Evidence on Recreational Marijuana Legalization and Traffic Fatali- ties.” Economic Inquiry, 2019. https://doi.org/10.1111/ ecin.12751.

Hellinger, F. J., and W. E. Encinosa. “The Impact of State Laws Limiting Malpractice Awards on the Geographic Distribution of Physicians.” Nation, 2003.

Henquet, C., L. Krabbendam, J. Spauwen, C. Kaplan, R. Lieb, H.-U. Wittchen, and J. vanOs. “Prospective Cohort Study of Cannabis Use, Predisposition for Psychosis, and Psychotic Symptoms in Young People.” BMJ (Clin- ical Research Edition), 330(7481), 2005, 11.

Jacobi, L., and M. Sovinsky. “Marijuana on Main Street? Estimating Demand in Markets with Limited Access.” American Economic Review, 106(8), 2016, 2009 – 45.

Jones, D., and I. Marinescu. “The Labor Market Impacts of Universal and Permanent Cash Transfers: Evidence from the Alaska Permanent Fund”. NBER Working Paper No. 24312, 2018.

Kang, S. K., J. O’Leary, and J. Miller. “From Forbidden Fruit to the Goose That Lays Golden Eggs: Marijuana Tourism in Colorado.” SAGE Open, 2016. https://doi .org/10.1177/2158244016679213.

Kaul, A., Kloeßner, S., Pfeifer, G. & Schieler, M. (2015). Syn- thetic Control Methods: Never Use All Pre-Intervention Outcomes Together With Covariates. https://mpra.ub .uni-muenchen.de/83790.

Malani, A. “Valuing Laws as Local Amenities.” The Harvard Law Review, 121(5), 2008, 1273 – 331.

Miron, J. “Marijuana Policy in Colorado.” Cato Working Paper No. 24_2, 2014.

ZAMBIASI & STILLMAN: POT RUSH 679

Molloy, R., C. L. Smith, and A. Wozniak. “Internal Migration in the United States.” Journal of Economic Perspectives, 25(3), 2011, 173 – 96.

Pacula, R. L. “Examining the Impact of Marijuana Legaliza- tion on Marijuana Consumption. Insights from the Eco- nomics Literature.” RAND Corporation, WR-770-RC, 2010.

Pew Research Center. “About Six-in-Ten Americans Sup- port Marijuana Legalization.” 2018. Accessed February 25, 2018. http://www.pewresearch.org/fact-tank/2018/ 01/05/americans-support-marijuana-legalization/

Roback, J. “Wage, Rents, and the Quality of Life.” Journal of Political Economy, 90(6), 1982, 1257 – 78.

Rodríguez-Pose, A., and T. D. Ketterer. “Do Local Amenities Affect the Appeal of Regions in Europe for Migrants?” Journal of Regional Science, 52(4), 2012, 535 – 61.

Rosen, S. “Hedonic Prices and Implicit Markets : Product Dif- ferentiation in Pure Competition.” Journal of Political Economy, 82(1), 1974, 34 – 55.

Saiz, A. “Immigration and Housing Rents in American Cities.” Journal of Urban Economics, 61(2), 2007, 345 – 71.

The Hemp Connoisseur. “The Great Marijuana Migration by DJ Reetz.” 2014. Accessed February 25, 2018. https:// www.thcmag.com/the-great-marijuana-migration/

Young, C., C. Varner, I. Z. Lurie, and R. Prisinzano. “Million- aire Migration and the Taxation of the Elite: Evidence from Administrative Data.” American Sociological Review, 81(3), 2016, 421 – 46.

Zambiasi, D.. “The Pot Rush: The Impact of Marijuana Legalization on Migration in Colorado.” Master in Economics and Management of the Public Sector, Free University of Bozen-Bolzano, 2017.

Zammit, S., P. Allebeck, S. Andreasson, I. Lundberg, and G. Lewis. “Self Reported Cannabis Use as a Risk Factor for Schizophrenia in Swedish Conscripts of 1969: Historical Cohort Study.” BMJ, 325(7374), 2002, 1199 – 9.

Copyright of Economic Inquiry is the property of Wiley-Blackwell and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.