Minimum Wage Pass-through to Wholesale and Retail Prices: Evidence from Cannabis Scanner Data

In November 2012, voters in Washington state approved the creation of a legal recreational marijuana market for adults 21 years and older.¹²¹²12Cannabis production and consumption remains prohibited at the federal level. However, in August 2013, the United States Department of Justice announced that it would not interfere with state-level legalization as long as distribution and sales were strictly regulated by states. This effectively green-lit legalization for U.S. states. Cannabis has since become a major agricultural industry in the state. In 2020, retail sales topped $1.4 billion and the industry contributed $1.85 billion to gross state product, making it the fourth most valuable agricultural crop in the state behind apples, wheat, and potatoes but ahead of timber, cherries, and hay (Nadreau et al., 2020). Cannabis is an important source of employment as the sector supports approximately 18,700 full-time equivalent (FTE) jobs in Washington (Nadreau et al., 2020). This mirrors the growing importance of cannabis employment in the U.S. more generally, where, according to one industry report, cannabis employs more than 428,000 workers (Barcott et al., 2022).¹³¹³13To add perspective, there are more cannabis workers than hair stylists, barbers, and cosmetologists combined (Barcott et al., 2022).

Figure 2 summarizes the minimum wage hikes used in my main analysis. In November 2016, Washington voters approved a ballot measure to scale up the state minimum wage from $9.47 to $13.50 by the year 2020. The measure spelled out predetermined, stepwise increases for January 1st of each year, with an initial increase to $11.00 in 2017, then $11.50 in 2018, $12.00 in 2019, followed by the final increase to $13.50 in 2020. Then, starting January 1st, 2021, the minimum wage was to adjust with the federal Consumer Price Index for Urban Wage Earners and Clerical Workers (CPI-W) on an annual basis. Besides the state minimum wage, there are two cities in Washington state with a binding citywide minimum wage. The city of Tacoma’s minimum wage took effect in early 2016 with a predetermined schedule of annual increases designed such that the city and state minimum wages converged in 2020, with the latter binding for all subsequent years. Seattle’s minimum wage went into effect in April 2015 and contained two sets of hikes depending on whether an employer paid towards an individual employee’s medical benefits.¹⁷¹⁷17Firms with over 501 employees are subject to a higher minimum wage than small employers. Since no cannabis business in Seattle has more than 500 employees, the large employer minimum wage does not apply. For employees earning $2.19 per hour in benefits (on top of their hourly wage), the minimum wage was identical to the state minimum wage except for a larger (predetermined) jump to $15 in 2021. In my main analysis, I assume that this is the schedule of hikes applicable to cannabis establishments in Seattle. However, in a series of robustness checks, I also consider the alternative schedule for employees earning less than $2.19 in benefits. In that schedule, the minimum wage increased more steeply and reached $15.75 in 2020, while in 2021 it adjusted according to a local CPI (this feature was written into the law in 2015). Due to the potential for reverse causality in that scenario, I drop Seattle establishments from the sample for the 2021 hike and find that results are unaffected (see appendix F for details).¹⁸¹⁸18I also consider potential wage spillovers from Seattle to surrounding areas, in which case estimates for establishments in neighboring cities would also suffer from reverse causality in 2021. See appendix F for details. For both Seattle and Tacoma, the citywide hikes occurred on the same day of the year as the statewide hikes (January 1st).

To monitor developments in the cannabis market, legalization came with stringent data reporting and sharing requirements for all licensed cannabis businesses. Producers and retailers are required to track every step of production from ‘seed to sale’ and they must regularly upload data feeds about plants, harvests, processing, transfers between businesses, and retail sales to the LCB. The data, which is usually reported weekly, contains detailed information on the price and quantity of each product sold by a producer-processor to a retailer, and the subsequent price and quantity of that very same product sold at the retail level.¹⁹¹⁹19Compliance with seed-to-sale traceability is strictly enforced by the LCB. When a business is issued a violation, it can receive a fine, a temporary license suspension, or both. In cases of repeated violations, a license can be revoked by the LCB board. Given such strict enforcement, violations are uncommon. In 2021 for example, the LCB issued 66 violations among approximately 1,192 licensees. See: https://lcb.wa.gov/enforcement/violations-and-due-process The LCB switched providers for its traceability system in October 2017 and again in December 2021, creating two structural breaks in the price data. My sample period lies between these breaks and spans August 2018 through July 2021, a period that covers three statewide and three citywide minimum wage hikes (see Figure 2). I obtained the data from Top Shelf Data, a data analytic firm that ingests the raw tracking data from the LCB and matches it with additional product information. The estimation sample covers sales from 1,192 distinct retail and producer establishments and contains an industry-wide average of 31,800 unique retail products and 18,268 unique wholesale products per month (see Table 1). To give an example, a 1.0 gram package and a 2.0 gram package of Sunset Sherbert usable marijuana (dried flower) produced by Northwest Harvesting Co are treated as different products in the data.²⁰²⁰20Similar to how wines can be distinguished by the grape (e.g. Riesling, Chardonnay, etc), cannabis comes in many strains, which is ’Sunset Sherbert’ in the given example. The LCB classifies products into 12 categories. As Table 1 illustrates, usable marijuana and concentrate for inhalation account for more than 80% of all retail sales. Another 14% of retail sales comes from solid edibles (chocolate bars, cookies, etc), liquid edibles (soda and other infused drinks), and infused mix (e.g. pre-roll joints infused with concentrates). The remaining categories make up less than 2% of total revenue; these are topical products (e.g. creams and ointments), packaged marijuana mix (e.g. pre-roll joints), capsules, tinctures, transdermal patches, sample jar, and suppository. Retailers are located in 37 counties while producers are located in 35 counties in Washington state (the state has 39 counties in total). Due to restrictions on the number of licenses a firm can hold, the vast majority are single-establishment firms. Over the entire sample period, the data contain $4.47 billion and $1.46 billion in retail and wholesale sales, respectively. Note that all retail cannabis sales are subject to a 37% excise tax, and retail prices and revenues reported in this paper are tax-inclusive. In contrast, there is no tax on wholesale transactions.

To estimate pass-through elasticities I follow previous studies (e.g. Renkin et al. (2022); Leung (2021)) and define the dependent variable as the natural logarithm of the monthly establishment-level price index:

$\pi_{j,t}$ is the inflation rate for establishment $j$ in month $t$; $I_{j,t}$ is an establishment-level Lowe price index that aggregates price changes across product subcategories $c$; the weight $\omega_{c,j,y(t)}$ is the revenue share of subcategory $c$ in establishment $j$ during the calendar year of month $t$.²¹²¹21As pointed out byRenkin et al. (2022), price indexes are often constructed using lagged quantity weights. Since product turnover is high in cannabis retail, lagged weights would limit the number of products used in constructing the price indexes. Thus, contemporaneous weights are used. To limit the potential impact of outliers, I trim inflation rates above the 99.5th and below the 0.5th percentile of the monthly distribution in my main specification (results are robust to keeping outliers). I describe the establishment-level price index in more detail in Appendix B.

My identification strategy rests on the idea that minimum wage hikes affect establishments with a high share of minimum wage workers more than those with a low share. Since wages are not observable at the establishment level, I follow previous studies and use geographic variation in the minimum wage bite as a proxy (see e.g. Card (1992); Bossler and Schank (2022); Leung (2021); Renkin et al. (2022); Dustmann et al. (2022)).²²²²22Firm-level wages are almost never observed in the minimum wage literature. An exception is Ashenfelter and Štepán Jurajda (2022). I define bite as the share of FTE workers in an industry-county earning below the new minimum wage two quarters prior to the hike. The industries are based on the North American Industrial Classification System (NAICS) which explicitly spells out classification for cannabis establishments of various types. NAICS 453 (”Miscellaneous store retailers”) captures all cannabis retailers since NAICS 453998 includes ”All Other Miscellaneous Store Retailers (except Tobacco Stores), including Marijuana Stores, Medicinal and Recreational” (US Census Bureau, 2017). NAICS 111 (”Crop production”) captures cannabis producers, since NAICS 111998 includes ”All Other Miscellaneous Crop Farming, including Marijuana Grown in an Open Field” and NAICS 111419 includes ”Other Food Crops Grown Under Cover, including Marijuana Grown Under Cover” (US Census Bureau, 2017).²³²³23Recall that in addition to growing cannabis, most producers are also processors (i.e. producer-processors). Processing falls under NAICS 424 which includes as a subcomponent ”Other Farm Product Raw Material Merchant Wholesalers, including Marijuana Merchant wholesalers” (NAICS 424590). However, NAICS classifies an establishment based on its primary activity, meaning that a cannabis producer only belongs to NAICS 424 if its revenue from processing activities exceeds that of its own crop production (US Census Bureau, 2017). Table 1 shows that unprocessed ”Usable Marijuana” accounts for the majority of revenue for producers, indicating that they belong to NAICS 111. Further evidence comes from Jiang and Miller (2022), who show that when cannabis was first legalized, the establishment count for NAICS 1114 in Washington increased by a similar count as the number of producer cannabis licenses. Moreover, the state saw a proportional increase in the number of workers and the total wages paid in NAICS 1114 (Jiang and Miller, 2022). Figure 4(b) depicts the average industry-by-county bite in the sample period for the NAICS industries containing cannabis establishments.

By defining bite at the level of the three-digit industry, I assume that variation in wages at cannabis establishments resembles variation in the corresponding NAICS industries.²⁴²⁴24In Appendix F, I construct an alternative bite variable at the five-digit NAICS level and show that my results do not depend on the chosen level of industrial classification. I document several facts to support this assumption. First, in Appendix A I show that average wages for cannabis retailers and producers are very similar to those in the corresponding NAICS industries.²⁵²⁵25I also find that average wages are homogenous among the industries contained in the relevant 3-digit NAICS industries. Moreover, for both the cannabis industry and the NAICS industries, average wages are remarkably close to the wage floor imposed by the minimum wage.²⁶²⁶26For producers, the gross average wage is 5%-10% above the minimum wage for the years 2018 to 2020, while for retailers it ranges from 15%-19% above the minimum wage. See appendix A. Thus, to the extent that the wage distributions differ between cannabis establishments and their NAICS industries, these differences should come from the upper part of the wage distributions rather than the lower part (since outliers are bounded from below by the minimum wage but unbounded from above). Furthermore, my regressions control for local labor market conditions (county-level average wage and unemployment) as well as establishment fixed effects, which implies that any remaining measurement error is likely to be random and will lead to conservative treatment effect estimates. Finally, the dynamic difference-in-differences framework allows me to closely examine treatment effect timing, meaning that for estimates to be biased, non-random measurement error would have to induce bias in the exact period that the minimum wage hike occurs, a scenario which I consider unlikely.

I obtained the bite data from the Washington Employment Security Department (ESD) which collects data on employment and wages in industries covered by unemployment insurance (about 95% of U.S. jobs).²⁷²⁷27The ESD data feeds into the better-known Quarterly Census of Employment and Wages (QCEW), a federal/state cooperative program that measures employment and wages in industries covered by unemployment insurance at the detailed-industry-by-county level. A similar dataset has been used in the recent minimum wage literature (see e.g. Dube et al. (2016); Renkin et al. (2022); Leung (2021)). It is important to note that, while the treatment intensity varies across time and space in my sample, the timing of the treatment does not vary (i.e. no staggered treatment). Thus, the six minimum wage hikes (three citywide and three statewide hikes) in the sample period amount to three minimum wage events, each spaced 12 months apart.

To estimate direct pass-through to wholesale and retail prices, I employ a difference-in-differences specification with a continuous treatment. This strategy, which is commonly applied in the literature investigating the effects of minimum wage hikes (e.g. Card (1992)), identifies the effects of aggregate shocks through cross-sectional variation in the fraction affected (the “treatment intensity”).²⁸²⁸28Other examples of this approach beyond the minimum wage literature are Lucca et al. (2019), who study the link between student loan credit expansion and college tuition, and Lindo et al. (2020), who investigate the effect of abortion clinic closures on abortion rates. The treatment intensity in my setting is the minimum wage bite variable constructed at the industry-by-county level. In contrast to a binary treatment, the continuous treatment setup identifies an average causal response parameter capturing the overall causal response of a small change in the minimum wage bite. Callaway et al. (2021) show that a key identifying assumption in this setting is no ”selection bias” among units with similar treatment intensities.²⁹²⁹29This is in addition to the standard parallel trends assumption. This implies that conditional on a set of controls and fixed effects, inflation in counties with slightly lower minimum wage bite provides a valid counterfactual for inflation in counties with slightly higher bite.³⁰³⁰30It is important to emphasize the local nature of the selection bias assumption: counties with vastly different minimum wage bite need not have the same potential outcomes. In appendix D, I describe the average causal response parameter in more detail and provide supportive evidence of the zero selection bias assumption.

Since firms may be forward-looking in their price setting, it is important to consider anticipatory effects that may cause price increases in the months leading up to the hike. Alternatively, firms may smooth price changes across several periods before and after a hike. The high frequency of the price data allows me to capture such dynamics, and I specify a nonparametric distributed lag model with leads and lags before and after each hike. Since the establishment-level price index and the minimum wage bite are first-differenced by construction, I specify the model in first differences. I estimate the following equation separately for retailers and producers:

Equation 2 relates the monthly establishment-level inflation rate, $\pi_{j,t}$, to the treatment intensity in county $k$, which is defined as the interaction between the percent change in the minimum wage applicable to establishment $j$, $\Delta MW_{j,t-l}$, and the minimum wage bite in the industry-county $k$ that establishment $j$ belongs to, $Bite_{k(j),t-l}$.³¹³¹31For establishments subject to a citywide minimum wage, $\Delta MW_{j,t-l}$ corresponds to the citywide hike. Note that $\Delta MW_{j,t-l}$ does not contribute to the identifying variation and simply scales the bite variable (the main identifying variation) such that the estimated coefficients are interpretable as pass-through elasticities at a given $Bite_{k(j),t-l}$.³²³²32I show in appendix F that results are similar when $Bite_{k(j),t-l}$ is not scaled by $\Delta MW_{j,t-l}$. The vector of control variables, $X_{k(j),q(t)}$, contains the average wage and unemployment rate for county $k$ in the quarter $q$ of month $t$. I include these to absorb variation in cannabis prices related to local macroeconomic factors that may covary with the minimum wage bite. Time fixed effects $\gamma_{t}$ account for monthly industry-wide changes in cannabis prices. Since the identifying variation is at the county level, standard errors are clustered by county to allow for autocorrelation in unobservables within counties, as in Bertrand et al. (2004).

For a given minimum wage hike, the parameter $\beta_{l}$ measures the percent change in establishment $j$’s prices resulting from a percentage point increase in minimum wage exposure $l$ months after the minimum wage hike (or $l$ months before when $l$ is negative). Though inflation is the dependent variable, I follow previous studies and present the estimates as the effect of the minimum wage on the price level (see e.g. Renkin et al. (2022); Leung (2021)). I thus normalize the effect to zero in a baseline period $m$ months before each hike and report the cumulative treatment effect as the sum of $\beta_{l}$ at various lags: $E_{L}=\sum_{l=-m}^{L}\beta_{l}$. The pre-treatment coefficients are reported in a similar manner with $P_{L}=-\sum_{l=m}^{-L-1}\beta_{-l}$.³³³³33Schmidheiny and Siegloch (2023) show that cumulative distributed lag coefficients are numerically equivalent to the parameter estimates from an event study design with binned endpoints. Since distributed lag coefficients measure treatment effect changes, one fewer lead has to be estimated compared to an event study specification. Thus, a 12 month event window requires estimating 11 distributed lag coefficients.

My approach resembles strategies that attempt to identify effects of aggregate shocks through cross-sectional variation in the fraction affected (see e.g. Bartik (1991); Lucca et al. (2019); Goldsmith-Pinkham et al. (2020)). This strategy is useful in the context of Washington’s cannabis market as it enables me to estimate pass-through despite the relatively small number of minimum wage hikes. Another advantage is that the estimated price level effects $E_{L}$ can be reformulated as pass-through elasticities at the average bite, allowing for direct comparison to elasticities found in the literature on minimum wage pass-through (e.g. Leung (2021); Renkin et al. (2022)).

An important consideration is the number of leads and lags to include in equation 2. One limitation is that minimum wage hikes occur in exact 12 month intervals, meaning event variables get highly collinear when $l$ is large. Another issue is that the establishment panel is not balanced, meaning that changes in the underlying sample may affect estimates when $l$ is large (Renkin et al., 2022). Therefore, in my baseline estimation I opt for a non-overlapping 12-month event window. While this may seem restrictive, I show in appendix F that treatment effects remain stable over a longer event window such that the 12-month event window adequately captures the short-run impact of the minimum wage on prices.³⁴³⁴34This mirrors the results from Renkin et al. (2022) and Leung (2021), who find that firms adjust prices at most three months prior to an event and that effects plateau within 1-2 months after the hike (Renkin et al., 2022; Leung, 2021).

A central concern with this research design is possible reverse causality. Since the treatment intensity is the product of two variables, $\Delta MW_{j,t-l}\times Bite_{k(j),t-l}$, the potential for reverse causality must be addressed for each of these variables in turn. $\Delta MW_{j,t-l}$ would suffer from reverse causality if policymakers were to increase the minimum wage in response to local inflation (e.g. in an effort to keep real wages constant). This is clearly not the case with the statewide hikes in my sample, since they are either predetermined or linked to the CPI-W, a national—not local—price index.³⁵³⁵35The city of Seattle has a citywide minimum wage that could be endogenous for some businesses for event 3 (January 1st, 2021). I address this possibility in Appendix F.4 and show that the main results are robust to dropping Seattle establishments for event 3. $Bite_{k(j),t-l}$ would suffer from reverse causality if county-level inflation drove wages. To account for this possibility, in a robustness check I include county fixed effects to absorb county-level differences in trend inflation.³⁶³⁶36Note that the baseline specification controls for establishment fixed effects because both the treatment and outcome variables are first-differenced by construction. Moreover, the distributed lag specification allows me to closely examine effect timing so that to the extent that differences in inflation trends remain, these can be easily distinguished from treatment effects.

It is important to highlight that when estimated for retailers, equation 2 uniquely identifies direct pass-through to retail prices and avoids picking up indirect pass-through effects. To see this, note that two conditions must be jointly met for the direct pass-through estimates to be contaminated by indirect pass-through. First, retailers would need to purchase predominantly from producers located in the retailer’s own county. Second, the bite variable for retailers would need to correlate with bite for producers within each county.³⁷³⁷37Importantly, both conditions must hold for the direct pass-through estimates to be contaminated by indirect pass-through effects. If the first condition is met but the second condition doesn’t hold, then the minimum wage effect on wholesale prices is part of the error term, but it is orthogonal to retail bite, and hence does not bias direct pass-through estimates. If the second condition holds but not the first, then producer bite and retail bite are not independent, but the minimum wage effect on wholesale prices in a given county has no impact on retail prices in that county since retailers don’t purchase from local producers. In appendix A, I show that the first condition does not hold since over 85% of retailers’ wholesale purchases are from producers located in other counties. Moreover, the within-county correlation between producer and retail bite is low: 0.18 (unconditional) and -0.03 (conditional on covariates). In other words, there is no systematic relationship between producer and retail bite, which implies that the second condition also does not hold.

One limitation is that my research design cannot distinguish between the effects of minimum wage legislation and implementation. If firms are forward-looking in their price setting, prices may adjust when a minimum wage hike is announced rather than when the hike actually takes effect.³⁸³⁸38Renkin et al. (2022), for example, find that price effects occur primarily in the three months following the passage of minimum wage legislation rather than after the hike itself. The first two hikes in my sample period were announced in 2016, two and three years prior to implementation, respectively. Because my sample runs from August 2018 through July 2021, any price effects from that announcement fall outside of the sample window and cannot be estimated. For the third event, the magnitude of the hike was announced three months prior to implementation, meaning price effects at announcement can be directly observed using my event study framework. As detailed in appendix F.8, I find no evidence of price effects at announcement but large effects at implementation for both wholesale and retail prices. This indicates that cannabis establishments wait until the cost shock hits before adjusting prices even if they have full prior knowledge about the magnitude of the shock.

I begin by estimating the direct effect of minimum wage hikes on wholesale prices and depict the results in Figure 5. In my preferred specification I include time FE but no county controls (results are robust to including controls). One question regarding the wholesale estimates is whether to control for a treatment-specific pre-trend since the baseline specification reveals a slight negative trend in the pre-treatment period. Though the trend is interrupted by a large and highly statistically significant treatment effect in the period that the minimum wage hike occurs, the contemporaneous treatment effect is slightly undone in subsequent periods as the pre-trend continues into the post-treatment period. Thus, while the trend does not mask the effect in period $t$, failure to account for the trend changes the interpretation of the results over a longer time horizon.³⁹³⁹39In appendix E I show that the pre-trend is entirely driven by event 2, a period corresponding to a wholesale supply glut and falling wholesale prices across the industry. It is therefore plausible that for event 2 unobserved confounders covary with treatment intensity and wholesale cannabis deflation. The trend persists despite the inclusion of county FE because county means are based on all three events and the trend is only present for a single event.

I apply two common strategies to control for the pre-trend, both of which yield similar results. First, I include region-time FE (i.e. interactions between time and region dummy variables) to account for regional economic trends that may covary with bite and inflation.⁴⁰⁴⁰40The regions are based on the three major socioeconomic regions in Washington state, where each region includes a subset of counties. See Appendix F.9 for details. To the extent that unobserved time-variant heterogeneity is common within regions, region-time FE will control for the treatment-specific trend (Neumark et al., 2014).⁴¹⁴¹41Unlike the baseline specification, I include county controls to further capture time-varying heterogeneity. Results are robust to including both county FE and controls or neither. This assumes that the proper counterfactual for inflation in counties with higher treatment intensity is inflation in counties with slightly lower treatment intensity located in the same region; that is, the identifying information comes from within-region variation in the treatment intensity.

Second, I apply the two-step procedure from Goodman-Bacon (2021) and re-estimate equation 2 using a trend-adjusted dependent variable. Specifically, I calculate the average of the distributed lag estimates (from equation 2) in the pre-baseline period and then extrapolate this pre-trend through the 12-month event window to obtain the treatment-specific linear trend $\hat{\pi}_{j,t}$. I then subtract the linear trend from the original dependent variable to get the trend-adjusted variable $\tilde{\pi}_{j,t(e)}=\pi_{j,t}-\hat{\pi}_{j,t}$. As argued by Rambachan and Roth (2023), this assumes that the observable linear pre-trend is a valid counterfactual for the unobservable post-trend. I view this as a valid assumption since the mean observable post-treatment trend is -0.00074 (95% CI: -0.00218, 0.00070) which is nearly identical to—and not statistically significantly different from—the pre-treatment trend of -0.00077, (95% CI: -0.00213, 0.00059).⁴²⁴²42In appendix E, I show that for all three specifications (unadjusted, trend-adjusted, region-time FE) the distributed lag coefficients are not statistically significantly different from zero for $t-5$ through $t-2$, and the period $t$ treatment effects are large and not statistically significantly different from each other.

Figure 5 illustrates that the period $t$ treatment effects are large and statistically significant at the 1-5% level for all three specifications. At the average bite (17.20%), a 10% increase in the minimum wage corresponds to a 1.07% increase in wholesale prices with the unadjusted dependent variable; 1.21% for the trend-adjusted specification; and 1.40% with region-time FE. In the latter two specifications, the pre-treatment period shows no significant trend and the large contemporaneous inflationary effect is no longer undone by the continuation of the pre-trend into the post-treatment period.⁴³⁴³43At higher lags, the price level effects from the specification with region-time FE are slightly lower than the trend-adjusted regression, but the difference is not statistically significant. Thus, it matters little how one controls for the trend, as the linear trend adjustment and region-time FE specifications both lead to a permanently higher wholesale price level effect.

Having obtained estimates for producers, I next estimate equation 2 for retail establishments. In my preferred specification for retailers, I include time FE and county controls (effect sizes are similar with county FE but estimates tend to be less precise at higher lags). Figure 6 illustrates that the effects for retailers differ from those of producers in several respects. First, effects for retailers show no pre-trend. Second, the treatment effect appears in $t-2$, i.e. one period prior to that for producers, suggesting that retailers may be more forward-looking in their pricing than producers.⁴⁴⁴⁴44This is consistent with the findings of Hollenbeck and Uetake (2021), who find that Washington’s cannabis retailers in have substantial market power and behave like local monopolists. Though producers’ market power has not been formally investigated in the literature, a common complaint among producers is their lack of market power compared to retailers (Washington State Liquor and Cannabis Board, 2021; Schaneman, 2021; Barbagallo, 2021). Given the earlier treatment effect, I normalize the baseline period in $t-2$ when calculating cumulative effects on retail prices. For retailers at the average bite (19.37%), a 10% increase in the minimum wage corresponds to a 0.64% jump in prices in period $t$.

In this subsection, I estimate the impact of minimum wage hikes on retail markups over marginal input cost (MIC) with the aim of determining the degree of wholesale cost pass-through to retail prices. I estimate the following equation for retail establishments:

The dependent variable, $\Delta\mu_{r,t}$, is the monthly percent change in MIC markup for establishment $r$ (see appendix B for details on the markup index). Table 6 displays the estimated markup effects and compares them to the estimates for direct and indirect pass-through to retail prices found in sections 4-5. Comparing $E_{4}$ in columns 1 and 2 reveals that the direct effect on the markup is very similar to the direct effect on the price level. This is not surprising, since direct pass-through to retail prices entails a price increase that is—by definition—independent of wholesale costs. In contrast, columns 3 and 4 reveal a large indirect effect on the price level in $E_{4}$ (significant at the 10% level) but a small and statistically insignificant indirect effect on markups. This indicates no markup adjustment to the wholesale cost shock on the part of retailers and implies a full pass-through of the wholesale cost shock to retail prices.

One concern in the cannabis industry is that rules governing production capacity for producers have created an uneven playing field, and a common complaint is that small producers operate on slim margins while large establishments enjoy a higher degree of market power (Washington State Liquor and Cannabis Board, 2021). If larger producers have more market power, then they may be able to absorb cost shocks along other margins (e.g. by adjusting profits), in which case pass-through elasticities may decrease with establishment size. Therefore, in this section I test for heterogeneous pass-through across different establishment sizes.

Producer licenses are based on a three-tier system governing the square footage of plant canopy that an establishment is legally permitted to operate. Tier 1 producers can grow up to 2,000 square feet of plant canopy, tier 2 can grow up to 10,000 square feet, while tier 3 can operate up to 30,000 square feet. Tiers were assigned to establishments before the market fist opened in 2014 and once assigned an establishment cannot switch tiers.⁵⁹⁵⁹59While producers can produce below the threshold for their tier, actual canopy usage has been found to be proportional to those thresholds (Washington State Liquor and Cannabis Board, 2019). Thus, the scale of production is effectively predetermined for cannabis producers during the sample period.

In contrast to producers, there is no tier system governing the scale of production for cannabis retailers. Nevertheless, since a retail firm can operate up to five establishments, one can approximate retail firm size by distinguishing between independent stores and those belonging to a chain.⁶⁰⁶⁰60Each establishment has it’s own cannabis business license. Since the number of licenses in the market is fixed, a firm must purchase an existing license in order to open another retail location. Retail establishments belonging to a chain may exhibit different cost pass-through compared to independent establishments for a variety of reasons. For example, if chains spread cost shocks across locations then pass-through effects will be attenuated compared to independent establishments. Alternatively, if chains have lower prices to begin with, then they may have more scope to increase prices via cost pass-through. Lower initial prices also imply that a dollar increase in price corresponds to a larger pass-through elasticity than that same dollar increase at a higher price point. Another possibility is that chains may have higher profit margins and hence may be able to offset cost shocks by reducing profits. Finally, a chain’s very existence may reflect superior management, better pricing strategies, or other unobservable differences that influence its ability to pass cost shocks through to prices. Ultimately, effect sizes for independent vs. chain stores are difficult to predict a priori.

I estimate equation 4 separately for each subsample and report the results in Table 7. Columns 1-3 show that pass-through to wholesale prices is monotonically decreasing with establishment size: small producers exhibit much larger price level effects than medium-sized producers, while effects for large producers are close to zero and statistically insignificant.⁶¹⁶¹61The difference between effect sizes for small and large producers is statistically significant for $E_{2}$ but not for $E_{0}$ or $E_{4}$. I interpret this as evidence that small producers may be less able to absorb the minimum wage cost shock via other margins of adjustment (e.g. by adjusting profits) compared to large producers.

For retailers, the opposite pattern holds, as direct pass-through effects monotonically increase with chain size (though effects are only statistically significant for establishments belonging to ’large’ chains). Results are less straightforward for indirect pass-through to retail prices. On the one hand, indirect pass-through monotonically increases with chain size for $E_{0}$ and $E_{2}$, thereby mimicking the pattern for direct pass-through. However, at $E_{4}$ effects resemble an inverse U-shape, with large effects for mid-sized chains and comparatively small effects for both independent stores and large chains. Taken together, retail establishments belonging to chains appear to have higher pass-through elasticities than independent stores.⁶²⁶²62In appendix B I show that average retail prices for product subcategories are similar for independent and chain stores. Average wholesale prices are similarly homogeneous across producer tiers. Thus, heterogeneous pass-through is not a result of different initial price levels across the scale of production.

In this section, I examine the employment effects of minimum wage hikes during the sample period. Since employment information is not available for cannabis establishments, I use monthly employment data from the QCEW at the 5-digit NAICS industry level. I estimate the following distributed lag equation:

The dependent variable is the first difference of (log) county employment at the 5-digit industry level.⁶³⁶³63Cannabis retailers belong to NAICS 45399 (”all other miscellaneous store retailers”). Cannabis producer-processors that grow indoors belong to NAICS 11141 (”food crops grown under cover”). I do not estimate employment effects for the NAICS industry containing outdoor growers because the majority of producer-processors in Washington state grow indoors. The treatment intensity $\sum_{l=-5}^{6}\beta_{l}\Delta MW_{t-l}\times Bite_{k,t-l}$ is the same as in the previous sections but for one difference: Since equation 7 is at the county level, $\Delta MW$ does not include citywide minimum wage hikes. Appendix Table G1 shows that employment effects are mostly insignificant. Overall, these results suggest that the minimum wage has no effect on employment for cannabis establishments. However, I caution against over-interpreting these results. Since cannabis workers are a subset of employees at the 5-digit NAICS level, one cannot definitively rule out employment effects at cannabis establishments.

This paper treats minimum wage hikes as a cost shock to cannabis establishments. However, it is conceivable that by raising the incomes of low-wage workers, minimum wage hikes affect demand for cannabis products which in turn could contribute to the retail price effects found in the main part of the paper.⁶⁴⁶⁴64Leung (2021) finds more than full minimum wage pass-through to grocery store prices in the U.S. and attributes part of the price effect to demand feedback. In this subsection, I test for such demand effects. If the retail price elasticity of demand for cannabis products is non-zero, then a regression of quantity sold on treatment intensity will suffer from simultaneity. Nevertheless, the resulting bias will lead to conservative estimates and hence provide a lower bound—and useful test—of possible demand effects.⁶⁵⁶⁵65Treatment intensity is endogenous because it simultaneously affects prices and quantity demanded. Since $cov(p_{r,t},\Delta MW_{r,t-l}\times Bite_{k(r),t-l})>0$, $cov(q_{r,t},\Delta MW_{r,t-l}\times Bite_{k(r),t-l}>0)$, and $cov(q_{r,t},p_{j,t})<0$, OLS estimates are negatively biased. To test for demand effects I estimate the following variant of equation 2 for retailers:

The dependent variable is the natural logarithm of the monthly establishment-level quantity index.⁶⁶⁶⁶66The quantity index is constructed similarly to the price index used in the previous sections. Since a quantity index requires products to have a common unit of measurement, equation 8 is restricted to the product category ’Usable Marijuana’ and product quantities are converted to grams. My results show that minimum wage hikes have no effect on cannabis consumption in the nine month period spanning $t-2$ through $t+6$ (see Appendix Table G2). However, a positive pre-trend is found for the periods $t-6$ through $t-2$. This pre-trend coupled with the flat treatment effects after $t-2$ preclude a strong conclusion on the effects of minimum wage hikes on cannabis demand.

Productivity is further channel through which the minimum wage may affect firms and workers, and several mechanisms have been proposed in the literature. Since the minimum wage reduces the price of physical capital relative to labor, firms may substitute machinery for workers. Mayneris et al. (2018) find evidence of this in the context of the 2004 minimum wage reform in China. However, I view this scenario as unlikely in the cannabis context, since cannabis production leaves little scope for technological adjustments. Most producers grow cannabis indoors in a setting averse to mechanization, and most cannabis is harvested, dried, and trimmed by hand to produce aesthetically pleasing, higher-priced buds (Jiang and Miller, 2022).

Ku (2022) proposes a different mechanism through which the minimum wage might affect productivity: if the minimum wage causes workers to anticipate potential layoffs, workers may increase their individual effort to avoid being laid off. Alternatively, firms may substitute out of low-skilled labor and into high-skilled labor. I do not observe employment at the establishment level so I cannot test either of these mechanisms.

The cost shock could also induce firms to adopt better management or organizational practices, which has been shown to increase productivity without the need for physical capital investment (Bloom et al., 2013; Atkin et al., 2017). To the extent that productivity-enhancing changes to business processes are implemented, they are likely long-term adjustments that carry a considerable lag before any productivity gains are realized. In that case, pass-through effects would decrease with successive minimum wage hikes. Yet I find the opposite to be true: of the three minimum wage events in my sample period, event 1 has the smallest pass-through effects while event 3 has the largest. Ultimately, however, I do not observe worker or firm productivity and therefore cannot rule out this channel.

One issue with this type of analysis is the degree to which results from one industry can be used to infer price dynamics in other industries. Some of the characteristics that make Washington’s cannabis industry an ideal laboratory for studying minimum wage pass-through also set the industry apart. The prevalence of small-scale indoor cultivation, along with rules governing the scale of cultivation, mean that cannabis production is likely to be more labor intensive than other agricultural industries (I discuss this topic in appendix A). Accordingly, the labor share of variable cost for cannabis producers is expected to be higher, and hence, the cost shock imposed by the minimum wage may be larger compared to other agricultural industries. At the retail level, however, the variable cost structure for cannabis establishments is remarkably similar to conventional retail settings (see appendix A.2), meaning the relative importance of the wholesale and labor cost shocks should be similar.

Another consideration when comparing industries is the price elasticity of demand, a key determinant of cost pass-through. Conditional on the degree of competition, a higher price elasticity reduces the ability of firms to pass cost shocks onto consumers (Weyl and Fabinger, 2013). This is an important consideration in the context of cannabis given its potentially addictive nature.⁶⁷⁶⁷67Epidemiological evidence on cannabis addiction is mixed. Recent evidence suggests that 9% of regular cannabis users become dependent, compared to 67.5% for nicotine, 22.7% for alcohol, and 20.9% for cocaine (Lopez-Quintero et al., 2011). Using a similar cannabis scanner dataset as this paper, Hollenbeck and Uetake (2021) estimate an overall retail price elasticity of 1.04 (relative to the outside good), which is higher than elasticities for cigarettes and beer but lower than that for spirits.⁶⁸⁶⁸68Estimates of aggregate price elasticities range from .16 (Gordon and Sun, 2015) to 0.8 (Becker et al., 1994) for cigarettes; 0.69 to 0.72 for beer (N. H. Miller and Weinberg, 2017); and 2.8 for spirits (Miravete et al., 2018). Interestingly, this elasticity is not substantially different from those found in non-addictive product markets.⁶⁹⁶⁹69N. H. Miller et al. (2017), for example, estimate an elasticity of 0.92 for the Portland cement industry. This suggests that based on the price elasticity alone, retail cost pass-through in cannabis is not expected to differ from other industries studied in the literature.

A final distinction is that the cannabis market operates under statewide autarky. This implies that cannabis retailers may be constrained in their response to the wholesale cost shock since the set of substitutable wholesale products is partly determined by geography. In contrast, retailers in other industries can leverage interstate trade networks to substitute out of products with high pass-through to wholesale prices. Therefore, I view my indirect retail pass-through estimates as more applicable for industries with home bias (e.g. grocery stores), but potentially less applicable for industries with a high degree of geographic substitutability along the supply chain (e.g. drugstores and general merchandise stores).⁷⁰⁷⁰70Using the 2007 Commodity Flow Survey, Renkin et al. (2022) provide evidence of substantial home bias in US grocery consumption.

In this paper, I study the effects of minimum wage increases on wholesale and retail prices in Washington state’s legal recreational cannabis industry. I use scanner-level data to estimate pass-through elasticities across a set of predetermined minimum wage hikes from 2018 to 2021. When ignoring pass-through to wholesale prices, I find that a 10% increase in the minimum wage raises retail prices by 0.77%. Yet, I also find substantial pass-through to wholesale prices: a 10% increase in the minimum wage raises wholesale prices by 1.72%. The existence of pass-through to wholesale prices implies that retailers face a wholesale cost shock in addition to the labor cost shock. When the empirical model is augmented to account for the wholesale cost shock, retail pass-through more than doubles to 2.04%. I find that retailers do not adjust markups to wholesale pass-through, indicating a full pass-through of the wholesale cost shock to retail prices. Moreover, pass-through to wholesale prices decreases with the scale of production, which suggests that large producers may adjust to the labor cost shock along other margins. The findings in this paper highlight the importance of examining the entire supply chain—beyond the final point of sale—when investigating the product market effects of minimum wage hikes.