• Quarterly Per Capita Sales by County
• Aggregate vs. Per Capita Sales across Counties
• Determinants of Aggregate Sales
• Conclusions

 

The California Department of Tax and Fee Administration (CDTFA) provides quarterly data by county on Total Cannabis Sales and Per Capita Cannabis Sales for the 3Q2018 - 2Q2023 period. 

For reference, Figure 1 shows a map of California’s counties.

Figure 1

Source: By Thadius856 - Self-made. Labels inspired by Image: California counties map1.gif., Public Domain, https://commons.wikimedia.org/w/index.php?curid=3277580

Quarterly Per Capita Sales by County

First, I looked at the time trends for quarterly per capita sales. There are too many counties to display them in a single graph, so I subdivided the counties into three separate graphs: Figure 2A shows quarterly per capita sales for the counties with the largest populations, Figure 2B shows per capita sales for medium-sized counties, and Figure 2C shows per capita sales for the smallest counties by population. 

Figure 2A

Figure 2B

Figure 2C

Counties with the Highest Per Capita Sales

The first pattern to note is that several counties have relatively large quarterly per capita sales (over $80): San Francisco leads the large counties, Stanislaus leads the medium-sized counties, and Mono, Humboldt, and Mendocino lead the small counties.

San Francisco is a well-known hub of cannabis activity. For example, it was home to Dennis Peron,^[1] the medical cannabis activist who, among other accomplishments, drafted Prop 215, the proposition legalizing medical cannabis in California in 1996.

Humboldt and Mendocino, along with Trinity County, form the Emerald Triangle,^[2] the largest cannabis producing region in the US. (There is only one cannabis retailer in Trinity, and the CDTFA only has a few quarters of data for the County, so Trinity has generally been excluded from my analyses.) It’s no surprise, then, that both these counties have high per capita sales.

As for Stanislaus, I have no immediate explanation for its high per capita use…

Let’s take a deeper dive. The measures of cannabis sales reported by the CDTFA are presumably those sales that occur in licensed cannabis retailers. It follows that, all else equal, we would expect sales to be higher in areas where there are more cannabis retailers. Figure 3 provides data for each county comparing the number of retailer licenses per one million population with 2Q2023 per capita sales (correlation = 0.762, P-value < 0.000). Sure enough, Mono County has more retail cannabis licenses than any other county, and it also has the highest per capita cannabis sales. Mendocino and Humboldt also have both high numbers of retailers and high per capita sales, as does San Francisco. 

Falling deep in the middle of the pack, Stanislaus remains a mystery…

Figure 3

Distributions of Quarterly Per Capita Sales by County over Time

The second pattern to note is that all three graphs start the period with per capita sales between $0 and $20. Per capita sales in the largest counties remain relatively flat and compressed – that is, there’s not much variation across counties, and sales increase a bit, but not much, over time. For medium sized counties, there’s a bit more cross-county spread, and sales increase over time from $0 - $20 per capita in the earlier quarters to $20 - $60 per capita in the later quarters. As for the smallest counties, they show the most spread and the most growth over the period. 

The last pattern to note is the general increase in per capita sales between 1Q202 and 2Q2020. 

Let’s take a closer look at the distributions of per capita sales across large, medium, and small population counties and over time. Figure 4A shows the distribution of quarterly per capita sales separately by large, medium, and small counties based on population. Figure 4A confirms what Figures 2A, 2B, and 2C suggest, that larger counties tend to have smaller per capita sales than larger counties. 

A quick aside to describe how the information in Figure 4A is calculated: There are 22 quarters of data between 2Q2018 and 3Q2023, and there are 15 counties in the large population category, yielding 22 x 15 = 330 county quarters of per capita sales data. There are 9 county quarters with missing data, leaving a total 321 county quarters of per capita sales for large counties. Of those 321 county quarters, 134 (42%) have sales under $20 per capita. In contrast, only 19% of the county quarters of data for medium-sized counties are under $20, and only 11% of the county quarters for small-sized counties are under $20.

Figure 4A

Figures 2A, 2B, and 2C suggest that per capita sales are increasing over time. So then we can take the analysis displayed in Figure 4A a step further by subdividing the distributions by time period. Figures 2A, 2B, and 2C suggest there’s a jump in sales between 1Q2020 and 2Q2020. Some counties also show a small jump between 1Q2022 and 2Q2022. So let’s try subdividing the period into 3 subperiods,

• 1^stSubperiod: 2Q2018 – 1Q2020 (8 quarters)
• 2^ndSubperiod: 2Q2020 – 1Q2022 (8 quarters)
• 3^rdSubperiod: 2Q2022 – 3Q2023 (6 quarters)

to see if the distributions differ across those periods. The results of this analysis are displayed in Figure 4B.

Figure 4B

The data shown in Figure 4B suggest that per capita sales increased between the first and second subperiod in all counties. Between the second and third subperiod, not much changed in large counties, and sales decreased somewhat in medium-sized counties. In small counties, per capita sales became less extreme across counties, that is, relative to the second period, in the 3^rd period there were fewer county quarters with either low sales or with high sales. 

The first pattern noted is that large counties tend to have lower, less-variable per capita sales than small counties. By their very nature, large populations tend to be more stable than small populations. Changes in averages over large populations require either small changes by large numbers of people and/or large changes by small numbers of people, where either one is difficult to achieve. In contrast, changes in averages over small populations only require small changes by a few people, which is not nearly as difficult to achieve. 

The second pattern is the change in all counties between 1Q2020 and 2Q2020 and the change in medium and small counties between 1Q2022 and 2Q2022. In March 2020, the Covid lockdowns took effect, and soon thereafter, cannabis activity was considered essential.^[3] Studies generally report at least a transitory increase in cannabis use during the Covid period.^[4] The decrease in per capita sales in medium-sized counties and the compression in per capita sales in small counties may be due to the ending of the Covid changes in consumption patterns, whereas the (less-extreme) Covid changes in the large counties seem to have persisted. 

Aggregate vs. Per Capita Sales across Counties

The next analysis I performed was to sum up Total Cannabis Sales across all the quarters (2Q2018 – 3Q2023) to get a measure of total aggregate sales for the entire period. The period at issue is 6 years long. Cannabis markets have evolved significantly over the past 6 years. By comparing Total Sales over the period with sales during the most recent quarter, 3Q2023, in the different analyses that follow, we can get a sense of what’s happened over the entire period and also what’s been happening more recently. 

Let’s start by comparing Total Per Capita Sales during the entire period with Per Capita Sales during the most recent quarter to see if relative sales across the counties have changed more recently. Figure 5A shows that there’s a high correlation (correlation = 0.828, P-value = 0.000) between per capita sales during the period and recent per capita sales.

Figure 5A

Figures 5B and 5C show aggregate sales over the period and per capita sales for the most recent quarter. The two figures show the same data, but Figure 5B is sorted on Total Sales in descending order from left to right, while Figure 5C is sorted on Per Capita Sales in descending order from left to right.

Figure 5B

Figure 5C

Figures 5B and 5C show that Los Angeles has by far the most aggregate cannabis sales ($6.7B) – presumably due to its large population (9.8M) – while tiny Mono County (population just over 13,000) has, by far, the highest quarterly Per Capita sales (just under $111). The figures also show that there’s no clear relationship between aggregate sales during the period and average per capita sales during the most recent quarter (correlation = 0.065, P-value = 0.686). In other words, it is not the case that counties with more sales also have more sales per person; rather, it’s more likely that counties with more sales simply have larger populations. In other words, population size may very well be a good predictor of cannabis sales. 

Determinants of Aggregate Sales

Thus far, the analysis suggests that higher per capita sales in a county may be due to a larger population and/or to more retail dispensaries. To better understand the determinants of cannabis sales, I pulled in data for numbers of licenses issued in California for various cannabis activities (for a more complete analysis of patterns of within- and cross-state cannabis licenses, see my earlier analysis, Cross-State Comparisons of Cannabis Licenses across Activities and Organizations) and ran some regressions. I used two measures of sales, Total Aggregate Sales for the entire reporting period, 2Q2018 – 3Q2023 and sales for the most recent quarter, 3Q2023. 

First, I looked at how well 3Q2023 population counts (derived from the data reported by CDTFA) predict Total Aggregate Sales and 3Q2023 Sales:

Sales = α + β • Population 

As seen in Figure 6, most of the variation in sales can be predicted by population. Specifically,

• The correlation between Total Aggregate Sales and Population is 0.961 (P-value < 0.000), and the correlation between 3Q2023 Sales and Population is 0.971 (P-value < 0.000). 
• Under these univariate regressions, a doubling of the population roughly doubles both measures of sales (columns [A] and [C]).

Figure 6A

Figures 7A and 7B plot Actual and Estimated values from the univariate estimations. (The outlier in the upper-right hand corner of the plots is Los Angeles.) The relatively tight clustering around the 45° line reflects the goodness of fit of the regression, that is, the high correlation between Population and Sales. 

Figure 7A

Figure 7B

Next, I considered whether population density (population divided by area) and the numbers of cannabis licenses issued in each county for different types of cannabis activity also predict sales: 

Sales = α + β₁ • Population + β₂ • Population Density + Σ β_3i • # Cannabis Licenses

Relative to the other counties in California, San Francisco (SF) has a very high population density and high per capita sales, while Los Angeles (LA) has a very large population and large total sales. I thought these 2 cities might skew the regression results, so I ran regressions on all the counties that had 3Q2023 sales, and I also ran regressions that exclude SF and LA to see if the results changed much. The results are presented in Figure 6B.

Figure 6B

As seen in Figure 6B:

• SF and LA do not change the relative orders of magnitude of the statistically significant variables in either the Total Sales or the 3Q2023 Sales regression.
• By accounting for cannabis license counts, we reduce the sensitivity of cannabis sales on population by about half, from an elasticity of about 100% to an elasticity of 40% - 50%. 
• Population Density has a marginally significant impact on Total Aggregate Sales but not 3Q2023 Sales.
• The numbers of cannabis licenses issued for Manufacturers, Retailers, and Testing Labs have statistically significant impacts on both Total Aggregate Sales and 3Q2023 Sales. 
• Cannabis sales are higher when there are more licenses for Cannabis Retailers and Testing Labs, while cannabis sales are lower when there are more licenses for Cannabis Manufacturers. More specifically, a doubling of the number of retail licenses leads to increases in cannabis sales by about 60%; a doubling of testing licenses leads to increases in cannabis sales by roughly a quarter; and a doubling of manufacturers’ licenses leads to decreases in cannabis sales by roughly a half. 

Figures 8A, 8B, 9A, and 9B plot Actual and Estimated values from the multivariate estimations. By adding in these additional variables, we get better predictions, that is, tighter clustering around the 45° lines in the figures. Note that Figure 8B is a zoom-in of Figure 8A (the axes have been cut off to show the detail around the data for the smaller counties), and Figure 9B is a zoom-in of Figure 9A

Figure 8A

Figure 8B

Figure 9A

Figure 9B

Population and Population Density vs. Sales

Of course, counties with larger populations, all else equal, will have greater sales. Higher population density generally also drives greater sales because larger portions of the population generally have better access to retailers in more densely populated areas, which encourages sales. In the regressions, population density was only a marginal predictor (in either significance or magnitude) of Total Sales and did not predict current 3Q2023 Sales. 

Let’s plot Population Density against Total Sales to see if that provides any clarity. As seen in Figure 10A, the low correlation between Population Density and Total Sales is largely driven by San Francisco (SF) and Los Angeles (LA). SF has relatively low sales and high population density, while LA has relatively high sales and moderate population density. 

Figure 10A

Figure 10B provides a zoom-in of Figure 10B, from which it’s clear that there’s at least a moderate correlation between total sales and population density.

Figure 10B

We can see from the figures what will happen if we exclude SF and/or LA from the regressions:

• From Figure 10A: If we keep SF in the regression but remove LA, we get high density drives sales.
• From Figure 10A: If we keep LA in but exclude SF, we get low density drives sales. 
• From Figure 10B: If we remove both we get little correlation.

The finding that as a whole population density does not drive sales suggests that the other variables in the regressions are capturing the effects of density. For example, more densely populated areas will tend to have more dispensaries, and more dispensaries are correlated with higher sales.

Retailer Licenses vs. Sales

The finding that larger numbers of retailer licenses are associated with higher cannabis sales was discussed earlier (see the discussion surrounding Figure 3). 

Testing Lab Licenses vs. Sales

Figure 11 shows there’s a high correlation between Total Cannabis Sales during the period and the numbers of testing lab licenses issued (correlation = 0.930, P-value = 0.000). One hypothesis consistent with the finding that larger numbers of lab licenses are associated with higher cannabis sales is that more testing labs don’t drive higher cannabis sales, but rather, sales drive more labs, since counties with higher cannabis sales need more labs to conduct all the required lab tests. A much less convincing hypothesis is that when there are more labs, consumers are more confident that cannabis products are being tested, that is, that cannabis products are safe to consume, which then drives more sales.

Figure 11

Manufacturing Licenses vs. Sales

Figures 12A and 12B show, respectively, the number of per capita manufacturer licenses and the percentage of manufacturer licenses against 3Q2023 quarterly sales. 

Figure 12A

Figure 12B

Figure 12A shows a small, marginally significant, positive correlation between per capita manufacturer licenses and quarterly per capita sales (correlation = 0.258, P-value = 0.103), but this measure of licenses does not control for any of the other variables. That is, it could be the case that larger counties simply have more of all types of cannabis activity, and in fact, the correlation between population and total license counts for all types of cannabis activity is 0.567 (P-value = 000).  

Figure 12B, which does account for other license types, shows a slightly larger, negative, statistically significant correlation between the portion of total licenses issued to manufacturers and quarterly per capita sales (correlation = (0.323), P-value = 0.039). This latter negative correlation between sales and manufacturer licenses matches the negative correlation in the regressions. But this still begs the question: Why do counties with more cannabis manufacturers, all else equal, tend to have lower cannabis sales?

Perhaps counties with relatively more manufacturer licenses tend to have relatively fewer retailer licenses, in which case sales would be lower due to consumer accessibility problems. We can test this hypothesis by looking at the correlation between manufacturer licenses and retailer licenses. Figure 13A shows that counties with more manufacturer licenses also tend to have more retailer licenses (correlation = 0.397, P-value = 0.007). However, taking total licenses into account, Figure 13B shows a small negative, but statistically insignificant relationship between portion of manufacturer licenses and portion of retailer licenses (correlation = (0.093), P-value = 0.545). So the lack of access to retailers hypothesis is not supported by the data.

Figure 13A

Figure 13B

As a second hypothesis, we know that in California, only manufacturers (and not also cultivators) extract cannabis,^[5] and, of course, manufactures produce non-flower cannabis products (dabs, vapes, edibles, etc.). So then perhaps cannabis consumers in counties with relatively more manufacturers tend to consume more extracted cannabis products and less flower. And perhaps consumers who consume non-flower products consume less cannabis than consumers who prefer flower. There’s also the possibility that non-flower products tend to be less expensive than flower products, which would generate lower dollar sales, but I don’t think this is the case. Unfortunately, since the CDTFA does not break out total sales by product type, I cannot test this hypothesis. 

Actual vs. Estimated Sales

The last analysis I performed was an examination of the differences between actual and estimated sales by county. Specifically, which counties had greater than estimated sales and which had less, that is, which counties are above and which below the 45° lines in Figures 8B and 9B?

Counties in which actual sales are greater than estimated sales (to the right and below the 45° lines) are counties where, holding constant all the various determinants of cannabis sales (population, population density, and numbers of cannabis activity licenses), sales are greater than expected – there’s either a greater prevalence of cannabis use (higher percentage of cannabis users in the population) and/or a greater intensity of cannabis use (higher cannabis sales per consumer). And, conversely, counties in which actual sales are less than estimated sales (to the left and above the 45° lines) are counties with either a lower prevalence and/or a lower intensity of cannabis use. These types of differences across counties are generally cultural in nature, that is, they depend on differences in the values and practices of county populations.

Figure 8B shows that high cannabis consumption counties over the entire 2Q2018 – 3Q2023 period include, for example: San Diego, Alameda, Sacramento, Santa Clara, Stanislaus, and Sonoma. Low cannabis consumption counties over the period include: Riverside, Orange, San Bernadino, Ventura, Santa Barbara, San Mateo, and Fresno.

Figure 9B shows that high cannabis consumption counties during the most recent quarter, 3Q2023, include, for example: San Diego, Santa Clara, Stanislaus, and Sonoma. Low cannabis consumption counties during the most recent quarter include: Riverside, Orange, Santa Barbara, San Mateo, and Marin.

A comparison of the high- and low-cannabis counties during the entire period and during the most recent quarter begs the question: Are the counties that are high- or low-consumption counties during the most recent quarter also high- or low-consumption counties over the entire period, or have counties changed more recently in their cannabis sales patterns? Figures 14A and 14B answer this question. Figures 14A and 14B reveal the following insights:

• Humboldt County, with actual cannabis sales during the most recent quarter of almost 25 times the estimated amount, and actual sales of over 3 times estimated sales during the entire period, is a big outlier, which dominates the results displayed in Figure 14A. Figure 14B is a zoom-in of Figure 14A so we can see the information better for the other counties. 
• In Figure 14B, the dashed lines and the circle divide the graph into 5 sections. 
  • In the center region, a good number of counties cluster around the point (1.0, 1.0), indicating that for most counties, actual sales coincide with estimated sales during both the aggregate period and the most recent quarter. 19/41 (46%) counties have both measures of sales within the range 0.75 – 1.25. This was expected, since (as seen in the regressions) the variables examined captured almost all the variance across counties in actual sales.
  • Counties in the upper-right quadrant outside the center region have sales that are higher than estimated in both the aggregate period and the most recent quarter. 9/41 (22%) counties (Humboldt, Kings, Stanislaus, Santa Clara, El Dorado, Lake, Solano, Shasta, and Sonoma) fall within this quadrant. These are the counties with high cannabis sales that cannot be explained by population or licensed activity.
  • Counties in the lower-left quadrant outside the center region have sales that are lower than estimated in both the aggregate period and the most recent quarter. 10/41 (24%) counties (Yolo, Nevada, Imperial, Fresno, Placer, Marin, Mendocino, Kern, San Mateo, and Santa Barbara) fall within this quadrant. 

A note on Mendocino: Mendocino, one of the Emerald Triangle counties, has a tiny population (89,000) relative to the number of retail licenses the county has issued, together with large per capita cannabis sales, as seen in Figure 3. As seen in the regressions (Figure 6), the number of retail licenses is the largest predictor of Total Sales over the period and the second largest predictor of 3Q2023 Sales. While Mendocino has very high sales per capita, the small population is not large enough to generate the total sales predicted by the large number of retailers. Hence, the regressions yield sales estimates that fall short of actual sales.

• Counties in the upper-left have greater-than-estimated sales over the total period, but less-than-estimated sales during the most recent quarter. None of the counties fall within this quadrant.
• Counties in the lower-right quadrant have less-than-estimated sales over the total period, but greater-than-estimated sales during the most recent quarter. 3/41 (7%) counties (Napa, Inyo, and Del Norte) fall within this quadrant.

• Counties along the 45°line have similar ratios of actual-to-estimated sales during both the aggregate period and the most recent quarter. 

Figure 14A

Figure 14B

The analysis shows that almost all counties are generally consistent in their cannabis sales patterns (either high sales, average sales, or low sales) over time. The exceptions seem to be Inyo, Del Norte, and Napa (Fresno is borderline), which have greater-than-expected sales in the most recent quarter, but less-than-expected sales over the longer period. Figure 14 shows the quarterly per capita sales data for these three (and a half) exception counties. As seen in Figure 14, the three counties at issue were late to the cannabis game and have been increasing consumption since they started. As for Fresno, it seems to simply have taken some time to ramp up sales.

Figure 15

Conclusions

After analyzing total cannabis sales and total per capita cannabis sales in California across counties and over time, what have we learned?

We know that 

Total Sales = Population x Per Capita Sales

So then total sales will be greater either when 

• The population is larger, 
• The prevalence of use within the population is larger, and/or 
• The intensity of use among cannabis users is larger.

A cursory view of the data indicates that total cannabis sales are higher, but per capita sales are lower, in counties with higher populations. The data also indicate that these relationships have persisted as cannabis sales have generally increased over time in all counties. No surprises here.

More interesting were the insights gained from analyzing determinants of cannabis sales.

• The biggest determinant of sales is number of retail licenses issued by the county. Since legal sales are only allowed in licensed retail dispensaries, it’s no surprise that counties with more retail licenses have higher sales. At the same time, there’s a strong correlation between population and number of retail licenses. So, then, while having more retail licenses enables greater points of access, it also reflects there being a larger population and/or higher demand, both of which lead to higher sales.
• Surprisingly, the second biggest – and negative – determinant of sales is the number of manufacturer licenses issued by the county. This is a puzzle. The best hypothesis I could come up with is that counties with more manufacturers might tend to sell relatively more non-flower cannabis products, and buyers of non-flower products might be less intense consumers than buyers of flower. This scenario would be consistent with lower cannabis sales in counties with more manufacturers.
• Third biggest determinant of sales is population.
• And the fourth biggest determinant is number of testing lab licenses. My guess is that causation in the relationship between cannabis sales and number of testing labs runs from sales to labs, and not from labs to sales. That is, my guess is that counties with more sales need more labs to test the cannabis products being sold. 

After using the regressions to better understand drivers of cannabis sales, the last analysis I performed was to use to the regression to understand which counties had greater-than-expected or less-than-expected sales, given the underlying determinants of sales. This analysis captured the unmeasured drivers of demand, namely culture. So what did we find? A comparison of actual with expected sales indicated that holding constant population, population density, and number of licenses for different types of cannabis activity,

• 19/41 counties (46%) had actual sales close to expected sales, that is, these counties have average patterns of cannabis sales.
• 9/41 counties (22%) – led far and away by Humboldt – had higher than expected cannabis sales. These are counties with cultures prone toward cannabis use.
• 10/41 counties (24%) had lower than expected cannabis sales. These are counties with cultures tending away from cannabis use.
• 3/41 counties (7%) had late starts, but sales have been increasing since they started. 

 

References

^[1] Barman J. A Brief History Medical Marijuana In California. Sparc. (2020, Jul 6). https://sparc.co/a-brief-history-medical-marijuana-in-california/

^[2] What is the Emerald Triangle? Weedmaps. https://weedmaps.com/learn/the-plant/emerald-triangle

^[3] Booker B. 'Illegal To Essential': How The Coronavirus Is Boosting The Legal Cannabis Industry. NPR. (2020, Apr 20). https://www.npr.org/2020/04/20/831861961/illegal-to-essential-how-coronavirus-is-boosting-the-legal-cannabis-industry

^[4] See, for example, Mehra K et al. Changes in self-reported cannabis use during the COVID-19 pandemic: a scoping review. BMC Public Health. (2023, Nov 1). https://bmcpublichealth.biomedcentral.com/articles/10.1186/s12889-023-17068-7 and Brenneke SG et al. Trends in cannabis use among U.S. adults amid the COVID-19 pandemic. Int J Drug Policy. (2022, Feb). https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8653409/

^[5] License Types. California Department of Cannabis Control. https://cannabis.ca.gov/applicants/license-types/