Adolescent Cannabis Use Disorder Following Recreational Cannabis Legalization in California, USA: A Population-Based Interrupted Time Series Study

Abstract

Aims:

To examine whether and how rates of adolescent cannabis use disorder (CUD) in California, USA changed after recreational cannabis legalization (RCL).

Design and Setting:

Population-based time-series study of adolescents aged 13 to 17 years with Kaiser Permanente Northern California healthcare system membership from January 2014 to February 2020.

Participants:

The sample (N=637,461) was 66.45% non-White, 50.93% male, with a mean (standard deviation) age of 14.07 (1.43) years.

Measurements:

Exposure variables included California state RCL passage (11/9/2016), which legalized possession, home grows, and reduced penalties for underage possession; and implementation of legal recreational sales (1/1/2018). Outcome was the count of incident CUD diagnoses in each month, extracted from the electronic health record. Interrupted time series analyses were fit using Poisson regression.

Results:

In the best fitting model, prior to RCL passage, CUD rates were decreasing (adjusted rate ratio [aRR] = 0.95 per year, 95% confidence interval [CI]: 0.92-0.97). Rates increased at the time of RCL passage + 6-month lag (adjusted slope change aRR = 1.26, 95%CI: 1.14-1.38). At RCL implementation + 6 months, there was a statistically significant negative slope change (adjusted slope change aRR = 0.86, 95%CI: 0.76-0.98). CUD incidence continued to increase, but was no longer statistically significant (adjusted trend aRR = 1.02, 95%CI: 0.96-1.09).

Summary:

Among adolescents in Northern California, USA, trends in incident cannabis use disorder rates appeared to increase following recreational cannabis legalization passage, with increasing rates persisting through retail implementation and three years post-passage. Results underscore the importance of monitoring cannabis use disorder, in addition to measures of cannabis use prevalence and frequency, when evaluating the public health implications of cannabis legalization.

INTRODUCTION

As cannabis legalization has spread across the U.S., national surveillance data have been widely interpreted as evidence that legalization has not increased youth cannabis use. By 2020, 24 states had legalized medical cannabis, and 16 had legalized recreational, or adult-use, cannabis, exposing 69% of the US population to a commercial cannabis market.¹ Yet, past-year cannabis use trends among adolescents nationwide have remained relatively flat,² hovering around 35% among 12^th graders, 28% among 10^th graders, and 13% among 8^th graders for roughly the past two decades. Youth use rates declined precipitously following the social distancing policies of the COVID-19 pandemic, a major endogenous disruption that complicates inference about legalization effects. By 2025, past-year use remained lower than pre-pandemic levels at 26% among 12^th graders, 16% among 10^th graders, and 7% among 8^th graders.²

However, aggregating to the national level and across patterns of use may conceal concentrated effects on smaller, higher-risk subgroups of youth. In particular, far less is known about whether legalization has affected clinically meaningful outcomes among youth, such as the development of cannabis use disorder (CUD). The risk of developing CUD increases sharply with greater frequency of cannabis use and with earlier age of initiation.^{3, 4} A recent systematic review and meta-analysis found that one out of three adolescents who use cannabis frequently meet criteria for CUD,⁵ with symptoms that are more persistent when onset occurs during adolescence rather than adulthood.⁶ Consistent with this concern, national trends in daily cannabis use have diverged from trends in any use, rising from 2016 to 2020 before declining during the COVID-19 period and subsequently increasing at varying rates from 2022 to 2025 among 8^th, 10^th, and 12^th graders.² Diverging patterns of use are consistent with developmental models showing that while experimentation with cannabis is common, it is often transient; escalation to frequent and sustained use occurs gradually and among a smaller subset of adolescents.⁷

In California, prior studies have found that rates of past-year, past 30-day, and daily or near-daily cannabis use among adolescents were stable or declining in the period preceding recreational cannabis legalization (RCL).^8–10 After RCL, trends again diverged by pattern of use: there was an immediate increase in past 30-day use, which then largely leveled off or reverted to a decline.⁹ In contrast, self-reports of past-year use and daily or near-daily cannabis use reversed course and increased significantly following RCL through 2020.^{9, 10}

In this study, we build on prior literature by examining trends in incident CUD diagnoses using 6 years of electronic health record (EHR) data from adolescents in California. Analyses span the periods before policy passage, between passage and implementation, and after the implementation of recreational retail sales until 2020. We evaluated whether trends in incident CUD diagnoses changed across these policy periods and whether observed patterns were consistent with prior evidence on frequent cannabis use among adolescents following legalization.

METHODS

Sample

Adolescents aged 13-17 years and enrolled in Kaiser Permanente Northern California (KPNC) for at least one month between January 1, 2014, and February 29, 2020, were eligible for inclusion (N = 641,556). KPNC is an integrated healthcare delivery system serving approximately 4.6 million members, demographically similar to the insured population of Northern California.¹¹ KPNC recommends adolescent well-child visits every 1-2 years, during which adolescents are routinely screened for substance use, including cannabis use. Adolescents reporting cannabis use receive further clinical assessment by their pediatrician and referral for specialty evaluation when indicated. CUD diagnoses are assigned by clinicians during routine care and may also be identified during other healthcare encounters, including primary care, behavioral health, emergency department visits, addiction treatment, and specialty care visits. Because the analytic focus was on incident CUD, adolescents with any documented CUD diagnosis prior to their first eligible month based on an unlimited lookback period (maximum: 12 years; median [IQR] 1.08 [0.54-1.91]) were excluded (N = 3,026, 0.47%). Individuals whose only CUD diagnosis during the study period was a remission code were also excluded (N = 4, 0.00%) (ICD-10: F12.11, F12.21, and F12.91; ICD-9: 304.33 and 305.23). Adolescents with other or unknown sex at baseline (N = 413, 0.06%), or who lacked address data needed to determine neighborhood deprivation index (NDI) throughout their entire eligibility window (N = 652, 0.10%), were excluded from the analysis. Because these exclusions represented less than 0.2% of the eligible cohort, they were unlikely to meaningfully influence aggregate monthly incidence estimates or overall study findings. The final analytic cohort comprised 637,461 adolescents (Supplemental Figure 1). We received KPNC Institutional Review Board approval with a waiver of informed consent and adhered to STROBE guidelines.¹²

Measures

Outcome

The primary outcome was incident cannabis use disorder (CUD), identified using ICD diagnostic codes for cannabis abuse, dependence, or unspecified cannabis use-related disorders (ICD-10: F12*; ICD-9: 304.3* and 305.2*), extracted from the EHR.

Exposure

We assessed exposure before and after two time points of RCL: RCL passage on November 9, 2016, when adult cannabis possession and home grows became legal and penalties for underage (< age 21) possession were reduced, and RCL implementation on January 1, 2018, when legal recreational retail sales commenced.¹³

Covariates

We adjusted for sociodemographic characteristics that may be associated with both CUD and healthcare utilization based on the literature. These included age (years, calculated at the beginning of each calendar month), sex (male and female), neighborhood deprivation index (NDI, quartiles, categorized using cut points based on distribution within the insured KPNC population),¹⁴ and race and ethnicity (Hispanic, and Non-Hispanic: Asian, Black, Multiple, other, or unknown, and White), extracted from the EHR.

Statistical analysis

We used descriptive statistics to characterize baseline sociodemographic characteristics, defined at each adolescent’s first month of eligibility, overall and across the three RCL-defined time periods. The pre-RCL passage period included adolescents eligible from January 2014 through October 2016; the post-RCL passage but pre-RCL implementation period included adolescents eligible from November 2016 through December 2017; and the post-RCL implementation period included adolescents eligible from January 2018 through February 2020, the month before COVID shutdowns took effect in California and affected regular check-ups of youth.

We conducted interrupted time series (ITS) analyses using segmented Poisson regression models.¹⁵ We specified two interruptions a priori: RCL passage on November 9, 2016, and RCL retail sales implementation on January 1, 2018. Because RCL passage was statewide, an unexposed comparison population was not available within California for evaluation of the legalization policy. Therefore, we used a single-group interrupted time series design to evaluate whether trends in incident CUD diagnoses changed following RCL passage and RCL implementation. For the crude model, the outcome was the monthly count of adolescents with incident CUD, and the offset was the natural logarithm of the eligible adolescent population at risk for each month. For adjusted models, the data were aggregated by month and covariate strata; the outcome was the monthly count of incident CUD cases within each covariate combination, and the offset was the natural logarithm of the corresponding at-risk population for that stratum. Once an adolescent was identified as an incident CUD case, they were removed from the denominator in all subsequent months.

Following standard model-fitting guidance for ITS models, preliminary models using data prior to the first interruption (RCL passage) were used to assess linearity, seasonality, and autocorrelation.^{15, 16} Based on these assessments, all models included seasonal adjustment using Fourier terms.¹⁷ We evaluated overdispersion in the adjusted primary model and found no evidence of overdispersion (dispersion parameter = 1.02); therefore, Poisson regression models were retained. Residuals showed no evidence of autocorrelation (Ljung-Box test, adjusted primary model p = 0.12). We assessed potential outliers by examining standardized residuals >3 SDs (Supplemental Figure 2).

We tested for interruptions at RCL passage and RCL implementation. Consistent with hypothesized gradual changes in the development of CUD, primary models specified immediate changes in slope at each interruption, but not level. We first conducted a model adjusting for time, a slope change at RCL passage, a slope change at RCL implementation, and seasonality using Fourier terms (Model 1), and a second model that additionally adjusted for age, sex, NDI, and race and ethnicity (Model 2). We reported adjusted rate ratios (aRR) with 95% confidence intervals (CIs) for changes associated with each interruption. Model 1 estimates were used to construct plots.

We conducted sensitivity analyses to assess whether the primary model assumptions were overly restrictive. These models included lags at 6 months and 3 months after RCL passage and RCL implementation (Models 3 to 4) and immediate level and slope changes at each interruption (Model 5). We compared model fit using Akaike information criterion (AIC) values, an established ITS approach to identify the best-fitting model.¹⁸ We used a two-sided P-value threshold of < .05 for statistical significance. Analyses were conducted with R version 4.5.1.

RESULTS

The sample (N = 637,461 adolescents) had a mean (SD) age of 14.07 (1.43) years at baseline and was 50.93% male. The cohort was 17.14% non-Hispanic Asian, 8.01% non-Hispanic Black, 29.58% Hispanic, 11.45% non-Hispanic multiple, other, or unknown race and ethnicity, and 33.55% non-Hispanic White (Table 1). Overall, 8,804 incident CUD diagnoses occurred, yielding a crude incidence rate of 5.03 per 10,000 adolescents over the study period.

Table 1.

Characteristics among N = 637,461 adolescents age 13 to 17 years before and after recreational cannabis legalization (RCL) passage and RCL implementation from January 1, 2014 to February 29, 2020, Kaiser Permanente Northern California

		Time periodb
Characteristic, no. (%) a	Overall	Pre-RCL passage	Post-RCL passage, pre-RCL implementation	Post-RCL implementation
	N = 637,461	N = 405,244	N = 312,538	N = 392,431
Age, years, mean (SD)	14.07 (1.43)	14.38 (1.49)	14.68 (1.49)	14.47 (1.50)
Sex
Female	312,783 (49.07%)	199,262 (49.17%)	153,183 (49.01%)	192,230 (48.98%)
Male	324,678 (50.93%)	205,982 (50.83%)	159,355 (50.99%)	200,201 (51.02%)
Neighborhood deprivation index, quartile
Q1: least deprivation	109,265 (17.14%)	62,248 (15.36%)	51,341 (16.43%)	80,738 (20.57%)
Q2	187,240 (29.37%)	121,405 (29.96%)	94,351 (30.19%)	110,675 (28.20%)
Q3	179,653 (28.18%)	115,587 (28.52%)	86,659 (27.73%)	113,157 (28.83%)
Q4: most deprivation	161,303 (25.30%)	106,004 (26.16%)	80,187 (25.66%)	87,861 (22.39%)
Race and ethnicity
Hispanic	188,558 (29.58%)	123,432 (30.46%)	94,519 (30.24%)	114,129 (29.08%)
Non-Hispanic
Asian	111,037 (17.42%)	68,954 (17.02%)	56,282 (18.01%)	72,428 (18.46%)
Black	51,038 (8.01%)	34,160 (8.43%)	25,139 (8.04%)	30,281 (7.72%)
Multiple, other, or unknownc	72,979 (11.45%)	39,831 (9.83%)	30,497 (9.76%)	44,779 (11.41%)
White	213,849 (33.55%)	138,867 (34.27%)	106,101 (33.95%)	130,814 (33.33%)

^aPercentages may not total to 100 percent due to rounding.

^bThe pre-RCL passage period included adolescents eligible from January 2014 – October 2016; the post-RCL passage but pre-RCL implementation period included adolescents eligible from November 2016 – December 2017; the post-RCL implementation period included adolescents eligible from January 2018 – February 2020.

^cMultiple, other, or unknown race and ethnicity includes American Indian, Alaska Native, Pacific Islander, Native Hawaiian, and multiracial individuals.

Average study enrollment duration was 27.40 months. Standardized mean differences (SMDs)¹⁹ indicated that the distributions of sex, NDI, and race and ethnicity were similar across time periods (absolute SMDs ≤ 0.04). The age distribution was also very similar over time (Supplemental Figure 3).

RCL Passage and Implementation

Findings were similar before and after covariate adjustment (Models 1 and 2; Table 2). Before RCL passage, the estimated mean rate of incident CUD was 4.96 per 10,000 adolescents (adjusted annualized trend aRR = 0.94, 95%CI: 0.91-0.97). At RCL passage, there was a positive slope change (adjusted annualized slope change aRR = 1.15, 95%CI: 1.04-1.26), after which CUD incidence increased (adjusted annualized trend aRR = 1.08, 95%CI: 1.00-1.15). Following RCL implementation, there was no statistically significant additional slope change (adjusted annualized slope change aRR =1.00, 95%CI: 0.90-1.11), and CUD incidence continued to increase at a similar rate (adjusted annualized trend aRR = 1.07, 95%CI: 1.03-1.12). By February 2020, rates of incident CUD reached 5.41 per 10,000 adolescents (Figure 1 and Table 2).

Table 2.

Model results examining changes in incidence rates of cannabis use disorder among adolescents age 13 to 17 years from January 1, 2014 to February 29, 2020, Kaiser Permanente Northern California

	Annual trend	Annual change in slope
Effect	Model 1a RR (95% CI)	Model 1a RR (95% CI)	p-value
Pre-RCL passage	0.94 (0.90 - 0.97)	—	—
RCL passage	1.07 (1.00- 1.15)	1.14 (1.04 - 1.26)	<.01
RCL implementation	1.07 (1.02 - 1.11)	1.00(0.90 - 1.11)	0.93
	Model 2b RR (95% CI)	Model 2b RR (95% CI)	p-value
Pre-RCL passage	0.94 (0.91 - 0.97)	—	—
RCL passage	1.08 (1.00 - 1.15)	1.15 (1.04 - 1.26)	<.01
RCL implementation	1.07 (1.03 - 1.12)	1.00 (0.90 - 1.11)	0.99
	Model 3c RR (95% CI)	Model 3c RR (95% CI)	p-value
Pre-RCL passage	0.95 (0.92 - 0.97)	—	—
RCL passage + 6 month lag	1.19 (1.10 - 1.28)	1.26 (1.14 - 1.38)	<.01
RCL implementation + 6 month lag	1.02 (0.96 - 1.09)	0.86 (0.76 - 0.98)	0.02

Abbreviations: CI – confidence interval. RR – rate ratio, RCL – Recreational Cannabis Legalization

^aModel 1 is adjusted for time, a slope change at RCL passage, a slope change at RCL implementation, and seasonality using Fourier terms.

^bModel 2 includes all Model 1 terms, additionally adjusted for age, sex, race and ethnicity, and neighborhood deprivation index quartile.

^cModel 3 includes all Model 2 terms, with 6-month lags for RCL passage and RCL implementation.

Bold indicates p-value < 0.05.

The pre-RCL passage period included adolescents eligible from January 2014 – October 2016; the post-RCL passage but pre-RCL implementation period included adolescents eligible from November 2016 – December 2017; the post-RCL implementation period included adolescents eligible from January 2018 – February 2020.

Figure 1.

Model 1 results examining changes in crude incidence rates of cannabis use disorder among adolescents age 13 to 17 years from January 1, 2014 to February 29, 2020, Kaiser Permanente Northern California

Abbreviations: RCL – Recreational Cannabis Legalization

Model selection and Sensitivity

The comparison of AIC values indicated that the best-fitting model was Model 3, which included 6-month lags for RCL passage and RCL implementation (Supplemental Table 1). In Model 3, the annualized trend was 0.95 (95%CI: 0.92-0.97) prior to RCL passage + 6 months, followed by a positive slope change and increase in CUD incidence at RCL passage + 6 months (annualized slope change = 1.26, 95%CI: 1.14-1.38; adjusted annualized trend aRR = 1.19, 95%CI 1.10-1.28). At RCL implementation + 6 months, there was a statistically significant negative slope change (adjusted annualized slope change aRR = 0.86, 95%CI: 0.76-0.98). CUD incidence continued to increase but was no longer statistically significant (adjusted annualized trend aRR = 1.02, 95%CI: 0.96-1.09) (Table 2). Outlier analysis identified two outliers during the post-RCL implementation period (October 2018 and May 2019; Supplemental Figure 2). Sensitivity analyses excluding those outlier months found similar results to the main analyses; however, the previously negative slope change at RCL implementation + 6 months was no longer statistically significant (Supplemental Table 2).

Results differed slightly in Model 5, the sensitivity analysis model that allowed for immediate sudden changes by including level changes at RCL passage and implementation (Supplemental Table 3). Before RCL passage, the annualized trend was 0.95 (95%CI: 0.91-0.99). However, at RCL passage, the annualized slope change was still positive but no longer statistically significant (annualized slope change aRR = 1.14, 95%CI: 0.98 -1.33), and at RCL implementation, the annualized slope change was still negative but no longer statistically significant (annualized slope change aRR = 0.97, 95%CI: 0.83-1.14). We note also that this model had the poorest fit based on the AIC values (Supplemental Table 1).

DISCUSSION

In this large cohort of adolescents receiving care in Northern California, rates of incident CUD declined steadily prior to RCL passage, but this trend reversed following passage of the law. This upward trend persisted beyond the implementation of recreational retail sales without evidence of an additional change in slope. By the end of the study period, over three years after RCL passage, CUD incidence had returned to levels comparable to those observed at the beginning of the evaluation period.

These findings are consistent with prior research documenting increasing trends in past-year and frequent cannabis use following legalization from both anonymized in-school surveys⁹ and routine clinical screening,¹⁰ and extend the literature by focusing on a clinically meaningful outcome that reflects sustained and impairing patterns of use in a large and diverse population. The increase in frequent use following legalization is of potential concern, given the relationship of frequent use to subsequent development of CUD and other adverse outcomes.⁴

The observed slope changes are consistent with gradual changes in risk accumulation over time rather than abrupt increases at specific policy milestones, as predicted by developmental models of substance use in which escalation to more frequent and sustained use occurs gradually among a subset of adolescents.⁷ Gradual changes in incidence are also consistent with theories of policy diffusion, which posit that the effects of major policy reforms emerge incrementally. This interpretation is also supported by the absence of evidence for an immediate increase in CUD incidence following either RCL passage or implementation in our sensitivity analysis, in which the estimated level changes were small and non-significant.

The timing of the change in CUD incidence following RCL passage, rather than at the point of retail implementation, may reflect changes in the broader legal, normative, and informational environments associated with policy adoption that influence transitions in youth substance use. For example, passage of the initiative reduced criminal penalties for cannabis-related offenses for juveniles, which may have altered perceived consequences of use among adolescents, particularly those already using frequently. Passage of the law may also have served as a normative signal, reinforcing emerging beliefs that cannabis use is socially acceptable or low-risk.²⁰ In addition, increased exposure to cannabis-related messaging during the period surrounding the 2016 ballot initiative may have heightened the salience of cannabis use and contributed to shifts in perceived acceptability or safety among adolescents. As the retail market matured, youth-appealing advertising and products proliferated,^{21, 22} along with higher-potency (Δ9-tetrahydrocannabinol) products,²³ which may increase the likelihood of progression to frequent use.²⁴ Together, these changes in legal context, messaging and marketing, normative signaling, and product diversification could plausibly facilitate gradual and sustained escalation in use intensity and subsequent CUD development without producing abrupt changes at the time of retail implementation.

Alternative explanations should also be considered. Because our outcome was based on clinically diagnosed CUD captured in electronic health records, increases in diagnosis rates may reflect changes in underlying disorder occurrence, disclosure of cannabis use, clinical recognition, or diagnostic practices. One possibility is that the increased acceptance of cannabis use post-legalization²⁰ increased adolescents’ willingness to disclose use to healthcare providers. Another is that clinicians may have become more aware of and concerned about the potential for legalization to affect use rates, changing diagnostic practices. However, to date, prior studies have generally found that legalization is associated with declines in CUD diagnosis and treatment utilization at the population level.^25–27 Qualitative analyses report lower perceived need for treatment and reduced recognition of cannabis use as problematic among adolescents, parents, and physicians.^28–30 These findings suggest that shifts in social norms and awareness could influence identification of CUD in multiple, potentially opposing directions, making it difficult to distinguish changes in underlying disorder occurrence from changes in disclosure or clinical recognition.

Several limitations warrant consideration. Our sample is limited to adolescents seeking care at KPNC, and although KPNC members are representative of the insured population of Northern California,¹¹ the findings may not generalize to adolescents in Southern California, in other states, or to uninsured adolescents. In addition, because Proposition 64 was implemented statewide, no concurrent unexposed comparison group was available to help distinguish policy-related changes from underlying temporal trends. Furthermore, because local jurisdictions differed in whether they permitted cannabis retailers, the retail sales interruption may not fully capture variation in adolescents’ exposure to legal retail availability. Although we implemented a strong quasi-experimental design and adjusted for key sociodemographic characteristics and modeled seasonal patterns, this single-group interrupted time series design cannot fully account for unmeasured secular trends occurring during the study period. Broader national trends in adolescent cannabis use behaviors and norms, adolescent mental health, clinical screening and diagnostic practices, healthcare utilization, or other contemporaneous factors may have influenced diagnosis rates independent of RCL passage or implementation. Like most observational studies, the analytic protocol was not preregistered; therefore, the findings should be interpreted with appropriate caution. Given the absence of a concurrent comparison group and potential influence of unmeasured secular trends, observed changes in diagnosis rates should be interpreted as associations with policy implementation periods rather than definitive estimates of causal effects. In addition, because CUD was ascertained from diagnoses recorded in EHR and administrative data rather than standardized research assessments, cases that were not disclosed, clinically recognized, diagnosed, or documented were not captured, and the date of diagnosis may not correspond to the onset of the disorder. Adolescents who developed CUD before enrollment in KPNC may also have been misclassified as incident cases, potentially affecting estimated incidence rates. Finally, the study period ended prior to the COVID-19 pandemic, limiting inference about longer-term post-implementation trends but avoiding the substantial confounding introduced by pandemic-related disruptions.

In summary, in this large adolescent population, incident CUD rates exhibited a distinct shift in trend following RCL passage, with continued increases that persisted through retail implementation, fully reversing earlier progress observed from 2014 to 2016. These results underscore the importance of monitoring disorder-level outcomes alongside measures of cannabis use prevalence and frequency when evaluating the public health implications of cannabis legalization. Future research should examine post-pandemic trajectories, assess whether similar patterns are observed in other populations and settings, and further investigate factors that may contribute to changes in incidence of adolescent CUD following cannabis policy changes.