BP was measured during the resting state, in triplicate with 1 min intermissions, using a random-zero sphygmomanometer at Y0–15 with the first- and fifth-phase Korotk off sounds corresponding to systolic and diastolic BP, respectively, with the average of the last two measurements used. BP at Y20 and Y25 was measured with an Omron HEM907XL oscillometer and calibrated to the random-zero readings. Body weight was measured using a calibrated balance-beam scale to the nearest 0.2 kg, with participants in light clothing. Height was measured to the nearest 0.5 cm using a vertical ruler, with BMI calculated as the weight in kilograms divided by the squared height in metres. WC was measured midway between the iliac crest and the lowest lateral portion of the rib cage. Diet was assessed using an interviewer-administered CARDIA diet questionnaire at examination Y0, Y7 and Y20 and a diet score was created, as previously described. We used two distinct methods to evaluate the metabolic effects of self-reported marijuana use in CARDIA participants. First, to assess the cross-sectional association between self-reported marijuana use and prediabetes and diabetes, data from examination Y25 were analysed. For this cross-sectional evaluation, of the 3,496 individuals present at examination Y25, we excluded those who had fasted for less than 8 h prior to the visit and those with an undeterminable diabetes status or missing relevant covariate information , resulting in a sample of 3,034 participants. Individuals who had diabetes at Y25 were excluded from prediabetes analyses , and prediabetes status was undetermined for one individual. Therefore, 2,676 individuals were included in prediabetes analyses. The second analytical approach was to prospectively evaluate the association between self-reported marijuana drying rack use and incident prediabetes and diabetes. Fasting glucose was not measured at CARDIA Y2 and Y5, and diabetes status was determined by medication use. In order to include the ADA criteria in determining incident diabetes at each examination, data from examination Y7 were used as the analysis baseline.

Individuals were excluded from analysis if they did not participate in the Y7 examination , presented with a fasting time of less than 8 h prior to the Y7 examination , did not return for follow-up in all of Y10–25 or were missing covariate information at the Y7 examination . When assessing marijuana use and incident diabetes, individuals with prevalent diabetes at Y7 or those whose diabetes status was undetermined on follow-up were excluded , resulting in an analysis sample of 3,174 participants. For the association between marijuana use and incident prediabetes, 468 people were excluded based on baseline prediabetes, diabetes and underdetermined prediabetes status on follow-up, giving a final analysis sample of 2,758 participants. Those excluded were on average older and were more likely to be male, African-American and less educated, with a longer history of smoking, higher levels of fasting glucose and CRP, and greater lifetime frequency of marijuana use compared with the included participants. Categories of all unique forms of self-reported drug use were determined by status and total use . Former use was defined as an affirmative response to the question ‘Ever use?’, but with no reported use in the previous 30 days. Current use was determined by a report of use on one or more of the last 30 days. Along with other illicit drug use, we considered several additional covariates as potential confounders. Cigarette smoking status was based entirely on current use. Regular alcohol consumption was classified as none, up to one drink daily and more than one drink daily. Educational attainment was characterised into three groups: ≤12 , 13–16 or >16 years of education . Systolic BP, BMI, WC, LDL- and HDL-cholesterol, and CRP variables were modelled continuously, as were physical activity and diet scores. Antihypertensive and lipid-lowering medication use was taken into account in models that included adjustment for BP and cholesterol levels. Given the strength of association between BMI and diabetes and to reduce potential residual confounding, all adjusted models containing BMI also included a BMI2 term to account for a possible nonlinear relationship. Participant characteristics were calculated across categories of self-reported marijuana use. Univariate models were used to assess the crude direction and magnitude of each association, with sequential models adjusting for the potential confounders noted above. The association between marijuana use and the presence of prediabetes and diabetes at CARDIA examination Y25 was estimated with logistic regression, obtaining crude and adjusted ORs and 95% CIs. For longitudinal analyses, crude and adjusted HRs and 95% CIs were estimated using Cox proportional hazards models.

Contributed person-time to the study was calculated as the duration from date of examination Y7 to either: the examination at which the event of interest was ascertained; or administrative censoring of the participant’s last examination visit. The proportional hazards assumption was assessed by including a product term between marijuana use category and natural log of contributed person-time. To investigate whether the risk of prediabetes and diabetes according to marijuana use differed by sex or race, separate multiplicative interactions were tested by adding product terms to the proportional hazards model. Sensitivity analyses were also performed, repeating the main analyses with data from different CARDIA examination years to confirm whether associations were similar regardless of the examination from which participant data were used. For example, for cross-sectional analyses, marijuana use and prevalence of prediabetes and diabetes were assessed using data from each CARDIA examination Y0–20. For prospective analyses, we assessed marijuana use at each CARDIA examination Y0–20 and incidence of prediabetes and diabetes through to Y25. Statistical analyses were performed using SAS statistical software version 9.3 . The self-reported marijuana use status of the individuals present at each examination is displayed in Fig. 1. The per cent of individuals reporting ‘never’ or ‘current’ use of marijuana declined over time, while the per cent who reported ‘former’ use of marijuana increased, particularly in the early years. Baseline participant characteristics for the prospective analysis are presented in Table 1 by category of lifetime frequency of marijuana use. In both the cross-sectional and longitudinal analyses, older age, male sex, white race, current smoking, greater daily alcohol consumption,vertical grow rack system current use of marijuana, other illicit drug use and greater participation in physical activity were all associated with a greater lifetime frequency of marijuana use, while longer time in education and greater BMI were associated with lower frequency of marijuana use. At CARDIA examination Y25, 45% of the analysis population had prediabetes . Unadjusted analysis found marijuana use was associated with higher odds of prediabetes, regardless of status or frequency of use . Specifically, individuals who reported current use and those who reported a lifetime use of ≥100 times had significantly higher odds of prediabetes compared with those who reported never using marijuana. The greatest attenuation of estimates was observed with adjustment for age, sex, and race, while the greatest strengthening of estimates was observed when use of other illicit drugs was included. There were 357 cases of prevalent diabetes identified at Y25 for the cross-sectional analysis. Without adjustment for covariates, individuals who reported a history of marijuana use when marijuana use was modelled by status or lifetime frequency had marginally lower odds of diabetes compared with never-users . Adjustment for demographic and lifestyle characteristics reversed the apparent direction of the association from <1 to >1, although the 95% CIs continued to span 1 .

Estimates were most sensitive to adjustment for alcohol use, field centre, BP and use of other illicit drugs. The results for the prediabetes and diabetes analyses did not materially change when CRP level was excluded from the models or when BMI, BMI2 and WC were inserted into the models. More than half of the participants without prediabetes or diabetes at the start of follow-up developed prediabetes over an average of 13.8 years of follow-up . Table 3 presents the crude and fully adjusted HRs with 95% CIs and crude incidence rates for prediabetes and diabetes according to self-reported marijuana use category. Unadjusted models for the association between marijuana use and incident prediabetes found a suggestive increase in the hazard for prediabetes for individuals with the greatest frequency of use at baseline . Adjustment for covariates strengthened the observed association in this group, with the 95% CIs no longer spanning 1 after adjustment for demographics, tobacco use, alcohol intake and dietary pattern. Compared with those who reported never using marijuana, individuals who reported use of ≥100 times had a significantly increased risk for prediabetes , after adjustment for demographic, lifestyle and clinical characteristics. There were 351 incident cases of diabetes identified during 50,569 years of follow-up in the prospective analysis, giving an overall crude incidence of 694 cases per 100,000 personyears. In unadjusted analysis, a decreased risk of diabetes was found for those who reported marijuana use compared with never-users, but this did not attain statistical significance. The associations were attenuated after adjustment for basic demographic and lifestyle characteristics; further adjustment for dietary pattern and BP resulted in the greatest attenuation of estimates. Irrespective of the outcome , the results did not differ when fasting glucose, BMI and pack-years of cigarette smoking were included in the final model. For all prospective analyses, inclusion of age and illicit drug use at baseline in the model resulted in considerable strengthening of estimates. Otherwise, any strengthening of the associations with the incremental inclusion of individual variables was far less in magnitude and balanced by the covariates that attenuated the associations. Formal tests of interaction were not significant for any of the potential effect modifiers for any of the analyses in this study. No violations to the proportional hazards assumption were detected. Results from sensitivity analyses confirmed the primary analyses ; patterns of the associations were similar and did not depend on the year from which participant data were used. Electronic supplementary material Table 1 shows the fasting glucose levels at the time of censoring: either the examination at which prediabetes and diabetes was ascertained or administrative censoring of the last examination visit. There was no observable linear trend in glucose levels at the time of censoring across marijuana use categories for diabetes. However, a statistically significant positive linear trend was observed for prediabetes, although this was no longer apparent after adjustment for illicit drug use. In this cohort of healthy men and women, marijuana use was associated with a higher prevalence of prediabetes during middle adulthood after controlling for potential confounding variables, but was not associated with the presence of diabetes at this age. Similarly, marijuana use in young adulthood was associated with the incidence of prediabetes in middle age. The greatest lifetime frequency of use at baseline was associated with the highest incidence of prediabetes over the study’s follow-up, compared with participants who reported never using marijuana. Marijuana use was not associated with the incidence of diabetes. Marijuana use was modelled categorically in two different ways , contributing to the interpretation and robustness of these findings. The findings of this study are important, given the previously reported associations of marijuana use with various metabolic outcomes. The impact of BMI on the association between marijuana use and incident diabetes and prediabetes is unclear . In this study, the results were unchanged with the addition of BMI, BMI2 and WC to the statistical model, consistent with the minimal estimate shift observed in a recent meta-analysis, and we found no cross-sectional association between marijuana use and BMI , in contrast to previous findings on marijuana use and metabolic health. A previous study assessed marijuana use in relation to obesity status in two population-based, nationally representative samples of US adults. Using the National Epidemiologic Survey on Alcohol and Related Conditions, researchers found that individuals who reported cannabis use on ≥3 days per week had 39% lower odds of obesity compared with individuals who reported no use in the past 12 months, after adjustment for demographics, education, marital status, religion and tobacco smoking status. This association was attenuated when researchers studied individuals from the National Comorbidity Survey—Replication; adjusted estimates no longer attained statistical significance. The prevalence of current marijuana use was <8% in this study, and the prevalence of use among young adults was below the national average and that found in our cross-sectional analysis.