Predicting Short Time-to-Crime Guns: a Machine Learning Analysis of California Transaction Records (2010-2021).

Gun-related crime continues to be an urgent public health and safety problem in cities across the US. A key question is: how are firearms diverted from the legal retail market into the hands of gun offenders? With close to 8 million legal firearm transaction records in California (2010-2020) linked to over 380,000 records of recovered crime guns (2010-2021), we employ supervised machine learning to predict which firearms are used in crimes shortly after purchase. Specifically, using random forest (RF) with stratified under-sampling, we predict any crime gun recovery within a year (0.2% of transactions) and violent crime gun recovery within a year (0.03% of transactions). We also identify the purchaser, firearm, and dealer characteristics most predictive of this short time-to-crime gun recovery using SHapley Additive exPlanations and mean decrease in accuracy variable importance measures. Overall, our models show good discrimination, and we are able to identify firearms at extreme risk for diversion into criminal hands. The test set AUC is 0.85 for both models. For the model predicting any recovery, a default threshold of 0.50 results in a sensitivity of 0.63 and a specificity of 0.88. Among transactions identified as extremely risky, e.g., transactions with a score of 0.98 and above, 74% (35/47 in the test data) are recovered within a year. The most important predictive features include purchaser age and caliber size. This study suggests the potential utility of transaction records combined with machine learning to identify firearms at the highest risk for diversion and criminal use soon after purchase.

Trends and Sources of Crime Guns in California: 2010-2021.

Firearm-related interpersonal violence is a leading cause of death and injury in cities across the United States, and understanding the movement of firearms from on-the-books sales to criminal end-user is critical to the formulation of gun violence prevention policy. In this study, we assemble a unique dataset that combines records for over 380,000 crime guns recovered by law enforcement in California (2010-2021), and more than 126,000 guns reported stolen, linked to in-state legal handgun transactions (1996-2021), to describe local and statewide crime gun trends and investigate several potentially important sources of guns to criminals, including privately manufactured firearms (PMFs), theft, and "dirty" dealers. We document a dramatic increase over the decade in firearms recovered shortly after purchase (7% were recovered within a year in 2010, up to 33% in 2021). This corresponds with a substantial rise in handgun purchasing over the decade, suggesting some fraction of newly and legally acquired firearms are likely diverted from the legal market for criminal use. We document the rapid growth of PMFs over the past 2-3 years and find theft plays some, though possibly diminishing, role as a crime gun source. Finally, we find evidence that some retailers contribute disproportionately to the supply of crime guns, though there appear to be fewer problematic dealers now than there were a decade ago. Overall, our study points to temporal shifts in the dynamics of criminal firearms commerce as well as significant city variation in the channels by which criminals acquire crime guns.

SuperMICE: An Ensemble Machine Learning Approach to Multiple Imputation by Chained Equations.

Researchers often face the problem of how to address missing data. Multiple imputation is a popular approach, with multiple imputation by chained equations (MICE) being among the most common and flexible methods for execution. MICE iteratively fits a predictive model for each variable with missing values, conditional on other variables in the data. In theory, any imputation model can be used to predict the missing values. However, if the predictive models are incorrectly specified, they may produce biased estimates of the imputed data, yielding inconsistent parameter estimates and invalid inference. Given the set of modeling choices that must be made in conducting multiple imputation, in this paper we propose a data-adaptive approach to model selection. Specifically, we adapt MICE to incorporate an ensemble algorithm, Super Learner, to predict the conditional mean for each missing value, and we also incorporate a local kernel-based estimate of variance. We present a set of simulations indicating that this approach produces final parameter estimates with lower bias and better coverage than other commonly used imputation methods. These results suggest that using a flexible machine learning imputation approach can be useful in settings where data are missing at random, especially when the relationships among the variables are complex.

Neighborhood Racial and Economic Segregation and Disparities in Violence During the COVID-19 Pandemic.

Objectives. To describe associations between neighborhood racial and economic segregation and violence during the COVID-19 pandemic. Methods. For 13 US cities, we obtained zip code-level data on 5 violence outcomes from March through July 2018 through 2020. Using negative binomial regressions and marginal contrasts, we estimated differences between quintiles of racial, economic, and racialized economic segregation using the Index of Concentration at the Extremes as a measure of neighborhood privilege (1) in 2020 and (2) relative to 2018 through 2019 (difference-in-differences). Results. In 2020, violence was higher in less-privileged neighborhoods than in the most privileged. For example, if all zip codes were in the least privileged versus most privileged quintile of racialized economic segregation, we estimated 146.2 additional aggravated assaults (95% confidence interval = 112.4, 205.8) per zip code on average across cities. Differences over time in less-privileged zip codes were greater than differences over time in the most privileged for firearm violence, aggravated assault, and homicide. Conclusions. Marginalized communities endure endemically high levels of violence. The events of 2020 exacerbated disparities in several forms of violence. Public Health Implications. To reduce violence and related disparities, immediate and long-term investments in low-income neighborhoods of color are warranted. (Am J Public Health. 2022;112(1):144-153. https://doi.org/10.2105/AJPH.2021.306540).

Unemployment and Crime in US Cities During the Coronavirus Pandemic.

Unemployment and violence both increased during the coronavirus pandemic in the United States (US), but no studies to our knowledge have examined their association. Using data for 16 US cities from January 2018 to July 2020, we estimated the association between acute changes in unemployment during the coronavirus pandemic and violent and acquisitive crime. We used negative binomial regression models and parametric g-computation to estimate average differences in crime incidents if the highest and lowest levels of unemployment observed in each city had been sustained across the exposure period (March-July 2020), compared with observed unemployment in each city-month. During the pandemic, the percentage of the adult population who were unemployed was 8.1 percentage points higher than expected, on average. Increases in unemployment were associated with increases in firearm violence and homicide. For example, we estimated an average increase of 3.3 firearm violence incidents (95% CI: - 0.2, 6.7) and 2.0 homicides (95% CI: - 0.2, 3.9) per city-month from March to July 2020 if all cities experienced their highest versus observed level of unemployment. There was no association between unemployment and aggravated assault or any acquisitive crime. Findings suggest that the sharp rise in unemployment during the pandemic may have contributed to increases in firearm violence and homicide, but not other crime. Additional research is needed on mechanisms of association, generalizability, and modifying factors.

Physical Distancing, Violence, and Crime in US Cities during the Coronavirus Pandemic.

Violent crime increased and most property crime decreased in many United States (US) cities during the coronavirus pandemic. Using negative binomial regressions, we examined the association between physical distancing (a central coronavirus containment strategy) and crime within 16 large cities (in 12 US states and the District of Columbia) through July 2020. Physical distancing was measured with aggregated smartphone data and defined as the average change in the percentage of the population staying completely at home. Outcome data were obtained from the Gun Violence Archive and city open data portals. In multivariable models, increases in the percentage of the population staying home were associated with decreases in reported incidents of aggravated assault, interpersonal firearm violence, theft, rape, and robbery, and increases in arson, burglary, and motor vehicle theft. Results suggest that changes in the frequency of interpersonal interactions affected crime during the coronavirus pandemic. More research is needed on the specificity of these assocations and their underlying mechanisms.

Armed and prohibited: characteristics of unlawful owners of legally purchased firearms.

BACKGROUND: A substantial proportion of individuals who lawfully purchase firearms later become unlawful owners ('prohibited firearm owners'), usually following events associated with an increased risk for future violence. This high-risk population has not previously been described. We aimed to characterise all individuals in California's Armed and Prohibited Persons System (APPS), a statewide programme for recovering firearms from individuals who legally purchased them and later became prohibited from ownership. METHODS: We used univariate and bivariate statistics to describe and compare prohibited firearm owners in APPS with a random sample of non-prohibited firearm owners in relation to age, sex, race/ethnicity and type of firearms owned as of 1 February 2015. We also characterised the geographical distribution of prohibited firearm owners and described their prohibitions. RESULTS: Of the 18 976 prohibited firearm owners, most were men (93%), half were white (53%) and the mean age was 47 years. Prohibited firearm owners were more likely to be male and to be black or Hispanic people than non-prohibited owners. Both prohibited and non-prohibited firearm owners had an average of 2.6 firearms, mostly handguns. Nearly half (48%) of prohibited firearm owners had a felony conviction. Extrapolating from our findings, we estimated that there are approximately 100 000 persons in the USA who unlawfully maintained ownership of their firearms following a felony conviction. CONCLUSIONS: Retention of firearms among persons who become lawfully prohibited from possessing them is common in California. Given the nationwide dearth of a programme to recover such weapons, this is likely true in other states as well.

Individual, firearm, and purchasing characteristics associated with risk of firearm-related violent crime arrest: a nested case-control study.

BACKGROUND: Firearm purchasing records offer a potentially important administrative data source to identify individuals at elevated risk of perpetrating firearm violence. In this study, we describe individual, firearm, and transaction characteristics of purchasers in California who were arrested for a firearm-related violent crime (FRV) as compared to the general population of registered purchasers in the state. METHODS: Relying on a dataset of all individuals with transaction records in California (1996-2021), linked to criminal records (1980-2021), we enrolled a cohort of individuals for whom we could capture the legal firearm purchase history. We identified those arrested for FRV post purchase, and using incidence density sampling, gender-matched cases to ten purchasers (controls) who remained "at risk" at the time the case was arrested. We focused on the purchase closest in time prior to the arrest ("index" purchase). We implemented conditional logistic regression and included models with controls for individual- and community-level demographics, as well as interactions between firearm and purchasing characteristics and criminal history. RESULTS: The cohort included 1,212,144 individuals, of whom 6153 were arrested for FRV (0.5%). Cases were matched to 61,530 controls to form the study sample. The largest risk factor was a prior criminal history: purchasers had 5.84 times the risk of FRV if they had a prior arrest within three years of the index purchase (CI 5.44-6.27). Several transaction and firearm characteristics were also associated with FRV. For example, risk increased if the firearm was redeemed at a pawn shop (aIRR: 1.37, CI 1.05-1.77) and decreased if the transaction was a registered private party transfer (vs. retail purchase) (aIRR: 0.83, CI 0.76-0.90) or the firearm was a bolt action firearm (vs. semi-automatic) (aIRR: 0.64, CI 0.51-0.79). In the interaction models, most of the purchase and firearm features only remained significant among those with no criminal history. CONCLUSIONS: Given limited data on firearm transactions, there has been little research on whether the type of firearm an individual purchases or the nature of the purchase might serve as indicators of risk for FRV. We found several transaction and firearm features were associated with risk of FRV. Notably, these features provided little evidence of additional risk for those with a prior criminal record.

Purchaser, firearm, and retailer characteristics associated with crime gun recovery: a longitudinal analysis of firearms sold in California from 1996 to 2021.

BACKGROUND: Firearm violence is a major cause of death and injury in the United States. Tracking the movement of firearms from legal purchase to use in crimes can help inform prevention of firearm injuries and deaths. The last state-wide studies analyzing crime gun recoveries used data from over 20 years ago; thus, an update is needed. METHODS: We used data for 5,247,348 handgun and 2,868,713 long gun transactions and law enforcement recoveries from California crime gun recovery (2010-2021) and California's Dealer Records of Sales records. Covariates included characteristics of dealership sales, firearms and their transactions, and purchaser's demographic characteristics, purchasing history, criminal history (from firearm purchaser criminal history records), and neighborhood socioeconomic status. Analyses for handguns and long guns was conducted separately. In multivariable analysis, we included correlates into a Cox proportional hazard model accounting for left truncation and clustering between the same firearm, purchaser, dealerships, and geographic location. Covariates that remained significant (P < 0.05) were retained. For handguns, we evaluated associations of violent and weapons crimes separately. In supplementary analyses, we examined interactions by purchasers' race and ethnicity. RESULTS: In total, 38,441 handguns (0.80%) and 6,806 long guns (0.24%) were recovered in crimes. A firearm dealer's sales volume, percent of transactions that were denials, pawns, pawn redemptions, and firearms that became crime guns were each positively associated with firearm recovery in crime. Handguns that were inexpensive, larger caliber, and that had been reported lost or stolen were positively associated with recovery in crimes. Purchaser characteristics associated with crime gun recovery included: being younger, female, Black, Hispanic, Native American or Pacific Islander, or other race/ethnicity (vs white), having previous arrests, living in close proximity to the firearm dealership, and living in a more socially vulnerable census tract. Associations with race and ethnicity were modified by previous infraction-only arrests. CONCLUSIONS: This study confirms that many previously studied correlates of firearm recovery are still relevant today. We were able to expand on previous research by examining novel associations including purchasers' criminal history and previous firearm transaction history. These results provide evidence that can be used to disrupt firearm use in crimes.

Association of medical conditions and firearm suicide among legal handgun purchasers in California: a case-control study.

BACKGROUND: Suicide is a pressing public health problem, and firearm owners are at especially elevated risk. Certain health conditions are markers of suicide risk, but more research is needed on clinical risk markers for suicide among firearm owners specifically. Our goal was to examine associations of emergency department and inpatient hospital visits for behavioral and physical health conditions with firearm suicide among handgun purchasers. METHODS: This was a case-control study of 5415 legal handgun purchasers in California who died between January 1, 2008, and December 31, 2013. Cases were firearm suicide decedents; controls were motor vehicle crash decedents. Exposures were emergency department and hospital visits for six categories of health diagnoses in the 3 years prior to death. To account for selection bias due to deceased controls, we used probabilistic quantitative bias analysis to generate bias-adjusted estimates. RESULTS: There were 3862 firearm suicide decedents and 1553 motor vehicle crash decedents. In multivariable models, suicidal ideation/attempt (OR 4.92; 95% CI 3.27-7.40), mental illness (OR 1.97; 95% CI 1.60-2.43), drug use disorder (OR 1.40; 95% CI 1.05-1.88), pain (OR 1.34; 95% CI 1.07-1.69), and alcohol use disorder (OR 1.29; 95% CI 1.01-1.65) were associated with higher odds of firearm suicide. When adjusting for all conditions simultaneously, only the associations for suicidal ideation/attempt and mental illness remained significant. Quantitative bias analysis indicated that observed associations were generally biased downward. For example, the bias-adjusted OR for suicidal ideation/attempt was 8.39 (95% simulation interval 5.46-13.04), almost twice that of the observed OR. CONCLUSIONS: Diagnoses for behavioral health conditions were markers for firearm suicide risk among handgun purchasers, even for conservative estimates that did not adjust for selection bias. Encounters with the healthcare system may provide opportunities to identify firearm owners at high risk of suicide.

Machine Learning Analysis of Handgun Transactions to Predict Firearm Suicide Risk.

Importance: Evidence suggests that limiting access to firearms among individuals at high risk of suicide can be an effective means of suicide prevention, yet accurately identifying those at risk to intervene remains a key challenge. Firearm purchasing records may offer a large-scale and objective data source for the development of tools to predict firearm suicide risk. Objective: To test whether a statewide database of handgun transaction records, coupled with machine learning techniques, can be used to forecast firearm suicide risk. Design, Setting, and Participants: This prognostic study used the California database of 4 976 391 handgun transaction records from 1 951 006 individuals from January 1, 1996, to October 6, 2015. Transaction-level random forest classification was implemented to predict firearm suicide risk, and the relative predictive power of features in the algorithm was estimated via permutation importance. Analyses were performed from December 1, 2020, to May 19, 2022. Main Outcomes and Measures: The main outcome was firearm suicide within 1 year of a firearm transaction, derived from California death records (1996-2016). With the use of California's Dealer's Records of Sale (1996-2015), 41 handgun, transaction, purchaser, and community-level predictor variables were generated. Results: There are a total of 4 976 391 transactions in the California's Dealer's Record of Sale database representing 1 951 006 individuals (1 525 754 men [78.2% of individuals]; mean [SD] age, 43.4 [13.9] years). Firearm suicide within 1 year occurred in 0.07% of handgun transactions (3278 transactions among 2614 individuals). A total of 38.6% of observed firearm suicides were among transactions classified in the highest-risk ventile (379 of 983 transactions), with 95% specificity. Among the small number of transactions with a random forest score above 0.95, more than two-thirds (24 of 35 [68.6%]) were associated with a purchaser who died by firearm suicide within 1 year. Important features included known risk factors, such as older age at first purchase, and previously unreported predictors, including distance to firearms dealer and month of purchase. Conclusions and Relevance: This prognostic study presented the first large-scale machine learning analysis of individual-level handgun transaction records. The results suggested the potential utility of such records in identifying high-risk individuals to aid suicide prevention efforts. It also identified handgun, individual, and community characteristics that have strong predictive relationships with firearm suicide and may warrant further study.

Does recreational cannabis legalization change cannabis use patterns? Evidence from secondary school students in Uruguay.

BACKGROUND AND AIMS: In 2013, Uruguay became the first country to legalize and regulate the production and distribution of cannabis for recreational use. We measured whether Uruguay's non-commercial model of recreational cannabis legalization was associated with changes in the prevalence of risky and frequent cannabis use among secondary school students. DESIGN: We used data from repeated cross-sectional surveys of secondary students in Uruguay and Chile (2007-2018). Using a difference-in-difference approach, we evaluated changes in the prevalence of past-year, past-month, any risky and frequent cannabis use following enactment (2014) and implementation (2016) of cannabis legalization among the full sample of secondary students and among students who reported past-year/month use. We examined changes separately for students ages 12 to 17, and students for whom cannabis became legally accessible, ages 18 to 21. SETTING: Uruguay and Chile (2007-2018). PARTICIPANTS: Secondary school students in 8th, 10th and 12th grade (n = 204 730). MEASUREMENTS: Past-year and past-month cannabis use; any risky cannabis use measured with the Cannabis Abuse Screening Test (CAST); and frequent cannabis use (10+ days in the past-month). FINDINGS: We found a decrease in past-year and past-month use following enactment or implementation. Among students ages 18 to 21, post-enactment, we observed a transitory increase in 2014 that decreased thereafter for: any risky use among those who reported past-year use (prevalence difference [PD] = 13.5%; 95% CI: 2.0, 24.9), frequent use in the full sample (PD = 4.5%; 95% CI: 1.0, 8.1), and frequent use among those who reported past-month use (PD = 16.8%; 95% CI: 1.9, 31.8). CONCLUSION: The legalization of recreational cannabis in Uruguay was not associated with overall increases in either past-year/past-month cannabis use or with multi-year changes in any risky and frequent cannabis use among young people.

Handgun purchasing characteristics and firearm suicide risk: a nested case-control study.

BACKGROUND: Firearms are the most lethal method of suicide and account for approximately half of all suicide deaths nationwide. We describe associations between firearm purchasing characteristics and firearm suicide. METHODS: Data on all legal handgun transactions in California from 1996 to 2015 were obtained from the California Department of Justice Dealer's Record of Sale database. Handgun purchasers were linked to mortality data to identify those who died between 1996 and 2015. To account for variation in timing and duration of observation time, analyses were stratified by birth cohort. The primary analysis focused on those aged 21-25 in 1996. A secondary analysis tested associations among those aged 50-54 in 1996. Using incidence density sampling, purchasers who died by firearm suicide (cases) were each gender-matched to 5 purchasers (controls) who remained at risk at the case's time of death. We examined the characteristics of purchasers and transactions, focusing on the transaction closest in time to the case's death. Data were analyzed with conditional logistic regression. RESULTS: There were 390 firearm suicides among the younger cohort and 512 firearm suicides among the older cohort. Across both cohorts, older age at first purchase and the purchase of a revolver were associated with greater risk of firearm suicide. For example, among the younger cohort, those who purchased a revolver versus semiautomatic pistol had 1.78 times the risk of firearm suicide (95% CI 1.32, 2.40) in multivariable models. Other associations varied across cohorts, suggesting cohort or age effects in purchasing patterns. CONCLUSIONS: Findings add to the evidence on firearm suicide risk and may help inform prevention strategies and future research.

Firearm purchasing and firearm violence during the coronavirus pandemic in the United States: a cross-sectional study.

BACKGROUND: Firearm violence is a significant public health problem in the United States. A surge in firearm purchasing following the onset of the coronavirus pandemic may have contributed to an increase in firearm violence. We sought to estimate the state-level association between firearm purchasing and interpersonal firearm violence during the pandemic. METHODS: Cross-sectional study of the 48 contiguous states and the District of Columbia from January 2018 through July 2020. Data were obtained from the National Instant Criminal Background Check System (a proxy for firearm purchasing) and the Gun Violence Archive. Using negative binomial regression models, we estimated the association between cumulative excess firearm purchases in March through July 2020 (measured as the difference between observed rates and those expected from autoregressive integrated moving average models) and injuries (including nonfatal and fatal) from intentional, interpersonal firearm violence (non-domestic and domestic violence). RESULTS: We estimated that there were 4.3 million excess firearm purchases nationally from March through July 2020 and a total of 4075 more firearm injuries than expected from April through July. We found no relationship between state-level excess purchasing and non-domestic firearm violence, e.g., each excess purchase per 100 population was associated with a rate ratio (RR) of firearm injury from non-domestic violence of 0.76 (95% CI: 0.50-1.02) in April; 0.99 (95% CI: 0.72-1.25) in May; 1.10 (95% CI: 0.93-1.32) in June; and 0.98 (95% CI: 0.85-1.12) in July. Excess firearm purchasing within states was associated with an increase in firearm injuries from domestic violence in April (RR: 2.60; 95% CI: 1.32-5.93) and May (RR: 1.79; 95% CI: 1.19-2.91), though estimates were sensitive to model specification. CONCLUSIONS: Nationwide, firearm purchasing and firearm violence increased substantially during the first months of the coronavirus pandemic. At the state level, the magnitude of the increase in purchasing was not associated with the magnitude of the increase in firearm violence. Increases in purchasing may have contributed to additional firearm injuries from domestic violence in April and May. Results suggest much of the rise in firearm violence during our study period was attributable to other factors, indicating a need for additional research.

The impact of spikes in handgun acquisitions on firearm-related harms.

BACKGROUND: Research has documented sharp and short-lived increases in firearm acquisitions immediately following high-profile mass shootings and specific elections, increasing exposure to firearms at the community level. We exploit cross-city variation in the estimated number of excess handgun acquisitions in California following the 2012 presidential election and the Sandy Hook school shooting 5 weeks later to assess whether the additional handguns were associated with increases in the rate of firearm-related harms at the city level. METHODS: We use a two-stage modeling approach. First, we estimate excess handguns as the difference between actual handgun acquisitions, as recorded in California's Dealer Record of Sales, and expected acquisitions, as predicted by a seasonal autoregressive integrated moving-average (SARIMA) time series model. We use Poisson regression models to estimate the effect of city-level excess handgun purchasing on city-level changes in rates of firearm mortality and injury. RESULTS: We estimate there were 36,142 excess handguns acquired in California in the 11 weeks following the election (95% prediction interval: 22,780 to 49,505); the Sandy Hook shooting occurred in week 6. We find city-level purchasing spikes were associated with higher rates of firearm injury in the 52 weeks post-election: a relative rate of 1.044 firearm injuries for each excess handgun per 1,000 people (95% CI: 1.000 to 1.089). This amounts to approximately 290 (95% CI: 0 to 616) additional firearm injuries (roughly a 4% increase) in California over the year. We do not detect statistically significant associations for shorter time windows or for firearm mortality. CONCLUSION: This study provides evidence for an association between excess handgun acquisitions following high-profile events and firearm injury at the community level. This suggests that even marginal increases in handgun prevalence may be impactful.

Investigating how perceived risk and availability of marijuana relate to marijuana use among adolescents in Argentina, Chile, and Uruguay over time.

AIMS: Amid changing marijuana policies in the Southern Cone, we examined relationships between marijuana-related risk factors and marijuana use among adolescents in Argentina, Chile, and Uruguay from 2001 to 2016. METHODS: Using cross-sectional surveys from 8th, 10th, and 12th graders and weighted time-varying effect models, we estimated associations between perceived risk (no/low risk versus moderate/great risk) and perceived availability (easy/very easy versus difficult/very difficult/not able to obtain) of marijuana, and any past-month marijuana use. RESULTS: In all countries, marijuana use increased over time and adolescents who perceived no/low risk and easy availability had higher odds of use. In Argentina, the bivariate risk/use association weakened from 2001 (OR = 15.24, 95%CI = 9.63, 24.12) to 2004 [OR = 3.86 (2.72, 5.48)] and strengthened until 2011 [OR = 8.22 (7.56, 10.30)]; the availability/use association strengthened from 2005 [OR = 5.32 (4.05, 6.98)] to 2009 [OR = 20.77 (15.57, 27.70)] and weakened until 2014 [OR = 11.00 (9.11, 13.27)]. In Chile, the risk/use association weakened from 2001 [OR = 7.22 (6.57, 7.95)] to 2015 [OR = 5.58 (4.82, 6.48)]; the availability/use association weakened from 2001 [OR = 5.92 (4.96, 7.06)] to 2015 [OR = 4.10 (3.15, 5.34)]. In Uruguay, the risk/use association weakened from 2003 [OR = 34.22 (22.76, 51.46)] to 2016 [OR = 6.23 (4.96, 7.83)]; the availability/use association weakened from 2005 [OR = 29.13 (13.39, 63.39) to 2007 [OR = 9.42 (3.85, 23.07)], and strengthened until 2016 [OR = 22.68 (12.03, 42.76)]. CONCLUSIONS: Overall, the association between risk and use weakened in all countries, suggesting risk perceptions became a weaker determinant of marijuana use. Perceived availability remained strongly associated with use and may become an increasingly important driver of use (particularly in Uruguay and Argentina).