Skip to main content
American Journal of Epidemiology logoLink to American Journal of Epidemiology
. 2008 Aug 6;168(6):592–601. doi: 10.1093/aje/kwn184

Adolescent Cannabis Problems and Young Adult Depression: Male-Female Stratified Propensity Score Analyses

Valerie S Harder 1,, Elizabeth A Stuart 1, James C Anthony 2
PMCID: PMC2727198  PMID: 18687663

Abstract

Cannabis use and depression are two of the most prevalent conditions worldwide. Adolescent cannabis use is linked to depression in many studies, but the effects of adolescent cannabis involvement on young adult depression remain unclear and may differ for males versus females. In this cohort study of youth from a mid-Atlantic metropolitan area of the United States, repeated assessments from 1985 (at age 6 years) through 2002 (at age 21 years) were made for 1,494 individuals (55% female). Measured covariate differences between individuals with and without cannabis problems were controlled via propensity score techniques. The estimated risk of young adult depression for adolescents with cannabis problems was not significantly different from that for comparison adolescents for either females (odds ratio = 0.7, 95% confidence interval: 0.2, 2.3) or males (odds ratio = 1.7, 95% confidence interval: 0.8, 3.6). The evidence does not support a causal association linking adolescent-onset cannabis problems with young adult depression.

Keywords: causal inference, comorbidity, mental disorders, sex factors, substance-related disorders


In the United States and elsewhere, adolescent cannabis use and problems continue to be public health concerns (1, 2), and there is speculation that cannabis use or associated problems might be contributing to an increased prevalence of young adult depressive disorders (3, 4). Whether cannabis use is actually a cause of depressive disorders remains an open question (2, 5). Some prospective studies support the idea of a link from cannabis use or problems to later depression (610). Other studies fail to support a cause-effect relation (1113). Hall and Degenhardt (14) argue that more prospective research with improved statistical analyses to better control for confounders is needed to test the potential causal relation linking cannabis and depression. The studies that have made the best attempts to control for confounders have been limited by their use of more traditional statistical methods, such as multivariable regression models (6, 9). In practice, these standard statistical approaches may be suboptimal and may lead to ill-founded causal inferences (15, 16). In contrast, this study makes use of statistical techniques specifically designed for causal inference, known as propensity score techniques (17, 18). We use propensity score techniques to estimate a suspected causal effect of adolescent-onset cannabis problems on later depression in young adulthood.

Based upon the most recent world literature on the age-of-onset distributions for cannabis problems, this study has a focus upon the developmentally important onsets during early-mid adolescence (i.e., before the age of 17 years). Epidemiologic evidence for the United States indicates that the greatest mass of these onsets will be found after the age of 17 years but that nearly 20 percent of the onsets occur before 17 years (19, 20). As such, in this study, the idea is that adolescent-onset cannabis problems might influence excess risk of depression in young adulthood. Also based upon the most recent world literature, the onset of depression and other mood disorders remains relatively uncommon until after the mid teens, with an age-related linear increase thereafter through late middle age. Epidemiologic evidence from the United States indicates that fewer than 25 percent of major depression onsets occur prior to age 19 years and that age 32 years is the median age of onset for major depression in the United States (19, 20). As such, this study's motivation might be characterized by evidence on the hypothesized possibility that adolescent-onset cannabis problems might account for an excess risk of depression during young adulthood.

The hypothesized causal effect of cannabis problems on depression is estimated through propensity score adjustment of data from a cohort study of youths followed from first grade into young adulthood, with stratification to shed light on possible male-female variation in the association, as suggested in prior research (21). The male-female stratification is motivated by prior epidemiologic studies in which there is a male excess in cannabis involvement but a female excess in occurrence of depression (2227).

In this research, we estimate a suspected causal association linking cannabis problems with depression using propensity score techniques to achieve balanced distributions of measured covariates between those with adolescent-onset cannabis problems and comparison individuals. This ensures the comparison of individuals with and without cannabis problems who are as similar as possible on the other measured covariates. Propensity score techniques are growing in popularity, and a large variety of methods are available. This paper provides an introduction to a few of the methods and our criteria for how to select among the many methods available.

MATERIALS AND METHODS

Study population

In the mid-1980s, the Prevention Research Center at The Johns Hopkins University enrolled 2,311 first-grade children in a randomized trial of two preventive interventions. The children came from 43 classrooms in 19 urban elementary schools in a Mid-Atlantic metropolitan area of the United States. All study protocols gained institutional review board approval. The resulting data are those of a longitudinal cohort study nested within the randomized prevention trial. Annual follow-up interviews were conducted from elementary school to late-middle school with subsequent follow-up in young adulthood of over 75 percent of the surviving participants. The young adult interviews contributed information for this research—a telephone interview and a subsequent in-person interview, with similar assessment protocols used in both of these two assessments (28).

Not all young adults were interviewed in both ways, but information from both young adult assessments was used in this research. Separate analyses conducted for data from the telephone and in-person assessments resulted in similar inferences and are not presented. After exclusion of individuals with missing data on adolescent cannabis involvement or young adult depression, 1,494 individuals remained in the analysis data set (826 females, 668 males). Other publications provide more detail on these young adult follow-up assessments (2832), as well as sensitivity analyses about how missing data due to attrition may affect the results of studies that use these data (33).

Measures

Exposure.

In this study, we are interested in the effect of an “exposure” variable—adolescent onset of cannabis problems. Specifically, this exposure is defined as the occurrence of cannabis problems during adolescence (before 17 years of age), where “problems” indicate either cannabis dependence or nondependent abuse. The exposure variable was dichotomized, indicating cannabis problems versus none, with onset ranging mainly between the ages of 12–16 years with four individuals with onset at age 11. The comparison group contained individuals who had never used cannabis and individuals who used cannabis but did not experience problems before age 17. The assessment was made on the basis of recall of age of first cannabis problems during the young adult interview, by using standardized items from the Composite International Diagnostic Interview (CIDI) (34) or a CIDI-like interview, depending on whether the young adult was interviewed over the telephone or in-person, respectively. The CIDI is a comprehensive, structured, diagnostic interview used by trained lay interviewers for the assessment of mental disorders as defined by the Diagnostic and Statistical Manual of Mental Disorders: DSM-IV (35).

Outcome.

The outcome was defined as a depressive episode occurring the year prior to the first available young adult interview date between the ages of 19 and 24 years. The telephone and in-person depression assessments followed the DSM-IV diagnostic criteria for major depression in that they asked a series of questions about depression and allied clinical features that occurred in the year prior to assessment.

Covariates.

Potentially confounding covariates included in the analyses consisted of demographic, socioeconomic status, other drug use, childhood disturbances of psychological well-being, parental monitoring, and behavioral intervention status variables. All covariates were modeled as either categorical or binary factors. Race was categorized into Black, White, and other, which included Hispanic, Asian, and Native American groups. Family income was categorized as low (<$5,001), moderate ($5,001–$20,000), or high (>$20,000). Free or subsidized lunch eligibility was based on school records at the time of school entry. Parental supervision and monitoring (36) were assessed via a summary score divided into four categories: low, moderate, high, or higher. Concentration problems, behavior problems, and shyness were encoded as summary scores from the Teacher Observed Classroom Adjustment—Revised questionnaire. Depression and anxiety levels from a child self-reported “How I Feel” questionnaire were categorized into low, moderate, and high. The aforementioned covariates were all assessed before the age of 12 years. Tobacco involvement was indicated by onset of daily tobacco use, alcohol involvement was assessed by indications of alcohol abuse or dependence, and other drug use refers to using any illegal drug besides cannabis. Tobacco, alcohol, and drug use covariates were assessed before the age of first cannabis problems for the cannabis problem users or before age 17 years for comparison individuals.

Missing data.

A missing category was generated within each covariate when needed, and no individuals were removed from analyses if they had missing data on a covariate. In the male-female stratified analyses, the “other” race category was combined with the “White” category because of zero values in some cells. For females, the small sample size (66 female cannabis problem users) resulted in four other cells with zero female cannabis problem users. Because these covariate values would perfectly predict cannabis problem use, those cells were combined with either the null category (daily tobacco, other illegal drug use) or with a neighboring category (behavior problems, shyness).

Statistical analyses

This study focuses on estimating the causal effect of adolescent-onset cannabis problems on the odds of young adult depression by using propensity score techniques, with stratification to capture possible male-female variation in the link between cannabis problems and depression. Results from more traditional epidemiologic analyses (multivariable logistic regressions) are also presented. Two parametric models and one nonparametric model were used to estimate the propensity score. We then applied the estimated propensity score to the final outcome logistic regression using three different application models. Details of these propensity score estimation and application techniques are included in the Models and software section below.

Decision criteria.

Although this study builds and tests nine combinations of estimation and application techniques for the propensity score-adjusted models, the reported results are limited to estimates from the better performing propensity score techniques, as determined through decision criteria based on the assessment of covariate effect sizes (37). The propensity score techniques that perform well may vary for other data sets and research questions. These decision criteria can help researchers select which method is better for their particular study. The “effect size” for a particular covariate is the difference in average covariate values between the exposed and comparison groups divided by the standard error in the exposed group. In brief, the decision criteria identify the techniques that yield the smallest effect size across the majority of the covariates and across a few theoretically critical confounding covariates, while minimizing the extreme values of effect size for all covariates. Of importance, the propensity score techniques that meet the decision criteria are chosen prior to running the final outcome regressions, thus preventing bias through the selection of a method that yields a desired result.

Average causal effect.

This article presents the estimated average causal effects of the “treatment on the treated.” In the causal inference methodology literature, the exposure variable is referred to as a “treatment,” but in this article we retain the epidemiologic terminology of exposure. The “treatment on the treated” is an estimate of the average causal effect that would be seen if everyone in the exposed group had been exposed versus no one in the exposed group being exposed. The other commonly reported average causal effect is referred to simply as the “average treatment effect” and is described elsewhere (38). In this article, we present the “treatment on the treated” estimate.

Models and software.

Multivariable logistic regression (MLR), MLR with critically chosen interaction terms (39, 40), and generalized boosted modeling (GBM), a nonparametric regression tree technique (41), were used to estimate the propensity score. Each of these techniques models cannabis problem use as a function of the measured covariates. The propensity scores are the resulting predicted probabilities of cannabis problems for each individual. One to one (1:1) matching (18), full matching (42, 43), and weighting by the odds (44) were used to apply the propensity score to the final regression. Prior to running the final logistic regressions predicting young adult depression, we compared the resulting covariate effect sizes from each of the nine combinations of estimation and application techniques utilizing the aforementioned decision criteria. For females, the two propensity score techniques that performed well were MLR paired with full matching and MLR paired with weighting by the odds. For males, GBM paired with weighting by the odds and MLR paired with weighting by the odds both performed well. For the combined sample, GBM paired with weighting by the odds performed well with regard to the decision criteria.

All statistical analyses were conducted in the R language (45). Two propensity score packages written for the R environment were used: MatchIt (46) and Twang (47). The two parametric propensity score estimation techniques used MatchIt, while the nonparametric estimation technique used Twang, which utilized the GBM package in R (48). The final logistic regression models were adjusted for the preexposure covariates used in the propensity score models to account for residual confounding. The results presented below are the propensity score-adjusted odds ratios from these logistic regressions, run for males and females separately, as well as for the combined sample.

RESULTS

Preexposure differences

Cannabis problem users (the “exposed” group) are different from comparison individuals on many measured preexposure covariates. Across males, females, and the combined sample, the cannabis problem users and comparison individuals do not appear to have markedly different preexposure depression or anxiety levels. However, a higher percentage of the cannabis problem users were daily tobacco users, had problem alcohol use, or had slightly higher concentration and behavior problems than the comparison individuals (tables 1 and 2). The application of the propensity score corrected for these imbalances, as evidenced by the decrease in all measured covariate effect sizes below 0.25 and by nonsignificant chi-squared test statistics for all covariates after propensity score adjustment (table 2).

TABLE 1.

Baseline characteristics of 1,494 adolescent-onset cannabis problem users and comparison individuals from the original 2,311 individuals in the Prevention Research Center cohort, United States, 1985–2001

Cannabis problem users
Comparison individuals
Chi square* p value, two sided
No. % No. %
Sex 63.54 <0.005
    Male 151 70 517 40
    Female 66 30 760 60
Race 19.05 <0.005
    Black 132 61 948 74
    White 84 39 315 25
    Other 1 0 14 1
Family income 1.90 0.59
    Low 19 9 115 9
    Middle 54 25 373 29
    High 71 33 381 30
    Missing 73 34 408 32
Free lunch 1.68 0.43
    No 62 29 314 25
    Yes 149 69 931 73
    Missing 6 3 32 3
Daily tobacco smoker 336.08 <0.005
    No 69 32 1,089 85
    Yes 146 67 163 13
    Missing 2 1 25 2
Alcohol abuse or dependence 376.62 <0.005
    No 113 52 1,210 95
    Yes 93 43 39 3
    Missing 11 5 28 2
Other illegal drug use 21.06 <0.005
    No 206 95 1,262 99
    Yes 7 3 5 0
    Missing 4 2 10 1
Parental monitoring 0.91 0.92
    Low 39 18 220 17
    Moderate 42 19 238 19
    High 38 18 232 18
    Higher 43 20 285 22
    Missing 55 25 302 24
Concentration problems 22.11 0.001
    Lowest 15 7 157 12
    Lower 28 13 235 18
    Low 39 18 262 21
    Moderate 47 22 256 20
    High 30 14 99 8
    Higher 7 3 17 1
    Missing 51 24 251 20
Behavior problems 41.98 <0.005
    Lower 32 15 397 31
    Low 69 32 377 30
    Moderate 30 14 171 13
    High 25 12 60 5
    Higher 10 5 21 2
    Missing 51 24 251 20
Shyness 3.24 0.70
    Lower 7 3 65 5
    Low 49 23 310 24
    Moderate 76 35 457 36
    High 31 14 171 13
    Higher 3 1 23 2
    Missing 51 24 251 20
Depression symptoms§ 4.40 0.22
    Low 29 13 182 14
    Moderate 117 54 764 60
    High 14 6 64 5
    Missing 57 26 267 21
Anxiety symptoms§ 5.47 0.14
    Low 45 21 235 18
    Moderate 94 43 657 51
    High 21 10 118 9
    Missing 57 26 267 21
Intervention status (classroom) 0.31 0.86
    Standard setting 129 59 736 58
    Good behavior game 42 19 266 21
    Mastery learning 46 21 275 22
Intervention status (school) 0.04 0.98
    Standard setting 60 28 347 27
    Good behavior game 75 35 439 34
    Mastery learning 82 38 491 38
*

First category in each covariate is the reference. Fisher's exact tests used when cells have less than five individuals.

Other race includes Hispanics, Asians, and Native Americans.

Based on the standardized Teacher Observed Classroom Adjustment-Revised teacher's rating summary score.

§

Based on a child's self-reported mood questionnaire summary score.

TABLE 2.

Balance of baseline characteristics by males and females separately, before and after propensity score adjustment of 1,494 adolescent-onset cannabis problem users and comparison individuals from the original 2,311 individuals in the Prevention Research Center cohort, United States, 1985–2001

Males
Females
Cannabis problem users
Comparison individuals
Chi square, unadjusted Chi square, propensity score adjusted Cannabis problem users
Comparison individuals
Chi square, unadjusted Chi square, propensity score adjusted
No. % unadjusted No. % unadjusted % propensity score adjusted No. % unadjusted No. % unadjusted % propensity score adjusted
Race 2.42 0.11 19.54* 0.32
    Black 99 66 372 72 67 33 50 576 76 52
    Other§ 52 34 145 28 33 33 50 184 24 48
Family income 0.97 0.56 2.36 5.61
    Low 17 11 48 9 12 2 3 67 9 0
    Middle 38 25 137 26 23 16 24 236 31 24
    High 49 32 157 30 33 22 33 224 29 34
    Missing 47 31 175 34 33 26 39 233 31 41
Free lunch 0.58 2.78 4.40 2.26
    No 39 26 135 26 33 23 35 179 24 39
    Yes 110 73 370 72 67 39 59 561 74 58
    Missing 2 1 12 2 0 4 6 20 3 3
Daily tobacco smoker 145.60** 6.62 165.73** 2.95
    No 51 34 428 83 44 18 27 675 89 33
    Yes 98 65 78 15 56 48 73 85 11 67
    Missing 2 1 11 2 0
Alcohol abuse or dependence 202.00** 8.59 81.27** 9.47
    No 67 44 481 93 55 46 70 729 96 72
    Yes 77 51 24 5 39 16 24 15 2 17
    Missing 7 5 12 2 7 4 6 16 2 10
Other illegal drug use 6.61 0.52 25.59 3.71
    No 145 96 512 99 96 61 92 757 100 95
    Yes 2 1 2 0 2 5 8 3 0 5
    Missing 4 3 3 1 2
Parental monitoring 4.06 8.42 9.4 8.73
    Low 29 19 82 16 16 10 15 138 18 21
    Moderate 33 22 101 20 18 9 14 137 18 14
    High 29 19 82 16 16 9 14 150 20 10
    Higher 31 21 123 24 25 12 18 162 21 14
    Missing 29 19 129 25 25 26 39 173 23 41
Concentration problems 6.49 6.98 9.43 13.12
    Lowest 4 3 36 7 2 11 17 121 16 12
    Lower 19 13 82 16 16 9 14 153 20 20
    Low 30 20 93 18 20 9 14 169 22 12
    Moderate 36 24 117 23 25 11 17 139 18 17
    High 24 16 64 12 11 6 9 35 5 7
    Higher 6 4 14 3 2 1 2 3 0 0
    Missing 32 21 111 21 23 19 29 140 18 32
Behavior problems 20.92* 3.36 5.11 5.72
    Lower 13 9 104 20 10 19 29 293 39 33
    Low 50 33 158 31 36 19 29 219 29 24
    Moderate 24 16 92 18 12 6 9 79 10 7
    High 22 15 37 7 12 3 5 29 4 3
    Higher 10 7 15 3 7
    Missing 32 21 111 21 24 19 29 140 18 33
Shyness 3.39 7.5 5.27 8.97
    Lower 2 1 20 4 2 5 8 45 6 7
    Low 36 24 102 20 19 13 20 208 27 25
    Moderate 55 36 197 38 40 21 32 260 34 27
    High 23 15 78 15 14 8 12 107 14 8
    Higher 3 2 9 2 0
    Missing 32 21 111 21 24 19 29 140 18 32
Depression symptoms 3.09 2.06 5.65 17.48
    Low 22 15 81 16 16 7 11 101 13 19
    Moderate 83 55 300 58 54 34 52 464 61 45
    High 10 7 18 3 5 4 6 46 6 2
    Missing 36 24 118 23 25 21 32 149 20 34
Anxiety symptoms 1.30 3.87 10.48 12.56
    Low 36 24 104 20 20 9 14 131 17 10
    Moderate 70 46 261 50 45 24 36 396 52 44
    High 9 6 34 7 9 12 18 84 11 12
    Missing 36 24 118 23 25 21 32 149 20 34
Intervention status (classroom) 2.88 0.06 4.38 0.03
    Standard setting 89 59 305 59 59 40 61 431 57 60
    Good behavior game 24 16 108 21 16 18 27 158 21 28
    Mastery learning 38 25 104 20 25 8 12 171 23 12
Intervention status (school) 0.83 6.38 3.05 0.03
    Standard setting 37 25 144 28 28 23 35 203 27 34
    Good behavior game 51 34 175 34 26 24 36 264 35 36
    Mastery learning 63 42 198 38 47 19 29 293 39 29
*

p < 0.01;

**

p < 0.005, two sided.

Cells with zero individuals are combined with the reference group if binary or combined with the nearest neighbor if categorical.

First category in each covariate is the reference. Fisher's exact tests are used when cells have less than five individuals.

§

Other race includes Whites and other non-Blacks.

Propensity score adjustment

The final estimated odds ratios from the propensity score-adjusted regression models that met the decision criteria are presented in table 3. Female cannabis problem users experienced a modestly lower prevalence of major depression, while male problem users experienced a modestly higher prevalence of major depression, but the variation was not statistically significant by conventional frequentist standards (p > 0.05). The other propensity score-adjusted “treatment on the treated” models mentioned above (those not selected by the decision criteria) produce similar odds ratio estimates (male odds ratio range = 1.6–2.1; female odds ratio range = 0.6–1.1), with only one odds ratio of 18 with p < 0.05. The final odds ratios from the propensity score-adjusted models for the combined sample are all slightly above the null (odds ratio range = 1.1–1.8) with two of the nine with p < 0.05.

TABLE 3.

Estimated association by males and females separately, linking young adult depression with adolescent-onset cannabis problems, with covariate adjustment and use of propensity score techniques for 1,494 individuals from the Prevention Research Center cohort, United States, 1985–2001

Propensity score adjustment models No. of adolescent cannabis problem users Odds ratio 95% confidence interval p value, two sided
Males
    GBM* and weighting by the odds 151 1.72 0.77, 3.86 0.19
    MLR* and weighting by the odds 1.67 0.77, 3.60 0.19
Females
    MLR and full matching 66 0.63 0.25, 1.58 0.32
    MLR and weighting by the odds 0.68 0.20, 2.34 0.54
Combined sample
    GBM and weighting by the odds 217 1.33 0.76, 2.33 0.32
*

GBM, generalized boosted modeling propensity score estimation technique; MLR, multivariable logistic regression propensity score estimation technique.

Traditional adjustment

The more traditional epidemiologic regression model for these data, multivariable logistic regression, produces results that are similar in the combined sample but slightly different in the male-female stratified subgroups. For males, the traditional odds ratio estimate comparing young adult depression among adolescents with and without cannabis problem use is over 2 and is statistically significant (odds ratio = 2.6; p < 0.01). For females, the result is essentially null (odds ratio = 0.9; p = 0.72). The cannabis-depression estimate for the combined sample is positive but not significant (odds ratio = 1.5; p = 0.11). The propensity score-adjusted analyses are generally preferred because they ensure the similarity of covariates between the exposed and comparison groups.

DISCUSSION

Studying a sample of youth followed from childhood through adolescence and into adulthood, we found essentially null associations linking adolescent-onset cannabis problems with later young adult depression. Propensity score techniques were used to estimate the causal effects in the combined sample as well as separately for males and females. Results were confirmed through sensitivity analyses by using traditional multivariable logistic regressions. The magnitude of the association was found to be lower for females than males, but with little evidence of statistically robust associations for either males or females. Previous research examining similar questions suggested that female cannabis users might be slightly more likely than males to experience depression in adulthood (8, 13, 49). Our relatively small sample size of female cannabis problem users (n = 66) may be responsible for the qualitatively different results between males and females. There is a clear need for study replication with larger sample sizes.

Our study does not support the hypothesis that adolescent-onset cannabis problem use causes young adult depression. Two other causal hypotheses remain: 1) Depression causes individuals to manage their symptoms through self-medication by use of cannabis, and 2) a common genetic or environmental influence causes both depression and cannabis use. Although there is consistent evidence that depression does not cause cannabis use (50), there is evidence in support of the common cause hypothesis through the use of co-twin methodology to control for genetic influences (51). Our findings do not rule out the common cause hypothesis and, in fact, may add support to it by virtue of ruling out the hypothesis that adolescent cannabis problem use might be functioning as a causal factor for young adult depression.

Results from a recent large national survey in the United States suggest that the relation between cannabis use and depression may be explained by associations between cannabis use and bipolar disorders, which are also associated with other drug use disorders (5). In our study, preexposure reports of other drug problem use (tobacco, alcohol, or illegal drugs) were controlled. This additional statistical control might explain why this report is not entirely consistent with what has been observed by others, such as Stinson et al. (5). Together with these findings and the findings from other studies also reporting null associations between cannabis use and depression (1113), it appears that there is mounting evidence against the hypothesized causal association. Even if this study lacked power to detect a causal association, there may be a small association for males and for the combined sample, but it is unlikely that there is a direct causal pathway linking adolescent-onset cannabis problems to young adult depression.

Continued efforts to resolve the debate over the nature of the association between cannabis and depression are warranted, given recent reports of the link between cannabis and psychosis. Schizophrenia (psychosis) has been linked with cannabis problems in longitudinal studies (5255). Four recent reviews (5659) agree, and one review disagrees (60) with the claim of causality. It should be noted that none of the aforementioned studies applied propensity scores or other causal inference statistical techniques. Regardless, there is a general overall consensus that cannabis use may be a contributory cause of psychosis among individuals with susceptibility to psychosis (61). This situation of uncertainty motivates a more complete examination of the evidence about whether cannabis problems cause other mental health disorders, such as depression.

The findings of this study must be interpreted in light of potential methodological limitations. For example, in this study, the age at the occurrence of first cannabis problems was assessed retrospectively during the young adult interview, concurrently with the assessment of the age at first depressive episode. As such, the link from cannabis problems to depression is not strictly longitudinal, although we used ancillary information about age at onset to be sure that the cannabis problems had occurred before the age of 17 years, and depression was assessed afterwards. Fortunately, there is some evidence that recalled age of cannabis involvement can be measured reliably (62, 63). Another limitation of the study is that we could not balance unobserved (unmeasured) covariates. Researchers may be hesitant to apply propensity score techniques because of the major limitation that they do not control for unmeasured covariates. However, this concern over potential unmeasured confounders is common to both propensity score techniques and traditional multivariable regression applied to observational data. Therefore, because the same limitation applies, it is not a special concern for this study in particular. Given some findings that males experience onset of depression at a later age than females do (64), it is possible that our limited range and relatively young age for assessing depression have resulted in a downwardly biased estimate for the occurrence of male depression and a resulting effect estimate biased toward the null. This question can be addressed in the ongoing follow-up of this cohort in the future. Another limitation of this study is the length of time between surveys. Our ability to define the exposure during adolescence and the outcome in young adulthood leaves a gap of as little as 3 years to as large as 11 years between an individual's age at first cannabis problems and that individual's age at the young adult interview. Future research applying longitudinal propensity score techniques (65) to questions of causation linking cannabis and depression may be possible if an appropriate data set is identified with repeated measures of cannabis use and depression across the developmental period from adolescence into adulthood.

One of the strengths of this study involves its capacity to control for many potential confounding variables measured over a long span of time (i.e., over 15 years). As such, the nature of these data has allowed us to control for numerous critical preexposure confounding covariates (other drug use, childhood psychological distress, and socioeconomic factors). In addition, the use of structured diagnostic interviews allowed for the assessment of clinically relevant definitions of both cannabis problems and depression. A final strength of the present study involves the use of propensity scores, a fairly recent causal inference statistical method. The process of first estimating the propensity score and then later applying the propensity score allows for the evaluation of several propensity score estimation techniques based on the balance of the preexposure covariates prior to running the final propensity score-adjusted outcome model.

In conclusion, the evidence from propensity score adjusted analyses does not support the hypothesized causal link between adolescent-onset cannabis problem use and young adult depression. If adolescent cannabis problem use is causing some cases of young adult depression, the causal link is modest at best and may be limited to males. On the basis of this study's evidence, cannabis problems do not appear to contribute to depression among young adult females. If we are able to prevent adolescents from developing problem cannabis use, we may see very little reduction in the occurrence or prevalence of depressive episodes. If there is any cause-effect relation, it might be more readily found among males compared with females.

Acknowledgments

This research and resulting article were supported by an Individual Ruth L. Kirschstein (F31) National Research Service Award, DA021956 (Principal Investigator: V. Harder), National Institute of Drug Abuse (NIDA). Dr. Stuart's time was supported by the Center for Prevention and Early Intervention, jointly funded by the National Institute of Mental Health and NIDA (grant MH066247; Principal Investigator: N. Ialongo), and by the Centers for Disease Control and Prevention-funded Center for the Prevention of Youth Violence (grant 5U49CE000728; Principal Investigator: P. Leaf). Dr. Anthony's time was supported by his K05 Senior Scientist Award from NIDA, DA015799, and his R01 research grant, DA015799. He also was a Co-Director of the original Prevention Research Center and served as Principal Investigator for the NIDA research grant awards that supported assessment of cannabis involvement, major depression, and other constructs between 1989 and 2002 (R01DA004392 and R01DA009897).

The authors thank Dr. Nicholas Ialongo and the Prevention Research Center at The Johns Hopkins University for use of the data from the National Institutes of Mental Health-sponsored study (MH57005).

Conflict of interest: none declared.

Glossary

Abbreviations

CIDI

Composite International Diagnostic Interview

GBM

generalized boosted modeling

MLR

multivariable logistic regression

References

  • 1.Johnston LD, O'Malley PM, Bachman JG, et al. Bethesda, MD: National Institute on Drug Abuse; 2007. Monitoring the future national results on adolescent drug use: overview of key findings, 2006. (NIH publication no. 07-6202) [Google Scholar]
  • 2.Degenhardt L, Hall W, Lynskey M. Exploring the association between cannabis use and depression. Addiction. 2003;98:1493–504. doi: 10.1046/j.1360-0443.2003.00437.x. [DOI] [PubMed] [Google Scholar]
  • 3.Compton WM, Conway KP, Stinson FS, et al. Changes in the prevalence of major depression and comorbid substance use disorders in the United States between 1991 –1992 and 2001–2002. Am J Psychiatry. 2006;163:2141–7. doi: 10.1176/ajp.2006.163.12.2141. [DOI] [PubMed] [Google Scholar]
  • 4.Moore TH, Zammit S, Lingford-Hughes A, et al. Cannabis use and risk of psychotic or affective mental health outcomes: a systematic review. Lancet. 2007;370:319–28. doi: 10.1016/S0140-6736(07)61162-3. [DOI] [PubMed] [Google Scholar]
  • 5.Stinson FS, Ruan WJ, Pickering R, et al. Cannabis use disorders in the USA: prevalence, correlates and co-morbidity. Psychol Med. 2006;36:1447–60. doi: 10.1017/S0033291706008361. [DOI] [PubMed] [Google Scholar]
  • 6.Brook D, Brook J, Zhang C, et al. Drug use and the risk of major depressive disorder, alcohol dependence, and substance use disorders. Arch Gen Psychiatry. 2002;59:1039–44. doi: 10.1001/archpsyc.59.11.1039. [DOI] [PubMed] [Google Scholar]
  • 7.Bovasso G. Cannabis abuse as a risk factor for depressive symptoms. Am J Psychiatry. 2001;158:2033–7. doi: 10.1176/appi.ajp.158.12.2033. [DOI] [PubMed] [Google Scholar]
  • 8.Patton GC, Coffey C, Carlin JB, et al. Cannabis use and mental health in young people: cohort study. BMJ. 2002;325:1195–8. doi: 10.1136/bmj.325.7374.1195. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Fergusson DM, Horwood LJ, Swain-Campbell N. Cannabis use and psychosocial adjustment in adolescence and young adulthood. Addiction. 2002;97:1123–35. doi: 10.1046/j.1360-0443.2002.00103.x. [DOI] [PubMed] [Google Scholar]
  • 10.Georgiades K, Boyle MH. Adolescent tobacco and cannabis use: young adult outcomes from the Ontario Child Health Study. J Child Psychol Psychiatry. 2007;48:724–31. doi: 10.1111/j.1469-7610.2007.01740.x. [DOI] [PubMed] [Google Scholar]
  • 11.Brook JS, Cohen P, Brook DW. Longitudinal study of co-occurring psychiatric disorders and substance use. J Am Acad Child Adolesc Psychiatry. 1998;37:322–30. doi: 10.1097/00004583-199803000-00018. [DOI] [PubMed] [Google Scholar]
  • 12.Fergusson DM, Horwood JL. Early onset cannabis use and psychosocial adjustment in young adults. Addiction. 1997;92:279–96. [PubMed] [Google Scholar]
  • 13.Harder VS, Morral AR, Arkes J. Marijuana use and depression among adults: testing for causal associations. Addiction. 2006;101:1463–72. doi: 10.1111/j.1360-0443.2006.01545.x. [DOI] [PubMed] [Google Scholar]
  • 14.Hall W, Degenhardt L. Prevalence and correlates of cannabis use in developed and developing countries. Curr Opin Psychiatry. 2007;20:393–7. doi: 10.1097/YCO.0b013e32812144cc. [DOI] [PubMed] [Google Scholar]
  • 15.Rubin DB. Estimating causal effects of treatments in randomized and nonrandomized studies. J Educ Psychol. 1974;66:688–701. [Google Scholar]
  • 16.Rubin DB. On principles for modeling propensity scores in medical research. Pharmacoepidemiol Drug Saf. 2004;13:855–7. doi: 10.1002/pds.968. [DOI] [PubMed] [Google Scholar]
  • 17.Rosenbaum PR, Rubin DB. The central role of the propensity score in observational studies for causal effects. Biometrika. 1983;70:41–55. [Google Scholar]
  • 18.Stuart EA, Rubin DB. Matching methods for causal inference: designing observational studies. Best practices in quantitative methods. In: Osborne J, editor. Thousand Oaks, CA: Sage Publications; 2007. pp. 155–76. [Google Scholar]
  • 19.Kessler RC, Berglund P, Demler O, et al. Lifetime prevalence and age-of-onset distributions of DSM-IV disorders in the National Comorbidity Survey Replication. Arch Gen Psychiatry. 2005;62:593–602. doi: 10.1001/archpsyc.62.6.593. [DOI] [PubMed] [Google Scholar]
  • 20.Kessler RC, Angermeyer M, Anthony JC, et al. Lifetime prevalence and age-of-onset distributions of mental disorders in the World Health Organization's World Mental Health Survey Initiative. World Psychiatry. 2007;6:168–76. [PMC free article] [PubMed] [Google Scholar]
  • 21.Chen CY, Wagner FA, Anthony JC. Marijuana use and the risk of major depressive episode. Epidemiological evidence from the United States National Comorbidity Survey. Soc Psychiatry Psychiatr Epidemiol. 2002;37:199–206. doi: 10.1007/s00127-002-0541-z. [DOI] [PubMed] [Google Scholar]
  • 22.Eaton WW, Kramer M, Anthony JC, et al. The incidence of specific DIS/DSM-III mental disorders: data from the NIMH Epidemiologic Catchment Area Program. Acta Psychiatr Scand. 1989;79:163–78. doi: 10.1111/j.1600-0447.1989.tb08584.x. [DOI] [PubMed] [Google Scholar]
  • 23.Anthony JC, Warner LA, Kessler RC. Comparative epidemiology of dependence on tobacco, alcohol, controlled substances, and inhalants: basic findings from the National Comorbidity Survey. Exp Clin Psychopharmacol. 1994;2:244–68. [Google Scholar]
  • 24.O'Malley PM, Johnston LD, Bachman JG. Adolescent substance abuse: epidemiology and implications for public policy. Pediatr Clin North Am. 1995;42:241–60. doi: 10.1016/s0031-3955(16)38945-3. [DOI] [PubMed] [Google Scholar]
  • 25.Van Etten ML, Anthony JC. Male-female differences in transitions from first drug opportunity to first use: searching for subgroup variation by age, race, region, and urban status. J Womens Health Gend Based Med. 2001;10:797–804. doi: 10.1089/15246090152636550. [DOI] [PubMed] [Google Scholar]
  • 26.Regier DA, Narrow WE, Rae DS, et al. The de facto US mental and addictive disorders service system. Epidemiologic catchment area prospective 1-year prevalence rates of disorders and services. Arch Gen Psychiatry. 1993;50:85–94. doi: 10.1001/archpsyc.1993.01820140007001. [DOI] [PubMed] [Google Scholar]
  • 27.Kessler RC, Berglund P, Demler O, et al. The epidemiology of major depressive disorder: results from the National Comorbidity Survey Replication (NCS-R) JAMA. 2003;289:3095–105. doi: 10.1001/jama.289.23.3095. [DOI] [PubMed] [Google Scholar]
  • 28.Reed PL, Storr CL, Anthony JC. Drug dependence enviromics: job strain in the work environment and risk of becoming drug-dependent. Am J Epidemiol. 2006;163:404–11. doi: 10.1093/aje/kwj064. [DOI] [PubMed] [Google Scholar]
  • 29.Ialongo NS, Werthamer L, Kellam SG, et al. Proximal impact of two first-grade preventive interventions on the early risk behaviors for later substance abuse, depression, and antisocial behavior. Am J Community Psychol. 1999;27:599–641. doi: 10.1023/A:1022137920532. [DOI] [PubMed] [Google Scholar]
  • 30.Storr CL, Reboussin BA, Anthony JC. Early childhood misbehavior and the estimated risk of becoming tobacco-dependent. Am J Epidemiol. 2004;160:126–30. doi: 10.1093/aje/kwh184. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Wilcox HC, Anthony JC. The development of suicide ideation and attempts: an epidemiologic study of first graders followed into young adulthood. Drug Alcohol Depend. 2004;76(suppl):S53–67. doi: 10.1016/j.drugalcdep.2004.08.007. [DOI] [PubMed] [Google Scholar]
  • 32.Reed PL, Anthony JC, Breslau N. Incidence of drug problems in young adults exposed to trauma and posttraumatic stress disorder: do early life experiences and predispositions matter? Arch Gen Psychiatry. 2007;64:1435–42. doi: 10.1001/archpsyc.64.12.1435. [DOI] [PubMed] [Google Scholar]
  • 33.Scharfstein DO, Manski CF, Anthony JC. On the construction of bounds in prospective studies with missing ordinal outcomes: application to the good behavior game trial. Biometrics. 2004;60:154–64. doi: 10.1111/j.0006-341X.2004.00158.x. [DOI] [PubMed] [Google Scholar]
  • 34.World Health Organization. Composite. Geneva, Switzerland: World Health Organization; 1990. International Diagnostic Interview. [Google Scholar]
  • 35.American Psychiatric Association. 4th ed. Washington, DC: American Psychiatric Association; 1994. Diagnostic and statistical manual of mental disorders: DSMIV. [Google Scholar]
  • 36.Chilcoat HD, Anthony JC. Impact of parent monitoring on initiation of drug use through late childhood. J Am Acad Child Adolesc Psychiatry. 1996;35:91–100. doi: 10.1097/00004583-199601000-00017. [DOI] [PubMed] [Google Scholar]
  • 37.Ho D, Imai K, King G, et al. Matching as nonparametric preprocessing for reducing model dependence in parametric causal inference. Polit Anal. 2007;15:199–236. [Google Scholar]
  • 38.Imbens G. Nonparametric estimation of average treatment effects under exogeneity: a review. Rev Econ Stat. 2004;86:4–29. [Google Scholar]
  • 39.Dehejia RH, Wahba S. Cambridge, MA: National Bureau of Economic Research; 1998. Propensity score matching methods for non-experimental causal studies. (NBER working paper, no. 6829). ( http://www.nber.org/papers/w6829.pdf). (Accessed May12, 2008) [Google Scholar]
  • 40.Dehejia RH, Wahba S. Causal effects in nonexperimental studies: reevaluating the evaluation of training programs. J Am Stat Assoc. 1999;94:1053–62. [Google Scholar]
  • 41.McCaffrey DF, Ridgeway G, Morral AR. Propensity score estimation with boosted regression for evaluating causal effects in observational studies. Psychol Methods. 2004;9:403–25. doi: 10.1037/1082-989X.9.4.403. [DOI] [PubMed] [Google Scholar]
  • 42.Hansen BB. Full matching in an observational study of coaching for the SAT. J Am Stat Assoc. 2004;99:609–18. [Google Scholar]
  • 43.Stuart EA, Green KM. Using full matching to estimate causal effects in non-experimental studies: examining the relationship between adolescent marijuana use and adult outcomes. Dev Psychol. 2008;44:395–406. doi: 10.1037/0012-1649.44.2.395. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44.Hirano K, Imbens G, Ridder G. Efficient estimation of average treatment effects using the estimated propensity score. Econometrica. 2003;71:1161–89. [Google Scholar]
  • 45.R Development Core Team. Vienna, Austria: R Foundation for Statistical Computing; 2003. R: a language and environment for statistical computing. ( http://www.cran.r-project.org). (Accessed May12, 2008) [Google Scholar]
  • 46.Ho D, Stuart E, Imai K, et al. Vienna, Austria: R Foundation for Statistical Computing; 2008. MatchIt: MatchIt. ( http://cran.r-project.org/web/packages/MatchIt/index.html). (Accessed July18, 2008) [Google Scholar]
  • 47.Ridgeway G, McCaffrey D, Morral A. Vienna, Austria: R Foundation for Statistical Computing; 2006. Twang: Toolkit for Weighting and Analysis of Nonequivalent Groups. R package version 1.0–1. ( http://www.cran.r-project.org). (Accessed May12, 2008) [Google Scholar]
  • 48.Ridgeway G. Vienna, Austria: R Foundation for Statistical Computing; 2007. GBM 1.6-3 package manual. ( http://cran.r-project.org/web/packages/gbm/index.html). (Accessed July18, 2008) [Google Scholar]
  • 49.Rowe MG, Fleming MF, Barry KL, et al. Correlates of depression in primary care. J Fam Pract. 1995;41:551–8. [PubMed] [Google Scholar]
  • 50.Kalant H. Adverse effects of cannabis on health: an update of the literature since 1996. Prog Neuropsychopharmacol Biol Psychiatry. 2004;28:849–63. doi: 10.1016/j.pnpbp.2004.05.027. [DOI] [PubMed] [Google Scholar]
  • 51.Lynskey MT, Glowinski AL, Todorov AA, et al. Major depressive disorder, suicidal ideation, and suicide attempt in twins discordant for cannabis dependence and early-onset cannabis use. Arch Gen Psychiatry. 2004;61:1026–32. doi: 10.1001/archpsyc.61.10.1026. [DOI] [PubMed] [Google Scholar]
  • 52.van Os J, Bak M, Hanssen M, et al. Cannabis use and psychosis: a longitudinal population-based study. Am J Epidemiol. 2002;156:319–27. doi: 10.1093/aje/kwf043. [DOI] [PubMed] [Google Scholar]
  • 53.Zammit S, Allebeck P, Andreasson S, et al. Self reported cannabis use as a risk factor for schizophrenia in Swedish conscripts of 1969: historical cohort study. BMJ. 2002;325:1199–201. doi: 10.1136/bmj.325.7374.1199. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 54.Fergusson DM, Horwood LJ, Swain-Campbell N. Cannabis use and psychosocial adjustment in adolescence and young adulthood. Addiction. 2002;97:1123–35. doi: 10.1046/j.1360-0443.2002.00103.x. [DOI] [PubMed] [Google Scholar]
  • 55.Andreasson S, Engstrom A, Allebeck P, et al. Cannabis and schizophrenia: a longitudinal study of Swedish conscripts. Lancet. 1987;2:1483–6. doi: 10.1016/s0140-6736(87)92620-1. [DOI] [PubMed] [Google Scholar]
  • 56.Arseneault L, Cannon M, Witton J, et al. Causal association between cannabis and psychosis: examination of the evidence. Br J Psychiatry. 2004;184:110–17. doi: 10.1192/bjp.184.2.110. [DOI] [PubMed] [Google Scholar]
  • 57.Henquet C, Murray R, Linszen D, et al. The environment and schizophrenia: the role of cannabis use. Schizophr Bull. 2005;31:608–12. doi: 10.1093/schbul/sbi027. [DOI] [PubMed] [Google Scholar]
  • 58.Smit F, Bolier L, Cuijpers P. Cannabis use and the risk of later schizophrenia: a review. Addiction. 2004;99:425–30. doi: 10.1111/j.1360-0443.2004.00683.x. [DOI] [PubMed] [Google Scholar]
  • 59.Semple DM, McIntosh AM, Lawrie SM. Cannabis as a risk factor for psychosis: systematic review. J Psychopharmacol. 2005;19:187–94. doi: 10.1177/0269881105049040. [DOI] [PubMed] [Google Scholar]
  • 60.MacLeod J, Oakes R, Copello A, et al. Psychological and social sequelae of cannabis and other illicit drug use by young people: a systematic review of longitudinal, general population studies. Lancet. 2004;363:1579–88. doi: 10.1016/S0140-6736(04)16200-4. [DOI] [PubMed] [Google Scholar]
  • 61.Hall WD. Cannabis use and the mental health of young people. Aust N Z J Psychiatry. 2006;40:105–13. doi: 10.1080/j.1440-1614.2006.01756.x. [DOI] [PubMed] [Google Scholar]
  • 62.Shillington AM, Clapp JD. Self-report stability of adolescent substance use: are there differences for gender, ethnicity and age? Drug Alcohol Depend. 2000;60:19–27. doi: 10.1016/s0376-8716(99)00137-4. [DOI] [PubMed] [Google Scholar]
  • 63.Shillington AM, Cottler LB, Mager DE, et al. Self-report stability for substance use over 10 years: data from the St. Louis Epidemiologic Catchment Study. Drug Alcohol Depend. 1995;40:103–9. doi: 10.1016/0376-8716(95)01176-5. [DOI] [PubMed] [Google Scholar]
  • 64.Marcus SM, Young EA, Kerber KB, et al. Gender differences in depression: findings from the STAR*D study. J Affect Disord. 2005;87:141–50. doi: 10.1016/j.jad.2004.09.008. [DOI] [PubMed] [Google Scholar]
  • 65.Robins JM, Hernan MA, Brumback B. Marginal structural models and causal inference in epidemiology. Epidemiology. 2000;11:550–60. doi: 10.1097/00001648-200009000-00011. [DOI] [PubMed] [Google Scholar]

Articles from American Journal of Epidemiology are provided here courtesy of Oxford University Press

RESOURCES