Part 1. Review of published data and meta-analysis

Methods

Method of selecting studies for inclusion in the review

Articles appropriate for the review were selected by several methods. First, literature searches were conducted in PsycINFO, Medline, and Web of Science using combinations of the keywords "adolescent," "adolescence," "drug or substance," "use or abuse," and "psychiatric comorbidity." The search was limited to articles in the English language. Literature that focused on parental drug abuse as a predictor of adolescent drug abuse were excluded, as were manuscripts from Dissertation Abstracts International. Second, the bibliographies of these articles were examined for the purpose of yielding additional articles. Third, the in-house reference library of the Center for Developmental Epidemiology was searched. Finally, advisors to the Center were asked for references that might be appropriate for our purposes. One hundred forty-one papers were found that contained all the keywords or were found through the other methods described. The list can be found in appendix A.

Selection criteria for meta-analysis

Two criteria had to be met before papers were included in the list for meta-analysis: community-based sampling and formal psychiatric diagnostic procedures. We had hoped to include only studies that contained information on the temporal ordering or risk and protective factors, but too few met this criterion, so it was dropped.

Back to Top

1. Community-based samples

Information about psychiatric disorders that co-occur with substance use or abuse comes from two main sources: clinical studies and studies using representative population-based samples. This meta-analysis is based on the latter. While clinic-based studies of comorbidity can be valuable for suggesting developmental pathways, causal patterns, or likely interventions, they cannot be used as sources of information about the extent of comorbidity, its distribution in different groups, or attributable risk. The reasons are that:

1. People with two illnesses are more likely to seek treatment than people with either one of those illnesses separately (Berkson, 1946). This means that a clinic-based sample is likely to have a higher proportion of comorbid people in it than does the general population, and so one cannot use them to estimate the size of the problem.

2. Some combinations of disorders may bring people into treatment settings more often than others. For example, youth with substance abuse disorder and conduct disorder might be referred to clinics in higher proportions than youth with substance abuse disorder and an anxiety disorder. This would give clinicians the impression that comorbidity with conduct disorder is very common and comorbidity with anxiety very rare. This may or may not be true, but the point cannot be established from clinic samples.

3. Clinic cases may or may not be more severely affected with either of the comorbid conditions than community cases. Alternatively, comorbid cases may seek clinical treatment even though they have less severe symptoms of one or other disease than most community cases.

4. The temporal ordering of comorbidity may or may not be the same in clinical as in community cases.

5. Comorbidity seen in clinic cases may or may not be precipitated by the same risk factors as those that precipitate cases in the general population. These potential differences can be checked empirically for diseases where almost all cases get into treatment. However, in the case of drug abuse there is very strong evidence that this is not so. We also know that many people with psychiatric disorders never receive treatment. For these reasons we cannot assume that the patterns of association and risk seen in clinic samples mirror those seen in the population.

6. What appear to be risk factors for one or other disorder may in fact be predictors of treatment referral. For example, it might appear from a clinic study that poverty was associated with one or more disorders, where in fact stiff managed care regulations restricted access to treatment for children from privately insured families, while Medicaid regulations were more generous. This could mean that only children on Medicaid had much chance of getting treatment, and so poor children were overrepresented in the clinic sample.

For all these reasons, the scientific study of prevalence, incidence, development, and risk for comorbidity cannot rely solely on clinic-based samples.

Among community-based samples, one would of course like to use only those that employed appropriate (random or stratified random) sampling of representative population samples. These are hard to find. Many studies of adolescent drug use and abuse, for example, have relied on school samples, which tend to miss children who are absent or have dropped out — often a high-risk group. However, we have included studies using school-based samples to increase the number suitable for analysis.

Back to Top

2. Formal diagnostic rules and procedures

This meta-analysis includes studies that use structured interviewer-based or respondent-based interviews and formal scoring algorithms or best-estimate diagnostic procedures to generate psychiatric diagnoses using one of the recent taxonomies: ICD-9 or ICD-10, DSM-III, DSM-III-R, or DSM-IV. Three exceptions to this criterion that were included because authors made an attempt to generate diagnostic-like symptom clusters from survey questionnaires were the Ontario study (Offord et al., 1987), the National Household Survey of Drugs and Alcohol (SAMHSA, 1993), and the Middle Adolescent Vulnerability Study (Windle & Davies, 1999). Five diagnostic groupings were available for analysis of comorbidity in a large enough number of studies: conduct disorder (CD), oppositional defiant disorder (ODD), attention deficit hyperactivity disorder (ADHD), depressive disorders, and anxiety disorders.

When it came to drug abuse and dependence the requirement of a formal diagnosis was somewhat relaxed. The DSM-IV rules for dependence rely heavily on physical or work-related incapacities that most adolescents have not had time or been in a position to experience. In this meta-analysis we have taken researchers' definitions of abuse and dependence at face value, even if they do not demonstrate that they are using strict adult diagnostic criteria.

We have where possible distinguished between any use and abuse/dependence. However, it is important to note that the studies that report "any use" often do not report abuse/dependence separately; thus, the rates for any use usually include those for the nested category of abuse/dependence. Too few studies were available with the necessary information for us to model psychiatric comorbidity with specific drugs.

Comorbidity is used differently in different studies (Angold, Costello, & Erkanli, 1999), to refer to everything from concurrent co-occurrence of two disorders at the time of interview to their both having been diagnosed at some time during the lifetime of the individual. Because there were so few papers that met even basic criteria for inclusion in the analysis, we chose to include papers across the entire spectrum, from lifetime to concurrent comorbidity.

Back to Top

3. Risk factors and correlates

We had hoped to include only studies that could advance our understanding of risk factors for comorbidity, but there were not enough studies to make this possible. The tables and figures present the data for both sexes combined and separately, but the available data did not permit analysis at any more fine-grained level. There are not enough studies for separate analyses by race/ethnicity, or even by age.

Analytic method

In combining information from multiple studies we used random-effects hierarchical Bayesian regression models to account for study-to-study variability. Specifically, we assumed that the estimated log-odds ratios obtained from available studies were distributed normally and independently with each study having its own unknown mean, and a variance equal to the (estimated) variance of the log-odds ratios in each sample. This constituted the first level of the hierarchy. At the second level, the unknown means of study effects (i.e., the "true" log-odds ratios) were assumed to be independently and identically distributed with a normal distribution having a common (unknown) population mean and (unknown) variance. At the third and final level these common population parameters were given independent noninformative priors. Posterior computations were then performed using Gibbs sampling.

Among the available studies, that is, the studies that provided estimates of odds ratios and their standard errors, two only published the estimated odds ratios. So before applying the hierarchical model described above we imputed the unknown standard errors by imposing a statistical model on them: they are independently distributed as Gamma with unknown scale and shape parameters a and ß, respectively. Once again assuming a Bayesian model, we treated these parameters as random having independent uniform prior distributions between 0 and 100. These values were uninformative so the prior distributions had no influence on the imputed values of unknown standard errors.

Assuming that any missing data were missing at random, these unknown standard errors were simulated from their posterior predictive distributions, and the arithmetic means of the simulated values were treated as imputed values, which in turn replaced the missing values in the data set needed to run the hierarchical model as discussed above.

All the necessary computations for imputing the unknown standard errors and estimating the combined (population) estimates of the odds ratios for each substance use group, each male-female group, and the overall population estimates were performed using the latest version of the Bayesian software WinBUGS1.3, which implements a Markov chain Monte Carlo approach based on the Gibbs sampler. Documentation and more information on the technical background to Gibbs sampling and other related issues are available at http://www.mrc-bsu.cam.ac.uk/bugs. During these computations, we used 1,000 simulations to initialize the posterior distributions (burn-in time), and 10,000 iterations for convergence. More details including the WinBUGS codes used for these analyses can be obtained from Dr. Erkanli at al@psych.mc.duke.edu.

Back to Top

Conditional comorbidity estimates

The analyses described above were applied to each comorbidity pair, e,g., substance use/depression, independent of other comorbidity pairs. In other words, the estimates of odds-ratio for one pair was not adjusted for the other pairs. We know that in reality the estimated odds ratios obtained from each study were correlated across comorbidity pairs (Angold et al., 1999). However, most published papers did not provide cross-tabulations between substance use and all the other diagnoses considered simultaneously, so there was no way of knowing the true correlation between any comorbidity pair and another. In the absence of this information, the best we could do was to use a statistical model to try to recover these correlations, adjusting the population odds ratios for other pairs of comorbid diagnoses. We used an extension of the model we described above. At the first level, estimated log-odds ratios from each data set and for each comorbidity pair were assumed to be statistically independent conditional on unknown mean and variance equal to the (estimated) variance of the log-odds ratios in each study and comorbidity pair. At the second level, the unknown log-odds ratios were assumed to be independently distributed with a normal distribution having a common (unknown) population mean and (unknown) variance, for each comorbidity pair. At the third and final level these common population parameters were given dependent priors, using a slightly modified version of the conditional auto-regressive (CAR) model of Besag (1974). These priors, in effect, induced correlations between the population estimates of the log-odds ratios for each comorbidity pair. Computations were again performed using Gibbs sampling in WinBUGS1.3.

Results

Articles included in the meta-analysis

Twenty-one of the 141 articles evaluated met the basic criteria for inclusion in the meta-analysis. Articles were excluded if they did not meet the criteria for population-based, community studies (n=34), were reviews or conceptual papers (n=28), provided insufficient comorbidity data (n=24), included adults ages 20 or older in the sample in such a way that they could not be excluded for analytic purposes (n=19), did not provide lifetime or current prevalence rates (n=10), or provided data only on comorbid use, abuse, or dependence regarding one specific substance (n=5). Appendix B provides a summary of the studies used in the analyses. Appendix C summarizes the results from each individual study in both tabular and descriptive form. Table 1 lists the eight studies that provided only information on comorbidity with any substance use, including abuse or dependence. Table 2 lists studies that provided information on abuse or dependence.

The tables of comorbidity findings presented in this paper show that (1) not many studies have published the data needed to assess comorbidity with substance abuse, and (2) the estimates generated from these studies show wide variability. It is easy to see why this happens. Apart from all the issues of different diagnostic methods and scoring algorithms, estimates of comorbidity are likely to be small and therefore unstable when derived from small samples. As an example, if we were to assume that the population base rate of drug abuse in adolescents was 2 percent and that of major depressive disorder 2 percent (Costello et al., 1996), then the two would co-occur by chance four times in 1,000 observations. If they co-occured twice as often as expected by chance there would still only be eight cases in 1,000 observations. Given the uncertainties of measurement, a single study with 1,000 subjects could easily overestimate or underestimate the "real" association. Combining estimates from different studies using the methods described above enabled us to provide much more stable estimates than most individual studies can offer.

Back to Top

Results of meta-analysis

Figure 1 provides a general picture, showing the extent of comorbidity with the five diagnostic groupings, for any use, abuse, or dependence, for both boys and girls together. Results are presented in the form of odds ratios with their standard errors. An odds ratio of one indicates that the comorbid condition Y is equally likely to occur in persons with the index condition X and in those without X. An odds ratio of 2 indicates that Y is twice as likely in those with X as it is in those without X. The rule of thumb is that an odds ratio whose 95-percent confidence interval (CI) does not include 1 is "statistically significant." In general, an odds ratio of two or more is statistically significant. There are no standard procedures for statistically comparing two odds ratios derived from the modeling procedures described above. However, it is reasonable to assume that in the figures in this report two odds ratios are "significantly" different if one odds ratio is outside the 95-percent CI of the other (i.e., in the figures the CI bar on the odds ratio bar for one type of comorbidity does not overlap the top of the odds ratio bar for the other type of comorbidity).

Figure 1 shows that the extent of comorbidity varied widely from one type of psychiatric disorder to another. The extent of comorbidity with anxiety disorders was low (OR 1.9, 95% CI 1.4-2.2), while that for disruptive behavior disorders (DBDs: conduct disorder, oppositional defiant disorder, attention deficit hyperactivity disorder) was high (ORs 5.6-6.9), and not significantly different from one DBD to the other. The odds ratio for comorbidity with depression (OR 4.2, 95% CI 2.9-6.1) fell between those for anxiety (OR 1.9 95% CI 1.4-2.2) and ADHD (5.6, 95% CI 3.2-8.6). It was significantly higher than that for anxiety, and lower than those for CD and ODD, but not significantly lower than that for ADHD.

Comorbidity with any substance use and with substance abuse/dependence

Some of the studies available for analysis presented rates of any substance use, a category that included abuse and dependence. Others provided data for abuse/dependence only, or for the two categories separately. We wanted to use the data to test the hypothesis that comorbidity was more likely in cases of abuse/dependence than in cases of use. Our ability to do so was limited by the fact that many studies reporting any use include in that category abuse/dependence. From the data sets available it was not possible generate odds ratios for use alone. We divided the data sets into 13 that provided information on comorbidity with substance abuse and/or dependence, and 8 that provided information on any use, abuse, or dependence.

Figure 2 shows the odds ratios associated with abuse/dependence (grey bars) and any use (white bars). Comparing figure 2 with figure 1, we see that the rank ordering of odds ratios for comorbidity with the various diagnostic categories remained the same for both substance abuse/dependence and any use, although the smaller number of studies in each group led to wider confidence intervals. In the studies of abuse/dependence, the odds ratios for comorbidity with the DBDs were not significantly different from one another, but were significantly higher than the odds ratios for comorbidity with depression or anxiety; the latter two were not significantly different. In the case of any use, the odds ratios for comorbidity with CD and ODD were not significantly different, but they were higher than that for ADHD, which was higher than that for depression, while depression was in turn higher than anxiety. The odds ratio for comorbidity between any use and anxiety was not itself much above the level that might be seen by chance alone (OR 1.6, 95% CI 1.3-2.1).

For every diagnosis except CD, the odds of a comorbid psychiatric condition was significantly higher for adolescents with abuse/dependence than for those with any use. The difference was most extreme for ADHD, where abuse/dependence involved a 2.5-fold increase of risk compared with any use, followed by anxiety and depression (twofold increase of risk). The increase was lowest for conduct disorder, where the increase in risk was only around 20 percent, and ODD, where it was 45 percent.

In summary, the meta-analysis showed a stable pattern of high comorbidity with DBDs, low comorbidity with anxiety disorders, and intermediate comorbidity with depression. It is worth noting that even at the lowest level, substance use/abuse/dependence was associated with a doubling of the likelihood of an anxiety disorder, while the risk of a DBD increased six- to eightfold. However, these analyses do not permit us to draw any conclusions about the temporal ordering or causal relationship between one disorder and the other, only that the association between them greatly exceeds what would be expected by chance alone.

Effects of comorbidity among other diagnoses

If the comorbid conditions were themselves comorbid, this could affect the estimates shown in figures 1 and 2.

Figure 3 presents the results after attempting to model other forms of comorbidity in the data. It is immediately clear that (1) the rates of comorbidity maintained the same ordering as before, with higher comorbidity with DBDs and lower comorbidity with emotional disorders, but (2) the odds ratios were very much lower than they were for the simple two-way comorbidity. Only two were greater than 2: comorbidity with CD and ODD. In the case of depression and anxiety comorbidity was around 1; that is, substance abuse was no more likely in the presence of the other diagnosis than in its absence. The 95-percent confidence intervals were very wide, reflecting the extreme conservatism of the model and the small number of data sets available. In no case did the confidence interval exclude 1, an indication that the odds ratios were not significantly different from 1.

Given our limited ability to model the conditional comorbidities in the available data, the rest of these analyses are based on the two-way analyses, uncorrected for comorbidity among other diagnoses. It is, however, important to bear in mind that the relationship between drug abuse and any of the psychiatric disorders is likely to be influenced by comorbidity among the latter.

Back to Top

Correlates of and risk factors for comorbidity

We might get closer to understanding the causes of comorbidity if we could establish clear associations with putative risk factors. The published literature provides almost no data to make a meta-analytic review possible at this level. Sex was the only variable on which information was available in all studies.

Comorbidity in boys and girls

Figures 4 and 5 show the odds ratios associated with abuse/dependence (figure 4) and any use (figure 5) by sex. Figure 4 shows that for abuse/dependence the odds ratios were higher for girls than for boys, for every diagnosis, although in the case of depression the difference was very slight. The higher odds ratios for girls' comorbidity were statistically significant for CD and anxiety.

Figure 5 shows a similar pattern for any use. Girls who used alcohol or drugs were at greater risk of comorbidity than boys, for every type of disorder except depression; the difference was statistically significant for CD and anxiety. Thus the pattern of sex differences in comorbidity was similar for both types of substance use.

Figures 6 and 7 look at the question another way, comparing the odds ratios associated with any use versus abuse/dependence within sex. These figures show that although, as seen in figures 4 and 5, the odds ratios were lower for boys than for girls, the increase in risk associated with abuse/dependence relative to any use was similar across the sexes. Thus, the increase in risk associated with abuse/dependence compared with any use was highly significant for every diagnostic group except for anxiety, where comorbidity was actually slightly higher (though not significantly so) with any use than with abuse. A comparison of the size of the difference in odds ratios between users and abusers shows that for both boys and girls it was highest for ADHD (a 2.5-fold increase in risk for with abuse/dependence compared with any use) and depression (2.4-fold increase in risk). CD showed the smallest difference in risk between users and abusers: a 40-percent excess risk for girls, and a 60-percent excess risk for boys.

In summary, patterns of comorbidity were similar for boys and girls, but more extreme in girls. Substance-using girls were at higher risk than boys for every type of comorbidity except for that with depression. Although DBDs are less common in girls than in boys, the higher odds ratios for girls imply that when they do occur, they are more likely to be associated with other harmful behaviors, such as drug abuse. In both sexes, for every diagnosis except anxiety comorbidity was higher in association with abuse/dependence than with any use, an effect that was strongest for ADHD and depression and weakest for conduct disorder.

Other risk factors

Although individual studies discuss other risk factors (see appendix C), too few studies provide data on any one factor for meta-analysis to be possible. This is not to say that the data were not collected, only that they have not yet been published in a form that makes aggregation feasible. Thus we could draw no general conclusions from the published literature on the association between different diagnostic comorbidities and, for instance, age, race/ethnicity, poverty, family history, early development, or (most importantly for this meeting) the effects of treatment. Nor could meta-analysis help with teasing out the temporal ordering of different disorders because of a lack of published data to address this question.

Back to Top

Discussion

This analysis of published studies makes it absolutely clear that adolescents who use or abuse substances are significantly more likely than other youth to have one or more of the five major groups of psychiatric disorder. The risk ranges from twofold for anxiety disorders to sevenfold for CD. The risk of comorbidity is higher for girls than for boys for everything except depression. Controlling for comorbidity among psychiatric diagnosis the pattern is similar, although the odds ratios are lower. However, with so few studies the confidence intervals are very wide.

It is important to emphasize that these findings are based on cross-sectional (correlational) analyses. Few published studies provide the data needed to examine temporal relationships, and no attempt has been made to do so here. Without this dimension, we cannot talk about cause or prediction; even the word risk carries etiologic implications that the data cannot support. These analyses merely confirm that when substance use or abuse is present, other psychiatric diagnoses are present to a degree not expected by chance alone.

The data contain two hints of causal explanations, although these are very weak. First, the higher comorbidities associated with abuse/dependence than with any use, for every psychiatric diagnosis except anxiety, suggest a dose-response relationship, which is one step on the path to establishing causality (Robins & Guze, 1970). Second, the fact that girls, who are less likely to have DBDs (Loeber & Keenan, 1994), are more likely to use or abuse substances if they do have a DBD raises the possibility of a causal arrow from DBDs to substance use/abuse rather than vice versa. However, we have also to account for the fact that girls are both more likely to have anxiety disorders (Costello & Angold, 1995) and more likely to show comorbidity.

It is important to bear in mind the limitations of these analyses. Apart from those already discussed, there are many reasons why comorbidity itself could be the result of methodologic artifacts rather than a real phenomenon (Angold et al., 1999). Also, a meta-analysis can only be as good as the data that go into it, and even stretching our criteria to the limits there were many problems with the data; in particular, the construction of DSM diagnoses from instruments not designed for that purpose. We could stratify by sex but by no other variable. The number of psychiatric diagnoses was limited, and these had to be grouped into what may be misleading categories. For example, analysis of the Great Smoky Mountains data set showed no association between substance use or abuse/dependence and anxiety disorders as a group, controlling for other psychiatric comorbidities (Costello, Erkanli, Federman, & Angold, 1999). However, further analysis showed that separation anxiety reduced the likelihood of alcohol use, and increased the age at first use, whereas generalized anxiety disorder increased the risk of alcohol use and lowered the age at first use (Kaplow, Curran, & Costello, 2000). This suggests the importance of being able to look at the association of substance abuse with different psychiatric disorders separately. One would also want to examine the association of different psychiatric disorders with specific substances (alcohol, cannabis, etc.). Further, investigation of the association between substance abuse, psychiatric comorbidity, and mental health service use is desirable but not possible given the limits of the studies included in the meta-analysis. Finally, it would be fascinating to be able to use the power of Markov chain Monte Carlo modeling to examine the temporal relations among disorders.

None of these is possible using published papers. Comorbidity with psychiatric disorder has seldom been the focus of publications from the data sets we used for the meta-analysis. But for longitudinal psychiatric epidemiologic surveys a great deal of information has to be collected that is usually not published in scientific papers. For example, a study has to collect detailed information on a wide range of drugs that adolescents might use, even to publish a simple analysis of "drug abuse." The detailed information that lies behind the single published figure may never appear in print. The literature on child psychopathology contains reports from several large studies of adolescent substance abuse with representative samples, careful, consistent methods of data collection and data aggregation, and at least some information about co-occurring psychiatric disorders. Several are panel studies with repeated assessments of the same participants over years or even decades. This raised for us the question of whether these data sets could potentially yield more than they have so far on psychiatric comorbidity with drug abuse. Recent advances in data transfer, computing speed, and statistical methods raise the possibility of exploiting this wealth of information. In the next section we report on a preliminary exploration of what is "out there" to address the question of psychiatric comorbidity as a risk for later drug abuse.

Back to Top