Clusters of Marijuana Use in the United States

Transcription

1 American Journal of Epidemiology Copyright 1998 by The Johns Hopkins University School of Hygiene and Public Health All rights reserved Vol. 148, 12 Printed in U.S.A. Clusters of Marijuana Use in the United States G. V. Bobashev and J. C. Anthony In this study, the authors test for and estimate the clustering of marijuana use within United States neighborhoods, making use of data from annual nationally representative household sample surveys conducted during the period A recently developed statistical method, alternating logistic regression, was used to quantify the clustering of marijuana users in neighborhoods. The resulting estimates of pairwise odds s ranged from 1.3 ( ) for the lifetime history of marijuana use to 2.0 ( ) for recent sharing of marijuana from one person to another. Exploratory analysis showed a slight decrease of clustering effects after adjustment for individual-level covariates: age, sex, race, education, annual family income, and history of tobacco use. Nevertheless, the main factors that account for clustering remain to be determined. Alternating logistic regression provided useful estimates of marijuana use clustering and can be used to estimate clustering of the other drug-related behavior, including sharing of needle injection equipment and other human immunodeficiency virus risk behaviors. As a form of multilevel analysis, the alternating logistic regression can accommodate shared, community-level characteristics that might influence drug taking (e.g., collective efficacy), as well as individual-level covariates, such as age and sex. Am J Epidemiol 1998;148: cannabis; logistic models; marijuana abuse; social environment; space-time clustering; spatial behavior; substance use disorders Less than 5 years ago, Katz et al. (1, 2) introduced a new statistical method to study the clustering of diarrheal disease within villages and households. This method, termed "alternating logistic regression" (ALR), yields a readily interpretable statistical index of disease clustering in the form of a "pairwise odds " (PWOR) or "pairwise cross-product." In nontechnical terms, the PWOR reflects how strongly a disease occurs in clusters. In more technical terms, the PWOR reflects odds of the disease for an individual in a cluster given that another randomly chosen individual from that cluster has a disease relative to the odds if that randomly chosen individual does not have a disease. (Details on the definition and interpretation of PWOR are presented in appendix 1.) At the village level, the clustering of diarrheal disease has been seen in PWOR estimates ranging from a Received for publication December 11, 1997, and accepted for publication May 5, Abbreviations: ALR, alternating logistic regression; Cl, ; NHSDA, National Household Survey on Drug Abuse; PWOR, pairwise odds. From The Johns Hopkins University, School of Hygiene and Public Health, Department of Mental Hygiene, Baltimore, MD. Reprint requests to Dr. James C. Anthony, Johns Hopkins School of Hygiene and Public Health, Department of Mental Hygiene, 624 North Broadway, Baltimore, MD lower value of 1.03 up to 2.1. Within households, the clustering of diarrheal disease has been manifest in PWOR estimates ranging from 1.9 to 10.1 (1). In this investigation, we use ALR methods to study the clustering of marijuana use in US neighborhoods, analyzing national survey data gathered during the period 1990 through Consistent with a previously demonstrated person-to-person spread of drug-taking behavior (3, 4), we expected to see some evidence that marijuana use occurs in clusters. In addition, we test whether the magnitude of clustering might be influenced by sociodemographic or other characteristics of the population. To provide a frame of reference, we use the village-level PWOR estimates for diarrheal disease (1) and clustering of total family income in neighborhoods. These estimates serve as benchmark values against which we can compare the magnitude of clustering of marijuana use within neighborhoods. This is not the first time that concepts and methods of infectious disease epidemiology have been applied in the epidemiologic study of drug-taking behavior (3, 5,6). However, it is the first time that multilevel analysis in the form of ALR has been used for research on the epidemiology of drug involvement. For this reason, we have restricted this first investigation on marijuana to a small set of characteristics, such as sex, age, and educational attainment of the respondent; personal his- 1168

2 Clustering of Marijuana 1169 tory of tobacco smoking; and total family income. If we are successful in this more-restricted context, we will seek to extend our investigation to more-challenging causal hypotheses, with attention to neighborhoodlevel census indicators of social advantage (e.g., "percent of owner-occupied housing") and individual-level survey characteristics (e.g., history of childhood delinquent behavior and criminal activity), which also might help to account for clustering and the occurrence of marijuana use. MATERIALS AND METHODS This investigation uses publicly available data from the National Household Surveys on Drug Abuse (NHSDA), conducted from 1990 through The NHSDA sample is designed to be representative of civilian, noninstitutionalized US residents age 12 years and older. The numbers of participants in these nationwide epidemiologic field surveys are as follows: n = 9,259 in 1990; n = 32,594 in 1991; n = 28,832 in 1992; n = 26,489 in 1993; n = 17,809 in 1994; and n = 17,747 in Given sample sizes of this magnitude, the period of data gathering lasts for 6 months or more each year. The survey response rate during these years is at the 80 percent level (±3 percent), as described elsewhere (7-12). Descriptions of the NHSDA research protocols appear both in detailed monographs and in journal articles (7-13). In brief, during each cycle of the NHSDA, the research team draws a probability sample of noninstitutionalized residents of the United States aged 12 years or more. The sampling plan used to identify these participants involves multistage area probability sampling down to the level of block groups, census tracts, or enumen districts, followed by sampling of households and then individual respondents within households. It is this multistage sampling plan that allows us to study clustering. Here, block groups, census tracts, and enumen districts represent discrete, geographically contiguous, and relatively homogeneous aggregations of human habitations, which, for ease of presentation, we have termed "neighborhood" as in past scientific articles (14-16). Once a designated respondent is identified and recruited according to IRB-approved human subjects consent procedures, the assessment of drug involvement involves one of two methods. Most often, respondents choose to mark their own replies to standardized interview questions on an answer sheet that the interviewer cannot see. Less often, the interviewer reads the questions aloud, listens to the respondent's answer, and then records it. The same procedures are used to gather information about sociodemographic characteristics of the participants, such as sex and race. Measures In this investigation, we focus attention upon the participants' reports about their marijuana use during the month just prior to the assessment. We give secondary attention to marijuana use in the year prior to assessment and to a lifetime history of marijuana use (without respect to the timing of most recent use). Our expectation is that the magnitude of clustering will depend upon recency of use, with greater clustering of most recent use. We expect lower clustering for lifetime history of marijuana use because residential mobility might separate a respondent from earlier environments of marijuana use (e.g., college). For comparative purposes, we also investigate the magnitude of clustering of having shared marijuana with other persons as well as the magnitude of clustering of the economic status of a household. In 1992 and 1993 (but not in the other years), the NHSDA included questions about the sharing of marijuana and asked whether the participant had been involved in sharing marijuana with others. The NHSDA assessments of the economic status of a household are based upon answers to standardized items about gross family income. In advance of the analysis, we chose to dichotomize gross family income at a value of $24,000 per year. This threshold value for gross family income was not set in relation to any a priori specification of poverty or income adequacy. In fact, specifications for poverty levels vary from state to state. Nonetheless, given social stratification within the United States, we anticipated some clustering of family income in relation to this dichotomy. Other individual-level covariates under study include self-reported sex (male vs. female), age in years, educational attainment in years, race-ethnicity, and history of tobacco use during the same period prior to assessment as marijuana use. Tobacco involvement is included because we hypothesize that some degree of clustering of marijuana use might depend upon the participants' tobacco smoking histories. Statistical analysis Following procedures outlined by Katz et al. (1, 2) for the study of diarrheal diseases, we estimate the PWOR as an index of clustering of marijuana smoking within neighborhoods surveyed as part of the NHSDA protocol. In this context, the PWOR is a specific parameter in the equation for conditional expectation of marijuana use for an individual respondent, with conditioning on the occurrence of marijuana use in another individual respondent chosen within the same cluster (neighborhood). Within the ALR framework developed by Katz et al. (1,2), the logarithm of the PWOR can be expressed as

3 1170 Bobashev and Anthony a function of an indicator variable coded to show whether subjects j and k in a pair belong to the same or different neighborhoods: log(pwor jk ) = az jk, where Z jk, takes values 1 or 0, depending on whether the pair (j,k) belongs to the same neighborhood. Hence, in this context, the PWOR estimate reflects whether there is an increase in probability that marijuana use in a pair of respondents within a neighborhood cluster is concordant versus discordant. Estimation The statistical method for ALR simultaneously accounts for the suspected influence of clustering and the suspected influence of covariates. This is accomplished via iterative recalculation of PWORs and logistic regression on the outcomes. The logistic part for the jth respondent is modeled as log(oddsj) = where X, denotes a covariate, and p, is the corresponding odds. Summation is taken over all the covariates. Here, the ALR method allows one to take into account neighborhood-level associations involving marijuana use during estimation of the hypothesized associations of marijuana use with covariates in the equation, while providing more-efficient estimates for variances. Available ALR software (17), implemented as S-Plus routines (MathSoft, Inc., Seattle, Washington), provides estimates for the PWOR standard errors as well as for corresponding standard errors for the regression coefficients (betas) associated with each covariate. As point estimates tend to be asymptotically normal, the information about standard errors allows derivation of corresponding 95 percent s. Two features of the NHSDA survey design require an adaptation in this analysis. First, in a relatively small number of households, the survey designated not one, but two, respondents per household. This happened too infrequently for us to estimate household-level clustering of drug use. Instead, we accommodate this feature of sampling by making a random draw of one, and only one, respondent per household. As a result, the withinhousehold probability of selection always depends upon the number of potentially eligible participants per household. As a sensitivity analysis, we repeated this random-draw process and estimation procedure 20 times. The resulting regression coefficients showed little variation (i.e., within ±0.08 of the original estimate; data available upon request to the corresponding author). Second, to achieve an adequate number of minority participants, the NHSDA is designed with what amounts to an "oversampling" of US residents who belong to minority groups. This oversampling is in addition to the variation in the standard sample selection probabilities that are due to varying numbers of potentially eligible respondents within each household. For this investigation, this variation in sample selection probabilities could not be modeled explicitly within the framework of the ALR methodology. Carey (V. Carey, personal communication, 1997) has adapted the ALR methods to allow for cluster-level variation in sample selection probabilities, but not for individual-level variation. As a result, to hold constant variation in sample selection probabilities, we have studied the PWOR estimates for sample strata within which the sample selection probabilities are roughly constant. RESULTS The NHSDA study samples are large enough to reduce serious concerns about statistical precision and power with respect to the estimates reported in this paper. Description of the study samples each year from 1990 through 1995 shows that thousands of participants reported use of marijuana, and hundreds reported sharing of marijuana (table 1). The anticipated pattern in clustering of marijuana use can be seen in figure 1 for each of the survey years under study. The estimated PWORs tend to be smaller for lifetime history of marijuana use, slightly greater for use of marijuana in the year prior to interview, and somewhat larger for use of marijuana in the month prior to interview. The PWOR estimates for marijuana use in the month prior to interview range from a value of about 1.4 in survey years 1990, 1992, 1994, and 1995 to values of in 1991 and For each of these years, the estimated 95 percent s for the PWOR are consistent with statistical significance of the observed clustering (i.e., they do not trap the PWOR null value of 1.0). As anticipated, a slightly greater estimated magnitude of clustering is seen when we shift focus from reported marijuana use to reported sharing of marijuana. For sharing, the PWOR point estimates are larger than the values observed for use of marijuana in the previous month (figure 1). The estimated PWOR magnitude for clustering of annual family income (less than $24,000 vs. $24,000 or more) is between 1.5 and 2.0 for (data not shown). Hence, family income, which also is expected to cluster within neighborhoods, shows a magnitude of clustering not too distant from that seen for last month

4 Clustering of Marijuana 1171 TABLE 1. Characteristics of the samples used for the analysis of lifetime, last year, and last month use and sharing of marijuana, NHSDA* data, Sample size 8,828 29,060 21,577 23,672 16,003 15,855 Age (years) <12t ^0 >40 1,966 1,973 2,886 2, ,649 7,113 9,444 5, * ,359 5,797 7,336 3, ,027 4,962 9,406 3, ,167 3,325 6,098 2, ,847 3,617 5,944 2, Sex Male Female 3,986 4, ,821 16, ,724 11, ,740 12, ,149 8, ,795 9, Race/ethnicity Black, not Hispanic White, not Hispanic Hispanic origin Other 1,735 5,168 1, ,910 14,664 6, ,799 10,668 5, ,322 11,430 6, ,521 7,938 4, ,570 8,033 3, Recency of marijuana use Ever in life Last year Last month Shared last month (available only for data) 3,166 1, ,915 3,673 1, ,583 2,679 1, * NHSDA, National Household Surveys on Drug Abuse. Data from nationally representative sample surveys on noninstitutionalized residents of the United States, age 12 years and older, in each of the listed years, t Not included in NHSDA sample, i Because of rounding, some of the percentages do not sum to ,899 3,339 1, ,608 2,075 1, ,385 2,033 1, i X O lifetime use last year use last month use last month share s Year FIGURE 1. Estimates of pairwise odds s with s for clustering of marijuana use by year and time. Data from NHSDA, For the years , clustering of the last month sharing the drug is also presented. All estimates are adjusted for age, race, sex, and education. The NHSDA gathered data on sharing only in

5 1172 Bobashev and Anthony use and sharing of marijuana. For the years , family income PWORs are 1.6 (95 percent (CI) ) and 1.8 (95 percent CI ), respectively, and for recent sharing of marijuana in these years, the PWOR estimates are 2.0 (95 percent CI ) and 1.9 (95 percent CI ). When ALR methods are used to estimate the pairwise odds and then to make statistical adjustments for individual-level covariates, we find some modest attenuation in the magnitude of clustering for recent marijuana use. As shown in figure 2, with inclusion of terms for age, sex, race, and educational attainment, the clustering estimate drops on average by 0.1. When terms for family income and recent tobacco use are added, it drops another 0.1. Regression coefficients and their s are presented in table 2. On the basis of data from , there is some variation for the magnitude of clustering in relation to quartiles of individual sampling selection probabilities, especially in the 1990 and 1995 surveys, where clustering completely disappears for a strata corresponding to the third quartile of sample weight distribution. For the years , there is no statistically significant variation over the quartiles of sample weights (data not shown; figure available upon request to corresponding author). DISCUSSION To our knowledge, this study presents the first epidemiologic evidence of neighborhood-level clustering of marijuana use and sharing of marijuana, with data from 6 years of a nationwide survey. The strength of this evidence comes, in part, from the nationally representative character of the survey and from derivation of clustering estimates for 6 separate years. We see evidence of neighborhood clustering for marijuana involvement in all of the years under study, with the weakest clustering observed for lifetime history of marijuana use and the strongest for sharing of marijuana by one person with another. We observe little association between the magnitude of clustering and individuallevel characteristics, such as age, sex, race, educational attainment, history of tobacco use, and family income. Before we can give a more-detailed interpretation of these results, several limitations need to be discussed. First, these are results from a sample designed to be representative of the United States civilian, noninstitutionalized population during the 1990s. Whether these results pertain to other populations or other times is a matter for future research. We are especially concerned that the estimates of clustering might be understated due to neglect of drug users who are not affiliated with households (e.g., homeless persons). Indeed, homeless persons are often drug users and tend to congregate in specific metropolitan areas (15). Second, the survey response rate is respectable, but not superb. To the extent that drug taking is associated with survey nonresponse (18), the degree of clustering might be understated, especially if survey nonresponse is greater within some neighborhoods than in others. This issue cannot be examined in the frame of NHSDA public use data. Third, reliance upon self-report data about a sensitive and illegal topic, such as marijuana use, is always O A X unadjusted adjusted for age, race, sex, education adjusted for age, race, sex, education, tobacco use, family income o A Year FIGURE 2. Unadjusted and adjusted pairwise odds s with s for the clustering of last month of marijuana use. Data from NHSDA, Adjustment was done sequentially for the covariates age, race, sex, education, recent tobacco use, and family income.

6 Clustering of Marijuana 1173 TABLE 2. Intercept Sex* Estimated odds s for covariates used in ALR* model for the "last month" marijuana use, NHSDAf data O Race/ethnicity Black, non- Hispanic Hispanic Other Age Education Family income!] Tobacco use last * ALR, alternating logistic regression. t NHSDA, National Household Surveys on Drug Abuse. Data from nationally representative sample surveys of noninstitutionalized residents of the United States, age 12 years and older, in each of the listed years. $ Males were used as a reference category. White, non-hispanic were used as a reference category. H Family income was coded as an ordinal variable with unit steps corresponding to the breakpoints of $10,000, $20,000, $40,000, and $75,000. a concern in epidemiologic research of this type (5). Fourth, in this initial investigation of marijuana clusters, we have had to be selective in our choice of covariates under study. If successful in this line of research, we will be able to extend the statistical models, including neighborhood-level covariates (e.g., percent of owner-occupied housing) and individual-level covariates (e.g., level of delinquent behavior). Finally, the ALR methods do not yet allow for individual-level variation in sampling weights. In our paper, we are able to accommodate this limitation by studying estimates for different strata of the sampling weights, but a moresatisfactory solution remains to be developed as this biostatistical method is refined. Notwithstanding limitations such as these, we note that the magnitude of clustering of marijuana use and marijuana sharing are within the range previously reported for village clustering of diarrheal diseases within developing countries (1). This is noteworthy. In addition, the level of marijuana clustering is not too distant from clustering of annual family incomes, a social index of wealth and economic prosperity. The magnitude of marijuana clustering is not altered substantially when we use the statistical model to hold constant other suspected determinants of marijuana use, including the individual's history of tobacco use. Indeed, a history of tobacco use is found to have a strong association with the estimated probability of recent marijuana use. However, tobacco use does not seem to be a determinant of the magnitude of marijuana clustering. Hence, the determinants of marijuana clustering remain an open field for new research. We are continuing to apply the ALR methods to these national survey data in order to study other forms of drug involvement, including use of tobacco, alcohol, cocaine, heroin, and injection drug use. Nonetheless, the real challenge will be to discover the neighborhood- or individual-level characteristics that might account for clustering, as well as to identify public health techniques that will reduce both clustering and occurrence of hazardous illicit drug use. In this content, a recently published article in Science on neighborhood levels of "collective efficacy" and occurrence of delinquent behavior (15) has some importance. The concepts of social cohesion and collective efficacy within neighborhoods now have entered the domain of considen with respect to marijuana use as well. Whether neighborhood-level characteristics such as collective efficacy also will influence clustering and occurrence of illicit drug use remains to be seen. ACKNOWLEDGMENTS Supported by a research grant award from the National Institute on Drug Abuse DA09592 to Dr. James C. Anthony, principal investigator.

7 1174 Bobashev and Anthony The authors thank the Substance Abuse and Mental Health Services Administn (SAMHSA) for sharing these public use data and for providing answers to our questions about details in the survey design. They thank Scott Hubbard for research assistance and Dr. Vincent Carey for statistical consultation. REFERENCES 1. Katz J, Carey VJ, Zeger SL, et al. Estimation of design effects and diarrhea clustering within households and villages: Am J Epidemiol 1993;138: Katz J, Zeger SL, West KP Jr, et al. Clustering of xerophthalmia within households and villages. Int J Epidemiol 1993;22: De Alarcon R. The uses of clinical epidemiology. In: Slaby AE, ed. Studying drug abuse. New Brunswick, NJ: Rutgers University Press, 1985:25^5. 4. Wolfson M, Forster JL, Claxton AJ, et al. Adolescent smokers' provision of tobacco to other adolescents. Am J Public Health 1997;87: Anthony JC, Helzer JE. Epidemiology of drug dependence. In: Tsuang MT, Tohen M, Zahner GEP, eds. Textbook in psychiatric epidemiology. New York, NY: John Wiley & Sons, 1995:361^ Hawks DV, Ogborne AG, Mitcheson MC. The strategy of epidemiological research in drug dependence, Br J Addict 1970;65: United States. Department of Health and Human Services. National Institute of Drug Abuse (NIDA): National Household Survey on Drug Abuse, main findings: Rockville MD: Department of Health and Human Services, (DHHS publication no. 1788). 8. United States. Department of Health and Human Services. Substance Abuse and Mental Health Services Administn findings: Rockville MD: Department of Health and Human Services, (DHHS publication no. 1980). 9. United States. Department of Health and Human Services. Substance Abuse and Mental Health Services Administn findings: Rockville MD: Department of Health and Human Services, (DHHS publication no. 3012). 10. United States. Department of Health and Human Services. Substance Abuse and Mental Health Services Administn findings: Rockville MD: Department of Health and Human Services, (DHHS publication no. 3020). 11. United States. Department of Health and Human Services. Substance Abuse and Mental Health Services Administn findings: Rockville MD: Department of Health and Human Services, (DHHS publication no. 3085). 12. United States. Department of Health and Human Services. Substance Abuse and Mental Health Services Administn findings: Washington, DC: Department of Health and Human Services, (DHHS publication no. 3127). 13. United States. Substance Abuse and Mental Health Services Administn (SAMHSA). National household survey on drug abuse, population estimates: Washington, DC: US GPO, (DHHS publication no. 3095). 14. Crum RM, Lillie-Blanton M, Anthony JC. Neighborhood environment and the opportunity to use cocaine and other drugs in late childhood and early adolescence. Drug Alcohol Depend 1996;43: Sampson RJ, Raudenbush SW, Earls F. Neighborhoods and violent crime: a multilevel study of collective efficacy. Science 1997;277: Ennett ST, Flewelling RL, Lindrooth RC, et al. School and neighborhood characteristics associated with school rates of alcohol, cigarette and marijuana use, J Health Soc Behav 1997;38: Carey VJ, Zeger SL, Diggle P. Modelling multivariate binary data with alternating logistic regressions. Biometrika 1993;80: Cottier LB, Compton WM, Ben-Abdallah A, et al. Achieving a 96.6 percent follow-up rate in a longitudinal study of drug abusers. Drug Alcohol Depend 1996;41: APPENDIX 1 The definition of PWOR comes from the analysis of a 2 x 2 table for paired outcomes where the first and the second rows, respectively, correspond to positive and negative responses of the first subject in a pair, and the first and the second columns correspond to positive and negative responses of the second subject in the pair. Technically, denoting probabilities of a pair to be found in one of the quadrants (left-to-right, top-to-bottom) as Pu, P\o, Po\> Poo> PWOR is defined as a PuPoo/PwPoi- Because of its construction, PWOR also can be termed as pairwise cross-product. The latter term might be more preferable (V. Carey, Harvard University, personal communication, 1997). However, in this paper, we chose to adopt the term introduced by Katz et al. (1) in their first paper about this new method.