Annals of Epidemiology 25 (2015) 486e491
Contents lists available at ScienceDirect
Annals of Epidemiology journal homepage: www.annalsofepidemiology.org
Original article
Estimating the association between metabolic risk factors and marijuana use in U.S. adults using data from the continuous National Health and Nutrition Examination Survey Christin Ann Thompson BS a, b, Joel W. Hay PhD a, b, * a b
Department of Clinical Pharmacy, Pharmaceutical Economics, and Policy, University of Southern California, Los Angeles Leonard Schaeffer Center for Health Policy and Economics, University of Southern California, Los Angeles
a r t i c l e i n f o
a b s t r a c t
Article history: Received 22 August 2014 Accepted 20 January 2015 Available online 31 January 2015
Purpose: More research is needed on the health effects of marijuana use. Results of previous studies indicate that marijuana could alleviate certain factors of metabolic syndrome, such as obesity. Methods: Data on 6281 persons from National Health and Nutrition Examination Survey from 2005 to 2012 were used to estimate the effect of marijuana use on cardiometabolic risk factors. The reliability of ordinary least squares (OLS) regression models was tested by replacing marijuana use as the risk factor of interest with alcohol and carbohydrate consumption. Instrumental variable methods were used to account for the potential endogeneity of marijuana use. Results: OLS models show lower fasting insulin, insulin resistance, body mass index, and waist circumference in users compared with nonusers. However, when alcohol and carbohydrate intake substitute for marijuana use in OLS models, similar metabolic benefits are estimated. The Durbin-Wu-Hausman tests provide evidence of endogeneity of marijuana use in OLS models, but instrumental variables models do not yield significant estimates for marijuana use. Conclusion: These findings challenge the robustness of OLS estimates of a positive relationship between marijuana use and fasting insulin, insulin resistance, body mass index, and waist circumference. Ó 2015 Elsevier Inc. All rights reserved.
Keywords: Instrumental variables Marijuana use Metabolic health Multivariate linear regression
Introduction Marijuana use has become increasingly prevalent in the United States. In a 2010 survey from the Substance Abuse and Mental Health Services Administration, an estimated 7.3% of Americans aged 12 and older used marijuana in 2012, more than any other illicit substance. Of those who reported using marijuana, an estimated 18.9 million used in the past month and 7.6 million could be considered chronic users [1]. With the rapidly changing policy landscape surrounding the control of marijuana use and its application in healthcare, more research is needed on the short- and long-term health effects of marijuana. Evidence on the effect of marijuana on common disease processes, such as diabetes, would be useful in health care decision-making. Metabolic syndrome is a set of clinical criteria associated with increased risk of type II diabetes and cardiovascular disease [2]. Metabolic syndrome *Corresponding author. Leonard Schaeffer Center for Health Policy and Economics, University of Southern California, Verna and Peter Dauterive Hall, 635 Downey Way, Los Angeles, CA 90089-3333. Tel.: þ1-818-338-5433; fax: þ1-213-740-3460. E-mail address: [email protected] (J.W. Hay). http://dx.doi.org/10.1016/j.annepidem.2015.01.013 1047-2797/Ó 2015 Elsevier Inc. All rights reserved.
includes excess fat around the abdomen, low high-density lipoprotein cholesterol (HDL-C), high triglyceride levels, high blood pressure, high blood sugar level, and insulin resistance [3]. Previous examinations of the relationship between marijuana use and health outcomes have provided conflicting results. In a small controlled study of male research volunteers, periods of marijuana smoking increased daily caloric intake and body weight [4]. In another study that used data from National Health and Nutrition Examination Survey (NHANES) III, total caloric intake was higher in current users but body mass index (BMI) was lower in current users compared with nonusers [5]. One large-sample, retrospective analysis found an association between marijuana use and higher caloric and alcohol consumption but did not find an association between current use and BMI or cardiovascular risk factors [6]. In a case-control study matched for age, sex, ethnicity, and BMI, marijuana use was associated with higher abdominal visceral fat, lower HDL-C, and lower adipocyte insulin resistance; however, there were no differences in total body fat, hepatic steatosis, insulin insensitivity, measures of beta-cell function, or glucose intolerance [7]. A study that used two large U.S. data sets found no significant difference in the multivariate-adjusted odds of obesity in marijuana users compared with abstainers, with the exception of users
C.A. Thompson, J.W. Hay / Annals of Epidemiology 25 (2015) 486e491
who smoked marijuana more than three times a week [8]. Prevalence of overweight or obesity in young adults from the Mater-University of Queensland Study of Pregnancy and its Outcomes was significantly lower in marijuana users in multivariate-adjusted analyses [9]. In contrast, a study using the National Longitudinal Survey of Youth found that, compared to nonuse or low use in adolescence, consistent or increasing patterns of marijuana use in adolescence are associated with an increased risk of obesity [10]. In another study among youth in the United States, frequent marijuana use was associated with overweight status but not obesity in young girls [11]. In a recent study, Penner et al. [12] evaluated the association between self-reported marijuana use and components of metabolic syndrome using the National Health and Nutrition Survey from 2005 to 2010. Surprisingly, the results of ordinary least squares estimation (OLS) of the multivariate linear models suggested reduced fasting insulin, increased HDL-C levels, and a smaller waist circumference in “current users” of marijuana compared with “never users”. Results from the analysis of survey data that indicate improved BMI and other factors of metabolic health contradict what is known about the role of cannabis compounds in the cannabinoid system in humans. It is well known that marijuana contains appetitestimulating compounds known as cannabinoids, which attach to cannabinoid receptors in the brain and other parts of the body [13]. This physiological effect has motivated its use in the treatment of cachexia (wasting syndrome) in cancer and human immunodeficiency virus patients [14,15]. To make sense of the conflicting and sometimes surprising results from the observational studies reviewed in the previous sections, the analytic methods used in these studies should be carefully considered. The models used in prior observational research are based on OLS regression models that disregard the potential endogeneity of marijuana use in health outcome risk equations. When used as explanatory variables, endogenous variables lead to biased regression estimates because of the correlation with missing or unknown control variables [16]. For example, tobacco use is inversely related to obesity [17e19] and directly proportional to marijuana use [20]. If one were to exclude tobacco use inappropriately as a confounder in a model explaining the relationship between marijuana use and health outcomes, the estimated relationship would be potentially biased. Empirical studies have previously acknowledged the endogeneity of substance in the health economics literature [21]. Models that account for endogeneity should be considered, as well as different specifications and checks on model validity. This study aims to explore the relationship between marijuana use and the clinical factors of metabolic syndrome by critically evaluating OLS regression analysis of NHANES data from 2005 to 2012. Materials and methods Sample This study sample included participants from the continuous NHANES from 2005 to 2012, a cross-sectional survey which oversamples young children, older persons, and certain ethnic groups in two-year cycles and applies weights for a nationally representative sample. Participants underwent an in-home interview and laboratory tests, which included blood and urine samples. This investigation focused on the 6281 participants surveyed between 2005 and 2012 who responded to questionnaire items regarding the use of marijuana.
487
continuous NHANES data. Outcome variables included fasting insulin, fasting glucose, homeostasis model assessment of insulin resistance (HOMA-IR; [fasting serum insulin (mU/mL) fasting plasma glucose (mg/dL)/405)], HDL-C, triglycerides, blood pressure (average of three blood pressure readings), BMI (weight in kilograms/height in meters2), and waist circumference. All laboratory measurements were taken at a medical examination center, and the methods are reported in detail in the NHANES Laboratory Procedures Manual [22]. Insulin, HOMA-IR, and triglycerides were log transformed. Explanatory variables were chosen to control for demographic characteristics and risk factors for metabolic syndrome. The variable of interest, marijuana use, was defined using data on selfreported marijuana use from the drug use questionnaire component of the NHANES survey: never users (never smoked marijuana, n ¼ 2861), past users (smoked marijuana at least once but not in the past 30 days, n ¼ 2589), and current users (smoked marijuana at least once in the prior 30 days, n ¼ 831). Significant risk factors for metabolic syndrome include age, sex, race, BMI, tobacco use, physical activity, income, alcohol consumption, carbohydrate intake, and postmenopausal status [23]. Education level was also included as a control variable. As in the study by Penner et al [12], because a significant portion of the income data were missing (z7%), missing values were generated using multiple imputations. The imputations (m ¼ 10) included multivariate normal models to allow for variation in the observations generated by the model. The multivariate imputation model incorporated NHANES stratum effects to account for survey design [24]. Rubin’s [25] rules for combination were applied to adjust estimates and standard errors for variability between imputations. A squared term for age was also added to the full model to capture any nonlinear relationships between age and the health outcomes under examination. Multivariate analysis The appropriate weights were applied based on guidelines from the CDC’s National Center for Health Statistics to account for the complex survey design of NHANES data in regression coefficient and standard error estimates [26]. Separate multiple linear regression models were fit to each outcome variable and to the log-transformed insulin, HOMA-IR, and triglyceride variables. A measure of carbohydrate intake was also added to each model, which has been included in empirical models dealing with metabolic risk factors; however, nutritional data were not available for 2011 to 2012, so the final models did not include carbohydrate intake [23,27,28]. The final regression models adjusted for age, sex, race, education, income, BMI, tobacco use, alcohol use, and physical activity. BMI was left out of the models for BMI and waist circumference. To test for effect modification, multivariate models were stratified by age and sex. Estimating the effects of other risk factors on metabolic syndrome using the multivariate linear models provides an additional robustness check. Alcohol use categories (nondrinkers, 1 drink per week, >1 to 14 drinks per week, and >14 drinks per week) and carbohydrate intake (low, medium, high) were examined in place of marijuana use as the variables of interest in the linear models for each outcome variable. Instrumental variables analysis Instrumental variable (IV) methods can be used to explore the potential endogeneity of marijuana use. The conceptual model is as follows:
Measures
Yi ¼ AXi þ Bmi þ εi
The outcomes of interest were cardiometabolic risk factors. Penner et al. [12] have previously defined these outcomes using
where Yi is a vector of all outcomes, mi is the treatment effect variable for marijuana use, Xi denotes other relevant
488
C.A. Thompson, J.W. Hay / Annals of Epidemiology 25 (2015) 486e491
exogenous personal characteristics, and εi is an error term. The parameter on marijuana use (B) may be (asymptotically) biased if models do not account for endogeneity. Marijuana use can be defined in terms of both relevant exogenous variables from the first equation (Xi ) and variables that instrument for marijuana use (zi ):
mi ¼ gzi þ dXi þ ni : Appropriate exogenous instruments (zi ) will be correlated with a propensity to use marijuana but uncorrelated with the health outcomes of interest in each model. The error terms from the first and second equations, εi and ni , respectively, are assumed to be asymptotically uncorrelated. The method for estimation using this model is typically two-stage least squares (2SLS), which estimates fitted values for marijuana use (mi ) in the first stage using the exogenous instruments [16]. As in OLS estimation, the 2SLS estimation was adjusted based on CDC’s National Center for Health Statistics guidelines to account for complex survey design. IV models were stratified by age and sex to test for effect modification. Selection of instruments for drug use that are both theoretically and statistically sound is challenging, and weak instruments can cause inference problems of their own [29]. IVs that have been chosen in past studies for substance use include household characteristics (e.g., history of substance use), personal beliefs (e.g., religiosity), policies (e.g., state legalization of marijuana), and prices [21]. Variables were chosen that could be correlated with attitudes toward risk but that are not correlated with the specific health outcomes from the analysis. Preliminary instrument testing was performed on variables that reflected religious service attendance (as a measure of religiosity) and two sexual behavior measures (as a proxy for attitudes toward risk). The social support questionnaire, which includes religious service attendance, is only available for years 2005 to 2008; therefore, the final analysis did not use religious service attendance as an instrument. The two IVs for past and current marijuana use were defined as follows to capture risk-taking attitudes: “IVfirstsex,” an indicator for first having sex at 16 years or younger and “IVnocondom,” an indicator for having sex without a condom more than 3 times a year. Past marijuana use was considered a viable instrument for current marijuana use; this IV analysis, which considered current use alone as the risk factor of interest, was performed separately. Once instruments were selected and defined from the survey data, they were tested for correlation with the endogenous indicators for marijuana use. In the IV analysis with two endogenous variables, both current and past marijuana use were regressed on the instruments and control variables using least squares methods to verify a correlation with the endogenous variables. The instruments were evaluated in these models using the F statistic for joint significance; the instruments were judged to be significant if the F statistics were greater than a threshold of 10 [30]. Endogeneity of marijuana use in the model for each outcome variable was tested using the augmented regression test or the Durbin-Wu-Hausman (DWH) test. This test examines the explanatory power of the residuals from the first-stage equation, which predicts the probability of marijuana use when added to the original equation [31]. IV analysis was also performed using past marijuana use as an IV for current marijuana use. This analysis followed the same steps for instrument testing, endogeneity testing, and 2SLS estimation outlined for the analysis with the sexual behavior instruments.
Results As summarized in Table 1, there were significant differences across marijuana use categories for demographic characteristics in the sample, including sex, race or ethnicity, age, education level, tobacco use, alcohol consumption, income, BMI category, and postmenopausal status (p < .0001). In unadjusted analyses of the association between marijuana use and cardiometabolic risk factors, significant differences were also observed in the means for fasting insulin, insulin resistance, BMI, and waist circumference (p < .0001; Table 1). Fasting insulin, insulin resistance, BMI, and waist circumference were all significantly lower in current marijuana users compared with lifetime nonusers in multivariate models adjusted for age, sex, race, education, income, BMI, tobacco use, alcohol use, and physical activity (Table 2). Stratification by age and sex appeared to modify the results of some of the OLS regressions, although reduced sample sizes in subgroups may have diminished the precision of estimates. Significant effects of marijuana disappeared entirely in the subgroup aged 40 years and older. Significant effects persisted only for insulin, insulin resistance, and waist circumference in persons younger than 40 years. When analyses were broken down by sex, significant effects disappeared in women. In men, the effect estimates for current users were different from nonusers in models for HDL-C, BMI, and waist circumference. Estimates of the effects on insulin and insulin resistance became nonsignificant in both males and females. As an additional check on the models estimated using OLS from the original analysis, the same multivariate models were used to demonstrate the relationship between metabolic syndrome and four categories of alcohol use, ranging from abstinence to heavy drinking. When alcohol use replaced marijuana use as the risk factor of interest in multivariate linear models, the patterns of significance were similar to the estimates of the effects of marijuana use; however, the estimated effects of heavy alcohol drinking on fasting insulin, HOMA-IR, BMI, and waist circumference were greater than the estimated effects obtained for marijuana use. HDLC was also higher for heavy drinkers compared with nondrinkers (Table 3). When carbohydrate intake was considered as the risk factor of interest in this same manner (limited to 2005e2010 waves of NHANES data), fasting glucose, BMI, and waist circumference were lower in individuals with high carbohydrate consumption compared to low carbohydrate consumption (Table 4). IV analysis was then performed to assess the endogeneity of marijuana use in the OLS regressions and adjust for potential endogeneity. Two sets of IV analyses were performed; the first used two sexual behavior variables as instruments for past and current marijuana use in the first stage of 2SLS regressions, and the second used past marijuana use as an instrument for current use in the first stage. The F test for joint significance of the two sexual behavior instruments in the model models for past use and current use fell below the prespecified threshold of 10 (F ¼ 7.04, 4.52). When past use was included in the model for current use, the F value was well above the threshold (F ¼ 469.17). The results of the DWH test using the two sex instruments were significant in the models for waist circumference and BMI (p < .05). The test statistic approached significance in the model for triglycerides (p ¼ .0544). When past use was included as the instrument for current use, the test was significant in models for waist circumference and BMI again (p < .0001). The test approached significance in the models for insulin resistance (p ¼ .0667) and fasting insulin (p ¼ .0819). When 2SLS was performed using the sexual behavior instruments, the coefficients for current and past marijuana use were nonsignificant (Table 5). The results for current marijuana users
C.A. Thompson, J.W. Hay / Annals of Epidemiology 25 (2015) 486e491
489
Table 1 Characteristics of the study sample and cardiometabolic risk factors from NHANES 2005 to 2012 by marijuana use category Variables Sex Male Female Age (y) 20e29 30e44 45e59 Race or ethnicity Hispanic Non-Hispanic white Non-Hispanic black Other Income (per year) $75,000 Educational level Less than high school High school Some college Alcohol use (past year) Lifetime abstinence 0e1 per wk 14 per wk Tobacco use Never Past Current Physical activity (past month) Inactive Active BMI (kg/m2) category
Contents lists available at ScienceDirect
Annals of Epidemiology journal homepage: www.annalsofepidemiology.org
Original article
Estimating the association between metabolic risk factors and marijuana use in U.S. adults using data from the continuous National Health and Nutrition Examination Survey Christin Ann Thompson BS a, b, Joel W. Hay PhD a, b, * a b
Department of Clinical Pharmacy, Pharmaceutical Economics, and Policy, University of Southern California, Los Angeles Leonard Schaeffer Center for Health Policy and Economics, University of Southern California, Los Angeles
a r t i c l e i n f o
a b s t r a c t
Article history: Received 22 August 2014 Accepted 20 January 2015 Available online 31 January 2015
Purpose: More research is needed on the health effects of marijuana use. Results of previous studies indicate that marijuana could alleviate certain factors of metabolic syndrome, such as obesity. Methods: Data on 6281 persons from National Health and Nutrition Examination Survey from 2005 to 2012 were used to estimate the effect of marijuana use on cardiometabolic risk factors. The reliability of ordinary least squares (OLS) regression models was tested by replacing marijuana use as the risk factor of interest with alcohol and carbohydrate consumption. Instrumental variable methods were used to account for the potential endogeneity of marijuana use. Results: OLS models show lower fasting insulin, insulin resistance, body mass index, and waist circumference in users compared with nonusers. However, when alcohol and carbohydrate intake substitute for marijuana use in OLS models, similar metabolic benefits are estimated. The Durbin-Wu-Hausman tests provide evidence of endogeneity of marijuana use in OLS models, but instrumental variables models do not yield significant estimates for marijuana use. Conclusion: These findings challenge the robustness of OLS estimates of a positive relationship between marijuana use and fasting insulin, insulin resistance, body mass index, and waist circumference. Ó 2015 Elsevier Inc. All rights reserved.
Keywords: Instrumental variables Marijuana use Metabolic health Multivariate linear regression
Introduction Marijuana use has become increasingly prevalent in the United States. In a 2010 survey from the Substance Abuse and Mental Health Services Administration, an estimated 7.3% of Americans aged 12 and older used marijuana in 2012, more than any other illicit substance. Of those who reported using marijuana, an estimated 18.9 million used in the past month and 7.6 million could be considered chronic users [1]. With the rapidly changing policy landscape surrounding the control of marijuana use and its application in healthcare, more research is needed on the short- and long-term health effects of marijuana. Evidence on the effect of marijuana on common disease processes, such as diabetes, would be useful in health care decision-making. Metabolic syndrome is a set of clinical criteria associated with increased risk of type II diabetes and cardiovascular disease [2]. Metabolic syndrome *Corresponding author. Leonard Schaeffer Center for Health Policy and Economics, University of Southern California, Verna and Peter Dauterive Hall, 635 Downey Way, Los Angeles, CA 90089-3333. Tel.: þ1-818-338-5433; fax: þ1-213-740-3460. E-mail address: [email protected] (J.W. Hay). http://dx.doi.org/10.1016/j.annepidem.2015.01.013 1047-2797/Ó 2015 Elsevier Inc. All rights reserved.
includes excess fat around the abdomen, low high-density lipoprotein cholesterol (HDL-C), high triglyceride levels, high blood pressure, high blood sugar level, and insulin resistance [3]. Previous examinations of the relationship between marijuana use and health outcomes have provided conflicting results. In a small controlled study of male research volunteers, periods of marijuana smoking increased daily caloric intake and body weight [4]. In another study that used data from National Health and Nutrition Examination Survey (NHANES) III, total caloric intake was higher in current users but body mass index (BMI) was lower in current users compared with nonusers [5]. One large-sample, retrospective analysis found an association between marijuana use and higher caloric and alcohol consumption but did not find an association between current use and BMI or cardiovascular risk factors [6]. In a case-control study matched for age, sex, ethnicity, and BMI, marijuana use was associated with higher abdominal visceral fat, lower HDL-C, and lower adipocyte insulin resistance; however, there were no differences in total body fat, hepatic steatosis, insulin insensitivity, measures of beta-cell function, or glucose intolerance [7]. A study that used two large U.S. data sets found no significant difference in the multivariate-adjusted odds of obesity in marijuana users compared with abstainers, with the exception of users
C.A. Thompson, J.W. Hay / Annals of Epidemiology 25 (2015) 486e491
who smoked marijuana more than three times a week [8]. Prevalence of overweight or obesity in young adults from the Mater-University of Queensland Study of Pregnancy and its Outcomes was significantly lower in marijuana users in multivariate-adjusted analyses [9]. In contrast, a study using the National Longitudinal Survey of Youth found that, compared to nonuse or low use in adolescence, consistent or increasing patterns of marijuana use in adolescence are associated with an increased risk of obesity [10]. In another study among youth in the United States, frequent marijuana use was associated with overweight status but not obesity in young girls [11]. In a recent study, Penner et al. [12] evaluated the association between self-reported marijuana use and components of metabolic syndrome using the National Health and Nutrition Survey from 2005 to 2010. Surprisingly, the results of ordinary least squares estimation (OLS) of the multivariate linear models suggested reduced fasting insulin, increased HDL-C levels, and a smaller waist circumference in “current users” of marijuana compared with “never users”. Results from the analysis of survey data that indicate improved BMI and other factors of metabolic health contradict what is known about the role of cannabis compounds in the cannabinoid system in humans. It is well known that marijuana contains appetitestimulating compounds known as cannabinoids, which attach to cannabinoid receptors in the brain and other parts of the body [13]. This physiological effect has motivated its use in the treatment of cachexia (wasting syndrome) in cancer and human immunodeficiency virus patients [14,15]. To make sense of the conflicting and sometimes surprising results from the observational studies reviewed in the previous sections, the analytic methods used in these studies should be carefully considered. The models used in prior observational research are based on OLS regression models that disregard the potential endogeneity of marijuana use in health outcome risk equations. When used as explanatory variables, endogenous variables lead to biased regression estimates because of the correlation with missing or unknown control variables [16]. For example, tobacco use is inversely related to obesity [17e19] and directly proportional to marijuana use [20]. If one were to exclude tobacco use inappropriately as a confounder in a model explaining the relationship between marijuana use and health outcomes, the estimated relationship would be potentially biased. Empirical studies have previously acknowledged the endogeneity of substance in the health economics literature [21]. Models that account for endogeneity should be considered, as well as different specifications and checks on model validity. This study aims to explore the relationship between marijuana use and the clinical factors of metabolic syndrome by critically evaluating OLS regression analysis of NHANES data from 2005 to 2012. Materials and methods Sample This study sample included participants from the continuous NHANES from 2005 to 2012, a cross-sectional survey which oversamples young children, older persons, and certain ethnic groups in two-year cycles and applies weights for a nationally representative sample. Participants underwent an in-home interview and laboratory tests, which included blood and urine samples. This investigation focused on the 6281 participants surveyed between 2005 and 2012 who responded to questionnaire items regarding the use of marijuana.
487
continuous NHANES data. Outcome variables included fasting insulin, fasting glucose, homeostasis model assessment of insulin resistance (HOMA-IR; [fasting serum insulin (mU/mL) fasting plasma glucose (mg/dL)/405)], HDL-C, triglycerides, blood pressure (average of three blood pressure readings), BMI (weight in kilograms/height in meters2), and waist circumference. All laboratory measurements were taken at a medical examination center, and the methods are reported in detail in the NHANES Laboratory Procedures Manual [22]. Insulin, HOMA-IR, and triglycerides were log transformed. Explanatory variables were chosen to control for demographic characteristics and risk factors for metabolic syndrome. The variable of interest, marijuana use, was defined using data on selfreported marijuana use from the drug use questionnaire component of the NHANES survey: never users (never smoked marijuana, n ¼ 2861), past users (smoked marijuana at least once but not in the past 30 days, n ¼ 2589), and current users (smoked marijuana at least once in the prior 30 days, n ¼ 831). Significant risk factors for metabolic syndrome include age, sex, race, BMI, tobacco use, physical activity, income, alcohol consumption, carbohydrate intake, and postmenopausal status [23]. Education level was also included as a control variable. As in the study by Penner et al [12], because a significant portion of the income data were missing (z7%), missing values were generated using multiple imputations. The imputations (m ¼ 10) included multivariate normal models to allow for variation in the observations generated by the model. The multivariate imputation model incorporated NHANES stratum effects to account for survey design [24]. Rubin’s [25] rules for combination were applied to adjust estimates and standard errors for variability between imputations. A squared term for age was also added to the full model to capture any nonlinear relationships between age and the health outcomes under examination. Multivariate analysis The appropriate weights were applied based on guidelines from the CDC’s National Center for Health Statistics to account for the complex survey design of NHANES data in regression coefficient and standard error estimates [26]. Separate multiple linear regression models were fit to each outcome variable and to the log-transformed insulin, HOMA-IR, and triglyceride variables. A measure of carbohydrate intake was also added to each model, which has been included in empirical models dealing with metabolic risk factors; however, nutritional data were not available for 2011 to 2012, so the final models did not include carbohydrate intake [23,27,28]. The final regression models adjusted for age, sex, race, education, income, BMI, tobacco use, alcohol use, and physical activity. BMI was left out of the models for BMI and waist circumference. To test for effect modification, multivariate models were stratified by age and sex. Estimating the effects of other risk factors on metabolic syndrome using the multivariate linear models provides an additional robustness check. Alcohol use categories (nondrinkers, 1 drink per week, >1 to 14 drinks per week, and >14 drinks per week) and carbohydrate intake (low, medium, high) were examined in place of marijuana use as the variables of interest in the linear models for each outcome variable. Instrumental variables analysis Instrumental variable (IV) methods can be used to explore the potential endogeneity of marijuana use. The conceptual model is as follows:
Measures
Yi ¼ AXi þ Bmi þ εi
The outcomes of interest were cardiometabolic risk factors. Penner et al. [12] have previously defined these outcomes using
where Yi is a vector of all outcomes, mi is the treatment effect variable for marijuana use, Xi denotes other relevant
488
C.A. Thompson, J.W. Hay / Annals of Epidemiology 25 (2015) 486e491
exogenous personal characteristics, and εi is an error term. The parameter on marijuana use (B) may be (asymptotically) biased if models do not account for endogeneity. Marijuana use can be defined in terms of both relevant exogenous variables from the first equation (Xi ) and variables that instrument for marijuana use (zi ):
mi ¼ gzi þ dXi þ ni : Appropriate exogenous instruments (zi ) will be correlated with a propensity to use marijuana but uncorrelated with the health outcomes of interest in each model. The error terms from the first and second equations, εi and ni , respectively, are assumed to be asymptotically uncorrelated. The method for estimation using this model is typically two-stage least squares (2SLS), which estimates fitted values for marijuana use (mi ) in the first stage using the exogenous instruments [16]. As in OLS estimation, the 2SLS estimation was adjusted based on CDC’s National Center for Health Statistics guidelines to account for complex survey design. IV models were stratified by age and sex to test for effect modification. Selection of instruments for drug use that are both theoretically and statistically sound is challenging, and weak instruments can cause inference problems of their own [29]. IVs that have been chosen in past studies for substance use include household characteristics (e.g., history of substance use), personal beliefs (e.g., religiosity), policies (e.g., state legalization of marijuana), and prices [21]. Variables were chosen that could be correlated with attitudes toward risk but that are not correlated with the specific health outcomes from the analysis. Preliminary instrument testing was performed on variables that reflected religious service attendance (as a measure of religiosity) and two sexual behavior measures (as a proxy for attitudes toward risk). The social support questionnaire, which includes religious service attendance, is only available for years 2005 to 2008; therefore, the final analysis did not use religious service attendance as an instrument. The two IVs for past and current marijuana use were defined as follows to capture risk-taking attitudes: “IVfirstsex,” an indicator for first having sex at 16 years or younger and “IVnocondom,” an indicator for having sex without a condom more than 3 times a year. Past marijuana use was considered a viable instrument for current marijuana use; this IV analysis, which considered current use alone as the risk factor of interest, was performed separately. Once instruments were selected and defined from the survey data, they were tested for correlation with the endogenous indicators for marijuana use. In the IV analysis with two endogenous variables, both current and past marijuana use were regressed on the instruments and control variables using least squares methods to verify a correlation with the endogenous variables. The instruments were evaluated in these models using the F statistic for joint significance; the instruments were judged to be significant if the F statistics were greater than a threshold of 10 [30]. Endogeneity of marijuana use in the model for each outcome variable was tested using the augmented regression test or the Durbin-Wu-Hausman (DWH) test. This test examines the explanatory power of the residuals from the first-stage equation, which predicts the probability of marijuana use when added to the original equation [31]. IV analysis was also performed using past marijuana use as an IV for current marijuana use. This analysis followed the same steps for instrument testing, endogeneity testing, and 2SLS estimation outlined for the analysis with the sexual behavior instruments.
Results As summarized in Table 1, there were significant differences across marijuana use categories for demographic characteristics in the sample, including sex, race or ethnicity, age, education level, tobacco use, alcohol consumption, income, BMI category, and postmenopausal status (p < .0001). In unadjusted analyses of the association between marijuana use and cardiometabolic risk factors, significant differences were also observed in the means for fasting insulin, insulin resistance, BMI, and waist circumference (p < .0001; Table 1). Fasting insulin, insulin resistance, BMI, and waist circumference were all significantly lower in current marijuana users compared with lifetime nonusers in multivariate models adjusted for age, sex, race, education, income, BMI, tobacco use, alcohol use, and physical activity (Table 2). Stratification by age and sex appeared to modify the results of some of the OLS regressions, although reduced sample sizes in subgroups may have diminished the precision of estimates. Significant effects of marijuana disappeared entirely in the subgroup aged 40 years and older. Significant effects persisted only for insulin, insulin resistance, and waist circumference in persons younger than 40 years. When analyses were broken down by sex, significant effects disappeared in women. In men, the effect estimates for current users were different from nonusers in models for HDL-C, BMI, and waist circumference. Estimates of the effects on insulin and insulin resistance became nonsignificant in both males and females. As an additional check on the models estimated using OLS from the original analysis, the same multivariate models were used to demonstrate the relationship between metabolic syndrome and four categories of alcohol use, ranging from abstinence to heavy drinking. When alcohol use replaced marijuana use as the risk factor of interest in multivariate linear models, the patterns of significance were similar to the estimates of the effects of marijuana use; however, the estimated effects of heavy alcohol drinking on fasting insulin, HOMA-IR, BMI, and waist circumference were greater than the estimated effects obtained for marijuana use. HDLC was also higher for heavy drinkers compared with nondrinkers (Table 3). When carbohydrate intake was considered as the risk factor of interest in this same manner (limited to 2005e2010 waves of NHANES data), fasting glucose, BMI, and waist circumference were lower in individuals with high carbohydrate consumption compared to low carbohydrate consumption (Table 4). IV analysis was then performed to assess the endogeneity of marijuana use in the OLS regressions and adjust for potential endogeneity. Two sets of IV analyses were performed; the first used two sexual behavior variables as instruments for past and current marijuana use in the first stage of 2SLS regressions, and the second used past marijuana use as an instrument for current use in the first stage. The F test for joint significance of the two sexual behavior instruments in the model models for past use and current use fell below the prespecified threshold of 10 (F ¼ 7.04, 4.52). When past use was included in the model for current use, the F value was well above the threshold (F ¼ 469.17). The results of the DWH test using the two sex instruments were significant in the models for waist circumference and BMI (p < .05). The test statistic approached significance in the model for triglycerides (p ¼ .0544). When past use was included as the instrument for current use, the test was significant in models for waist circumference and BMI again (p < .0001). The test approached significance in the models for insulin resistance (p ¼ .0667) and fasting insulin (p ¼ .0819). When 2SLS was performed using the sexual behavior instruments, the coefficients for current and past marijuana use were nonsignificant (Table 5). The results for current marijuana users
C.A. Thompson, J.W. Hay / Annals of Epidemiology 25 (2015) 486e491
489
Table 1 Characteristics of the study sample and cardiometabolic risk factors from NHANES 2005 to 2012 by marijuana use category Variables Sex Male Female Age (y) 20e29 30e44 45e59 Race or ethnicity Hispanic Non-Hispanic white Non-Hispanic black Other Income (per year) $75,000 Educational level Less than high school High school Some college Alcohol use (past year) Lifetime abstinence 0e1 per wk 14 per wk Tobacco use Never Past Current Physical activity (past month) Inactive Active BMI (kg/m2) category
