University of Southern California Dissertations and Theses

Selected papers on methods for evaluating the impact of marijuana use on BMI and other risk factors for metabolic syndrome

(USC Thesis Other)

Selected papers on methods for evaluating the impact of marijuana use on BMI and other risk factors for metabolic syndrome

PDF

Download

Open document

Flip pages

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content 1

SELECTED PAPERS ON METHODS FOR EVALUATING THE IMPACT OF MARIJUANA
USE ON BMI AND OTHER RISK FACTORS FOR METABOLIC SYNDROME

By

CHRISTIN THOMPSON

A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALFIORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(PHARMACEUTICAL ECONOMICS AND POLICY)

DECEMBER 2015

2
Dedication
First, I would like to dedicate this work to the family members who have been a part of this
process: Dave Thompson, Claudia Thompson, Heather Robertson, Suzy Thompson, James
Robertson, and Bradley Thompson, as well as my boyfriend, Jon Tringale. These amazing
people have provided me with the resilience to overcome my doubts and frustrations over the
past few years. With their love and generosity, I always know that I have a soft place to land, and
that knowledge gives me the courage to do anything! Writing a dissertation requires endurance
above anything else, and these individuals provide me with my strength.
A couple of my friends have been kind enough to help me and to love me, even when I was tired,
grumpy, boring, and sober. I have to thank these friends for giving me moments of laughter and
play that brightened my weeks, especially Darci Hiemstra and Stephanie Hawkins.
I must also acknowledge Dr. Jeffrey McCombs for not only his contribution to my academic
development, but for his kindness and emotional support over the past 5 years. His friendly face
made every day easier. He is truly an advocate for the students in our program and we are all
extremely fortunate to have him.
A few of my classmates were also helpful with the process of completing this dissertation,
especially Shraddha Chaugule and Barbara Blaylock. Justin McGinnis was also very generous
with his time and emotional support in helping me prepare for my defense.
Finally, I would like to thank Dr. Francine Lederer for helping me grow as a person so that I
could arrive at this place.

3
Acknowledgements
First, I would like to acknowledge Dr. Joel Hay, my advisor, for being an amazing educator over
the past five years. I initially chose Dr. Hay as an adviser because I knew that I would be
challenged by his wisdom and passion for health economics, and I was correct. I am a better
person for it. Each paper in this dissertation was created under his regular guidance and input.
I would also like to acknowledge my committee member Dr. Yang Lu, a teacher and mentor
throughout my entire dissertation writing process; without Dr. Lu I would not have become
involved in substance use research in the first place. My other committee member, Dr. Neeraj
Sood, lent his brilliance in econometrics and health policy to the methods used throughout this
dissertation. Dr. Jason Doctor guided the direction of my Add Health analysis by taking the time
to have brainstorming discussions with me and provide thoughtful input. His outside-the-box
thinking planted the seed for Paper 2 of this dissertation. Dr. Jeff McCombs helped organize and
focus the message a across all parts of the dissertation. He helped me see the forest for the trees.
In addition, I would like to recognize Dr. Orison Woolcott, a scientist specializing in diabetes
and obesity research, for taking the time to answer my questions regarding the cannabinoid
receptor system in the human body. On a similar note, I would like to thank my brother, Bradley
Thompson, PhD candidate, for lending his medical expertise to the interpretation of the results of
my analyses.
Last but not least, I would like to recognize Jillian Wallis, Patricia St Clair, and Brian Tysinger
for continuous support with data and code issues.

4
Table of Contents
1. Introduction ................................................................................................................................. 6
2. Estimating The Association Between Metabolic Risk Factors And Marijuana Use In US
Adults Using Data From The Continuous National Health And Nutrition Examination Survey . 11
Abstract ..................................................................................................................................... 11
Materials and Methods .............................................................................................................. 15
Results ....................................................................................................................................... 20
Discussion ................................................................................................................................. 22
Conclusions ............................................................................................................................... 25
Acknowledgements ................................................................................................................... 26
References ................................................................................................................................. 27
Tables ........................................................................................................................................ 30
3. Does marijuana use lead to weight loss? Exploring the role of effect moderation in the
association between marijuana use and BMI and waist circumference in US adults ................... 36
References ................................................................................................................................. 46
Supplementary Materials ........................................................................................................... 49
References ................................................................................................................................. 54
4. The Role of Behavioral Confounding In the Association between Marijuana Use and BMI in
US Adults ...................................................................................................................................... 55
Background ............................................................................................................................... 55
Objectives .................................................................................................................................. 55
Methods ..................................................................................................................................... 55
Results ....................................................................................................................................... 56
Conclusions ............................................................................................................................... 59
References ................................................................................................................................. 60
5. Methods for addressing endogeneity of marijuana use in models for metabolic health ........... 61
5
Introduction ............................................................................................................................... 61
Materials and Methods .............................................................................................................. 63
Results ....................................................................................................................................... 72
Discussion ................................................................................................................................. 81
References ................................................................................................................................. 84
Appendix A. Univariate Statistics ............................................................................................. 89
Appendix B. Results by gender ................................................................................................. 97

6
1. Introduction
For the past decade marijuana use has increased significantly and has been more prevalent than
the use of any other illicit substance in the US. Of those who reported using marijuana in 2013,
an estimated 19.8 million, or 7.5%, used in the past month (1). The acceptance of marijuana use
for both medical and recreational purposes is increasing. As of July 2015, 23 states and the
District of Columbia have legalized medical marijuana and four states and the District of
Columbia have legalized non-medical use (2). As marijuana is legalized in more states across the
US for non-medical use, we now have to concern ourselves with how the drug impacts the health
of the general population. At present, there is limited well-designed research on the relationship
between recreational marijuana use and health. Marijuana is a schedule 1 drug, which strictly
limits marijuana research. Experimental research on the effects of marijuana use applies to
certain patient populations, such as cancer and aids patients, and not recreational users.
Our current understanding of the cannabinoid receptor system in human beings explains the
pharmacologic effects of cannabis compounds, such as THC, on energy balance, appetite, insulin
sensitivity, pancreatic β-cell function, lipid metabolism, and other processes. Cannabis use and
cannabinoid receptor activation may negatively affect metabolic actions such as insulin
resistance, glucose moderation, HDL-Cholesterol, and triglyceride levels through reduced
adiponactin levels. Cannabinoids also lead to appetite stimulation by attaching to cannabinoid
receptors in the brain (3). This pharmacologic effect has motivated its use in the treatment of
wasting syndrome in cancer and HIV patients (4, 5). Despite our understanding of the
endocannabinoid system and its implications for metabolic health, relatively little is known about
the metabolic effects of chronic cannabis use in humans. Estimating the impact of marijuana use
on metabolic risk would be useful information to policy-makers, clinicians, and recreational
7
marijuana users, given that the diseases associated with metabolic health- heart disease, stroke,
and diabetes- are chronic and expensive health conditions in the United States.
Both experimental and observational research support the proposition that marijuana use is
associated with increased caloric intake. At the same time, some observational studies report that
marijuana users have lower BMI, decreased rates of obesity, and improved insulin resistance.
This raises the question of whether marijuana affects metabolic processes in a paradoxical
fashion: if marijuana use can stimulate appetite while simultaneously preventing weight gain and
poor metabolic health, this would have incredible therapeutic implications.
Observational studies that imply a positive relationship between marijuana use and metabolic
health estimate multivariate models using ordinary least squares or other standard estimation
techniques, such as multinomial logit. Given that the decision to engage in marijuana use is
inherently tied to underlying attitudes/behavioral characteristics of an individual, and at the same
time BMI and other factors of metabolic health are affected by underlying attitudes/behavioral
characteristics, the potential for confounding should be investigated. Furthermore, these studies
rely on self-reported marijuana use, which is subject to measurement error due to underreporting.
So given the trends in marijuana use, its suspected role in certain physiological processes in
humans, and conflicting findings in current research, the general research question this
dissertation will attempt to answer is: what is the actual relationship between marijuana use and
cardio-metabolic health? This research will improve upon the existing research by exploring
alternative methods for estimating the relationship between marijuana use and metabolic health,
finding better data to model the relationship, and exploring a more suitable definition of
marijuana using behavior.
8
I hypothesize that marijuana use does not improve metabolic health, as indicated by several
recent studies, but rather, leads to worse metabolic health. Furthermore, to bolster this hypothesis
I will demonstrate that there are methodological issues in studies that find improved metabolic
risk factors in marijuana users, which may be responsible for the findings. The “true” estimate of
the relationship between marijuana use and metabolic health will come from a model that
accounts for the unique statistical challenges faced when observational data is used to investigate
this relationship.
The data required for my research objectives will be observational data that contains marijuana
use variables and any factors of metabolic syndrome, such as BMI, excess fat around the
abdomen, low HDL Cholesterol, high triglyceride levels, high blood pressure, high blood sugar,
and insulin resistance. The two data sets that fit these criteria are the Continuous National Health
and Nutrition Examination Survey (NHANES) and the National Longitudinal Study of
Adolescent Health (Add Health). NHANES has been used in previous research, including the
study from Penner and colleagues, which found a positive impact of marijuana use on metabolic
health using OLS estimation. Because of this, NHANES is ideal for exploring the validity of
previously reported results. However, the Add Health data, longitudinal study that follows
individuals from adolescence through adulthood and has rich behavioral and contextual
variables, may be better suited to developing a model that accounts for underlying or
unobservable differences between users and non-users. This would address our main
methodological concerns.
Overview of papers: To ultimately test the hypothesis that marijuana use has a negative impact
on metabolic health, it is necessary to investigate the validity of studies that have concluded the
opposite. Hence, a preliminary hypothesis is that such studies are invalid due to methodological
9
issues. This can be demonstrated by creating OLS models that resemble the models used in the
studies in question and testing the validity and robustness of those models. In section 2, I recreate
multivariate linear models of the impact of marijuana use on fasting insulin, fasting glucose,
homeostasis model assessment of insulin resistance (HOMA-IR), high-density lipoprotein
cholesterol (HDL-C), triglycerides, blood pressure, BMI, and waist circumference using
NHANES. The flaws in these models and potential endogeneity are first explored in a
manuscript titled "Marijuana use in Models for Health Outcomes," which has been published in
the American Journal of Medicine. In section two, I also demonstrate the flaws in these models
using other self-reported personal consumption variables in place of marijuana in the models.
Next I formally test endogeneity in OLS models with IV estimation and DWH test and attempt to
correct for the endogeneity using instrumental variables models estimated with two-stage least
squares (2SLS). The results of this analysis are included in a manuscript titled, "Estimating the
association between metabolic risk factors and marijuana use in US adults using data from the
continuous national health and nutrition examination survey," which has been published in the
Annals of Epidemiology (6). In an effort to reinforce the importance of defining marijuana use
variables to capture intensity of use, section 3 explores effect modification by variables that
distinguish chronic marijuana use from casual use in OLS models using NHANES. Section 4
demonstrates effect modification in Add Health in order to explore underlying
behavioral/attitude differences as evidence for endogeneity in multivariate linear models
estimated using OLS. These results were presented as an abstract titled, “Role Of Behavioral
Confounding In The Association Between Marijuana Use And BMI In US Adults” presented at
the ISPOR 20th Annual Meeting in Philadelphia, PA (7). Ultimately, my aim is to develop a
model that “does it all”: addresses endogeneity, causal direction, and measurement error for
10
marijuana use. In section 5, I build on multivariate linear models estimated using OLS and
instrumental variables models estimated with 2SLS by exploiting the panel structure of the Add
Health data. This paper employs panel fixed effects techniques to address endogeneity caused by
unobserved fixed individual differences. Finally, a first-differenced dynamic panel model
estimated using GMM (diff-GMM) is employed to address endogeneity caused by both fixed and
time-varying unobserved heterogeneity.

11
2. Estimating The Association Between Metabolic Risk Factors And Marijuana Use In US
Adults Using Data From The Continuous National Health And Nutrition Examination
Survey
1

Abstract

Purpose: More research is needed on the health impacts of marijuana use. Results of previous
studies indicate that marijuana could alleviate certain factors of metabolic syndrome, such as
obesity. Methods: Data on 6,281 persons from National Health and Nutrition Examination
Survey (NHANES) from 2005 to 2012 was used to estimate the impact of marijuana use on
cardio-metabolic 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 variables methods were utilized 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 to non-users. However, when alcohol and
carbohydrate intake substitute for marijuana use in OLS models, similar metabolic benefits are
estimated. The Durbin-Watson-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.

1
Thompson CA and Hay JW. Annals of Epidemiology (2015)
12
Key Words: Instrumental variables; Marijuana Use; Metabolic health; Multivariate linear
regression
13
Introduction
Marijuana use has become increasingly prevalent in the United States. In a 2010 survey from the
Substance Abuse and Mental Health Services Administration of Americans aged 12 and older, an
estimated 7.3% of Americans 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 (8).
With the rapidly changing policy landscape surrounding the control of marijuana use and its
application in health care, more research is needed on the short and long term health impacts of
marijuana. Evidence on the impact 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 2 Diabetes and cardiovascular disease (9). Metabolic
syndrome includes excess fat around the abdomen, low HDL-Cholesterol, high triglyceride
levels, high blood pressure, high blood sugar, and insulin resistance (10).
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 (11). In another study that
used data from National Health and Nutrition Examination Survey (NHANES) III, total caloric
intake was higher in current users but BMI was lower in current users compared to non- users
(12). 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 (13). In a case-control study matched for age, gender,
ethnicity, and BMI, marijuana use was associated with higher abdominal visceral fat, lower
14
HDL-Cholesterol, 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 (14). A study that used two large US datasets found no significant difference in the
multivariate-adjusted odds of obesity in marijuana users compared to abstainers, with the
exception of users who smoked marijuana more than three times a week (15). Prevalence of
overweight/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 (16). In contrast, a study using the National Longitudinal Survey of Youth found that,
compared to non-use or low-use in adolescence, consistent or increasing patterns of marijuana
use in adolescence are associated with an increased risk of obesity (17). In another study among
youth in the US, frequent marijuana use was associated with overweight status but not obesity in
young girls (18). In a recent study, Penner et al. evaluated the association between self-reported
marijuana use and components of metabolic syndrome using the National Health and Nutrition
Survey from 2005-2010. Surprisingly, the results of ordinary least squares estimation (OLS) of
the multivariate linear models suggested reduced fasting insulin, increased high-density
lipoprotein cholesterol levels, and a smaller waist circumference in “current users” of marijuana
compared to ”never users” (19).
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 appetite-stimulating
compounds known as cannabinoids, which attach to cannabinoid receptors in the brain and other
parts of the body (3). This physiological effect has motivated its use in the treatment of cachexia
(wasting syndrome) in cancer and HIV patients (4, 5). To make sense of the conflicting and
15
sometimes surprising results from the observational studies reviewed above, the analytic
methods employed 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 due to
correlation with missing or unknown control variables (20). For example, tobacco use is
inversely related to obesity (21-23), and directly proportional to marijuana use (24). 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 (25). 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 National Health and Nutrition
Examination Survey (NHANES) from 2005 to 2012, a cross-sectional survey which over-
samples 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 include blood and urine samples. This investigation focused on the 6,281
16
participants surveyed between 2005 and 2012 who responded to questionnaire items regarding
the use of marijuana.
Measures
The outcomes of interest were cardio-metabolic risk factors. Penner et al. have previously
defined these outcomes using continuous NHANES data (19). Outcome variables included
fasting insulin, fasting glucose, homeostasis model assessment of insulin resistance (HOMA-IR)
(fasting serum insulin (muV/mL)*fasting plasma glucose (mg/dL)/405)), high-density
lipoprotein cholesterol (HDL-C), triglycerides, blood pressure (average of 3 blood pressure
readings), BMI (weight in kilograms/height in meters
2
), 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(26). 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 self-
reported marijuana use from the drug use questionnaire component of the NHANES survey:
never users (never smoked marijuana, n= 2861), past users (smoked marijuana at least once but
not in the past 30 days, n= 2589), and current users (smoked marijuana at least once in the prior
30 days, n= 831).
Significant risk factors for metabolic syndrome include age, sex, race, BMI, tobacco use,
physical activity, income, alcohol consumption, carbohydrate intake, and postmenopausal status
(27). Education level was also included as a control variable. As in the study by Penner and
colleagues, because a significant portion of the income data was missing (≈7%), missing values
17
were generated using multiple imputations. The imputations (m=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 (28). Rubin’s rules for combination were applied in order to adjust estimates and standard
errors for variability between imputations (29). A squared term for age was also added to the full
model to capture any non-linear relationships between age and the health outcomes under
examination.
Multivariate Analysis
The appropriate weights were applied based on guidelines from the CDC/National Center for
Health Statistics to account for the complex survey design of NHANES data in regression
coefficient and standard error estimates (30). 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 was
not available for 2011-2012, so the final models did not include carbohydrate intake (27, 31, 32).
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
gender.
Estimating the impacts 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 & <=14 drinks per week, >14 drinks per week) and carbohydrate intake (low,
18
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:
!
"
=$%
"
+ ()
"
+ *
"

Where +
,
is a vector of all outcomes, -
,
is the treatment effect variable for marijuana use, .
,

denotes other relevant exogenous personal characteristics, and /
,
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 (.
,
) and variables that instrument for marijuana use (1
,
):
)
"
=23
"
+ 4%
"
+ 5
"

Appropriate exogenous instruments (1
,
) 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, /
,
and 6
,
, 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 (-
,
) in the first stage using the exogenous instruments
(20). As in OLS estimation, the 2SLS estimation was adjusted based on CDC/National Center for
Health Statistics guidelines to account for complex survey design. IV models were stratified by
age and gender to test for effect modification.
19
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 (33). Instrumental
variables 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 (25). Variables were chosen that could be correlated
with attitudes towards 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 2005-2008; therefore the final analysis did not use
religious service attendance as an instrument. The two instrumental variables for past and current
marijuana use were defined as follows in order to capture risk-taking attitudes: “IVfirstsex” is an
indicator for first having sex at 16 years or younger and “IVnocondom” is 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 (34). Endogeneity of marijuana use in the model for each outcome variable was
20
tested using the augmented regression test, or the Durbin-Wu-Hausman test (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 (35).
Instrumental variables analysis was also performed using past marijuana use as an instrumental
variable 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 shown in Table 1, there were significant differences across marijuana use categories for
demographic characteristics in the sample, including sex, race/ethnicity, age, education level,
tobacco use, alcohol consumption, income, BMI category, and postmenopausal status (P <
0.0001). In unadjusted analyses of the association between marijuana use and cardio-metabolic
risk factors, significant differences were also observed in the means for fasting insulin, insulin
resistance, BMI, and waist circumference (P < 0.0001) (Table 1).
Fasting insulin, insulin resistance, BMI, and waist circumference were all significantly lower in
current marijuana users compared to lifetime non-users in multivariate models adjusted for age,
sex, race, education, income, BMI, tobacco use, alcohol use, and physical activity (Table 2).
Stratification by age and gender 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 and older.
Significant effects persisted only for insulin, insulin resistance, and waist circumference in
persons aged younger than 40. When analyses were broken down by gender, significant effects
21
disappeared in females. In males, the effect estimates for current users were different from non-
users in models for HDL-C, BMI, and waist circumference. Estimates of the effects on insulin
and insulin resistance became non-significant 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 impacts 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. HDL-C was
also higher for heavy drinkers compared to non-drinkers (Table 3). When carbohydrate intake
was considered as the risk factor of interest in this same manner (limited to 2005-2010 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).
Instrumental variables analysis was then performed in order 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 pre-
specified threshold of 10 (F = 7.04, 4.52). When past use was included in the model for current
use, the F-value was well above the threshold (F = 469.17).
22
The results of the DWH test using the two sex instruments were significant in the models for
waist circumference and BMI (P < 0.05). The test statistic approached significance in the model
for triglycerides (P = 0.0544). When past use was included as the instrument for current use, the
test was significant in models for waist circumference and BMI again (P < 0.0001). The test
approached significance in the models for insulin resistance (P = 0.0667) and fasting insulin (P =
0.0819).
When 2SLS was performed using the sexual behavior instruments, the coefficients for current
and past marijuana use were non-significant (Table 5). The results for current marijuana users
were also non-significant in the full models with past marijuana use as an instrument (Table 6).
Stratification by age and gender did not modify results.
Discussion
The results of multivariate linear regressions estimated using OLS were consistent with prior
research (19). Current marijuana users appeared to have lower fasting insulin, improved insulin
resistance, lower BMI, and a smaller waist than non-users. The analysis that follows
demonstrates the fallibility of OLS methods that have been employed in studies that estimate the
relationship between marijuana use and BMI and other factors of the metabolic profile (11-16,
19, 36).
When marijuana use is replaced by alcohol use in OLS regression models, alcohol use is also
associated with lower fasting insulin, HOMA-IR, BMI, and waist circumference. The association
between these health outcomes and moderate to heavy drinking is at odds with the widely
acknowledged detrimental health effects of alcohol consumption (37, 38). Given the positive
relationship between alcohol consumption and marijuana use (P < .00), these results raise
23
concerns for the validity of the multivariate linear models for the association between marijuana
use and health outcomes such as factors of metabolic syndrome. Similarly, the findings that
increased carbohydrate consumption reduces BMI and other metabolic risk factors is also highly
implausible (27, 39, 40).
In addition, despite weak instruments for past and current marijuana use, IV methods offer an
endogeneity bias explanation for the surprising marijuana treatment effect estimates from OLS
regressions. The results of the DWH test suggested that marijuana use was endogenous in the
models for waist circumference and BMI. The test approached significance in models or
triglyceride levels, fasting insulin, and insulin resistance. The endogeneity of marijuana use at
the theoretical level and as evidenced using instrumental variables analysis is the major
conceptual issue for the models of the effects of marijuana use on metabolic syndrome risk
factors. In the face of endogeneity, the OLS regression estimates are biased by the correlation
between marijuana use and any unobserved characteristics also affecting these risks.
The inclusion of dietary variables mediates the estimated relationship between marijuana use and
cardio-metabolic risk factors. Dietary data was not available in the final wave of evidence, so
dietary variables were left out of the final model; however, when the models were estimated
using a sample limited to the first three waves of data, carbohydrate intake mitigated the effects
of marijuana use on waist circumference. Concurrent drug use also has the potential to alter the
estimates of the association between marijuana use and health outcomes; however, data on illicit
drug use was limited in NHANES. Longitudinal data that captures duration, sequence, and
intensity of use of various illicit substances would enhance the analysis.
24
Another weakness of the empirical research in this field is imprecise marijuana use variables. We
were constrained by the NHANES data in our ability to create a more precise measure of
marijuana use. Variables are based on self-reported marijuana use and generated from
questionnaire items that fail to capture intensity of marijuana use. Future research in this field
would benefit immensely from data on drug use based on urinalysis or blood tests in order to
precisely capture drug use behavior. When that is not possible, survey data with questionnaire
items that capture the intensity of use could improve the accuracy of marijuana use variables.
Finding high-quality instrumental variables for marijuana use in the NHANES data was also
challenging. The sexual behavior questionnaire was missing a large portion of responses and was
no doubt subject to measurement error issues. The sexual behavior instrument that was a dummy
for engaging in sex without a condom may reflect attitudes towards risk differently in
married/cohabiting persons. In an attempt to account for this, first, marital status was added as a
control variable in the IV models. Then, the IV models were performed in a sub-sample of
unmarried individuals. These analyses did not strengthen the instruments nor did they alter the
significance of the estimates from two-stage least squares. Given the limited selection of
available instruments in the NHANES data, it was not feasible to test and instrument for all
possible endogenous explanatory variables. Alcohol use, tobacco use, and BMI are also
potentially endogenous to these models and may further bias results. Instrumental variables
analysis would be improved by data with questionnaire items that capture behavioral or policy
variables associated with the propensity to use marijuana.
Nevertheless, estimates of the impact of marijuana use from IV analysis that used the most
reasonable instruments available from the NHANES data were inconsistent with the OLS
estimates; these results undermine the robustness of prior findings. This study provides evidence
25
that OLS models are inadequate in this context and that models that address the problem of
endogeneity should be considered in future research.
Conclusions
Previous observational research on the impact of marijuana use on metabolic syndrome has
offered conflicting results. The reliability of the epidemiologic methods employed and
underlying data in these studies must be critically examined to makes sense of these conflicting
results and, ultimately, to arrive at reliable estimates of the impact of marijuana use. This study
demonstrates that OLS models for estimating this relationship are flawed. Given that alcohol use
and carbohydrate consumption are risk factors for metabolic health (37-40), the results of OLS
models that examine these risk factors raise concerns for the validity of the multivariate linear
models that have been used to estimate the impact of marijuana use on metabolic syndrome in
previous studies. The theoretical and empirical evidence of endogeneity of marijuana use in the
multivariate linear models for metabolic syndrome also undermines the use of OLS models for
capturing this relationship, and offers an explanation for the ostensible bias in OLS estimates.

26
Acknowledgements
We thank Patricia St. Claire and Brian Tysinger for their assistance with the NHANES data and
Jeff McCombs and Andrew Messali for their helpful advice in the interpretation of the results.
We would also like to thank Dr. Orison Woolcott for lending his expertise regarding the
cannabinoid system and its relationship to metabolic processes.
This work was not supported through any grants or financial support.

27
References
1. Substance Abuse and Mental Health Services Administration, Results from the 2010
National Survey on Drug Use and Health: Summary of National Findings, NSDUH
Series H-41, HHS Publication No. (SMA) 11-4658. Rockville, MD: Substance Abuse
and Mental Health Services Administration, 2011.
2. Alberti KG, Zimmet P, Shaw J, et al. The metabolic syndrome--a new worldwide
definition. Lancet 2005;366(9491):1059-62.
3. Kassi E, Pervanidou P, Kaltsas G, et al. Metabolic syndrome: definitions and
controversies. BMC Med 2011;9:48.
4. Foltin RW, Fischman MW, Byrne MF. Effects of smoked marijuana on food intake and
body weight of humans living in a residential laboratory. Appetite 1988;11(1):1-14.
5. Smit E, Crespo CJ. Dietary intake and nutritional status of US adult marijuana users:
results from the Third National Health and Nutrition Examination Survey. Public Health
Nutr 2001;4(3):781-6.
6. Rodondi N, Pletcher MJ, Liu K, et al. Marijuana use, diet, body mass index, and
cardiovascular risk factors (from the CARDIA study). Am J Cardiol 2006;98(4):478-84.
7. Muniyappa R, Sable S, Ouwerkerk R, et al. Metabolic effects of chronic cannabis
smoking. Diabetes Care 2013;36(8):2415-22.
8. Le Strat Y, Le Foll B. Obesity and cannabis use: results from 2 representative national
surveys. Am J Epidemiol 2011;174(8):929-33.
9. Hayatbakhsh MR, O'Callaghan MJ, Mamun AA, et al. Cannabis use and obesity and
young adults. Am J Drug Alcohol Abuse 2010;36(6):350-6.
10. Huang DY, Lanza HI, Anglin MD. Association between adolescent substance use and
obesity in young adulthood: a group-based dual trajectory analysis. Addict Behav
2013;38(11):2653-60.
11. Farhat T, Iannotti RJ, Simons-Morton BG. Overweight, obesity, youth, and health-risk
behaviors. Am J Prev Med 2010;38(3):258-67.
12. Penner EA, Buettner H, Mittleman MA. The impact of marijuana use on glucose, insulin,
and insulin resistance among US adults. Am J Med 2013;126(7):583-9.
13. Kumar RN, Chambers WA, Pertwee RG. Pharmacological actions and therapeutic uses of
cannabis and cannabinoids. Anaesthesia 2001;56(11):1059-68.
14. Pertwee RG. Cannabis and cannabinoids: Pharmacology and rationale for clinical use.
Forsch Komplementmed 1999;6(Suppl. 3):12-5.
15. Gorter R. Cancer cachexia and cannabinoids. Forschende
Komplementärmedizin/Research in Complementary Medicine 2004;6(Suppl. 3):21-2.
16. Angrist JD, Pischke J-S. Mostly harmless econometrics: an empiricist's companion.
Princeton: Princeton University Press; 2009.
17. Patel K, Hargreaves MK, Liu JG, et al. Relationship Between Smoking and Obesity
Among Women. American journal of health behavior 2011;35(5):627-36.
18. Clair C, Chiolero A, Faeh D, et al. Dose-dependent positive association between cigarette
smoking, abdominal obesity and body fat: cross-sectional data from a population-based
survey. BMC PUBLIC HEALTH 2011;11(1):23.
19. Matsushita Y, Nakagawa T, Yamamoto S, et al. Associations of smoking cessation with
visceral fat area and prevalence of metabolic syndrome in men: the Hitachi health study.
Obesity (Silver Spring) 2011;19(3):647-51.
28
20. Rigotti NA, Lee JE, Wechsler H. US college students' use of tobacco products - Results
of a national survey. Jama-J Am Med Assoc 2000;284(6):699-705.
21. French MT, Popovici I. That instrument is lousy! In search of agreement when using
instrumental variables estimation in substance use research. Health Econ 2011;20(2):127-
46.
22. Centers for Disease Control and Prevention. National Health and Nutrition Examination
Survey: Laboratory Procedures Manual,
http://www.cdc.gov/nchs/data/nhanes/nhanes_11_12/2011-
12_Laboratory_Procedures_Manual.pdf; 2011 [accessed on 10.16.14]. [electronic
article].
23. Park YW, Zhu S, Palaniappan L, et al. The metabolic syndrome: prevalence and
associated risk factor findings in the US population from the Third National Health and
Nutrition Examination Survey, 1988-1994. Arch Intern Med 2003;163(4):427-36.
24. Reiter JP, Raghunathan TE, Kinney SK. The importance of modeling the sampling
design in multiple imputation for missing data. Survey methodology 2006;32(2):143-50.
25. Little RJA, Rubin DB. The Analysis of Social-Science Data with Missing Values.
Sociological Methods & Research 1989;18(2-3):292-326.
26. When and how to construct weights when combining survey cycles [Internet].
Hyattsville, MD: National Center for Health Statistics.
(http://www.cdc.gov/nchs/tutorials/nhanes/surveydesign/Weighting/intro_i.htm).
(Accessed 17 September 2013).
27. Willett WC. Will high-carbohydrate/low-fat diets reduce the risk of coronary heart
disease? Proc Soc Exp Biol Med 2000;225(3):187-90.
28. Wirfalt E, Hedblad B, Gullberg B, et al. Food Patterns and Components of the Metabolic
Syndrome in Men and Women: A Cross-sectional Study within the Malmo Diet and
Cancer Cohort. American journal of epidemiology 2001;154(12):1150-9.
29. Bound J, Jaeger DA, Baker R. The Cure Can Be Worse than the Disease: A Cautionary
Tale Regarding Instrumental Variables. NBER Working Paper Series 1993;t0137(Journal
Article).
30. Stock JH, Wright JH, Yogo M. A Survey of Weak Instruments and Weak Identification
in Generalized Method of Moments. Journal of Business & Economic Statistics,
2002:518-29.
31. Davidson R, MacKinnon JG. Estimation and inference in econometrics. New York:
Oxford University Press; 1993.
32. Rajavashisth TB, Shaheen M, Norris KC, et al. Decreased prevalence of diabetes in
marijuana users: cross-sectional data from the National Health and Nutrition Examination
Survey (NHANES) III. BMJ Open 2012;2:e000494.
33. National Institute on Alcohol Abuse and Alcoholism. Beyond hangovers: Understanding
alcohol's impact on your health . NIH; 2012. (NIH Publication No. 10–7604).
34. American Heart Association. "Heavy, Lifetime Alcohol Users May Be Toasting
Metabolic Syndrome." ScienceDaily.
www.sciencedaily.com/releases/2004/11/041116233400.htm (accessed August 21, 2014).
(Accessed).
35. Foster GD, Wyatt HR, Hill JO, et al. A randomized trial of a low-carbohydrate diet for
obesity. The New England journal of medicine 2003;348(21):2082-90.
29
36. Volek JS, Phinney SD, Forsythe CE, et al. Carbohydrate Restriction has a More
Favorable Impact on the Metabolic Syndrome than a Low Fat Diet. Lipids
2009;44(4):297-309.

30
Tables
Table 1. Characteristics of the Study Sample and Cardio-metabolic Risk Factors from NHANES 2005-2012 by
Marijuana Use Category
a

Never
use
(n =
2861) 95% CI
Past use
(n =
2589) 95% CI
Current
use
(n = 831) 95% CI P-value
Sex <0.0001
Male 0.43 (0.41-0.45) 0.53 (0.5-0.55) 0.66 (0.61-0.7)
Female 0.57 (0.55-0.59) 0.47 (0.45-0.5) 0.34 (0.3-0.39)
Age (years) <0.0001
20-29 0.23 (0.2-0.25) 0.23 (0.21-0.25) 0.41 (0.36-0.47)
30-44 0.39 (0.37-0.41) 0.35 (0.32-0.37) 0.33 (0.29-0.37)
45-59 0.38 (0.35-0.41) 0.42 (0.4-0.45) 0.26 (0.2-0.31)
Race/ethnicity <0.0001
Hispanic 0.23 (0.19-0.27) 0.10 (0.08-0.11) 0.09 (0.06-0.11)
Non-Hispanic white 0.55 (0.5-0.6) 0.76 (0.74-0.79) 0.70 (0.65-0.75)
Non-Hispanic black 0.12 (0.1-0.14) 0.10 (0.09-0.12) 0.17 (0.14-0.21)
Other 0.10 (0.08-0.12) 0.04 (0.03-0.05) 0.04 (0.02-0.06)
Income (per year) <0.0001
<$20,000 0.14 (0.12-0.16) 0.10 (0.09-0.12) 0.21 (0.18-0.24)
$20,000-$44,999 0.26 (0.23-0.29) 0.23 (0.2-0.26) 0.26 (0.22-0.31)
$45,000-$74,999 0.26 (0.23-0.29) 0.25 (0.22-0.28) 0.22 (0.18-0.27)
>$75,000 0.34 (0.3-0.37) 0.42 (0.38-0.46) 0.30 (0.25-0.35)
Educational Level <0.0001
Less than high school 0.19 (0.16-0.21) 0.12 (0.1-0.14) 0.22 (0.17-0.26)
High school 0.21 (0.19-0.24) 0.21 (0.18-0.23) 0.29 (0.24-0.34)
Some college 0.60 (0.57-0.63) 0.67 (0.64-0.71) 0.50 (0.44-0.56)
31
Alcohol Use (past year) <0.0001
Lifetime abstinence 0.42 (0.39-0.44) 0.19 (0.17-0.22) 0.09 (0.06-0.12)
0-1 per week 0.31 (0.29-0.33) 0.28 (0.26-0.3) 0.20 (0.16-0.23)
< 1 and <=14 per week 0.25 (0.23-0.28) 0.44 (0.41-0.47) 0.52 (0.48-0.56)
>14 per week 0.02 (0.01-0.03) 0.08 (0.06-0.1) 0.19 (0.15-0.23)
Tobacco Use <0.0001
Never 0.79 (0.76-0.81) 0.44 (0.41-0.47) 0.27 (0.22-0.31)
Past 0.10 (0.09-0.12) 0.28 (0.25-0.31) 0.17 (0.13-0.21)
Current 0.11 (0.09-0.12) 0.28 (0.25-0.31) 0.56 (0.5-0.62)
Physical Activity (past
month) 0.12
Inactive 0.41 (0.38-0.43) 0.37 (0.34-0.41) 0.37 (0.31-0.42)
Active 0.59 (0.57-0.62) 0.63 (0.59-0.66) 0.63 (0.58-0.69)
BMI (kg/m
2
) category <0.0001
bmi<25 0.31 (0.29-0.34) 0.32 (0.29-0.34) 0.45 (0.39-0.5)
bmi>=25 & <30 0.31 (0.29-0.33) 0.34 (0.32-0.37) 0.31 (0.26-0.35)
bmi>30 0.38 (0.36-0.4) 0.34 (0.31-0.37) 0.25 (0.21-0.29)
Postmenopausal Status <0.0001
yes 0.84 (0.82-0.86) 0.88 (0.86-0.89) 0.94 (0.91-0.96)
no 0.16 (0.14-0.18) 0.12 (0.11-0.14) 0.06 (0.04-0.09)
Cardio-metabolic risk factors
Insulin (µU/mL)† 10.30 (9.94-10.66) 9.17 (8.78-9.57) 7.79 (7.29-8.33) <0.0001
Glucose (mg/dL) 102.85
(101.61-
104.09) 100.74 (99.53-101.95) 100.62 (98.27-102.97) 0.05
HOMA-IR† 2.55 (2.45-2.66) 2.24 (2.13-2.36) 1.90 (1.77-2.04) <0.0001
Triglycerides (mg/dL)† 104.66
(101.15-
108.29) 107.07 (103.5-110.77) 106.11
(100.86-
111.64) 0.57
HDL-C (mg/dL) 53.14 (52.38-53.89) 53.21 (52.25-54.17) 53.56 (52.3-54.81) 0.82
BMI (kg/m2) 29.14 (28.77-29.52) 28.64 (28.28-29) 26.93 (26.38-27.48) <0.0001
Waist circumference (cm) 97.43 (96.55-98.31) 98.05 (97.16-98.94) 93.40 (91.89-94.92) <0.0001
32
SBP (mm Hg) 117.41
(116.48-
118.34) 117.36
(116.59-
118.13) 118.53
(117.18-
119.87) 0.23
DBP (mm Hg) 70.56 (69.81-71.31) 70.90 (70.16-71.63) 69.31 (68.14-70.47) 0.05
Abbreviations: BMI, body mass index; CI, Confidence interval; HDL-C, high-density lipoprotein cholesterol; HOMA-IR
homeostasis model assessment of insulin resistance; NHANES, National Health and Nutrition Examination Survey
a
Characteristics reported as proportions; cardio-metabolic risk factors reported as means; all analyses weighted to account for
survey design. Samples from NHANES 2005-2012.
b
Variables are log transformed

Table 2. Multivariate Linear Regression for Association Between Marijuana Use
and Cardio-metabolic Risk Factors from NHANES 2005-2012
a

Dependent Variables Past Use Current Use
Insulin (uU/mL) -0.046 (0.027) -0.116** (0.037)
HOMA-IR -0.059 (0.030) -0.108** (0.039)
Glucose (mg/dL) -1.503 (0.866) 1.324 (1.472)
Triglycerides (mg/dL) -0.009 (0.022) 0.022 (0.035)
HDL-C (mg/dL) -0.110 (0.503) 0.726 (0.621)
BMI (kg/m2) 0.051 (0.223) -0.771* (0.340)
Waist (cm) 0.475 (0.528) -2.120* (0.885)
SBP (mm Hg) -1.045 (0.560) 0.251 (0.785)
DBP (mm Hg) 0.080 (0.448) 0.513 (0.672)
Abbreviations: BMI, body mass index; HDL-C, high-density lipoprotein cholesterol;
HOMA-IR homeostasis model assessment of insulin resistance; NHANES, National
Health and Nutrition Examination Survey
** p<0.01; * p<0.05
a
All analyses weighted to account for survey design; standard errors in parentheses

33
Table 3. Multivariate Linear Regression for Association Between Alcohol Consumption and Cardio-metabolic
Risk Factors from NHANES 2005-2012 (Non-drinkers Omitted)
a

Dependent Variables ≤1 drink per week <1 & ≤14 drinks per week >14 drinks per week
Insulin (uU/mL) 0.004 (0.025) -0.056 (0.031) -0.180** (0.048)
HOMA-IR 0.008 (0.027) -0.053 (0.031) -0.172** (0.052)
Glucose (mg/dL) -0.032 (1.314) -0.066 (1.107) 0.945 (1.649)
Triglycerides (mg/dL) 0.050* (0.023) 0.005 (0.027) 0.045 (0.043)
HDL-C (mg/dL) 0.259 (0.457) 4.171** (0.584) 10.935** (1.046)
BMI (kg/m2) -0.600 (0.328) -1.665** (0.280) -2.022** (0.431)
Waist (cm) -1.020 (0.805) -3.544** (0.674) -4.686** (1.032)
SBP (mm Hg) 0.208 (0.585) 0.282 (0.626) 5.295** (1.244)
DBP (mm Hg) 0.152 (0.395) 0.089 (0.470) 1.822* (0.897)
Abbreviations: BMI, body mass index; HDL-C, high-density lipoprotein cholesterol; HOMA-IR homeostasis model
assessment of insulin resistance; NHANES, National Health and Nutrition Examination Survey
** p<0.01; * p<0.05
a
All analyses weighted to account for survey design; standard errors in parentheses
34

Table 4. Multivariate Linear Regression for Association Between Carbohydrate
Intake and Cardio-metabolic Risk Factors from NHANES 2005-2012 (Low Intake
Omitted)
a

Dependent Variables Medium intake High intake
Insulin (uU/mL) 0.015 (0.031) 0.065 (0.042)
HOMA-IR -0.011 (0.035) 0.041 (0.047)
Glucose (mg/dL) -4.817** (1.623) -4.589* (1.960)
Triglycerides (mg/dL) 0.036 (0.028) 0.168** (0.036)
HDL-C (mg/dL) -2.329** (0.736) -4.940** (0.853)
BMI (kg/m2) -0.202 (0.344) -1.347** (0.376)
Waist (cm) 0.041 (0.727) -2.705** (0.943)
SBP (mm Hg) -1.276 (0.779) -1.275 (0.781)
DBP (mm Hg) -0.494 (0.607) 0.659 (0.784)
Abbreviations: BMI, body mass index; HDL-C, high-density lipoprotein cholesterol;
HOMA-IR homeostasis model assessment of insulin resistance; NHANES, National
Health and Nutrition Examination Survey
** p<0.01; * p<0.05
a
All analyses weighted to account for survey design; standard errors in parentheses

Table 5. Instrumental Variables Analysis with Sexual Behavior Variables as
Instruments for Marijuana Use from NHANES 2005-2012
a

Dependent Variables Past Use Current Use
Insulin (uU/mL) 0.147 (0.611) -0.910 (1.291)
HOMA-IR 0.018 (0.553) -0.547 (1.178)
Glucose (mg/dL) -19.713 (30.563) 62.384 (55.716)
Triglycerides (mg/dL) 0.456 (0.406) -0.152 (0.853)
HDL-C (mg/dL) -10.533 (12.000) 1.035 (23.981)
BMI (kg/m2) 0.165 (5.549) 10.370 (12.794)
Waist (cm) 5.762 (11.373) 4.641 (25.474)
SBP (mm Hg) 0.255 (10.162) -6.792 (21.349)
DBP (mm Hg) 8.982 (9.719) -7.614 (17.120)
Abbreviations: BMI, body mass index; HDL-C, high-density lipoprotein cholesterol;
HOMA-IR homeostasis model assessment of insulin resistance; NHANES, National
Health and Nutrition Examination Survey
** p<0.01; * p<0.05
a
All analyses weighted to account for survey design; standard errors in parentheses

35

Table 6. Instrumental Variables Analysis with Past
Marijuana Use as an Instrument for Current Use
from NHANES 2005-2012
a

Dependent Variables Current Use
Insulin (uU/mL) 0.033 (0.084)
HOMA-IR 0.075 (0.096)
Glucose (mg/dL) 5.447 (2.939)
Triglycerides (mg/dL) 0.012 (0.058)
HDL-C (mg/dL) 2.030 (1.618)
BMI (kg/m2) -0.629 (0.773)
Waist (cm) -2.958 (1.870)
SBP (mm Hg) 2.204 (1.646)
DBP (mm Hg) -0.641 (1.487)
Abbreviations: BMI, body mass index; HDL-C, high-
density lipoprotein cholesterol; HOMA-IR homeostasis
model assessment of insulin resistance; NHANES,
National Health and Nutrition Examination Survey
** p<0.01; * p<0.05
a
All analyses weighted to account for survey design;
standard errors in parentheses

36
3. Does marijuana use lead to weight loss? Exploring the role of effect moderation in the
association between marijuana use and BMI and waist circumference in US adults
2

As the U.S. undergoes considerable state-level relaxation of legal restrictions on cannabis, it is
important to have better understanding of the long-term effects of cannabis use. In addition to
other physiologic effects, marijuana and related cannabinoids cause appetite stimulation by
attaching to cannabinoid receptors in the central nervous system and other parts of the body (3,
5). This effect has motivated FDA-approved marijuana-based cannabinoid therapies (e.g.,
dronabinol and nabilone) for anorexia and cachexia in cancer and HIV patients (4, 41-43).
However, because obesity, diabetes and metabolic syndrome are among the most rapidly
increasing and burdensome health problems in the U.S. population, expanded marijuana access
and use in the general population may have negative public health implications (44).
In both clinical and observational research, non-medical marijuana use has been associated with
increased caloric intake (11-13, 45). Due to appetite stimulating effects, recreational marijuana
use may contribute to obesity and visceral obesity, which, in turn, are risk factors for metabolic
health (46). However, relatively little is known about the effects of non-medical marijuana use
on metabolic health in humans. Rigorous clinical studies of cannabinoid treatment effects are
based on short term, small sample human trials (47). Previous estimates of the relationship
between marijuana use and body mass index (BMI), waist circumference, and obesity rates offer
conflicting results. While the outcome of some studies are consistent with the appetite-
stimulating effects of marijuana (14, 17), in others, marijuana users appear to have lower BMI,
waist circumference, and rates of obesity (6, 12, 15, 16, 19, 48). This study aims to explore that

2
Thompson
,
CA, Hay JW, Doctor JN
37
contradiction by identifying whether or not marijuana use is consistently related to weight loss as
indicated by BMI and waist circumference, or if the relationship between marijuana use and
weight varies at different levels of marijuana consumption.
Observational research presents unique challenges for analyzing the relationship between
marijuana use and BMI, waist circumference, and other metabolic risk factors. For example,
marijuana use variables often fail to capture the duration or frequency of use (6, 12, 17, 19, 48).
As a result, the impact of occasional marijuana use is conflated with the impact of chronic
marijuana use when, in fact, there may be an important distinction. Light et al. reported that the
vast majority of marijuana consumption following legalization in Colorado was attributable to
21.8% of users, considered frequent users (account for 66.9% of demand for marijuana), with
almost one-third of all users identifying as “rare-users” (49). These statistics demonstrate the
potential variability in the frequency of marijuana use, even among “current users” identified in
the research. There are certain variables in NHANES that we argue are capable of distinguishing
chronic or heavy users from mild users. In this study, we explore how those variables moderate
the effect of marijuana use on the dependent variables. A moderator variable impacts the strength
or direction of the relationship between the variable of interest and the dependent variable (50).
We analyzed data from 6,238 respondents of the nationally representative Continuous National
Health and Nutrition Examination Survey (NHANES) between 2005 and 2012. This sample
included respondents who completed questionnaire items regarding marijuana use behavior and
for whom height, weight, and waist circumference measurements were available (as well as
additional covariates). Complete marijuana use data restricted the sample age to 20-59.
38
We describe the basic multivariate linear models in detail elsewhere (51). BMI was calculated
using respondents’ measured height and weight (weight in kilograms/height in meters
2
). Waist
circumference was measured in centimeters. The marijuana use variable has four categories
based on self-reported use: non-user (never once tried marijuana), past user (reported using
marijuana, but not in the past 30 days), low current user (used less than 10 times in the past 30
days), and high current user (used 10 or more times in the past 30 days). Among the sample
cohort, 45.6% were non-users, 41% were past users, 7% were low current users, and 6% were
high current users. In contrast with previously reported results, when “high current use” is
distinguished from all current use, only high use is significantly associated with reductions in
BMI and waist circumference (6, 19). Even when users are identified by the number of days of
recent use, this variable may fail to capture the duration or intensity of marijuana use among
recent users. Moreover, self-reported marijuana use is likely to be underreported (1, 52).
We further examined whether gender, the age of onset of use, and chronic user status “moderate”
the relationships between marijuana use and BMI and waist circumference. A moderator variable
may influence the strength or direction of the relationship between the outcome and the predictor
variable of interest (marijuana use, in this case.) We explored the influence of moderator
variables by including interaction terms between each of those variables and marijuana use in
separate models. Moderator variables were chosen based on their potential to distinguish among
users with different levels of intensity of marijuana consumption. This goal is obviously
achieved by exploring the interaction between marijuana use and indicators for “chronic user
status” and “early age of onset.” However, the motivation behind interacting marijuana use with
gender in order to distinguish between mild users and chronic users is less straightforward.
Gender was identified as a moderator variable because male users have been known to have
39
higher or more problematic levels of marijuana consumption than females, even among current
users (53-55). Preliminary univariate estimates using the NHANES 2005-2012 data show that,
among daily users, males have higher levels of marijuana consumption than females. In addition,
research suggests that females have lower rates of support for marijuana use and legalization (56,
57). Accordingly, when an interaction between gender and marijuana use is added to the model,
coefficient estimates for that interaction term are significant or nearly significant (p = 0.059 in
the model for BMI and p = 0.0058 in the model for waist circumference), suggesting that gender
moderates the effect of past month marijuana use on BMI and waist circumference. The
regression coefficients from the multivariate linear models are described in detail and displayed
in the supplementary materials. Figures 1 and 2 below illustrate how the predicted levels of BMI
and waist circumference compare among the different gender-marijuana user combinations. In
these figures, the “base” category represents predicted BMI in males versus females for the full
sample. The remaining four categories represent differences in predictive margins between males
and females for each category of marijuana user. The negative correlation between marijuana use
and BMI estimated in the original models appears to exist in males only, as represented by the
trend in the blue point estimates in Figure 1. The significant negative correlation is not observed
in females from the sample. Effect moderation by gender appears to exist for the relationship
between marijuana use and waist circumference as well, as illustrated by Figure 2.

40
Figure 1.

Figure 2.

41
The younger the age of onset of marijuana use, the higher the risk of cannabis abuse or
dependence (58, 59). Additionally, the age of onset of marijuana use may serve as a proxy for
duration of use. Therefore, a separate set of models includes an interaction term to capture the
relationship between user status and age of onset of marijuana use (age 16 or younger versus
later than 16). The sample included in these models is isolated to individuals who have used
marijuana at any time in the past (non-users excluded). In these models, the age of onset
moderates estimates of the effects on waist circumference and BMI: in users with an age of onset
17 or older, there is no significant difference between marijuana user status and estimated BMI
or waist circumference (test for significance of interaction terms: p = 0.13 in BMI model and p =
0.03 in waist circumference model). The trend in BMI and waist circumference over levels of
increasing marijuana use appears to be negative in users who initiated marijuana use prior to 17
years old and neutral or positive in users who initiated marijuana use at age 17 or older (Figures
3 and 4).

42
Figure 3.

Figure 4.

43
Another proxy for duration of marijuana use or chronic marijuana use is the questionnaire item
that indicates whether or not a user has used every month for a year at any time in the past.
Among all past and current marijuana users (nonusers again excluded due to the moderator
variable of interest), the negative relationship between current marijuana use and both waist
circumference and BMI appears to be isolated to users who reported having used every month
for a year (test for significance of interaction terms: p = 0.10 in the model for BMI and p = 0.14
in the model for waist). Based on the results displayed in Figures 5 and 6, there does not appear
to be any significant relationship between marijuana use and BMI or waist circumference among
users who reported never having used with consistency for a year. This suggests that the
relationship is stronger in or isolated to users who may be considered chronic users.
Figure 5.

44
Figure 6.

Why would chronic marijuana use be associated with lower BMI and waist circumference? A
plausible explanation is a dulling or switching of the “agonization” effect of cannabinoids on
CB1 receptors, which are associated with appetite and other metabolic factors. Over time, the
CB1 receptors may develop tolerance to the appetite stimulating effects of the phytocannabinoids
from marijuana, and, as a result, more cannabinoids, from both within the body and from
marijuana, are required to achieve baseline levels of appetite stimulation. Evidence for a
switching effect is provided by studies on withdrawal symptoms in marijuana dependent
subjects: subjects in these studies experience weight loss following cessation (60, 61).
In previous work we postulated that estimates of a negative relationship between marijuana use
and metabolic risk factors, such as BMI, were biased due to endogeneity of the marijuana use
variable in the regression, meaning that marijuana use may be influenced by unobserved
confounders that also influence metabolic health but are not controlled for in the regression (6).
We still suspect that endogeneity biases estimates downwards and that it is plausible that
45
marijuana use actually leads to weight gain in most users, except for a subgroup of “chronic”
users who experience habituated appetite suppression (62). Future research should account for
both issues: researchers should use methods to test and correct for the potential for an
endogeneity bias and they should define marijuana use in a manner that allows for the possibility
that the relationship is modified by intensity and duration of marijuana use.
The effect moderation by gender, age of onset of use, and chronic user status in this analysis
suggests that there are distinct effects at different levels of marijuana consumption, because these
variables are proxies for more intense levels of marijuana use. Results showing reduced weight,
waist circumference, or rates of obesity in marijuana users may be attributable to altered appetite
regulation following chronic and/or high amounts of marijuana use. Marijuana is known to be
addictive, leading to chronic habitual cannabinoid use (63, 64). If cannabinoids suppress normal
appetite in chronic habitual users, they are unlikely to be a practical therapy for weight loss,
diabetes or metabolic syndrome, as some accounts suggest (65, 66), since this beneficial effect
may be largely confined to those with chronic excessive use and/or cannabinoid addiction.

46
References
1. R. N. Kumar, W. A. Chambers, R. G. Pertwee, Pharmacological actions and therapeutic
uses of cannabis and cannabinoids. Anaesthesia 56, 1059-1068 (2001).
2. R. G. Pertwee, Cannabis and cannabinoids: Pharmacology and rationale for clinical use.
Forsch Komplementmed 6, 12-15 (1999).
3. T. F. Plasse et al., Recent clinical experience with dronabinol. Pharmacology,
biochemistry, and behavior 40, 695-700 (1991).
4. R. Gorter, Cancer cachexia and cannabinoids. Forschende
Komplementärmedizin/Research in Complementary Medicine 6, 21-22 (2004).
5. J. E. Beal et al., Dronabinol as a treatment for anorexia associated with weight loss in
patients with AIDS. Journal of pain and symptom management 10, 89-97 (1995).
6. K. Nelson, D. Walsh, P. Deeter, F. Sheehan, A phase II study of delta-9-
tetrahydrocannabinol for appetite stimulation in cancer-associated anorexia. Journal of
palliative care 10, 14-18 (1994).
7. American Diabetes Association, Economic costs of diabetes in the U.S. in 2012. Diabetes
Care 36, 1033-1046 (2013).
8. R. W. Foltin, J. V. Brady, M. W. Fischman, Behavioral analysis of marijuana effects on
food intake in humans. Pharmacology, biochemistry, and behavior 25, 577-582 (1986).
9. R. W. Foltin, M. W. Fischman, M. F. Byrne, Effects of smoked marijuana on food intake
and body weight of humans living in a residential laboratory. Appetite 11, 1-14 (1988).
10. N. Rodondi, M. J. Pletcher, K. Liu, S. B. Hulley, S. Sidney, Marijuana use, diet, body
mass index, and cardiovascular risk factors (from the CARDIA study). Am J Cardiol 98,
478-484 (2006).
11. E. Smit, C. J. Crespo, Dietary intake and nutritional status of US adult marijuana users:
results from the Third National Health and Nutrition Examination Survey. Public Health
Nutr 4, 781-786 (2001).
12. K. G. Alberti, P. Zimmet, J. Shaw, The metabolic syndrome--a new worldwide definition.
Lancet 366, 1059-1062 (2005).
13. P. F. Whiting et al., Cannabinoids for Medical Use: A Systematic Review and Meta-
analysis. JAMA 313, 2456-2473 (2015).
14. R. Muniyappa et al., Metabolic effects of chronic cannabis smoking. Diabetes Care 36,
2415-2422 (2013).
15. D. Y. Huang, H. I. Lanza, M. D. Anglin, Association between adolescent substance use
and obesity in young adulthood: a group-based dual trajectory analysis. Addict Behav 38,
2653-2660 (2013).
16. M. R. Hayatbakhsh et al., Cannabis use and obesity and young adults. Am J Drug Alcohol
Abuse 36, 350-356 (2010).
17. Y. Le Strat, B. Le Foll, Obesity and cannabis use: results from 2 representative national
surveys. Am J Epidemiol 174, 929-933 (2011).
18. E. A. Penner, H. Buettner, M. A. Mittleman, The impact of marijuana use on glucose,
insulin, and insulin resistance among US adults. Am J Med 126, 583-589 (2013).
19. C. A. Thompson, J. W. Hay, Estimating the association between metabolic risk factors
and marijuana use in US adults using data from the continuous National Health and
Nutrition Examination Survey. Annals of epidemiology, (2015).
47
20. G. Ngueta, R. E. Belanger, E. A. Laouan-Sidi, M. Lucas, Cannabis use in relation to
obesity and insulin resistance in the inuit population. Obesity (Silver Spring) 23, 290-295
(2015).
21. Market Size and Demand for Marijuana in Colorado (2014
https://www.colorado.gov/pacific/sites/default/files/Market Size and Demand Study, July
9, 2014%5B1%5D.pdf).
22. R. M. Baron, D. A. Kenny, The moderator-mediator variable distinction in social
psychological research: conceptual, strategic, and statistical considerations. J Pers Soc
Psychol 51, 1173-1182 (1986).
23. Materials and methods are available as supplementary materials on Science Online.
24. J. Zhao, T. Stockwell, S. MacDonald, Non-response bias in alcohol and drug population
surveys. Drug and Alcohol Review 28, 648-657 (2009).
25. Results from the 2013 National Survey on Drug Use and Health: Summary of National
Findings (2014).
26. D. Kandel, K. Chen, L. A. Warner, R. C. Kessler, B. Grant, Prevalence and demographic
correlates of symptoms of last year dependence on alcohol, nicotine, marijuana and
cocaine in the U.S. population. Drug Alcohol Depend 44, 11-29 (1997).
27. C. Coffey, J. B. Carlin, M. Lynskey, N. Li, G. C. Patton, Adolescent precursors of
cannabis dependence: findings from the Victorian Adolescent Health Cohort Study. The
British journal of psychiatry : the journal of mental science 182, 330-336 (2003).
28. L. A. Warner, R. C. Kessler, M. Hughes, J. C. Anthony, C. B. Nelson, Prevalence and
correlates of drug use and dependence in the United States. Results from the National
Comorbidity Survey. Archives of general psychiatry 52, 219-229 (1995).
29. J. Carroll, Who Supports Marijuana Legalization? Gallup Poll Tuesday Briefing, (2005).
30. The Pew Research Center for People and the Press, "In Debate Over Legalizing
Marijuana, Disagreement Over Drug’s Dangers," (2015; Available from
http://www.people-press.org/2015/04/14/in-debate-over-legalizing-marijuana-
disagreement-over-drugs-dangers/2/).
31. K. E. Leonard, G. G. Homish, Changes in marijuana use over the transition into marriage.
Journal of drug issues 35, 409-429 (2005).
32. G. G. Homish, K. E. Leonard, Spousal influence on general health behaviors in a
community sample. Am J Health Behav 32, 754-763 (2008).
33. J. G. Bachman, P. M. Omalley, L. D. Johnston, Drug-Use among Young-Adults - the
Impacts of Role Status and Social-Environment. Journal of Personality and Social
Psychology 47, 629-645 (1984).
34. R. M. Warner, Applied statistics: from bivariate through multivariate techniques.
(SAGE Publications, Thousand Oaks, Calif, 2008).
35. C. Y. Chen, M. S. O'Brien, J. C. Anthony, Who becomes cannabis dependent soon after
onset of use? Epidemiological evidence from the United States: 2000-2001. Drug Alcohol
Depend 79, 11-22 (2005).
36. S. Behrendt, H. U. Wittchen, M. Hofler, R. Lieb, K. Beesdo, Transitions from first
substance use to substance use disorders in adolescence: is early onset associated with a
rapid escalation? Drug Alcohol Depend 99, 68-78 (2009).
37. A. J. Budney, B. A. Moore, R. G. Vandrey, J. R. Hughes, The time course and
significance of cannabis withdrawal. J Abnorm Psychol 112, 393-402 (2003).
48
38. A. J. Budney, J. R. Hughes, B. A. Moore, R. Vandrey, Review of the validity and
significance of cannabis withdrawal syndrome. Am J Psychiatry 161, 1967-1977 (2004).
39. D. A. Savitz, Interpreting epidemiologic evidence: strategies for study design and
analysis. (Oxford University Press, New York, 2003).
40. A. J. Budney, R. Roffman, R. S. Stephens, D. Walker, Marijuana dependence and its
treatment. Addiction science & clinical practice 4, 4-16 (2007).
41. M. C. Greene, J. F. Kelly, The prevalence of cannabis withdrawal and its influence on
adolescents' treatment response and outcomes: a 12-month prospective investigation. J
Addict Med 8, 359-367 (2014).
42. J. S. Alpert, Marijuana for diabetic control. Am J Med 126, 557-558 (2013).
43. M. Szalavitz, Marijuana: The Next Diabetes Drug. Time Magazine, (2013).

49
Supplementary Materials
Materials and Methods
Data: The Continuous National Health and Nutrition Examination Survey (NHANES) is a
complex, multistage probability survey of noninstitutionalized civilians in the US that over-
samples young children, older persons, and certain ethnic groups and applies weights for a
nationally representative sample.
Statistical methods: Multivariate linear regression analysis estimates the linear, additive
relationship between marijuana use and BMI or waist circumference. The estimated coefficients
from the regressions represent the change in the expected level of BMI or waist circumference
for each category of marijuana use relative to non-use. The original models labeled in the tables
provide coefficient estimates for the predictor of interest (marijuana) and are controlled for other
variables that may effect the estimate of those coefficients. Models with interaction terms include
a moderator variable along with the predictor of interest; moderator variables interact with
predictor to “partition” the effect on the outcome, which can lead to changes in magnitude or
direction of the relationship reflected in coefficient estimates. The figure below illustrates how
effect moderation influences the relationship between two variables.

50
Figure 1.

In these multivariate linear regression analyses, age, gender, race/ethnicity, education level,
income, smoking status, alcohol use, and physical activity level variables are included in order to
estimate the effect of marijuana use while controlling for the presence of these other variables. In
addition, a square term for age was included to capture any nonlinearity in the relationship
between age and body size. To account for the complex survey design, analyses were weighted
using guidelines from the CDC/National Center for Health Statistics (67). The basic model and
multiple imputation for missing income data have been described elsewhere (6).
Supplementary text
We found only three observational studies that distinguished frequent users separately from
infrequent or experimental users (13, 15, 16). In two of these studies, lower odds of obesity and
overweight occur in or are isolated to the most frequent users (15, 16). The third study found no
difference in BMI between any category of users and non-users (13).
Outcome
Predictor
Moderator
Predictor x
Moderator
a
b
c
51
Table 1. Multivariate Adjusted BMI and Waist Circumference
VARIABLES BMI
Waist
Circumference

Past user 0.05 0.47
(0.22) (0.53)
Used less than 10 times in past
mo. -0.59 -2.05*
(0.41) (1.10)
Used 10 or more times in past
mo. -1.01** -2.23**
(0.43) (1.08)
Female -0.54** -6.85***
(0.21) (0.55)

Observations 6,554 6,517
R-squared 0.08 0.14
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

Table 2. Multivariate Adjusted BMI and Waist Circumference,
gender interaction
VARIABLES BMI
Waist
Circumference

Past user -0.06 0.34
(0.24) (0.68)
Used less than 10 times in past
mo. -1.57*** -4.82***
(0.50) (1.29)
Used 10 or more times in past
mo. -1.43*** -4.15***
(0.50) (1.28)
Female*Non-user 0.00 0.00
(0.00) (0.00)
Female*Past user 0.17 0.08
(0.40) (0.99)
Female*Used less than 10 times 2.23*** 6.32***
(0.82) (2.02)
Female*Used 10 or more times 1.23 6.22***
52
(0.97) (2.30)
Female -0.83*** -7.60***
(0.30) (0.75)

Observations 6,554 6,517
P-value for joint test of
interaction terms: 0.0598 0.0058
R-squared 0.09 0.14
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

Table 3. Multivariate Adjusted BMI and Waist Circumference,
age of onset interaction
VARIABLES BMI
Waist
Circumference

Used less than 10 times in past
mo. -0.90* -3.12**
(0.51) (1.35)
Used 10 or more times in past
mo. -1.62** -4.40***
(0.65) (1.58)
Later onset of marijuana use (17
or older)
-0.13 -0.39
(0.39) (0.90)
Later onset*Past user 0.00 0.00
(0.00) (0.00)
Later onset*Used less than 10
times 0.75 2.07
(1.09) (2.66)
Later onset*Used 10 or more
times 1.95** 6.35**
(0.97) (2.41)

Observations 4,001 3,986
P-value for joint test of
interaction terms: 0.13 0.03
R-squared 0.08 0.13
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

53

Table 4. Multivariate Adjusted BMI and Waist Circumference,
chronic user status interaction
VARIABLES BMI
Waist
Circumference

Identified as chronic user 0.97** 2.82***
(0.41) (1.00)
Used less than 10 times in past mo. -0.10 -1.35
(1.05) (2.77)
Used 10 or more times in past mo. 5.43 5.88
(4.57) (6.96)
Chrontic user*Past user 0.00 0.00
(0.00) (0.00)
Chronic user*Used less than 10
times -1.70 -3.62
(1.30) (3.45)
Chronic user*Used 10 or more
times -7.92 -12.13*
(4.84) (6.56)

Observations 1,984 1,977
P-value for joint test of interaction
terms: 0.0655 0.1835
R-squared 0.10 0.14
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

54
References
1. When and how to construct weights when combining survey cycles [Internet].
Hyattsville, MD, National Center for Health Statistics (2010; Available from
http://www.cdc.gov/nchs/tutorials/Nhanes/SurveyDesign/Weighting/Task2.htm).
2. C. A. Thompson, J. W. Hay, Estimating the association between metabolic risk factors
and marijuana use in US adults using data from the continuous National Health and
Nutrition Examination Survey. Annals of epidemiology, (2015).
3. W. R. Rice, Analyzing Tables of Statistical Tests. Evolution 43, 223 (Jan, 1989).
4. J. M. Bland, D. G. Altman, Multiple significance tests: the Bonferroni method. BMJ 310,
170 (Jan 21, 1995).
5. N. Rodondi, M. J. Pletcher, K. Liu, S. B. Hulley, S. Sidney, Marijuana use, diet, body
mass index, and cardiovascular risk factors (from the CARDIA study). Am J Cardiol 98,
478 (Aug 15, 2006).
6. M. R. Hayatbakhsh et al., Cannabis use and obesity and young adults. Am J Drug Alcohol
Abuse 36, 350 (Nov, 2010).
7. Y. Le Strat, B. Le Foll, Obesity and cannabis use: results from 2 representative national
surveys. Am J Epidemiol 174, 929 (Oct 15, 2011).

55
4. The Role of Behavioral Confounding In the Association between Marijuana Use and
BMI in US Adults
3

Background
A number of studies have attempted to estimate the impact of marijuana use on BMI and obesity.
It is well-known that the compounds in marijuana increase appetite(4, 5, 11, 13), but some
empirical studies have found that, surprisingly, marijuana users have lower BMI and rates of
obesity than non-users(12, 15, 16, 19). Failure of these studies to account for differences in
behaviors and attitudes increases the potential for confounding since the decision to engage in
marijuana use is tied to underlying behaviors and attitudes.
Objectives
Given conflicting results in the existing body of literature, what is the relationship between
marijuana use and BMI? This study investigated how estimates of that relationship change when
models are stratified by variables that reflect differences in attitude or behavior. We discuss the
implications of effect modification for previous estimates of the relationship between marijuana
use and BMI/obesity.
Methods
Data
Participants in Wave IV of the National Longitudinal Study of Adolescent Health (Add Health)
were interviewed between 2008 and 2009.(68) Participant ages ranged from 24-32.

3
Thompson CA and Hay JW (2015). Accepted Poster presentation for the ISPOR 20
th
Annual
Meeting, Philadelphia, PA.
56
Variables
The dependent variable is BMI, measured in kg/m
2
. The risk factor of interest is marijuana use,
defined as current user (smoked in the last 30 days) or non-current user (has not smoked in that
last 30 days). All analyses are controlled for age, age squared, gender, race income, physical
activity level, smoking status, and alcohol consumption. A squared term for age was added to
capture nonlinearity in effects of age.
Multivariate Adjusted Analyses and Effect Modification
Multivariate regressions were used to ascertain the relationship between marijuana use (as
indicated by use in that past 30 days) and BMI (kg/m2) using Add Health Wave IV data.
Regressions were then stratified by variables that might define subgroups based on behavioral
differences, including age of onset of marijuana use, concurrent alcohol use category, self-
assessed weight status, self-reported risk-propensity, and gender. All analyses were adjusted for
survey design.
Conceptual model:

Where Y_i is a vector of all outcomes, m_i is the variable for marijuana use, X _i denotes
other relevant exogenous personal characteristics, and ϵ_i is an error term.
Results
(See table) Consistent with findings reported in other empirical studies,
5,7
marijuana use was
significantly associated with lower BMI in the core model (P = 0.015). When the models are
stratified by variables that define behaviors and attitudes, the results change. Marijuana use
57
appears to be associated with lower BMI only in marijuana users who initiated use at 16 years or
younger (P = 0.040). The association disappears completely in users who initiated use older than
16 years (P = 0.937). Negative effect estimates persist only in individuals who considered
themselves to be underweight (P = 0.005); in individuals who reported being satisfactory or
overweight, there was no significant association between marijuana use and BMI (P = 0.313 and
P = 0.748, respectively). When subgroups are defined by risk-taking personality attributes,
results were significant only in the subgroups of individuals who identified themselves as risk-
takers (P = 0.549 and P = 0.401 in groups not identified as risk-takers, P= 0.037 in risk-takers).
When the models are stratified by alcohol consumption categories, the results change drastically.
Marijuana use remains significantly associated with lower BMI in individuals who drink more
than once a week (P < 0.05); however, in the subgroup of non-drinkers who have tried alcohol,
the magnitude and direction of the effect changes significantly (drinkers: -0.996, p = 0.015 vs.
non-drinkers: 7.96, p = 0.007). Results are not significant in females (P = 0.976). See table for
effect estimates for each model.

58
Association between Wave IV marijuana use and BMI: multivariate linear model and
stratified multivariate linear models
Sub-group Stratification Variables Marijuana
Use
(SE) N
Original Model N/A -1.00** (0.41) 2,770
Gender
Female 0.02 (0.69) 1,533
Male -1.83*** (0.47) 1,237
Age first tried marijuana
Tried marijuana 16 or
younger -1.56** (0.75) 574
Tried marijuana older than 16 -0.04 (0.54) 761
Alcohol use
Never tried a drink -1.35 (1.70) 515
Tried drinking, not in past
year 7.96*** (2.91) 228
Drink less than once a week -0.13 (0.79) 1,168
Drink 1-2 times a week -2.34*** (0.66) 548
Drink 3+ times a week -1.17** (0.58) 311
Tobacco Use
Never tried smoking 0.46 (1.80) 1,039
Tried smoking, never
regularly -1.67** (0.79) 543
Smoked regularly in the past -1.42 (1.29) 328
Current smoker -0.96* (0.57) 864
Income Category
Less than $30,000 -0.06 (1.04) 294
$30,000-$49,999 -0.18 (0.92) 649
$50,000-$74,999 -1.45 (1.03) 668
$75,000 or more -1.60** (0.75) 850
Self-assessed weight status
Underweight -2.23*** (0.78) 212
About right -0.30 (0.29) 941
Overweight 0.19 (0.60) 1,617
Social Personality
Disagree -0.58 (0.88) 744
Neutral -0.34 (0.77) 474
Agree -1.13** (0.54) 1,550
"Don't consider
consequences"
Disagree -0.55 (0.67) 1,729
Neutral -1.11 (0.91) 470
Agree -2.13** (0.93) 570
"Risk-taker"
Disagree -0.55 (0.92) 1,086
Neutral -0.67 (0.80) 739
Agree -1.62** (0.77) 945
All analyses controlled for age, age squared, gender, race income, physical activity level,
smoking status, and alcohol consumption. Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
59
Conclusions
Our results suggest that the extent to which marijuana use influences BMI may depend on other
behavioral factors. This finding requires further validation in more representative cohorts.
Further investigation is also required to pinpoint the behavioral differences between individuals
who may gain weight versus individuals who are safe from weight gain with marijuana use, as
well as the intensity of marijuana use associated with changes in BMI. The results have
implications for assessing the role of marijuana use in strategies to address obesity.

60
References
1. Foltin RW, Fischman MW, Byrne MF. Effects of smoked marijuana on food intake and body
weight of humans living in a residential laboratory. Appetite 1988;11:1-14.
2. Rodondi N, Pletcher MJ, Liu K, et al. Marijuana use, diet, body mass index, and
cardiovascular risk factors (from the CARDIA study). Am J Cardiol 2006;98:478-84.
3. Pertwee RG. Cannabis and cannabinoids: Pharmacology and rationale for clinical use. Forsch
Komplementmed 1999;6:12-5.
4. Gorter R. Cancer cachexia and cannabinoids. Forschende Komplementärmedizin/Research in
Complementary Medicine 2004;6:21-2.
5. Penner EA, Buettner H, Mittleman MA. The impact of marijuana use on glucose, insulin, and
insulin resistance among US adults. Am J Med 2013;126:583-9.
6. Hayatbakhsh MR, O'Callaghan MJ, Mamun AA, et al. Cannabis use and obesity and young
adults. Am J Drug Alcohol Abuse 2010;36:350-6.
7. Smit E, Crespo CJ. Dietary intake and nutritional status of US adult marijuana users: results
from the Third National Health and Nutrition Examination Survey. Public Health Nutr
2001;4:781-6.
8. Le Strat Y, Le Foll B. Obesity and cannabis use: results from 2 representative national
surveys. Am J Epidemiol 2011;174:929-33.
9. Harris KM. The add health study: Design and accomplishments. Chapel Hill, NC 2013.

61
5. Methods for addressing endogeneity of marijuana use in models for metabolic health
Introduction
Given marijuana use trends in the US, estimating the impact of marijuana use on metabolic risk
would be useful information for policy-makers, clinicians, and recreational marijuana users (1).
At the same time, the diseases associated with metabolic syndrome- heart disease, stroke, and
diabetes- are chronic and expensive health conditions in the United States (44).
A few studies have attempted to estimate the impact of marijuana use on metabolic risk factors,
such as obesity rates, BMI, waist circumference, and insulin resistance. It is well known that the
compounds in marijuana can increase appetite (4, 5, 11, 13, 45); however, some empirical
studies have found that, surprisingly, marijuana users have lower BMI and rates of obesity than
non-users (12, 15, 16, 19). Muniyappa et al. found that chronic marijuana users had higher
abdominal visceral fat, lower HDL-Cholesterol, and lower adipocyte insulin resistance, but no
differences in total body fat, hepatic steatosis, insulin insensitivity, or glucose intolerance (14).
Rodondi et al. found no association between marijuana use and systolic blood pressure,
triglyceride levels, BMI, lipid levels, or glucose levels (13). Conversely, a study using data from
the continuous National Health and Nutrition Examination Survey found lower fasting insulin,
higher HDL-C levels, and a smaller waist circumference in self-reported marijuana users
compared to non-users (19). Results from adjusted analyses using a survey of adult Inuit
residents of Nunavik, Quebec BMI, found lower percent fat mass, fasting insulin, insulin
resistance, and LDL-cholesterol in past-year marijuana users while HDL-cholesterol was higher
(these results were dissipated when models controlled for BMI) (48). These studies have been
summarized in greater detail elsewhere (6).
62
The literature up to this point lacks agreement. Past studies have relied on naïve OLS regressions
and don’t account for unobserved differences between marijuana users and non-users, which
may confound regression estimates. To answer the research question using survey data, a model
should account for unobserved individual differences, potential endogeneity due to selection
bias, and direction of causality. Due to the failure of current research to address such
methodological considerations, this study will employ some novel models that account for the
unique statistical challenges faced by empirical studies investigating the relationship between
marijuana use and metabolic health.
We have previously analyzed this problem using data from the Continuous National Health and
Nutrition Examination Survey (NHANES), which provided multiple years of cross-sectional
survey data (6). Cross-sectional OLS models, which predicted lower fasting insulin, insulin
resistance, body mass index, and waist circumference in users compared to non-users, suffered
from the same methodological shortcomings and the potential endogeneity bias. Cross-sectional
instrumental variables models, estimated using two-stage least squares, were considered to
remedy the bias caused by the endogeneity of marijuana use. However, in the absence of strong
and good available instruments in the NHANES data, instrumental variables models estimated
with two-stage least squares are unable to overcome the issue of bias due to endogeneity.
This study compares estimates of the effect of marijuana use computed with different models.
The goal is to determine the magnitude and direction of the impact of marijuana use behavior on
metabolic health. This paper employs four different models. First, we recreate the results found
in previous research using linear models estimated with ordinary least squares. Then, because
many of the dependent variables are only available in Wave IV, we attempt to instrument for
marijuana use with exogenous instrumental variables. These models are estimated using two-
63
stage least squares (2SLS). Because Add Health is longitudinal, and BMI is available in all
waves, next we take advantage of multiple panels with a panel fixed effects model, which
eliminates unobserved individual fixed differences as the cause of endogeneity. Finally, to
address any remaining concerns over endogeneity and selection bias, we employ a dynamic
panel model that is estimated using the first differences GMM method. This model eliminates
fixed effects through first difference transformation and, additionally, uses GMM-style lagged
instruments to address the correlation between covariates and the idiosyncratic error. These
models will be described in more detail in the following sections.
Materials and Methods
Sample
The data for this empirical analysis comes from Waves I, III, and IV of the National
Longitudinal Study of Adolescent Health (Add Health). Conducted by the Carolina Population
Center of University of North Carolina (UNC), Add Health follows a nationally representative
sample of adolescents starting in grades 7-12 in 1994. Add Health includes longitudinal survey
data on respondents’ health, demographic, behavioral, and economic characteristics as well as
contextual data on family, school, and community. This analysis employs Add Health restricted-
use datasets, which are more extensive than the publicly available data. Adolescents were first
interviewed in 1994-1995 (Wave I), followed up in 1996 (Wave II), interviewed again in 2001-
2002 (Wave III), and last interviewed in 2007-2008 (Wave IV), at ages 24-32. All Wave I
respondents were eligible for Waves III and IV. Only respondents from Wave I who did not
matriculate or drop-out of school were eligible one year later to interview for Wave II. Due to
this attrition, Wave II was excluded from this study.
64
There are a few important advantages of Add Health over survey data used in previous studies,
such as NHANES, to address the relationship between marijuana use and BMI. First, because it
is panel data, panel models, which eliminate unobserved time-invariant individual differences,
can be used to address the previously mentioned concerns about the endogeneity of marijuana
use. Additionally, geographic and contextual variables offer a rich selection of instruments and
control variables for models.
Measures
Marijuana use
Marijuana use is defined in Waves I, III, and IV using questionnaire items that correspond to
whether or not a person has used marijuana and how often marijuana users have consumed
marijuana in the previous 30 days. Marijuana use is distinguished with four categories in the
following analysis: individuals who have never used marijuana, individuals who have used
marijuana but not in the previous 30 days, individuals who have used marijuana less than eight
days out of the previous 30, and individuals who have used marijuana eight or more days out of
the previous 30.
Basic demographic information and contextual variables
Respondent gender, age, and race were provided for each wave. Wave I also contained
information regarding respondent’s family background, such as parent education level, family
income, and household composition. A variable indicating whether a person’s mother was obese
was also included as a control, since maternal obesity is a risk factor for obesity (69). In Waves
III and IV, in which the full sample had reached adulthood, the respondent’s own socioeconomic
variables were collected. Marital status from Waves III and IV was also included.
65
Health behavioral variables
A categorical variable for physical activity level was created using instances of mild to moderate
physical activity, as defined by the “Compendium of Physical Activity” (Ainsworth 1992).
Activity levels were low (0-2 times per week), moderate (3-4 times per week), or high (5 or more
times per week). Four levels of alcohol use were defined: non-drinkers (0 times in the past year),
infrequent drinker (less than once a week), casual drinker (1-2 days per week), and heavy drinker
(3-7 days per week. Tobacco smoking behavior was defined using three categories: non-lifetime
non-smoker (tried less than 100 times in lifetime), past smoker (has tried more than 100 times or
smoked regularly in the past but not currently), and current smoker (regular smoker in the past
30 days).
Metabolic Health Variables
In waves III and IV, BMI (weight in kilograms/height in meters
2
) was constructed using
respondent’s in-home measured height and weight data. Wave I BMI was constructed using self-
reported height and weight, which introduces a potential measurement error. To explore the
differences between measured and reported height and weight in this sample, measured and
reported height and weight from Wave II were compared. The differences were small but
significant, with the primary difference driven by under-reported weight in females by an
average of two pounds; therefore self-reported weight may contain measurement error that
deflates the estimated effects in regressions.
Because child and adolescent BMI changes naturally with age, and the change occurs differently
across gender, it is necessary to account for age and gender when using BMI to assess weight
status in adolescents. Accordingly, we standardized anthropomorphic measures for age and
66
gender in Waves I, III, and IV. Because there were so few observations for ages 11 and 12 in
Wave I, the reference population for creating standardized z-scores at these ages is considered
too small, so these respondents aged 11 or 12 in Wave I are excluded from analysis.
A few variables related to metabolic health were found only in Wave IV, precluding the
possibility of panel analyses for such variables. These include waist circumference (cm), fasting
and non-fasting glucose, HbA1c, triglycerides, HDL cholesterol, and LDL cholesterol.
Instruments for Marijuana Use
Contemporaneous instruments must be strictly exogenous, which is theoretically difficult. With
regards to personal characteristics, it is challenging to find variables that predict marijuana use
but are unrelated to behaviors that might contribute to BMI. Therefore, behavioral variables will
be considered bad instruments, suspected of contributing to asymptotic bias in 2SLS. A potential
strong instrument candidate is political ideology. A Gallup National Poll from 2010 reported that
among self-identified liberals, 72% supported marijuana legislation; conversely, only 30% of
identified conservatives supported marijuana legislation.
Policy/contextual variables may be superior instruments, but a similar dilemma exists: what
political or social characteristics of an individual’s environment might predict the decision to use
marijuana but not the decision to diet, exercise, and lead an active lifestyle? Demographic
characteristics of a neighborhood would pose a risk for correlation with unobserved
environmental factors that impact BMI. Variables that directly affect the availability or demand
for marijuana may be promising instruments (25). Evidence suggests that marijuana and alcohol
use are compliments (70). Therefore, policies associated with variations in alcohol use may, in
67
turn, result in variations in marijuana use. However, alcohol use itself is associated with BMI and
possibly endogenous in these same models.
Geographic variation in marijuana policy and support for marijuana may provide favorable
instruments. Numerous national polls capture the variation in support for medical marijuana and
marijuana legalization (71). Individuals who live in states that have medical marijuana
legislation are more likely to use marijuana and have a lower perception of risk regarding the
drug. Some studies have also found that people in states with medical marijuana legislation have
higher usage and dependency rates, which may be related to the perception of societal norms and
its influence on behavior.
As long as sequential endogeneity can be justified, the lagged values of endogenous variables are
good instrument candidates (72). We will explore a panel model that takes advantage of lagged
instruments.
Basic Statistics
In light of the methodological concerns raised by previous econometric models for the
relationship between marijuana use and factors of metabolic health, a specific aim of this paper is
to determine the “best” models for estimating these relationships. This section describes
alternative models, justifies the choice of a “best” model for each indicator of metabolic health,
and details the statistical plan. All estimates account for complex survey design, including
observation weights and clustering. All results with values of p < 0.05 were considered
significant and no adjustments were made for multiple comparisons due to the exploratory nature
of this study.
68
The eligible sample differs for each set of cross-sectional and panel models based on the number
of observations in each included wave and the independent variables included in each model.
Means of characteristics of individuals in each set of analyses were compared with means from
excluded respondents to establish any patterns in the missing data. A large number of
observations were missing for parental income, parental education level, and maternal obesity
status. Missing income was imputed using a linear prediction. Parent education and maternal
obesity status could not be imputed with a linear prediction since the data does not include parent
characteristics to use as regressors. These are categorical variables, so in order to avoid losing
observations, a category was added for missing observations.
Cross-sectional Analyses
Cross-sectional ordinary least squares (OLS) models, which have been used previously to
explore this relationship, do not account for potential bias due to unobserved individual
heterogeneity (6, 12, 13, 15, 16, 18, 19). The results from cross-sectional linear models estimated
using OLS are presented for the sake of comparison with panel and other models that address
unobserved heterogeneity. In this way, OLS estimates will provide evidence of the postulated
endogeneity bias. Because BMI is also available in Waves I and III, OLS results for models of
BMI using Waves I and III are also presented. Cross-sectional models control for individual and
household level variables. Wave IV models control for age, age squared, gender, race, maternal
obesity, individual income, individual education, Wave I parental education, Wave I household
income, marital status, physical activity level, smoking status, and alcohol consumption. Wave
III analyses control for age, age squared, gender, race, maternal obesity, Wave I parental
education, Wave I household income, physical activity level, smoking status, and alcohol
consumption. Wave I analyses control for age, age squared, gender, race, maternal obesity,
69
parental education, household income, physical activity level, smoking status, and alcohol
consumption. The obesity status of each individual’s biological mother, reported by parents in
Wave I, was included as a control because maternal obesity status is a known risk factor for
adolescent obesity (73). An additional set of wave IV OLS models is also estimated using wave
III BMI levels as a control. Inclusion of this control may take care of some selection bias, since
healthier (lower BMI) individuals may have a higher probability of selecting to use marijuana,
leading to selection bias. The theoretical model is presented below:
!
"
=$%
"
+ ()
"
+ *
"
(1)
Where !
"
are all the dependent variables, described above; %
"
are the control variables, which
differ for models in wave I, wave III, and wave IV; and )
"
is contemporaneous self-reported
marijuana use, which can be defined as either past month use or eight or more days in the past
month.
Most metabolic health variables were available only in Wave IV, eliminating the possibility of
panel models for those outcome variables. As a result, cross-sectional instrumental variables
models estimated using two-stage least squares (2SLS) were employed to estimate the effects on
waist circumference, glucose level, HbA1c, triglycerides, HDL-C, and LDL-C. Instrumental
variables models can overcome the issue of bias due to endogeneity if instruments are strong and
good; if not, estimates will also be biased. The theoretical model is:
!
"
=$%
"
+ ()
"
+ *
"
(2)
Where !
"
are dependent variables for which data are only available in Wave IV, %
"
are Wave IV
exogenous control variables, and marijuana use, )
"
=+,
"
+ -%
"
+ .
"
, is estimated using ,
"
, a
vector of exogenous instruments for marijuana use. *
"
and .
"
are assumed to be asymptotically
70
uncorrelated. The selection of instruments for marijuana use that are both theoretically and
statistically sound is challenging, and weak instruments can cause inference problems of their
own (33).
Panel Models
An alternative to IV models for addressing unobserved individual heterogeneity is to take
advantage of longitudinal panel data. Since BMI is available in multiple panels, variation over
time can be used to estimate model parameters in panel models. The basic model is given below.
/01
"2
=) ′
"2
4
)
+5′
"2
4
5
+ 6
"2
(3)
In this model, /01
"2
represents the BMI of individual i during wave year t, where t=1, 2, 3
corresponds to Wave I, Wave III, or Wave IV, respectively; )
"2
is contemporaneous marijuana
use; 5
"2
are individual and household time-varying exogenous control variables (age, age
squared, individual income, individual education, marital status, physical activity level, smoking
status, and alcohol consumption); and 6
"2
is an error term, which includes individual fixed
effects, 7
"
, and an idiosyncratic error, .
"2
. There are some potential econometric issues with (5).
First, the right-hand side variables may be correlated with the fixed effects, 7
"
, in the error term.
Transforming the equation by differencing or demeaning the model, as in first-differenced (FD)
and fixed effects panel models (FE), will address this problem by eliminating the individual
unobserved fixed effects. Second, the right-hand side variables could also be endogenous due to
correlation with the idiosyncratic errors, .
"2
. Transforming the model does not take care of this
type of endogeneity. This is certainly true if the model includes lagged values of the dependent
variable as a control, because, once transformed, predetermined regressors are not strictly
71
exogenous. To address the latter issues, panel models that correct for endogeneity and selection
bias should be considered.
Panel FE
The first step in addressing the econometric issues outlined above is to transform the model (5)
and eliminate individual fixed effects. We strongly suspect that unobserved individual
characteristics may be responsible for the endogeneity bias observed in previous research (6, 7,
74). For the sake of comparison with dynamic panel models, which will be described later, and
since many equation variables exist only in waves III and IV, the panel fixed effects (FE) model
of the impact of marijuana use on BMI is given for T=2 time periods (below). With two time
periods, this model is equivalent to the first-differenced model, which uses first-differences to
eliminate individual fixed effects term. In the panel models, we control for all individual and
household level variables that vary over time.
First-difference transform of (5) eliminates fixed effect 7
"
:
/01
"2
−/01
",2;<
=()
"2
−)
",2;<
)′4
)
+ (5
"2
−5
",2;<
)′4
5
+ (?
"2
−?
",2;<
)(@)
Dynamic Panel Model: Difference-GMM
The model described above does not address all of the potential sources of endogeneity. Thus,
we propose a method that will be superior to both panel fixed effects and cross-sectional
instrumental variables for handling endogeneity and selection bias in the model. Difference
generalized method-of-moments (difference-GMM) estimation first transforms the data to
remove fixed effects, and then it instruments endogenous (e.g. )
"2
) and predetermined variables
(e.g. /01
",2;<
) with their own lags and with other strictly exogenous instruments.
72
There are a few significant benefits to using the difference-GMM model framework in this study.
First, difference-GMM allows for estimation of a dynamic panel model, which controls for
lagged values of the dependent variable. It is reasonable to assume that past levels of BMI will
significantly predict current levels. Additionally, the method allows for both contemporaneous
instruments that are correlated with the endogenous variables and orthogonal to .
"2
to be entered
into the instrument matrix as traditional, IV-style instruments, as well as lagged values of the
instrumented variables to enter into the instrument matrix as GMM-style instruments. The
difference-GMM estimator, which originated in a paper by Arellano and Bond (75), is described
in greater detail in Wooldridge Chapter 11.6 (76). The theoretical model is presented below.
∆/01
"2
=B∆/01
",2;<
+ ∆) ′
"2
4
)
+∆5′
"2
4
5
+ ∆?
"2
(C)
Where ∆/01
"2
is the differenced dependent variable, which equals the BMI of individual i
during wave year t-1 subtracted from the BMI of individual i during wave year t, where t=1, 2, 3
corresponds to Wave I, Wave III, and Wave IV, respectively. ∆/01
",2;<
is the differenced
lagged dependent variable, which is considered a “predetermined” variable and is therefore not
strictly exogenous. ∆)
"2
is the differenced endogenous marijuana use variable, and ∆5
"2
includes
the differenced explanatory variables. ∆?
"2
: is the transformed idiosyncratic error term; the
unobserved individual effects have been eliminated from the model.
Results
Cross-Sectional analyses
See Appendix A. for tables with characteristics of marijuana users and univariate statistics. In
Wave IV, 50.63% of respondents reported never using marijuana, 29.64% reported having used,
but not in the past 30 days, 8.33% reported using less than eight days in the past 30, and 11.41%
73
reported using eight or more days. Those proportions were 53.69%, 23.23%, 11.63%, and
11.45%, respectively, in Wave III and 72.35%, 13.21%, 9.30%, and 5.15%, respectively, in
Wave I.
The OLS results for Wave IV dependent variables mirror estimates reported in previous studies;
recent marijuana use (any amount in the past month) is associated with lower BMI (standardized
BMI score) and smaller waist circumference. Triglycerides were lower for individuals who
reported using more than eight times in the past month. Wave IV OLS results are presented in
Table 1.

In OLS results for Wave III BMI, only the category of marijuana users with the highest reported
levels of use have significantly lower standardized BMI scores (Table 2).

Table 1. Wave IV OLS (comparison: non-users)
VARIABLES BMI
Waist
Circumference Glucose HBA1C Triglycerides
HDL-
C
LDL-
C

Used marijuana,
not in past month -0.051 -0.852 -0.451 -0.040* -0.018 0.104 -0.063
(0.038) (0.652) (1.097) (0.022) (0.101) (0.114) (0.111)
Past month use,
less than 8 times -0.165*** -3.111*** -2.581 -0.040 -0.105 0.066 0.110
(0.050) (0.873) (1.781) (0.042) (0.148) (0.192) (0.180)
8 or more times
in past month -0.174*** -2.448*** -2.047 -0.014 -0.413*** -0.057 0.010
(0.047) (0.841) (1.431) (0.033) (0.148) (0.151) (0.185)
Observations 9230 9280 8413 8624 8185 8245 7879
R-squared 0.114 0.115 0.023 0.070 0.087 0.067 0.011
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

74

Table 2. Wave III OLS (comparison: non-
users)
VARIABLES BMI

Used marijuana, not in past month 0.010
(0.037)
Past month use, less than 8 times -0.071*
(0.042)
8 or more times in past month -0.141***
(0.047)
Observations 8,536
R-squared 0.088
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

Similar to Wave III results, using marijuana eight or more times in the past month is associated
with lower standardized BMI in Wave I (Table 3).
Table 3. Wave I OLS (comparison: non-users)
VARIABLES BMI

Used marijuana, not in past month -0.015
(0.035)
Past month use, less than 8 times -0.030
(0.050)
8 or more times in past month -0.194***
(0.054)
Observations 11,374
R-squared 0.075
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

In an additional analysis, we took advantage of the availability of BMI in multiple waves to
control for lagged BMI in the a cross sectional model in an efforst to isolate the independent
contribution of marijuana use on current BMI. This was done in the full Wave IV sample as well
as a subsample restricted to non-users in Wave III in order to rule out the possibility of past use
75
corellation with past BMI. The results are displayed in Table 4 below.

To support the use of instrumental variables models, a series of tests were performed to
determine the adequacy of the instruments and the endogeneity of marijuana use in the models
for Wave IV dependent variables. A few potentially suitable instruments were identified from the
data: political beliefs, risk-taking behavior, and lagged marijuana use. Initially, we might be
concerned the lagged marijuana use is not strictly exogenous, which is a requirement for
instrumental variables. First, we tested the association between each potential instrument and
variables from the models that might proxy health attitudes or behaviors (which, in turn, affect
the dependent variables). These correlations might suggest that the instruments are not strictly
exogenous. The following significant correlations may be of concern: political beliefs with
marital status, alcohol use, and tobacco use; risk-taking with marital status, alcohol use, tobacco
use, and physical activity level; and lagged marijuana use with marital status, alcohol use, and
tobacco use. The F-test was performed for instruments and combinations of instruments. A
cutoff value of 10 for the F-statistic has been suggested to determine whether instruments/sets of
instruments are significant (34). The F-statistic was above the threshold for the following sets:
Table 4. Wave IV OLS, base model and model with past BMI control
VARIABLES
Base, no
lagged BMI
Controlled
for Wave III
BMI; Full
Sample
Controlled
for Wave III
BMI; Non-
users in
Wave III

Used marijuana, not in past month -0.051 -0.024 -0.017
(0.038) (0.022) (0.024)
Past month use, less than 8 times -0.165*** -0.017 -0.035
(0.050) (0.038) (0.070)
8 or more times in past month -0.174*** -0.067** -0.121**
(0.047) (0.030) (0.057)
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
76
political beliefs (on its own); political beliefs and risk-taking; political beliefs, risk-taking, and
lagged marijuana use; lagged marijuana use (24.89, 19.73, 96.27, and 209.13, respectively). The
Durbin-Wu-Hausman test, or augmented regression test, is used to judge whether or not there is
evidence of endogeneity in the instrumented variable. The DWH test reveals evidence of
endogeneity in the models for BMI, waist circumference, and glucose. The lagged user
instrument on its own does not detect endogeneity. Below are the results of 2SLS with
instrumental variables for different instrument combinations (Tables 5-7).
Table 5. Instruments: Political beliefs, risk-
taking*
VARIABLES BMI
Waist
Circumferenc
e Glucose
HBA1
C
Triglycerid
es HDL-C LDL-C
Used but not in
past month
1.088 14.948 8.098 0.458 1.132 2.092 0.014
(1.328) (15.556) (35.760) (0.577) (2.401) (3.641) (1.957)
Past month use,
less than 8 times
-2.861 -22.695 -81.220 -1.268 -3.349 -9.722 -2.174
(4.870) (48.978) (102.141) (1.986) (8.617) (11.757) (7.127)
8 or more times
in past month
2.464 27.159 20.155 -0.037 4.011 5.531 -0.447
(2.484) (25.316) (52.052) (1.009) (5.023) (6.251) (3.392)
Observations 8992 9040 8174 8390 7947 8007 7655

Table 6. Instruments: Political beliefs, risk-taking, lagged marijuana
use*
VARIABLES BMI
Waist
Circumferenc
e Glucose
HBA1
C
Triglycerid
es HDL-C LDL-C
Used but not in
past month
0.056 -2.264 -1.917 -0.095 -0.090 0.013 -0.131
(0.136) (2.307) (3.972) (0.077) (0.363) (0.408) (0.373)
Past month use,
less than 8 times
-0.336 8.784 -7.403 0.067 0.068 2.964 -2.499
(0.638) (11.495) (29.198) (0.515) (2.308) (2.340) (2.525)
8 or more times
in past month
-0.032 -4.965 4.829 0.025 -0.489 -0.875 0.873
(0.280) (4.844) (12.677) (0.224) (0.925) (0.965) (1.064)
Observations 7449 7489 6846 7024 6653 6705 6411

77

Table 7. Instrument: Political
beliefs*
VARIABLES BMI
Waist
Circumferenc
e Glucose HBA1C
Triglycerid
es HDL-C LDL-C
Used but not in
past month
1.970 23.124 -6.003 0.612 0.721 0.995 2.916
(4.190) (45.087) (35.949) (2.402) (3.903) (4.417) (10.176)
Past month use,
less than 8 times
-11.504 -124.437 -61.122 -5.802 -5.189 -7.308 -31.426
(21.927) (230.854) (140.732) (25.660) (19.890) (24.627) (53.134)
8 or more times
in past month
7.268 88.121 25.270 3.070 6.062 5.512 17.156
(12.065) (134.970) (81.999) (16.543) (12.677) (14.237) (32.827)
Observations 8683 8730 7904 8110 7686 7743 7406

*results from regressions without potentially endogenous covariates (e.g. alcohol, tobacco) do not change

The estimates from all instrumental variables models are non-significant. Standard errors are
anticipated to increase with 2SLS estimation; however, some parameter values seem
unreasonable. It is possible that bad instruments have led to bias in this estimation method.
Instruments based on policy or geography would be less likely to cause problems due to
correlation with the error; however no such variables were available in this data.
Appendix B breaks down these results by gender.
Panel FE model
When panel data is available, fixed effects can be removed by transforming the variables in a
model, thereby eliminating any correlation between predictor variables and unobserved, time-
invariant individual heterogeneity. Panel FE estimates of the effects of smoking eight or more
times in the past month are similar to OLS estimates (Table 8). Estimates for the effects of using
less than eight times become non-significant in panel models, suggesting a bias caused by
unobserved fixed effects in the cross-sectional OLS models. However, bias caused by correlation
78
with the idiosyncratic error may still be a concern. Leaving out potentially endogenous
covariates, such as alcohol use, smoking, physical activity, does not alter results (second column
of results).
Table 8. Panel
FE
VARIABLES BMI

Past user -0.016
(0.026)
Past month use, less than 8
times -0.011
(0.037)
8 or more times in past month -0.094**
(0.036)

Observations 12,116
R-squared 0.021
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

Difference-GMM
Fixed effects with instrumental variables could take care of the issue of endogeneity, but with
weak instruments, the fixed-effects IV estimators are likely to be biased in the way of the OLS
estimators. Diff-GMM allows for lags of the endogenous variables to instrument for themselves
in the model. These models may also include a lagged dependent variable. Estimating dynamic
panel data models the usual way (as in FE) leads to biased and inconsistent estimates.
Below are the results of the model. The exogenous “political beliefs” instrumental variable is
included. Parameter estimates for marijuana use are not significant in this model. Instruments
pass the Sargan test, indicating that the instruments, as a group, are exogenous.
79
For comparison, the panel FE model from the previous section was performed on the sample
from diff-GMM estimation. Estimates from this model are presented in the last column of Table
9.
Table 9. Diff GMM
VARIABLES BMI
BMI Panel
FE, restricted
sample⁺

Past user -0.028 -0.023
(1.227) (0.028)
Past month use, less than 8
times 0.549 -0.004
(1.575) (0.039)
8 or more times in past month -1.069 -0.093**
(1.603) (0.038)
Lagged BMI 0.226
(0.128)

Observations 6,841 6,841
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
⁺restricted to the sample used in Difference-GMM regression

Sensitivity analysis
Subgroup analysis was performed based on gender and age, as those variables may modify
estimates of the effects of marijuana on BMI and other factors of metabolic health (7). Women
and men differ in drug consumption patterns as well as body composition (53, 55). Additionally,
because each wave contains a range of ages, we might be concerned that the trajectory of effects
from one wave to the next may not be congruent across each starting value of age. A subgroup
analysis by age at Wave I in the panel models will examine this possibility. For cross-sectional
and panel models, age does not appear to modify results, except that many results appear non-
80
significant, most likely due to loss of observations. Separate models for males and females are
presented in Appendix B. The significant OLS estimates for Waves I, III, and IV are either
isolated to or larger in magnitude in the male subgroup compared to the female subgroup. The
results of IV analyses remain non-significant in both female and male subgroups. Panel FE
estimates are only significant in the male subgroup. Diff-GMM estimates remain non-significant
when stratified by gender. These results indicate
The rate of other illicit drug use was very low in this sample; however, we added indicator
variables for non-prescribed stimulants, non-prescribed pain killers, cocaine, meth, other illegal
drugs, and injected illegal drugs in the models to determine how controlling for other illicit drug
use changes the effect estimates. Controlling for other illicit drug use reduces sample size due to
the higher rate of non-response for these items. However, controlling for other drug use does not
impact estimates, possibly due to the very small number of individuals with other reported illicit
drug use.
Angrist and Krueger argue against including endogenous variables as controls for other
endogenous variables (77). To test the impact of the potentially endogenous control variables
tobacco use, alcohol use, and physical activity, we compare estimates from all models with and
without controlling for potentially endogenous control variables. This exercise does not affect
the significance or interpretation of FE or diff-GMM estimates. However, effect estimates for
marijuana use from OLS appear to increase in magnitude when endogenous control variables are
excluded; presumably due to the correlation between the effects of marijuana use and the effects
of other excluded health behaviors leading to a greater endogeneity bias.

81
Discussion
This study aims to augment the understanding of the relationship between marijuana use and
metabolic health by employing cross-sectional and panel models that address the endogeneity of
marijuana use. Cross-sectional OLS models provide a baseline for comparison with these models
to illustrate the bias that is introduced using more common estimation techniques, such as OLS.
As expected, results from cross-sectional multivariate linear models estimated using OLS
suggest that marijuana users have lower BMI, waist circumference, and triglyceride levels. The
purpose of controlling for Wave III BMI in the Wave IV BMI cross-sectional model was to
attempt to rule out the influence of past BMI on current BMI. When past BMI is controlled for,
high marijuana use is still associated with lower BMI and waist circumference, but low
marijuana use becomes non-significant. This suggests that high marijuana use might be an
independent contributor to decreases in bodyweight, but there might also be some selection bias
in the original estimates.
For the dependent variables available only in the fourth wave of data (waist circumference,
glucose, HBA1C, triglycerides, HDL-C, and LDL-C), IV models estimated using 2SLS might be
inadequate to model the effects of marijuana use due to weak or invalid available instruments for
marijuana use. However, when the only dependent variable available in multiple waves, BMI, is
modeled using panel FE techniques, which removes unobserved individual fixed effects that may
lead to endogeneity, the estimate for low recent use loses significance. When the diff-GMM
technique is employed, which further addresses endogeneity caused by time-varying unobserved
effects, the estimated effects of all levels of marijuana use become non-significant. It is
important to note that the standard errors for diff-GMM estimation are larger than with other
estimation techniques. The diff-GMM model controls for lagged BMI, unlike the baseline OLS
82
results and Panel FE. However, Diff-GMM results remain non-significant when the lagged
dependent variable is removed from the model. Another potential reason for the difference
between panel FE and diff-GMM results is the decreased sample size in diff-GMM estimation.
As a sensitivity analysis, the panel-FE model was estimated using the same sample as the diff-
GMM model, but this did not meaningfully alter FE estimates.
These results suggest that estimates of the effect of marijuana use on BMI (and possibly other
factors of metabolic health) may be spurious due to the endogeneity of marijuana use.
There were a few limitations to this research. First, panel FE and diff-GMM models rely on
within-individual variation in key variables for estimation. A large portion of the sample may be
lost due to lack of within-individual variation in patterns of marijuana use, which leads to
imprecise estimates. Additionally, since the diff-GMM method also involves exogenous
contemporaneous instruments, the estimates are subject to the same problems as cross-sectional
IV estimation: using weak instruments risks producing estimates that are unreliable and biased
toward the original estimate.
There are also some drawbacks in using the Add Health data. Although the data provided a
longitudinal panel of individuals with rich behavioral information, some key variables were only
available in certain waves, such as waist circumference, limiting the use of panel estimation
techniques for such variables. In addition, substance use was self-reported, which may lead to
measurement error. Furthermore, this dataset lacked great candidates for instrumental variables,
such as state variation in medical marijuana policy.
These findings alone don’t indicate that there is no meaningful relationship between marijuana
use and weight gain or other metabolic risk factors. Indeed, research on the endocannabinoid
83
system in humans provides strong evidence that the compounds in marijuana directly impact
appetite and peripheral systems involved in metabolic health (5). In a separate forthcoming
paper, Thompson and Hay argue that this relationship may look different depending on the level
of marijuana use. Further research is needed to explore how marijuana use impacts weight gain
and metabolic health at varying durations and quantities of marijuana consumption.

84
References
1. Results from the 2013 National Survey on Drug Use and Health: Summary of National
Findings Rockville, MD: Substance Abuse and Mental Health Services Administration,
2014, ( NSDUH Series H-48 no. Publication No. (SMA) 14-4863)(Administration
SAaMHS
2. A look at U.S. marijuana laws. CNN.
(http://www.cnn.com/interactive/2014/01/politics/map-marijuana/). (Accessed 8/7/15).
3. Kumar RN, Chambers WA, Pertwee RG. Pharmacological actions and therapeutic uses of
cannabis and cannabinoids. Anaesthesia 2001;56(11):1059-68.
4. Gorter R. Cancer cachexia and cannabinoids. Forschende
Komplementärmedizin/Research in Complementary Medicine 2004;6(Suppl. 3):21-2.
5. Pertwee RG. Cannabis and cannabinoids: Pharmacology and rationale for clinical use.
Forsch Komplementmed 1999;6(Suppl. 3):12-5.
6. Thompson CA, Hay JW. Estimating the association between metabolic risk factors and
marijuana use in US adults using data from the continuous National Health and Nutrition
Examination Survey. Annals of epidemiology 2015.
7. Thompson C, Hay JW. PSY12 - Role of behavioral confounding in the association
between Marijuana use and BMI in us Adults [Abstract]. Presented at Value in Health:
ISPOR 20th Annual International Meeting, Philadelphia, PA2015.
8. Substance Abuse and Mental Health Services Administration, Results from the 2010
National Survey on Drug Use and Health: Summary of National Findings, NSDUH
Series H-41, HHS Publication No. (SMA) 11-4658. Rockville, MD: Substance Abuse
and Mental Health Services Administration, 2011.
9. Alberti KG, Zimmet P, Shaw J, et al. The metabolic syndrome--a new worldwide
definition. Lancet 2005;366(9491):1059-62.
10. Kassi E, Pervanidou P, Kaltsas G, et al. Metabolic syndrome: definitions and
controversies. BMC Med 2011;9:48.
11. Foltin RW, Fischman MW, Byrne MF. Effects of smoked marijuana on food intake and
body weight of humans living in a residential laboratory. Appetite 1988;11(1):1-14.
12. Smit E, Crespo CJ. Dietary intake and nutritional status of US adult marijuana users:
results from the Third National Health and Nutrition Examination Survey. Public Health
Nutr 2001;4(3):781-6.
13. Rodondi N, Pletcher MJ, Liu K, et al. Marijuana use, diet, body mass index, and
cardiovascular risk factors (from the CARDIA study). Am J Cardiol 2006;98(4):478-84.
14. Muniyappa R, Sable S, Ouwerkerk R, et al. Metabolic effects of chronic cannabis
smoking. Diabetes Care 2013;36(8):2415-22.
15. Le Strat Y, Le Foll B. Obesity and cannabis use: results from 2 representative national
surveys. Am J Epidemiol 2011;174(8):929-33.
16. Hayatbakhsh MR, O'Callaghan MJ, Mamun AA, et al. Cannabis use and obesity and
young adults. Am J Drug Alcohol Abuse 2010;36(6):350-6.
17. Huang DY, Lanza HI, Anglin MD. Association between adolescent substance use and
obesity in young adulthood: a group-based dual trajectory analysis. Addict Behav
2013;38(11):2653-60.
18. Farhat T, Iannotti RJ, Simons-Morton BG. Overweight, obesity, youth, and health-risk
behaviors. Am J Prev Med 2010;38(3):258-67.
85
19. Penner EA, Buettner H, Mittleman MA. The impact of marijuana use on glucose, insulin,
and insulin resistance among US adults. Am J Med 2013;126(7):583-9.
20. Angrist JD, Pischke J-S. Mostly harmless econometrics: an empiricist's companion.
Princeton: Princeton University Press; 2009.
21. Patel K, Hargreaves MK, Liu JG, et al. Relationship Between Smoking and Obesity
Among Women. American journal of health behavior 2011;35(5):627-36.
22. Clair C, Chiolero A, Faeh D, et al. Dose-dependent positive association between cigarette
smoking, abdominal obesity and body fat: cross-sectional data from a population-based
survey. BMC PUBLIC HEALTH 2011;11(1):23.
23. Matsushita Y, Nakagawa T, Yamamoto S, et al. Associations of smoking cessation with
visceral fat area and prevalence of metabolic syndrome in men: the Hitachi health study.
Obesity (Silver Spring) 2011;19(3):647-51.
24. Rigotti NA, Lee JE, Wechsler H. US college students' use of tobacco products - Results
of a national survey. Jama-J Am Med Assoc 2000;284(6):699-705.
25. French MT, Popovici I. That instrument is lousy! In search of agreement when using
instrumental variables estimation in substance use research. Health Econ 2011;20(2):127-
46.
26. Centers for Disease Control and Prevention. National Health and Nutrition Examination
Survey: Laboratory Procedures Manual,
http://www.cdc.gov/nchs/data/nhanes/nhanes_11_12/2011-
12_Laboratory_Procedures_Manual.pdf; 2011 [accessed on 10.16.14]. [electronic
article].
27. Park YW, Zhu S, Palaniappan L, et al. The metabolic syndrome: prevalence and
associated risk factor findings in the US population from the Third National Health and
Nutrition Examination Survey, 1988-1994. Arch Intern Med 2003;163(4):427-36.
28. Reiter JP, Raghunathan TE, Kinney SK. The importance of modeling the sampling
design in multiple imputation for missing data. Survey methodology 2006;32(2):143-50.
29. Little RJA, Rubin DB. The Analysis of Social-Science Data with Missing Values.
Sociological Methods & Research 1989;18(2-3):292-326.
30. When and how to construct weights when combining survey cycles [Internet].
Hyattsville, MD: National Center for Health Statistics.
(http://www.cdc.gov/nchs/tutorials/nhanes/surveydesign/Weighting/intro_i.htm).
(Accessed 17 September 2013).
31. Willett WC. Will high-carbohydrate/low-fat diets reduce the risk of coronary heart
disease? Proc Soc Exp Biol Med 2000;225(3):187-90.
32. Wirfalt E, Hedblad B, Gullberg B, et al. Food Patterns and Components of the Metabolic
Syndrome in Men and Women: A Cross-sectional Study within the Malmo Diet and
Cancer Cohort. American journal of epidemiology 2001;154(12):1150-9.
33. Bound J, Jaeger DA, Baker R. The Cure Can Be Worse than the Disease: A Cautionary
Tale Regarding Instrumental Variables. NBER Working Paper Series 1993;t0137(Journal
Article).
34. Stock JH, Wright JH, Yogo M. A Survey of Weak Instruments and Weak Identification
in Generalized Method of Moments. Journal of Business & Economic Statistics,
2002:518-29.
35. Davidson R, MacKinnon JG. Estimation and inference in econometrics. New York:
Oxford University Press; 1993.
86
36. Rajavashisth TB, Shaheen M, Norris KC, et al. Decreased prevalence of diabetes in
marijuana users: cross-sectional data from the National Health and Nutrition Examination
Survey (NHANES) III. BMJ Open 2012;2:e000494.
37. National Institute on Alcohol Abuse and Alcoholism. Beyond hangovers: Understanding
alcohol's impact on your health . NIH; 2012. (NIH Publication No. 10–7604).
38. American Heart Association. "Heavy, Lifetime Alcohol Users May Be Toasting
Metabolic Syndrome." ScienceDaily.
www.sciencedaily.com/releases/2004/11/041116233400.htm (accessed August 21, 2014).
(Accessed).
39. Foster GD, Wyatt HR, Hill JO, et al. A randomized trial of a low-carbohydrate diet for
obesity. The New England journal of medicine 2003;348(21):2082-90.
40. Volek JS, Phinney SD, Forsythe CE, et al. Carbohydrate Restriction has a More
Favorable Impact on the Metabolic Syndrome than a Low Fat Diet. Lipids
2009;44(4):297-309.
41. Plasse TF, Gorter RW, Krasnow SH, et al. Recent clinical experience with dronabinol.
Pharmacology, biochemistry, and behavior 1991;40(3):695-700.
42. Beal JE, Olson R, Laubenstein L, et al. Dronabinol as a treatment for anorexia associated
with weight loss in patients with AIDS. Journal of pain and symptom management
1995;10(2):89-97.
43. Nelson K, Walsh D, Deeter P, et al. A phase II study of delta-9-tetrahydrocannabinol for
appetite stimulation in cancer-associated anorexia. Journal of palliative care
1994;10(1):14-8.
44. American Diabetes Association. Economic costs of diabetes in the U.S. in 2012. Diabetes
Care 2013;36(4):1033-46.
45. Foltin RW, Brady JV, Fischman MW. Behavioral analysis of marijuana effects on food
intake in humans. Pharmacology, biochemistry, and behavior 1986;25(3):577-82.
46. Alberti KG, Zimmet P, Shaw J. The metabolic syndrome--a new worldwide definition.
Lancet 2005;366(9491):1059-62.
47. Whiting PF, Wolff RF, Deshpande S, et al. Cannabinoids for Medical Use: A Systematic
Review and Meta-analysis. JAMA 2015;313(24):2456-73.
48. Ngueta G, Belanger RE, Laouan-Sidi EA, et al. Cannabis use in relation to obesity and
insulin resistance in the inuit population. Obesity (Silver Spring) 2015;23(2):290-5.
49. Light MK, Orens A, Lewandowski B, et al. Market Size and Demand for Marijuana in
Colorado. The Marijuana Policy Group., 2014, (Revenue CDo
50. Baron RM, Kenny DA. The moderator-mediator variable distinction in social
psychological research: conceptual, strategic, and statistical considerations. J Pers Soc
Psychol 1986;51(6):1173-82.
51. Materials and methods are available as supplementary materials on Science Online.
52. Zhao J, Stockwell T, MacDonald S. Non-response bias in alcohol and drug population
surveys. Drug and Alcohol Review 2009;28(6):648-57.
53. Kandel D, Chen K, Warner LA, et al. Prevalence and demographic correlates of
symptoms of last year dependence on alcohol, nicotine, marijuana and cocaine in the U.S.
population. Drug Alcohol Depend 1997;44(1):11-29.
54. Coffey C, Carlin JB, Lynskey M, et al. Adolescent precursors of cannabis dependence:
findings from the Victorian Adolescent Health Cohort Study. The British journal of
psychiatry : the journal of mental science 2003;182:330-6.
87
55. Warner LA, Kessler RC, Hughes M, et al. Prevalence and correlates of drug use and
dependence in the United States. Results from the National Comorbidity Survey.
Archives of general psychiatry 1995;52(3):219-29.
56. Carroll J. Who Supports Marijuana Legalization? Gallup Poll Tuesday Briefing 2005.
57. The Pew Research Center for People and the Press. In Debate Over Legalizing
Marijuana, Disagreement Over Drug’s Dangers. 2015; Available from
http://www.people-press.org/2015/04/14/in-debate-over-legalizing-marijuana-
disagreement-over-drugs-dangers/2/.
58. Chen CY, O'Brien MS, Anthony JC. Who becomes cannabis dependent soon after onset
of use? Epidemiological evidence from the United States: 2000-2001. Drug Alcohol
Depend 2005;79(1):11-22.
59. Behrendt S, Wittchen HU, Hofler M, et al. Transitions from first substance use to
substance use disorders in adolescence: is early onset associated with a rapid escalation?
Drug Alcohol Depend 2009;99(1-3):68-78.
60. Budney AJ, Moore BA, Vandrey RG, et al. The time course and significance of cannabis
withdrawal. J Abnorm Psychol 2003;112(3):393-402.
61. Budney AJ, Hughes JR, Moore BA, et al. Review of the validity and significance of
cannabis withdrawal syndrome. Am J Psychiatry 2004;161(11):1967-77.
62. Savitz DA. Interpreting epidemiologic evidence: strategies for study design and analysis.
New York: Oxford University Press; 2003.
63. Budney AJ, Roffman R, Stephens RS, et al. Marijuana dependence and its treatment.
Addiction science & clinical practice 2007;4(1):4-16.
64. Greene MC, Kelly JF. The prevalence of cannabis withdrawal and its influence on
adolescents' treatment response and outcomes: a 12-month prospective investigation. J
Addict Med 2014;8(5):359-67.
65. Alpert JS. Marijuana for diabetic control. Am J Med 2013;126(7):557-8.
66. Szalavitz M. Marijuana: The Next Diabetes Drug. Time Magazine 2013.
67. When and how to construct weights when combining survey cycles [Internet].
Hyattsville, MD, National Center for Health Statistics (2010; Available from
http://www.cdc.gov/nchs/tutorials/Nhanes/SurveyDesign/Weighting/Task2.htm).
68. Harris KM. The add health study: Design and accomplishments. Chapel Hill, NC 2013.
69. Whitaker RC, Wright JA, Pepe MS, et al. Predicting obesity in young adulthood from
childhood and parental obesity. The New England journal of medicine 1997;337(13):869-
73.
70. Haberstick BC, Young SE, Zeiger JS, et al. Prevalence and correlates of alcohol and
cannabis use disorders in the United States: results from the national longitudinal study of
adolescent health. Drug Alcohol Depend 2014;136:158-61.
71. Votes and Polls, 2000-Present. procon.org; 2015.
(http://medicalmarijuana.procon.org/view.additional-resource.php?resourceID=149).
(Accessed August 27 2015).
72. Roodman D. How to do xtabond2: An introduction to difference and system GMM in
Stata. Center for Global Development working paper 2006(103).
73. Gordon-Larsen P, Adair LS, Suchindran CM. Maternal obesity is associated with
younger age at obesity onset in U.S. adolescent offspring followed into adulthood.
Obesity (Silver Spring) 2007;15(11):2790-6.
88
74. Thompson CA, Hay JW. Evaluating the impact of cannabis use on metabolic syndrome
using data from the continuous national health and nutrition examination survey. Value in
health : the journal of the International Society for Pharmacoeconomics and Outcomes
Research 2014;17(3):A210.
75. Arellano M, Bond S. Some Tests of Specification for Panel Data: Monte Carlo Evidence
and an Application to Employment Equations. The Review of Economic Studies
1991;58(2):277-97.
76. Wooldridge JM. Econometric analysis of cross section and panel data. Cambridge,
Mass.: MIT Press; 2002.
77. Angrist J, Pischke J-S. Multiple endogenous variables – now what?! ; 2010.
(http://www.mostlyharmlesseconometrics.com/2010/02/multiple-endogenous-variables-
what-now/). (Accessed August 27 2015).

89
Appendix A. Univariate Statistics
Wave IV Univariate

Never used
Used but not in
the past month
Past month use,
less than 8 times
8 or more times
in past month Total
Variables
Col
% 95% CI
Col
% 95% CI
Col
% 95% CI
Col
% 95% CI
Col
% 95% CI
F-Test P-
value
Gender P = 0.000
Female (n=6,800) 54.7 [53.2,56.2] 46.8 [44.4,49.2] 41.2 [36.9,45.6] 33.5 [30.3,36.8] 48.8 [47.5,50.1]
Male (n=6,049) 45.3 [43.8,46.8] 53.2 [50.8,55.6] 58.8 [54.4,63.1] 66.5 [63.2,69.7] 51.2 [49.9,52.5]
Total (n=12,849) 100 100 100 100 100
Race/ethnicity P = 0.000
Hispanic (n=2,066) 14.3 [10.1,19.8] 10.4 [7.9,13.5] 10 [7.3,13.5] 9.6 [6.8,13.5] 12.2 [9.1,16.2]
Non-Hispanic
white
(n=6,764)
58.7 [51.7,65.4] 73.3 [68.0,78.0] 70 [64.2,75.2] 67.6 [60.9,73.7] 65 [59.0,70.5]
Black (n=2,713) 20 [15.1,26.2] 9.7 [7.0,13.3] 12.6 [9.2,17.0] 15.9 [11.8,21.1] 15.9 [12.1,20.6]
American
Indian
(n=308)
2.1 [1.5,2.9] 2.3 [1.6,3.4] 3 [1.8,4.8] 3.3 [2.3,4.7] 2.4 [1.8,3.0]
Asian/Pacific
Islander
(n=833)
3.9 [2.5,6.1] 3.1 [1.8,5.1] 2.9 [1.5,5.5] 2.6 [1.6,4.2] 3.4 [2.2,5.3]
Other (n=153) 1 [0.7,1.5] 1.2 [0.7,1.8] 1.5 [0.8,3.1] 0.9 [0.5,1.8] 1.1 [0.8,1.4]
Total (n=12,837) 100 100 100 100 100
Maternal obesity status P = 0.004
Not obese (n=8,823) 69 [67.0,70.9] 72 [69.8,74.2] 77.2 [73.6,80.4] 72.8 [69.3,76.0] 71 [69.4,72.6]
Obese (n=2,010) 17.3 [16.0,18.7] 15.6 [14.0,17.4] 12.6 [9.8,16.1] 14.4 [12.2,16.9] 16.1 [15.0,17.2]
Missing (n=2,016) 13.7 [11.9,15.7] 12.4 [10.8,14.1] 10.2 [7.9,13.2] 12.8 [10.5,15.6] 12.9 [11.6,14.4]
Total (n=12,849) 100 100 100 100 100
Parent/household Income in Wave I (thousands) P = 0.000
$0-20 (n=2,316) 28.1 [24.6,31.8] 19.4 [16.3,22.9] 21.6 [16.9,27.2] 24.1 [20.1,28.7] 24.4 [21.4,27.7]
$21-38 (n=2,542) 27 [25.1,29.1] 26.2 [23.5,29.1] 21.2 [17.7,25.2] 28.4 [25.1,31.9] 26.4 [24.7,28.3]
90
$39-59 (n=2,407) 25.1 [23.1,27.2] 24.5 [22.0,27.2] 23.7 [19.8,28.2] 24.6 [20.5,29.1] 24.8 [23.1,26.5]
$60 (n=2,454) 19.7 [17.0,22.8] 29.9 [25.6,34.6] 33.5 [28.4,38.9] 22.9 [18.5,28.0] 24.4 [21.2,27.8]
Total (n=9,719) 100 100 100 100 100
Parent education level in Wave I P = 0.000
Less than H.S. (n=1,806) 17.6 [14.7,20.9] 10.4 [8.3,12.9] 11.1 [8.3,14.7] 10.7 [8.6,13.2] 14.1 [11.9,16.6]
H.S. or
equivalent
(n=3,351)
29.3 [26.8,31.9] 28.5 [25.3,31.8] 27 [22.9,31.4] 29.6 [25.8,33.7] 28.9 [26.7,31.2]
Some college
or vocational
(n=3,243)
23.1 [21.3,25.1] 27 [24.6,29.5] 26.3 [22.6,30.4] 30.4 [26.7,34.5] 25.4 [23.8,27.0]
College and
beyond
(n=2,629)
16.1 [14.1,18.3] 21.9 [18.7,25.4] 24.9 [20.5,29.9] 17.5 [14.3,21.2] 18.7 [16.4,21.2]
Missing (n=1,820) 13.9 [12.1,16.0] 12.3 [10.7,14.2] 10.8 [8.3,13.8] 11.8 [9.6,14.4] 12.9 [11.6,14.5]
Total (n=12,849) 100 100 100 100 100
Personal education level P = 0.000
Less than H.S. (n=1,020) 9.5 [7.9,11.4] 7.3 [5.8,9.2] 11.8 [8.9,15.5] 14.9 [12.2,18.1] 9.7 [8.2,11.3]
H.S. (n=2,125) 19 [16.9,21.4] 16 [13.9,18.3] 16.8 [13.1,21.2] 21.9 [18.5,25.7] 18.3 [16.5,20.2]
Some college
or vocational
(n=5,717)
41.3 [39.3,43.4] 44.8 [41.8,47.9] 43.2 [38.6,48.0] 46.4 [42.9,50.1] 43.1 [41.4,44.9]
College and
beyond
(n=3,986)
30.1 [27.1,33.3] 31.9 [27.5,36.7] 28.2 [22.9,34.2] 16.7 [13.0,21.3] 29 [25.8,32.3]
Total (n=12,848) 100 100 100 100 100
Personal income P = 0.000
<= $7,000 (n=1,867) 16.8 [15.0,18.7] 13.6 [12.0,15.2] 12.7 [9.9,16.1] 13.5 [11.3,16.1] 15.1 [13.8,16.5]
>$7,000 &
<=20,000
(n=2,246)
17 [15.6,18.6] 15.7 [14.0,17.7] 22.7 [19.0,26.8] 24.4 [21.5,27.6] 18 [16.7,19.3]
>$20,000 &
<=$35,000
(n=3,762)
29 [27.2,30.9] 28.3 [26.2,30.6] 31 [26.8,35.5] 35.6 [32.7,38.7] 29.7 [28.3,31.2]
>$35,000 (n=4,974) 37.2 [34.9,39.6] 42.4 [39.3,45.5] 33.7 [29.4,38.3] 26.4 [22.9,30.3] 37.2 [35.0,39.5]
Total (n=12,849) 100 100 100 100 100
Marital Status P = 0.000
Unmarried (n=6,455) 45 [41.3,48.8] 49.8 [46.6,53.1] 66.6 [62.0,70.9] 67.8 [64.1,71.2] 50.9 [48.1,53.6]
91
Married once
or more
(n=6,383)
55 [51.2,58.7] 50.2 [46.9,53.4] 33.4 [29.1,38.0] 32.2 [28.8,35.9] 49.1 [46.4,51.9]
Total (n=12,838) 100 100 100 100 100
Activity level (days active per week) P = 0.008
0-2 times/wk (n=4,241) 35.3 [33.3,37.3] 32.8 [30.4,35.4] 30.8 [26.5,35.4] 28.3 [25.2,31.6] 33.4 [31.8,35.0]
3-4 times/wk (n=2,052) 17.1 [15.7,18.6] 16.5 [15.0,18.2] 16.9 [13.7,20.7] 16.8 [14.5,19.5] 16.9 [15.9,17.9]
>=4 times/wk (n=6,142) 47.6 [45.9,49.3] 50.6 [47.7,53.5] 52.3 [47.6,57.0] 54.8 [51.5,58.2] 49.7 [48.1,51.3]
Total (n=12,435) 100 100 100 100 100
Tobacco Use P = 0.000
Non-smoker (n=6,773) 70.7 [68.4,72.8] 30.7 [28.4,33.2] 23.5 [20.2,27.1] 18.2 [15.3,21.7] 48.9 [46.7,51.2]
past smoker (n=1,459) 6 [5.1,7.1] 23.1 [21.5,24.9] 12.6 [10.2,15.4] 10.3 [8.3,12.8] 12.1 [11.1,13.2]
current
smoker
(n=4,575)
23.3 [21.5,25.3] 46.2 [43.8,48.5] 63.9 [59.5,68.1] 71.4 [67.7,74.9] 38.9 [37.0,40.9]
Total (n=12,807) 100 100 100 100 100
Alcohol Use P = 0.000
Non-drinker (n=7,022) 68.9 [66.8,71.1] 40.6 [37.5,43.8] 27.8 [23.5,32.6] 34.3 [30.7,38.2] 53.2 [50.7,55.6]
<1 day a week (n=1,994) 13.8 [12.5,15.1] 18.2 [16.5,20.2] 15.4 [12.6,18.8] 17 [14.6,19.8] 15.6 [14.7,16.6]
1-2 days a
week
(n=2,393)
12.4 [11.0,14.1] 25.8 [23.5,28.2] 31.5 [27.8,35.5] 24.5 [21.2,28.1] 19.4 [17.9,20.9]
3-7 days a
week
(n=1,414)
4.9 [4.1,5.8] 15.4 [13.3,17.7] 25.3 [21.3,29.7] 24.2 [21.2,27.3] 11.9 [10.7,13.2]
Total (n=12,823) 100 100 100 100 100

92
Wave III Univariate
Never used
Used but not in
the past month
Past month use,
less than 8 times
8 or more times
in past month Total
Variables
Col
% 95% CI
Col
% 95% CI
Col
% 95% CI
Col
% 95% CI
Col
% 95% CI
F-Test P-
value
Gender P = 0.000
Female (n=7,397) 54.3 [52.7,55.9] 48.3 [45.6,50.9] 44.9 [41.7,48.0] 32.6 [29.0,36.4] 49.3 [48.0,50.6]
Male (n=6,557) 45.7 [44.1,47.3] 51.7 [49.1,54.4] 55.1 [52.0,58.3] 67.4 [63.6,71.0] 50.7 [49.4,52.0]
Total (n=13,954) 100 100 100 100 100
Race/ethnicity P = 0.000
Hispanic (n=2,246) 13.1 [9.3,18.1] 11.2 [8.7,14.3] 9.2 [6.9,12.3] 9.4 [6.8,12.8] 11.8 [8.8,15.6]
Non-Hispanic
white
(n=7,320)
60.5 [53.8,66.9] 72.2 [67.2,76.6] 70.5 [64.6,75.7] 71.9 [65.6,77.5] 65.7 [59.8,71.1]
Black (n=2,813) 19 [14.3,24.6] 9 [6.6,12.0] 13.3 [10.1,17.4] 12.3 [8.9,16.9] 15.2 [11.6,19.7]
American
Indian
(n=341)
2.1 [1.6,2.8] 2.5 [1.6,3.9] 2.3 [1.4,3.7] 2.7 [1.6,4.4] 2.3 [1.7,3.0]
Asian/Pacific
Islander
(n=1,057)
4.5 [2.9,6.8] 4 [2.6,6.2] 2.8 [1.8,4.4] 2.1 [1.2,3.8] 3.9 [2.6,5.8]
Other (n=167) 0.9 [0.6,1.4] 1.2 [0.8,1.7] 1.8 [1.0,3.4] 1.6 [0.8,3.2] 1.2 [0.9,1.5]
Total (n=13,944) 100 100 100 100 100
Maternal obesity status P = 0.048
Not obese (n=9,548) 69.3 [67.3,71.2] 70.1 [67.7,72.4] 73.1 [69.6,76.3] 74.6 [71.3,77.7] 70.5 [68.9,72.1]
Obese (n=2,199) 17.4 [16.1,18.8] 16.8 [15.1,18.8] 15.5 [13.4,17.9] 13.9 [11.6,16.5] 16.7 [15.6,17.7]
Missing (n=2,207) 13.3 [11.6,15.2] 13.1 [11.3,15.0] 11.4 [9.0,14.3] 11.5 [9.2,14.2] 12.8 [11.4,14.4]
Total

(n=13,954) 100 100 100 100 100
Parent/household Income in Wave I P = 0.000
$0-20 (n=2,440) 26.8 [23.5,30.3] 19.7 [16.6,23.2] 17.9 [14.1,22.4] 20.5 [16.8,24.8] 23.3 [20.4,26.5]
$21-38 (n=2,711) 26.3 [24.5,28.2] 24.6 [21.9,27.6] 25.4 [21.9,29.2] 24.3 [21.5,27.4] 25.6 [23.8,27.4]
$39-59 (n=2,618) 25.6 [23.6,27.8] 24.7 [22.2,27.4] 23.8 [20.3,27.6] 26.7 [23.3,30.4] 25.3 [23.5,27.2]
$60
(n=2,785)
21.3 [18.5,24.3] 31 [26.8,35.5] 33 [28.0,38.5] 28.5 [23.9,33.6] 25.8 [22.6,29.3]
93
Total (n=10,554) 100 100 100 100 100
Parent education level in Wave
I 100 P = 0.000
Less than H.S. (n=1,943) 16.3 [13.7,19.3] 10.7 [8.7,13.1] 10.5 [8.1,13.5] 9.9 [7.6,12.8] 13.6 [11.5,16.0]
H.S. or
equivalent
(n=3,574)
28.5 [26.3,30.9] 29.3 [26.2,32.7] 26.9 [23.6,30.4] 27.4 [23.7,31.4] 28.4 [26.2,30.7]
Some college
or vocational
(n=3,502)
24 [22.3,25.8] 25.6 [23.3,28.0] 27.6 [24.7,30.6] 28.6 [25.5,31.9] 25.3 [23.8,26.9]
College and
beyond
(n=2,949)
17.7 [15.6,19.9] 21.8 [18.1,26.1] 23.2 [19.7,27.2] 23.7 [19.6,28.3] 20 [17.5,22.7]
Missing (n=1,986) 13.5 [11.8,15.4] 12.6 [10.9,14.5] 11.9 [9.4,14.9] 10.5 [8.3,13.2] 12.7 [11.3,14.3]
Total (n=13,954) 100 100 100 100 100
Personal education level P = 0.000
Less than H.S. (n=1,677) 13.1 [11.4,15.0] 12.9 [10.7,15.5] 15.6 [12.8,19.0] 21 [17.6,24.9] 14.3 [12.5,16.2]
H.S. (n=4,584) 35.4 [32.6,38.3] 29.8 [26.4,33.3] 28.4 [24.7,32.5] 33.3 [29.3,37.6] 33 [30.5,35.6]
Some college
or vocational
(n=5,515)
37.5 [35.0,40.1] 39.7 [36.8,42.6] 43.1 [39.2,47.1] 37.7 [33.0,42.8] 38.7 [36.3,41.1]
College and
beyond
(n=2,169)
14 [11.6,16.8] 17.6 [14.4,21.4] 12.8 [9.7,16.8] 7.9 [5.8,10.7] 14 [11.7,16.7]
Total (n=13,945) 100 100 100 100 100
Personal income P = 0.000
<= $7,000 (n=4,748) 36.2 [33.0,39.6] 30 [27.0,33.2] 36.2 [32.1,40.5] 30.8 [26.8,35.0] 34.1 [31.4,37.0]
>$7,000 &
<=20,000
(n=5,770)
40.5 [38.5,42.6] 44.3 [41.6,47.1] 42.4 [38.6,46.3] 47.4 [43.6,51.3] 42.4 [40.8,44.1]
>$20,000 &
<=$35,000
(n=2,826)
19 [16.8,21.4] 20.2 [17.7,22.9] 17.6 [14.6,21.1] 19.2 [16.0,22.8] 19.1 [17.1,21.3]
>$35,000 (n=610) 4.3 [3.6,5.2] 5.5 [4.4,6.9] 3.8 [2.6,5.5] 2.6 [1.7,4.0] 4.3 [3.7,5.1]
Total (n=13,954) 100 100 100 100 100
Marital Status P = 0.000
Unmarried (n=11,269) 77 [74.3,79.5] 81.4 [78.8,83.7] 90.7 [88.1,92.9] 92.3 [90.1,94.0] 81.4 [79.2,83.4]
94
Married once
or more
(n=2,674)
23 [20.5,25.7] 18.6 [16.3,21.2] 9.3 [7.1,11.9] 7.7 [6.0,9.9] 18.6 [16.6,20.8]
Total (n=13,943) 100 100 100 100 100
Activity level (days active per week) P = 0.000
0-2 times/wk (n=6,875) 53.4 [51.2,55.6] 47.1 [44.6,49.7] 44.1 [39.9,48.5] 46.6 [42.4,50.8] 50.1 [48.3,52.0]
3-4 times/wk (n=2,030) 14.7 [13.6,15.9] 15.1 [13.5,16.9] 19.6 [16.9,22.5] 17.3 [14.9,20.1] 15.7 [14.8,16.5]
>=4 times/wk (n=4,709) 31.9 [30.1,33.6] 37.8 [35.5,40.1] 36.3 [32.2,40.5] 36.1 [32.4,39.9] 34.2 [32.7,35.8]
Total (n=13,614) 100 100 100 100 100
Tobacco Use P = 0.000
Non-smoker (n=6,579) 69.3 [66.5,71.9] 37.2 [34.2,40.2] 31.4 [28.2,34.7] 20.8 [17.8,24.2] 51 [48.7,53.4]
past smoker (n=966) 5.8 [5.0,6.8] 14.4 [12.9,16.1] 8.5 [6.7,10.7] 8.9 [6.9,11.2] 8.5 [7.8,9.3]
current smoker (n=4,514) 24.9 [22.6,27.3] 48.4 [45.4,51.4] 60.2 [56.4,63.9] 70.3 [66.7,73.7] 40.4 [38.2,42.7]
Total (n=12,059) 100 100 100 100 100
Alcohol Use P = 0.000
Non-drinker (n=7,684) 70.6 [67.7,73.4] 39.8 [36.8,42.9] 29.1 [25.9,32.6] 25.9 [22.2,29.9] 53.5 [50.6,56.3]
<1 day a week (n=2,293) 13.7 [12.3,15.2] 19.5 [17.6,21.7] 20.8 [18.2,23.8] 17.4 [14.7,20.5] 16.3 [15.3,17.4]
1-2 days a
week
(n=2,592)
12 [10.4,13.8] 28.7 [26.1,31.4] 30.4 [27.3,33.8] 30.7 [27.4,34.3] 20.2 [18.5,22.0]
3-7 days a
week
(n=1,272)
3.7 [3.0,4.5] 12 [10.5,13.8] 19.6 [16.8,22.7] 26 [22.8,29.6] 10.1 [9.1,11.1]
Total (n=13,841) 100 100 100 100 100

95
Wave I Univariate
Never used
Used but not in
the past month
Past month use,
less than 8 times
8 or more times in
past month Total
Variables
Col
% 95% CI
Col
% 95% CI
Col
% 95% CI
Col
% 95% CI
Col
% 95% CI
F-Test P-
value
Gender P = 0.000
Female (n=9,445) 50.8 [49.6,52.0] 47.7 [44.8,50.5] 49.1 [45.6,52.5] 36.6 [31.5,42.0] 49.5 [48.4,50.6]
Male (n=8,991) 49.2 [48.0,50.4] 52.3 [49.5,55.2] 50.9 [47.5,54.4] 63.4 [58.0,68.5] 50.5 [49.4,51.6]
Total (n=18,436) 100 100 100 100 100
Race/ethnicity P = 0.071
Hispanic (n=3,125) 11.9 [8.8,15.9] 13.2 [9.7,17.7] 10.4 [7.8,13.8] 13.2 [9.3,18.3] 12 [9.1,15.7]
Non-Hispanic
white
(n=9,469)
65.1 [59.1,70.6] 66.3 [59.3,72.7] 65.9 [59.3,71.9] 65.8 [58.3,72.5] 65.3 [59.5,70.7]
Black (n=3,763) 15.5 [11.9,20.1] 13.1 [9.3,18.2] 15.7 [11.5,21.1] 14.3 [9.9,20.3] 15.2 [11.6,19.6]
American
Indian
(n=446)
2 [1.6,2.5] 3 [2.0,4.4] 3.6 [2.2,5.8] 3.5 [2.1,5.8] 2.3 [1.8,3.0]
Asian/Pacific
Islander
(n=1,376)
4.3 [2.9,6.3] 3 [1.7,5.4] 3 [1.7,5.3] 2.2 [1.2,4.1] 3.9 [2.6,5.8]
Other (n=247) 1.2 [0.9,1.6] 1.4 [0.9,2.2] 1.4 [0.7,2.7] 1 [0.4,2.4] 1.3 [1.0,1.6]
Total (n=18,426) 100 100 100 100 100
Maternal obesity
status P = 0.001
Not obese
(n=12,534)
71.1 [69.4,72.8] 66.4 [63.1,69.6] 70.1 [67.1,73.0] 70.1 [65.7,74.1] 70.3 [68.7,71.9]
Obese (n=2,749) 16.2 [15.1,17.3] 16 [13.7,18.6] 14.8 [12.2,17.7] 13.8 [11.2,16.8] 15.9 [15.0,16.9]
Missing (n=3,153) 12.7 [11.2,14.4] 17.6 [15.1,20.3] 15.1 [12.5,18.2] 16.1 [13.2,19.5] 13.7 [12.3,15.3]
Total (n=18,436) 100 100 100 100 100
Parent/household Income in
Wave I P = 0.638
$0-20 (n=3,340) 24.3 [21.1,27.7] 24.6 [20.5,29.2] 24.6 [20.0,29.8] 27.5 [21.7,34.2] 24.5 [21.4,27.8]
$21-38 (n=3,460) 25.4 [23.6,27.2] 26.9 [24.0,30.0] 22.6 [18.8,26.8] 26.3 [21.9,31.2] 25.4 [23.7,27.1]
$39-59 (n=3,358) 25.1 [23.4,26.9] 23.5 [20.7,26.5] 25.5 [21.5,30.0] 22.6 [18.3,27.6] 24.8 [23.2,26.5]
96
$60 (n=3,509) 25.3 [22.0,28.8] 25 [20.8,29.7] 27.4 [23.0,32.2] 23.6 [19.0,28.9] 25.3 [22.2,28.8]
Total (n=13,667) 100 100 100 100 100
Parent education level Wave I 100 P = 0.004
Less than H.S. (n=2,614) 14 [11.8,16.6] 14.2 [11.3,17.6] 12.5 [10.1,15.3] 12.9 [10.1,16.3] 13.8 [11.7,16.3]
H.S. or
equivalent
(n=4,702)
29.1 [26.9,31.4] 27.5 [24.8,30.4] 26.7 [23.2,30.4] 29.9 [25.5,34.7] 28.7 [26.7,30.8]
Some college or
vocational
(n=4,597)
24.9 [23.4,26.5] 26.5 [23.9,29.2] 27.5 [23.7,31.6] 25.8 [22.0,30.0] 25.4 [23.9,27.0]
College and
beyond
(n=3,659)
19.3 [17.0,21.8] 15 [12.6,17.8] 18.4 [15.1,22.3] 15.7 [11.4,21.2] 18.4 [16.2,20.9]
Missing (n=2,864) 12.7 [11.1,14.4] 16.8 [14.3,19.7] 15 [12.3,18.1] 15.8 [12.9,19.1] 13.6 [12.1,15.2]
Total (n=18,436) 100 100 100 100 100
Activity level (days active per
week) P = 0.000
0-2 times/wk (n=3,312) 16.6 [15.3,18.1] 19.4 [17.2,21.7] 19.9 [17.4,22.7] 23.5 [19.7,27.8] 17.6 [16.4,19.0]
3-4 times/wk (n=3,198) 16.9 [15.9,17.9] 18 [16.2,19.9] 17.5 [15.0,20.3] 18.7 [15.6,22.4] 17.2 [16.3,18.1]
>=4 times/wk (n=11,924) 66.5 [64.7,68.3] 62.6 [59.9,65.3] 62.6 [59.4,65.7] 57.8 [53.1,62.3] 65.2 [63.4,66.9]
Total (n=18,434) 100 100 100 100 100
Tobacco Use P = 0.000
Non-smoker (n=10,561) 79.7 [77.6,81.6] 35.4 [31.7,39.4] 23.2 [20.2,26.4] 14.8 [11.9,18.2] 64.5 [61.8,67.1]
past smoker (n=543) 2.9 [2.4,3.4] 9.4 [8.0,11.2] 3.6 [2.4,5.3] 2.7 [1.6,4.4] 3.8 [3.4,4.3]
current smoker (n=4,712) 17.5 [15.7,19.4] 55.1 [51.4,58.8] 73.3 [69.7,76.5] 82.5 [78.8,85.7] 31.7 [29.3,34.2]
Total (n=15,816) 100 100 100 100 100
Alcohol Use P = 0.000
Non-drinker (n=15,139) 91.5 [90.4,92.4] 70.5 [67.3,73.5] 51.2 [47.3,55.2] 32.7 [28.1,37.6] 81.9 [80.3,83.5]
<1 day a week (n=1,458) 4.3 [3.8,4.9] 14 [11.6,16.8] 22.5 [19.6,25.6] 19.5 [16.7,22.6] 8.1 [7.2,9.0]
1-2 days a week (n=1,165) 2.8 [2.3,3.3] 10 [8.3,12.0] 19.1 [16.7,21.7] 25.8 [21.9,30.1] 6.4 [5.7,7.2]
3-7 days a week (n=612) 1.4 [1.2,1.8] 5.5 [4.4,6.9] 7.2 [5.6,9.3] 22.1 [18.7,25.9] 3.6 [3.2,4.0]
Total (n=18,374) 100 100 100 100 100

97
Appendix B. Results by gender
Wave IV OLS, females (comparison: non-users)
VARIABLES BMI
Waist
Circumference Glucose HBA1C Triglycerides HDL-C LDL-C

Used marijuana,
not in past month
-0.071 -1.026 1.337 -0.028 0.034 0.291** 0.032
(0.051) (0.902) (1.470) (0.027) (0.152) (0.136) (0.143)
Past month use,
less than 8 times
-0.149** -3.837*** -2.535 -0.062* -0.142 0.055 0.072
(0.067) (1.313) (1.670) (0.035) (0.278) (0.226) (0.253)
8 or more times
in past month
-0.072 -1.717 -1.180 0.031 -0.087 0.098 0.125
(0.078) (1.428) (2.863) (0.064) (0.217) (0.242) (0.267)
R-squared 0.143 0.143 0.025 0.078 0.057 0.049 0.015
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

Wave IV OLS, males (comparison: non-users)

VARIABLES BMI
Waist
Circumference Glucose HBA1C Triglycerides
HDL-
C LDL-C

Used marijuana,
not in past month
-0.037 -0.566 -2.476 -0.057* -0.079 -0.064 -0.132
(0.051) (0.778) (1.609) (0.032) (0.142) (0.162) (0.160)
Past month use,
less than 8 times
-0.185*** -2.719*** -3.039 -0.029 -0.104 0.060 0.181
(0.068) (1.033) (2.689) (0.065) (0.230) (0.267) (0.242)
8 or more times in
past month
-0.213*** -2.532** -3.071 -0.043 -0.607*** -0.246 -0.070
(0.060) (1.066) (1.925) (0.042) (0.184) (0.165) (0.231)
R-squared 0.120 0.111 0.014 0.062 0.077 0.036 0.024
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

Wave III OLS, females
(comparison: non-users)

VARIABLES BMI

Used marijuana,
not in past month
-0.026
(0.047)
0.061
98
Past month use,
less than 8 times (0.063)
8 or more times
in past month
-0.009
(0.066)
R-squared 0.116
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

Wave III OLS, males
(comparison: non-users)

VARIABLES BMI

Used marijuana,
not in past month
0.033
(0.053)
Past month use,
less than 8 times
-0.186***
(0.057)
8 or more times in
past month
-0.210***
(0.065)
R-squared 0.096
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

Wave I OLS, females
(comparison: non-users)

VARIABLES BMI

Used marijuana, not
in past month
-0.014
(0.051)
Past month use, less
than 8 times
0.064
(0.062)
8 or more times in
past month
-0.171**
(0.070)
R-squared 0.111
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

99
Wave I OLS, males
(comparison: non-users)

VARIABLES BMI

Used marijuana, not
in past month
-0.016
(0.048)
Past month use, less
than 8 times
-0.100
(0.079)
8 or more times in
past month
-0.207**
(0.082)
R-squared 0.066
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

100

Instruments: Political beliefs, risk-taking*, females
VARIABLES BMI
Waist
Circumference Glucose HBA1C Triglycerides HDL-C LDL-C
Used but not in
past month
0.523 8.942 0.309 0.528 0.288 1.725 0.637
(0.462) (9.026) (15.161) (0.491) (1.667) (2.560) (1.378)
Past month
use, less than 8
times
0.050 11.628 13.355 -1.828 -4.991 -13.284 -1.858
(3.227) (54.384) (77.829) (3.011) (8.517) (14.772) (6.535)
8 or more
times in past
month
0.713 -3.576 -9.388 0.911 4.661 7.180 0.565
(1.958) (33.974) (51.636) (2.191) (5.599) (9.792) (4.076)

Instruments: Political beliefs, risk-taking, lagged marijuana use*, females
VARIABLES BMI
Waist
Circumference Glucose HBA1C Triglycerides HDL-C LDL-C
Used but not in
past month
-0.154 -4.395 -0.085 -0.026 0.037 0.686 0.021
(0.144) (2.932) (4.394) (0.099) (0.337) (0.432) (0.408)
Past month
use, less than 8
times
0.790 23.576 15.824 -0.059 -1.179 0.035 -1.327
(0.739) (15.826) (34.160) (0.600) (2.287) (2.257) (2.266)
8 or more
times in past
month
-0.309 -10.172 -7.992 -0.039 0.146 -0.071 0.217
(0.407) (8.950) (14.100) (0.253) (0.938) (1.063) (1.196)

Instrument: Political beliefs, females
VARIABLES BMI
Waist
Circumference Glucose HBA1C Triglycerides HDL-C LDL-C
Used but not in
past month
2.926 19.747 -12.442 0.716 -0.693 -0.036 2.400
(16.529) (60.949) (27.542) (1.369) (2.345) (2.282) (7.851)
-24.811 -109.109 84.646 -2.660 -0.396 -2.958 -19.895
101
Past month
use, less than 8
times (167.024) (554.662) (169.542) (11.629) (12.590) (12.815) (54.162)
8 or more
times in past
month
16.429 78.693 -43.468 1.327 2.773 2.099 11.861
(106.655) (369.886) (104.942) (9.124) (7.570) (7.231) (29.985)

*results from regressions without any potentially endogenous covariates (alcohol, tobacco, etc.) do not
change

Instruments: Political beliefs, risk-taking*, males

VARIABLES BMI
Waist
Circumference Glucose HBA1C Triglycerides HDL-C LDL-C
Used but not in
past month
-0.053 -4.562 -76.325 -1.446 1.809 0.251 0.567
(2.435) (39.095) (62.256) (1.576) (3.395) (2.816) (3.498)
Past month use,
less than 8
times
-0.514 -5.897 33.641 2.444 -0.877 -1.395 -7.339
(4.005) (56.928) (115.441) (3.447) (7.979) (6.685) (7.290)
8 or more times
in past month
1.667 30.931 -60.722 -1.748 3.803 3.085 0.944
(2.092) (27.802) (68.850) (1.657) (4.218) (3.259) (3.891)

Instruments: Political beliefs, risk-taking, lagged marijuana use*, males
VARIABLES BMI
Waist
Circumference Glucose HBA1C Triglycerides HDL-C LDL-C
Used but not in
past month
0.240 -0.158 -5.342 -0.205* -0.398 -0.879 -0.063
(0.259) (3.421) (5.974) (0.119) (0.653) (0.744) (0.855)
Past month use,
less than 8
times
-1.487 -7.983 -30.074 0.456 2.114 6.240 -6.538
(1.392) (18.111) (36.893) (0.789) (3.314) (4.264) (5.759)
0.321 0.254 14.638 -0.061 -1.336 -2.136 2.259
102
8 or more times
in past month (0.616) (7.612) (17.543) (0.336) (1.401) (1.686) (2.311)

Instrument: Political beliefs, males
VARIABLES BMI
Waist
Circumference Glucose HBA1C Triglycerides HDL-C LDL-C
Used but not in
past month
2.797 30.886 -66.159 -1.590 2.405 0.299 -1.022
(9.569) (110.193) (84.941) (6.689) (9.976) (5.246) (5.557)
Past month use,
less than 8
times
-8.188 -106.664 80.594 10.809 -13.506 -8.247 -11.665
(20.021) (237.811) (218.038) (36.817) (29.112) (26.654) (21.804)
8 or more times
in past month
5.949 87.803 -90.761 -5.692 12.283 7.514 4.116
(11.541) (137.396) (141.551) (17.400) (22.118) (16.693) (13.930)

*results from regressions without any potentially endogenous covariates (alcohol, tobacco, etc.) do not
change

103

Panel FE, females
VARIABLES BMI
BMI, no
endogenous
covariates

Past user -0.013 -0.022
(0.038) (0.039)
Past month use, less than 8
times 0.013 -0.015
(0.055) (0.050)
8 or more times in past month -0.045 -0.089**
(0.045) (0.042)

Observations 6,390 6,584
R-squared 0.033 0.014
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

Panel FE, males

VARIABLES BMI
BMI, no
endogenous
covariates

Past user -0.017 -0.013
(0.036) (0.030)
Past month use, less than 8
times -0.026 -0.025
(0.049) (0.044)
8 or more times in past month -0.121** -0.116**
(0.050) (0.044)

Observations 5,320 5,533
R-squared 0.027 0.010
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

104

Diff GMM, female
VARIABLES BMI
BMI, no
endogenous
covariates

Past user 0.598 0.379
(0.708) (0.345)
Past month use, less than 8
times 0.438 -0.129
(2.02) (0.686)
8 or more times in past month 0.010 0.382
(1.10) (0.970)
Lagged BMI 0.202 0.236
(0.130) (0.090)

Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

Sargan test of overid. restrictions: chi2(4)=2.33; Prob > chi2= 0.675

Diff GMM, male

VARIABLES BMI
BMI, no
endogenous
covariates

Past user -2.178 -0.237
(3.263) (1.108)
Past month use, less than 8
times -2.197 -1.872
(3.310) (1.955)
8 or more times in past month -3.127 0.319
(4.379) (0.877)
Lagged BMI 0.233 0.166
(0.249) (0.132)

Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

Sargan test of overid. restrictions: chi2(4)= 0.35; Prob> chi2 =
0.986

Abstract (if available)

Abstract Contents: 1) Estimating the association between metabolic risk factors and marijuana use in US adults using data from the continuous National Health And Nutrition Examination survey

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Hepatitis C in the post-interferon era: selected essays in health economics

PDF

Essays on the use of microsimulation for health and economic policy analysis

PDF

Advances and applications for economic evaluation methods in health technology assessment (HTA)

PDF

Essays in pharmaceutical and health economics

PDF

Selection and impacts of early life events on later life outcomes

PDF

Essays on the economics of infectious diseases

PDF

The impact of Patient-Centered Medical Home on a managed Medicaid plan

PDF

Three essays on comparative economic and clinical outcomes of antipsychotics using retrospective claims data

PDF

Essays on the equitable distribution of healthcare

PDF

The impact of adherence to guidelines on the health care expenditures of COPD patients

PDF

Burden of illness in hemophilia A: taking the patient’s perspective

PDF

Evaluating approaches to reduce inappropriate antibiotic use in the United States

PDF

The role of altered folate one-carbon metabolism (FOCM) in the development of colorectal cancer

PDF

The causal-effect of childhood obesity on asthma in young and adolescent children

PDF

Selected papers on the evaluation of healthcare costs of prematurity and necrotizing enterocolitis using large retrospective databases

PDF

Development of new approaches for antibody modification

PDF

Impact of pharmacy-based transitional care on healthcare utilization and costs

PDF

Delivering better care for children with special health care needs: analyses of patient-centered medical home and types of insurance

PDF

Inhibition of tumor cell growth by mefloquine via multimechanistic effects involving increased cellular stress, inhibition of autophagy, and impairment of cellular energy metabolism

PDF

Healthcare provider recommendations: a panacea to improving influenza vaccination rates?

Asset Metadata

Creator Thompson, Christin A. (author)

Core Title Selected papers on methods for evaluating the impact of marijuana use on BMI and other risk factors for metabolic syndrome

School School of Pharmacy

Degree Doctor of Philosophy

Degree Program Pharmaceutical Economics and Policy

Publication Date 12/17/2015

Defense Date 09/18/2015

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag BMI,cannabis,cardio-metabolic health,dynamic panel,effect modification,endogeneity,fixed effects,insulin,Marijuana,metabolic health,OAI-PMH Harvest,omitted variables bias,ordinary least squares,two-stage least squares,waist,weight

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Hay, Joel W. (committee chair), Lu, Yang (committee member), Sood, Neeraj (committee member)

Creator Email christinthompso@gmail.com,thompsoc@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c40-203961

Unique identifier UC11278165

Identifier etd-ThompsonCh-4066.pdf (filename),usctheses-c40-203961 (legacy record id)

Legacy Identifier etd-ThompsonCh-4066.pdf

Dmrecord 203961

Document Type Dissertation

Format application/pdf (imt)

Rights Thompson, Christin A.

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA