This manuscript has been reproduced from the microfilm master UMI films the text directly from the original or copy submitted Thus, some thesis and dissertation copies are in typewriter face, while others may be from any type of computer printer
The quality of this reproduction is dependent upon the quality of the copy submitted Broken or indistinct print, colored or poor quality illustrations and photographs, print bleedthrough, substandard margins, and improper alignment can adversely affect reproduction
in the unlikely event that the author did not send UMI a complete manuscript
and there are missing pages, these will be noted Also, if unauthorized copyright material had to be removed, a note will indicate the deletion
Oversize materials (e.g., maps, drawings, charts) are reproduced by sectioning the original, beginning at the upper left-hand corner and continuing from left to right in equal sections with small overiaps
Photographs included in the original manuscript have been reproduced xerographically in this copy Higher quality 6° x 9° black and white photographic prints are available for any photographs or illustrations appearing in this copy for an additional charge Contact UMI directly to order
ProQuest Information and Learning
300 North Zeeb Road, Ann Arbor, MI 48106-1346 USA 800-521-0600
&
Trang 3A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF ECONOMICS AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
Trang 4UMI Number: 3048517
UMI
UME Microform 3048517
Capyright 2002 by ProQuest Information and Learning Company All rights reserved This microform edition is protected against
unauthorized copying under Title 17, United States Code
ProQuest Information and Learning Company 300 North Zeeb Road
Trang 6Bỉ
I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy
Š
Dong Bernheim, Principal Advisor
I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy
⁄_⁄2 Anosểngg
I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy
2 —
: Shoven
Approved for the University Committee on Graduate Studies:
Trang 7emphasis on public policy implications In the first chapter I focus on people’s behaviors related to heart disease risk I speculate that the rise in overweight and obesity rates in the United States during the 1980s was at least partly related to the arrival of highly effective drugs to lower blood pressure and cholesterol [ construct this hypothesis from a simple economic model in which people regard heart disease risk factors
as substitutes [ test the hypothesis empirically by using insurance stams as an instrument for the
availability of the new drugs and employing a “diff-in-diff-in-diff” approach to individual level data From
a policy perspective this study could have implications for how we evalwate the benefits of drugs, in terms of health outcomes and overall utility
‘he second chapter presents new findings on the effectiveness (in terms of fatal crash reductions)
of state-level public policies related to drunk driving Given the possibility for correlation among policies,
it is important to look at all of these policies together rather than im isolation This study includes the most comprehensive set of drunk driving-related policy variables to date It raises the issue that conventional estimates of policy effects might be biased due to the endogeneity of policies, and addresses this concern by analyzing the time pattern of policy effects with respect to the date of adoption For the 0.08 BAC law the results suggest that a bias upwards exists, but the policy is still somewhat effective Graduated
licensing programs for young drivers and the Mothers Against Drunk Driving (MADD) organization are
also evaluated for the first time in this type of analysis The estimated time pattern of effects for graduated licensing suggest that its effects are also overstated in conventional analyses but the policy is still effective for young drivers The results for MADD do not imply an effect but this could be due to the crudeness of the variable used
In the third and final chapter I address the question of whether two particular types of peer effects
exist for substance use by adolescents [ argue in Section H of the chapter that to date there is no solid
empirical evidence on peer effects for substance use although a substantial body of research is present In
general the empirical strategies in the literature do not confront certain major estimation problems Most
notably they do not adequately address the possible presence of correlated unobservable factors due to
selection or common environments within peer groups [ propose and test two estimation strategies which
are arguably immune to the problems in the literature These strategies focus on two types of peer effects
which correspond to two particular policy interventions The first corresponds to the removal of a “bad
Trang 8Acknowledgements
I am grateful to many people for their help with this project [ thank Professors Doug Bernheim and Antonio Rangel for their many valuable ideas and consistent
guidance I thank Professors John Shoven, Jon Levin, Dan Kessler, and Mark McClellan
for helpful feedback at various stages in the process I thank my fellow graduate students in the Public Economics group for their many constructive comments on my work [| thank the Forman family for its financial support through the Forman Family Fellowship, the National Science Foundation for its support through the Graduate Research
Fellowship, and the Stanford Institute for Economic Policy Research (SIEPR) and Lynde and Harry Bradley Foundation for support through SIEPR’s Bradley Dissertation
Trang 9Introduction I Chapter LI: Are we substituting between heart disease risks? 4
Chapter 2: Evaluating the effectiveness of polices related 38 to drunk driving
Chapter 3: Peer effects for adolescent substance use: 97 do they really exist?
Trang 10List of Tables Chapter I
E.L: Trends in the major “controllable” heart disease risk factors in the U.S 1.2: Trends in obesity prevalence among American adults
1.3: Trends in heart disease risk factors, cont'd
L4: Drug Mentions (in thousands) From Office Visits in the U.S L.5: Dispensed Prescriptions (in thousands) in the United States L.6: Dispensed Prescriptions (in thousands) in the United States
1.7 Regressing overweight on whether taking blood pressure (cholesterol) med 1.8: [Impact of msurance status on health: "before” (71-75) & “after” ('88-'94) influx
of effective BP/cholesterol drugs
1.9: Impact of insurance status on health (continued): Repeating Table 4.2, w/ time*educ
& time*income imteractions added
1.10: Do we observe greater "diff-in-diff-in-diffs” for HS grads? I LI: Do we observe greater "diff-in-diff-in-diffs” for college grads? 1.12: Do we observe greater "diff-in-diff-in-diffs” for men 45+ years old?
1.13: Do we observe greater "diff-in-diff-in-diffs” for HS grad 45+ years old men?
Chapter 2
2.5: Summary of imitiatives related to drunk driving and traffic safety 2.2: Previous studies on the effectiveness af 0.08 percent BAC law
2.3: Summary of selected studies from general literature on policies related to drunk driving
2.4: Outcome measures (dependent variables) `
2.5: Policy variables
2.6: Timing variables for 0.08 BAC limit 2.7: Other control variables
2.8: States which adopted 0.08 BAC by 6/2000: when did they adopt other policies related
to drunk driving?
2.9: Estimated effects of policies on fatal crash rates
2.10: Timing of the effects of 08 BAC law on fatal crash rates 2.t1: Timing of self-reported drinking with respect to 0.08 BAC law 2.12: Timing of the effects of 08 BAC law on fatal crash rates (cont'd) 2.13: Timing of the effects of MADD on fatal crash rates
2.14: Estimated effects of policies on fatal crash rates for young drivers 2.15: Timing of the effects of zero tolerance on under-2!1 fatal crash rates 2.16: Timing of the effects of graduated licensing on under-21 fatal crash rates 2.17: Effects of policies on fatal crash rate, by gender
2.18: Predicted lives saved per year in NY state with 0.08 law
2.19: imputed non-fatal crashes avoided per year in N¥ with 0.08 law 2.20: Cost per incident, by severity type (thousands of 1999$)
2.21: Predicted total benefits (billions of 1999$) from: first ten years of 0.08 law in NY 2.22: Estimated # of licensed drivers who drink in NY state
Trang 11Chapter 3
3.1: Summary of literature on peer effects for substance use (and other risky behaviors)
3.2: Summary of literature on peer effects for other applications 3.3: Average frequency of substance use per month, by grade
3.4: Regressing individual behavior on school behavior (minus individual)
3.5: Regressing individual behavior on peer behavior, w/ different peer groups
3.6: Regressing individual behavior on peer behavior — with school fixed effects 3.7: Regression of individual behavior on school behavior, 2SLS
3.8: Regressing individual behavior on friends’ behavior
3.9: Regressing change in individnal behavior on change in school behavior 3.10: Regressing change in individual behavior on change in friend behavior 3.11: Regressing change in individual behavior on initial friend behavior
3.12: Number of observations for kids of different types
3 L3: Regressing change in individual behavior on “bad friend moves”
3.14: Regressing change in individual behavior on “worst friend moves” 3.15: Regressing change in individual behavior on "bad friend graduates” 3.16: Regressing change in individual behavior on "worst friend graduates” 3.17: Regressing change in individual behavior on change im average peer behavior
due to moving
3.18: Regressing change m individual (non-grad) behavior on change in average peer
behavior due to graduation
3.19: Evaluating "mover" and “grad” instruments 3.20: Evaluating “mover” and “grad” instruments (cont'd)
3.21: Correlation btwn "mover" ("grad") instrument and change in peer environment
3.22: Number of schools (and students) of each type
3.23: Do 8th and 7th graders in schools with older kids engage in more substance use?
3.24: Differences in observable variables between kids in "older" vs “younger” schools
3.25: Average behavior of kids in other grades, “older” vs “younger” schools 3.26: Regressing individual behavior on avg behavior of other grades in school
Trang 12List of Figures Chapter I
II: Heart disease risk factor trends in the U.S adult population
1.2: Simple model of response to availability of new, effective drugs L.3: “Diff-in-diff-in-diff” approach
Chapter 2
2.1: U.S fatal crash rates per 10,000 people, [982-2000 2.2: Timing of 0.08 policy effects
2.3: Timing of zero tolerance policy effects
2.4: Timing of graduated licensing policy effects
Trang 13common behaviors that endanger people’s health m the United States, and thereby gain some understanding of how public policies should address these behaviors The behaviors studied include those related to heart disease risk factors (e.g control of weight, blood pressure, and cholesterol), drunk driving, and substance use A large literature spanning the disciplines of medicine, public health, psychology, sociology, economics, and others has examined the nature of these behaviors, but some important questions remain unanswered [ address the following particular questions: 1) to what extent do people behave as if heart disease risk factors are substitutes?, 2) how effective have recent public policies been at reducing fatal crashes involving drunk driving?, 3) to what extent do adolescents exert certain types of “peer effects” on each other’s level of substance use?
Each of these questions relates to public policy on two levels The first level is paternalism People might not always do “what is best for themselves,” because their preferences are time-inconsistent or they have incorrect or insufficient information about consequences of behaviors To the extent that this is true, policymakers can improve people’s welfare by understanding the determinants of their behaviors, and then use that understanding to modify the behaviors The second level is protection of others in society In various ways the behaviors I study can negatively impact the wellbeing of people other than the doers Even behaviors related to heart disease risk fit this
description, as people who contract heart disease are potential burdens to the medical care system, for example In cases where the doers fail to internalize the externalities they impose on others, policymakers have 2 potential role for improving social welfare
Trang 14services research, sociology, public health, or psychology Economists have brought their traditional theoretical and statistical toolkits to nontraditional applications This dissertation is 2 continuation of that tradition
In the first chapter I focus on people’s behaviors related to heart disease risk 1 speculate that the rise in overweight and obesity rates in the United States during the
1980s was at least partly related to the arrival of highly effective drugs to lower blood pressure and cholesterol I construct this hypothesis from a simple economic model in which people regard heart disease risk factors as substitutes I test the hypothesis empirically by using insurance status as an instrument for the availability of the new drugs and employing a “diff-in-diff-in-diff’” approach to individual level data From a policy perspective this study could have implications for how we evaluate the benefits of drugs, in terms of health outcomes and overall utility
The second chapter has a much more direct connection to public policy, as [ explicitly evaluate the effectiveness of policies related to drunk driving Using a state and year fixed effects model with a panel of state-level observations, I test whether certain policies have reduced fatal crash rates related to drunk driving in the United States in the last twenty years Policies have varied substantially across states, and which set of policies is optimal for confronting drunk driving remains a controversial question The results from this chapter will hopefully help inform this debate
In the final chapter I address the question of whether two particular types of peer effects exist for substance use by adolescents The types of peer effects correspond to two particular policy interventions The first corresponds to the removal of a “bad apple” from a peer group; the second corresponds to the mixing of children with schoolmates who are several years older I propose and test two estimation strategies which are arguably immune to the problems common to the empirical peer effects literature Understanding the impact of these peer effects has implications for real-world policies, both public and private If we were to discover, for example, that the removal of “bad apples” from peer groups leads to significantly improved behavior for the remaining peers, then we might want to adjust accordingly the practice of expelling or transferring “bad apples” m schools Similarly, parents might want to adjust their “policies”
Trang 16Chapter I: Are we substituting between heart disease risks? 4
Chapter I: Are we substituting between heart disease risks?
I: Introduction
For over fifty years heart disease has been the number one cause of death for Americans, but we can take courage im the fact that we have a large amount of control over our individual vulnerabilities With our behavior and medical treatment we can modify several of the most important predisposing factors The National Heart, Lung, and Blood Institute identifies six major “controllable” risk factors for heart disease: high blood pressure, high cholesterol level, overweight, smoking, diabetes, and physical inactivity’ (NHLBI £999) These controllable risk factors have changed significantly in the American population in the last two decades According to data from the National Health
Examination Survey (NHES) and three editions of the National Health and Nutrition Examination Survey (NHANES E, Hf, and IED), the prevalences of high blood pressure, high cholesterol, and smoking have declined in the American adult population, while the
prevalences of overweight and diabetes have increased (see Figure 1.1) Physical inactivity is impossible to quantify reliably given available data
This study focuses on the contrast between the recent sharp declines in proportion of people with high blood pressure or high cholesterol, and the recent sharp increase in proportion of people who are overweight Why did these trends diverge? Increased awareness of both the presence and consequences of tisk factors has probably played a large role in reducing the rates of high blood pressure and high cholesterol The federal government launched the National High Blood Pressure Education Program in 1973 and the National Cholesterol Education Program in 1985 Between the late 1970s and late
1980s, the percentages of Americans with high blood pressure who were aware they had high blood pressure increased from SI to 73 Treatment of people with high blood pressure increased from 31 to 55, according to the sixth report of the the Joint National Committee on prevention, detection, evaluation, and treatment of high blood pressure
Trang 17
change in eating habits by the American people For example, according to various surveys and food data, during the 1970s Americans significantly reduced their consumption of fat, cholesterol, sugar, and salt (Stamler 1985)
Meanwhile, increased awareness of heart disease risk has clearly not been enough to prevent the increase in the proportion of overweight people This upward trend can probably be attributed largely to more sedentary lifestyles While it is difficult to observe objective trends in physical activity in any available data set, Steve Fortmann, the Director of the Stanford Center for Research in Disease Prevention, points out that one can make
reasonable inferences Average caloric intake has fallen while overweight prevalence has risen, indicating that calorie expenditure has fallen He surmises that decreased activity, particularly in jobs and transportation, is largely responsible for the fall in calorie
expenditure In particular, office jobs have replaced more active manual labor jobs, cars and other vehicles have replaced walking as means of transportation: and televisions and computers have replaced more calorie-burning sources of entertainment In an explorative study on the causes of recent increases in obesity, Philipson and Posner (1999) speculate that the decreasingly physical nature of jobs is indeed a significant factor, and they also note that there is little evidence in support of alternative explanations such as decreases in smoking rates and shifts from homecooked to fast food meals
While the above explanations certainly represent major pieces of the puzzle, this study proposes an additional explanation involving recent innovations in medical technology, which could help explain simultaneously the declines in people with high blood pressure and cholesterol as well as the increase im people who are overweight In the last two decades, effective drug treatments of high blood pressure and high cholesterol have surfaced, while comparably effective treatments for obesity have not This study poses and attempts to test the following hypothesis: economic utility maximization suggests that consumers, with the mtroduction of more cost-effective blood pressure and cholesterol reducing treatments, would begin to lead riskier lifestyles with respect to potential heart disease, in particular by controlling their weights less diligently”
Trang 18Chapter I: Are we substituting between heart disease risks? 6
In order to examine this hypothesis empirically, f employ a “diff-in-diff-in-diff” type approach I look at differentials in health outcomes across drug availability (proxied by insurance status) and time (before versus after the new drug influx), and I compare these differentials across different health measures (in particular, overweight versus other
measures)
The intuition for this approach ts as follows People with insurance when the effective blood pressure and cholesterol drugs are available (time = 1) can be thought as the “treatment” group; the effective drugs are in theory available to them Their
contemporary peers without insurance are the “controf” group Comparison of health outcomes across these two groups yields the first “diff.” Next, it is likely that there are unobservable differences in people’s characteristics across insurance status that would affect health outcomes I purge the portion of these differences in characteristics which are unchanged over time by comparing the health “diff” across insurance status at time | to the health “diff” across insurance status at time 0 (before the introduction of new, effective drugs), yielding a “diff-in-diff” Finally, [ am concerned that the “diff-in-diff” might simply represent a general trend in the differential across insurance status for all
health measures In order to confront this issue, | compare the “diff-in-diffs” across different health measures Overall the results appear to support the hypothesis of this study
Trang 19The hypothesis of this study is motivated by the fact that certain trends in high
blood pressure, high cholesterol, and overweight rates coincided with the introduction and
expanded popularlity of effective treatments for high blood pressure and high cholesterol In this section I examine these trends in detail I begin by presenting trend data on heart disease risk factors, and then I present trend data on drug prescriptions for these
conditions Finally, I evaluate the temporal links between the two sets of data and conclude that they are consistent with the hypothesis of this study
Trends in heart disease risk factors
Table 1.1 shows the percentage of U.S adults (over [8 years old) with the listed heart disease risk factors in surveys administered during the years shown The percentages were computed using data from four different surveys published by the federal
government The first is the National Health Examination Survey (NHES), conducted from £960 to 1962, and the three succeeding surveys are versions of the National Health and Nutrition Examination Survey (NHANES [, I, and I), which were conducted in
1971-75, 1976-80, and 1988-94 respectively The surveys contain nationally
representative samples of individuals who responded to surveys and completed medical examinations They contain the only national data on blood pressure levels, cholesterol levels, and body measurements which come from doctors’ examinations rather than self- reporting All four data sets are freely available through the Inter-university Consortium for Political and Social Research (ICPSR) at the University of Michigan
Trang 20Chapter L: Arc we substitnting between heart đisease risks7 8 are less reliable, due to the fact that it is self-reported (respondents’ blood pressure, cholesterol, and weight are physically measured by NHES and NHANES) Improvements im detection rates could have accounted in part for the rise in diabetes prevalence
According to the Centers for Disease Control (CDC), currently almost [6 million Americans have diabetes, but about one-third of them are not aware of their condition Finally, cigarette smoking decreased during the 1980s, but not so markedly as high blood pressure and high cholesterol Cross-tabulations by sex and age groups (not shown) reveal that the basic trends above hold for each sex, as well as for each age group
Further investigation of the increasing trend for overweight prevalence reveals that the rise in the 1980s was even steeper for the more severe cases of overweight A study by Flegal, Carroll, Kuczmarski, and Johnson classified obesity as class I if person’s body
mass index’ (BMI) was between 30.0 and 34.9, class EE if between 35.0 and 39.9, and
class HI if over 40.0 Their results showed dramatic increases in the proportion of American falling into each category between NHANES IE and NHANES III (see Table 1.2) (Flegal et al 1998)
Table 1.3 shows the trends in average levels of blood pressure, cholesterol, and body mass index, rather than the proportions of people who exceed particular cut-off values These data confirm the directions of the trends in Fable 1.1, although the magnitudes now appear less dramatic A likely explanation is that a large proportion of the population was either just above or just below the cut-off for high blood pressure, high cholesterol, and overweight in the third survey, a modest shift in the distribution then resulted in dramatic changes in the proportions of people who exceeded the cut-off values im the last survey In any case, the changes highlighted by the shaded boxes in Table 1.3 are still substantial
Trends in drugs for heart disease risk factors
Trang 21subsets of the surveys The fourth survey, NHANES BIT, also contains similar information on whether an individual uses any cholesterol-lowermg drugs
The National Ambulatory Medical Care Survey (NAMC), on the other hand, offers
a better picture of trends in the use of the drugs of interest The prescription numbers in Table 1.4 come from three different publications from the National Center for Health Statistics’ Advance Data From Vital and Health Statistics, each of which uses the NAMC as its data source (Koch 1982, Koch and Knapp 1987, Nelson 1993) The NAMC
conducts annual, year-long surveys of office-based medical care providers and projects its findings to national figures
The Advance Data publications provide numbers on national “drug mentions” (see note below table for definition of a “mention”) to treat hypertension and obesity, but no data on drugs to treat high cholesterol, diabetes, or smoking In Table 1.4 one can see that these two categories of drugs show markedly divergent trends during the 1980s Antihypertensive drug mentions increased significantly, particularly during the second half of the decade, while anti-obesity drug mentions plummeted The increase in
antihypertensive drugs was propelled largely by the emergences of calcium channel blockers and angiotensin converting enzyme (ACE) inhibitors, which both increased five to six fold in number of prescriptions between 1986 and 1995 During that same time, prescriptions for diuretics, which had been a popular treatment for high blood pressure for
many years, fell by about 20 % (Mulrow £998)
The increasing acceptance of blood pressure medication during the 1980s and early 1990s was reflected not only in the number of drug mentions but also in the
recommendations of the 1993 Joint National Committee on Detection, Evaluation, and
Treatment of High Blood Pressure In his JAMA review of the report, Alderman writes “Perhaps the most significant new departure in this fifth edition of the Joint National Committee Reports is reflected in the therapeutic algorithm The committee continues to recommend that imitial efforts to control blood pressure should be through lifestyle modification However, for that large majority of hypertensive patients in whom
Trang 22
Chapter I: Are we substituting between heart disease risks? 10
behavioral change does not produce a sustained and significant drop in pressure, antihypertensive drugs are indicated.” Later in his review, Alderman notes “It was also agreed that side effects, while different for the classes, were, for most patients, not 2 persistent problem” (Alderman 1993) Furthermore, the blood pressure lowering drugs’ popularity was bolstered by studies which validated their effectiveness For example, a
1993 meta-analysis of clinical trials and observational studies estimated that drug therapy for mild to moderate hypertension decreases the risk of coronary heart disease by 16% (95% CI spans 8 to 23%) (Hebert et al 1993)
Cholesterol-reducing drugs also progressed rapidly during the 1980s An article appearing in JAMA in 1990 by Wysowski et al (1990), which derives its numbers from IMS America Ltd.’s National Prescription Audit, details the rising use of cholesterol lowering drugs, as seen in Table 1.5 The number of prescriptions fell between 1978 and
1983, but then exploded between 1983 and 1988 The article also highlights the dramatic shift in which cholesterol lowering drugs were being used In 1978, clofibrate dominated the cholesterol lowering drug market with 80.9 % of prescriptions, and dextrothyroxine was second at 8.9% By 1988, these two drugs only accounted for 3.5 and 0.4 % of prescriptions respectively, while the two leaders were lovastatin at 29.7 % and gemfibrozil at 29.4% Lovastatin was introduced to the market in 1987 and shot to the top in only one year, while gemfibrozil was introduced in 1982
These newer drugs have generally been shown to be both effective and free of adverse side effects According to the 1993 Expert Panel on Detection, Evaluation, and Treatment of High Blood Cholesterol in Adults, statins are highly effective in lowering LDL cholesterol (“bad cholesterok”) levels, and are relatively safe A meta-analysis of all published trial testing statin drugs between £985 and 1995 found the following: an average
of 22% decrease in total cholesterol level and 30% in LDL level; 2 29% reduction in risk
Trang 23antihypertensive drugs, the emergence of new drugs, and of cholesterol reducing drugs in general, clearly suggest that the 1980s was a decade of great innovation for this class of drugs
More recent data suggest that prescriptions for blood pressure lowering drugs and cholesterol drugs have continued to increase steadily in the 1990s, while anti-obesity drug prescriptions remained stable during the early 1990s and then rose precipitously in 1995 and 1996” These data, obtained from National Data Corporation (NDC) Health Information Services, are presented in Table 1.6 NDC collects data from participating pharmacies and then projects figures out to the known U.S pharmacy universe The data in Table 1.5 and 1.6 are not directly comparable to the data in Table 1.4, since drug mentions in Table 1.4 are compiled from doctors’ responses to surveys, whereas the prescriptions in Table 1.5 and 1.6 are compiled directly from the pharmacies Also, some drugs which are sometimes but not always used to lower blood pressure may have been counted in one survey as blood pressuring lowering, but not in the other Due to these inconsistencies, it is more instructive to examine the trends within each data source rather than trying to construct trends by bridging them
The NDC data also breaks down the blood pressure lowering and cholesterol lowering drugs into sub-classes (not shown here) The leading blood pressure lowering drugs changed little from 1991 to 1997, with diuretics, calcrum channel blockers, ACE inhibitors, and beta blockers being the top four in each of the years Meanwhile, the cholesterol lowering drugs are classified as statins and “others.” Statins continued their emerging dominance from the 1980s, with approximately 50 % of prescriptions in 1991 and over 80 % of prescriptions in 997
Do risk factor trends and drug trends coincide?
Trang 24Chapter I: Are we substituting between heart disease risks? 12 in Table 1.6, we can at least note that a great emergence of blood pressure lowering and cholesterol lowering drugs coincided roughly with considerable decreases in the
proportion of people with high blood pressure and high cholesterol, as well as an increase im the proportion of overweight people Each of these events occurred during the early and mid 1980s Furthermore, there was no sustained emergence of effective weight lowering drugs These observations conform to the basic hypothesis posited by this study, but of course cannot be considered evidence In the next section I elaborate on the theory behind the hypothesis, and then in the following sections I attempt to test it empirically IH: Economic Theory
How we manage our heart disease risks is not only a biological issue but also a distinctly economic one We can devote various limited resources, such as our time, energy, and money, to controlling or lowering our risks Every day we make decisions, whether they are conscious or not, about how much of these resources to allocate and where to allocate them Do we take half an hour to go for ajog? Do we take the time and spend the money to see the doctor for a checkup? Another cost we bear when we manage our heart disease risk is the sacrifice involved from abstaining from certain pleasures, such as eating tasty but fatty foods
Certainly not every action aimed at controlling heart disease risk involves only costs Many people enjoy exercise, for example, and derive benefits besides the lowering of their heart disease risk However, for the purpose of the model outlined below, the important point is that most decisions related to heart disease risk involve definite tradeoffs, and to some extent these tradeoffs can be understood and analyzed in an economic framework
One possible economic model of people’s behavior with respect to heart disease risk focuses on the tradeoff between the probability of avoiding heart disease and a risky lifestyle (Figure 1.2) In this model, both concepts are “goods” which enter positively into the consumer’s utility function A “risky lifestyle” would include eating high-fat and high- cholesterol foods, smoking, and avoiding exercise, for example The straight line AB
Trang 25
possible, and thus the highest probability of avoiding heart disease By contrast, at point B the person has the riskiest lifestyle possible, and thus the lowest probability of avoiding heart disease The person maximizes utility, given the budget constraint, at point Ì
Such a model is analagous to the model outlined by Sam Peltzman in his 1975 Journal of Political Economy article, “The Effects of Automobile Safety Regulation.”
Peltzman argues that for automobile drivers, “driving intensity” is a good and “probability of death” is a bad, and there is a positive correlation between the two Thus, if a driver wants to consume more driving intensity (which refers to speed, most notably), then he or she must assume an increased risk of death In the heart disease risk model in Figure 1.2, a risky lifestyle is analagous to driving intensity in Peltzman’s model, and the probability of avoiding heart disease is analagous to the probability of death except that the former has been transformed from a bad to 2 good by inserting the word “avoiding.” This departure from Peltzman’s model seems preferable for the simple reason that the two good
framework is a more familiar one than the one good, one bad framework, and can be understood more easily in terms of tradeoffs
To evaluate the effect of the availability of new drugs that lower blood pressure and cholesterol more effectively per cost to the consumer than previous drugs, [ consider what happens to a person’s consumption behavior in the context of Figure 1.2 With the new drugs, the person’s tradeoff between the probability of avoiding heart disease and riskiness of lifestyle becomes less steep, and he faces the new budget constraint AC This constraint applies to any person for whom the new drugs are available, and reflects the fact that for any amount of riskiness in lifestyle, his probability of avoiding heart disease has increased In essence the price of a risky lifestyle has fallen, because adding a given amount of lifestyle change (e.g consuming more fatty foods) results in less additional heart disease risk than before
Trang 26Chapter 1: Are we substituting between heart disease risks? 14 lifestyle has essentially become cheaper, so the person “consumes” more of it Next, the nature of the income effect depends on whether or not we assume the goods m the model are normal goods If we assume that they are normal goods, then the income effect will cause the person to consume more of both goods, and the person will end up on a point within the dark segment on AC Thus, we can determine in this model that the use of new, more effective drugs increases the riskiness of lifestyle, but whether or not the probability of avoiding heart disease rises or falls is ambiguous
One could certainly criticize the above model for being too simplified For
example, it does not consider explicitly how the person’s management of heart disease risk through drugs fits into his overall budget constraint and consumption of other goods However, for the purposes of this model, the important point is that the person’s consumption of heart disease risk management through drugs increases with the
introduction of more cost-effective drugs What happens to consumption of other goods does not influence this relationship
A more serious simplification in the model is that it does not consider what happens if someone does contract heart disease Heart disease ranges widely in terms of its gravity; persistent uncomfortable chest pain and a fatal heart attack are both considered heart disease, for example In the context of the model presented in Figure 1.2, one can see how this issue could affect the analysis by comparing points 3 and 4 The two points represent the same probability of avoiding heart disease, but the probabilities result from different behaviors Point 3 involves a less risky lifestyle and less devotion of resources to heart disease risk management through drugs, than point 4, which involves a very risky lifestyle and more resources to control risk One might guess that the person at point 4 would generally experience more serious consequences, given that heart disease occurs, than the person at point 3, because at point 4 he is relying on medications rather than a healthy lifestyle to control the risk For now, however, this idea remains speculation until empirical medical evidence passes judgment In any case, it should not affect the
qualitative predictions based on the model; we still should expect the change in riskiness of lifestyle to be positive and the change in probability of avoiding heart disease to be
Trang 27disease risk factors If, for example, people do not understand that both high blood pressure and being overweight are risk factors, then it is unlikely that they will behave in response to the availability of effective blood pressure lowering drugs in the manner predicted in the model Fortunately there is some empirical evidence that shows that at least a significant minority of the adult population is aware of the major “controllable” risk factors that are prominent in this study Kirkland et al analyzed a survey of Canadian adults aged 55-74 who were asked to name risk factors for heart disease in 1986 and 1992
and found the following: 23% listed high blood cholesterol; 16% listed hypertension; 42%
listed smoking; and 31% listed overweight Also, Mosca et al conducted a similar survey of American women over the age of 25 in 1997, and found that about 30% listed high blood cholesterol, smoking, and overweight, and about 15% listed high blood pressure While these numbers are not overwhelming, they suggest that the behavioral response hypothesized in this study could occur for at least a substantial fraction of the population
The key prediction in this study is that riskiness of lifestyle with respect to heart disease should increase with the introduction of more cost-effective drugs to lower heart disease risk More specifically, the hypothesis is that the introductions during the 1980s of new drugs to lower blood pressure and cholesterol were partly responsible for the
increasing proportion of overweight people in the American adult population In addition we would expect that this substitution effect was responsible for keeping smoking rates from falling more than they actually did As for blood pressure and cholesterol, the theory presented above does not make a clear prediction regarding the effect of the new drugs; the benefits of the drugs and the substitution towards riskier behaviors oppose each other, and the net effect is ambiguous In the following section I describe my strategy for examining empirically the impact of the new drugs on each of the risk factor conditions IV: Empirical strategy
Trang 28Chapter 1: Are we substituting between heart disease risks? 16
or whether a person smokes I begin the empirical analysis with this approach, using a regression equation of the following form (for “overweight,” for example):
(1) overweight; = Bo + B; X; + Bz (med; } + e;
The variable “overweight” is dummy for whether the person is overweight (as
defined in Figure I.1), X represents socioeconomic variables, and (med) is a dummy
variable for whether a person is taking medication to lower blood pressure (cholesterol) In these regressions the hypothesis of this study suggests that people who have access to effective medications to lower blood pressure or cholesterol! will control their weights less and smoke more, and thus the coefficient on the (med) dummy variable should be positive and significant for these behaviors Of course, one cannot necessarily interpret such results as evidence that the medications cause people to control their weight less The direction of causality undoubtedly runs the other way as well, people who are more overweight tend to use the medications more frequently, because the lowering of their blood pressure and cholesterol is more urgent Thus the regressions equation above can establish correlations, but not unambiguous directions of causality
In the theory of this study, the force which drives the substitution towards riskier lifestyles is the availability of effective medications, as opposed to the actual use of them The use of medications is highly correlated with availability, but as a RHS variable it suffers from obvious endogeneity problems, as mentioned above A better proxy for availability of medications is insurance status It still represents a reasonable measure of the availability of medications for 2 person, and furthermore it is less likely to be
Trang 29that for health measures that are not specific to heart disease risk Essentially this is a “diff-in-diff-in-diff” analysis, with the three dimensions of comparison being insurance status (yes or no), time (1971-75 versus 1988-94), and health measure (overweight (smoking) versus other measures)
In the regression framework for this strategy | combine the NHANES |! and the NHANES 3 samples, and estimate the following equation (with BMI as the example health measure):
(2) overweight, = Bo + B, X; + Bz (NHANES3;, ) + Bs (ins; ) + Bg (ins; * NHANES3;) + e;
NHANES is a dummy variable which is equal to one for observations from the NHANES 3 survey (post-drugs), and zero for observations from NHANES | (pre-drugs) The variable “ins” is a dummy for whether the person has any health insurance’ In this equation, the coefficient B, represents the change over time in the health differential across insurance status, or the “diff-in-diff.” The third differential is then a comparison of this coefficient for the overweight equation to the corresponding coefficient for other health measures, such as a self-reported general health condition
The purpose of the “diff-in-diff-in-diff’” approach used here is to isolate as much as possible the relationship of interest, that between the availability of effective blood
pressure and cholesterol medications and riskier lifestyles with respect to heart disease Figure 1.3 helps illustrate how the approach accomplishes this goal The vertical axis measures the differential in health (by a particular measure) between insured people and uninsured people Given that insurance status proxies availability of drugs, why not simply look at this differential for overweight (smoking) in the cross-section sample taken after the influx of the drugs (NHANES 3)? A likely problem with this tack would be that the msured population differs from the uninsured population in unobservable ways which influence health outcomes such as overweight For example, people who are insured
Trang 30Chapter I: Are we substituting between heart disease risks? 18 perhaps tend to be more health-conscious than uninsured people, even after I control for observable characteristics such as race, income, and education Thus if I just look at the NHANES 3 sample, I might see that the health differential in terms of overweight between insured and uninsured people is positive (i.e insured people are “heathier,” or less
overweight), even if the hypothesis of this study were true The effect I am looking for would be offset by the unobserved differences in people across insurance status
I can address this problem by comparing the overweight differential across
insurance status for people before the arrival of effective drugs to the differential after the arrival of effective drugs, yielding a “diff-in-diff.” If the unobservable differences affecting health outcomes between insured and uninsured people are constant over this time span, then the “diff-in-diff” would isolate the effect of interest We would expect that the health differential, in terms of overweight, between insured and uninsured people would be worse in the post-drug period than in the pre-drug period This situation is depicted in Figure
L.3 by the negatively sloped line for “health in terms of overweight.”
With the “diff-in-diff’ approach described thus far, there remains the concern that unobserved health-related differences between insured and uninsured people have in fact changed over time For example, perhaps sured people have become relatively more health conscious than their uninsured counterparts over time, and these changes cannot be explained by observable characteristics I look at a third differential in order to address this possibility: the difference in the “diff-in-diff” for health in terms of overweight (smoking) and the “diff-in-diff” for health measures nonspecific to heart disease risk There are two reasons why [ expect the former “diff-in-diff” to be more negative than the latter, if the hypothesis of this study is true First, the time span for this analysis was one of technological advancement which was exceptional for medications to treat high blood pressure and cholesterol”, even in comparison to general advances in health care Second, overweight and smoking are examples of health measures that are far more “controllable” than most health measures; thus, the substitution towards riskier lifestyles could not occur to the same extent for many other facets of health For these reasons I look at the third
Trang 31nonspecific to heart disease risk, as illustrated in Figure 1.3 V: Results
Results for equation | using probit estimations for each of the different NHES and NHANES surveys are shown in Fable 1.7 As predicted, it appears that people who take medications for high blood pressure are more likely to be overweight, controlling for socioeconomic characteristics In the surveys taking such medications predicts between a 39 and 49 % greater chance of bemg overweight In contrast taking a cholesterol
lowering medication does not correspond to a significant difference in the probability of being overweight’
As noted in the previous section, estimation of equation | is unlikely to reflect a true causal relationship Results using equation 2 are more appropriate for evaluating the hypothesis of this study Table 1.8 shows results from estimating equation 2 using nine different health-related dependent variables The coefficients for the interaction between health insurance status and the NHANES 3 dummy (in the grey shaded row) correspond to “diff-in-diff” estimates Then, the “diff-in-diff-in-diffs” are obtained by comparing these coefficients for different health measures
In the first column the coefficient for the “overweight” regression has the
following interpretation: the overweight differential for the insured versus the uninsured increases by 14 % (in terms of probability) between NHANES | (pre-drugs) and
NHANES 3 (post-drugs) This “diff-in-diff” result is consistent with the hypothesis of the study The relative health of the insured versus the uninsured, in terms of overweight, becomes worse after the effective blood pressure and cholesterol drugs arrive The second column shows a similar result for the regression using the dependent variable “obese,” which is a category reserved for more extreme cases of overweight
In the third column we see that relative outcomes for the imsured also worsens over time for high blood pressure This fact might seem surprising considering the influx
Trang 32Chapter I: Are we substituting between heart disease risks? 20 of effective blood pressure lowering drugs However, there are a couple ways in which this result can be reconciled with the predictions of this study First, people’s substitution towards riskier lifestyles with respect to heart disease could have more than offset the improvements due to the new drugs For example, increases in overweight generally result in increases in blood pressure Recall from the discussion of Figure 1.2 in Section III that the predicted change of overall heart disease risk is ambiguous, with the
introduction of more effective drugs Another factor is that improvements in heart disease treatments would cause people to allow their blood pressures to become higher, in much the same way that improvements im blood pressure lowering drugs would cause people to allow their weights to become higher
Why do we not observe a similar worsening of relative outcomes for the insured in the case of high cholesterol (column 4)? One possibility is that the cholesterol reducing drugs were generally more effective than the blood pressure reducing ones Another possibility is that cholesterol is less easily modified by behavior than blood pressure or overweight, in that case we would not observe the same magnitude of behavioral response Next, for smoking (column 5) the coefficient for the “diff-in-diff” is positive, but is only significantly different from zero at the 10% level (column 5) Like overweight, smoking ts clearly a health measure that can be modified, although addiction often makes quitting very difficult
One could summarize the first five columns of Table 1.8 by saying that for each of the four controllable risk factors (overweight, high blood pressure, high cholesterol, smoking) insured people are relatively healthier than uninsured people in NHANES I, but this differential disappears by NHANES 3 These changes are statistically significant at the 10% level for all but high cholesterol The next question is whether this pattern holds for health measures that are not specific to heart disease risk If so, then there is cause for concern that the patterns estimated for heart disease risk factors are simply byproducts of a general trend in unobservable health-related differences between the insured and
Trang 33I survey, and there is no statistically significant change in that differential between NHANES | and NHANES 3 Thus, comparison of this result with those for the heart disease risk factors (the “diff-in-diff-in-diff” approach) is reassuring for the hypothesis of this study
In the remaining three columns (7-9) of Table 1.8 I look at indicators that people have been diagnosed with high blood pressure by a doctor in the past We would expect the “diff-in-diff’ coefficients to be positive for these variables under the theory of this study, insured people would become relatively more likely to reach dangerous levels of blood pressure with the availability of an effective drug “safety net.” The estimated coefficients are indeed positive, but not statistically different from zero Ít is possible that changes over time in the relative amounts and types of contact with doctors across insurance status could confound the relationships I am trying to estimate in columns 7-9
In the results of Table 1.8 one could raise the concern that the estimated “diff-in- diffs” for heart disease risk factors could be related to a combination of the following two factors: differences between characteristics of the insured population and the uninsured population, and a change over time in the relationship between these characteristics and the risk factors For example, suppose the insured population is more educated and has higher income on average.* Furthermore, suppose more education and income generate lower levels of risk factors in the earlier sample (NHANES 1), but they are less important factors in the later sample (NHANES 3), perhaps because information about heart disease risks has permeated more to the less-educated and wealthy segment of society Table 1.7 appears to attest to this pattern
These two factors in tandem would provide an alternative explanation for the results in Table 1.8 I address this issue in Table 1.9 by adding interaction terms between
time and education, and time and income In this specification [ allow the effects of
education and income on the health outcomes to vary over time The results appear largely unchanged The notable differences in Table 1.9, compared to Table 1.8, are that
Trang 34Chapter 1: Are we substituting between heart disease risks? 22 the “diff-in-diff” coefficient for smoking increases and is now significant at the 5% level, and the magnitudes for the overweight, obese, high blood pressure, and high cholesterol all decline somewhat
Another way to check the validity of the results is to see if the “diff-in-diffs” and “diff-in-diff-m-diffs” consistent with the hypothesis of this study are more pronounced for certain segments of the insured population who are most likely to exhibit the predicted behavior In particular, we might expect people who are better informed about heart disease risk factors, or people for whom heart disease is a greater concern due to
exogenous characteristics, to exhibit the hypothesized behavior more strongly To study this question, [ interact the “diff-in-diff” term with a high school grad dummy (Table
1.10), a college grad dummy (Table 1.11), a “over 45 year old male” dummy (Table 1.12), and a “over 45 year old male high school grad” dummy (Table 1.13) The high school grad and college grad variables represent people who are likely to be better informed about heart disease risk factors, the “over 45 year old male” dummy represents people who are especially vulnerable to heart disease for exogenous reasons’, and the “over 45 year old male high school grad” variable represents a combination of these two factors
In the first two columns of the the bottom row of Table 1.10, the positive and significant coefficients indicate that high school grads do indeed exhibit higher “diff-in- diffs” for overweight and obesity than non-high school grads Meanwhile these people show an improvement in the “diff-in-diff” for general health (column 6) Taken together these results are strongly suggestive of this study’s hypothesis However, the results are opposite for smoking, which has a significant negative coefficient in the bottom row Thus the results are not consistent with the study’s hypothesis for smoking
For college graduates (Table 1.11) the results are not at all supportive of the hypothesis The mteraction term in the bottom row has negative and significant
coefficients for overweight, obese, and high cholesterol Evidently the insured and highly educated population changed in some significant, unobservable way, or else their
environment related to these health measures did_
Trang 35
are at least 45, provide some mixed evidence For 45+ men (Table 1.12), the “diff-in-diff” for smoking is larger than it is for other people By comparison, the trend in general health is the same as it is for others However, the coefficients for overweight and obese are not significantly different from zero, and those for high blood pressure and high cholesterol are negative For 45+ high school grad men (Table 1.13), the results are more consistent with the study’s hypothesis, m the case of overweight and obesity The
coefficients for the bottom row term are positive and significant However, these results might be related to a negative change in general health for this population (column 6) V: Conclusion
In Section II, I derived a simple theory of how people respond to the availability of new, more cost-effective drugs to lower blood pressure and cholesterol, where “cost” incorporates not only monetary price but also adverse side effects The theory suggests that people should shift to lifestyles that are riskier with respect to heart disease, by controlling their weight less and smoking more, for example This theory could help explain why the prevalences of high blood pressure and high cholesterol fell substantially starting in the late 1970s, while the prevalences of overweight and obesity rose
substantially The aggregate data in Section I demonstrate that these trends were tightly connected temporally
Trang 36Chapter I: Are we substituting between heart disease risks? 24 Assuming that people are im fact raising their heart disease risks through behavior in response to drug innovations, as suggested by the hypothesis of this study, what are the implications for policymakers, and for doctors and patients for that matter? People are merely maximizing their own utilities by choosing to engage in riskier lifestyles, so why should we worry how they respond to the new drugs? One source of concern is that patients, and even doctors, do not understand enough about how risk factors combine to result m a particular heart disease outcome For example, perhaps, as suggested in Section II, having an equal probability of heart disease as someone else does not mean that the expected outcomes will be of the same gravity That is, person A, who is overweight and using medication to control blood pressure, might have the same probability of heart disease as person B, who is not overweight and uses no medication, but person A could be at risk for more serious cases of heart disease People tend to focus on numbers such as blood pressure level, but a certain blood pressure level achieved through a healthy diet is most likely better than that same level achieved through an unhealthy diet and a medication to lower blood pressure This issues deserves attention from medical researchers, because there is a good possibility that people are making tradeoffs at “exchange rates” that are suboptimal for their bodies Until we gain further understanding of heart disease risk, physicians should emphasize, as they undoubtedly do in most cases, that lowering blood pressure or cholesterol through medications is not a perfect substitute for a healthy lifestyle
Another implication is that we should regard a drug’s estimated effectiveness in lowering heart disease risk in a clinical trial or other medical study with a certain amount of caution We should consider the fact that many patients might increase their riskiness of lifestyle with respect to heart disease in response to the availability of the drug, much like people started to drive more recklessly as automobile safety equipment proliferated in the early 1970s, as shown by Peltzman In cases where clinical trials carefully prescribe healthy lifestyles and diets, the real impact on health of a new wonder drug on heart disease risk will likely not be as great as the trial suggests On the flip side of this issue, if
Trang 37utility, if not health, benefits (through substitution to riskier, higher utility lifestyles) An important step for future study will be to quantify in 2 more precise and direct way the extent to which the behavioral response to new drugs exists A panel data set which extends for a period of several years would be ideal for this purpose One could essentially compare how people behave after beginning to take medications to lower heart disease risk, versus how similar people who do not take the medications behave
Assuming that the drug interventions are not part of a randomized trial, some instruments, such as insurance status, would be necessary to account for the endogeneity of drug treatment Differences in government surance coverage for drugs across time and geographic areas would probably address endogeneity concerns even more effectively Another possibility would be to model carefully people’s behaviors related to heart disease risk, choose reasonable preference parameters, and obtain quantitative data on the
Trang 38Chapter I: Are we substitoting between heart disease risk factors? 26
Table £.1: Trends in major “controllable” heart disease risk factors for U.S adults
Risk factors 1960-1962 LSTI-1975 1976-980 1988-1994
High blood pressure 3LE 40.6 354 178
High cholesterol level 31.9 28.1 26.8 17.9
Overweight 26.7 26.9 24.8 34.2
Has diabetes L9 3.4 3.2 5.3
Has smoked 100+ cigarettes wa 59.1 58.6 53.3
Sousces: NHES, NHANES I, I and HE
Note: Definitions of the risk factors are noted in the notes befow Figure L, which is 2 scatter plot graph of the data in Table 3.1
Fable 1.2: Trends in obesity prevalence among American adults
1960-1962 1971-1975 1976-1980 1988-1994 Class I (BMI 30.0-34.9) 96% 10.1 10.1 14.4 Class 1 (BMI 35.0-39.9) 24 28 3.1 5.2
Class IIE (BMI >=40.0) 0.8 13 L3 2.9
Source: Flegal KM, Carroll MD, Kuczmarski RJ Johnson CL “Overweight and obesity in the United States: prevalence and trends,
1960-1994.” Internanonal Journal of Obesity Related Matabolic Disorders 1998 Jan-22( 1)-39-47
Table 1.3: Trends in heart disease risk factors, cont’d
1960-1962 ESTE-1975 [976-1980 1988-1994
Systolic blood pressure 131 132 126 t2L
Diastolic blood pressure 79 82 $I 72
Total cholesterol level 221 255 213 203
Body mass index 25.3 25.3 25.2 26.4
Trang 39
Table 1.4: Drug Mentions (in thousands) From Office Visits in the U.S
1980 1985 1990
Drugs to treat hypertension 38,463 39,011 47,309
Drugs to treat obesity 13,554 3,470 2,926
Note: A “drug mention” is defined by the ordering or providing of 2 drug by 2 physician for a patient as a resuit of an office visit Non-
Table £.5: Dispensed Prescriptions (in thousands) in the United States
1978 1983 1988
Drugs to lower cholesterol 439% 249 12,900
Source: Wysowski ct al, “Prescnbed Use of Cholesterol Lowering Drugs in the United States.” JAMA April 25, 1990; Volume 263(16):
2185-2188 (13)
Table 1.6: Dispensed Prescriptions (in thousands) in the United States
Type of drug — 1991 1992 1993 1994 1995 1996 1997 Blood pressure 241,749 247,781 253.326 270,261 288.957 308.638 326495 lowering Cholesterol 21,844 24,254 26,353 30,316 = 35,355 42,968 54,980 lowering Anti-obesity 4,546 4,329 4380 4.616 7,939 22,639 19,454
Trang 40
Chapter I: Are we substituting between heart disease risk factors? 28
Table 1.7: Regressing overweight on whether taking blood pressure (cholesterol} med
model: I 2 3 4 5
- NHES NHANES | NHANES 2 NHANES 3 NHANES 3 (1988)
7 ( 1960-62) (1971-75) (1976-80) (1988-94) 94)
method: OLS OLS OLS OLS OLS
|# obs : 5,955 2,959 LL957 15,203 15,203
dep var: overweight overweight overweight overweight overweight
mean of dep var: 0.265 0.278 0.271 0.353 0.353
female 0.0147" 0.0080" 0.0074" 0.0071* 0.0103* (0.0013) (0.0020) (0.0008) (0.0007) (0.0007) jage 0.0823* 0.0195 0.1375* 0.2121* 0.2231* (0.0366) (0.0508) (0.0255) (0.0212) (0.0211) nonwhite 0.1422* 0.1886* 0.0763* 0.1254" 0.1491* (0.0546) (0.0824) (0.0313) (0.0229) (0.0228)
education grade level -0.0253* -0.0302* -0.0354° -0.0127% -0.0123*