2.1. NLSY97
In this study, we used National Longitudinal Survey of Youth 1997 (NLSY97) from 1997 to 2009 to identify smoking initiation/relapse and to conduct the analysis. NLSY97 is a longitudinal survey that tracks a nationally representative sample of 8984 youth who were 12–16 years old as of 31 December 1996 over time. Therefore, the age range for NLSY97 during 1997–2009 is 12–30 years. Given that respondents were interviewed on an annual basis, smoking initiation/relapse can be identified through comparing self-reported smoking status in each year. For example, initiation occurred when respondents who had never smoked by the previous year’s survey reported to have smoked in the past 30 days in this year’s survey. Moreover, smoking status is reported both in frequency (During the past 30 days, on how many days did you smoke a cigarette?) and in intensity (When you smoked a cigarette during the past 30 days, how many cigarettes did you usually smoke each day?), which allows us to categorize smoking initiation/relapse into four distinguishable stages: (1) transition into non-daily smoking; (2) transition into daily smoking; (3) transition into light smoking(<10 cigarettes per day); and (4) transition into heavy smoking (≥10 cigarettes per day).
In addition to smoking status, NLSY contains demographics of respondents and tracks their socioeconomic characteristics over years. The demographic and socioeconomics controls we used in the analysis are gender, race/ethnicity (non-Hispanic White-referent, Hispanic, non-Hispanic Black, and non-Hispanic other races), employment status (worked in employee-type job since last interview vs. not), marital status (married vs. not), and education/school enrollment levels constructed using enrollment status and highest degree received (college degree or higher-referent, enrolled in college, high school degree not enrolled in college, enrolled in high school, and high school dropout). Because the effects of time-variant socioeconomic variables on smoking initiation/relapse are likely to differ by youth and young adults (e.g., being single may have very different impacts on smoking among teenagers than among young adults), we also control for interactions of these socioeconomic variables and a dichotomous variable for being younger than age 21.
Furthermore, NLSY97 Geocode file contains state of residence for each respondent in each year that allows for a linkage to state-level cigarette taxes, smoke-free policies, and other controls. After stacking NLSY97 data from 1997–2009, we dropped those who did not report smoking status or whose state of residence were missing, which combined were about 13% of the sample. This attrition rate is similar to the one reported in Aughinbaugh and Gardecki (2007) [
44]. For the missing marital or education/enrollment status, we compared the non-missing values in the preceding and following years and, if they are the same or can be inferred (e.g., missing grade between 10 grade in the preceding and 12 grade in the following year would be filled with 11 grade), we replaced missing values with the non-missing or inferred values; otherwise, we replaced the missing values using the non-missing values in the preceding year. Sensitivity analyses by dropping these missing covariates were also conducted.
2.2. State-Level SFA Laws, Cigarette Taxes, and Other Controls
The SFA laws in 13 venues, state tobacco control funds in 1982–1984 dollars, and the number of youth access laws prohibiting minors from possessing, using, and purchasing tobacco products (PUP) for each state and year during 1997–2009 were obtained from Impacteen. The SFA venues include bars, restaurants, private work places, public work places, public transit, recreational centers,
etc. For each venue, the strength of SFA laws is rated on a scale of 0–5 or 0–3 annually for each state, which leads to a total score of 49. For restaurants (including bar areas of restaurants) and bars, where 0 equates no laws, 1–2.5 equates SFA laws with exemptions such as designated smoking areas or ventilation standards, and 3 equates 100% smoke free, no exemption.
Figure 1 shows, over years, the percentages of states having any SFA laws with exemptions and states having the 100% SFA law with no exemption in bars during 1997–2009. Most states imposed SFA laws with certain exemptions in earlier years and had gradually moved towards 100% SFA law without exemptions.
Figure 1.
The Prevalence of smoke-free air (SFA) Laws in Bars in 50 States and DC.
Figure 1.
The Prevalence of smoke-free air (SFA) Laws in Bars in 50 States and DC.
In order to fully study the effect of SFA in bars, we created dichotomous indicators for different strengths of laws (none-referent, SFA laws with exemptions, and 100% SFA law with no exemption). Because the legal drinking age in the US is 21, we also included interaction terms for the two law indicators and a dichotomous indicator of being younger than 21 and controlled for them in the analyses. The SFA law in restaurants measured using a 0–3 scale is also controlled in all analyses. Although in most previous studies, SFA laws in restaurants and bars are studied/considered the same, the data show that about 46% of the states in the study period have imposed a different SFA law in bars than in restaurants. Therefore, it is important to study the effects of SFA laws in bars and in restaurants separately. In addition to SFA laws in bars and restaurants, laws in other venues were controlled using an average SFA policy index, which was constructed using the sum of the scores of the rest 11 venues divided by the highest possible total score of 43. In a separate sensitivity check, we controlled for individual SFA law scores of these 11 venues instead of the average SFA policy index.
State-level average cigarette taxes for the calendar year were obtained from the annual
Tax Burden on Tobacco [
45] and include the average federal tax of the year. We converted the taxes into 2009 dollars using the Consumer Price Index from Bureau of Labor Statistics. Finally, we merged SFA laws, state tobacco control funds, PUP index, and average cigarette taxes to the NLSY97 data using the state identifier and year.
2.3. Methodology
The Discrete-time Hazard Model is employed to analyze smoking initiation (one-puff, daily-smoking initiation, and heavy-smoking initiation) and relapse (relapse to nondaily/daily smoking and relapse to light/heavy smoking). This method has been widely used in analyzing smoking dynamics in the literature [
46,
47] and will estimate the impact of explanatory variables on the “hazard” probability of the outcome. In order to identify the model, all individuals start from a baseline status (outcome = 0) and are followed over time. Once an individual transits into the studied status (outcome = 1), the “hazard” occurs and he or she will be subsequently dropped out of sample. Those who remained at the baseline status 0 and had never made the transition by the end of the study period will be kept in the sample at all times. After the data is shaped into such format, logistic regressions are usually used for estimation. Additionally, this model can be extended to the Competing Risk Model that estimates transitions into two or more “competing” outcome statuses, and analyzed using multinomial regressions. The following equation shows the Competing Risk model with two competing or mutually exclusive outcome status (outcome = 1 or 2) that are different from baseline 0.
More specifically, smoking initiation (one-puff, daily-smoking initiation, and heavy-smoking initiation) was estimated using the Hazard model with one potential risk while smoking relapse (relapse to nondaily/daily smoking and relapse to light/heavy smoking) was estimated using the Competing Risk Model with two competing risk statuses—Those are, in smoking frequency, either relapsing into nondaily or into daily smoking; and in smoking intensity, either relapsing into light or into heavy smoking.
The study samples were obtained in the following ways: (1) for smoking initiation with even one-puff of a cigarette, we first dropped ever-smokers in the baseline year 1997, then assigned never-smokers 0 in the following years until they initiated; (2) for daily-smoking initiation, we first identified non-daily smokers or non-smokers in the baseline year 1997, then assigned non-daily smokers and non-smokers 0 in the following years until they started smoking daily; (3) for heavy-smoking initiation, we first identified light smokers or non-smokers in the baseline year 1997, assigned light smokers and non-smokers 0 in the following years until they started heavy-smoking; (4) for relapsing in smoking frequency, we first identified eve-smokers who were not smoking, took the first year when we observed them to be abstinent (either 1997 or if smokers quit during 1997–2009, the year when they quit) as the baseline year, assigned abstinent non-smokers 0, and followed them until they relapsed to smoking. Then, we coded those who relapsed to non-daily smoking as 1 and those relapsed to daily smoking as 2; 5) for relapsing in smoking intensity, the procedure is the same with the procedure for relapsing in frequency. After identifying relapsing, those who relapsed to light smoking were coded as 1 and those who relapsed to heavy smoking were coded as 2. Because some smokers may report positive smoking intensity with a 0 frequency or vice versa, the samples sizes for two relapsing samples are slightly different.
Finally, the effects of SFA laws in bars (none-referent, SFA laws with exemptions, and 100% SFA law with no exemption) on smoking initiation/relapse and their interactions with age were examined along with state fixed effects, year fixed effects, individual-level socioeconomic and demographic controls and state-level controls described in the data section. In addition, duration dependence that measures how the hazards of outcomes change over time [
46,
47] is controlled using the natural log form of age since 10. Similar to previous studies [
8,
9,
10,
11], because we selected sub-populations who were at risk of making the particular transitions, NLSY weights that adjust for non-responses become inappropriate and thus were not used. For all analyses, standard errors were clustered at the state level. Stata Special Edition 13.1 was used to implement the analyses.
In order to examine the validity of the estimates, sensitivity analyses using different specifications or samples were conducted. These include: (1) instead of year fixed effects, a linear year trend that is less collinear with SFA policies was used; (2) instead of the average SFA policy index, individual SFA policy scores that are less restrictive in modeling the impacts of other policies were estimated with a linear year trend and other covariates; (3) samples were restricted to respondents who reached the drinking age of 21, and 4, samples were restricted to respondents with non-missing covariates.