3.1 Smoking and other outcome variables

In each of the survey periods, students were asked about past smoking and other drugs’ behaviour. Specifically, respondents were asked ‘within the last 30 days, on how many days did you use cigarettes, cigars, etc?’. Possible responses were ‘never used’, ‘have used, but not in the last 30 days’, ‘1–2 days’, ‘3–5 days’, ‘6–9 days’, ‘10–19 days’, 20–29 days’ and ‘used daily’. For each category of cigarette, cigar and other drugs, two dichotomous indicators were created as measures of participation: one for ever participated and another for participated only in the past 30 days. For example, if a student has ever smoked cigarette, then the variable takes the value of 1 if he or she indicates ‘have used [cigarettes], but not in the last 30 days’ or if such a person has used cigarettes at least 1 day in the past 30 days. On the other hand, if the student has smoked in the past 30 days, the dummy variable is given the value of 1 for the use of cigarettes at least 1 day in the past 30 days; otherwise, the value of this variable is 0. Following DeCicca et al. (Reference DeCicca, Kenkel, Mathios, Shin and Lim2008), a measure of unconditional demand is also constructed by using the mid-point of each of the above ranges of days during which the student used the substance as follows: a value of 0 indicates that the student did not smoke on a day in the past 30 days or never used cigarettes, while 1.5, 4, 7.5, 14.5, 24.5 and 30.5 represent the average number of days smoked in the past 30 days.Footnote ⁴

The list of outcome variables also includes perception of students’ participation in and demand for smoking. Our measure of the perception of students’ participation in smoking is based on the following question: ‘Within the last 30 days, what percent of students at your school used [cigarettes or cigars as the case may be]: State your best estimate’. We assess the perception of students’ demand for smoking from the question, which asked ‘Within the last 30 days, how often do you think the typical student at your school used [cigarettes, cigars]: (State your best estimate…)’. In a fashion similar to DeCicca et al. (Reference DeCicca, Kenkel, Mathios, Shin and Lim2008), we create a variable to take the value of 0 if the respondent thinks a typical student may have smoked, but not in the last 30 days; the mid-point value for each of the following estimated range of the number of days a typical student may have used cigarettes/cigars: 1–2 days, 3–5 days, 6–9 days, 10–19 days, 20–29 days; and a value of 30.5 for those students who are perceived to have smoked daily.

This measure is important for outcomes such as cigarette smoking where it is well known that survey respondents are more likely to under-report the true extent of their smoking (Hatziandreu et al., Reference Hatziandreu, Pierce, Fiore, Grise, Novotny and Davis1989; Wasserman et al., Reference Wasserman, Manning, Newhouse and Winkler1991). In addition to information on smoking and tobacco products, the survey contains other important outcomes such as academic performance (cumulative and term GPA) and being ‘healthy’, which takes the value of 1 if the student’s self-assessed health is excellent, very good or good, and 0 if otherwise. Data on e-cigarettes are excluded in the cigarette smoking outcomes, since this information was first collected in spring 2016 and the resolution supporting smoking in designated areas on campus does not apply to e-cigarettes.

3.2 Control variables

The WKU ACHA-NCHA provides standard individual-level socio-economic and demographic characteristics to create independent variables that are important correlates of youth smoking behaviour. Our included covariates are designed to be strictly exogenous since any variable that is highly correlated with the policy and smoking may induce bias in the estimates. These variables include age, age squared, dichotomous variables for gender (female), race (Black, Asian, Native American, Bi-racial, other race), ethnicity (Hispanic of any race), on campus residence, membership in a Greek-letter organisation, employed, student classification (freshman, sophomore, junior), health knowledge and self-described overweight, and continuous variables for depression index and the number of drinks consumed when partying or socialising.Footnote ⁵

After excluding data with missing cases, the sample is reduced to approximately 2283 students, with 932 of them being exposed to the smoking policy. The sample is defined and summarised for all students, as well as for students pre- and post-policy according to outcome variables (Table 1) and covariates (Table 2). If the means are statistically different before and after policy, at least one asterisk (*) is placed next to it. With respect to the outcome variables, students spend fewer days on average consuming cigarettes after the policy was introduced. Additionally, students are less likely to participate in smoking cigarette, cigars and hookah. They are also likely to have higher term and cumulative GPAs since the policy began. In terms of the explanatory variables, students who are exposed to the policy are older on average and more likely to be females, white and non-Hispanic. Since the policy began, students are less likely to live on campus but more likely to be part of Greek-letter organisations and sophomores. Students’ consumption of alcohol since the policy began also significantly declined.

Table 1

Means of dependent variables on smoking, health and academic outcomes before and after smoking policy

Note: Standard deviations in parentheses.

Significance level *** indicates statistically significant difference from the mean value in the pre-policy period at p<0.1, p<0.05 and p<0.01, respectively.

Table 2

Means of control variables before and after smoking policy

Note: Standard deviations in parentheses.

Significance levels *, ** and *** indicate statistically significant difference from the mean value in the pre-policy period at p<0.1, p<0.05 and p<0.01, respectively.

3.3 Empirical strategy

In assessing the effects of a university smoking policy on smoking behaviour, health and academic outcomes, we estimate variants of the equation:

(1) $$Y_{{it}} \;{\equals}\;\beta _{0} {\plus}\beta _{1} \,Policy_{t} {\plus}{\bf{X}} _{{it}} \Omega {\plus}{\mib {\varepsilon}} _{{it,}} $$

where subscripts denote individual i in time period t. The variable Y _it denotes one of several smoking behaviours, as well as health and academic outcomes, as described above. Policy _t , our key variable of interest, is a dichotomous variable that equals 1 if a smoking policy is in place in survey period t, and 0 otherwise. Since the resolution to support a limited smoking campus was passed in 2015, Policy _t takes the value of 1 for spring 2016 and 0 for previous periods (namely, 2012 and 2014) in the sample. We consider those who are surveyed in 2016 as the treatment group. Since the policy was implemented in 2015, those who are surveyed in 2012 and 2014 comprise the control group as they were never exposed to the programme at the time. From the full sample of approximately 2281 respondents, about 60%, or a total of 1349, came from the control group while the remaining 932 constitute the treatment group.

For the outcome variables, the coefficient β ₁ measures the impact of the policy on smoking (e.g. cigarette and cigars) participation, as well as the average change in the amount of cigarettes and cigars smoked in the past 30 days. The coefficient also captures the impact on self-assessed health and student academic outcomes since the introduction of the policy. The vector of controls (X _it ) comprises observable individual and socio-economic characteristics that are associated with a penchant for smoking as described above.

There are a number of empirical challenges in estimating the impact of a smoking policy on smoking behaviour, as well as health and academic outcomes. For instance, an important identifying assumption of equation (1) is that minus the limited smoking policy, there should be no differences in smoking behaviour and other outcomes for those exposed to the policy and the other students (i.e. β ₁=0). It further assumes the absence of other policies being implemented around the same time and which may have had a similar impact on the outcomes of interest. Also, as part of our robustness exercise, results of a placebo test are presented, in which the policy is switched to a period where it never occurred. Another econometric concern is the potential for the policy to be highly correlated with other variables in the model, resulting in spurious results. Results from the variance inflation factor show no evidence of strong collinearity in the observable measures in the model. This result suggests that multicollinearity is not a problem in the model estimation.

Other identification issues may be encountered in investigating a causal effect of the policy on smoking and other outcomes related to smoking behaviour. Given the quasi-experimental design of our difference-in-difference strategy, an important concern is the possibility of biased results arising from omitted variables and unobserved heterogeneity. We attempt to control for omitted variable bias by directly controlling for factors, such as whether an individual has received health knowledge of any form from the university, and show that our results are robust after excluding those students who have received some form of health information. Even then, failure to properly account for important characteristics related the outcome and the policy can result in biased estimates in ordinary least squares (OLS) regression. An alternative estimation procedure that can address endogeneity due to unobserved characteristics is instrumental variables method, which requires finding at least one instrument that is highly correlated with the smoking policy but does not have a direct effect on students’ smoking behaviour, health and academic outcomes. An effective instrumental variable strategy often requires finding a range of policy-related variables with significant variation. Given the cross-sectional nature of the data, this is empirically challenging. One useful alternative strategy that may not suffer from these limitations is PSM, which is described below.

3.4 PSM

PSM methods, attributable to the seminal work of Rosenbaum and Rubin (Reference Rosenbaum and Rubin1983), are designed to create subsets or blocks of similar individuals based on a set of pre-treatment characteristics. Within each block, there are two groups comprising the treatment (those students who were exposed to the smoking policy) and a control (students who were not exposed to the policy). In traditional PSM techniques, each observation in the treatment unit is matched with at least one observation in the untreated or control group. The differences in outcomes between both groups are calculated for each block and then averaged to obtain an overall effect attributable to the treatment (smoking policy). This average is called the Average Treatment Effect on the Treated (ATT). There are two key advantages to using PSM over standard regression techniques. First, PSM is non-parametric in nature and therefore requires no specific functional form of the model; second, PSM relaxes the assumption of homogeneity in treatment across the two groups (Dehejia and Wahba, Reference Dehejia and Wahba1999; Zhao, Reference Zhao2005).

To obtain the ATT, we first obtain propensity scores defined as the probability of receiving treatment conditional on a set of pre-treatment covariates. That is, p(X)≡Pr(Policy=1 | X), where p(X) is the propensity score and Policy and X are as defined in equation (1). The propensity scores, which are obtained from a logit regression on a fairly large number of covariates, can be used to calculate the ATT as follows:

(2) $$\eqalignno{ &#x0026; E\left\{ {Y_{i} {\equals}0, \,Y_{i} {\equals}1\,\mid\,Policy{\equals}1} \right\} \cr &#x0026; {\equals}E\,[E\{ Y_{i} {\equals}0, \,Y_{i} {\equals}1\,\mid\,Policy{\equals}1,\,p\left( {\rm X} \right)\} {\rm }] \cr &#x0026; {\equals}E[E\{ Y_{i} {\equals}1\,\mid\,Policy{\equals}1,p\left( {\rm X} \right)\} {\minus}E\{ Y_{i} {\equals}0\,\mid\,Policy{\equals}0,\,p\left( {\rm X} \right)\} \,\mid\,Policy{\equals}1], $$

where Y _i is outcome already defined earlier and E is expectation. Both the PSM and the ATT rely on two main assumptions: (1) the assumptions of unconfoundedness (conditional independence), where the assignment to treatment is independent of the outcomes, given the covariates (i.e. Y _i=0, Y _i=1 ⊥ Policy=1, X), and (2) balancing, where observations with same propensity score follow the same distribution of observable and unobservable covariates independent of treatment status [i.e. Policy ⊥ X | p(X)]. To ensure that each treatment unit is matched to at least one corresponding control, there must be a common support that implies that the probability of assignment is bounded between 0 and 1, 0< Pr(Policy=1 | X) <1.

We employ several methods that can be used to evaluate the ATT through the matching process. Each of these matching algorithms possesses unique advantages (Mocan and Tekin, Reference Mocan and Tekin2006). Further, within finite samples there is a trade-off between quality (the efficiency of the estimates) and quantity (the extent of biasness in the estimates). In order to properly evaluate the robustness of the matching estimates, consideration is given to several matching algorithms. The following commonly employed matching algorithms were used: variants of the nearest neighbour matching method, caliper matching and radius matching with a caliper and local linear regression and kernel matching, the latter two being non-parametric in nature. Dehejia and Wahba (Reference Dehejia and Wahba2002) and Mocan and Tekin (Reference Mocan and Tekin2006) provide excellent discussions on the various methods.