Sample
A total of 999 adult respondents were recruited in Omaha, Nebraska, using random digit dialing (47.2 % with a response rate of 22.4 %) and placement of local advertisements (52.8 %) in media such as the major daily newspaper and Craigslist, in 2014. All data were collected using telephone interviews that took an average of 20 min. Those included in the study spoke English, were 18 years of age or older, were current smokers meaning that they had smoked more than 100 cigarettes in their life, [23] and smoked five or more cigarettes a day at the time of recruitment. We excluded very light smokers, i.e. current smokers who smoked less than 5 cigarettes a day, because they appear to be notably different from other smokers in relation to important smoking-related factors such as tobacco dependence, cravings to smoke before or after smoking cessation, [24] likelihood to make a quit attempt, post-cessation withdrawal symptoms, [25] and smoking motives [26]. Those who responded “never” to the following question were excluded from the study: “How often do you visit the stores in the neighborhood where you live? By stores, we mean such places as convenience stores, gas stations, grocery stores, supermarkets, drug stores, liquor stores, and tobacco stores.” Response options were 1 = never, 2 = sometimes, 3 = frequently, and 4 = always. The University of Nebraska Medical Center Institutional Review Board provided ethics approval for the study. Informed consent was obtained from each participant verbally as the data collection was done through telephone interviews.
While the study sample was not a random sample, its socio-demographic distribution was similar to the subsample of smokers in the center city of Nebraska Metropolitan Statistical Areas in the Behavioral Risk Factor Surveillance System (BRFSS) [27]. For example, the gender distribution in our sample and BRFSS were identical. The mean age was 47.8 years in our sample and 53 years in BRFSS. The percentage of Whites was 71.7 in our sample and 86.1 in BRFSS. The percentage of respondents with a high school diploma or a lower level of education was 49.9 in our sample and 46.3 in BRFSS. The median income was $22,500 in our sample and $30,000 in BRFSS.
Measurement
Latent variables
We measured exposure to POS marketing with the following three survey items, which are adapted from previous studies: [8, 28] “When you are in a store in your neighborhood, how often do you notice tobacco ads?”; “When you are in a store in your neighborhood, how often do you notice tobacco promotions such as special prices, multi-pack discounts, or free gift with purchase of cigarettes?”; and “When you are in a store in your neighborhood, how often do you notice cigarette pack displays?” Possible responses to each question were: 1 = never, 2 = rarely, 3 = sometimes, 4 = often, 5 = always. Before asking these questions, respondents were told that in the study “store” refers to convenience store, gas station, grocery store, supermarket, drug store, liquor store, tobacco store, etc. where tobacco products are sold.
We measured cravings to smoke with the following three survey items: “When you are in a store in your neighborhood that sells tobacco products, how often do you (1) feel a craving for a cigarette? (2) feel like nothing would be better than smoking a cigarette? (3) feel like all you want is a cigarette?”. The response options were: 1 = never, 2 = rarely, 3 = sometimes, 4 = often, 5 = always [29–32].
Observed variables
SID was measured with the following question: “In the last six months, has there been a time when the money you spent on cigarettes resulted in not having enough money for household essentials such as food?” [17, 21, 33].
We measured urges to buy cigarettes and unplanned purchases of cigarettes using the following two questions, respectively, which were adapted from previous studies: [8, 28]: “When you are in a store in your neighborhood, how often do you get an urge to buy cigarettes?”; and “When you are in a store in your neighborhood to shop for something other than cigarettes, how often do you decide to buy cigarettes?” Possible response options were: 1 = never, 2 = rarely, 3 = sometimes, 4 = often, 5 = always.
We included the following control variables in the analyses: Heaviness of Smoking Index (HSI), which is an indicator of nicotine dependence, [34, 35] gender, age in years, race/ethnicity, household income, education, frequency of visiting stores, and method of recruitment (random digit dialing versus other). Race was categorized as non-Hispanic White, non-Hispanic Black, Hispanic, and other. Education was categorized based on highest grade or year of school completed as follows: less than high school, high school graduate, some college, and college graduate and higher.
Statistical analysis
We used Stata v. 13 for descriptive statistics [36] and Mplus [37] to perform structural equation modeling (SEM) to address the aim of the study. SEM is a multivariate technique that estimates parameters in structural equations in order to simultaneously examine complex relationships between several independent and dependent variables and estimate direct, indirect, and total effects. SEM can accommodate latent variables with multiple indicators to isolate and remove measurement error and enhance predictive power [38]. Observations that had a missing value on any of the study variables, which constituted 6 % of the original sample, were omitted from the analysis. The analysis sample size was 939. We examined the association of missingness with sociodemographic factors and the study outcome. There was very little evidence that missingness was associated with gender (n = 999; p = 0.669), race/ethnicity (n = 996; p = 0.057), education (n = 998; p = 0.769), or SID (n = 998; p = 0.054). However, there was evidence that the mean age of the individuals in the analysis was lower than that of those not included in the analysis (n = 994; p = 0.003).
We performed SEM in two stages [39]. First, using maximum likelihood parameter estimation with standard errors and mean- and variance-adjusted (MLMV) chi-square test statistics for continuous data, [37] we estimated a measurement model involving POS cigarette marketing and cravings to smoke, each with three indicators. Next, using probit regression and robust weighted least square parameter estimation with standard errors and mean- and variance-adjusted (WLSMV) chi-square test statistic for binary outcome, [37] we estimated a structural model representing both latent (POS marketing and cravings to smoke) and observed variables of interest (urge to buy cigarettes, unplanned purchase of cigarettes, and SID) in the study and specifying the pathways connecting these variables [37]. Where appropriate, we used the modification index to estimate additional parameters to enhance the fit of a model. We used the DIFFTEST procedure, which provides a χ2 difference test, to compare nested models. We included all of the control variables as exogenous observed variables in structural equations. If the p-value for the effect of a control variable was greater or equal to 0.05 in any of the equations, it was removed from that equation. We used comparative fit index (CFI), Tucker-Lewis index (TLI), and root mean square error of approximation (RMSEA) to assess the fit of the models. A model was considered to have a good fit with the observed data if the following were true: CFI > = 0.95, TLI > =0.95, and RMSEA < = 0.05 [40, 41]. Standardized coefficients representing the direct effects were presented in the structural equation diagram. Standardized coefficients are expressed in terms of standard deviation units and as such provide a measures of the strength of association, are used as an effect size index, and allow a comparison of different effects within the same model [42]. Estimates of total and indirect effects (Muthén B: Applications of causally defined direct and indirect effects in mediation analysis using SEM in Mplus, unpublished working paper) of POS on SID were also provided. All reported coefficients are standardized regression coefficients orβs. It should be noted that, as is customary in describing the results of SEM, we use the word “effect” to describe the association between variables rather than to ascribe a causal nature to the observed pattern of associations.