Study design and sample

We conducted a cross-sectional secondary data analysis using data from Wave (W) 5 of the Population Assessment of Tobacco and Health (PATH) Study, a national longitudinal cohort survey of US youth and adults. Data were collected in respondents’ households using computer-assisted self-interviews administered in English or Spanish, as appropriate. Survey data were collected from adults (aged ≥18 years) in 2013–2014 (W1), 2014–2015 (W2), 2015–2016 (W3), 2016–2017 (W4), and 2018–2019 (W5). Further details about the PATH Study design, methods, and instruments are publicly available²¹. The current secondary data analysis used the W5 Adult Restricted Use Files²¹ and was limited to adults aged ≥40 years at W5 (n=12126). See Supplementary file Figure 1 for a flow diagram of the study sample. This study qualified as exempt per guidelines of the Westat Institutional Review Board and the Dartmouth Health Human Research Protection Program.

Measures

Current cigarette smoking

The primary outcome was current (vs never/former) cigarette smoking at W5. We used the PATH Study-derived variables for never, current established, and former established cigarette smoking to create a two-category smoking status variable capturing never/former (coded 0) or current (coded 1) smoking. Current established cigarette smoking was defined as smoking at least 100 cigarettes in lifetime and currently smoking every day or some days. Former established cigarette smoking was defined as smoking at least 100 cigarettes in lifetime and not smoking in the past 12 months or currently smoking not at all. Never smoking was defined as smoking fewer than 100 cigarettes in lifetime.

Geographical location

The primary exposure variable was geographical location (released in July 2024 in the PATH Study Restricted Use Files)²¹, which was categorized as urban, suburban, town, or rural. Categories were assigned to individual respondents at W5 at the US Census block level using 2021 Census block definitions. Respondents’ geolocations were coded to match 12-level locale classifications prior to being collapsed into the four categories listed²¹. Additional information can be found at: https://nces.ed.gov/programs/edge/Geographic/LocaleBoundaries²².

Cigarette risk perceptions

Cigarette risk perceptions at W5 were assessed through four separate survey items. Three of the items were related to beliefs about the health risks of cigarette smoking: ‘Based on what you believe, how much do you agree or disagree with the following statements? Smoking can cause: a) lung cancer in smokers, b) lung disease (such as emphysema) in smokers, and c) lung disease in non-smokers from secondhand smoke’. Response options for all three of these items ranged from 1 (Strongly agree) to 5 (Strongly disagree). The fourth item (d) was more general: ‘How harmful do you think cigarettes are to health?’, with response options ranging from 1 (Not at all harmful) to 5 (Extremely harmful). The four individual cigarette harm perception items were combined into a 4-item scale (Cronbach’s alpha=0.78). To create the scale, items were reverse coded as necessary such that across all four items, higher values indicated lower perceptions of risk. Then, an average risk perception score across the four items was calculated for each respondent (ranging from 1 to 5).

After examining the 4-item scale in a primary multivariable regression model (Model 3), we examined each individual risk perception item in four secondary separate models (Models 3a–3d) during sensitivity analyses to determine which element(s) had the strongest association with current smoking and which impacted the geographical location association most. For sensitivity analyses, response options for each item were dichotomized into: ‘Agree’ (strongly agree/agree, coded as 0) versus ‘Not agree’ (neither/disagree/strongly disagree, coded as 1) or ‘Harmful’ (very/extremely, coded as 0) versus ‘Not harmful’ (slightly/somewhat/not at all, coded as 1) as appropriate.

Sociodemographic covariates

Sociodemographic covariates were categorized as shown in Table 1. From W5, these included age, sex, race, education level, and total household income in the past 12 months. Missing data on age, sex, and race were imputed as described in the PATH Study Restricted Use Files User Guide²¹.

Statistical analysis

Weighted descriptive statistics and Poisson regressions were used to evaluate associations between geographical location, cigarette risk perceptions, and current smoking²³. Given that current smoking is not a rare outcome, we chose to use Poisson regression with more easily interpreted risk ratios (vs logistic regressions with odds ratios) to generate conservative estimates for associations between geographical location and current smoking²². Unadjusted models and models adjusted for age, sex, race, education level, and household income were examined to estimate risk ratios (RRs) and adjusted RRs (ARRs). Separate regression models were conducted in a stepwise fashion. Model 1 examined the unadjusted association between geographical location and current smoking. Model 2 examined this same association adjusted for sociodemographic covariates. Model 3 examined ARRs for geographical location and mean cigarette risk perceptions (using the 4-item scale), also adjusted for sociodemographic covariates. Stepwise models were built based on previous empirical findings. Specifically, prior work has shown that age, gender, race, education level, and income are all independently associated with current smoking when concurrently present in a multivariable model ²³. Therefore, after examining the unadjusted association between geographical location and current smoking (Model 1), these known sociodemographic covariates were entered into an additional step (Model 2), prior to examining how accounting for cigarette risk perceptions (Model 3) would impact the association between geographical location and current smoking above and beyond covariates known to be associated with current smoking.

Respondents who had valid data on the variables used in specific regression models were included in the models. Respondents who were missing data on one or more of the variables in a specific regression model were removed from the respective model. All analyses were weighted using the W5 single-wave survey weights for the W4 cohort, which included full sample and 100 replicate weights, to produce nationally representative estimates. The weights employed in our analysis accounted for attrition and non-response bias, as well as replenishment of the sample at W4, in order to obtain estimates for the US population²¹. Variances were computed using the balanced repeated replication (BRR) method²⁴ with Fay’s adjustment set to 0.3 to increase estimate stability²⁵ using the following code in Stata: svyset [pweight=R05_A_S04WGT], brr(R05_A_S04WGT1 – R05_A_S04WGT100) vce(brr) mse fay(0.3). Extensive details on how replicate weights were constructed by the PATH Study can be found in the PATH Study Restricted Files User Guide²¹. All analyses were conducted using Stata/MP 17.0 (www.stata.com/statamp/).

Sensitivity analysis

Four additional multivariable models (Models 3a–3d) examined ARRs for geographical location and each individual risk perception item from the 4-item scale to determine whether certain scale element(s) had a greater impact on the association between geographical location and current smoking. For sensitivity analyses, response options for each item were dichotomized into ‘Agree’ (coded 0) vs ‘Not agree’ (coded 1) or ‘Harmful’ (coded 0) vs ‘Not harmful’ (coded 1), as appropriate.

Sample characteristics

Table 1 shows sociodemographic characteristics overall and by geographical location. Approximately one-third (30.0%) of the analytic sample were located in urban areas, 34.8% were in suburban areas, 9.8% were in town areas, and 25.5% were in rural areas. Participants were mainly female (53.3%), of White race (80.1%), and the mean age was 59.65 years (SD=12.30). There was a relatively even distribution of respondents across total household income levels and age categories. When comparing respondents across geographical locations, rural (vs urban) areas had a higher proportion of males (50.8% vs 44.4%) and White residents (89.9% vs 70.9%) and lower proportions of Black (6.4% vs 18.7%) or Other/Multiple race residents (3.8% vs 10.5%). Rural areas also had more respondents aged ≥60 years whereas urban areas had more respondents aged 40–49 years. Rural (vs urban) areas had fewer residents who reported an annual household income of ≥$100000 (18.0% vs 22.4%) and fewer residents who reported having a Bachelor’s or advanced degree (21.9% vs 34.9%). Respondents from urban and suburban areas were similar across sociodemographic characteristics, whereas respondents from rural and town areas were similar.

Geographical location, cigarette risk perceptions, and current smoking

Current smoking was more common among rural (20.6%) and town (20.2%) versus urban (17.6%) and suburban (14.6%) residents (Table 1). Shown in Table 2, in the unadjusted model (Model 1), rural (vs urban) residence was significantly positively associated with current (vs never/former) smoking (RR=1.17; 95% CI: 1.03–1.32) whereas suburban (vs urban) residence was significantly negatively associated with current smoking (RR=0.83; 95% CI: 0.74–0.92). The RR for rural (vs urban) residence when the model was adjusted for sociodemographic covariates (Model 2) decreased but remained statistically significant (ARR=1.14; 95% CI: 1.01–1.29) while the RR for suburban (vs urban) residence became non-significant. Further adjustment for cigarette risk perceptions (Model 3) confounded the association for rural residence (ARR=1.08; 95% CI: 0.96–1.21). Lower cigarette risk perceptions (Model 3) were also associated with significantly increased risk of current cigarette smoking, independent of geographical location and sociodemographic covariates (ARR=2.05; 95% CI: 1.94–2.18).

Sensitivity analysis

Overall, adults who currently (vs never or formerly) smoked cigarettes perceived cigarettes as being less harmful regardless of geographical location (overall mean=1.33 for never/former and 1.82 for current) (Supplementary file Table 1). Among adults who reported never/former smoking, risk perceptions were similar for those in rural (mean=1.33) and urban (mean=1.32) locations. However, among adults who reported current smoking, those in rural areas perceived cigarettes as less harmful (mean=1.91) than those in urban areas (mean=1.78). For the individual scale items, risk perceptions were similar between adults in urban and non-urban areas who reported never/former smoking. However, among adults who currently smoked, respondents in rural areas strongly agreed or agreed with beliefs about cigarette health risks less often than respondents in urban areas. The largest difference was observed for the statement ‘smoking can cause lung disease in non-smokers from secondhand smoke’, with which 73.6% of urban versus 63.4% of rural residents strongly agreed/agreed. A similar trend was observed for ‘smoking can cause lung cancer in smokers’ (90.2% vs 85.6% strongly agreed/agreed for urban and rural, respectively). However, the perception that ‘smoking can cause lung disease in smokers’ (92.3% vs 89.3% strongly agreed/agreed for urban and rural, respectively) and reporting that cigarettes are very/extremely harmful to health (69.0% vs 66.2%, respectively) were similar for urban and rural respondents (Supplementary file Table 1).

Shown in Table 3, across the four individual items of perceptions (Models 3a– 3d), two individual items confounded the rural-current smoking association whereas two did not. Specifically, believing that cigarettes are ‘Harmful’ (vs ‘Not harmful’) to health and responding ‘Agree’ (vs ‘Not agree’) to the statement that secondhand smoke causes lung disease in people who do not smoke, confounded the rural association with current smoking. Responding ‘Agree’ (vs ‘Not agree’) to statements about smoking causing lung cancer or lung disease in people who smoke, did not confound the association. In addition, the largest ARR for risk perceptions on current smoking was observed for ‘how harmful do you think cigarettes are to health?’ (ARR=3.02; 95% CI: 2.81–3.26), which was significantly larger than the ARRs for ‘smoking can cause lung cancer in smokers’ (ARR=2.46; 95% CI: 2.16–2.80), ‘smoking can cause lung disease in smokers’ (ARR=2.17; 95% CI: 1.93–2.44), and ‘smoking can cause lung disease in non-smokers’ (ARR=2.57; 95% CI: 2.39–2.77). For each of these associations, a lower perception of risk was significantly positively associated with current smoking independent of other factors.

Limitations

Our study has several strengths, including use of a nationally representative sample of a group at high risk for tobacco-related diseases and use of a newly released (in July 2024) PATH Study geographical location variable. However, this study also has limitations. First, there are other factors that put people in rural areas of the US at increased risk for tobacco use (e.g. lower tobacco excise taxes³⁴; greater social norms around tobacco³⁵), many of which were not measured or included in this study. It is therefore possible that the associations found in our study could be accounted for by unmeasured factors and residual confounding. Second, causality cannot be determined based on our cross-sectional study. Rigorous longitudinal studies that examine changes in respondents’ geographical location and how it might be associated with changes in cigarette risk perceptions or smoking status are needed. Finally, we intentionally limited the analytic sample to US adults aged ≥40 years given that this subpopulation is at greatest risk for developing smoking-related lung disease; findings may not generalize to younger populations or older adults in other countries. Future work could also consider examining interactions between geographical location and other sociodemographic characteristics (e.g. sex) to determine whether geographical location has a larger association with current smoking for specific sociodemographic groups.