Using place-based characteristics to inform FDA tobacco sales inspections: results from a multilevel propensity score model

Abstract

Background: Conducting routine inspections for compliance with age-of-sale laws is essential to reducing underage access to tobacco. We seek to develop a multilevel propensity score model (PSM) to predict retail violation of sales to minors (RVSM).

Methods: The Food and Drug Administration compliance check of tobacco retailers with minor-involved inspections from 2015 to 2019 (n=683 741) was linked with multilevel data for demographics and policies. Generalised estimating equation was used to develop the PSM using 2015-2016 data to predict the 2017 RVSM. The prediction accuracy of the PSM was validated by contrasting PSM deciles against 2018-2019 actual violation data.

Results: In 2017, 44.3% of 26 150 zip codes with ≥1 tobacco retailer had 0 FDA underage sales inspections, 11.0% had 1 inspection, 13.5% had 2-3, 15.3% had 4-9, and 15.9% had 10 or more. The likelihood of having an RVSM in 2017 was higher in zip codes with a lower number of inspections (adjusted OR (aOR)=0.988, 95% CI (0.987 to 0.990)) and penalties (aOR=0.97, 95% CI (0.95 to 0.99)) and a higher number of violations (aOR=1.07, 95% CI (1.06 to 1.08)) in the previous 2 years. Urbanicity, socioeconomic status, smoking prevalence and tobacco control policies at multilevels also predicted retail violations. Prediction accuracy was validated with zip codes with the highest 10% of the PSM 3.4 times more likely to have retail violations in 2019 than zip codes in the bottom decile.

Conclusion: The multilevel PSM predicts the RVSM with a good rank order of retail violations. The model-based approach can be used to identify hot spots of retail violations and improve the sampling plan for future inspections.

Keywords: environment; public policy; surveillance and monitoring.

Figure 1

Flow chart for sample selection in FDA tobacco retail compliance inspections and multilevel data linkage with community characteristics and tobacco control policies. a: the final analysis is limited to 50 states and the District of Columbia (DC), so inspections in US territories were removed. These territories include American Samoa (AS), Guam (GU), Northern Marianas Islands (MP), Puerto Rico (PR) and US Virgin Islands (VI). b: multilevel factors, including zip code-level past inspection results and SES along with other contextual factors, such as county-level smoking prevalence, state and local adoption of Tobacco 21, and other state tobacco control policies using geographic information system. These multilevel factors were treated as covariates (explanatory variables) in the PSM development and validation. ACS, American Community Survey; PSM, propensity score model; RUCA, Rural–Urban Commuting Area; RVSM, retail violation of sales to minors; SES, socioeconomic status; T21, Tobacco 21.

Figure 2

Multilevel propensity score model (PSM) prediction of RVSM at the zip code level (n=26 131 with ≥1 tobacco retailer): 2017^a vs 2018^b vs 2019^c. a: 2017 (PSM development): multilevel propensity score was developed where the dependent variable was RVSM (yes vs no) in 2017 and the explanatory variables included the past 2-year (2015–2016) inspection results along with other contextual factors (see equation 1). The propensity score was ranked in the order from the lowest (1, the bottom decile) to the highest (10, the top decile) in the 2017 data. b: 2018 (PSM validation): the developed PSM was prospectively validated using the 2018 FDA data where the dependent variable was RVSM (yes vs no) in 2018 and the explanatory variables included the past 2-year (2016–2017) inspection results along with other contextual factors (see equation 2). The propensity score was ranked in the order from the lowest (1, the bottom decile) to the highest (10, the top decile) in the 2018 data. c: 2019 (PSM validation): the developed PSM was prospectively validated using the 2019 FDA data where the dependent variable was RVSM (yes vs no) in 2019 and the explanatory variables included the past 2-year (2017–2018) inspection results along with other contextual factors (see equation 2). The propensity score was ranked in the order from the lowest (1, the bottom decile) to the highest (10, the top decile) in the 2019 data. FDA, Food and Drug Administration; RVSM, retail violation of sales to minors.