Variation in smoking attributable all-cause mortality across municipalities in Belgium, 2018: application of a Bayesian approach for small area estimations

Smoking is one of the main causes of preventable mortality and morbidity worldwide. According to the World Health Organization (WHO), globally 12% of all deaths among adults aged 30 years and over were attributed to tobacco [1]. Tobacco use has a proven causal relationship with a number of chronic conditions, like cancer, cardiovascular disease, chronic respiratory disease and others [2]. In the COVID-19 pandemic, current smokers who contracted coronavirus were shown twice as likely to be hospitalized and tended to report more symptoms than non-smokers [3], although some controversy remains whether smokers are less likely to get infected with this virus [3].

Despite notable successful efforts by WHO and national governments worldwide to limit tobacco use [4], smoking prevalence in European countries remains a public health concern. Among the WHO regions, Europe has the highest prevalence of tobacco smoking among adults (28% in overall population, 38% in males and 19% in females) and one of the highest prevalence of tobacco use by adolescents [5]. It is thus not surprising that, compared to the rest of the world, the WHO European Region has one of the highest proportions of deaths attributable to tobacco use. WHO has estimated that tobacco use is currently responsible for 16% of all deaths in adults over 30 in European Region, with many of these deaths occurring prematurely [5].

Smoking prevalence in Belgium is slightly lower than the EU-15 average and was estimated at 15% in 2018. This proportion is higher in men (18%) than in women (12%). On the positive side, the prevalence of daily smoking has decreased by 40% between 1997 and 2018 [6]. Substantial regional variation in smoking rates was reported with higher prevalence in Wallonia (18%), followed by Brussels (16%) and Flanders (13%) and socio-economic disparities in smoking are persistent in Belgium similar to other countries [7‐9]. Differences at lower geographical level (e.g. municipalities) may thus also exist while these have never been studied. These differences, if substantial, may be highly relevant for public health policies and prevention campaigns. Absence of municipality level statistics on smoking prevalence further hinders estimates of spatial distribution of smoking attributable mortality (i.e., the number of deaths that can be attributed to tobacco) in support of local health policy-making and understanding regional variations in mortality.

Since 1997, the Belgian Health Interview Surveys (BHIS) has been regularly collecting data on tobacco use in Belgium. The sampling design allows reliable estimates of many health indicators including smoking on national and subregional level, however, as only approximately one fourth of nearly 600 municipalities has been effectively sampled each survey round, direct estimates of smoking at municipality level are precluded. Recently, novel statistical methods have been proposed to address the growing demand for estimating small-area level indicators in situations when few or no data in some area precludes direct estimations [10]. In such cases, indirect estimates may be computed. Such estimates borrow strength from related areas through linking models based on auxiliary data such as recent census and administrative records [11]. Accounting for the complex survey design in computing these indirect estimates poses additional methodological challenges.

The aim of this study was to obtain estimates of smoking prevalence in each Belgian municipality using the BHIS 2018 data and to calculate smoking attributable mortality at municipality level using all-cause mortality statistics.

The objective of this study was to apply the Bayesian hierarchical model to obtain indirect small area estimations of smoking prevalence in Belgium and compute municipality-level smoking attributable mortality. The final estimates of current/former smoking prevalence ranged between 11 [Credible Interval 3;23] and 27% [21;34] between the municipalities. Across Belgian municipalities, smoking was assessed to be the cause of 10 [7;15] to 47% [34;59] of all deaths. This demonstrates that smoking poses a serious public health burden and, given that smoking-related deaths are avoidable, it also presents a prevention opportunity to improve quality and duration of life of many people.

To our knowledge, this is the first study to compute municipality level estimates for smoking and smoking attributable mortality in Belgium. Information about spatial distribution of these health indicators should support local policy-making as well as contribute to explaining the regional differences in mortality. The methodology used in the paper can be broadly applicable in the context of national health surveys in other countries.

While a number of small-area estimation methods have been proposed in the literature, it was out of scope of this work to provide a comprehensive comparative analysis of these methods and we rather focused on practical application of the method proposed by Mercer et al. [12]. This approach had the advantage in a sense that it is suitable for complex survey design with large amounts of off-sample areas. A crucial question is the validity of the estimates in the off-sample areas as well as in areas with very few respondents.

With respect to off-sample areas and in an absence of the true estimates in each municipality it is hard to make strong claims about the validity of the produced estimates. Existing methodological literature including simulation studies suggests this method yields valid estimations [12], however, it is reasonable to expect for its accuracy to decrease with increasing number of off-sample areas. In the current setting, data were available from 154 of 589 municipalities, which implied a lot of smoothing and extrapolation. To our knowledge, no simulation studies are available to assess the performance of this method in context where almost three quarters of the areas were off-sample (Watjou et al. have omitted up to 56% of the areas in their simulation study [10]). To improve the accuracy of estimations, auxiliary data from three municipality-level socio-economic covariates (age, gender and education) were used. These factors have been shown to be associated with smoking in the BHIS data but also from previous research [24], and thus, accounting for distribution of municipality population across age, gender and educational groups was expected to further fine-tune the estimates. In fact, only minor improvements in the fitted models were observed according to DIC which may be explained by ecological fallacy – area-level covariates may fail to pick up the existing associations with the outcome on individual level. Incorporating individual level covariates in the model would likely improve the predictions; however, more methodological research is needed to accommodate challenges related to complex survey data (e.g. how to incorporate the individual weights when part of areas is off-sample) and missing data (e.g. missing outcome or covariate data in survey data and in off-sample areas) in this context. In current analyses, provided estimates of smoking prevalence are likely conservative given lower educated respondents (how are known to be more frequently smokers) were more likely to leave the question on smoking blank. For missing survey data multiple imputation (MI) could provide a remedy under assumption of ‘missing at random’ (MAR) [25]. Smoothed estimates method should be further developed to accommodate the theoretical and practical implementation of MI in this context. Last but not least, given the increasing policy relevance of small area estimates, future rounds of the BHIS survey may reconsider the tradeoff between increasing the number of in-sample areas, which improve possibilities of small area analyses, and the cost implications of such design changes.

With respect to areas with very few respondents, model diagnostics revealed substantial smoothing towards the mean prevalence. Where direct smoking prevalence estimates (based on 2 to 249 respondents) ranged between – likely unplausible – 0 and 100%, the smoothed indirect estimates had a range of 11 to 27%. This is subject to large uncertainty if the sample size is very small. On the other hand, omitting important covariates can result in over smoothing, so obtained estimates should be interpreted with caution and validated in future studies.

Another important methodological step in this work was calculating the population attributable fractions that are subsequently used to obtain smoking attributable mortality. These widely used in public health measures are subject to strong limitations. First, these calculations rely on assessment of relative risk of smoking for all-cause mortality from existing studies across the world or indirect methods such as Peto-Lopez method that bases relative risks on the excess rates of lung cancer mortality [26]. For this study, we have used the estimates from four meta-analyses of existing studies [20‐23] as meta-analyses provide more robust and stable estimates of relative risks compared to single studies. However, the obtained estimates did not cover in sufficient detail all age and gender groups of general adult population. Analyses under the two scenarios (standardizing the population for age and gender vs summing up estimates from six strata's by age and gender) resulted in a wide range of estimates with wide confidence intervals reflecting substantial uncertainty (Table 4). Second, while relative risks were obtained after adjusting for several important confounders, the calculations of population attributable fraction may still fail to account for a number of possible exposures such as, for example, air quality or work-related exposure. Third, important considerations were made in respect of standardization of number of deaths per municipality to age and gender structure of general Belgium population. On the one hand, estimates of relative risks may have already captured the age and gender distribution in the study cohort (which may or may not be comparable to Belgian population), on the other hand, obtaining PAF estimates on standardized or stratified number of death per age and gender group was supposed to explicitly eliminate differences in number of deaths between relatively ‘green’ (with large proportion of young population) and ‘grey’ (with large proportion of older population) municipalities. To further improve the accuracy of PAF calculations, data on age and gender specific relative risks of all-cause mortality from smoking in Belgian population are needed. Last but not least, the estimates of smoking prevalence and death statistics from 2018 were used, while there is a known large lag between the onset of smoking, length of exposure to smoking and death. Given that smoking rates have been declining substantially over the last 20 years [6], provided estimates of smoking attributable mortality are likely to be conservative as deaths occurring in 2018 are related to past smoking.

In conclusion, substantial within-country variation in smoking and subsequently smoking attributable mortality was observed. This study provided an illustration of the methodological approach to obtain small area estimates from national health survey that can be used for a broad range of health indicators.