We used data from the Comprehensive Survey of Living Conditions in Japan [8], which sought to ascertain the income and health status of households [23]. This survey investigates the health status of Japanese citizens every 3 years, using over 5,000 districts in Japan chosen by stratified random sampling, and all households in the districts are subject to the survey. The numbers of responses (the response rate) for each of the analyzed years were as follows: 2004, 220,836 households (80%); 2007, 229,821 households (80%); 2010, 228,864 households (79%); 2013, 234,383 households (79%); 2016, 224,208 households (77%); 2019, 217,179 households (72%) [8]. These data have been used in several studies in Japan [24,25]. The subjects received a questionnaire asking about their smoking status, with the question “Do you smoke tobacco?” Respondents could indicate that they smoked every day, that they smoked sometimes, that they had quit smoking (and had not smoked in more than a month), or that they did not smoke. Cigarettes and other types of tobacco products were included. The estimated number of each category in Japan was calculated using the prevalence of each category and the estimated number of household members in the survey [23]. We categorized the data for the estimated number of household members and the estimated number of smokers (smoked every day or sometimes) by age group, gender, and area from 2004 to 2019. The age groups categorized into 5-year intervals from 20-24 years to 75-79 years of age were used. The estimated numbers for each prefecture and designated cities (the special wards in Tokyo and government ordinance-designated municipalities) are publicly available. We used the numbers for the special wards in Tokyo and government ordinance-designated municipalities in 2004 as those of urban areas, as in a previous study [14]. The government ordinance-designated municipalities in 2004 were Sapporo, Sendai, Chiba, Saitama, Yokohama, Kawasaki, Nagoya, Kyoto, Osaka, Kobe, Hiroshima, Kitakyushu, and Fukuoka. Although the data for 2001 are also publicly available, the aggregation method for each age group was different from that of the other years, and the data therefore could not be used for an APC analysis. Those who were 75-79 years old in 2004 (born between 1925 and 1929) were the initial cohort, as they were the oldest birth cohort in the data set. Using a 1-year shift in the birth year, starting from the first cohort, the group aged 20-24 years in 2019 (born between 1995 and 1999) comprised the last cohort. Therefore, a total of 71 birth cohorts (1925-1929, 1926-1930, ... , 1994-1998, 1995-1999) were analyzed.

We calculated smoking prevalence based on age groups in urban and non-urban areas. The estimated number of smokers and household members in non-urban areas was calculated by subtracting the number of households in urban areas from those in all of Japan. Age-standardized smoking prevalence was calculated using the population ratio of the total household members in 2019 as the standard population.

We used a Bayesian binomial APC model for analysis, with y_ij being the estimated number of smokers for the age group i (1, ... , I) in year j (1, ... , J) [26]:

y_ij~Binomial (n_ij, p_ij),

where n_ij denotes the corresponding estimated number of household members, and p_ij denotes the corresponding smoking prevalence. The log odds of the smoking prevalence η_ij= log{p_ij/ (1-p_ij)} become:

η_ij=δ+α_i+β_j+γ_k+z_ij,

where δ is the intercept, α_i is the effect of the age group, β_j is the period effect, and γ_k (k= 1, ... , K) are cohort effects. We added the heterogeneity term z_ij to the equation [26]. I, J, and K are the number of time points for age groups, periods, and cohorts, respectively. To identify each parameter, the sum of each effect was restricted to zero. A first-order random-walk was used as a prior for each effect [26]. The programs for the Bayesian APC model were written by the author using rstan software (https://rdrr.io/cran/rstan/). The convergence of the estimated parameters was confirmed based on the values of R-hat. Next, the estimated smoking prevalence was calculated for each age, period, and cohort. For example, the estimated smoking prevalence for age group i can be calculated as expit (δ^+α^i), where δ^ and α^ are the estimated values for the intercept δ and the age effect α_i, and expit denotes a sigmoid function. Furthermore, the smoking prevalence ratio of non-urban areas and urban areas was calculated for each age group, period, and cohort by dividing the estimated smoking rate of non-urban areas by that of urban areas. To ascertain the fit of the APC model, we calculated the deviance information criterion for the age model, age-period model, and age-cohort model in addition to the APC model.

All statistical analyses were conducted using R version 3.5.1 (https://cran.r-project.org/bin/windows/base/old/3.5.1/).

Institutional review board approval was not required because we analyzed data that are available to the public.

The relationship between the areas was different depending on age, and the smoking prevalence ratio decreased as age increased in both men and women. In men, the estimated smoking prevalence was higher in non-urban areas for those aged 20-64, before falling in older ages, as shown in Figure 2. These phenomena resulted from differences between urban and non-urban areas in the magnitude of the decrease in smoking prevalence with increasing age. Smoking prevalence tended to decrease with increasing age, as shown in Table 1 and Figure 1, and this phenomenon has been observed in other countries [27-29]. Smoking cessation by older people is one of multiple possible explanations for this phenomenon. A study investigating predictors of smoking cessation in Japan found that age, health check-up participation, and physical activity were predictors [13]. It is also thought that older people quit smoking because they are more conscious of the health risks associated with smoking [29]. Furthermore, a program of specific health check-ups and specific health guidance was started in 2008 in Japan [30], and if a person aged over 40 years has a high-risk of developing a lifestyle-related disease, the person receives specific health guidance on improving their lifestyle behaviors. Therefore, it is possible that the extent of smoking cessation in older ages differed between urban and non-urban areas. As another possibility, there might have been differences in the prevalence of older people who have never smoked between both areas, particularly for women. A survey conducted in 2001 showed that smoking prevalence in urban areas was higher than in non-urban areas for women [14], and it discussed how urbanization and social participation for women tended to increase the likelihood of smoking in Japan [14,31,32]. These factors might account for the difference in smoking prevalence in older women between urban and non-urban areas.

Although the estimated smoking prevalence showed a decreasing trend during the analyzed period for both genders and areas, the magnitude of the decrease was larger in urban areas for both men and women. Although it is known that smoking prevalence is on the decline in Japan [7], differences were found in the rate of decrease according to the region. A difference in the rate of decrease of the smoking prevalence according to SES was also observed in Germany, and some explanations have been proposed [22]. These include low exposure to public health campaigns and a poor ability to process relevant information among people with low SES [22,33]. Smoking-related public health campaigns in Japan were conducted before the analyzed period [34,35], and manufacturers of tobacco products are obliged to attach information on the harmful effects of smoking on their packages [6]. These factors might have led to differences in the perceived risk of smoking between urban and non-urban areas. Another explanation for the observed difference in Germany is that opportunities to smoke may still be greater for low-SES groups, and the impact of smoking bans was different between SES strata [22]. In Japan, tobacco control policies are unevenly implemented across prefectures [36], and there might be differences in smoking cessation measures, particularly between urban and non-urban areas. Furthermore, a lack of social support is also considered to be a causal factor for a lower likelihood of smoking cessation in low-SES groups [10]. Thus, the difference in the number of medical institutions that provide smoking cessation outpatient services between urban and non-urban areas might be another factor contributing to the difference between the areas. Smoking cessation expenses are not high and are covered by medical insurance in Japan [37]. However, the distribution of physicians in Japan is not balanced [38], and access to smoking cessation outpatient services may be relatively limited in non-urban areas.

Regarding the relationship between smoking prevalence and tobacco price increases, it should be noted that tobacco prices and taxes were raised twice during the analyzed period [4], and sales figures dropped significantly in 2010. However, it is not certain to what extent sales changed for low-SES individuals or in non-urban areas, and the results of this study suggest that the effects may have differed depending on the area. Furthermore, data from the Comprehensive Survey of Living Conditions (2007-2010) [39] showed that the increase in tobacco prices had a significant impact on smoking cessation regardless of SES, which is consistent of the results of this study.

A significant cohort effect was found for estimated smoking prevalence. An increase in cohort effect for prevalence of smoking was observed starting in the cohorts born in the 1930s in both areas for both genders. In Japan, tobacco consumption per person per day increased from the end of World War II to the 1970s [34], and after 1949, women were targeted in advertising produced by the Japan Tobacco and Salt Public Corporation [7]. Smoking prevalence increased, especially for young women, in this rapid economic growth era, and there was a difference in the time points of an increase in smoking prevalence between regions. The degree of the increase in smoking prevalence was more rapid among non-urban women, and the smoking prevalence ratio increased starting in the cohorts born in the 1960s for women, as shown in Figure 2. Previous studies also have found relationships between urbanization and smoking prevalence for women in other countries [15]. The increase of smoking prevalence across cohorts is considered to have led to a decrease in the smoking prevalence difference between regions for women in Japan. However, estimated smoking prevalence began to fall starting in the cohorts born in the 1970s for men and women. Tobacco consumption in Japan was also shown to decrease from the 1970s [34]. This decline can be attributed to the increase in tobacco prices and public awareness of the dangers of smoking [34]. Several smoking control measures have been introduced since the late 1970s [35]. The smoking prevalence ratio for cohorts born after the 1970s was above 1 for men and women, and smoking prevalence was higher for non-urban areas in younger cohorts.

This study was limited by the absence of data on the smoking prevalence of non-respondents to the Comprehensive Survey of Living Conditions, which had an approximately 80% response rate. It is necessary to conduct a similar analysis using nationwide epidemiological cohort data to check the robustness of the results. Furthermore, although we focused on the difference between urban and non-urban areas (i.e., regional SES), these differences do not necessarily reflect individual or personal SES. By analyzing the difference in smoking prevalence according to individual-level SES, disparities based on SES can be more accurately revealed. Furthermore, we could only obtain the estimated numbers of smokers for each area, and this might have affected the precision of the estimates of the APC analysis. Finally, APC analysis is a descriptive method, and the reasons for the changes in each effect are uncertain. Based on the results of this study, we need to reconsider the reasons for the non-decrease in smoking prevalence in non-urban areas. However, the main strength of this study lies in its use of nationwide national survey data for the analysis, and the results can be generalized to all of Japan.

In conclusion, although the age-standardized smoking prevalence was higher in urban areas than in non-urban areas in 2004 for women, there was almost no difference between the areas in 2019. The APC analysis showed that the smoking prevalence ratio between non-urban and urban areas was significantly affected by age for men. The magnitude of the decrease in the period effects on smoking prevalence was also larger in urban areas than in non-urban areas for men and women. Furthermore, although the smoking prevalence ratio of non-urban areas compared with urban areas was below 1 until the cohorts born in the 1970s for women, the trend reversed thereafter. The smoking prevalence ratio between non-urban and urban areas was significantly affected by age for men and by cohort for women, and smoking cessation measures are needed particularly for older cohorts in urban women and for younger ages in non-urban men.