Influence of major non-communicable chronic disease diagnoses on Chinese adult males’ smoking behavior: is there a moderating role for the home environment?

2.1 Data

The data (second-hand, publicly available) used for this study were retrieved from the CHNS, a large-scale longitudinal household survey conducted in more than ten provinces in China that exhibit significant variations in geographical conditions, social and cultural characteristics and levels of economic development. The CHNS was launched in 1989, as the product of an international collaborative effort between the University of North Carolina at Chapel Hill and the Chinese Center for Disease Control and Prevention.

The CHNS originally involved nine provinces, including Guangxi, Guizhou, Heilongjiang, Henan, Hubei, Hunan, Jiangsu, Liaoning and Shandong. The first-round survey initiated in 1989 employed a stratified, random sampling strategy to select sample communities and households. Counties in each of the nine provinces were stratified by income levels (low, middle and high), and four counties were randomly selected from each project province. Whenever possible, a lower-income city and the provincial capital city were also chosen. In addition, residential districts within selected cities and villages within selected rural counties were randomly chosen to be the survey sites. Between 1991 and 2015, the CHNS conducted nine rounds of follow-up surveys. In 2011, three municipalities (Beijing, Chongqing and Shanghai) joined the project; in 2015, three more provinces (Shaanxi, Yunnan and Zhejiang) became part of the project.

Overall, the CHNS involved approximately 7200 households and 30,000 individuals in its comprehensive data collection efforts. The surveys gathered rich data on their family configuration, demographic and socio-economic characteristics, food and nutrient intake, health status, and the conditions of basic infrastructure and healthcare facilities in their communities. Of particular relevance to the present study, the CHNS began collecting detailed data on sample respondents’ cigarette consumption starting from the second-round survey in 1991. Information on hypertension diagnosis was added in the third round in 1993, and data on the diagnoses of other NCDs were collected in the fourth round in 1997. More detailed information about the CHNS can be found at https://www.cpc.unc.edu/projects/china.

2.2 Sample restrictions

To ensure the suitability of the analytical sample for our intended analysis, several sample restrictions were applied. Firstly, to control for one’s “smoking history”, we excluded all observations collected before the fifth round (2000) to ensure that each sample observation had information on daily cigarette consumption lagged for one period. Secondly, we focused only on the nine original CHNS provinces, excluding observations from municipalities and provinces that joined the CHNS from 2011 onwards, to maximize the available information in the time dimension for each sample respondent. Thirdly, we excluded female respondents from the sample due to the low prevalence of smoking among them, as only 6.2% had ever smoked. Lastly, we focused on males aged between 45 and 85 years old. The lower bound (age 45) was chosen because very few observations under 45 (—the cut-off between “youth” and “middle age” commonly used in China) in the data had been diagnosed with any NCD. The upper bound (age 85) was chosen because individuals over 85 smoked significantly less than their younger counterparts [20]. To address missing information, especially on NCD diagnosis status, we imputed the data using available information from other rounds as much as possible. For instance, if a male respondent reported a diabetes diagnosis in 2000 but did not provide information on diabetes diagnosis in 2004, we inferred that the missing diabetes diagnosis status in 2004 should be filled in as “already diagnosed with diabetes”. After applying these sample restrictions, we obtained a pooled analytical sample comprising 16,188 individual-year observations for 5388 adult males residing in 228 communities across the nine original CHNS provinces. Each adult male appeared in at least two rounds (out of six) of the surveys conducted between 2000 and 2015. Fig. 1 presents a flowchart documenting the sample construction steps mentioned above.

Fig. 1.Flowchart for sample construction.

2.3 Variables

2.4 Statistical model

As a starting point for our empirical analysis, we consider a statistical model that examines the relationship between a Chinese adult male’s daily cigarette consumption (N_smoke) and its determinants:

Where the variables may vary across individuals (denoted with subscript i) and time periods (denoted with subscript t). S_it, the main explanatory variable of interest, represents whether an adult male i had been diagnosed with an NCD (i.e., one of diabetes, hypertension, myocardial infarction and stroke) by time t. F_it is a set of variables measuring an adult male i’s home environment (whether his parent or spouse smokes at t and whether his parent or spouse had been diagnosed with an NCD by t). Z_it represents the set of control variables discussed in Section 2.3.3. ϵ_it is the disturbance term. Before estimating equation (1), a correlation analysis is performed to check if there are multicollinearity issues among the explanatory variables.

It may be tempting to estimate equation (1) using linear regression, but linear models are unsuitable for this study due to the count-data nature of the outcome variable, N_smoke (subscripts omitted for ease of exposition), the number of cigarettes smoked per day. As this variable is in the form of counts, a Poisson regression framework could be considered to estimate equation (1). However, a “built-in” property of Poisson distributions, the so-called “equal-dispersion” property, i.e., Var(N_smoke∣X) = E(N_smoke∣X), for X = [S, F, Z], usually fails to hold in practice [28, 29, 30, 31]. To address this concern, we adopt a Negative Binomial regression framework, which accounts for the “over-dispersion” issue [28, 29, 30, 31]. Further, taking into account the fact that many sample males did not smoke, which generates a cluster of zeros in N_smoke, we adopt a zero-inflated Negative Binomial regression approach to estimating equation (1). More specifically, the “zero-inflation” part is estimated by a Logit model predicting “certain zeros” (i.e., “definitely not smoking”), jointly with a Negative Binomial model predicting the positive number of cigarettes consumed daily, conditional on one being a smoker. More details on the zero-inflated Negative Binomial model are provided in Supplementary material A.

Another estimation issue is that, as noted in section 2.3, our analytical sample involves repeated observations over time for a subset of sample males. To address this issue, we adopt the approach suggested by Cameron and Trivedi (2005) [28], originated in Liang and Zeger (1986) [32], and adjust the standard errors of the estimated parameters for intra-community clustering.

It should be noted that different (sub)samples are used in the analysis below for different purposes. For example, the full sample (including both smokers and non-smokers) is used to ensure that our estimation results can be generalized to a larger population, improving the external validity of our study. Admittedly, analysis based on this sample faces a potential reduction in its internal validity, as in that case, some cross-sectional variation must be utilized to identify the parameters in the model. By contrast, excluding never-smokers from the analytical sample improves the transparency of the identification mechanism, which may increase the internal validity of the analysis. But focusing on the smoker subsample could potentially reduce the external validity of our study, as that sample may not be representative of the Chinese adult male population. In any case, comparing results obtained using different subsamples can help inform the driving force of our empirical findings and assess the robustness of these findings.

All procedures and estimations described above were performed using the statistical package Stata (Version SE 17.0, StataCorp LLC, College Station, TX, United States).

3.1 Summary statistics

Table 1 presents summary statistics of all variables used in our analysis. Column (1) displays the statistics for the full sample. The average age of all sampled males (across all year-individual observations) was 59 years. Approximately 2/3 of the sample observations resided in rural areas, and 12% belonged to ethnic minority groups. On average, the sampled males had about 8 years of formal schooling and an annual per capita household income of approximately 13,700 yuan in 2015 values (equivalent to about 2200 U.S. Dollars). Regarding health insurance coverage, 78% of the observations reported being enrolled in a health insurance program during the survey period as a result of the rapid expansion of China’s health insurance industry since the early 2000s [24, 33, 34, 35]. Moreover, 58% of the sampled males reported being actively employed. And 57% reported consuming alcohol during the survey period. Additionally, 52% of the sampled males rated their overall health as “good” or “very good”.

Concerning sampled males’ smoking behavior, column (1) of Table 1 reveals that 53% of the observations reported smoking during the survey period. The average number of cigarettes smoked daily in the full sample was 8.9. Among smokers (column 3), approximately 29% were categorized as heavy smokers, who consumed at least 20 cigarettes per day [36]. Regarding health conditions, the prevalence of NCDs was 28% in the full sample: among smokers, 23% reported having received NCD diagnoses (column 3), while among non-smokers, the prevalence was 34% (column 2). It is interesting to note that the incidence of NCDs is lower among smokers (column 3) than non-smokers (column 2). While this finding may seem counterintuitive, it could be explained by the possibility that NCD diagnoses have induced some sample males to cut back on their cigarette consumption. This explanation is, in fact, confirmed by the comparison of one’s “before NCD diagnosis” (column 5) and “after NCD diagnosis” status (column 6) for available observations. Before NCD diagnoses, 58% of the sample males smoked, but this proportion dropped to 47% after diagnoses. Echoing this, sample males consumed 9.67 cigarettes daily before NCD diagnoses (column 5) while consuming 6.74 after NCD diagnoses (column 6). Also, as can be seen from Fig. 2, a Chinese adult male’s smoking probability (measured on the right vertical axis) and the number of cigarettes smoked daily (measured on the left vertical axis) both decrease with the number of NCD diagnoses.

Fig. 2.Unadjusted associations between the number of major chronic disease diagnosed and smoking probability and average daily cigarette consumption. NCDs: non-communicable chronic diseases.

The home environment of a sampled male is characterized by two binary indicators: Fam_smoke, which equals one if there are other smokers in the household and zero otherwise; Fam_NCD, which equals one if one or more family members of a sampled male have been diagnosed with NCDs and zero otherwise. As Table 1 shows, 19% of the sampled males have family members who smoke, and 26% had been diagnosed with NCDs by the survey. To better depict the relationship and pin down the direction of causality, we consider a large number of potential confounding variables in our analysis below. Before carrying out the regression analysis, a correlation analysis was performed to check for multicollinearity issues among the explanatory variables. Supplementary Table 1 in Supplementary material B reports the results: all correlations between explanatory variables are small, revealing no concern about multicollinearity issues in our study.

3.2 Main findings

To rigorously evaluate how NCD diagnoses may impact Chinese adult males’ smoking behavior while simultaneously addressing the presence of many non-smokers (N_smoke = 0), we employed zero-inflated Negative Binomial regressions to estimate equation (1). Alternative models, including standard Poisson, zero-inflated Poisson, and standard Negative Binomial models, are also estimated. Yet the results of Vuong’s (1989) likelihood ratio test [37] suggest that the zero-inflated Negative Binomial model is the most preferred one (—detailed test results are not reported here but are available upon request).

Table 2 presents zero-inflated Negative Binomial estimates of the effects of one’s own NCD diagnoses and the home environment, as well as how the latter may moderate the former’s effect. Model (1) has the simplest specification, including only the key explanatory variables introduced in section 2.3.2, i.e., Self_NCD, Fam_smoke, Fam_NCD, the interactions between Self_NCD and the other two, as well as province and survey-round fixed effects (dummy variables). Model (2) further includes a set of personal and family-level control variables, as discussed in section 2.3.3, to reduce confounding issues. Model (3) excludes the observations of never-smokers to ensure the identification of model parameters comes entirely from subjects with smoking experiences. Model (4) used only the observations whose first and last appearances form a balanced panel dataset to better control for one’s smoking history. Lastly, to avoid sample truncation due to the death of some sample males (possibly due to smoking-related issues), Model (5) excluded sample respondents above the age of 60.

The analysis of the impact of NCD diagnoses received by oneself reported in Table 2 suggests that being diagnosed with an additional NCD is negatively associated with the probability of smoking (δ_{Self_NCD} = −0.591–−0.230). (It should be noted that all models in Table 2 include interaction terms; thus, the interpretation of the main effects of the variables discussed thus far should also consider the interaction terms involving these variables. But as discussed below, the interaction terms are, in general, statistically insignificant). To put the estimates into perspective, consider the estimate from Model (2): δ_{Self_NCD} = −0.241, which means that being diagnosed with an additional NCD reduces the odds of smoking relative to not smoking by 27.3% (=e^0.241 − 1), a sizable effect. In contrast, one’s own NCD diagnosis has virtually no effect on the number of cigarettes smoked daily, conditional on the individual continuing smoking.

Regarding the roles of family members’ smoking status and NCD diagnoses, our analysis reveals that the presence of other smokers in the family (Fam_smoke) has a positive impact (albeit only marginally significant at the 10% level) on both a Chinese adult male’s smoking incidence (δ_{fam_smoke} = 0.023–0.033). The impact of having a smoking member in the family on one’s daily number of cigarettes smoked (η_{fam_smoke} = 0.304–0.402) is statistically more significant (at the 1% level). In contrast, the effects of family members’ NCD diagnoses (Fam_NCD) are less robust. The effect of Fam_NCD is significant at the 1% level in Model (1) (δ_{fam_NCD} = −0.228) but became statistically insignificant when a large set of control variables were added to Model (2) (δ_{fam_NCD} = −0.083), suggesting the possibility of confounding that the significant effect of Fam_NCD in Model (1) captures the impact of some control variables, such as family income (HH_income) and the number of cigarettes smoked per day in the previous period (N_smoke_lag1). Models (3)–(5) all include Fam_NCD as a predictor, but its effect remains statistically insignificant in these models.

To investigate whether the effect of one’s own NCD diagnosis (Self_NCD) may be moderated by family members’ smoking behavior and health shocks (NCD diagnoses), we follow the standard practice in the literature [38] and include relevant interaction terms—i.e., those between the presence of other smokers in the family and one’s own NCD diagnosis (Fam_smoke × Self_NCD) and between family members’ NCD diagnoses and one’s own NCD diagnosis (Fam_NCD × Self_NCD)—in the models. The results indicate that neither family members’ smoking behavior nor their NCD diagnoses had a moderating effect on the relationship between one’s own NCD diagnosis and smoking behavior, even though family members’ smoking behavior has a significant direct impact on one’s smoking behavior. As robustness checks, excluding sampled males who never smoked (Model 3), keeping the observations whose first and last appearances form a balanced panel data (Model 4), and dropping males aged above age 60 (Model 5) resulted in estimates of the impacts of one’s own NCD diagnosis and those of the home environment that are similar to those in Model (2), the largest model with the largest sample size.

To help visualize the effects of NCD diagnoses, Fig. 3 presents line plots linking the predicted probability of smoking (measured on the right vertical axis) and the predicted number of cigarettes smoked daily (measured on the left vertical axis) with the number of NCD diagnoses (the horizontal axis), based on estimates from a version of Model (2) in Table 2 that uses four dummy variables of the number of NCD diagnoses (instead of a single dummy variable Self_NCD) as the key predictors. The figure indicates that as the number of NCD diagnoses increases, both the probability of smoking and the number of cigarettes smoked daily reduce significantly, but the decline in smoking probability (from 0.56 to essentially zero) is much more significant than that in smoking intensity (from 15 to 9 cigarettes per day).

Fig. 3.Estimated Effects of the number of major chronic disease diagnoses on smoking probability and average daily cigarette consumption. NCDs: non-communicable chronic diseases.

To further see if the effect size of NCD diagnosis varies across different NCDs, we estimated zero-inflated Negative Binomial models separately for the diagnoses of diabetes, hypertension, myocardial infarction and stroke. The results, reported in Supplementary Table 2, indicate that the diagnosis of each of these four NCDs on smoking incidence is at least marginally significant (at the 10% level); the effect of myocardial infarction diagnosis has the largest and most statistically significant effect (at the 1% level), followed by the diagnosis of stroke, diabetes and hypertension. Although the effect sizes differ somewhat across NCDs, they all point in the same direction: NCD diagnoses exert deterring effects on smoking. Again, the effects of NCDs on smoking intensity are mostly statistically insignificant if a smoker decides not to quit smoking, and the moderating effect of the home environment is essentially non-existent.

The estimated effects of the control variables also provide valuable insights. Among the basic demographic characteristics examined in Table 2, age exhibits a significant and nonlinear (inverted U-shaped) effect on Chinese adult males’ cigarette consumption, indicating that Chinese adult male smokers tend to reduce their cigarette consumption after reaching a certain age. Per capita household income and years of education also have deterring effects on smoking, but they operate through different mechanisms. Education does not influence one’s smoking incidence but significantly reduces the daily number of cigarettes consumed among smokers. On the other hand, higher income reduces the likelihood of smoking, but conditional on an individual being a smoker, income has little impact on the number of cigarettes smoked daily. Moreover, factors such as employment status, residing in rural areas, self-assessed health status, and alcohol consumption are all statistically significant contributors to Chinese adult males’ cigarette consumption, but their roles differ to some extent. For instance, working increases both the likelihood and the intensity of smoking, while alcohol drinking affects only the likelihood of smoking but not the intensity of smoking.

3.3 Role of smoking intensity

Given the addictive nature of smoking, smoking intensity may moderate the effects of one’s own NCD diagnosis and home environment on smoking behavior. To examine how smoking intensity might influence the impact of one’s own NCD diagnoses and home environment on smoking behavior, we excluded never-smokers in the analysis reported in Table 3. We used Negative Binomial models to separately estimate the effects of one’s own NCD diagnoses on smoking for non-heavy smokers (Models 1 and 2) and heavy smokers (Models 3 and 4). Models (1) and (3) were estimated without personal and family-level control variables, similar to Model (1) in Table 2. In contrast, Models (2) and (4) include the full set of controls, similar to Model (2) in Table 2.

Comparing the results between heavy smokers and non-heavy smokers reveals three notable findings. First, one’s own NCD diagnosis (Self_NCD) significantly reduces cigarette consumption among non-heavy smokers but is associated with increased cigarette consumption among heavy smokers (though this effect is not statistically significant). This difference may be attributed to heavy smokers’ higher levels of nicotine addiction and dependence, which makes them more likely to use smoking as a coping mechanism when facing stress from an NCD diagnosis [39, 40]. Second, family members’ smoking behavior (Fam_smoke) significantly increases daily cigarette consumption among non-heavy smokers, while family members’ NCD diagnoses (Fam_NCD) significantly decrease it. However, neither family members’ smoking behavior nor their NCD diagnoses significantly impact heavy smokers’ smoking behavior, likely due to their long-established smoking habits, and few other things may change their smoking behavior. Finally, for both heavy and non-heavy smokers, family members’ smoking behavior and NCD diagnoses do not moderate the relationship between one’s own NCD diagnosis and smoking behavior.