Study design and sampling procedure

This study employed data from a ‘tobacco control advocacy capacity building’ project that was conducted in medical universities across China²⁶. A series of related projects covered all provinces in China and included 103 universities located in 81 different cities²⁷. The data for this study originated in a baseline survey on smoking beliefs, attitudes and behaviors of university students in the fourth project entitled ‘Building tobacco treatment capacity in medical universities and affiliated hospitals in China’. Participants were health-profession students, which comprised medical, nursing and other students in health-related programs. The sampling procedure used an observational cross-sectional, multilevel approach with a multi-stage cluster sampling design. In Stage 1, this project included 50 universities – about 60% (31 universities) of them were involved in the baseline survey – that were selected from across China and differentiated by regional location. Selection of universities and colleges utilized probability proportional to size sampling applied to six different regions: the northeast, north, northwest, southwest, and east²⁸. Of the selected universities, 18 were medical universities offering mainly medical, nursing and other health-related programs, and 13 were comprehensive universities offering both medical and non-medical programs. In total, 31 participating universities and colleges are located across China, extending from Harbin in the north to Guangdong in the south, to Xinjiang and Xizang in the west and Hangzhou in the east. The geographical distribution of the sample of institutions aligns with the total distribution.

In Stage 2, the sampling strategy involved the selection of levels within each university. All levels that had medical/health courses were included. In Stage 3, one-third of the courses were randomly selected from each level. On average, two classes were selected to participate in the study at each university. In Stage 4, all students in these selected classes were surveyed.

Data collection

Individual variables were measured via a structured self-administered questionnaire. This questionnaire was administered during regular classes and took approximately 30 minutes to complete. All responses were anonymous. A common research protocol was used across all 31 universities to assure homogeneity of questionnaire administration and data collection procedures. The study was approved by the Ethics Committee at the Zhejiang University Medical Center, and verbal consent was obtained from all participants before data collection.

Measures

Dependent variable: HRQoL

In this study, HRQoL was measured using the EQ-5D instrument, which describes health in five dimensions²⁹. It is simple to operate and be comprehended by the subjects and manifests good reliability and validity. EQ-5D is increasingly being used to assess public health issues and is now a common instrument in population-based survey research on HRQoL^30-32. The Chinese version of the EQ-5D has been used in mainland China^33-35. However, its reliability and validity were unreported in most studies except that of Wang et al.³⁶. The study determined that the Chinese EQ-5D showed fair to moderate levels of test-retest reliability. The scores had moderate or strong correlations with all SF-36 scores. However, their study did not report the results of the test for internal reliability. In our study, the Cronbach alpha coefficient of the EQ-5D is 0.56, which approaches the lowest value (0.6) requirements for consistency⁴. In concert with the Wang et al.³⁶ study results, we deem that the Chinese EQ-5D has acceptable reliability and validity. It taps mobility, self-care, usual activities, pain or discomfort, and anxiety or depression. Each question allowed for three alternative responses: ‘no problems’, ‘some or moderate problems’, and ‘extreme problems’. They were coded 1, 2 and 3 and cumulated to form a total score for HRQoL²³.

Departing from prior practice, where assessment scores of HRQoL were measured as a continuous variable^18-23, HRQoL was dichotomized as normal and poor in our study. We employed HRQoL as a categorical variable because the quantitative comparisons produced highly sensitive differences in statistical difference tests in a large sample, and often the results obtained from a large sample statistical test could not be replicated in a small sample statistical test⁴. Furthermore, if the variable was represented numerically, HRQoL level related to assessment scores cannot clearly distinguish higher from lower. It is important to consider the characteristics of the variable instead of automatically assuming it is numeric. Finally, in considering professional significance, it is more comprehensible to stakeholders if HRQoL results are presented as normal or poor – similar to normal or high blood pressure and mental stress readings. For this analysis, responses to EQ-5D of at least two ‘moderates’ or one ‘extreme’ were classified as poor quality of life⁴.

Information regarding smoking status was assessed using standard methods. For the behavioral smoking measure, we employed the method recommended by the World Health Organization³⁷, as utilized in the Global Adult Tobacco Survey (GATS). We defined a current smoker as someone who smoked cigarettes at the time of interview, a daily smoker as someone who smoked every day, and an occasional smoker as someone who smoked on some days^1,37. The independent variable was smoking status and coded as: 1=non-smoker, 2=occasional smoker, and 3=daily smoker. University-level current smoking status was based on an aggregation of individual-level current smoking responses, with current smoking prevalence distinguished as: <5%, 5–9.9%, and ≥10% (Table 1).

Individual-level control variables

Many studies have found that individual-level demographic and socioeconomic variables may influence HRQoL^38-40. In this study, sociodemographic characteristics included age, sex, ethnicity, paternal and maternal occupation, academic major and achievement, and monthly expenses. Academic achievement was measured by asking respondents their class position: upper third (high), middle third (medium), or lower third (low). Monthly expenses were derived from the question: ‘How much do you spend each month?’.

Alcohol use

Smoking and alcohol use often co-occur, and both smoking and excessive alcohol consumption are harmful to health⁴. Alcohol use was ascertained by the question: ‘Do you drink alcohol?’; options were ‘no’, ‘drink alcohol on some days’ (occasionally), and ‘drink alcohol every day’ (daily). Responses were coded as: 1=non-drinker, 2=occasional drinker, and 3=daily drinker.

Contextual control variables

HRQoL may be influenced by the environment⁴.

Home and university environment control variables

Some studies show that family environment is associated with HRQoL^30,31. We included a measure of the location of the home because this can reflect the socioeconomic position of the family, and social and economic development varies greatly between rural and urban areas⁴. Family location of students was differentiated as: countryside, township, county town, or city.

Heterogeneity in social resources and academic prestige of universities may influence student HRQoL^4,39. University type was determined using the China university ranking system (low, middle, and high) implemented by the National Ministry of Education⁴¹.

City-level environment control variables

The first environmental contextual variable included in this study was the economic level of the cities in which the students were studying, measured by per capita Gross Domestic Product (GDP) in RMB (1000 Chinese Renminbi about US$140). Since larger city size is frequently accompanied by increased stress and poor HRQoL, this population variable was also included. Also frequently, population density is positively associated with scarce resources, environmental pollution, traffic congestion, and other urban problems, and hence also is often accompanied by increased stress. This variable was included as a contextual variable because it may lead to poor HRQoL. The source was the National Bureau of Statistics²⁸. The preceding categorizations were based on prior practice⁴². Moreover, our sensitivity analyses using different categorizations of these variables produced similar results.

Data analysis

All data were entered into a database using Microsoft Excel; data quality was ensured by double data entry. The dataset was imported into SAS (9.4 version) for statistical analyses. Descriptive statistics were calculated for poor HRQoL prevalence, together with 95% confidence intervals, across smoking use and individual and contextual control variables. Both unadjusted and adjusted analyses were conducted in assessing the association between smoking and poor HRQoL. SAS survey logistic procedures were applied in the unadjusted analysis, using the university as the clustering unit, in order to account for within-clustering correlation attributable to the complex sample. Associations were confirmed through the application of a multilevel logistic regression model, using the SAS GlIMMIX procedure⁴³.

We built several multilevel models for this analysis. We estimated both fixed and random effects; the former explains the variation of HRQoL at the individual level and the latter at the contextual level⁴³. We commenced with the Null Model, a three-level (individual, university, and university city) model with random intercepts, which did not include any predictors except a constant, in assessing variation in individuals experiencing poor HRQoL. From this base, we constructed four additional models. They were: the individual characteristics model (Model 1); individual-level model (Model 2), which adjusted demographic variables and alcohol status; university-level model (Model 3), which examined both individual and university level smoking status and poor HRQoL – the covariates adjusted in this model are the same as for Model 2; and the city-level model (Model 4), which is based on Model 3 with an added adjusted environmental covariate, population density; and the last model was conceived to examine any change in the association between smoking and poor HRQoL after alcohol status was excluded (Model 5). Model fitting was assessed by the likelihood of a change in the −2 log and significance of the parameter (fix and random) variance estimates by the t-test statistic⁴³.

All analyses were weighted: sampling weights were the inverse of the probability of selection of the university; post-stratification weights were calculated in relation to sex, based on the estimated distributions of this characteristic from a national survey⁴⁴. Final overall weights were computed as the product of the prior two weights. We did not consider using a non-response weight because non-response rates were low in this study.

Limitations

The study has limitations. First, our cross-sectional study design is an important research limitation and precludes the establishment of causal links between smoking and poor HRQoL. Nonetheless, we employed a large sample, and our findings met several criteria for inferring causality, including the strength of some associations, consistency between individual-level and contextual effects, and consideration of individual and contextual confounding and the biological plausibility of effects. Future studies need to collect longitudinal surveillance data on the HRQoL of university students. Secondly, another study limitation is that our respondents were health-profession students studying at universities. Consequently, our results are not generalizable to other populations or to the wider Chinese population. Thirdly, other studies that utilized the Chinese EQ-5D have not reported on internal reliability. This is the most basic indicator for evaluating the reliability of a questionnaire or scale. In our study, the Cronbach alpha coefficient of the EQ-5D only reached 0.56. This value emanated from a large sample—it would be less in small samples. Further research is necessary. Finally, although the distribution of our sample of universities and colleges is consistent with the total distribution across the country, medical and other health-profession education in China is complex. They exist either as a college (or department) of medicine within a comprehensive university or as an independent medical institution under the central or local jurisdiction; thus, our sample representation is limited. The impact of this organizational complexity on our research questions warrants in-depth investigation.

Implications

Many studies have confirmed that smoking is causally associated with various diseases, and some studies have found an association between smoking and human functional impairment. This study examined the association between smoking and poor HRQoL in university settings across China. The study provided additional evidence that smoking is harmful to health, such as the adverse effect of both occasional smoking upon poor HRQoL, but more importantly, helps to elucidate the smoking risk mechanism. This study suggests the importance of the enforcement of public tobacco control education and policy to better protect people from the harms of smoking in China.