2.1 The sample
The study used the China Health and Retirement Longitudinal Study (CHARLS) from a nationally representative dataset was collected, using a steady-state design, from 450 villages or communities in 150 counties or districts of 28 provinces. The baseline survey was conducted in 2011, through a four-stage, stratified, cluster probability sampling design 35. The surveys in 2014 and 2015 contained assessments of social, economic and health status, and recruited additional individuals who just became 45 years old. Further details are provided elsewhere 36, we pooled data from the baseline survey and new respondents in the rounds 2 and 3 surveys. For the analyses in this study, we combined the 2011–2015 data, the exclusion criteria were as follows: (i) those aged>60 years; (ii) those who did not report an edentulous status; and (iii) those with values missing from the questionnaires. A total of 8,641respondents were included in the initial data, edentulous status was measured for 7,938 respondents, and 479 were excluded. Ultimately, 7,459 respondents (3,308 males and 4151 females) were selected for the fully-adjusted model analysis.
2.2 Measure
Edentulism was assessed based on the response question: “Have you lost all of your teeth?”, the variable had two options (yes or no).
2.2.1 Childhood SES and father’s health-related behaviors
Parental SES levels were used as childhood SES 37, and were derived from CHARLS. Father’ education consisted of four responses (illiterate, elementary school, middle school, and high school or above). The occupation of the parents 38 were dichotomized in two groups (farmer and non-agricultural). The father’s working status was divided into “all of childhood” and “part of childhood”. Family’s financial situation 39 and self-perceived childhood health were classified as four groups (very good, good, fair, poor, or very poor). Relationship with father 40 was classified as four groups (excellent, very good, good, or poor). Mother’s effort to care for her child was classified into four groups (a lot, some, a little, or none at all). Mother’s expression of love and affection 41 were classified as four groups (often, sometimes, rarely, or never). Parents quarrel and fight 26 were classified as four groups (often, sometimes, not very often, or never). Residential community security 42 was divided into three groups (very safe, somewhat safe, or not safe). Primary residence 38 was divided into two groups (rural or urban), neighbors’ help included three groups (willing to, somewhat willing to, or unwilling to). Neighborhood relation included three groups (very close-knit, somewhat close-knit, or not close-knit). Enough food 43, medical convenience 43, and father’ smoking and drinking habits 44 had two options (yes and no).
2.2.2 Childhood health
Childhood hospitalization (≥1 month) had two options (yes and no). Self-perceived childhood health status was classified as five groups (very good, good, fair, poor, or very poor).
2.2.3 Adult SES
Adult SES was assessed by adult educational achievements 38, divided into four groups (illiterate, elementary school, middle school, or high school and above).
2.2.4 Health-related behaviors
Health-related behaviors included smoking (yes and no) and drinking 38 (>1 times /month, ≤1 times /month, and not drinking). Physical activities 45,46 every week had two options (yes and no). Sleep time 47 was divided into four groups (≤4 h /day, 4-6 h /day, 6-8 h /day and>8 h /day).
2.3 Directed acyclic graphs (DAGs)
We used DAGs to analyze mediators of the association between the main exposure and the outcome, DAGs are considered as sets of arrows between exposures and outcomes and specify the relationships among mediators of exposures and outcomes 48. In this study DAGs were prepared based on theory, literature review and empirical evidence, as shown in Figure 1. This conceptual model identifies the minimally sufficient set of mediators and makes explicit assumptions to account for mediators of the relationship of childhood SES and edentulism. In the CHARLS dataset, there were four set of mediators: (1) father’s health-related behaviors; (2) childhood health; (3) adulthood SES; (4) adulthood health-related behaviors. We identified 16 variables of childhood SES, 2 variables of both father’s health-related behaviors and child health, 4 variables of adult health-related behaviors, and only one variable of adult SES. We used DAGitty v2.3 49 to assess the postulated DAGs .
2.4 Statistical analysis
The differences in edentulism by sex, childhood SES, father’s health-related behaviors, childhood health, adult SES, and health-related behaviors were analyzed using the Pearson’s Chi-Square tests or Fisher’s Exact tests. To account multicollinearity problems among childhood SES, confounding factors and edentulism variable, we applied a multicollinearity test. The variance inflation factor (VIF) and tolerance (TOL) were analyzed for detecting multicollinearity problems 50,51 (VIF>5.0 or TOL<0.2) 52.
To adjust for potential confounders, three stepwise binary logistic regression models were run to examine the association between childhood SES and edentulism. Sex, childhood SES and father’s health-related behaviors childhood health were adjusted in model 1; adult SES was subsequently adjusted in model 2; finally, health-related behaviors were adjusted in model 3. The odds ratio (OR) and the 95% confidence interval (CI) were used to compare the effects on the independent variables. We checked the presence of interaction effects between independent variables, performed overall model evaluation and goodness-of-fit statistics of the model. The statistical significance was considered as two-sided P-value < 0.05. All analyses were conducted using the Stata 14 software (Stata Corp. LLC, College Station, Texas, USA).