An Observed-Variable Socio-Ecological Model Chain Analysis of Socioeconomic Status, Parental Beliefs and Early-Childhood Caries in Chinese.

Research design and sample

The outlined hypotheses systematically assess each component’s relative influence and sequential roles within the structure-cognition-behaviour-disease pathway. Comparing hypotheses H4 and H5 clarifies the relative contributions of structural residual and chain-mediated effects. Confirmation of these hypotheses would inform evidence-based interventions targeting caries among children from low-SES backgrounds. Specifically, it evaluates whether traditional health education alone is sufficient to reduce social disparities in oral health, or whether macrostructural resource allocation, such as improved access to preventive services and economic subsidies, is required to achieve equitable oral health outcomes.

A stratified cluster survey was conducted in Nantong City, Jiangsu Province. Kindergartens were first stratified into urban and suburban strata; within each stratum, 2 public and 2 private schools were randomly selected (6 public, 4 private). All classes-small, middle, and large-were invited, and oral examinations plus questionnaires were completed for 612 children; 595 usable records yielded a 97.2% response. The instrument, adapted from a pre-tested preschool oral health questionnaire, covered demographics and SES (parental education, household income), a nine-item POHBS with 5-point Likert and 3 reverse-scored items, children’s brushing and sugar intake habits, and dmft recorded to WHO 2013 standards. Two calibrated dentists conducted exams (inter-examiner _k = 0.92). School doctors reviewed forms on site, eliminating missing data. Ethical approval was granted by the Biomedical Ethics Committee of Nantong Stomatological Hospital (PJ 2021-002-01) in line with the Declaration of Helsinki. Parents received written study information and gave informed consent; withdrawal was permitted at any time. Examinations were non-invasive (visual inspection and probing only), and data were stored confidentially for research and policy use. The same approval covers secondary analyses.

Questionnaire structure and data processing

The nine-item POHBS uses a 5-point Likert format. Exploratory factor analysis of the 9-item belief pool yielded 2 correlated factors: the professional prevention (PP) and outcome appearance (OA) domains. PP comprises 4 positively worded statements; OA contains 2 positive and 3 reverse-coded items. Because a assumes unidimensionality, this multidimensionality likely contributed to the low POHBS a. We therefore analysed a 4-item preventive-belief subscale to improve coherence, noting its still-modest reliability. Internal consistency was poor (a total = 0.18; PP = 0.40; OA = 0.08), indicating marked item heterogeneity. Exploratory factor analysis of the polychoric matrix (KMO = 0.70; Bartlett x²(36) = 684.5; p < .001) revealed 2 factors explaining 61% of the variance. After rotation, the 4 positive PP items-regular check-ups, pit-and-fissure sealants, protecting 6-year molars, and “bacteria cause gingival inflammation”-loaded 0.46-0.71 on Factor 1, while the 3 reverse PP items loaded -0.42 to -0.67. The 2 OA items loaded 0.64 and 0.68 on Factor 2, matching expectations. To enhance reliability, only the 4 high-loading PP items (> 0.35) were averaged and standardised to form BeliefPrev_z, focusing analyses on the intervention-relevant prevention construct and minimising measurement error.

All continuous or semi-continuous variables were z-standardised. SES was the mean of three 1-5 items (father’s education, mother’s education, and annual income). Three misconception statements were reverse-coded in the nine-item POralHBS (6-x); only 4 professional-prevention items with factor loadings > 0.35 were averaged and standardised as BeliefPrev_z. Misconception and appearance items were excluded to reduce error. BehaviourRisk combined indicators of inadequate plaque removal and free-sugar exposure. Nightly toothbrushing frequency (1 = 0 times to 5 = >3 times) was reverse-coded to InvBrush. The weekly frequencies of sweets and sugar-sweetened drinks (0-7) were averaged with InvBrush; all components were z-standardised and signs aligned so that higher scores reflect higher risk, yielding BehaviourRisk_z. A sensitivity variant (BehaviourRisk_alt_z) additionally included snack and fried-food frequency; inferences were unchanged. This work conducted stratified models by sex (boys/girls) and single-year age groups (3, 4, 5, 6 years) as pre-specified sensitivity analyses. Caries were recorded according to the WHO 2013 criteria; teeth with caries and fillings were counted as carious. Despite positive skew, dmft was entered into linear models; a binary dmft_01 (0 vs > 1) was used for sensitivity analyses. All indicators conformed to model assumptions.

Statistical analysis

This work used an observed-variable SEM chain to decompose total effects into direct and indirect components and to test mediation with bias-corrected bootstrap 95% CIs, reporting standardised B for within-model comparison. A latent-variable SEM was not pursued because the belief item pool showed low internal consistency (POHBS a =0.18; BeliefPrev a=0.40), raising risks of unstable factor solutions and biased paths. This transparent observed-variable framework, therefore, best matched the measurement properties of the present data.

Standardised path coefficients B are reported to convey the expected change in the outcome per 1 SD change in the predictor, conditional on other variables. Statistical significance was judged by bias-corrected bootstrap 95% CIs (2000 resamples) and 2-sided p-values; CIs excluding zero indicate significance. While B facilitates within-model comparisons of relative pathways, cross-study comparisons of B should be made cautiously, as standardisation depends on sample-specific variances and measurement reliability. As contextual guidance (not rigid thresholds), B values of about 0.10, 0.30, and >0.50 can be viewed as small, moderate, and large effects, respectively. While dmft is a count outcome, we used chained OLS for transparency; future work will consider generalised SEM with count links to address zero-inflation and non-normality. Indirect (mediated) effects were assessed with bias-corrected bootstrap 95% CIs. Given the modest reliability of the belief scale and the cross-sectional design, small mediations may be downward-biased and should be interpreted cautiously. The statistical analysis process is as follows:

Stage 1: Reliability/validity. Cronbach’s a was calculated for the nine items and subscales. Scales with a < 0.70 underwent exploratory factor analysis (polychoric matrix, multiple imputation, principal-axis factoring). Factor number was set by parallel analysis and eigenvalues > 1, followed by Varimax rotation; loadings > 0.35 were retained. If a stayed low, BeliefPrev was recomputed from high-loading items;

Stage 2: Path modelling. Chained OLS regressions estimated: (a) SES – BeliefPrev; (b, c’) SES + BeliefPrev – BehaviourRisk; (d-f) SES + BeliefPrev + BehaviourRisk – dmft. BehaviourRisk_alt_z tested sensitivity. Gender- and age-stratified models compared standardised coefficients (95% CI);

Stage 3: Indirect effects. Bias-corrected bootstrapping (2000 resamples) quantified mediation; paths were significant when the 95% CI excluded zero. Analyses used Python 3.11 (pandas 2.2, matplotlib 3.8), R 4.3.2 (psych 2.3.6, GPArotation 2023.9), and statsmodels 0.14/scikit-learn 1.4 with a fixed seed (2025).

Demographic and socioeconomic status distribution

This study encompassed 595 children aged 3 to 6 years and their respective caregivers. The cohort consisted of 326 males and 269 females, yielding a male-to-female ratio of approximately 1.21:1, which aligns with the findings of the Fourth National Oral Health Epidemiological Survey of China. The age distribution among the participating children approximated a normal distribution, with a mean age of 4.5 years (SD = 1.0). The sample sizes across the different age groups were relatively evenly distributed, with the proportions of children aged 3, 4, 5, and 6 years being 23.7%, 27.2%, 26.4%, and 22.7%, respectively. This balanced distribution satisfies the statistical requirements necessary for detailed age-stratified analyses.

Figure 1 summarises night-time brushing, fluoride toothpaste use, and sugar intake among 595 preschoolers. Nightly brushing was everyday: 62.7% brushed every night, 34.8% brushed 3-6 times/wk, and 2.5% brushed <2 times/wk. Fluoride use lagged brushing: 53.6% of caregivers reported fluoride toothpaste, 22.2% did not, and 24.2% were unsure, signalling information gaps that may limit fluoride’s preventive benefit.

1. Download: Download high-res image (467KB)
2. Download: Download full-size image

Fig. 1. Descriptive statistics of oral-hygiene and sugar-intake behaviours in preschoolers.

Sugar exposure was skewed toward snacks. While 24.7% of children ate sweets >3 times/wk, only 4.8% drank sugary beverages > 3 times/wk; 79.2% consumed sweet drinks once weekly. High-frequency intake of fried or puffed snacks was rare (3.7% and 5.0%, respectively), suggesting that these are minor sources of sucrose. No sex differences emerged in any indicator (e.g., sugary drinks x² = 0.27, p = .60). A behavioural risk index combined 3 standardised components: infrequent night-time brushing (<2 times/wk), sweet-snack frequency, and sweet-drink frequency. Standardisation prevented any single item from dominating the score. The index captures the 2 principal caries risks-limited plaque removal and dietary sugars-and will be used in subsequent path models to test behaviour as a mediator between SES and dmft.

Validation

Factor 2 was supported mainly by perception items; caries affects appearance (0.311), and a mother’s teeth influence her child’s (0.363), which explains ~20% of the variance. Loadings were lower than those in Factor 1, indicating that appearance-related concerns are mainly independent of professional-prevention beliefs; misconception items showed negligible loadings here. The scale’s low overall a reflects heterogeneity between these 2 dimensions and the small item set for Factor 2. To improve measurement reliability, 4 high-loading items from Factor 1 (sealants, bacteria, deciduous-tooth protection, check-ups) were retained to construct BeliefPrev, while Factor 2 was used descriptively only. This strategy enhances reliability and theoretical focus, providing a stable base for subsequent path models.

Model estimation

The SES – BeliefPrev path was small and non-significant. The inverse SES – BehaviourRisk association refers to the composite index, not a single item; results were similar using an expanded behaviour index.

• (1) SES – BehaviourRisk

Figure 2 displays direct and indirect links among SES, BeliefPrev, BehaviourRisk, and dmft within the SEM. Significant paths are drawn as black solid lines; non-significant ones as grey dashed lines. SES directly affected dmft (B = -0.329, p = .044), indicating higher caries risk in lower SES children. SES also negatively affected BehaviourRisk (B = -0.187, p < .001), showing that limited economic and educational resources are associated with poorer brushing and dietary habits, raising caries risk. This dual pathway supports the structure– priority model: structural resources directly shape disease prevalence, and behavioural risk is shaped indirectly.

1. Download: Download high-res image (105KB)
2. Download: Download full-size image

Fig. 2. Path relationships among SES, BeliefPrev, BehaviourRisk, and dmft.

Sensitivity and Robustness

Figure 3 contrasts 2 dmft models that differ only in the BehaviourRisk index. The original index includes night-time brushing, sweet-snack, and sweet-drink frequency; the extended index adds snack and fried-food frequency (5 items total). Both models control for BeliefPrev_z and SES_z. SES_z remained influential: B = -0.329 (p = .044) in the 3-item model and B = -0.310 (p = .057) in the 5-item model, slightly attenuated yet directionally stable, indicating that socioeconomic gradients persist after accounting for extra behaviour items. BehaviourRisk lost strength after extension (B = 0.193 – 0.176), and both 95% CIs crossed zero, confirming limited predictive value. BeliefPrev_z stayed negligible (B = 0.02; p > .8). Adding snacks and fried-food frequency did not improve model fit, likely because these variables are collinear with sugary-snack and sweet-drink intake, while omitting critical aspects such as food stickiness and postmeal brushing. Consequently, the behavioural pathway contribution to caries remains understated, whereas the structural SES effect endures across specifications.

1. Download: Download high-res image (146KB)
2. Download: Download full-size image

Fig. 3. Sensitivity analysis with alternative behaviour-risk definition.

Figure 4 plots B (95% CI) for SES_z, BehaviourRisk_z, and BeliefPrev_z against dmft separately for boys and girls. SES_z showed a stronger negative link in boys (B = -0.16; CI excludes 0) than in girls (B = -0.08; CI touches 0), suggesting greater socioeconomic vulnerability among boys. BehaviourRisk_z and BeliefPrev_z were near zero in both sexes, with CIs spanning 0, indicating no independent or gender-specific effects. Thus, the structure-dominant, weak behaviour-belief pattern remains consistent after gender stratification.

1. Download: Download high-res image (95KB)
2. Download: Download full-size image

Fig. 4. Sex-stratified B with 95 % CI predicting dmft.

Figure 5 traces SES effects on dmft across ages 3 – 6. The SES B became more negative from -0.06 (age 3) to -0.13 (age 5), then eased to -0.03 (age 6). A linear trend (= -0.03 per year; p < .05) supports the cumulative disadvantage hypothesis: low SES children experience prolonged exposure to sugary diets and limited resources once primary molars erupt, with disparities peaking at ages 4-5. The attenuated effect at age 6 may reflect fewer cases and earlier school-based fluoride programmes, suggesting that intervention should begin around 4-5 years via subsidies, fluoride toothpaste distribution, and routine screenings.

1. Download: Download high-res image (116KB)
2. Download: Download full-size image

Fig. 5. Age-stratified B with 95 % CI predicting dmft.

BehaviourRisk_z B values rose slightly with age but remained non-significant (CIs straddling 0), implying a biological lag between high sugar exposure / skipped brushing and observable enamel loss, compounded by parental under-reporting and omission of factors such as food stickiness or post-meal cleaning. BeliefPrev_z showed unstable, non-significant effects (B = -0.13 at 4 years; B = 0.08 at 6 years), likely due to scale unreliability and potential reverse causation (parents heighten beliefs after noticing caries). Future work should use objective methods, such as 24-hour dietary recall, plaque biomarkers, and longitudinal follow-up to capture delayed behavioural impacts. Validated scales are also needed to track links between beliefs, behaviour, and caries at different ages, forming the evidence base for phased, precise interventions.

Main findings

Using cross-sectional data from 595 preschoolers, this observed-variable SEM showed that SES retained a significant negative link with dmft after controlling for beliefs and behaviours. The SES effect intensified with age, peaking at B = -0.13 at 5 years, then easing, likely due to a smaller six-year subsample and school fluoride programme. This trend supports cumulative disadvantage theory and identifies ages 4-5 as a key intervention window. Compared with earlier reports (OR: 0.65 – 0.80; unstandardised B = -0.20), our age-specific B values allow finer cross-study comparisons. Relative to other structural markers, the SES B mirrors the dmft coefficient reported in the Basque region (B = -0.14) but is weaker than community dentist density in the US Head-Start cohort (B = -0.22).^20,21 In urban settings where access is generally good, economic advantage alone may not capture the full gradient; neighbourhood dentist distribution and social capital likely interact with SES. Multilevel models incorporating public service density alongside SES could more precisely disentangle economic versus neighbourhood residual effects. Relating these paths to our prior LCA-derived groups—high-risk/high-sugar, medium-risk/mixed, low-risk/high-brushing—shows that even low-risk children remain exposed when SES is low, while the null Belief – Behaviour path implies education alone cannot shift high-risk profiles without simultaneous structural support. Thus, structural measures plus targeted coaching are required; knowledge-only programmes are unlikely to close the SES gap. Because B is sample-standardised, effect significance rests on CIs/p-values, whereas practical importance should additionally consider estimate precision, construct reliability, and policy relevance. Accordingly, our conclusions emphasise direction, statistical evidence, and measurement limitations, rather than B magnitude alone.

Behavior and and belief pathways: Lag, measurement, age differences

The standardised coefficient for BehaviourRisk rose from 0.06 at age 3 to 0.16 at age 6, yet all 95% CIs crossed zero. This suggests a biological or reporting lag between high sugar / missed brushing and observable caries, consistent with STRIP-cohort evidence that sugar at age 5 predicts molar lesions only 18-24 months later.^22,23 Self-report bias further weakens the signal: questionnaires omit stickiness, timing, and post-meal cleaning, and parents often mis-recall bottle-feeding or bedtime drinks.²⁴ Objective tools, multi-day diet logs, plaque ATP fluorescence, and pre-sleep plaque photos could more accurately capture behaviour in future work. The 4-item BeliefPrev composite showed modest internal consistency (Cronbach’s a = 0.40). Such low reliability attenuates regression coefficients, so the near-zero BeliefPrev – Behaviour and BeliefPrev – dmft paths cannot be taken as evidence that parental beliefs are unimportant. Social-desirability bias and potential reverse causation (heightened beliefs after caries onset) may also mask actual effects. Reverse-coded items comprised 33% of the original scale, exceeding the WHO <20% recommendation, and likely depressed a, echoing misclassification seen in Korean and Japanese adaptations.²⁵ Future studies should adopt higher-fidelity instruments, for example, IRT-calibrated items, cognitive interviews for cultural clarity, situational-judgement tests, and, where possible, corroborate parental reports with objective markers of home care.

Model optimisation and measurement priorities. The non-significant BehaviourRisk – dmft path likely reflects under-reporting in self-reports and latency between exposures and clinical caries. Future analyses could employ generalised SEM using a negative-binomial or zero-inflated link for the dmft path, test non-linearities and interactions (e.g., SES × BehaviourRisk), and include richer indicators: diet timing and stickiness (bedtime sugars, retentive snacks), supervised bedtime brushing and fluoride exposure (toothpaste strength, varnish/sealants), objective plaque status (visible plaque or photographic indices; optional plaque ATP fluorescence), and contextual access (kindergarten fluoride programmes; neighbourhood service density). These enhancements may reduce attenuation, improve model fit, and better capture behaviour-to-caries translation.

Interpretation of null mediation. The non-significant indirect paths via BeliefPrev and BehaviourRisk do not imply that beliefs or behaviours are irrelevant to ECC. First, the BeliefPrev subscale showed modest internal consistency (Cronbach’s a = 0.40; POHBS a = 0.18), suggesting attenuation of mediated effects. Second, our cross-sectional design cannot establish temporality and is vulnerable to reverse causality (e.g., existing caries may heighten parental concern and reported behaviours). Third, brief self-reports under-capture diet timing/stickiness and supervised bedtime brushing, and behaviour-to-caries translation may require 18–24 months to manifest clinically. Consequently, we interpret the null mediation as a measurement/timing limitation, while the direct SES association remained detectable under these conditions. Prospective studies using validated belief instruments, objective behaviour metrics (multi-day diet logs, fluoride exposure, visible/plaque photographic indices), and longitudinal mediation within generalised SEM for count outcomes are warranted.

Public health and clinical implications

The results point to 2 priorities. First, the cumulative SES effect argues for early structural action-free fluoridated toothpaste, economic subsidies, and mobile clinics aimed at 4-5-year-olds, before habits rigidify and caries accelerate. Second, although BehaviourRisk was insignificant, its age-linked rise suggests age 6 is an opportune moment for dietary counselling and reinforced bedtime brushing. Objective indicators (detailed sugar logs, plaque biomarkers) should track programmed impact. The Australian School Dental Service shows that coupling toothpaste distribution, sealants, nutrition monitoring, and one-off subsidies for low-SES kindergartens reduces mean dmft by 0.5-0.8 within 2 years, underscoring the synergy of structural and behavioural measures.²⁴

Consistent with the observed SES–dmft gradient and attenuated belief/behaviour paths, structural supports should accompany education, especially for families with constrained resources. Priority options include coverage and co-payment waivers for fluoride varnish and sealants, subsidised distribution of fluoride toothpaste to households with preschoolers, and the integration of preventive care with routine child health contacts (e.g., immunisation visits, kindergarten health checks). At the service/environment level, kindergarten-based supervised toothbrushing and school food standards that limit free sugars reduce reliance on parental knowledge alone. For underserved neighbourhoods, mobile/community dental teams and tracking of service density and fluoride-programme participation can better target resources. Where feasible, earmarked revenues from sugar-sweetened beverage policies can finance preschool prevention. In light of our age-stratified results, targeting ages 4–5 may maximise yield. These recommendations are presented as policy-relevant associations rather than causal claims and should be evaluated prospectively.

This work emphasises harmonised indicators and structural supports, consistent with Broomhead et al.,²⁶ which demonstrate how standardised international oral-health datasets enable equity monitoring and cross-country comparability in service and outcome measures. In parallel, the 2-year longitudinal study by Chou shows that parental health beliefs prospectively predict ECC, indicating that our cross-sectional null mediation is likely influenced by measurement reliability and temporality, rather than the true absence of effect. ²⁷

Limitations and future research

The non-significant BehaviourRisk – dmft path likely reflects under-measurement of key behaviours and latency between exposures and manifest caries. Brief self-reports omit diet timing and stickiness (e.g., bedtime sugars, retentive snacks) and post-meal/brushing supervision, which can attenuate path estimates. It should incorporate multi-day 24-hour diet logs with timing/frequency/stickiness, together with bedtime-brushing supervision logs and fluoride exposure (toothpaste concentration, varnish/sealants), together with objective plaque indices (visible or photographic; optionally plaque ATP fluorescence) to reduce measurement error and better capture behaviour-to-caries translation.

This cross-sectional, single-city study cannot establish causality, may be subject to self-report bias, and limits generalisability. Future work should adopt multi-centre longitudinal designs with repeated 24-h/3-d sugar logs, pre-sleep plaque staining, and salivary biomarkers to track cumulative behavioural effects. IRT-based refinement and invariance testing across SES strata would strengthen measurement. A cluster-randomised kindergarten trial comparing a comprehensive package of structural subsidies, objective monitoring, and parental decision training with standard education could test for a 1-year dmft reduction. Overall, our SEM findings highlight a structure-priority, behaviour-lag, and cognition-weak pattern, offering a quantitative basis for staged, hierarchical interventions.