Development of geographic inequality in dental caries and its association with socioeconomic factors over an 18-year period in Denmark

This was a nationwide, register-based repeated cross-sectional study. It was conducted on all individuals in Denmark born in 1980, 1988, and 1998, who consequently became 15 years old in 1995, 2003, and 2013, respectively. The age group of 15 years was chosen as it is a World Health Organization (WHO)-recommended index age group for conducting oral health assessments in children [9].

Dental caries information was compiled from the national dental database, Sundhedsstyrelsens Centrale Odontologiske Register (SCOR). SCOR has been administered since 1972 by the Danish Health Authority. In SCOR, dental professionals systematically record data from dental visits made by children and adolescents participating in the national dental service. The register covers about three-fourths of all Danes < 18 years of age, particularly the age groups of 5, 7, 12, and 15 years who are expected to undergo a mandatory dental examination [7]. At the time of conduct of the study, the most recent and updated information in SCOR was available for 2013 while the earliest good quality data were obtainable for 1995.

The registers used to obtain sociodemographic data in this study included the population [10], education [11], income and transfer payments [12], labour market affiliation [13], and household and family registers [14]. A unique personal identification number (PIN) is assigned to each resident in Denmark at birth or immigration and is managed by the Danish Civil Registration System (CRS) [15]. This PIN was used to identify, extract, and link each 15-year-old’s data from the national registers with clinical data from SCOR. The Danish CRS includes PINs of parents, which were used to obtain information on parental social variables. Data on neighbourhood characteristics were computed from information publicly available at Statistics Denmark’s Statbank [16].

Adolescents were excluded from the analyses if they lacked dental data in SCOR or had missing information on gender, residence municipality, or one of the SEP variables. Those with negative household income were also excluded to avoid misclassification of income, as they could either be from poor families or rich business households registering negative income in a particular year.

Data management and statistical analyses were accomplished using SAS (version 9.4) and R (version 3.5.0). Individual- and neighbourhood-level data were described using frequency distributions, or prevalence proportions. The associations between the SEP variables and caries were analysed using Bayesian zero-inflated negative binomial regression models. A zero-inflated model was chosen to account for the high proportion of DMFS counts being 0 [2]. The three individual-level SEP variables were included individually as fixed effects; the models were adjusted for individual- and neighbourhood-level confounders (as fixed effects). The models also included a random effect component, modelled using a Besag-York-Mollié (BYM) model [19]; this component included (i) a geographically (spatially) structured term (to account for any spatial autocorrelation of DMFS values) and (ii) an unstructured term to account for any overdispersion of the data. Spatial autocorrelation, i.e., clustering of DMFS values in neighbouring municipalities, is possible as in general, neighbouring geographic areas tend to have similar health outcomes than those farther apart [20]. Spatial autocorrelation was accounted for by creating a binary neighbourhood adjacency matrix representing the adjacency structure of the Danish municipalities (a 275 × 275 matrix in which municipalities that shared a common boundary were coded as neighbours; the Danish islands were coded as neighbours to the municipalities to which they were connected by ferry or bridge). The geographically structured random effect was assigned an intrinsic conditionally autoregressive (CAR) prior, the hyperparameter of which was assigned a log-gamma distribution with parameters (1, 0.001). The unstructured random effect was modelled assuming an exchangeable Gaussian prior [21]. The fixed effects were assigned default Gaussian (0, 0.001⁻¹) prior distributions.

Using these zero-inflated BYM CAR models, two separate analyses were performed. The first analysis was a spatial analysis, stratified by study year (Model 1). The second analysis was a spatiotemporal analysis in which data from the three study years were combined and a joint analysis performed with “year” additionally included as a fixed effect (Model 2). For both analyses, integrated nested Laplace approximations (INLA) in R were used for Bayesian parametric inference [22]. Parameter estimates of the fixed-effect variables were derived using posterior mean DMFS ratios (i.e., ratios of mean DMFS scores from the various SEP and other fixed-effect variable subgroups), while the 2.5 and 97.5 percentiles were used to define the corresponding 95% credible intervals (CIs). For each variable, the subgroup having the lowest mean DMFS score (i.e., the best-off category) was used as the reference for obtaining the ratios. The relative percentage contribution of the geographically structured random effect to the total variance was calculated using a simulation-based approach as described elsewhere [21].

Subsequently, one sensitivity analysis was performed to evaluate the impact of the chosen prior distributions on the parameter estimates. This was assessed by changing the parameters of the specified distributions to the default values supplied by INLA (Model 3).

The major strengths of this study are the large nationwide population and inclusion of all Danish municipalities, enhancing their representativeness of the Danish adolescent population. The use of national register data of proven validity decreases the scope for misclassification of exposures and outcomes and limits biases such as recall and non-response bias [25]. Another strength of the study is that all three commonly used individual-level SEP variables, namely, education, occupational social class, and income, were examined in the study. For a multidimensional construct like SEP, any single SEP indicator is unlikely to encompass the entirety of the effect of SEP on caries. Analysing multiple indicators helps minimise residual confounding through unmeasured socioeconomic circumstances [26]. Furthermore, in contrast to most of the earlier studies in this area (e.g., Ekstrand et al. 2010 [6], Nørrisgaard et al. 2016 [7]), advanced regression modelling was implemented that allowed simultaneous adjustment for several potential individual- and neighbourhood-level confounders as well as control for geographic clustering. This would potentially reduce residual confounding further and improve the precision of the estimates.

Potential limitations of this study include the possibility of selection bias despite the large nationwide population. This is because adolescents lacking dental data and data on the SEP indicators and residence municipality were excluded. Approximately 15% to 23% of the 15-year-olds in the three study years were excluded because they were not registered in SCOR. This could engender selection bias if being registered in SCOR, i.e., if undergoing a (mandatory) dental examination at age 15, depended on both the exposure (SEP) and the outcome (caries) [27]. The possibility that the exposure (SEP) determines whether or not an adolescent undergoes a dental examination is mitigated by the fact that when the included (i.e., had a dental examination) and excluded (no dental examination) 15-year-olds were compared, no marked differences were seen in terms of the proportions of individuals comprising each SEP subgroup (Appendix 1 a. 1995; b. 2003; c. 2013). Prior reports imply greater utilisation of dental services by children and adolescents from lower SEP backgrounds in the Nordic countries (presumably because of having higher disease levels and [universal] access to dental care) [2, 28]. Thus, it is less likely that diseased individuals from lower SEP backgrounds are not being able to access care and are systematically missing from SCOR. A possibility exists that the lack of a dental examination at age 15 may be due to relocation issues; however, these are unlikely to be systematically related to the study variables. It may also be that the adolescents had extended recall intervals (i.e., less frequent dental examinations) because of low risk of caries or other dental problems [29]. Absence of individuals without caries may reduce the representativeness of the study population. However, as indicated earlier, in terms of the study exposures and covariates, these individuals were not systematically different from those having dental caries data in SCOR. Therefore, because of its inherent property of scale invariance [30], the relative caries ratio estimates from this study would not change even if these individuals without disease were included. Thus, missing observations are not likely to induce significant selection bias in this study.

Misclassification of caries is a possibility in this study because many different uncalibrated dental examiners from all over Denmark perform diagnosis and recording of caries in SCOR. To alleviate misclassification, we limited caries diagnosis to cavitated lesions only (which are more clearly visible to the naked eye). Moreover, the risks of misclassification is lowered because the caries recording system in SCOR follows a relatively simple coding methodology based on detailed guidelines provided by the Danish Health Authority [31]. In addition, the diagnostic criteria for caries in Denmark are well defined and correspond with WHO standards [9]. Despite this, independent errors in caries diagnosis or data entry may occur. The impact of such errors depends on whether they are dependent and/or differential. In the Danish context, any such individual error would be independent and non-differential with respect to the exposure variables, as these variables were recorded autonomously in the registers [2]. Thus, misclassification is unlikely to be an important weakness in this study.

To minimise individual- and neighbourhood-level confounding, ethnicity, household size and type, and diverse factors related to neighbourhood income and deprivation were taken into consideration. Moreover, several important background factors were held constant for the study population in each assessment year, including universal access to dental and medical care and a basic welfare system. Nevertheless, residual confounding is still possible because of unmeasured potential common causes of the SEP exposures and the outcome (caries). These could be related to the family environment (e.g., parental somatic and mental disorders, family functioning etc.) as well as factors outside of the family environment. At the neighbourhood level, these could likely include contextual elements specific to municipalities (e.g., infrastructural aspects of dental care management such as children:dentist and auxiliary personnel:dentist ratios, the proximity of homes and schools to clinics, transportation systems affecting mobility and impetus to seek dental care, the quality of food available in stores and barriers to making healthier food choices, the degree of social support and social cohesion) [32, 33]. Owing to the complexity of measuring neighbourhood characteristics, such contextual aspects of the neighbourhood framework were likely not adequately taken into account by the variables included in study.

In terms of defining the neighbourhood framework per se, we used pre-2007 municipalities (n = 275), which provided a finer geographical resolution than the current set of municipalities (n = 98). We chose municipalities as geographical units because of their fundamental role in the organisation and implementation of dental care in Denmark. Moreover, data on neighbourhood income and deprivation, including the Gini index, are available for municipalities. However, it may be argued that some municipalities are relatively large and heterogeneous areas with considerable within-municipality variations in individual-level socioeconomic factors and caries levels (e.g., Copenhagen). A smaller, more homogenous unit might be more reflective of an individual’s immediate neighbourhood. Alternative definitions of neighbourhoods relevant to oral health may be based on locations of services such as dental care institutions or proximity to stores selling sugary foods.