Participants and procedure

For the purpose of this study, N = 565 online responses collected from the National Dental Telephone Interview Survey (NDTIS) 2021 Follow-up Questionnaire (between June 22, 2021, and May 13, 2022) were included. Participants were aged between 19 and 91 years (M = 54.50, SD = 16.16), the majority of which were female (60.9%), employed either full-time (37.65%) or part-time (22.02%), and had completed an undergraduate degree or higher (70.1%). The NDTIS 2021 was a national survey conducted by the Australian Research Centre for Population Oral Health (ARCPOH) that aimed to obtain information about the frequency of oral health problems among the Australian general population, as well as the psychosocial factors associated with oral health problems. These included the impact of the COVID-19 pandemic, socioeconomic position, quality of life, among others. NDTIS 2021 participants were randomly selected as a representative sample of Australian adults aged 18 years and over. Following completion of the primary survey (conducted either by telephone interview or online), participants were invited to respond to the follow-up questionnaire and given the option to provide their answers over the phone or by completing a paper or online version. The NDTIS 2021 was approved by the Human Research Ethics Committee at the The University of Adelaide (approval number H-2020–153). All participants provided verbal consent to participate in the survey, and datasets were de-identified to ensure anonymity.

Theoretical framework and the selection of variables

The theoretical framework applied to this study was grounded in Andersen’s (1995) Behavioral Model (ABM, see Fig. 2), which proposes that health behaviors, such as personal health practices and the use of health services, are influenced by both environmental factors (e.g., health care system and the external environment) and population characteristic factors (e.g., predisposing characteristics, enabling resources, and need). The ABM has been widely used in healthcare research and continues to be an important framework for understanding healthcare utilization patterns across populations and health conditions, including preventive dental visiting (Harris et al. 2017).

Fig. 2

Andersen’s Behavioral Model (ABM). Adapted from “Revisiting the Behavioral Model and Access to Medical Care: Does it Matter?” by R. M. Andersen 1995, Journal of Health and Social Behavior, 36, p. 8. The model proposes causes between the environment, population characteristics, health behaviors, and outcomes

We identified six domains specific to the COVID-19 pandemic influencing oral health self-management behavior, each of which aligned with components of the ABM (i.e., external environment, enabling factors, and predisposing factors). Fear of COVID-19 disease was identified as the major external environmental factor impacting dental visitation during the pandemic, driven by concerns over infection risks in dental settings. In addition, fear and anxiety associated with the pandemic may have disrupted routines or led to increased stress levels, which can affect both oral hygiene habits and dental visitation negatively (Sari and Bilmez 2021). Three enabling factors, including accessibility to dental appointments, availability of transportation to dental appointments, and affordability of dental care (Beck et al. 2021; Burgette et al. 2021; Cook et al. 2020) were identified as possible barriers to oral health behaviors, potentially preventing individuals from seeking timely preventive or therapeutic oral health interventions. According to Cook et al. (2020), during the COVID-19 pandemic, Australians experienced limited access to both dental services and transportation due to restrictions imposed during lockdown periods, and were also more likely to be negatively affected by the high costs associated with dental care due to higher levels of unemployment. Finally, the ability to discuss oral health issues was identified as a predisposing factor, since openly discussing oral health concerns with peers, family members, and/or non-dental healthcare professionals can lead to increased awareness, motivation, and adherence to oral health self-management practices (Cohen et al. 2000).

Fear of COVID-19 Scale

Fear of COVID-19 disease was measured using three items from the Fear of COVID-19 Scale (Ahorsu et al. 2022). These included "I am most afraid of COVID-19" (Afr), "I am afraid of losing my life because of COVID-19" (Dth), and "I cannot sleep because I’m worried about getting COVID-19" (Ins). Participants responded on a five-point response Likert scale (1 = strongly disagree, 2 = disagree, 3 = neither disagree nor agree, 4 = agree and 5 = strongly agree). This scale displayed good internal consistency reliability (ω = 0.80).

Oral health impacts related to COVID-19

Oral health impacts related to COVID-19 were measured by asking participants the following: "During the COVID-19 pandemic, were you able to …” (a) “discuss your dental or oral health with non-dental professionals during the pandemic?” (Discuss), (b) “make an appointment with a dentist or oral health practitioner for any dental or oral health needs?" (Access), (c) “afford dental care?" (Afford), (d) “find transport required to attend for a dental appointment?" (Transp), and (e) “find the energy to manage your dental or oral health?” (OHSM). These oral health-specific survey items were responded to on a five-point response Likert scale (1 = unable to do so, 2 = with a lot of difficulty, 3 = with some difficulty, 4 = with a little bit of difficulty, and 5 = without difficulty). This scale also displayed good internal consistency reliability (ω = 0.82).

Statistical analysisSoftware packages

All statistical analyses were conducted using R packages EGAnet (Golino and Christensen 2022), bnlearn (Scutari 2019), NCT (van Borkulo et al. 2022), lavaan (Rosseel 2012), and caret (Kuhn 2008).

Data preparation

All items measuring population characteristics were reverse scaled as they were written with a positive valence. For example, while the item “afford dental care?” was reverse scaled, the item “I am most afraid of COVID-19” was not, since it was written with the required negative valence. In this way, higher scores across all items indicated stronger negative impacts of the pandemic (being less able to afford dental care, stronger fear of COVID-19, etc.).

To understand the nature of missing data, missingness was first summarized by item and then demographic factors (i.e., sex, education achievement, and employment status). Complete cases made up 75% (426/565) of the observed data, with the remaining 15% (86/565) of cases missing one item, 2% (14/565) missing two items, 3% (18/565) missing three items, 2% (9/565) missing four items, and 2% (10/565) of cases missing five items. The item Discuss (“…discuss your dental or oral health with people other than a dentist?”) displayed a comparatively high level of missingness at 18% (103/565) and was more likely missing from older respondents (i.e., > 50 years) than from younger respondents (i.e., ≤ 50 years), χ2 (5) = 18.6, p = 0.002.

Because rates of item missingness were > 5% and found to be distributed non-randomly across demographic factors, missing values were imputed using a single imputation method, the “bagged-tree” model. In bagged tree imputation, for each variable with missing data, a tree model is trained based on all other variables in the set, including demographic variables, and then the missing values are imputed using a regression function. This robust imputation model has been shown to outperform other imputation methods (e.g., k-nearest-neighbor imputation and median imputation), as measured by root mean square error (RMSE) between imputed and true values. Following Kuhn and Johnson (2019), the bagged-tree imputation method was utilized with n = 25 trees. The imputation model included OHSM as the outcome variable and all other survey items as predictors.

Gaussian graphical model (GGM)

The GGM (also known as the undirected graphical model) was estimated under the assumption of Gaussian distributions and models the conditional associations between variables (Christensen et al. 2020). Due to the ordinal nature of item scales, the polychoric correlation matrix was used as input. The GGM was estimated across N = 10,000 non-parametric bootstrapped samples using the graphical least absolute shrinkage and selector operator (GLASSO) based on minimization of the extended Bayesian information criterion (see Epskamp and Fried 2018). This method estimates the typical network structure formed by the mean partial correlations across the bootstrap samples. The model fit was evaluated with the root mean squared error of approximation (RMSEA) and comparative fit index (CFI). Following Kline (2015), values of CFI ≥ 0.95 and RMSEA < 0.05 were used to indicate good model fit. Metric, network, and global invariance were assessed across two demographic groups: sex (males vs. females) and median age (younger ≤ 57 years vs. older > 57 years). Metric invariance was assessed by comparing the strength of item network loadings, while network and global strength invariance were assessed by comparing network edge weights and the absolute sum of network edge weights, respectively (van Borkulo et al. 2022). All comparisons were based on 1,000 permutations. To ensure correct identification of statistical differences, two types of p-values were examined: uncorrected and those corrected using the Benjamini–Hochberg procedure (Benjamini and Hochberg 1995). According to Jamison et al. (2022), if both types of p-values pointed to a network loading difference being statistically significant across groups, then that item is considered non-invariant.

Directed acyclic graphical (DAG) model

The DAG model was estimated by employing the PC-stable algorithm across N = 10,000 non-parametric bootstrap samples (following Colombo and Maathuis 2014). Only directed edges appearing in 70% of bootstrap samples (i.e., strength ratio ≥ 0.70) were included in the model, and the directionality of the edge was established according to the direction that was observed in more than 50% of samples (i.e., direction ratio ≥ 0.50) (as suggested by Briganti et al. 2022b). The structure of estimated DAG was then assessed against the theoretical structure described by Andersen’s Behavioral Model (ABM).

Structural equation modeling (SEM)

The SEM was specified by including the causal pathways present in the estimated DAG. Due to the non-normal distribution of the survey data, the specified SEM was estimated using a maximum likelihood estimation with robust standard errors and a Satorra–Bentler scaled test statistic. Robust incremental (CFI) and absolute (χ2, RMSEA) fit indices were calculated and assessed, with values of CFI ≥ 0.95 and RMSEA < 0.05 applied to indicate acceptable model fit (Kline 2015).