2.1

Sources

Electronic searches were carried out in MEDLINE via PubMed, EMBASE via OVID, LILACS via BIREME, Web of Science, and Google Scholar. The search strategy was three-pronged, combining the condition (hypomineralization OR hypomineralisation OR hypomineralized OR hypomineralized OR hypoplasia OR demarcated OR opacities OR MIH OR cheese molars), the study type/outcome (survey OR questionnaire OR cross-sectional OR prevalence OR frequency OR population OR sample OR sampling), and the teeth of interest (molar OR molars OR incisors). Blocks were combined using the Boolean operator AND.

The date of publication was restricted to studies published from 2000 to May 2017. No further restrictions or limitations to the search were made. The search was complemented by cross-referencing, by screening through available reviews on this topic and hand searches as well as web-pages of conferences, governments and international health organizations.

2.2

Study selection

We included the following studies:

• (1)

  Observational studies, regardless of the type (cross-sectional, cohort, case-control), published after peer-review. Studies which were so far only published as theses, reports, audits etc., but did not undergo peer-review, were not included (note that abstracts were included, if sufficient details were reported, as these usually undergo some kind of peer-review). Case-control studies where the case definition was MIH, studies where populations with general diseases possibly affecting the prevalence of MIH, and studies which focused on non-representative samples (like institutionalized populations, particular professions, those with specific dental outcomes like high caries experience) were also excluded.

• (2)

  Studies needed to report on the prevalence of MIH in a sample of individuals (for case-control studies, this population was usually the control group). Note that we had originally planned to only include studies which used random or comprehensive sampling, but found a large proportion of studies did not report on the sampling strategy. We have thus included these studies and tested the impact of this inclusion in a sensitivity analysis.

• (3)

  MIH needed to be either defined according to the European Academy of Paediatric Dentistry (EAPD) definition or its modifications, or as a component of other indices (e.g. DDE index) . The impact of the case definition was tested in a sensitivity analysis. Studies which evaluated the prevalence of enamel defects not restricted to MIH (e.g. diffuse opacities), and which did not allow the extraction of the MIH component were excluded. Only studies which reported on MIH, i.e. the described defect of permanent molars with or without affected permanent incisors, were included. Studies which only reported on primary molars were excluded.

No language restrictions were set; if needed, publications were translated in parts or total. A restriction of maximal non-response rates was not set, as many studies had not reported on these in detail. The search was performed by two reviewers (FS, KE) independently and in duplicate. Inclusion and exclusion were performed in consensus.

2.3

Data extraction

Data extraction was performed by two reviewers independently and in duplicate (FS, KE). A pilot-tested spreadsheet (Excel, Microsoft, WA, USA) was used for data entry. The following items were extracted: (1) First author’s name, (2) year of publication, (3) year of study conduct, (4) country and place (region, city) of sampling, (5) sampling frame (population, school, hospital), (6) sampling strategy (random, comprehensive, other or not reported), (7) case definition (EAPD and its modification, or others), (8) setting of oral examination (school, dental hospital, or other), (9) oral examination inspection aids used (i.e. dental lights etc.), (10) preparation of the teeth prior to inspection (drying or not), (11) sample size and non-response, (12) number of individuals affected by MIH, if possible separated into male and female sex (to test for possible sex differences in prevalence), (13) number of individuals in need of dental care, i.e. with symptoms (hypersensitivity) and/or post-eruptive breakdown. If needed, authors were contacted to gather missing information.

2.4

Quality assessment

The quality of studies was assessed using a modified version of the Newcastle Ottawa scale . Four domains were assessed, each receiving one score (max. number 4): (1) Representativeness of sample (was the sample randomly drawn from a population with representativeness for the general population of interest in the setting, or was a comprehensive sample used?), (2) sample size estimation (had a sample size estimation be performed a priori?), (3) non-response (did the authors pay consideration to non-response or attrition and did they allow to assess it quantitatively; if so, was it <50%, or was any analysis performed to gauge the impact of non-response or attrition?), (4) validity assessment method (was the examiner calibrated against a reference standard prior to examination, and was an accepted assessment method used, with defined and reproducible criteria for cases?). This information was not used to include or exclude studies from data synthesis, but to gauge the possible impact of selection, attrition and assessment bias on the prevalence.

2.5

Meta-analysis

Prevalence was synthesized on global, super-regional, regional and national level using random-effects meta-analysis (Comprehensive Meta-Analysis 2, Biostat, Englewood, NJ, USA). In a sensitivity analysis, we consecutively left out data from exactly one study (leave-one-out meta-analysis) and assessed how each individual study affected the overall estimate. Such analysis allows to assess if the synthesized estimates were robust against outliers. 95% confidence intervals were used to characterize uncertainty. Egger-test and funnel plot analyses were used to assess possible publication bias . Meta-analysis was also used to perform subgroup or sensitivity analyses according to sex (male versus female), treatment need (i.e. with subjective symptoms and/or posteruptive breakdown versus no need), MIH case definition (EAPD versus other), and sampling strategy (random/comprehensive versus other/unknown).

2.6

Predictor variables

In order to estimate MIH prevalence for countries where no observational study was reported (see below), the response variable (the pooled MIH prevalence on country scale) was log-transformed and regressed on various predictor variables from the GBD 1980–2015 covariates dataset ( www.ghdx.healthdata.org ), as described below. The GBD 1980–2015 covariates dataset provides country-level data for 296 variables from 195 countries and includes, for example, variables on socioeconomic and demographic aspects, health system access, climate, and food consumption. Most data are reported on annual scale and in many cases for both sexes. From this dataset, variables were divided into two subsets, one selective set informed by the expertise of the authors (GBD_sub1), and another more comprehensive set including most available covariates, except those referring to summary exposure values and those containing missing values (GBD_sub2). If possible we included covariates for each sex and summarized the data for the period from 2001 to 2015 by taking the arithmetic mean. Covariates on proportional or percentage scale were log-transformed before further processing. In order to account for potential covariate interactions, second order polynomial feature expansion on both covariate subsets was applied. However, this step did not alter the model performance significantly.

2.7

Regression analysis

For regression analysis we used Python 3.5 ( www.python.org ) and in particular the pandas and the scikit-learn packages, as well as the statistical programming language R ( www.r-project.org ) and in particular the glmnet and the mht packages. We applied regularized and non-regularized regression analyses to predict MIH prevalence on country scale ( Fig. 2 ): (1) A baseline model, which corresponds to the arithmetic mean of the response variable. (2) Ordinary least-square (OLS) regression. Regularized models such as (3) lasso (least absolute shrinkage and selection operator), also referred to as ℓ1-norm, and (4) ridge regression, also referred to as ℓ2-norm. Lasso performs variable selection by causing less predictive model coefficients to become zero and in ridge regression less predictive model coefficients are shrunken towards zero . For lasso, bootstrap-enhanced least absolute shrinkage operator, referred to as bolasso, was also used . Bolasso iteratively performs lasso over (in our case 10,000) bootstrap iterations and returns an appearance frequency which approximates the predictive power of each covariate. (5) Elastic net regression combines ℓ1 and ℓ2-norm and results in sparse covariate vectors while maintaining the regularization properties of ridge regression . (6) Principal component analysis (PCA) for dimensionality reduction as a preprocessing step for the high dimensional GBD_sub2 dataset (d = 210). The number of principal components to be kept was determined by accounting for an explained variance of at least 0.9.

Modeling scheme for estimation of MIH prevalence on country level. With respect to the available variables in the GBD 1980–2015 covariates dataset, we split the data set into two subsets (gray rectangular boxes). For one subset, variable selection was informed by the expertise of the authors (GBD_sub1) and for the other subset most available covariates were included (GBD_sub2). For each data subset, different linear modelling approaches and (if appropriate) preprocessing steps were applied (hexagonal boxes). The predictive power of the different model configurations (ellipsoids) was evaluated using leave-one-out cross validation. Abbreviations: OLS: ordinary least square, PCA: principal component analyses, RR: ridge regression, EN: elastic net, Lasso: least absolute shrinkage and selection operator; see main text and Appendix for model description.

Using the different subsets (GBD_sub1, 2), eight different models were resultant. For each model its predictive power was evaluated by applying leave-one-out cross validation (LOOCV) and we ranked the models with respect to their root-mean-square error (RMSE). The RMSE is a commonly used model accuracy measure and corresponds to the square root of the average of squared differences between the actually observed values and values predicted by the model. We used the best performing model with respect to RMSE (lowest RMSE) to predict the MIH prevalence for countries where no measurement points were reported (see Appendix for details).

2.8

Estimation of global burden

Building on the resulting prevalence dataset and the GBD 1970–2015 population estimate dataset , the number of prevalent cases in 2015 and incident cases in 2016 were calculated, assuming all individuals aged 6 years or above to be possible prevalent cases, and all individuals aged 5 years in 2015 to be possible incident cases in 2016. To assess the robustness of our estimates, we additionally summed up the number of prevalent and incident cases on regional, super-regional and global level and compared it with the number resulting from multiplying the synthesized prevalence rates on these spatial scales with the overall population estimates. We found both ways of imputation-based estimation to not yield a significantly different burden of MIH.

2.1

Sources

Electronic searches were carried out in MEDLINE via PubMed, EMBASE via OVID, LILACS via BIREME, Web of Science, and Google Scholar. The search strategy was three-pronged, combining the condition (hypomineralization OR hypomineralisation OR hypomineralized OR hypomineralized OR hypoplasia OR demarcated OR opacities OR MIH OR cheese molars), the study type/outcome (survey OR questionnaire OR cross-sectional OR prevalence OR frequency OR population OR sample OR sampling), and the teeth of interest (molar OR molars OR incisors). Blocks were combined using the Boolean operator AND.

The date of publication was restricted to studies published from 2000 to May 2017. No further restrictions or limitations to the search were made. The search was complemented by cross-referencing, by screening through available reviews on this topic and hand searches as well as web-pages of conferences, governments and international health organizations.

2.2

Study selection

We included the following studies:

• (1)

  Observational studies, regardless of the type (cross-sectional, cohort, case-control), published after peer-review. Studies which were so far only published as theses, reports, audits etc., but did not undergo peer-review, were not included (note that abstracts were included, if sufficient details were reported, as these usually undergo some kind of peer-review). Case-control studies where the case definition was MIH, studies where populations with general diseases possibly affecting the prevalence of MIH, and studies which focused on non-representative samples (like institutionalized populations, particular professions, those with specific dental outcomes like high caries experience) were also excluded.

• (2)

  Studies needed to report on the prevalence of MIH in a sample of individuals (for case-control studies, this population was usually the control group). Note that we had originally planned to only include studies which used random or comprehensive sampling, but found a large proportion of studies did not report on the sampling strategy. We have thus included these studies and tested the impact of this inclusion in a sensitivity analysis.

• (3)

  MIH needed to be either defined according to the European Academy of Paediatric Dentistry (EAPD) definition or its modifications, or as a component of other indices (e.g. DDE index) . The impact of the case definition was tested in a sensitivity analysis. Studies which evaluated the prevalence of enamel defects not restricted to MIH (e.g. diffuse opacities), and which did not allow the extraction of the MIH component were excluded. Only studies which reported on MIH, i.e. the described defect of permanent molars with or without affected permanent incisors, were included. Studies which only reported on primary molars were excluded.

No language restrictions were set; if needed, publications were translated in parts or total. A restriction of maximal non-response rates was not set, as many studies had not reported on these in detail. The search was performed by two reviewers (FS, KE) independently and in duplicate. Inclusion and exclusion were performed in consensus.

2.3

Data extraction

Data extraction was performed by two reviewers independently and in duplicate (FS, KE). A pilot-tested spreadsheet (Excel, Microsoft, WA, USA) was used for data entry. The following items were extracted: (1) First author’s name, (2) year of publication, (3) year of study conduct, (4) country and place (region, city) of sampling, (5) sampling frame (population, school, hospital), (6) sampling strategy (random, comprehensive, other or not reported), (7) case definition (EAPD and its modification, or others), (8) setting of oral examination (school, dental hospital, or other), (9) oral examination inspection aids used (i.e. dental lights etc.), (10) preparation of the teeth prior to inspection (drying or not), (11) sample size and non-response, (12) number of individuals affected by MIH, if possible separated into male and female sex (to test for possible sex differences in prevalence), (13) number of individuals in need of dental care, i.e. with symptoms (hypersensitivity) and/or post-eruptive breakdown. If needed, authors were contacted to gather missing information.

2.4

Quality assessment

The quality of studies was assessed using a modified version of the Newcastle Ottawa scale . Four domains were assessed, each receiving one score (max. number 4): (1) Representativeness of sample (was the sample randomly drawn from a population with representativeness for the general population of interest in the setting, or was a comprehensive sample used?), (2) sample size estimation (had a sample size estimation be performed a priori?), (3) non-response (did the authors pay consideration to non-response or attrition and did they allow to assess it quantitatively; if so, was it <50%, or was any analysis performed to gauge the impact of non-response or attrition?), (4) validity assessment method (was the examiner calibrated against a reference standard prior to examination, and was an accepted assessment method used, with defined and reproducible criteria for cases?). This information was not used to include or exclude studies from data synthesis, but to gauge the possible impact of selection, attrition and assessment bias on the prevalence.

2.5

Meta-analysis

Prevalence was synthesized on global, super-regional, regional and national level using random-effects meta-analysis (Comprehensive Meta-Analysis 2, Biostat, Englewood, NJ, USA). In a sensitivity analysis, we consecutively left out data from exactly one study (leave-one-out meta-analysis) and assessed how each individual study affected the overall estimate. Such analysis allows to assess if the synthesized estimates were robust against outliers. 95% confidence intervals were used to characterize uncertainty. Egger-test and funnel plot analyses were used to assess possible publication bias . Meta-analysis was also used to perform subgroup or sensitivity analyses according to sex (male versus female), treatment need (i.e. with subjective symptoms and/or posteruptive breakdown versus no need), MIH case definition (EAPD versus other), and sampling strategy (random/comprehensive versus other/unknown).

2.6

Predictor variables

In order to estimate MIH prevalence for countries where no observational study was reported (see below), the response variable (the pooled MIH prevalence on country scale) was log-transformed and regressed on various predictor variables from the GBD 1980–2015 covariates dataset ( www.ghdx.healthdata.org ), as described below. The GBD 1980–2015 covariates dataset provides country-level data for 296 variables from 195 countries and includes, for example, variables on socioeconomic and demographic aspects, health system access, climate, and food consumption. Most data are reported on annual scale and in many cases for both sexes. From this dataset, variables were divided into two subsets, one selective set informed by the expertise of the authors (GBD_sub1), and another more comprehensive set including most available covariates, except those referring to summary exposure values and those containing missing values (GBD_sub2). If possible we included covariates for each sex and summarized the data for the period from 2001 to 2015 by taking the arithmetic mean. Covariates on proportional or percentage scale were log-transformed before further processing. In order to account for potential covariate interactions, second order polynomial feature expansion on both covariate subsets was applied. However, this step did not alter the model performance significantly.

2.7

Regression analysis

For regression analysis we used Python 3.5 ( www.python.org ) and in particular the pandas and the scikit-learn packages, as well as the statistical programming language R ( www.r-project.org ) and in particular the glmnet and the mht packages. We applied regularized and non-regularized regression analyses to predict MIH prevalence on country scale ( Fig. 2 ): (1) A baseline model, which corresponds to the arithmetic mean of the response variable. (2) Ordinary least-square (OLS) regression. Regularized models such as (3) lasso (least absolute shrinkage and selection operator), also referred to as ℓ1-norm, and (4) ridge regression, also referred to as ℓ2-norm. Lasso performs variable selection by causing less predictive model coefficients to become zero and in ridge regression less predictive model coefficients are shrunken towards zero . For lasso, bootstrap-enhanced least absolute shrinkage operator, referred to as bolasso, was also used . Bolasso iteratively performs lasso over (in our case 10,000) bootstrap iterations and returns an appearance frequency which approximates the predictive power of each covariate. (5) Elastic net regression combines ℓ1 and ℓ2-norm and results in sparse covariate vectors while maintaining the regularization properties of ridge regression . (6) Principal component analysis (PCA) for dimensionality reduction as a preprocessing step for the high dimensional GBD_sub2 dataset (d = 210). The number of principal components to be kept was determined by accounting for an explained variance of at least 0.9.

Only gold members can continue reading. Log In or Register to continue

Share this:

• Click to share on Twitter (Opens in new window)
• Click to share on Facebook (Opens in new window)

Related

Related posts:

1. Restoration of severely damaged endodontically treated premolars: The influence of the endo-core length on marginal integrity and fatigue resistance of lithium disilicate CAD-CAM ceramic endocrowns
2. Influence of the addition of microsphere load amoxicillin in the physical, chemical and biological properties of an experimental endodontic sealer
3. Impact of prosthodontic rehabilitation on the masticatory performance of partially dentate older patients: Can it predict nutritional state? Results from a RCT
4. Higher experience of caries and lower income trajectory influence the quality of restorations: A multilevel analysis in a birth cohort
5. Experimental use of an acrolein-based primer as collagen cross-linker for dentine bonding
6. Cost-effectiveness analysis of two impression methods for the fabrication of mandibular complete dentures

Stay updated, free dental videos. Join our Telegram channel

Join