This analysis was based on data originating from an epidemiological study conducted during year 2000 in Santiago, Chile [23
]. The local Committee of Ethics of the University of Chile approved the study protocol. The target population comprised all students attending the four grades covering adolescence in the high schools of the Province of Santiago (N ≈ 250,000). This target population represents 85% of the adolescent population of the Province [25
]. We used a two-stage random-cluster sampling strategy. Using information on governmental support and the full list of high schools from the Province provided by the Ministry of Education of Chile (N = 618), we generated a list of high schools receiving funds from the public system and another with the private institutions. Each list was permuted at random [26
], and lists were then merged to get a random permutation of high schools with publicly funded schools alternating with privately funded schools. The first 133 high schools of the list were contacted to obtain information on the number of students in the last four grades and the number of classes. A total of 104 high schools provided the necessary information and were invited to participate in the study. Six institutions declined to participate, leaving 98 schools in the study.
Second-stage sampling: The size of the schools varied noticeably and a second sampling stage was applied. For schools with few students or fewer than 4 classes, all classes were included in the study. In larger schools, where the number of students in the last four grades was >100 and the number of classes was >3, three classes were randomly selected [23
]. A total of 310 classes were finally selected and included in the study.
Participation approval was obtained from the headmasters of each selected high school, and informed consent was obtained from the parents of the students. Students were informed about their right to withdraw from the study at any point in time.
A total of 9,203 students aged 12–21 years present in the selected classes were invited to participate and were offered a toothbrush for their participation. All students accepted to fill a brief questionnaire on oral health-related behaviors and conditions [23
], while 40 students refused to participate in the clinical examinations. The questionnaire included information on tooth brushing frequency (How often do you brush your teeth? Less than once a day, once a day, more than once a day), smoking habits (Do you smoke cigarettes? No; Yes, sometimes; Yes, daily), their last dental visit (When was the last time you visited a dentist? Less than 6 months ago; 6 to 12 months ago; more than a year ago; never seen a dentist) and the reason for the last dental visit. Why did you visit the dentist?) [23
]. The students who received a clinical examination also filled an additional questionnaire on several dimensions of their socioeconomic position. A full description of the questionnaire variables can be found in previous publications [23
]. Previous analyses showed that some of these indicators were associated with several poor oral health outcomes, including tooth loss, periodontal attachment loss and necrotizing ulcerative gingival lesions [27
] and these were therefore used in the present analysis. These socioeconomic indicators include the monthly paternal income in thousands of Chilean pesos (no income; <$100; $100–$299; $300–$499; $500–$999; ≥$1000); and the level of education achieved by each of the parents (no education; incomplete primary school; primary school completed; incomplete high school; high school completed; incomplete technical education; technical education completed; incomplete university education; university education completed). The headmasters of the participating schools provided information on the monthly tuition and the annual school fees and this information was used to derive a single 'annual education expense' variable, which was used as an indicator of wealth. Three dichotomous outcome variables were considered. DENTVIS was coded 0 if the student reported to have visited a dentist within the last year, and 1 if this was not the case. SYMPTOM was coded 1 if the student reported that the last dental visits had been due to symptoms, such as pain, bleeding and infection, and coded 0 if the visit was not prompted by the presence of symptoms. NEVER was coded 1 if the student reported to never have visited a dentist, and coded 0 if this was not the case.
Univariable logistic regression analyses were carried out for the three outcomes investigated and the covariates gender; tooth brushing frequency; smoking habits; self-perceived oral health status, and each of the socioeconomic indicators investigated. Variables showing a P-value < 0.25 in the univariable analyses were selected to be included as covariates in age-adjusted multivariable logistic regression analyses for each outcome. The models were built by the consecutive exclusion of one variable from each full model using the likelihood ratio test as described by Hosmer and Lemeshow [29
], refitting and verifying the stability of the model after each deletion.
In order to use this modeling approach, only subjects for whom complete data were available for all variables in all models were included in the analyses. At most this led to the exclusion of 11% of the total study population.
Non-significant variables were retained in the models as confounders if their exclusion would result in a change of the estimates by more than 15%. We assessed the interaction between age and gender and the socioeconomic variables included in the models and none was found. It was impossible to assess the presence of other interactions due to collinearity between the variables. Once the final models were built, the goodness-of-fit of each model was evaluated using the Hosmer-Lemeshow goodness-of-fit test using the command 'lfit' in Stata [30
]. The option 'robust cluster' for the procedure 'logit' in Stata [30
] was used to take into account the fact that the students were nested in classes, which were the ultimate sampling units [23