We conducted, from August to November 2019, a school-based cross-sectional study in 20 primary schools in the Kilimanjaro region, Northern Tanzania. Kilimanjaro is one of the 31 administrative regions in Tanzania and comprises seven administrative districts. We purposely selected two districts, Moshi urban and Moshi rural, to represent urban and rural settings. Moshi urban has 48 primary schools (35 government schools and 13 private schools) and 33,207 students. Moshi rural has 272 primary schools (252 government: 20 private) and 81,297 students.
We used a multistage sampling technique to select the study participants. We randomly selected eight wards from each of the two districts. Primary schools from the selected wards were stratified by type of schools (government or private). We used simple random sampling to select 20 primary schools (10 from each district). Enrolment in government or private schools was regarded as a proxy indicator for lower or higher socio-economic status. We excluded children from boarding schools as they usually follow the same school menu. We obtained a list of all students aged 9 to 11 years from classes 4 to 6 from the attendance registers. From the list, we randomly selected study participants from each school according to probability proportional to the size of schools. This age band was set based on the assumption that children aged 9 or more years can express themselves and fill in questionnaires with guidance. Nevertheless, we excluded 24 children who could not understand the questions and express themselves.
Sample size estimation
Sample size estimations were based on the study’s primary aim: to estimate the prevalence of overweight and obesity . We assumed a precision of +/-2.5% and a design effect of 1.3 because we recruited participants from clusters (schools). The final sample size was 1170 primary school children.
Dietary Intake Assessment
Pilot study: 24-hour dietary recall
We first conducted a pilot study, a 24-hour dietary recall among 51 children (9 -11 years) randomly selected from 4 primary schools, two from each district (urban/ rural), to explore the food items usually consumed by children. A trained nutritionist asked children to recall their previous intake in 24 hours. We followed the four-stage multipass technique to obtain information on detailed food items consumed in the past 24 hours . Information on snacking habits, especially during recess time at school, lunchtime, on the way back home, and at friends’ houses, were used to get specific details for sugary drinks, snacks, and high-calorie low nutrition snacks. Food models were used to aid in portion size estimations . We used the Tanzania food composition table (TFCT)  to assign nutrient values and estimate different foods' energy and nutrient intake. For the industrially manufactured foods mentioned by children that were not on the TFCT, e.g., Pringles (dehydrated processed potato crisps), sausages, we obtained the information from food labels where possible. Nutrition information from each food was computed per 100 grams and then converted to the consumed amount based on the reported amount. For children who could not estimate their portion size, the standardised portion sizes from TFCT were used to evaluate their intake.
Food frequency questionnaire
For the main study of dietary intake, we used the findings from the 24-hour recall to adapt and modify the International Study on Childhood Obesity, lifestyle, and environment (ISCOLE) FFQ . This standard FFQ has been used elsewhere to explore the association between diet and obesity [20–22]. We removed some food items which did not apply to Tanzanian children (e.g., low-fat milk, cheese, energy drinks: “Redbull, guru”, sports drinks “Powerade”) and added the local foods / snacks recalled eaten frequently by children, e.g., samosas, chips, fried cassava, fried plantain, locally made ice creams (tamarind-water with sugar & colour), sweetened squeezed fruit juice, mandazi, milk and milk products (whole cow’s milk and cultured milk). We asked children about their usual intake of different foods in a typical week. We generated scores from the FFQ responses to obtain servings per day as follows: never = 0 servings/ day, once per day = 1 serving per day, once or twice per week = 0.21 servings per day, 3 - 6 times per week = 0.64 servings per day, 2 - 3 times a day/ every day more than once = 2.5 servings per day . Trained interviewers guided children step by step on filling the FFQ to ensure they understood each food item and contents and answered them independently.
Defining food groups and subgroups
We used the apriori knowledge of traditional local foods and snacks reported by children from the 24-hour dietary recall to estimate the contribution of different foods to total fat, carbohydrates, and protein in the diet. Further, to create food categories, we aggregated the 15 food items from the FFQ into nine foods/ food groups based on their nutrient composition and culinary uses (percentage contribution in the children’s diet). For example, we grouped intake of chocolates, cakes, sweets, biscuits and doughnuts as sweets and sugars. The food groups/ subgroups were used for identifying dietary patterns.
Our study outcomes were dietary patterns, five adiposity measures: BMI z–scores, waist circumference, body fat percentage by bioelectrical impedance, triceps, subscapular skinfolds and associations between dietary patterns and measures of adiposity.
We performed all anthropometric measurements in duplicate following NHANES standard procedures , using calibrated equipment. Before taking any measures, we asked children to remove shoes, socks, hair ornaments, items from pockets, jewellery, and clothing other than their regular school uniform. Two trained research assistants took independent anthropometry and adiposity measurements from each child, and the average between two measurements was calculated. Routinely, we checked for standard operating procedures and assessed the technical error of measurement (TEM)  to estimate the inter-observer measurements errors.
Height was measured to the nearest 0.1 cm using the TANITA height rod with children standing straight with the head in the Frankfurt plane. Weight and total body fat percentage were measured using BIA (TANITA model DC 430 MA). The BIA machine was set to deduct 0.5kg (clothing weight) before measurements.
Waist circumference was measured to the nearest 0.1cm using a non-elastic circumference tape. Children were asked to stand straight with the abdomen relaxed, arms at the sides, feet together and the measurement was taken from the narrowest girth. For children in whom we couldn’t identify the measurement point, e.g., obese children, we asked them to bend to the side, and the measure was taken from the point where the trunk folds.
Triceps and subscapular skinfolds were measured by Harpenden precision thickness calliper to the nearest 0.1mm, on the right arm . Triceps skinfold was measured at the midpoint of the upper right arm’s back. Subscapular skinfold was measured below and laterally to the shoulder blade’s angle, with the shoulder and arm relaxed.
Analysis was performed by STATA version 15.1 (StataCorp, College Station, TX, USA). We used probability plots to assess the distribution of variables. Continuous variables were reported as means and standard deviations or as medians and interquartile range due to non-normal distribution. For categorical variables, frequencies and proportions were reported. WHO Anthroplus (STATA SE) was used to determine the BMI z-scores based on age and sex, that is thinness <-2 SD, normal weight between -2SD and 1SD, overweight >+1 and <+2SD, obese ≥ +2SD. Schools were considered clusters, as data were collected from different school settings.
Based on the scores generated from the FFQ, we performed factor analysis, a data driven method to derive the dietary patterns. The factor was loaded and rotated using varimax rotation to simplify the interpretation of the factors. The correlation between foods / food groups was examined, using a Kaiser Meyer Olkin (KMO) test = 0.83, as indicator that the correlation among the variables was sufficiently strong for factor analysis. Variables (food items/ groups) with a factor loading of absolute ≥0.3 , were retained and used for labelling dietary patterns. When food items loaded higher than 0.3 in more than one factor, the factor with the highest loading was considered for factor labelling. We categorised children into terciles of low, medium, or high adherence to each dietary pattern. Further, we performed a chi square test to compare the BMI z-score categories (normal, thinness and overweight/ obesity) with dietary patterns terciles.
We selected potential variables for adjustment based on a recent study on overweight/obesity conducted in Tanzanian primary schools  as these variables might be associated with the dietary patterns. In model 1, we adjusted for age and sex. In model 2, we adjusted for lifestyle factors: school type (private or government), time spent walking to school, district (rural or urban), availability of television and electronic gadgets at home, and neighborhood playground.
Next, we examined the association between the identified dietary pattern terciles and different adiposity measures using multilevel linear regression. The dependent variables were logarithmically transformed before linear regression analysis because of the non-normal distributions. All analyses were two-tailed, and the significance level was set at 5%.