Study Design
The WaSH in Metro Manila Schools study was a cluster-randomized controlled trial (cRCT) conducted in the Philippines’ National Capital Region. We used a parallel group cRCT design with unequal allocation (ratio 1:8.7) of schools to control and intervention groups (CG and IGs) to enable us to implement the intervention efficiently. We hypothesized that the intervention would improve children’s HL, nutrition, hydration, and HW. We adjusted the trial design to measure group-level differences in selected outcomes by including three clusters (cities), each with different numbers of schools and children who were assessed twice, at baseline and endline. No cluster corrections were used for schools, as they were the units for different treatments. Each assessment cycle lasted about one month and was balanced between the CG and IGs to reduce confounding due to seasonal factors.
For the CG, we delivered the “standard of care” consisting of a WaSH policy workshop for teachers and two HE sessions for children. For the IGs, we randomly assigned schools to one of three arms based on the intensity (low, medium, or high) of HE. We also provided policy workshops for teachers, hygiene supplies, and WaSH facilities repairs. Instead of using a double-size CG to increase power, we made the IGs larger than the CG to increase the precision of the intervention comparison. We previously reported our study design and rationale [
16].
The study protocol was approved by the Ethics Committees of the University of Bonn, Germany (Number 216/16) (September 28, 2016), and the University of the Philippines, Manila (Number 2017–0113) (February 23, 2017).
Participants, study sites, and sample size
In Metro Manila, public schools have inadequate WaSH, and WaSH-related diseases are endemic. During a previous cross-sectional study conducted in March - May 2017, we measured diarrhoea and helminth infection rates at 14% and 29.7%, respectively [
17]. We selected schools in Manila, Navotas, and Quezon City, cities that are geographically and demographically representative of Metro Manila. Schools were identified based on a complete census of schools managed by the Philippines Department of Education (DepEd). We recruited 15 public schools that previously participated in our cross-sectional study because of existing trust and cooperation with school principals and personnel. These factors facilitated communication and collaboration, which were important to us, as we wanted a long-term working relationship with schools despite limited time and resources. We offered no monetary reimbursement to participants. Instead we gave participants compensation packages comprised of school and hygiene supplies [
16].
The outcomes of HAZ and stunting prevalence are the basis of sample size estimation. It assumes a difference of 0.15 HAZ between the IG and CG, not adjusting for repeated measures within clusters, as well as a relative risk (RR) of stunting of 0.7 or smaller, with 10% prevalence in the CG. We assumed a type I error (α) of 0.05, power (1 − β) of 0.8, a one-sided test for a two-sample comparison of means, and a 10% dropout after baseline. Due to limited resources, the CG was not double sized, limiting our ability to account for multiple hypothesis tests. We previously described our sample size estimation [
16]. Briefly, to estimate the sample size, we considered the target population to be all the public school children in Metro Manila, where in School Year 2014–2015 a total of 2,059,447 public school children were enrolled [
17]. We inflated the sample by 30% and 45% to account for nonresponse and refusal, respectively, and then inflated the sample by another 5% to account for differences in schools’ enrolment sizes. The target sample size was N = 760; we enrolled 756 students at baseline and surveyed 701 students at endline eight months later (retention rate: 93%).
We previously described our multi-stage cluster sampling strategy [
16]. First we recruited schools using a list of 15 schools that participated in our cross-sectional study [
18]. Schools were eligible if they were public (i.e. managed by the government), had WaSH facilities available for inspection, and had no other on-going WaSH projects. Second we recruited class sections. At each school, we selected one or two class sections to obtain a target sample of ~ 50 students per school. To avoid interrupting classroom instruction, we recruited entire class sections as a whole rather than groups of students from multiple class sections. We did not re-recruit the students who participated in our previous cross-sectional study because a new school year had begun and caused some students to move to different class sections or schools. Third we recruited children. Participants were eligible for our study if they were in grades five, six, seven, or ten; able to complete our questionnaire independently or with minimal assistance; and able to provide relevant health data. We chose these grade levels to ensure children were developmentally mature enough to use and have perceptions about school WaSH facilities and able to actively participate in our intervention activities.
Prior to on-boarding, school principals gave written informed consent “in loco parentis”, i.e., in the place of parents, for children’s participation. We explained to children our study’s purpose and procedures, and stated that participating in our study was voluntary and that all data would be anonymized, confidential, and would not affect their school grades.
Randomization and masking
We have previously described how we randomly assigned clusters to treatment using Microsoft© Excel’s random number function (simple randomization) [
16]. Briefly, in an Excel worksheet, the names of the 13 schools were listed in the first column, wherein one row represented one school. The four IGs (A - D) were listed in the second column, wherein one row represented one IG. (We previously determined how many schools would be allocated to each group.) Schools were ranked and then assigned to groups A, B, C, or D in the third column using Excel’s random number function. IGA was known as the low-intensity health education (LIHE) group, IGB and IGD were known as the medium-intensity health education (MIHE) group, and IGC was known as the high-intensity health education (HIHE) group. The research supervisor and one research assistant (who was not involved in data collection) performed randomization and assigned schools to IGs. The research supervisor and research team enrolled participants. The investigators were unmasked, while all school principals, teachers, personnel, children, and parents, were masked to treatment assignment. It was not possible for participants to know the treatment assignment of nearby schools because any intervention materials distributed to schools did not uniquely identify treatment status.
We completed baseline surveys and then purposively assigned two schools to the CG. One school had a principal who directly asked to participate in our previous cross-sectional study, while the other school was integrated (i.e., it offered kindergarten through grade 12) and was the location of the pilot testing of our survey instruments.
Procedures
We designed interventions to increase children’s understanding about WaSH and improve health-related behaviours, specifically HW. Our goal was to empower children to be proactive about reducing their exposure to pathogens in the environment, thereby preventing disease and promoting well-being. During formative research, we learned that there was insufficient knowledge about the health benefits of HW and adequate WaSH management. Thus, we developed an intervention strategy to increase knowledge by engaging directly with children and increasing their enthusiasm about HE and capacity for practicing healthy behaviours, as well as creating enabling environments through the provision of necessary equipment and supplies. Our educational materials were based on the existing DepEd curriculum and open educational resources from the U.S. Environmental Protection Agency and the U.S. National Library of Medicine. We report the content of HE sessions in Additional file 1. We used findings from our cross-sectional study and baseline survey, as well as inputs from research assistants (with expertise in the local context), school principals, teachers, and janitors, and we conducted opinion polls with children. To assess whether our intervention would be acceptable and sustainable, we used participatory research and proactively engaged with stakeholders. We confirmed that intervention materials were delivered to study participants at the start of the trial and we made unannounced visits to schools to periodically assess intervention adherence. We implemented the intervention between June 2017 and March 2018 (Additional file 2).
We previously described the four parts of the intervention [
16]. Briefly, we provided: WaSH policy workshops for teachers, HE for children, hygiene supplies, and WaSH facilities improvements (Additional file 3).
The research supervisor conducted an in-person eight-hour training workshop for research assistants before conducting baseline school surveys. We previously reported details about training methods, including the time, place, and duration of training, and teaching aids and technologies [
16]. Baseline school surveys were conducted according to protocol (Additional file 4). Eight months later, we conducted endline school surveys, using the same methodology and measuring the same outcomes assessed at baseline. We also obtained school administrative data from the DepEd and conducted two cross-sectional surveys: a demographic questionnaire for children to assess household-level risk factors and a water quality study to assess exposures to waterborne pathogens in schools and homes.
Further details about our research procedures, including contents of training workshops, intervention components, and adherence promotion strategies were previously reported [
16].
Outcomes
All trial outcomes were observable, measurable, pre-specified, and assessed at baseline and endline (Additional file 5: Table S1). Trial outcomes were: height-for-age Z score (HAZ), body mass index-for-age Z score (BAZ), body mass index (BMI), height, and weight; prevalence of stunting (HAZ < -2), undernutrition (a composite of thinness [− 3 < BAZ < − 2] and severe thinness [BAZ < − 3]), and overnutrition (a composite of overweight [1 < BAZ < 2] and obesity [BAZ > 2]); urine specific gravity (U
sg) and prevalence of any (U
sg ≥ 1.020), mild (U
sg = 1.020), moderate (U
sg = 1.025), and severe dehydration (U
sg = 1.030) [
19]; scores for overall HL, HL about germs, and HL about HW. We calculated HAZ and BAZ using the WHO AnthroPlus (for children 5–19 years old) software (version 3.2.2., WHO, Geneva, Switzerland). We classified nutrition status using the 2007 WHO Growth Reference [
20]. During initial trial registration, we erroneously omitted dehydration from our study protocol’s list of outcomes; the study protocol has since been updated. We estimated HL scores via a 20-item questionnaire developed and refined by our research team. We asked children about their knowledge about general hygiene, germs, and handwashing. Examples of questions include: “What are germs?” “True or false: If I have germs, then I can have vomiting or diarrhea.” “How long should I wash my hands with soap and water to get rid of germs?” We previously reported details about the health literacy tool and provided a sample questionnaire [
16].
We observed the adequacy of schools’ WaSH facilities, assessing availability, accessibility, cleanliness, and functionality, according to guidelines from the DepEd and the Philippines Department of Health (DOH) [
21,
22]. We report data on schools’ WaSH facilities in Additional file 6: Table S2.
We pilot-tested and improved data collection tools before beginning this trial and ensured the safety of participants by adhering to research protocols. No contingency plan was deemed necessary for adverse events as our intervention involved no invasive procedures or provision of medications.
We will report additional (cross-sectional) outcomes (e.g., children’s self-reported health status, satisfaction with schools’ WaSH facilities) and associated risk factors in a forthcoming paper. We report the sample sizes of all surveys conducted during this trial in Additional file 7: Table S3.
Statistical analysis
The research supervisor conducted data analysis according to a pre-specified data analysis plan. We used intention-to-treat analysis, comparing each IG to the CG. We conducted descriptive analysis, pre- and post-intervention, measuring study participation, demographic characteristics, and outcomes of interest. We reported demographic characteristics and household risk factors, including food insecurity, according to study arm. For each outcome, we reported descriptive results (e.g., percentages, frequencies) for each arm, including the estimated effect size and precision. We measured frequencies and interquartile ranges (IQRs) relevant to homes’ demographic makeup. Data from school inspections were summarized at the school-level by measuring the mean scores of individual facility inspections. We measured prevalence rates using contingency tables with estimates of standard error (SE) and precision.
We assessed socioeconomic status (SES) by performing a factor analysis of variables that indicated the possession of household assets, e.g., computer, cell phone, refrigerator, car. The score of the first factor was then divided into three categories using the k-means procedure. Food security status was derived from a factor analysis of variables indicating access to a secure food source, e.g., enough food is available for all members of the household; eats a variety of food; rarely has asked/begged for food; rarely has gone to sleep feeling hungry. The score of the first factor was then divided into three categories using the k-means procedure.
We used multi-level mixed effects regression models to assess intervention effects. We used two-sided tests for primary outcomes to compare study arms. Paired t-tests were used for continuous variables to calculate the mean height and weight differences. For continuous outcomes, we used multilevel mixed-effects linear regression models to estimate intervention effects with measures of precision, i.e., 95% confidence intervals (CI), and p-values. We used regression models to analyse exposure-response: g = (E[Ai]) = β0 + β1Bi + γCi, where Ai is the primary outcome of interest, g is the appropriate link function (identity for height and weight, logistic for stunting and poor HL), Bi is the continuous exposure of interest, and Ci is the vector of confounders. In linear regression models, intervention effects can be interpreted as the adjusted differences in the mean changes of the desired follow-up outcome between the respective IG and the CG. The model included the respective IG, random intercept for the city, and robust standard errors. We adjusted for the child’s sex, age, and desired outcome at baseline, and the parent/caregiver’s education level and SES. In linear regression models that assessed school-level outcomes, we adjusted for other possible confounders, e.g. attendance in primary school, the school’s MOOE budget, handwashing basin-to-student ratio, and the availability of water in the school restroom.
For binary outcomes, we used multilevel mixed-effects logistic regression and Poisson regression models. In logistic regression models, intervention effects were expressed as the odds ratio (OR) of the prevalence at endline of the desired outcome between the respective IG and CG. The model included the respective IG, random intercept for the city, and robust standard errors. We adjusted for the child’s sex, age, and desired outcome at baseline, and the parent/caregiver’s education level and SES. In logistic regression models that assessed school-level outcomes, we adjusted for other possible confounders, e.g. child’s sex, attendance in primary school, the school’s maintenance and other operating expenses (MOOE) budget, handwashing basin-to-student ratio, availability of water in school restroom. In Poisson regression models, intervention effects can be interpreted as the incidence-rate ratio (IRR) of a desired follow-up outcome between the respective IG and the CG. The model included the respective IG, random intercept for the city, and robust standard errors. We adjusted all models for the children’s age and sex, parent’s education level, and family’s SES, as well as the outcome at baseline, when appropriate. We report details of our models in Additional file 8: Table S4.
To control the effect of confounding we used randomization (i.e. randomly assigning 13 schools to one of four intervention groups) and statistical methods (e.g. logistic regression models that were adjusted for possible confounding variables like age, sex, parental education, and SES.
During our assessment of missing data, if no statistically significant difference was found between children who were missing data and children who were not missing data for key outcomes, we concluded that data were missing at random (MAR), though not missing completely at random (MCAR). MAR means that, within groups defined by the observed data, all data had an equal chance of being missing and that the reason why data were missing is due to a known characteristic of the data themselves [
23]. This is a less realistic occurrence in field research. Possible reasons for MAR in our study included: nonresponse or loss to follow-up due to school absence or discontinued school enrolment. We used Stata, version 15 (StataCorp, College Station, Texas, US) for all statistical analyses.
We retroactively registered our trial in the German Clinical Trials Register (DRKS) (05/08/2020; number DRKS00021623). The trial protocol is available on the DRKS’s website:
https://www.drks.de/drks_web/.