2.1. Study setting and sampling

Around 70% of Maputo's inhabitants live in informal neighbourhoods covering approximately 35% of the city's area (UN‐HABITAT, 2010). These neighbourhoods include densely packed compounds, characterised by substandard housing, limited public health infrastructure, and high levels of poverty (Brown et al., 2015) and child mortality (Nhampossa et al., 2013).

Our sampling frame was drawn from the MapSan trial participant list, stratified by study arm (control/intervention). We enrolled equal numbers of eligible participants randomly from each arm. To be eligible for enrolment, participants needed to be resident in a MapSan trial compound and a caregiver of a child aged 6–60 months, with caregivers defined as the adult primarily responsible for feeding the child. Sixty‐four caregivers were enrolled, of which 58 completed a structured questionnaire and provided a food sample (29 in each trial arm) and 59 completed structured observations of child food preparation.

All data were collected between June 27 to July 24, 2018. Questionnaires and observations were recorded on tablets with forms developed using Open Data Kit. Questionnaires were written in English, translated to Portuguese and piloted in nonstudy households. All data were encrypted and kept on a secure server.

2.2. Caregiver questionnaire

The caregiver questionnaire addressed socio‐demographic characteristics of the household, asset ownership, compound characteristics, food preparation and feeding practices, perceptions of diarrhoeal disease risk, and whether the child had had diarrhoea in the preceding week. Potential risk factors included were predetermined prior to observations, based on food hygiene literature, and are listed in Appendix B.

To estimate relative wealth in our study population, a relative wealth index was created using principal component analysis on information on household assets and housing characteristics commonly used in the Demographic and Health Survey (MISAU/INE/INCFI, 2011). Nineteen variables were included to construct the index (Appendix A): household size, crowding, asset ownership (eight binary variables), and housing characteristics (nine variables) including water sources, building materials, and sanitation type based on MapSan trial classification, due to concordance between asset indices that include and exclude sanitation type (Rheingans, Anderson, Luyendijk, & Cumming, 2014). The first principal component was used to group households into relative wealth quintiles.

2.3. Food sample collection

Enumerators visited caregivers to enrol them in the study and determine when they were likely to first feed the child the following day. Enumerators returned at the designated time and asked caregivers to provide a sample of the child's meal. If no food was available, enumerators waited for it to be prepared until noon that day, or returned the next day. Enumerators administered the caregiver questionnaire alongside food sample collection. If food preparation could be observed at this visit, structured observations were conducted. Otherwise, enumerators returned to the household at a later date to observe food preparation.

Samples were collected according to standardised and pretested protocols. The following information was recorded for each sample: time of preparation, reheating, contact with other foods, and whether it had been stored in a covered container and/or refrigerated. Where the meal had multiple components, samples of all components were homogenised. Estimated 20 g of samples were placed in sterile plastic 100‐ml Whirl‐Pak bags (NASCO, Fort Atkinson, WI), up to the first line, using the same utensil used to feed the child, and temperature was taken on site. All samples were cooled, transported to the laboratory, and refrigerated at 4°C until processing less than 6 hrs after collection.

2.4. Structured observations

Trained enumerators conducted observations of caregivers to note ingredients and food preparation, feeding, and storage practices, using prepared guides outlining the following activities: Enumerators first asked caregivers to list the ingredients of the food that they were going to prepare and the origin and storage location of each ingredient and to briefly enumerate the steps of the preparation process; enumerators then encouraged caregivers to prepare food in their usual manner and recorded each step taken and its duration. Where possible, enumerators observed handwashing with soap (HWWS) at two moments, before food preparation and before feeding. When storage after cooking could not be observed, enumerators recorded whether it was the caregiver's intention to store the food. Observations lasted between 5 and 125 min.

2.5. Microbiological analysis of food samples

Enterococci were enumerated in food samples using standard membrane filtration procedures (United States Environmental Protection Agency, 2002) on Enterococcus indoxyl‐β‐d‐glucoside (mEI) selective medium, applied previously to measure faecal contamination (Pickering et al., 2012; Pickering, Julian, Mamuya, Boehm, & Davis, 2011). A 5‐g aliquot of each food sample taken was diluted in 50 ml of sterile water, homogenised by shaking at least 25 times and allowing to settle. The 5, 0.5, and 0.05 ml volumes of the supernatant were filtered through membranes in 50 ml of sterile water, transferred onto mEI plates, and incubated at 41°C for 20–24 hrs. Method blanks using sterilised water, positive controls using contaminated water, and replicates of one food sample were analysed each day. Colonies greater than 0.5 mm in diameter with a blue halo were identified as Enterococcus spp. and counted in colony forming units (CFU) per gram wet weight (CFU/g) of sample. Plates containing up to 200 colonies for the lowest dilution were counted for analysis. Counts were based upon the lowest dilution of samples except when plates for lower dilutions exceeded 200 colonies, in which case plates for the next highest dilution were counted and multiplied by the dilution factor to standardise the denominator across all samples.

2.6. Statistical analysis

All statistical analyses were conducted in Stata version 15.1 (StataCorp, College Station, TX, USA).

Arithmetic means were used to report colony counts (Haas, 1996). In the absence of food industry guideline values for enterococci, we interpreted results similarly to a previous food contamination assessment (Gibson, Sahanggamu, Fatmaningrum, Curtis, & White, 2017) and coliform guidelines used in the milk industry (U.S. Department of Health and Human Services, Public Health Service, & Food and Drug Administration, 2009), defining an outcome of >10 CFU/g as “contaminated.” We calculated the prevalence of Enterococcus spp. contamination (>10 CFU/g) in child foods alongside 95% confidence intervals (CIs).

Exposures of interest were caregiver demographics, including wealth and education; factors related to environmental contamination, including water source and treatment and sanitation type; HWWS; and type and preparation of food. The ratio of household members to rooms was used as a proxy measure for crowding. All variables were converted to categorical variables based on appropriate threshold values, and food types were grouped based on common ingredients of samples. These exposures were then cross‐tabulated with the binary contamination outcome (>10 CFU/g) and tested for crude associations using Fisher's exact test.

For regression analyses, counts lower than 10 CFU/g were coded as below safe limits. Categorical variables were ordinal, apart from food type, for which each food category was treated as an individual dummy variable. Overdispersion and excessive zeros (27/58) indicated appropriate use of a zero‐inflated negative binomial regression model. Due to the relatively small sample size for food samples (n = 58), nonparametric bootstrapping as described by Efron (1979) was applied to subsequent risk factor analyses.

Variables with a p value below .2 in Fisher's exact test, as well as those variables that perfectly predicted the outcome, were considered for inclusion in the logistic portion of the model. Variables were tested for associations with bacterial count using bivariate negative binomial regression, and variables with a p value below .2 were considered for inclusion in the negative binomial portion of the multivariable model. For each iteration of the model, each variable was added individually to the logistic portion. Akaike information criterion (AIC) was used to compare the base model with each new model, with the variable retained where this statistic was minimised. This process was repeated until no variables improved model fit. Variables were then added to the negative binomial portion in the same manner. Variables excluded at the screening stage were checked to determine if their inclusion in either portion of the model improved the fit; none met this criterion. Risk factors included in the final multivariable model were checked for multicollinearity, and variables exceeding a variance inflation factor of 10 were removed. The final zero‐inflated negative binomial model was compared with negative binomial, Poisson, and zero‐inflated Poisson regression models using AIC and likelihood ratio tests. There was very strong evidence supporting the preference of the zero‐inflated negative binomial model (AIC: 505) over Poisson (AIC: 44,400) and zero‐inflated Poisson (AIC: 42,700) models, and weak evidence preferring it over a negative binomial model (AIC: 523), on the basis of these model fit statistics.

2.7. HACCP analysis methodology

HACCP analyses typically include collection of detailed information on the food product (composition, pH, A_W, etc.), full assessment of physical, chemical and biological hazards at each stage in the operation, and specification of critical limits (e.g., time and temperature criteria) and monitoring procedures for each CCP. A range of microbiological organisms of concern and associated toxins would also be considered (FAO & WHO Codex Alimentarius Commission, 1997). Logistical constraints for this small, exploratory study led to modification of the standard HACCP to assess microbiological hazards using an indicator for the likely presence of faecal pathogens, restricted to prepared foods at the point of feeding as the critical time for the child's safety, and incorporated quantitative risk factor analysis to deduce the relative importance of intermediate steps.

We analysed structured observation data and created food preparation flow diagrams for each food type observed, with foods then grouped into categories based on ingredients and preparation processes, as well as a general food‐flow diagram for all foods. We identified hazards by prevalence and magnitude of contamination by Enterococcus spp. in child foods. We considered the steps observed in the food preparation process and used risk factors identified from multivariable analysis to identify a range of points in the process where a step might increase, or a control measure might reduce, the prevalence or magnitude of contamination. We also compared frequencies of these significant risk factors in observations and used temperature measurements at time of feeding to assess the extent of reheating. We then identified CCPs out of the range of control points based on whether (a) preventive control measures can be applied, (b) loss of control could increase contamination to unacceptable levels, and (c) no subsequent steps exist to control the hazard (NSF (National Sanitation Foundation), 2006). We annotated the food‐flow diagrams with identified CCPs and proposed feasible control measures to reduce risk of faecal contamination of child foods in this setting.

3.2. Characteristics of study participants and exposures

A full list of potential risk factors cross‐tabulated with the binary contamination outcome are presented in Appendix B. Fourteen exposures had a p value below .2 or perfectly predicted food contamination above safe limits (Table 1) and were considered for inclusion in the logistic portion of the multivariable regression model. The same risk factors were tested for associations with the bacterial count using bivariate negative binomial regression modelling (Appendix C). The seven variables with a p value below .2 (Table 1) were considered for inclusion in the negative binomial portion of the multivariable model.

Caregivers reporting responsibility for feeding the child were predominantly female (98%) and often the child's mother (74%). The mean caregiver age was 34 years (range 17–79 years), and most (86%) had at least primary education. The mean child age was 31 months (range 7–56 months). Ten percent of children in the sample had diarrhoea in the week preceding (95% CI [4, 21]). Ninety‐five percent of children in the study had been breastfed at some point, but only 5% were still breastfed at the time of the study. The median number of household members was six (range 2–13), and houses were crowded—52% of households had more than four members per room. There was weak evidence that households with more than one child under 5 years had more contaminated food samples (67% vs. 46%, p = .17). Prevalence of food contamination did not vary by relative wealth quintile (p = .46).

Soap was present in 91% of households, and 60% reported HWWS. All households without soap had contaminated food samples, and there was weak evidence that caregivers reporting HWWS had fewer contaminated food samples (46% vs. 65%, p = .18). Key moments caregivers reported HWWS were after defecation (47%), before eating (52%), and before preparing food (38%). The rarest mentioned moment, after changing nappies, perfectly predicted no contamination.

Respondents took water from a tap in the compound (57%) and stored it in a container with a closed lid (84%). Ten caregivers (17%) reported treating water, seven of these by boiling. There was no difference in proportion of samples contaminated by water source, storage or treatment, or between intervention and control arms of the MapSan trial (55% vs. 52%, p = 1.00). Children's faeces were observed in two compounds (3%) and perfectly predicted contamination. Animals were observed in 39 compounds (67%).

Most caregivers (74%) prepared food inside their home. Fifty‐eight percent of cooked food samples, taken at time of feeding, were taken immediately after cooking, and others taken after the food had been stored, and later reheated (17%) or given cold (25%). Porridge and fried egg were served immediately and never stored. Salad was the only food not cooked. Prevalence of contamination varied by storage after cooking (64% vs. 37%, p = .05), but not by subsequent reheating (67% vs. 62% served cold, p = .81). Mean temperature of food samples at time of feeding was 37°C in reheated foods and 45°C for those recently cooked, with weak evidence of a temperature difference (p = .08).

Because samples were taken in the morning, foods not recently cooked had been stored for over 11 hrs overnight (mean: 20 hrs; 95% CI [18, 23]). 73% of stored foods had been kept in a covered container. Although 26% of study households used a refrigerator/freezer to store raw ingredients, only one food sample had been refrigerated. Of food stored after cooking, covered storage and refrigeration perfectly predicted contamination.

3.3. Multivariable regression

The multivariable regression model comprised nine risk factors for child food contamination (Table 2), in two distinct sets: five associated with the odds of food samples being below safe limits (<10 CFU/g) or nonzero; and four associated with increased or decreased bacterial count in nonzero samples. For the logistic portion of the model, while the small sample size (n = 58) with several perfect predictors for the binary contamination outcome produced some extreme estimates, there was strong evidence that soap availability and HWWS after changing nappies were associated with increased log odds of excess zeros in samples (20 and 21 log [odds] increase, respectively; p < .01). Salad and food stored after cooking were associated with reduced odds of zero counts (21 and 22 log [odds] reduction, respectively; p < .01). For nonzeros, there was evidence of lower counts for porridge (3.2 log [count] reduction, p = .10) and for stored foods being refrigerated (25 log [count] reduction, p < .01). The two sets approximate to those that may affect pathogen presence or survival—including hygiene behaviours and cooking food to boiling temperatures—and those that affect bacterial proliferation—including safe storage and type of food that facilitated growth.

The authors would like to thank Centro de Investigação e Treino em Saúde da Polana Caniço (CISPOC) for the use of their laboratory space and, in particular, the assistance of Victoria Cumbane. We also would like to thank Zaida Adriano and acknowledge all the members of the WE Consult enumerator team for their support. Lastly, we are particularly grateful to the study participants who generously contributed their time to the study and, without whom, it would not have been possible.

This work was partly funded by the Bill & Melinda Gates Foundation.