Background: With millions of people experiencing malnutrition and inadequate water access, FI and WI remain topics of vital importance to global health. Existing unidimensional FI and WI metrics do not all capture similar multidimensional aspects, thus restricting our ability to assess and address food-and water-related issues. Methods: Using the Sanitation, Hygiene and Infant Nutrition Efficacy (SHINE) trial data, our study conceptualizes household FI (N = 3551) and WI (N = 3311) separately in a way that captures their key dimensions. We developed measures of FI and WI for rural Zimbabwean households based on multiple correspondence analysis (MCA) for categorical data. Results: Three FI dimensions were retained: ‘poor food access’, ‘household shocks’ and ‘low food quality and availability’, as were three WI dimensions: ‘poor water access’, ‘poor water quality’, and ‘low water reliability’. Internal validity of the multidimensional models was assessed using confirmatory factor analysis (CFA) with test samples at baseline and 18 months. The dimension scores were associated with a group of exogenous variables (SES, HIV-status, season, depression, perceived health, food aid, water collection), additionally indicating predictive, convergent and discriminant validities. Conclusions: FI and WI dimensions are sufficiently distinct to be characterized via separate indicators. These indicators are critical for identifying specific problematic insecurity aspects and for finding new targets to improve health and nutrition interventions.
Data for the development of the household measures of FI and WI were obtained from the Sanitation, Hygiene and Infant Nutrition Efficacy (SHINE) trial. The trial’s primary objectives were to test the independent and combined effects of an improved water, sanitation and hygiene (WASH) intervention, and an improved infant and young child complementary feeding (IYCF) intervention on stunting and anemia among rural Zimbabwean children. The design, protocol, and primary outcomes have been published elsewhere [50,51,52]. Briefly, SHINE was a four-arm cluster-randomized community-based 2 × 2 factorial trial conducted in two rural districts in Zimbabwe: Shurugwi and Chirumanzu. The two districts were divided into 212 clusters which were then randomly allocated to one of the four trial arms: (1) Standard of Care (SOC), (2) SOC + IYCF, (3) SOC + WASH and (4) IYCF + WASH. Recruitment occurred between 22 November 2012 and 27 March 2015. Village health workers (VHWs) employed by the Zimbabwe Ministry of Health and Child Care prospectively identified and referred eligible women for the trial. Only women residing permanently in a cluster and who were pregnant at the time of recruitment were enrolled. Written informed consent, in the language of their choice (English, Ndebele, or Shona), was obtained prior to data collection. SHINE was approved by the Medical Research Council of Zimbabwe and the Johns Hopkins University Bloomberg School of Public Health Institutional Review Board. SHINE included an extensive structured questionnaire to collect detailed information on household, maternal and child characteristics. Baseline data collection spanned the recruitment period mentioned above. A few weeks after obtaining consent, research nurses made home visits for face-to-face interviews with the women. Additional home visits for subsequent data collection were also made at one, three, six, 12- and 18-months post-partum, until the end of the study in July 2017. The questionnaire and data collection protocol are available on OSF at https://osf.io/w93hy/ (accessed on 17 May 2021). From the 5280 pregnant women who were recruited, 4675 took part in the baseline interview. For the following analysis, the sample was restricted to households with complete information on the selected food (N = 3551) and water (N = 3311) variables. Figure 1 illustrates participant inclusion. Sample selection for FI and WI factor analyses. The creation of FI and WI measures were carried out in a stepwise manner, starting with item variable selection for inclusion in the quantitative analyses. The next steps included descriptive analyses, item reduction, multiple correspondence analysis (MCA) with extraction and rotation of dimensions, and validity assessments. The starting point for item selection was the internationally accepted definitions and dimensions of FI [3] and WI [7]. The FAO [3] and Action Contre la Faim (ACF) [53] provide some recommendations for indicators of FI dimensions, while WaterAid [7], Global Water Partnership (GWP) [54] and JMP [20] suggest items for WI dimensions. Indicators relevant to rural Zimbabwe and available from SHINE were then selected. Table 1 provides detailed descriptions of all variables selected to represent each FI and WI dimension. Brief justifications are also provided below for the choice of item variables: Complete set of item variables from the Sanitation Hygiene and Infant Nutrition Efficacy (SHINE) trial considered for each dimension of household food insecurity and water insecurity, collected at baseline from November 2012 to March 2015. * All item variables were either dichotomous or ordered categorical, and reverse coded so that insecurity scored higher; ** Parameterization of variables as used in the subsequent quantitative analyses in this study; a Variables excluded in the subsequent steps of factor analysis if categories were too small (≤5%) or too common (≥95%). A.1. Food availability refers to the food supply aspect of food security [3]. This dimension considers whether food is actually present for the population [55]. At the national-level, this has historically been addressed via the use of food balance sheets of food production and imports. At the rural household-level, food availability may be captured by considering food stocks, presence of markets and ability to produce food. We used three variables to operationalize this dimension: (1) number of days of staple food stocks available for household members to eat according to their needs, (2) availability of a garden where the household grows fruits and vegetables, and (3) the availability of left-over food from the last cooking occasion. A.2. Food access concerns economic, physical and social resources that enable acquisition of sufficient, nutritious and preferred foods in a dignified manner [3]. Physical food access is linked to infrastructure and at the household-level can be captured by considering time spent, distance travelled and transportation to safe food sources. Economic access depends on the ability of households to purchase or barter resources to obtain food [55]. Social access concerns food preferences in terms of taste, health requirements and religious restrictions. It also implies that food is obtained in socially acceptable ways. The following seven household-level variables were considered for this dimension: (1) access to preferred food, (2) food sufficiency for all household members, (3) help required from family and/or friends to obtain food, (4) purchasing or borrowing food on credit, (5) selling assets for food, (6) time from home to food market, and (7) method of transportation to food market. A.3. Food utilization reflects differences in the intra-household allocation of food, nutritional quality of food, and food safety in terms of preparation, handling, and storage [8,55]. Within SHINE, four variables were available as proxies for food utilization: (1) household dietary diversity, (2) handwashing behavior prior to handling food, (3) whether food containers were covered, and (4) food storage location. No information was available as proxy for intra-household allocation, which also depends on age, work load, and other factors. A.4. Food stability covers the barriers and promotors of food security dimensions [8,55]. At the household-level, this can be captured by considering exposures to risks, shocks or vulnerabilities that influence the ability of household to consistently acquire food [55]. The variables most appropriate to represent this dimension from SHINE were household experiences of social, economic, agriculture and health shocks. B.1. Water availability depends on the physical presence of water resources or infrastructure that makes it available in sufficient quantity to households [56]. Sufficient quantities of water must be available for drinking to prevent dehydration (≥5 L per person/day) and for cooking, bathing, hygiene and sanitation (>100 L per person per day) [19]. Within SHINE, two variables were considered: (1) volume of water, calculated from storage capacity of water containers and water collection frequencies, and (2) whether the households had access to water for irrigation purposes. B.2. Water access refers to physical delivery and economic access to water. Methods for assessing water access include the distance to water points, fetching time, and water expenditures [19,57]. Water access is inadequate if households have to travel >1 km or >30 min (return journey) to collect water [19,58]. Water is affordable if households spend <3–5% of their total income on it [59]. Five variables were considered to assess water access: (1) whether the household purchases water, (2) drinking water collection time, (3) distance to drinking water point, (4) non-drinking water fetching time, and (5) distance to non-drinking water point. B.3. Water utilization is meant to reflect the quality and safety of water for drinking and other purposes. Physical quality can be measured by considering the color, smell and taste of the water. Chemical quality and microbiological safety are determined by testing turbidity, total dissolved solids, chlorine levels and the presence of bacterial coliforms in the water. In low-income settings, types of water sources are used as proxy for water quality and safety [19]. For instance, protected sources such as piped water, boreholes and wells are considered microbiologically and chemically safer compared to surface water from rivers or streams. To capture this dimension, three SHINE variables were used: (1) reported satisfaction with the water smell, color and taste, (2) water source for drinking, and (3) water source for non-drinking purposes. B.4. Water reliability refers to whether water supply is consistent or intermittent. Whether water is piped into dwellings or available off premises, it may be periodically or seasonally inaccessible [2]. To assess the reliability of water supply among SHINE households, two variables were considered: (1) whether drinking source and (2) non-drinking source ran dry over the past year. Separate multiple correspondence analysis (MCA) were conducted on the selected item variables to develop FI and WI measures. MCA for categorical variables is equivalent to exploratory factor analysis (EFA) or principal component analysis (PCA) designed for continuous variables [60]. Analyses were conducted as explained below using Stata Version 16 (StataCorp LLP, College Station, TX, USA) for descriptives, ‘FactoMineR’ [61] and ‘PCAmix’ [62] packages from the software R Version 4.0.2 for MCA and factor rotation, and MPlus Version 8.4 (Muthén & Muthén, Los Angeles, CA, USA) for validity tests. First, we looked at the distributions of participants across the categories of each item variable using frequencies and percentages. Variables with categories reporting frequencies of ≤5% or ≥95% were excluded. Second, we ran polychoric correlations on all variables. Items indicating negative correlations and those without adequate variance (<0.1) were dropped. We also used the Kaiser-Meyer-Olkin (KMO) measure for sampling adequacy and Barlett’s test of sphericity to ensure robustness of our approach. We then carried out MCA on the remaining variables. Scree plots were used to decide the number of dimensions for extraction. We investigated factor extraction using oblique (geomin) and orthogonal (varimax) rotations. Since correlations among the extracted factors were small (<0.5), we report varimax-rotated loadings in our results. Dimensions extracted were interpreted and named based on the variables that loaded on them from the theoretical framework (Table 1). We report the squared correlation ratios between each item variable and dimension, eigenvalues and percentage explained variances. Squared correlation ratios <0.20 were not considered relevant in explaining a dimension. We then used post-estimation commands in R to obtain standardized dimension scores for individual households. Validity refers to the extent to which certain measures are acceptable indicators for what they are intended to capture [63]. We tested four types of validity for our FI and WI measures: internal, predictive, convergent and discriminant. These are briefly described in Table 2 with an explanation of their purpose and statistical methods used. For internal validity, we assessed multidimensional model fit via confirmatory factor analysis (CFA) in two groups: (1) a sub-sample of the baseline participants constituting 60% of the dataset, and (2) the same baseline households more than 18 months after the baseline interview. We used model fit statistics such as root mean square error of approximation (RMSEA), standardized root mean square residual (SRMR), comparative fit index (CFI), and Tucker Lewis index (TLI). Satisfactory fit was determined using recommended arbitrary cut-offs of RMSEA ≤ 0.05, SRMR ≤ 0.08, CFA ≥ 0.95 and TLI ≥ 0.95 [64,65]. CFA was performed in MPlus using geomin rotation with diagonally weighted least squares estimator (WLSMV). Validity assessments for dimension scores of food insecurity and water insecurity. For predictive, discriminant and convergent validity, we used a group of exogenous variables, also obtained from the SHINE trial. Self-reported perceived health status of women was measured using an adapted version of the RAND Health Survey [66]. Scores for perceived health status ranged from 0 to 5 units, with 0 indicating least healthy and 5 most healthy [67]. The Zimbabwe-validated version of the 10-question Edinburgh Postnatal Depression Scale (EPDS) was used to assess depression among the women [68]; those with a score ≥12 out of 30 were classified as clinically depressed. Household receiving food aid over the past 12 months from government or other organizations (yes/no) was self-reported by women. Usual frequency of water collection was reported as daily, weekly or monthly. HIV-status of the participating women was determined via rapid blood tests performed by trained nurses [50]. Household socio-economic status (SES) was based on a household wealth index [69]. Seasonality was determined based on the date of interview; hungry season was from January through March and rainy season was from November through March. These variables were used as predictors in simple regressions to estimate associations with FI and WI dimension scores from MCA. We tested the robustness of the MCA results after accounting for missingness in the selected items. Almost all variables had <10% missing values (Table S1). Lower SES, HIV-status and interview months were found to influence missingness (Table S2). To account for missing data uncertainty, we imputed missing variables using the multiple imputation by chained equations (MICE) method via the ‘MICE’ function from the ‘missMDA’ package in R [70]. We then re-ran MCA by including the additional households with imputed variables. Only households with less than three imputed variables were used for sensitivity analysis. The sample size increased considerably (N = 4622 for FI and N = 4575 for WI).