A novel household water insecurity scale: Procedures and psychometric analysis among postpartum women in western Kenya

Our ability to measure household-level food insecurity has revealed its critical role in a range of physical, psychosocial, and health outcomes. Currently, there is no analogous, standardized instrument for quantifying household-level water insecurity, which prevents us from understanding both its prevalence and consequences. Therefore, our objectives were to develop and validate a household water insecurity scale appropriate for use in our cohort in western Kenya. We used a range of qualitative techniques to develop a preliminary set of 29 household water insecurity questions and administered those questions at 15 and 18 months postpartum, concurrent with a suite of other survey modules. These data were complemented by data on quantity of water used and stored, and microbiological quality. Inter-item and item-total correlations were performed to reduce scale items to 20. Exploratory factor and parallel analyses were used to determine the latent factor structure; a unidimensional scale was hypothesized and tested using confirmatory factor and bifactor analyses, along with multiple statistical fit indices. Reliability was assessed using Cronbach’s alpha and the coefficient of stability, which produced a coefficient alpha of 0.97 at 15 and 18 months postpartum and a coefficient of stability of 0.62. Predictive, convergent and discriminant validity of the final household water insecurity scale were supported based on relationships with food insecurity, perceived stress, per capita household water use, and time and money spent acquiring water. The resultant scale is a valid and reliable instrument. It can be used in this setting to test a range of hypotheses about the role of household water insecurity in numerous physical and psychosocial health outcomes, to identify the households most vulnerable to water insecurity, and to evaluate the effects of water-related interventions. To extend its applicability, we encourage efforts to develop a cross-culturally valid scale using robust qualitative and quantitative techniques. Data for scale development and validation were collected between June 2015 and August 2016 in Nyanza region, southwestern Kenya where the Luo, Kisi/Gusii, Kuria, and Luhya are the predominant ethnic groups. The major economic activities include fishing (on the nearby Lake Victoria), and mixed and agro-pastoral agriculture [42]. The region is typified by low crop yields and soil fertility, with a greater proportion of farmers engaged in subsistence farming [43]. Nyanza is one of the poorest regions in Kenya, with about 63% of the population living on less than $1 a day [44]. Research was conducted in the context of an observational cohort of 266 postpartum HIV-infected and HIV-uninfected women entitled “Pii en Ngima” [PEN], Luo for “water is life” (Clinicaltrials.gov# {“type”:”clinical-trial”,”attrs”:{“text”:”NCT02979418″,”term_id”:”NCT02979418″}}NCT02979418) who had been previously enrolled in a pregnancy cohort titled “Pith Moromo” (Clinicaltrials.gov # {“type”:”clinical-trial”,”attrs”:{“text”:”NCT02974972″,”term_id”:”NCT02974972″}}NCT02974972). HIV-infected women were over-sampled to achieve 1:1 ratio. The study took place in seven clinical catchment areas that span urban, peri-urban and rural sites across Nyanza region including Kisumu (urban), Migori (peri-urban), Nyahera (peri-urban), Rongo (peri-urban), Macalder (rural), Nyamaraga (rural), and Ongo (rural). Family AIDS Care and Education Services (FACES), an HIV care and treatment program in Nyanza region, supported each of the clinics in the medical sites. Nyanza region was an appropriate study site because of the high level of food and water scarcity [19], the high prevalence of HIV, which is currently 6.9% for pregnant women in western Kenya [45], and the presence of an excellent clinical care, research and laboratory infrastructure through FACES. Data collection for this study was structured in four phases (Table 1). The first phase, formative data collection, explored the experiences of water insecurity through “go-along interviews” (Activity A) [46,47], Photovoice photo-elicitation interviews (Activity B) [48,49], and the Delphi method (Activity C) [50], which was conducted concurrently with focus group discussions (FGDs; Activity D)) [35]. The second phase involved the assembly (Activity E) and revision of the household water insecurity (HHWI) scale questions using cognitive interviews with non-cohort Kenyan women (n = 10) who had similar characteristics as our target population (Activity F) [51–53]. The third phase entailed the administration of the survey to individual women (Activity G) and the final phase included collection of non-survey data for purposes of further scale validation (Activity H). Activities A, B, and D used non-cohort women (n = 40) with similar demographic characteristics as those used in the third phase for survey administration, Activities G and H (n = 241 and n = 186, respectively). Notes ahousehold water insecurity bS1 Table cS2 Table dmonths postpartum Although the results of our Phase 1 are presented elsewhere to avoid an excessively long manuscript [35], we briefly describe the formative methods used in order to convey the basis of the initial scale questions and to place our scale development activities within a broader context. Go-along interviews are a hybrid between participatory research and qualitative in-depth interviewing, that attempt to contextualize meaning within social and spatial contexts [46,47,54]. In this study, a Kenyan anthropologist (PM) accompanied twenty participants to and from water collection sites while asking questions. The interviews were translated (from Swahili or Luo to English), transcribed, coded and analyzed using Dedoose software (Los Angeles, CA: SocioCultural Research Consultants, LLC). Photovoice applies documentary photography and critical dialogue to explore the lived experiences of people and as a means of sharing knowledge [48,49]. In this study, twenty women were lent digital cameras to take photos of their experiences of household water acquisition, use and insecurity. On a second visit, these photos were used to conduct in-depth individual interviews. A subset of these photos became the core focus for dialogues about HHWI during FGDs at a third encounter. Dedoose was also used to code translated transcripts from in-depth interviews and FGDs. The Delphi method is a technique “for structuring a group communication process so that the process is effective in allowing a group of individuals, as a whole, to deal with a complex problem” [50]. Here, it was used to obtain feedback from international experts including those with expertise in hydrology and geographic research, WASH and water related programs, policy implementation, food insecurity and scale development, over the course of three rounds of surveys (S1 Fig). Each round was interspersed with FGDs in which questionnaires progressively became more closed ended. Questions included the definition of water insecurity, household water-related activities, barriers to water acquisition, consequences of water insecurity, and possible survey items that could constitute a HHWI scale (S1 Table). FGDs were conducted iteratively with the Delphi process (S1 Fig). To participate in FGDs, nurses and healthcare professionals purposively recruited postpartum women who were available and were either pregnant or had children less than 2 years of age in 4 study areas. After Delphi round 1, FGD participants (Kisumu; n = 8 and Rongo; n = 7) were asked to provide feedback on topics discussed in the online survey to build consensus around the definition of and questions related to HHWI. Another group of FGD participants (Migori; n = 5 and Macalder; n = 7) also provided information with which to revise questions for the survey. Based on themes that emerged from activities A-D [35], and the burgeoning literature on measurement of household water insecurity [2–4,37,36], we created 29 questions related to HHWI. Thirteen (13) of the questions (marked with asterisk in S2 Table) on the psychological, social, economic, and health consequences of water insecurity were adapted and modified from previous water insecurity scales [2,3,36,40]; the other 16 questions originated from activities A-D. The 29 questions were then ordered in what we considered to be least to most severe manifestations of water insecurity. They were phrased as Likert-type items, with the following response options: never (0), rarely (1–2 times), sometimes (3–10 times), and often (more than 10 times) in the last four weeks. Once the questions were developed, we conducted cognitive interviews (n = 10). Cognitive interviewing was used to assess: whether participants perceived the intent of the water insecurity questions as intended, whether participants were able to repeat questions they had been asked and the thought processes behind their responses, and whether the response options were appropriate and/or adequate [51–53]. This interviewing approach resulted in only some minor rephrasing of the 29 items (S2 Table). The HHWI module was then administered as part of the PEN study at 15 and 18 months postpartum. Participants were also asked other survey questions pertaining to water and their physical and psychological health for purposes of scale validation. Water acquisition questions were asked to help assess convergent validity. Participants were asked to indicate how long it took for them to travel to the water source, queue, fetch water and return to their houses and how much they spent on water. Additionally, we assessed access to safe water by using the WHO/UNICEF [11] survey questions for improved and unimproved drinking water sources. Food insecurity was assessed for purposes of predictive validity. We used the Individual Food Insecurity Access Scale (IFIAS) [51], which is a 9-item scale analogous to the Household Food Insecurity Access Scale (HFIAS) [55] but asks participants about their own individual experiences with access to food in the prior month. The intensity of food insecurity was assessed with follow-up questions asking whether this condition was experienced never, rarely, sometimes, or often (coded 0, 1, 2 or 3) with a range of 0–27. Maternal stress was also assessed to examine predictive validity. We used Cohen’s Perceived Stress Scale [56], These questions asked about the feelings and thoughts of women in the prior month, i.e. the frequency they felt upset, nervous or worried. The intensity of perceived stress was assessed by Likert-type response format of never, almost never, sometimes, fairly often and very often (coded 0, 1, 2, 3 or 4) with a range of 0–40. Data on water quality were collected at 5 randomly selected PEN participant households in each of the 7 catchment areas (n = 35). Drinking water quality was assessed by aseptically collecting triplicate 100ml water samples using Whirl-Pak Thio-Bags to test for total coliforms and Escherichia coli (E.coli) using Compartment Bag Tests (CBT) (Aquagenx) and Colilert (Idexx Laboratories) [57]. Samples collected were analyzed for total coliform and E.coli most probable number (MPN). Water quality was dichotomized according to WHO standards showing the presence of E.coli (≥1 MPN/100 ml) in household drinking water [58–60]. In the same 35 households, we measured the amounts of stored water for drinking and non-drinking purposes at a single time point in a given day. The volume was measured in liters based on the size of the storage containers and the amount of water in the container. For instance, a half-full 20-liter jerry can was measured as 10 liters of stored water. We also measured the amount of water used daily by the household in liters based on estimates of the amount of water used in cooking, drinking, washing foods, washing clothes, bathing, washing face, brushing teeth, washing hands, washing utensils/dishes, and washing toilets. By dividing the total amount of water used by number of individuals in the household, we were able to estimate per capita household daily water use. For purposes of analysis, complete data were available for 27 households. Of the eight households dropped, 3 had no stored drinking water and 5 had data available from a different time point. To assess intra-respondent reliability, we administered a subset of the 29 items (20-item version of the water insecurity module) daily for 30 days. We used 20 items to reduce respondents’ fatigue as it was being asked continuously for a month. Thirty-five participants were asked each day if they had that experience of water insecurity in the prior day and could respond yes or no. On the 31st day, participants were asked to indicate the number of days they had experienced that particular aspect of HHWI over the prior 30 days. Correlation coefficients were calculated between cumulative daily recall and responses from the 31st day. Quantitative data analyses were conducted in six phases including descriptive analyses, item reduction, extraction of factors, tests of dimensionality, scale reliability, and validity (Table 2). Software packages used included MPlus version 7.40 (Los Angeles, CA: Muthén & Muthén) [61], SPSS version 20.0 (Armonk, NY: IBM Corp.) and STATA version 14 (College Station, TX: StataCorp LP). Tests of dimensionality of the factor structure developed using the 15 months postpartum data were conducted using data from 18 months postpartum (n = 186); the rest of the analyses were done using data from 15 months postpartum (n = 241). First, we estimated proportions, means, and standard deviations of the HHWI module and participant characteristics. Although there were 5-response categories for the scale items originally, the sample distribution was skewed to the right (<5%) for “always” for each item. Therefore, “often” and “always” were collapsed for subsequent analyses. We first assessed adequate variance for all HHWI items [62]. This was followed by polychoric (inter-item) and polyserial (item-total) correlation of scale items [61,63,64]. Items without adequate variance, very low inter-item (<0.3) and item-total correlations (<0.3), very high residual variances (>0.50), and high missing cases (>10%) were dropped. We also estimated item communalities for degree of common variance between items [62], the Kaiser-Meyer-Olkin measure for sampling adequacy [65], and the Bartlett test of sphericity [66–68] to ensure our item reduction approach was robust. Furthermore, one item was dropped for any two items that suggested collinearity (≥0.98). Multiple approaches were used to determine the number of factors to retain. Exploratory factor analysis (EFA) was used together with Guttman’s [94] eigenvalue rule of lower bound, Kaiser’s [71] eigenvalue >1 rule, Cattell’s [72] scree test, and Horn’s [73] parallel analysis (PA) to determine the optimal number of factors that fit the data at 15 months postpartum [73,61,64,70]. For the scree tests, the root of the scree was used as a point of extraction for the true number of factors [72]. The extraction process in all the models used oblique rotation with a diagonally weighted least squares (Mplus WLSMV) estimator except for Horn’s PA, which employed the maximum likelihood (ML) estimator. For sensitivity analysis, we employed principal axis factors. A number of model fit statistics were used to determine meaningful model fitness for both traditional factor and parallel analyses (S3 Table). The fit indices included the chi-square test of model fit, the Tucker Lewis Index (TLI≥0.95), the Comparative Fit Index (CFI≥0.95), the Root Mean Square of Error of Approximation (0≤RMSEA≤0.10), and the Standardized Root Mean Square Residual (SRMR≤0.08) [74–81]. Consistent with the factor structure of previous household water insecurity scales elsewhere [2–4,36,40], we assumed our model would produce similar a factor structure for our scale. In order to test the factor structure obtained from the EFA, a test of scale dimensionality (i.e., the number of factors and factor loading pattern) was conducted using confirmatory factor analysis (CFA) and bifactor or nested factor modeling on an independent sample at 18 months postpartum. While EFA is a dimension generating method, the CFA allows for a formal test of dimensionality of the hypothesized factors developed by the EFA [61,69,82,83]. To assess whether the resulting factor structure was unidimensional versus multidimensional, a bifactor analysis was performed. The bifactor model allows researchers to extract a primary unidimensional construct while recognizing the multidimensionality of the construct [84–86]. The bifactor model assumes each item loads on two dimensions. The first is a general latent trait or factor that underlies all the scale items and the second, a secondary group factor. This approach allows researchers to assess whether the secondary factor explains a trivial amount of variance (i.e., adds only noise to the analysis, in which case it should be dropped and the solution be declared unidimensional) or whether the secondary factor contributes meaningful variance (in which case it should be retained and the solution declared to be multidimensional) [84–86]. To determine whether to retain a construct as unidimensional or multidimensional, the factor loadings from the general factor are compared to those from the group factors (sub scales) [85,86]. Where the factor loadings on the general factor are significantly larger than the group factors, a unidimensional scale is implied [85,95]. The model fitness of both the confirmatory factor and bifactor models were assessed using CFI, RMSEA and the Weighted Root Mean Square Residual [74–81] (S3 Table). In addition, we assessed a number of bifactor indices of variance explained. These included the Explained Common Variance (ECV), Omega, Omega Hierarchical (OmegaH), and Factor Determinacy [87,88,96,97]. We assessed intra-respondent reliability of scale items retrospectively by comparing daily recall across 30 days with the sum score of a retrospective recall on the 31st day. This was to assess the stability and consistency of responses on scale items [89]. The reliability of the scale itself was estimated using coefficient alpha and the coefficient of stability. First, Cronbach’s coefficient alpha of scale items was calculated for the samples at 15 and 18 months postpartum to compare and correlate the observed score variation between each of the items in the scales for both samples [81]. Second, we assessed the coefficient of stability (test-retest reliability), which involved the correlation of scale scores at 15 and 18 months postpartum [63,64,91]. We used predictive (criterion), construct (convergent and discriminant) validity and differentiation between ‘known groups’ to assess scale validity. Predictive (criterion) validity was assessed by examining the associations between HHWI and perceived maternal stress as well as food insecurity [56,57,89]. Convergent validity was measured against time to and from water source and amount of money spent on purchasing water in the past month. We calculated Pearson product-moment correlations based on Fisher’s transformation [89,91–93]. Discriminant validity was assessed by correlating HHWI with per capita water used daily [91–93], which has similarly been used in previous studies [2,3,36]. Consistent with the findings of Tsai et al [3], Hadley and Wutich [36], and Stevenson et al. [2], we assumed that there would be no meaningful effect between HHWI and per capita water used, i.e. that HHWI is distinct from household water use. This was calculated using an equivalence testing “two-sided test” (TOST) approach [98]. As a final measure of validity, we assessed the scale score by differentiating the position of ‘known groups’. In other words, we expected to have significantly higher HHWI scores for participants whose water was contaminated with E.coli, who were HIV positive, who used unimproved water sources, and were interviewed during the dry season. We used Student’s t-test for this analysis. We obtained approval for this study from the Institutional Review Boards at Cornell University, Northwestern University, and the Kenyan Medical Research Institute (KEMRI) Scientific and Ethics Review Committee (SERU). Also, we obtained written informed consent from all participants in this study.