Description of study area
This study was conducted in the Karamoja subregion in the districts of Nakapiripirit and Kotido (Fig. 1). The subregion intensely experiences climate variability owing to its semiarid conditions. [10] Annual rainfall in the region varies from around 400–1000 mm depending on location. Meanwhile, mean temperature ranges from 16 to 24 °C. Owing to intense rainfall variability and high evapotranspiration in the region, intermittent drought episodes have become characteristic in the subregion [23].
The DEWS implementation was first piloted in the two districts of Kotido and Nakapiripirit. These two districts have a predominance of agro-pastoralism as the main subsistence production practice. Kotido subsists in the central sorghum and livestock zone while Nakapiripirit lies in the western mixed crop-farming zones; these are the two predominant zones in the Karamoja subregion [25]. The population statistics of Nakapiripirit and Kotido districts are 69,691 and 178,909, respectively [22]. A mixture of cereals and legumes are grown including some pulses such as sorghum, pearl millet, maize, groundnuts and sunflower. Meanwhile, livestock breeds such as sheep, goats, sheep and camels are kept.
Data collection
A cross-sectional survey was conducted in 2014 from purposively selected sub-counties based on the DEWS coverage. A multistage sampling criterion was used sequentially across two hierarchical levels. At the first level, sub-counties were purposively selected from each district based on the DEWS project selection criteria. At the second level, three to four parishes were purposively selected from each sub-county based on project coverage. Three participating parishes and non-participating parishes were selected from each sub-county. This amounted to a total of four parishes per district from which household respondents were purposively selected. All the parishes covered by the DEWS were considered for the study. Of the selected households, 173 participating and 132 non-participating households were used in the survey. One hundred and seventy-three (173) households were considered in order to ensure a total coverage of all DEWS participating households in the two project districts. While the sample size for the control households was initially the equivalent of the DEWS participating households included in the study, the number reduced because during the process of data cleaning, some households were dropped.
The study used guided semi-structured questionnaires to collect data. Guided questionnaires were used because of high illiteracy rates in the study communities at 90 and 83% for Kotido and Nakapiripirit districts, respectively [22]. Under such conditions, it is inconceivable to undertake self-administered questionnaires in data collection as the response rate and data quality would be extremely low. Data were collected by research assistants who knew the local language (Karamojong). Information and records of government and non-government reports were also reviewed from the national archives in Entebbe, Makerere University, Meteorology Department, in the Ministry of Water and Environment and other government departments. During data collection, the interviewers were able to make direct field observations on crop and livestock management practices at household level.
In this study, a household was taken as “the basic unit of society involved in production, reproduction, consumption and socialization” [26]. This meant that household members share a residence and meals, and make coordinated decisions, resource allocation and income pooling in some cases [26]. In addition, they recognize the authority of a single head of the household in major decisions relating to drought preparedness and response actions. Data captured were on the wider and specific social contexts of food security and DEWS information utilization at household level. Further, within the semi-structured questionnaire, the nine questions that helped in the construction of the household food insecurity access scale (HFIAS) model included: (1) worrying about getting enough food; (2) actual failure to get enough food; (3) eating poor quality foods; (4) relying on a few kinds of foods; (5) reducing the amounts of food eaten; (6) skipping meals; (7) eating less than what one feels they should have eaten; (8) not eating for a whole day because of lack of food; and (9) growing thinner because of not eating enough food were embedded in the instrument. These questions were asked based on a dummy approach with the respondents expected to either say yes or no. In this case, the responses were used to generate the raw food security scores ranging between 0 and 9 points with 0 representing the most food secure households while 9 indicates an extremely food insecure household [27, 28]. Attribution of DEWS to food security was determined by a comparison between the type of household herein called DEWS household and control household who did not participate in the DEWS intervention.
Similarly, in developing a link between DEWS and food security, the survey questionnaire contained a component of dietary diversity assessment [29]. Dietary diversity is a qualitative measure of food consumption reflecting the households’ access to a variety of foods. It is a proxy of nutrition adequacy. Dietary diversity also reflects a snapshot of the economic ability of a household to consume a balanced diet. Respondents were asked to score on the food items they consumed in the last 24 h. The food items included cereals, vitamin-rich vegetables, roots and tubers, dark leafy vegetables, other vegetables; vitamin A fruits; other fruits; meat, poultry, offal; eggs; fish; pulses/groundnuts/legumes; milk and milk products; oil/fats; sugar/honey. The expected response was either yes = 1 or no = 0 following Kefasi Nyikahadzoi [29] method.
Data analysis
Determination of dietary diversity
The household dietary diversity score (HDDS) was calculated as: HDDS (0–14) = Sum (A + B + C + D + E + F + G + H + I + J + K + L + M + N, reflecting the total number of food groups consumed by members), where A = cereals; B = vitamin-rich vegetables; C = roots and tubers; D = leafy vegetables; E = other vegetables; F = vitamin A fruits; G = other fruits; H = meat, poultry, offal; I = eggs; J = fish; K = pulses/groundnuts/legumes; L = milk and milk products; M = oil/fats; N = sugar/honey. The results of analysis of the HDDS were to support the HFIAS model to determine the food security status of the household. For this study, all the vegetables were merged as one food item, and all the fruits were merged as one, making a total of 12 food items. This was because the vegetables eaten in Karamoja were mainly wild vegetables (there were no domesticated vegetable varieties) to be categorized as leafy vegetables and other vegetables. The same was true in fruits. People mostly ate wild fruits.
Descriptive and econometric methods were used to analyze the data. Descriptive analysis involved generating means, percentages and correlations. A generalized linear model was used to determine which factors influence food security and dietary diversity. Dietary diversity is a qualitative measure of food consumption that reflects household access to a variety of foods, and is also a proxy for nutrient adequacy of the diet taken. Household dietary diversity scores (HDDSs) were used as a measure of household nutrition security [30].
Effects of drought early warning systems on food security
A generalized linear model (GLM) was used to determine the factors in household food security outcome, with proportion of HDDS on a range of 12 as the dependent variable. The model assumed a relationship between observations y of the random response variable Y and a probability (density) function [31]. The GLM supports in the development of a strategy for approaching statistical problems that involve non-normally distributed data, in a way that retains much of the simplicity of linear models. The model used the assumption of exchangeability in that if Y is the dependent variable, Xs are explanatory variables, i.e.,
$$Y \equiv Y\left[ U \right] = Y\left( {u_{1} } \right), \ldots , Y\left( {u_{i} } \right) \; {\text{on}}\;{\text{sample}}\;{\text{units}}\;{\text{and}}\;X \equiv X\left[ U \right] = X\left( {u_{1} } \right), \ldots ,X\left( {u_{i} } \right).$$
Assuming exchangeability, \(X\left[ U \right] = X\left[ {U^{\prime}} \right]\) implies for all U, U′ \(\subset {U}\).
The GLM also assumes independence of error terms of the various sampling units in a way that \(Y\left( {u_{1} } \right), \ldots , Y\left( {u_{n} } \right)\) are independent. The Y and the error term tend to a normal distribution so that \(Y \sim N\left( {\mu = X\beta ,\delta^{2} ln} \right), E\left( {Y\left( U \right)} \right) = \beta_{1} X_{1} + \cdots + \beta_{p} X_{p} \left( u \right)\) [32].
The model was therefore specified as in Eq. (1) and estimated as in Eq. (2).
$$Y = \beta_{0} + \beta_{i} X_{i} + \cdots + \mu$$
(1)
$$HDDS = \beta_{0} + \beta 1_{1} X_{1} + \beta_{2} X_{2} + \beta_{3} X_{3} + \beta_{4} X_{4} + \beta_{5} X_{5} + \beta_{6} X_{6} + \beta_{7} X_{7} + \beta_{8} X_{28} + \beta_{9} X_{9} + \varepsilon_{i}$$
(2)
where HDDS is the proportion score of a household on the household dietary diversity score scale, X1 is gender of the respondent, X2 whether or not the respondent has participated in DEWS intervention, X3 is age of the respondent, X4 is the educational level of the household head, X5 is the distance to the water source, X6 is the total land owned by the household, X7 is the land utilized by the respondent, X8 is the household labor and X9 is the distance to the local trading center and ɛ is the error term.
The GLM was also used to determine the factors affecting household food security outcomes, with proportion of the household food insecurity access scale (HFIAS) score as the dependent variable and independent variables as in Eq. (2).
A multi-collinearity test was done to check for associations among continuous variables and discrete variables which seriously affect the parameter estimates. As Gujarati [33] indicates, multi-collinearity refers to a situation where it becomes difficult to identify the separate effect of independent variables on the dependent variable due to existing strong relationship among them. In other words, multi-collinearity is a situation where explanatory variables are highly correlated. There are two measures that are often suggested to test the existence of multi-collinearity. These are variance inflation factor (VIF) for association among the continuous explanatory variables and contingency coefficients (CC) for dummy variables. In this study, both VIF and CC were used to check multi-collinearity of continuous variables. The computerized Statistical Package for Social Scientists (SPSS), version 18, was used to compute both VIF and CC. The results from multi-collinearity test show that there is multi-collinearity between the use of the DEWS information and households who were participating in DEWS intervention. A variance inflation factor (VIF) greater than 2 is usually considered problematic and there were two variables in the model with VIF more than 3 indicating possible effect of multi-collinearity between the variables. Therefore, the use of DEWS information was dropped from the model.