We employed a cohort study design by collecting data from a panel of 473 rural households in the Wolaita area. Rural villages in the area represent two agro-ecological divisions: Lowlands (with hot and semi-dry conditions) and Midlands (with relatively cooler and sub-humid conditions) [5, 26]. The study involved same participants at four data points in time (rounds): June 2017, September 2017, December 2017, and March 2018.
This study was conducted in two rural districts (Woredas), namely Humbo and Sodo Zuria in the Wolaita area [25, 27]. About 400,000 people live in the districts often suffering from chronic food insecurity problem [5, 28]. The households were recruited from lowland (< 1600 m) and midland (> 1600 m) areas.
Based on the amount and timing of seasonal rains, farming activities, crop harvest, and other area-specific contexts, we identified and accounted for four distinct agricultural seasons in our survey rounds [25, 29]. Survey round 1 (R1) was conducted in the month of June in the heavy rainy season. The second round (R2) was conducted in September when the main cereal crop harvest takes places; small rains in this season give opportunity for growing root varieties. The third round (R3) was conducted in December, a late post-harvest month. The dry season lasts from late December to February or March. The fourth round (R4) was conducted late in the dry season; the second rainy season. Accordingly R1 and R4 were in pre-harvest season, R2 was at main harvest and R3 late in post-harvest seasons [5, 25, 30].
As a subsample of the broader study, the current analysis included 473 households who were not taking part in the PSNP. We excluded households taking part in the programme [31, 32]. As the programme member households get periodic cash support, their income basis would differ from households not taking part in the programme [31, 33, 34]. We further intended to maintain comparability with previous studies [21,22,23].
We adapted the HFIAS questionnaire, which was recently validated in the rural Butajira district in central Ethiopia . The tool comprises nine questions which are based on the respondent’s recall of food insufficiency and related psychological responses in the past 30 days [4, 18]. The questions (and their shortened versions) are as follows:
Q1: ‘Did you worry that your household would not have enough food?’ (‘Worry for food’)
Q2: ‘Were you or any household member not able to eat the kinds of foods you preferred because of a lack of resources?’ (‘Unable to eat preferred foods’)
Q3: ‘Did you or any household member eat just a few kinds of food day after day because of a lack of resources?’ (‘Eat a limited variety of foods’)
Q4: ‘Did you or any household member eat food that you did not want to eat because of a lack of resources to obtain other types of food?’ (‘Eat foods that you did not want’)
Q5: ‘Did you or any household member eat a smaller meal than you felt you needed because there was not enough food?’ (‘Eat a smaller meal’)
Q6: ‘Did you or any household member eat fewer meals in a day because there was not enough food?’ (‘Eat fewer meals in a day’)
Q7: ‘Was there ever no food at all in your household because there were no resources to get more?’ (‘No food to eat of any kind’)
Q8: ‘Did you or any household member go to sleep at night hungry because there was not enough food?’ (‘Go to sleep at night hungry’)
Q9: ‘Did you or any household member go a whole day without eating anything because there was not enough food?’ (‘Go day and night without eating’); the capital letter ‘Q’ denotes a question and subscripts 1–9 are item numbers in increasing severity. Each question in the tool includes a follow-up item to determine the frequency of occurrence whose responses are coded as often ‘3’, sometimes ‘2’, rarely ‘1’, or not at all ‘0’
A language expert and one of the investigators (BYK) translated the questionnaire into the local Wolaita language. We then interviewed five women in lowland and five in midland areas to ensure that each item was understandable and not easily misinterpreted. These women were later not included here in the main analysis. We asked each woman all the nine questions and recorded their responses. Afterwards, each woman was asked if she understood a particular question. When limitations that could compromise the intended meanings were noted, we asked the women how such items could be improved. Finally, the 10 respondents, two field supervisors, and one of the investigators (BYK) discussed the results, and finally BYK compiled these into a modified module .
The preliminary test of the HFIAS revealed several contextual and cultural sensitivities. For example, all the 10 women answered ‘yes’ to Q1 (‘worry for food’). They identified the term ‘worry’ in Q1 as an ordinary situation, this required contextual modification. Respondents also described Q4 (‘Eat foods that you did not want’) as intrusive and embarrassing, thus requiring the addition of a brief discussion about area-specific food taboos, including a list of foods that are consumed only during extreme food shortages. Some of the women were shy or afraid of replying ‘yes’ for Q7 (‘No food to eat of any kind’), Q8 (‘Go to sleep at night hungry’), and Q9 (‘Go day and night without eating’). This could be due to religious perceptions that disclosing such extreme situations would be perceived as insubordination to God. These three items thus required a focused training to interviewers on probing skills. For this to be effective, having interviewers who were better aware of extreme food shortages conditions had a particular importance.
Ten data collectors and two supervisors who are native speakers of the local language were recruited and given training on the different modules of the questionnaire. The same data collectors (in most cases) interviewed the same respondents in all the four rounds. However, in some rare cases a similarly trained one data collector was substituted. Women were mainly recruited as respondents in this study. However, when a woman was unavailable, any adult who was present and ate food in the household in the previous day was asked. Women are commonly responsible for food preparation and child feeding roles in their households in the study area [23, 35, 36]. Moreover, the women commonly remain at home more often than any other family member in this area.
The household food insecurity was measured by using responses to the nine FI occurrence questions (Q1-Q9) and their follow-up frequency of occurrence items [18, 23]. Accordingly, households were grouped into four categories (levels). A household is food-secure if it scored ‘0’ or ‘1’ in the first FI frequency of occurrence question and ‘0’ in Q2 to Q9; Mildly food-insecure if the first FI frequency of occurrence item has ‘2’ or ‘3’ or the second item has ‘1,’ ‘2,’ or ‘3’ or the third item has ‘1’ or the fourth item has ‘1’ and items Q5 to Q9 score ‘0’; Moderately food insecure if item three = ‘2’ or ‘3’ or item four = ‘2’ or ‘3’ or item five = ‘1’ or ‘2’ or item six = ‘1’ or ‘2’ and item seven to nine = ‘0’; and.
Severely food insecure if item five = ‘3’ or item six = ‘3’ or item seven = ‘1,’ ‘2,’ or ‘3’ or item eight = ‘1,’ ‘2,’ or ‘3’ or item nine = ‘1,’ ‘2,’ or ‘3’ [18, 37].
The overall household FI prevalence was computed as the proportion of food-insecure households out of the total interviewed. The mean differences of each consecutive pair of data time points were considered to estimate seasonal variations of the outcome measure.
A wealth index was constructed using a principal component analysis of the data on household-level assets. These assets included the housing structure (upper most cover, interior roof, floor, and wall) based on construction materials as observed by the interviewers, as well as possession of items such as radios, mobile telephones, beds, mattresses, kerosene lamps, watches, electric or solar panels, chairs, tables, wooden boxes, and carts. The four components of the housing structure had ordinal responses ranked mostly from 0 to 3. However, household assets had responses from 0 to 1 only. We standardized these scores to reduce the tendencies that variables with greater response would underestimate the others. Through a dimension-reduction analysis these inventories with standardized scores were summarized into logical dimensions. Based on standardized scores, the households were lastly categorized into relative wealth quintiles: Poorest (20th percentile), Poor (40th percentile), Medium (60th percentile), Rich (80th percentile), and Richest (>80th percentile).
The dietary diversity was measured by using Household Dietary Diversity Scale (HDDS) . The HDDS was derived from previous day’s consumptions of households based on 12 food groups; ranging between 0 and 12 scores [39, 40]. Respondents were asked qualitatively about their entire households’ food intake in the 24 h preceding each data collection round of the survey, focusing on consumption of 12 food groups: (i) meat; (ii) fish; (iii) vegetables; (iv) fruits; (v) eggs; (vi) potatoes and other roots or tubers; (vii) dairy products; (viii) pulses (ix) cereals and breads; (x) oil, fat, or butter; (xi) sugar or honey; and (xii) other foods, such as coffee and tea . When the respondent was asked if her family had diet from a particular food group, cereal for example, she would reply either ‘yes’ or ‘no’. Accordingly a ‘yes’ response was coded as ‘1’ and if ‘no’ it was coded as 0. The sum of the ‘yes’ response codes for the 12 food groups gives us the HDDS.
We used SPSS software (version 25 Inc., Chicago.IL) and Stata (Version 14, Stata Corporation, College Station, TX) for data analyses. Reliability analysis was conducted to estimate Chronbach’s alpha values. Likelihoods of affirmative responses were evaluated for parallelism across wealth quintiles. Extended Mantel-Haenszel chi square for linear trend was used to check for dose-response relationships between wealth and FI strata. Reproducibility of item responses was evaluated for pairs of related seasons (between pre-harvest seasons and also between post-harvest seasons) through paired t-test for equality of means of the HFIAS scores. We applied one-way analysis of variance (ANOVA) with robust tests of equality of means for multiple comparisons [41, 42].
We did an exploratory analysis involving the nine items through a varimax rotation of responses. A Horn’s parallel analysis (PA) was used to determine the number of factors to retain based on observed eigenvalues compared with that obtained from uncorrelated normal variables.
The following criteria were used for validation of the nine HFIAS tool: Chronbach’s alpha values approaching 0.85 to assert internal consistency [21,22,23]; parallelism of item-responses across wealth quintiles; and the presence dose-response relationship between wealth and FI strata, and between dietary intake and food security [41, 42].