Home gardening improves dietary diversity, a cluster-randomized controlled trial among Tanzanian women

Homestead food production (HFP) programmes improve the availability of vegetables by providing training in growing nutrient-dense crops. In rural Tanzania, most foods consumed are carbohydrate-rich staples with low micronutrient concentrations. This cluster-randomized controlled trial investigated whether women growing home gardens have higher dietary diversity, household food security or probability of consuming nutrient-rich food groups than women in a control group. We enrolled 1,006 women of reproductive age in 10 villages in Pwani Region in eastern Tanzania, split between intervention (INT) and control (CON) groups. INT received (a) agricultural training and inputs to promote HFP and dietary diversity and (b) nutrition and public health counselling from agricultural extension workers and community health workers. CON received standard services provided by agriculture and health workers. Results were analysed using linear regression models with propensity weighting adjusting for individual-level confounders and differential loss to follow up. Women in INT consumed 0.50 (95% CI [0.20, 0.80], p = 0.001) more food groups per day than women in CON. Women in INT were also 14 percentage points (95% CI [6, 22], p = 0.001) more likely to consume at least five food groups per day, and INT households were 6 percentage points (95% CI [−13, 0], p = 0.059) less likely to experience moderate-to-severe food insecurity compared with CON. This home gardening intervention had positive effects on diet quality and food security after 1 year. Future research should explore whether impact is sustained over time as well as the effects of home garden interventions on additional measures of nutritional status. This analysis considers data from a pair‐matched cluster‐randomized trial (ClinicalTrials.gov). The study design and population have been described in detail elsewhere (Mosha et al., 2018). Briefly, the study was implemented in Rufiji District, Pwani Region, Tanzania. Ten villages were randomly sampled from the Rufiji Health and Demographic Surveillance System (HDSS), a database that provided demographic and descriptive data on households in the study area (Mwageni et al., n.d.). The villages were matched into five pairs based on location, proximity to water source that could be used for irrigation (for instance, river, well or running tap water) and population size. Villages in each pair were randomly assigned to INT or CON. Randomization was done by colleagues with no prior knowledge of the villages. Households within each village were selected based on their eligibility and were approached by field workers for enrolment and informed consent. Households that met the following eligibility criteria were recruited into the study: (a) had a woman between 18 to 49 years of age at time of recruitment and at least one child younger than 36 months of age, (b) household had access to a plot of land where vegetables can be grown and (c) provided informed consent. Across INT and CON, a total of 1,006 women were recruited into the study between August and October 2016, and a follow‐up assessment was conducted at 12 months postintervention initiation. For this pilot trial, we aimed to assess the feasibility and preliminary effectiveness of the intervention on dietary diversity and have not done formal power calculations in establishing sample size. The sample size of 1,000 households was predetermined by what was both logistically feasible for a time period of 1 year and also sufficient for pilot implementation, but no formal sample size calculation was performed a priori (for more details, please see Mosha et al., 2018). The intervention included two main components: (a) agricultural training and inputs to promote HFP and (b) nutrition and public health counselling to improve diet and health‐related behaviours. CON received the standard of care for the area, where agricultural extension services are offered in a nonstandardized way and community health services had not been established. Agricultural training consisted of 15 main topics: (a) overview of homestead gardening, (b) fertilizer management, (c) different types of fertilizer, (d) agronomical practices and irrigation, (e) controlling harmful insects and vegetable diseases, (f) pests and pest management, (g) harmful effects of pesticides, (h) irrigation support, (i) farm processes, (j) crop harvesting, (k) composting, (l) marketing of vegetables, (m) nursery preparation, (n) raised bed preparation and (o) transplanting. Participating households received three or four types of seeds out of six local crop varieties (African eggplant, amaranth, spinach, tomato, okra and Chinese cabbage, all from the Kibo Seeds Company) at least three times during the study period. Seed types were selected based on local climatic and production conditions and local preferences. Households also received urea and cow manure fertilizer and watering cans. The nutrition and health counselling for behaviour change consisted of 15 main topics: (a) food and water safety and hygiene, (b) food preparation and storage, (c) immunization, (d) handwashing, (e) key nutrition terms and importance of human nutrition, (f) maintaining a balanced diet, (g) physical activity, (h) food equivalents to meet energy needs, (i) causes and consequences of malnutrition, (j) breastfeeding, (k) complementary feeding, (l) advantages of vitamin A supplementation and deworming, (m) maternal and child health seeking behaviour, (n) importance of anthropometric measures and (o) importance of consuming iodized salt. The intervention engaged the existing community workforce of agricultural extension workers (AEWs), livestock extension workers (LEWs) and community health workers (CHWs) to deliver agriculture training and behaviour change messages to participants. Messages were delivered through two main mechanisms: households received visits from either AEWs/LEWs or CHWs every 2 weeks on a rotating basis. AEWs, LEWs and CHWs all received training on all topics for the intervention according to the training manual, including agricultural practices, basic health messages (such as water, sanitation and hygiene) and nutrition (including the importance of dietary diversity). All workers were trained on nutrition messaging so that while CHWs focused on health and nutrition messages, AEWs and LEWs were able to reinforce these messages. Additionally, approximately every 2 weeks, farmer field schools (FFS) were held in collaboration by AEWs, LEWs and CHWs. The FFSs were held at the garden of a participating household, with a typical attendance of 10–15 programme participants from the nearest hamlet. During the FFSs, messages from the household visits were reinforced, benefits of improved agricultural practices were demonstrated (with nonparticipant households welcome to attend), and community knowledge about local availability of nutritious crops was shared. The field schools also served as a forum for collaboration and discussion and as a platform for women’s empowerment as successful model farmers shared their experience and taught their peers best practices for home gardening. The study manager continuously monitored the AEWs and CHWs to ensure routine delivery of the intervention components, both at household visits and FFS sessions. Each week, two participants were randomly drawn from each intervention village to be visited by the field manager for monitoring purposes. Data were collected at baseline and after 1 year. Surveys were administered on electronic tablets by trained interviewers. The survey questionnaires were developed by the research team and included modules on household socio‐economic status, food frequency intake, HFP and food security. Each household was assigned a composite wealth score derived from household assets (roof type, whether roof leaks, floor type, electricity, couch, television and bike ownership) using principal components analysis (Filmer and Pritchett, 2001). Outcomes for this analysis include dietary diversity and food security. The primary outcome for the trial was dietary diversity, prespecified in the ClinicalTrials.gov registration, while the secondary outcome food security was added post hoc. Dietary diversity was measured as the number of food groups consumed out of 10 using a locally adapted food frequency questionnaire (FFQ) that has been tested for validity (Zack et al., 2018) and used in previous trials (Bliznashka et al., 2020; Gerber et al., 2020). Participants were asked the average frequency of consumption of a given food item over the past 30 days using options ‘0 times in a month, ‘1–3 times per month’, ‘1 time per week’, ‘2–4 times per week’, ‘5–6 times per week’, ‘1 time per day’, ‘2–3 times per day’, ‘4–5 times per day’ and ‘6+ times per day’. The responses to these questions were used to calculate daily frequencies of consumption for each item. Frequency of consumption of each food item was aggregated into consumption of the following 10 food groups according to FAO guidelines: starchy staples (e.g. maize, bread and rice), flesh foods (e.g. beef, fish and chicken), vitamin A‐rich dark green vegetables (e.g. spinach, Chinese cabbage and sweet potato leaves), other vegetables (e.g. lettuce, eggplant and cucumber), other fruits (e.g. banana, guava and watermelon), other vitamin A‐rich vegetables and fruits (e.g. mango, papaya and tomato), dairy products (e.g. milk and ice cream), beans and peas (e.g. kidney beans, chickpeas and green peas), eggs, and nuts and seeds (e.g. Bambara nuts and ground nuts). Food items included in each category are detailed in the supporting information. A participant was considered to consume a given food group if the sum of daily frequencies for all foods in that group equalled or exceeded one. The dietary diversity score was then defined as the number of food groups consumed out of 10. Minimum dietary diversity for women (MDD‐W) was defined as a participant consuming five or more food groups per day (FAO & FH360, 2016). Food security, a secondary outcome, was measured using the Household Food Insecurity Assessment Scale (HFIAS), a measure of food insecurity produced by Coates et al. (2007). The questionnaire operates under the assumption that levels of food access and insecurity produce predictable responses that can be expressed and quantified in a score. Such responses include quantity, quality and intake of food; perceived uncertainty or anxiety for food situations; and associated consequences. The HFIAS scale was also categorized to no food insecurity or mild, moderate and severe food insecurity (SFI) (Coates et al., 2007). We report descriptive statistics of the study population using means and frequencies in Table 1. Statistical analysis was conducted using Stata 15.1 (StataCorp LP). Data were analysed based on the intent‐to‐treat principle. Our analytic strategy is threefold: First, we present ‘unadjusted’ estimates from our pair‐matched trial design, which conditions exclusively on village‐level characteristics captured by the three criteria used for matching villages (location, proximity to water and population size) per the methods recommended by Imai et al. (2009). Second, we present pair‐matched results weighted by the probability (propensity) of treatment conditional on measured individual‐level confounders. These methods have been comprehensively described by Hernán and Robins (2020). We included inverse probability of treatment weights because we felt that there was not sufficient balance on individual covariates after the pair‐matched cluster randomization, and treatment weighting allows for adjustment by measured individual‐level confounders. Third, we present results weighted by propensity of both treatment and censoring. We included weights for censoring because we detected a differential loss to follow up between INT and CON. Of the 1,006 (504 INT and 502 CON) households enrolled in the study at baseline, 455 in INT and 421 in CON were reached at 12 months (2017). The sample was subjected to a 12.9% loss to overall follow up: 9.7% in INT and 16.1% in CON. Reported reasons for loss to follow‐up were out‐migration (45%), travel during data collection (30%), married out of village (13%), divorced out of village (6%), refusal (4%) or other reasons (5%) (see CONSORT chart, Figure S1). Baseline characteristics of women in CON and INT villages Abbreviations: BMI, body mass index; CON, control; INT, intervention. To understand the implications of these three analytic strategies, we report results from (a) unadjusted regressions with fixed effect for matched pair (the design‐based estimator), (b) weighted regressions with inverse probability of treatment weights and (c) weighted regressions with combined treatment‐and‐censoring weights. We focus the reporting and interpretation on the final (c) model results but present the unadjusted and inverse probability of treatment weighted estimates in Tables 3 and ​and44. Differences in dietary diversity score and household food insecurity score between intervention (INT) and control (CON) after 12 months of follow up Note: Adjusted models show risk differences from linear models with fixed effects for matched village pair, inverse probability of treatment and censoring weights, and robust standard errors. Confounders adjusted for by treatment weights include baseline response variable, baseline wealth quintiles, baseline education level and baseline livestock ownership. The Bonferroni corrected critical p value is 0.003. Interpretations: β1, effect estimate. Abbreviations: LCI, lower bound for 95% confidence interval; SE, standard error; UCI, upper bound for 95% confidence interval. Risk differences between intervention (INT) and CON after 12 months of follow up Note: Adjusted models show risk differences from linear regression models with fixed effects for matched village pair, inverse probability of treatment and censoring weights, and robust standard errors. Confounders adjusted for by treatment weights include baseline response variable, baseline wealth quintiles, baseline education level and baseline livestock ownership. The Bonferroni corrected critical p value is 0.003. Abbreviations: CON, control; LCI, lower bound for 95% confidence interval; RD, risk difference; SE, standard error; UCI, upper bound for 95% confidence interval. To evaluate effects of the intervention on women’s dietary diversity scores and household food insecurity scores, we fit linear regression models with fixed effects for matched‐village pair and inverse probability of treatment and censoring weights with robust variance estimators (Thoemmes and Ong, 2016). We computed inverse probability of treatment weights by predicting probabilities of being treated from logistic regression models stratified by matched pair controlling for baseline response variable, baseline wealth quintiles, baseline education level and baseline livestock ownership. Censoring weights were similarly obtained from pair‐stratified logistic regression models that included any significant predictors of loss to follow up: baseline marital status, baseline livestock ownership and baseline wealth quintiles (supporting information). To evaluate the effects of the intervention on the dichotomous outcomes of (a) achieving MDD‐W, (b) consuming a given food or (c) having a certain level of food insecurity, we fit linear probability models such that the intervention effect is expressed as a risk difference (RD). We do not report risk ratios, as there was a noticeable amount of effect heterogeneity across village pairs and a summary risk ratio would be misleading. We control for the family‐wise error rate—the probability of getting at least one false positive result—for multiple comparisons using the Bonferroni correction. We performed 16 hypothesis tests and therefore set our critical p value (probability level, p value) to 0.003. Missing data among surveyed households were less than 5%; therefore, we performed a complete‐case analysis. Models that account for pair‐matched cluster‐randomized design estimate the effect while assuming that observations within a pair are independent (Imai et al., 2009). To evaluate potential additional within‐village correlation in outcomes within pair, we fit mixed‐effects models with matched pair dummy variables and village‐level random effects and estimated an intraclass correlation coefficient (ICC) for dietary diversity scores of 0.004. From this, we conclude that additional within‐village clustering is minimal, and we do not further adjust for clustering beyond reporting Huber–White robust standard errors. We also estimated and report bootstrapped confidence intervals using wild bootstrap (null imposed, 999 replications, Wald test and Rademacher weights) and find similar values. The study protocol was approved by the institutional review boards of Ifakara Health Institute, the National Institute for Medical Research of Tanzania (NIMR/HQ/R. 8 a/Vol. IX/2262) and the Harvard T.H. Chan School of Public Health.