Can nutrition education mitigate the impacts of COVID‐19 on dietary quality? Cluster‐randomised controlled trial evidence in Myanmar's Central Dry Zone

Abstract We evaluate the immediate impact of a nutrition and gender behaviour change communication on dietary quality in rural communities in Myanmar and assess whether the communication helped mitigate the effect of the COVID‐19 crisis on dietary quality. The intervention was designed and implemented as a cluster‐randomised controlled trial in which 15 villages received the intervention and 15 control villages did not. The intervention was implemented from June to October 2020. This paper provides an assessment of the intervention's impact on dietary quality based on the results of two phone surveys conducted in August and October 2020. Immediate impacts of the intervention indicate an improvement in women's dietary diversity scores by half a food group out of 10. At baseline, 44% of women were likely to have consumed inadequately diverse diets; results indicate that 6% (p‐value: 0.003, SE: 0.02) fewer sample women were likely to have consumed inadequately diverse diets. More women in treatment villages consumed pulses, nuts, eggs and Vitamin A‐rich foods daily than in control villages. In response to economic shocks related to COVID‐19, households in the treatment villages were less likely to reduce the quantity of meat and fish consumption than in control villages. The long‐term impacts of the intervention need to be continuously evaluated.

slowing harvests, constraining food transport, forcing market closures and resulting in food loss (Aldaco et al., 2020;Amicarelli & Bux, 2020;IFPRI, 2020;Resnick, 2020). From a consumer perspective, reduced income may lead to less purchasing power for food, particularly among the vulnerable. It is estimated that, by the end of 2020, COVID-19 resulted in an additional 130 million people becoming food insecure (Anthem, 2020); 45 million people becoming acutely food insecure, 33 million of whom live in South and Southeast Asia (UNSDG, 2020) and 6.7 million children younger than 5 years wasted and stunted (Headey, Heidkamp, et al., 2020). Myanmar's first round of lockdown and other measures implemented in April 2020 were initially able to control the spread of the infection, but the crisis led to a sharp spike in poverty followed by a modest economic recovery in mid-2020 Headey, Heidkamp, et al., 2020).
A second, far worse phase of the crisis began in August 2020, accompanied by the imposition of a second lockdown and other measures in mid-September 2020. Results from a high-frequency International Food Policy Research Institute (IFPRI) survey show that Myanmar's second wave of COVID-19 infections led to severe reductions in income, resulting in deeply worrying increases in poverty and food insecurity (Headey, Zaw Oo, et al., 2020). In a national household survey, 83% of households reported lower than usual income between March and October 2020, with an average decline in income of 43%; poor and vulnerable households were more likely than those that were financially secure to experience reductions in income (CSO, MoPFI, & UNDP, 2020).
As of January 2021, in a country with a population of 53.7 million, the number of people with insufficient food consumption increased to 15 million, with an increase of 0.37 million from October to December 2020 (WFP, 2021). Due to a lack of money, 30.7% and 40.5% of households urban and rural households, respectively, reported eating less than usual between March and October (CSO, MoPFI, & UNDP, 2020). Healthier diets cost more than unhealthy diets; as household income declines, it is likely for diets to shift toward cheaper, more energy-dense, less micronutrient-dense foods (Alemu et al., 2019;Green et al., 2013;Miller et al., 2016). In a sample of women with young children in Myanmar's Central Dry Zone, Maffioli et al. (2021) find that those with incomes affected by COVID-19 experienced significantly lower food security and diet diversity.
Social protection programmes, particularly cash transfers, have become common strategies to mitigate the negative effects of the COVID-19 crisis. However, studies have also shown that nutrition communication with and without cash transfers is effective in improving diets and nutritional status (Dewey & Adu-Afarwuah, 2008). While behaviour change communication (BCC) interventions are recognised to be effective with additional material inputs or financial resources (Ferré & Sharif, 2014;Hoddinott & Skoufias, 2004;Lagarde et al., 2007;Zhang et al., 2018), research shows that intensive and well-designed interventions without input or financial support, even in households that have challenges with affording food, can also lead to better dietary quality and other nutrition indicators (Frongillo et al., 2019;Murendo et al., 2018;Reinbott et al., 2016;Warren et al., 2020). More research is needed to evaluate nutrition-only interventions and should include assessment of specific topics, BCC format, intervention context, strength of service delivery systems, recipient characteristics and length of exposure (Warren et al., 2020).
In Myanmar's Central Dry Zone, Maffioli et al. (2021)  • Further evaluation is needed to determine whether the immediate impacts are sustained over time.
Recovery Plan. 1 This paper presents a midterm assessment of the nutrition and gender intervention implemented in June through October 2020 in randomly selected rural communities and its causal impact on dietary quality.

| Study site
Myanmar's Central Dry Zone, an area with an estimated 12 million people or about 23% of Myanmar's total population, covers roughly a third of the country's grain cropping area (Herridge et al., 2019). This study focuses on 30 rural communities in the Central Dry Zone, which are located in the catchment areas of two irrigation sites: (1) Sinthe in Tatkone township in the Nay Pyi Taw region and (2) North Yamar in Pale and Yinmabin townships in the Sagaing region ( Figure 1). Rice is primarily grown in the irrigation sites, and other crops such as legumes and oilseeds are grown in nonirrigation (upland) areas. Growing rainfed crops has high risks, and access to water for irrigation is critical to reduce risks as well as to expand farmers' options (Boutry et al., 2017). Households in the sample communities are relatively better-off in terms of dwelling indicators compared to the average in the Central Dry Zone. This may be associated with greater economic opportunities coming from irrigation water access, relatively good infrastructure and access to markets. Most households have access to electricity and roads and are roughly 4-5 km or 20-30 min by motorbike from the nearest town centre, where household members might work in wage or salary employment or sell their produce. The main sources of purchased foods are vendors who come to the villages.

| Intervention
The intervention was a BCC on gender and nutrition implemented in June, July and October 2020 ( Figure 2). To minimise the risk of contamination, we adopted a cluster-randomised controlled trial in which the 30 relatively homogeneous sample villages are the clusters. We anticipate minimal differences across villages, strong information spillover within each village and minimal spillover across villages, justifying clustering and randomisation at the village level. We used a computer programme (Stata) to select villages and to randomly assign 15 villages into the treatment group (those receiving the BCC) and 15 villages into the control group (those not receiving the BCC). In each village, 29-32 households were randomly selected for the survey. In total, at baseline, there were 453 treatment households (those receiving BCC) and 465 control households (those not receiving BCC).
The Myanmar Institute of Gender Studies (MIGS) managed and implemented the BCC with the assistance of Save the Children and its partners who helped train nutrition facilitators. Nutrition messaging was designed around materials developed by Leveraging Essential Nutrition Actions to Reduce Malnutrition (LEARN). 2 LEARN's modules are designed specifically for the Myanmar context and cover nutrition basics, family nutrition, access to nutritious foods, food preparation and diet-related taboos (LEARN, 2015). The complementary gender component of the intervention is a potential pathway for improving dietary quality by shifting aspects of intrahousehold dynamics between women and men; those aspects include division of work and decision-making, particularly food decisions related to purchases, preparation and allocation (Ahmed et al., 2018;Farnworth et al., 2015). The study also aims to measure and track the impact of gender messaging on women's empowerment indicators to be measured during the endline evaluation survey. The gender messaging addresses gender and nutrition, as well as gender equality and equity, gender socialisation, gender stereotyping, gender-based violence, sex and gender, power and patriarchy. The actual content of the nutrition and gender materials were adjusted on the basis of gaps and 1 The amount of cash transfer in the MCCT programme was 10 000 kyat/month (about USD6.50/month) and increased to 15 000 kyat/month (about USD10.00/month) since October 2017 (Field & Maffioli, 2021). The COVID-19-related cash transfer is 20 000 kyat/ month. In comparison, the minimum daily wage is 3000 kyat. The amount of COVID-19-related cash transfer was equivalent to 6-7 days of work at the minimum wage. One-on-one household visits with preventive measures, distribution of materials and coaching took place in June 2020, followed by additional household visits and coaching sessions in July 2020. As cases began to rise in August, renewed restrictions prevented in-person engagements and training converted to phone-based monitoring and coaching in October 2020.
To enable comparison between women and men on agency, decision-making and nutrition-related knowledge and practices, one female adult and, when possible, one male adult were selected in each household to participate in the trainings and survey. Selections were based on decision-making in terms of livelihoods and food purchases and preparation. 3 There were no households without female adults in the sample, and 12% of sample households have only female adults.

| Survey sample and data collection
This assessment was based on three rounds of phone surveys conducted in June, August and October 2020 (see Figure 2 and Table S1). Each round of the phone survey inquires about respondents' experiences, activities and livelihoods in the preceding months, as well as food groups consumed the day before (24-h food recall) and frequency in eating meat, fish and vegetables in the week before the phone interview and how these indicators changed compared to the year before COVID-19 crisis (2019).
Phone surveys with village leaders of four communities were also conducted to inquire about the main COVID-19 mitigation measures implemented in these communities and about other shocks experienced during the survey period.
The sample at baseline was 918 households, and the subsequent three phone survey rounds included 606, 543 and 503 households, respectively. Household attrition from baseline to the first round of the phone survey was 39% and was primarily driven by phone-related issues. At baseline, 7% of households did not have a mobile phone, and an additional 7% did not provide their telephone numbers during the baseline survey. Furthermore, 25% of telephone numbers were not working during the first round. 4 Additional attrition from the first to third rounds occurred because of out-of-service telephone numbers, unanswered calls and interview refusals. 5 In total, household attrition was 45% from the baseline to the third round of survey. The GoM mandatory registration of SIM cards by 30 June 2020. This order likely led to a large share of phone numbers being blocked and users changing phone numbers. (https:// www.irrawaddy.com/news/burma/myanmar-mobile-users-told-re-register-sim-cards-june-30.html). 5 Attrition from phone survey round 1 to round 2 was 10%. Attrition from phone survey round 1 to 3 was 17%.

| Outcome indicators
F I G U R E 2 Time frame of the study. Source: Authors' illustration. Note: *Phone survey 1 asks respondents' experiences from February to May 2020 (onset of the pandemic), Phone survey 2 asks about experiences in June-July 2020 and Phone survey 3 asks about experiences in August-September 2020. The food group recall for measurement of dietary diversity pertains to the day before the interview. See also Table S1 for the details of the phone survey periods. **Lockdown and mobility restrictions are ongoing as of the writing of this paper 10 healthy food groups are consumed in a 24-h period (FAO & FHI360, 2016;Martin-Prével et al., 2015). There is no validated cut-off for other age groups. In this study we consider adult women primary decision-makers, including those above 49 years, and adopt the MDD-W threshold of five food groups as an indication of improved probability of adequate dietary diversity. The specific women's dietary diversity outcomes monitored are (1) a DDS measured by the number of healthy food groups consumed (0-10), (2) inadequate dietary diversity (<5 food groups) and (3)

| Statistical analysis
Given random assignment to the treatment, intention-to-treat (ITT) effects are estimated by ordinary least squares (OLS), where the variable of interest is the indicator variable equals to one if the village was assigned to the treatment group. The outcome can then be written as where Y is the outcome indicator (see the previous subsection), T is the assignment to the treatment, b measures the average effect of the treatment and a is the constant or intercept (which can also be interpreted as the value of Y if T = 0). We test the null hypothesis (b = 0).
If rejected, we conclude that the intervention has significant effect to the magnitude of b. Because the unit of randomisation is the village, standard errors from the regression models are clustered at the village level.
For the DDS, we performed a Poisson regression commonly used for count data (i.e. number of food groups consumed). As a robustness test, we also perform OLS mimicking a continuous outcome variable.
For the dummy variables measuring likelihood of inadequate dietary diversity, consumption of each of the 10 food groups and reported household reduction in meat and fish consumption due to the pandemic, we performed probit models.
We took advantage of the randomised experimental design and conducted an ITT analysis using single-difference estimation in the midline assessment data. The randomised assignment and balance in baseline characteristics minimise concerns of bias in the singledifference treatment estimates. As a robustness check, we performed additional regressions that include various demographic controls (see Table S5 for a summary of control variables). We also performed a regression that controls for self-reported household income loss, receipt of cash transfers and borrowing money during the COVID-19 crisis.
Throughout our analyses, we apply inverse probability weighting to attenuate attrition bias (i.e. we computed the inverse of the probability of the attrition probit [1/pr] and used it as attrition weight). This procedure gives more weight to households that have similar initial characteristics as households that subsequently dropped out than to households with characteristics that make them more likely to remain in the panel.

| Ethical considerations
This study was approved by Myanmar's Ministry of Agriculture, Livestock and Irrigation. Ethical approval for this study was obtained through the IFPRI Internal Review Board. The study is registered and reviewed by The American Economic Association's registry for randomised controlled trials. Informed oral consent was obtained from the study participants and recorded by trainers and enumerators, respectively.

| RESULTS
There is no difference between those in the treatment and control villages in terms of cell phone access or in-service telephone numbers, which indicates that the treatment effects among those with working telephone numbers and successful interviews can be extended to the full baseline sample. Attrition probit regressions were estimated to identify demographic factors that could explain the likelihood of attrition (Table S4). The characteristics of the full sample and phone survey subsample are similar, and the attrition probit shows low Pseudo R2, indicating low attrition bias.
Using a set of questions developed by Ragasa et al. (2020) to assess respondent's prior knowledge of topics covered in the LEARN modules, the baseline survey found that respondents scored particularly low on questions related to nutrition basics and family nutrition practices. Baseline assessment of respondents' food consumption in the previous 24 hours also found both men and women to be lacking the dietary diversity associated with healthy diets. Notably, the vast majority of respondents did not report consuming dairy or vitamin A-rich fruits and vegetables . Even when beans and nuts were seemingly available in the study areas, many households did not consume them on a daily basis. At baseline, 44% of women were likely to have consumed inadequately diverse diets (<5 out of 10 food groups). Finally, households in treatment villages were 11% (p-value: 0.000, SE: 0.02) less likely to reduce the quantity and 7% (p-value: 0.020, SE: 0.07) less likely reduce the frequency of meat and fish consumption during the COVID-19 crisis than those in control villages. However, the lower reduction in the number of days per week meat or fish that was consumed in treatment villages is significant only in the specification that controls for demographics together with changes in income.
Households in the treatment villages were 10% (p-value: 0.045, SE: 0.05) more likely to report income loss than those in control villages, whereas there was no difference in receipt of transfers and borrowing money during COVID-19 (Table S6). Controlling for selfreported income loss, receipt of transfers and borrowing increased the average treatment effect on nearly all statistically significant outcomes (Table 1) F I G U R E 4 Portion of sample households reporting reduced meat or fish consumption compared to a usual year (%). Source: IFPRI/MSR phone surveys (June, August and October 2020). Note: Meat or fish consumption pertains to the 7-day period before the phone interview. Statistical significance of the difference between control and treatment households at the 1%, 5% and 10% levels is indicated with ***, ** and *, respectively 6 Results are consistent with model estimates using attrition weights (Table S7). DDS estimates using OLS are nearly identical to those from the Poisson model and are thus not reported.
F I G U R E 3 Average dietary diversity score of women, by treatment and control groups. Source: IFPRI/MSR phone surveys (June, August and October 2020). Note: Dietary diversity pertains to the 24-h recall of food eaten the day before the phone interview in the same month as the survey. Statistical significance of the difference between control and treatment households at the 1%, 5% and 10% levels is indicated with ***, ** and *, respectively. Numbers in parenthesis are standard deviations reported income loss, with farming, business and nonfarm wage income impacted more heavily than farm wage income. Income loss not only hurts households directly but also has the potential compounding impact of reducing future income by diminishing funds available for production and investment Ragasa et al., 2020). With the implementation of the GoM's COVID-19 Economic Relief Plan, almost all respondents in treatment and control villages received transfers (20 000 kyat per household per month) during the study period (78% in June and July and 96% between August and September). In addition, a good proportion of the respondents is applied for more loans compared to usual years. In the June-July period, 57% of households borrowed money, and 43% borrowed in the following 2 months. The borrowings were mainly regular loans (150 000 kyat/acre) and COVID-19-related loans (200 000 kyat/acre) from the Myanmar Agricultural Development Bank (MADB).
As income decreases, consumption of more expensive foods, such as meat and fish, is expected to fall; indeed, reduced income was the main reason cited for consuming less meat or fish by about 76% of respondents in rounds 2 and 3 of the phone survey.
Transfers and loans were expected to have eased the income effect of the pandemic. Yet income loss is so extensive that transfers and loans are most likely insufficient to fully offset losses. In the second survey round (June-July) and third survey round (August-September), 38% and 57% of households, respectively, reported that either business or other income was at least 20% lower than usual.
Thus, we see most sample households reporting reduced meat and fish consumption due to income loss. Moreover, at the onset of the pandemic, 15% of respondents had false perceptions about the spread of COVID-19 through meat or fish consumption. Those false perceptions, as well as the reported reductions in availability of protein-rich foods such as meat and fish during periods of stringent mitigation measures, were worrying.
We demonstrate through this study that there is a role for nutrition education in complementing the financial assistance received Note: Source: IFPRI/MSR phone survey (August and October 2020). All regressions use rounds 2 and 3 of the phone survey sample, with survey round fixed effects, and clustering at village. Controls in column 2 include baseline head of household characteristics (age, education level, and occupation, such as agricultural farmer, labour, or other jobs) and baseline household demographic characteristics (type of household, township, household size, dummy indicating household is a water user), and dummies indicating whether the household has income loss due to COVID-19, has accepted transfers from government or nongovernment organisations, and has borrowed money during the COVID-19 crisis. a Ten MDD-W food groups. b Ten MDD-W food groups; score < 5 indicates lower likelihood of adequate dietary diversity. c We do not report the staple food group results because all respondents report consuming staples. *Statistical significance of coefficient estimates at the 10% level. **Statistical significance of coefficient estimates at the 5% level. ***Statistical significance of coefficient estimates at the 1% level.
during periods of severe shocks. Even in the presence of significant pressures to reduce consumption of more costly nutritious foods, women decision-makers in treatment villages had greater DDS and were more likely to consume some nutrient-dense foods (pulses, nuts and seeds, eggs, and vitamin A-rich fruits and vegetables) than those who did not receive the nutrition and gender messaging. The treatment effect on the DDS was half a food group. Among women who received the messaging compared to those who did not, the intervention also resulted in 6% (p-value: 0.003, SE: 0.02) fewer women who were likely to have inadequately diverse diets. This is an important improvement relative to the 44% of women at baseline who did not consume adequately diverse diets.
Our findings also highlight the potential for nutrition education in households with reduced income. Women living in households that reported income loss had significantly lower DDS than those who did not (Table S6). Furthermore, women in treatment households had a higher likelihood of income loss but still had higher DDS than those in the control group (Tables S6 and S8 only small positive effects on healthy eating. Furthermore, in many countries, literacy levels and access to smartphones are also lower in rural areas than in urban areas, making mobile-based nutrition messaging more challenging.
In our study area, relying solely on mobile nutrition messaging was not possible because of the lack of phone ownership or nonfunctional phone numbers for about 39% of the baseline survey households. Moreover, the functionality of some phones was insufficient to allow users to access flyers, pictures, videos and interactive features of MIGS's mobile messaging. Adaptations to more conventional BCC techniques involved distributing materials directly to households and greater use of individual-level coaching, both in person and via phone.
To conclude, this study shows that BCC delivered by through a range of tools, including household visits and phone-based coaching,

DATA AVAILABILITY STATEMENT
Data are available on request from the authors.