This cross-sectional study using secondary data followed a sequential, mixed-methods design in which findings from the quantitative analyses informed the direction of the qualitative analyses.

Study design & intervention background

This study used existing data from the Food and Agricultural Approaches to Reducing Malnutrition (FAARM) trial (ClinicalTrials.gov: NCT02505711). FAARM was one of 13 projects across Africa and South Asia involved in the Gender, Agriculture, and Assets Project – Phase 2 (GAAP2). GAAP2 partners collaboratively developed the tools utilized in this study, including the pro-WEAI and accompanying modules, discussed below.

FAARM was a cluster-randomized controlled trial designed to determine the impact of a homestead food production (HFP) intervention on women’s and children’s nutritional status. The intervention was implemented by Helen Keller International from 2015 to 2018 in rural Habiganj District, Sylhet Division, Bangladesh. Further information about the setting and the design of the FAARM trial is available in the study protocol [36]. In brief, households were eligible for enrollment in the FAARM trial if there was a woman under the (self-reported) age of 30, who had access to at least 40 square meters of land and who was married to a husband who spent the night at home at least once in the previous year. After eligible women were identified, 96 settlements (villages or sections of villages) of 10-65 eligible women were formed; these settlements were then randomly assigned to intervention and control arms. Overall, approximately 2,700 women were enrolled in the trial [36].

The HFP intervention was implemented through women farmer groups; group members selected a lead farmer family who was then tasked with overseeing the group’s model farm, producing and distributing seeds and seedlings, hosting regular meetings, and facilitating training. Women’s empowerment was one pathway through which it was hypothesized that the intervention would lead to improved nutritional outcomes. Specifically, certain intervention activities were intended to transfer assets, knowledge, and skills to women; to increase women’s control over productive resources; and to create opportunities for women’s group participation and leadership [36].

Quantitative data collection

The main pro-WEAI [17] and modified health and nutrition modules [37] were administered in April and May 2019 (4-5 months after the intervention was completed). Detailed methods for the survey are provided elsewhere [38]; information relevant to the current study is summarized below.

The survey targeted five households in each of the 96 FAARM trial settlements for a total sample of 480 households. Targeted households were randomly selected from a list of households with fewer than 10 members and only one woman enrolled in FAARM at baseline (for ease of survey administration). Four female enumerators were trained and administered the survey in Bengali to the enrolled woman in each selected household. Written informed consent by signature, or thumbprint and the signature of an independent witness (i.e., a subject that is not part of the study team) if the subject was illiterate, was obtained from all FAARM participants at the beginning of the trial in 2015.

The pro-WEAI survey tool was designed to identify and evaluate key empowerment outcomes of agricultural development interventions [17]. The tool consists of 12 indicators representing three domains of agency: intrinsic agency, instrumental agency, and collective agency. Self-efficacy is classed as an indicator of intrinsic agency while most decision-making indicators fall under instrumental agency. The main pro-WEAI includes decisions related only to the productive sphere (i.e., agricultural and income-related decisions) while the supplemental health and nutrition module – designed to measure aspects of instrumental agency more directly related to health and nutrition – includes decisions related to the domestic sphere [37]. Our study focuses on 15 decision-making topics/activities (i.e., diet, nutrition, health, and healthcare seeking behaviors related to the woman respondent herself or to her children), which were captured by the health and nutrition module (See Supplementary Table A). The pro-WEAI also includes the 8-item New General Self-Efficacy scale (NGSE), which was used to generate a GSE score for each woman, as described below.

Some of the demographic data used as control variables in the final models was collected during the FAARM baseline survey, which was administered to all enrolled women from March to May 2015. Additional information on the FAARM baseline survey is available elsewhere [39].

Quantitative data analysis

Measures

The NGSE scale includes statements about generalized self-efficacy (e.g., “I will be able to achieve most of the goals I have set for myself”) to which the respondent indicates level of agreement on a 5-point, Likert-type response scale that ranges from strongly disagree to strongly agree. The GSE scores for each respondent were summed for a total possible score of 40, which was used in our models as a continuous measure [18].

To assess participation in decision-making, women were asked, “When decisions are made about [activity], who normally takes the decision?” Women who listed themselves as the only decision-maker for a given decision were classified as sole decision-makers for the given activity. Women who listed themselves and at least one other household member were classified as joint decision-makers for the given activity. Only joint decision-makers were then asked, “How confident do you feel to make decisions about [activity]?” This question, which we used as a measure of SSE, had an ordinal response scale of not at all, somewhat, and very confident. Very few women reported that they were not at all confident for any of the health and nutrition activities. Therefore, we assigned a value of 0 to not at all and somewhat confident responses and a value of 1 to very confident responses. These values were then summed across the three activities assigned to each of five activity categories (personal health and diet; personal diet during pregnancy; personal health during pregnancy; child’s diet; healthcare seeking) (See Supplementary Table A for the full list of domain-specific self-efficacy items and the indices to which they were assigned). Women were given an SSE score for each activity category ranging from 0, for describing themselves as not at all or somewhat confident across all activities, to 3 for describing themselves as very confident across all activities in a given activity category.

Decision-making autonomy was measured using a dummy variable that assessed agreement between two items: “When decisions are made about [activity], who normally takes the decision” and “When decisions are made about [activity], who would you prefer made the decision?” The decision-making autonomy variable for each activity was only calculated for the women who were involved in the relevant decision where they took a value of 1 if the woman preferred to be involved (i.e., classified as having decision-making autonomy) and a value of 0 if she did not prefer to be involved (i.e., classified as not having decision-making autonomy). We created decision-making autonomy indices by summing the dummy variables for the three activities within each of the five activity categories. Women were given a score ranging from 0 (non-autonomous decision-making across the category) to 3 (complete autonomous decision-making across the category) for each category. Questions regarding health and nutrition during pregnancy and about children’s diet were only asked to women with relevant experiences.

Regression modeling

An ordinal logistic regression modeling approach was used to account for the inherent ordering in levels of decision-making autonomy [40]. An ordinal logistic regression model (Equation 1) was estimated to assess the association between GSE and autonomous decision-making:

Where g is equal to the number of autonomous decisions per decision-making index (1, 2, or 3) (See Supplementary Table A for the items included in each index) and where γ₁ to γ_k represent control variables capturing individual and household characteristics that may have influenced self-efficacy and decision-making autonomy (i.e., intervention arm, religion, years since marriage, age at baseline, women’s highest class passed, family type, and wealth decile). Wealth was calculated based on household ownership of selected assets using standard DHS techniques for the FAARM sample as a whole [39, 41]. We estimated Equation 1 for sole and joint decision-makers together, as well as for each decision-making type separately.

Another ordinal logistic regression equation (Equation 2) was then estimated to assess the association between SSE and decision-making autonomy in joint decision-makers. These analyses were conducted only for women who were joint decision-makers across all three activities in a given category. For example, if a woman reported joint decision-making in one of the questions comprising a category, but sole decision-making in the other two questions, she would not be included in the analysis for that category, due to the fact that SSE data was only available for joint decision-makers. The results for Equation 2 are also adjusted for the above-mentioned individual and household characteristics.

No violations of the parallel regression assumption were noted for either model and analysis continued using proportional odds (cumulative logit) models [42]. All models also included robust standard errors to adjust for settlement-level clustering. All quantitative analyses were conducted using Stata v16.1 (StataCorp LP, College Station, TX, USA).

Qualitative data collection

Detailed methods for the qualitative sampling and data collection are provided elsewhere [43]; information relevant to the current study is summarized below. In-depth interviews (IDIs) were conducted in Bengali with 12 women from 6 intervention settlements and 10 women from 5 control settlements in June and July, 2019. These settlements were randomly selected from a list of settlements where previous FAARM qualitative data collection had not taken place. Women from these settlements who had completed the quantitative pro-WEAI survey were then randomly selected and invited to participate in an IDI. Female data collectors, who were trained and experienced in qualitative research, conducted each interview in a private space either inside or directly adjacent to the interviewee’s home. The IDI question guide was designed to capture the influence of the HFP intervention on women’s individual and household experiences and changes in empowerment, including self-efficacy. The guides were piloted in an intervention settlement prior to roll-out. These guides were not explicitly designed to answer the research question that our study poses, but provide useful insights to contextualize the quantitative findings.

Qualitative data analysis

The IDI audio recordings were transcribed verbatim in Bengali and the transcripts were then translated into English. A codebook was developed deductively using key concepts of interest from the quantitative analysis. The codebook consisted of several codes that were each assigned to one of four domains: self-efficacy, decision-making, agency (other than decision-making), and sociocultural context. The transcripts used in this analysis were previously coded as part of a grounded theory analysis that sought to describe the pathway to women’s empowerment through the HFP intervention [43]. Codes used in the previous analysis, which were developed inductively, were leveraged in our analysis where relevant. All coding was conducted in MAXQDA2020 (VERBI Software, 2019). Coded text segments were then reviewed to identify patterns in the ways in which the 22 women in the sample experienced, perceived, and spoke of the four key domains. This thematic analysis yielded in-depth descriptions of each of the four key domains; the themes most relevant for answering the research questions posed above are discussed in the qualitative results section below [44].