Figures & Tables
Figures
- FIGURE 1
Poshan Didi Chatbot Development and Testing Process Steps
- FIGURE 2
Poshan Didi Implementation Timeline
- FIGURE 3
Theory of Change for Poshan Didi Chatbot to Improve Child Nutrition
Abbreviations: BCC, behavior change communication; CAS, Common Application System; HCW, health care worker.
- FIGURE 4
User Interface of Poshan Didi Chatbot
- FIGURE 5
User Enrollment for Poshan Didi Pilot Phase 2 Implementation
Tables
Pilot Phase Users, No. Sample Bias Engagement 1 10 Users are manually chosen Medium: WhatsApp.
The intervention interaction is manual (users engage with Poshan Didi, but a researcher reads and responds manually to user questions).
Evaluation is purely qualitative.
2 100 Users are not representative of the larger population Medium: Telegram.
Intervention is increasingly automated.
Results are increasingly quantitative, though not necessarily widely generalized.
Questions Hypothesis Expectations From Learnings How will users receive the chatbot? The chatbot would be well received (i.e., users would find it useful to receive information through their phones via a persona-based, conversational interaction). Understand smartphone penetration and usage
Test feasibility of text-based messaging
Explore initial reactions from beneficiaries on counseling information sent via technology
Build content to include key topics of age-appropriate counseling
How will users respond to incoming messages? Users would respond when they were engaged and interested. How will the chatbot affect AWW services? Accountability of AWW services can be increased by asking users about recent AWW visits. Abbreviation: AWW, Anganwadi worker.
- TABLE 3.
Phase 2 of the Poshan Didi Chatbot Pilot Phase 2 Research Questions and Hypotheses
Questions Hypothesis Expectations From Learnings We found mothers used very formal and polite language in Phase 1, demonstrating a high level of respect; how does users' respect compare during the semiautomated hybrid version of Poshan Didi? We would continue to see high levels of respect from users. Evaluate response rates from beneficiaries based on content and message format
How will users respond to incoming messages? At least 30% or more of the users will engage with the automated menu-based interaction menu system.
At least 10% of users will use the nurse-escalation system to ask a free-form question. Messages that users receive but do not explicitly reply to could still be read by users and be valuable.
Understand the value of real-time responses to beneficiaries that can be achieved through automation
Learn more about key counseling topics/queries that are relevant to mothers
Responsive Feedback Analyzed Implementation Change Feedback Source Multiple queries on nutrition for lactating mothers. Added new module called “Nutrition for lactating mothers” to menu-based script. Messages escalated to nurse Seasonal changes during duration of deployment prompted a request for counseling information on commonly occurring childhood illnesses that cause setbacks in child nutrition. Added new module called “Commonly occurring childhood diseases” to menu-based script. Requested by local government partner Users were not familiar with menu-based script and time gap between registration and push message meant users did not engage with system. Built and introduced “echo mode” feature to increase familiarity with system and to practice typing numbers.a Chatbot logs and midline interviews Unintelligible queries (e.g., included random letters, incomplete words, and greetings) that did not necessitate a response. Built “no reply” command that nurses could use to skip responding. Messages escalated to nurse Escalated queries that could be answered by menu-based script content that was available. Built “state” command that nurses could use to redirect to menu-based script. Messages escalated to nurse User questions on capabilities of Poshan Didi and on existing content that was not yet shared. Provided global menu with a complete list of modules to beneficiaries; project staff received additional training when necessary. Midline interviews Phone ownership and usage did not always correspond to user competency in using phone keyboard (∼70 women owned a personal phone). Added keyboard training, including installation of local language for typing, to onboarding process. Midline interviews Users with children identified as severely malnourished did not access content on growth charts. Sent push messages with growth charts to specific users. Midline interviews ↵a Echo mode refers to a chatbot feature in which the bot validated the number entered by the user in the beginning for practice purposes. This had no impact on the actual conversation the participants had with the chatbot.
