Odisha Ama Krushi Call-time Customization
IND -21 -1679Last modified on December 19th, 2025 at 10:41 am
-
Abstract
PxD operates Ama Krushi, a free agriculture information service delivered over mobile phones, in collaboration with the State Government of Odisha Department of Agriculture, using a two-way Interactive Voice Response (IVR) platform with “outbound” push calls and an “inbound” hotline service.
This study evaluated whether personalizing the timing of information delivery can increase user engagement with IVR content. Specifically, we tested a targeted delivery strategy where the timing of advisory calls was personalized based on a machine learning model trained on farmers’ past engagement patterns and characteristics; the targeted delivery was compared to a random delivery schedule. By combining predictive modeling with a randomized controlled trial, this study aimed to generate actionable insights on the role of service customization in improving farmer engagement with digital extension services. We find that the targeted timing strategy increased pick-up rates by 1.7 percentage points (pp) overall, with a larger 3.2 pp gain by low-engagement users. These results highlight the potential of data-driven customization to enhance farmer engagement with digital advisory services. -
Status
Completed
-
Start date
Q4 Oct 2021
-
Experiment Location
Odisha, India
-
Partner Organization
Golub Capital Social Impact Lab, Government of Odisha
-
Agricultural season
Kharif
-
Experiment type
A/B test
-
Sample frame / target population
Farmers receiving outbound calls from the Ama Krushi service (non-Odisha-RCT sample)
-
Sample size
1,300,000
-
Outcome type
Platform engagement, Service engagement
-
Mode of data collection
PxD administrative data
-
Research question(s)
How does service engagement compare for a targeted call-time policy (targeted at the individual using a predictive model of past behavior and individual characteristics) versus a random call-time policy (scheduled randomly)?
-
Research theme
Communication technology, Message timing and frequency
-
Research Design
This project ran over a period of six weeks, from October 5 to November 23, 2021. The first three weeks were pilot weeks, when we observed farmers’ responses to randomly scheduled IVR calls. We trained a machine learning model to predict the best timing of advisory calls for individual farmers based on their engagement patterns and characteristics. The last three weeks of the study were evaluation weeks, when we evaluated a targeted call-time policy.
The sample size for the intervention was about 1.3 million farmers, but we restricted the analysis to a subsample of about 900,000 farmers whose data we had for all the covariates.
During the evaluation weeks, we separated the Ama Krushi service operation time into 91 time bins: 7 days (Monday to Sunday) X 13 hour-long blocks per day (8AM–9PM). We randomized farmers at the individual level. Control group farmers received IVR calls at a random time—a randomly chosen specific minute in the 8AM to 9PM period for 7 days per week. Treatment group farmers received the IVR calls during their predicted optimal time bin. The specific minute within the bin was randomly chosen. As there were technical constraints on the number of calls that could be sent per hour, some treatment farmers were assigned to their second-best time bin.
There were three call attempts per message when farmers did not pick up the first attempts. Re-tries happened 24 hours after the previous attempt. Messages were prioritized in the following order: paddy advisories, non-paddy crop advisories, and livestock advisories. We use administrative data to measure pick-up rates as a binary outcome of interest.
For further information see Athey et al. (2024)
-
Results
We see a modest 1.7 pp gain in pick-up rate when the targeted policy was compared to the random policy. We also conducted a heterogeneity analysis by engagement group and find that low-engagement users (i.e., farmers whose pick-up rate in the pilot data was lower than the median pick-up rate) had a 3.2 pp gain in pick-up rate in the targeted-policy group compared to the random-policy group.