GiveDirectly MobileAid Phone-Data (CDR) Targeting Pilot, Cox's Bazar, Bangladesh

Overview

GiveDirectly, in partnership with the Government of Bangladesh's Aspire to Innovate (a2i) programme and researchers at the University of California Berkeley, piloted a machine-learning-based approach to poverty targeting in three sub-districts of Cox's Bazar in southern Bangladesh: Ramu, Teknaf, and Ukhia. The programme, known as MobileAid, sought to determine whether anonymised mobile phone call detail records (CDRs) could be used to remotely identify the poorest households faster and more cheaply than traditional targeting methods such as proxy means testing (PMT) and community-based targeting (CBT), in order to select beneficiaries for unconditional cash transfers delivered via mobile money.

The pilot was conducted in 2023 as part of a broader research study documented in a 2025 Cowles Foundation Discussion Paper (No. 2443) authored by Aiken, Ashraf, Blumenstock, Guiteras, and Mobarak. The researchers first conducted a census of approximately 106,000 households across 201 randomly chosen villages in the three sub-districts, collecting phone numbers and basic household information. A representative random sample of 5,006 households from 180 neighbourhoods was then surveyed in March 2023 using the standardised consumption module from the 2016 Bangladesh Household Income and Expenditures Survey (HIES), generating detailed per capita consumption expenditure data as the ground-truth measure of poverty. Community-based targeting exercises were also conducted in November 2023 in each of the 180 neighbourhoods, assembling 12 to 25 community members to collectively identify the poorest 20 per cent of households.

For the phone-based targeting arm, complete mobile phone metadata was obtained from all four mobile network operators active in Cox's Bazar (Grameenphone, Robi, Banglalink, and Teletalk) for all consenting survey households, covering the period from 1 March to 31 July 2023. The CDR data included pseudonymised identifiers for caller and recipient, date, time, and duration of calls, GPS coordinates of cell towers used, and information on daily mobile data usage. From these records, the researchers calculated 1,578 engineered features for each pseudonymised phone number, including statistics on call and text frequency, heterogeneity in contact networks, recharge patterns (indicating how much money subscribers add to their prepaid SIM cards), mobility traces based on cell tower usage, and mobile data consumption patterns. These features were matched to the household survey data for the 94 per cent of households that had at least one phone number present in the CDR records, and a gradient boosting model (an ensemble of decision trees with explicit regularisation) was trained to predict log per capita consumption expenditure.

The programme received IRB approval from the University of California Berkeley Committee for the Protection of Human Subjects (Protocol No. 2023-02-16103) and the Institutional Review Board of the Institute of Health Economics at the University of Dhaka. Informed consent was obtained from survey participants before accessing their CDR data. To minimise the risk of data misuse, the research team provided telecom operator staff with a set of phone numbers from consenting households along with code to extract the 1,578 CDR features. The telecom staff performed the feature extraction internally and merged the results with a redacted version of the household surveys. The resulting anonymised dataset was securely stored on an isolated server on the premises of a2i, an entity of the Government of Bangladesh. This multi-step protocol ensured that the research team, GiveDirectly, and the Bangladesh government never accessed CDR data with personal identifiable information, and the telecom operators never accessed unencrypted household survey data.

The study compared the accuracy of phone-based targeting (PBT), PMT, and CBT using three metrics: Spearman rank correlation with consumption, precision and recall for identifying the poorest 21 per cent of households, and area under the ROC curve (AUC). PMT achieved the highest accuracy (AUC of 0.82, precision/recall of 52 per cent, Spearman correlation of 0.65). Phone-based targeting was the second most accurate method (AUC of 0.61, precision/recall of 32 per cent, Spearman of 0.23), outperforming CBT (AUC of 0.58, precision/recall of 26 per cent, Spearman of 0.15). All differences between the three methods were statistically significant at p less than 0.001 using a Wilcoxon signed-rank test.

However, the study's central contribution was demonstrating that the welfare-maximising targeting approach depends on programme scale and budget. Using a social welfare framework adapted from Hanna and Olken (2018), the authors showed that for programmes with a relatively small budget screening a large number of households for eligibility, phone-based targeting is the most cost-effective method because its marginal cost per additional household screened is near zero once CDR infrastructure is established, whereas PMT requires expensive in-person surveys. For programmes with larger budgets relative to the number of households screened, PMT is more efficient due to its superior accuracy. CBT was dominated by both alternatives in terms of cost and accuracy across all programme scales.

GiveDirectly partnered with a2i to deploy the phone-based targeting results, selecting 22,000 households from the 106,000 listed in the census to receive unconditional cash transfers via mobile money. Each household received BDT 15,000 (approximately 136 USD) paid in two equal instalments between November 2023 and March 2024 via bKash mobile money, as documented in the GiveDirectly implementer note. Households selected through CBT received smaller transfers of 1,100 BDT (35 USD PPP). Selected households were contacted through Bangladesh's national 333 helpline, where trained agents verified identities and collected additional socioeconomic information. A staffed call centre supported recipients throughout the disbursement process. The academic paper (Aiken et al. 2025) reports a higher figure of BDT 30,000, which may reflect a different programme phase, total household allocation across multiple rounds, or a drafting error in the working paper. The GiveDirectly implementer note is treated as authoritative for the actual disbursement amount.

The project was funded by USAID and Google.org. The broader MobileAid initiative has been piloted across four countries since 2021 (Togo, DRC, Malawi, and Bangladesh), with the Bangladesh pilot representing the most rigorous head-to-head comparison of algorithmic versus traditional targeting methods to date. The programme aligns with the Government of Bangladesh's Smart Bangladesh Vision for leveraging technology in governance and social protection delivery.

Classification

Programme Details

Pilot programme using machine learning trained on mobile phone call detail records (CDRs) from all four Bangladeshi MNOs to predict household poverty and select beneficiaries for unconditional cash transfers delivered via mobile money (bKash) in three sub-districts of Cox's Bazar hosting Rohingya refugee communities.

Implementation Details

Risk & Oversight

Deployment & Outcomes

Sources

1. SRC-001-BGD-004 Aiken, E., Ashraf, A., Blumenstock, J., Guiteras, R., & Mobarak, A. M. (2025). Scalable Targeting of Social Protection: When Do Algorithms Out-Perform Surveys and Community Knowledge? Cowles Foundation Discussion Paper No. 2443. New Haven: Yale University.
   https://cowles.yale.edu/sites/default/files/2025-06/d2443.pdf
2. SRC-003-BGD-004 GiveDirectly (2025). In Bangladesh, AI targeted aid faster, cheaper, and often more accurately than manual methods, 2 July 2025. New York: GiveDirectly. Available at: https://www.givedirectly.org/ai-targeting-bangladesh/ (Accessed 24 March 2026).
   https://www.givedirectly.org/ai-targeting-bangladesh/
3. SRC-004-BGD-004 GiveDirectly (2025). Five things we've learned using mobile data and AI/ML to identify people in need (MobileAid). New York: GiveDirectly. Available at: https://www.givedirectly.org/mobileaid/ (Accessed 24 March 2026).
   https://www.givedirectly.org/mobileaid/
4. SRC-002-BGD-004 The Business Standard (2024). a2i & GiveDirectly joined forces to combat poverty by the usage of AI. The Business Standard, 9 March 2024. Available at: https://www.tbsnews.net/economy/corporates/a2i-givedirectly-joined-forces-combat-poverty-usage-ai-805994 (Accessed 31 October 2025).
   https://www.tbsnews.net/economy/corporates/a2i-givedirectly-joined-forces-combat-poverty-usage-ai-805994