1 of 73

Scalable Targeting of Social Protection:

Algorithms vs. Community Knowledge

Joshua Blumenstock (U.C. Berkeley)

Joint with: Emily Aiken (UCSD), Anik Ashraf (LMU Munich), Raymond Guiteras (NC State), Mushfiq Mobarak (Yale)

RA’s: Leo Selker (Berkeley) and Soumi Chandra (NC State)

1

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

2 of 73

Motivation

Social protection programs: Currently cover 52.4% of world’s population

Over $1 trillion spent annually; 13% of GDP (on average, excluding health costs)

The challenge of targeting: Who should be eligible for benefits?

Especially difficult in LMICs, where up-to-date data on poverty is rare (Jerven 2023)
In one review, 1 out of 4 anti-poverty programs were regressive (Coady et al. 2004)

2

0-3 YEARS

3-6 YEARS

6-9 YEARS

9-26 YEARS

NEVER

Avg. interval between economic surveys, 1993 - 2021

Burke et al. 2021

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

3 of 73

Motivation: Targeting

Dominant paradigms for targeting social protections:

Proxy-means testing (PMT): Eligibility based on ~20 survey questions

Grosh and Baker (1995), Brown et al. (2018), Banerjee et al. (2020)

Community-based targeting (CBT): Community members identify needy households

Alatas et al. (2012), Beaman et al. (2022), Trachtmanet al. (2022), Sumarto et al. (2025)

COVID-19 precipitated shift to passively collected “big” data, including satellite imagery (Smythe and Blumenstock 2022) and mobile phone data (Aiken, Blumenstock et al. 2022)

3

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

4 of 73

This Paper

If and when does big data + machine learning work better than more traditional alternatives (PMT, community-based targeting)?

First set of results: Compares accuracy of different methods: PMT > Phone > CBT

Remarkably little heterogeneity in this pattern

Second set of results: Considers welfare gains, accounting for targeting costs

PMT is most accurate, but also most expensive: what maximizes welfare?

Main takeaway: “best” approach depends on scale and scope of program

Algorithmic targeting works best when total budget is small relative to the number of households that must be screened

4

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

5 of 73

Outline of Talk

Motivation and Background
Research setting, Data, and Methods
Results: Accuracy
Results: Welfare
Discussion and Conclusion

5

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

6 of 73

Research setting

Primary setting: Southern Bangladesh

GiveDirectly program to support “host communities”

5 monthly mobile money transfers, starting Jan. 2024

30,000 Taka total (USD $300 nominal; $955 PPP)
Targeted to 21% poorest in rural Cox’s Bazar
22,000 beneficiaries

Based on “phone-based” targeting: ML and phone data

Households without phones were (4%) ineligible

Smaller cash transfer of 1,100 Taka ($35 USD PPP) allocated via Community based targeting (CBT)

6

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

7 of 73

Research setting

Secondary setting: Rural Togo

(Mostly used to assess external validity)

COVID-19 relief program
3-5 monthly mobile money transfers, 2020-2021

USD $20/month
Targeted to 29% poorest in rural Togo
140,000 beneficiaries

Details: Aiken, Bellue, Karlan, Udry, & Blumenstocket al. (2022). Machine Learning and Mobile Phone Data Can Improve Targeting of Humanitarian Aid. Nature 7(902)

7

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

8 of 73

Data: Bangladesh

Household Census: 106,000 HH in 201 randomly selected villages (of 300)

Household characteristics, asset ownership (to compute PPI), phone numbers

Household consumption survey: 5006 hh randomly selected from census

Consumption, expenditures, demographics, assets, durables peer rankings, consent
Peer rankings: each household ranks 8 randomly selected households in their neighborhood, and states how “well off” the household is (1-5) (Details ⮱)

Mobile phone data (call detail records)

Complete mobile phone metadata from all consenting survey respondents from all 4 mobile network operators active in study region in 2023 (Details ⮱)

Community-based targeting in 180 neighborhoods

Following BRAC’s protocol: 12-25 community members (all types) collectively identify the 20% poorest-ranked who would receive a one-time cash transfer (Details ⮱)

8

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

9 of 73

Data: Togo

Household Census: 106,000 households in 200 randomly selected villages

Household characteristics, asset ownership (to compute PPI), phone numbers

Household consumption survey: 6171 hh randomly selected from country

Consumption expenditures, demographics, assets, peer rankings, consent
Peer rankings: each household ranks 8 randomly selected households in their neighborhood, and states how “well off” the household is (1-5)

Mobile phone data (call detail records)

Complete mobile phone metadata from all consenting survey respondents from two mobile network operators active in Togo

Community-based targeting in 180 neighborhoods

Assembled 12-25 community members from all walks of life to collectively identify the 20% poorest-ranked who would receive a one-time cash transfer

9

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

10 of 73

Methods: Targeting Method Comparison

Our first set of results compares accuracy of targeting methods at identifying consumption poor (PPP-adjusted PCE):

Proxy-means test (PMT) with basic machine learning

LASSO optimized out of sample (selected 45 covariates) [Details ⮱]
(Also tried step-wise forward selection, random forest)

Community-based targeting (CBT) in Bangladesh

Following BRAC’s CBT protocol

Phone-based targeting based on mobile phone metadata and ML

Details on next slides
HH without phones (4%) were ineligible for program; these count as targeting errors

Others: PPI scorecard (10 questions), geographic targeting, peer rankings

10

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

CBT in each of the 180 neighborhoods included in our study, following the protocol used by BRAC.
First, in neighborhoods of more than 100 households, enumerators split neighborhoods into contiguous segments of 50-100 households
Enumerators worked with senior community members to identify 12-25 households to join the meeting, inviting households from all walks of life and ensuring participation from women, students, farmers, businessmen, and laborers.
Each meeting began with a ``social mapping'' exercise in which a community map was drawn with each household identified by name and occupation.
Meeting attendees then worked together to rank the wealth of all households in the community by placing index cards representing each household on a string in the order of wealth.
To make the CBT exercises consequential, participants were informed at the start of the meeting that the 20\% poorest-ranked households would receive a one-time cash transfer of 1,000 Taka (\$31.88 USD PPP) following the meeting.
The normalized wealth ranking within each village is used to identify the poorest households for our community-based targeting method. The implicit assumption of this approach is that poverty distributions across villages are comparable.
Households that are not ranked in the community-based targeting approach (0.4\% of households) are considered to be targeted last for benefits by the CBT approach.

11 of 73

Phone-Based Targeting: Intuition

Idea: Wealthy people use phones differently than poor people

These differences can be used to predict which subscribers are wealthy and poor

11

Non-poor

Ultra-Poor

# of days with activity

Ultra-Poor

Non-poor

Aiken, Bedoya, Blumenstock, and Coville (2023 JDE)

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

12 of 73

Phone-Based Targeting: Methods

How does it work?

12

Ai image

Locations of

surveyed

households

Supervised Learning: Y_i = f (X_i)

In-person household survey
1.5 hrs, 100’s of questions
Includes consumption (Y_i), PMT, and wealth info

Training Data (“labels”)

Each respondent

Vector of CDR features (X_i)

Feature engineering

Call Detail Records (CDR)

from surveyed households,

obtained from operator

CDR metadata includes:

Billions of transactions
Millions of subscribers
Activity (calls, SMS, recharges, data use)
Locations (from towers)
Network structure (inferred)

Each mobile phone

[ Details⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

Also Neeti Pokhriyal has a fantastic paper in PNAS

The combination of high-resolution satellite imagery and machine learning have proven useful in many sustainability-related tasks, including poverty prediction, infrastructure measurement, and forest monitoring. However, the accuracy afforded by high-resolution imagery comes at a cost, as such imagery is extremely expensive to purchase at scale. This creates a substantial hurdle to the efficient scaling and widespread adoption of high-resolution-based approaches. To reduce acquisition costs while maintaining accuracy, we propose a reinforcement learning approach in which free low-resolution imagery is used to dynamically identify where to acquire costly high-resolution images, prior to performing a deep learning task on the high-resolution images. We apply this approach to the task of poverty prediction in Uganda, building on an earlier approach that used object detection to count objects and use these counts to predict poverty. Our approach exceeds previous performance benchmarks on this task while using 80% fewer high-resolution images. Our approach could have application in many sustainability domains that require high-resolution imagery.

13 of 73

Phone-Based Targeting: Evidence

13

Rwanda 2010 (N=856)

Afghanistan 2015 (N=1,234)

Togo 2018 (N=4,296)

Togo 2020 (N=15,044)

Details: Blumenstock, Cadamuro, On (2015). Predicting poverty and wealth from mobile phone metadata. Science

Blumenstock (2018). Estimating Economic Characteristics with Phone Data. AER: P&P

Aiken, Bellue, Karlan, Udry, & Blumenstock (2022). Machine Learning and Mobile Phone Data Can Improve Targeting of Humanitarian Aid. Nature

Aiken, Bedoya, Coville, & Blumenstock (2023). Targeting Development Aid with Machine Learning and Mobile Phone Data. Journal of Development Economics

Across countries, phone-based predictions capture 25-50% of variation

Lower R² for consumption, higher for wealth
Models do not generalize from one country to another [Details ⮱]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

14 of 73

Comparing targeting accuracy: Intuition

Our analysis compares the predictions from different targeting methods

Generally, we observe PMT > Phone-based > CBT

14

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

15 of 73

From predictions to targeting: Details

15

GiveDirectly’s objective: Target the poorest 21% of households

We compare targeting methods on

Correlation with true consumption (Spearman and Pearson)

Accuracy at targeting the poorest 21% (also exclusion error rate, inclusion error rate)

AUC (Area under curve): a method for evaluating a classifier that is agnostic to the inclusion threshold

Correct inclusions

Exclusion errors

Correct inclusions

Exclusion errors

Inclusion errors

Correct inclusions

Exclusion errors

Inclusion errors

Correct exclusions

Desired

threshold

Implemented

threshold

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

Error rate is the simplest metric, but considers only binary errors, not the magnitude of error, and depends on the specific threshold of a particular program.
The second metric is the Spearman rank correlation between the rank assigned to a household by a particular method and the household's true rank in the distribution of consumption per capita. This puts less weight on the exact classification of households near the cutoff, and penalizes large errors in ranking households. The third is the Area under the ROC (receiver operating characteristic) curve, or \emph{AUC score}, which considers the accuracy at each possible value of the share of households targeted, rather than just the specific share for a particular program (21\% in our case). The ROC curve plots the true positive rate (recall) as a function of the false positive rate for each threshold value, and the AUC score is then the area under this curve. As the targeting threshold increases, the true positive rate (recall) increases, with more of the poorest households eligible for benefits, but the false positive rate also increases as additional non-poor households become eligible. More accurate targeting methods will achieve a higher area under the curve, representing a slow increase in false positives relative to true positives as the targeting threshold is raised.

16 of 73

Outline of Talk

Motivation and Background
Research setting, Data, and Methods
Results: Accuracy
Results: Welfare
Discussion and Conclusion

16

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

17 of 73

Targeting accuracy: Summary (Bangladesh)

17

[ Details⮱ ]

Exclusion errors (Bangladesh): PMT < Phone-Based < Community

Details on other metrics ⮱

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

18 of 73

What type of HH does each method target?

CBT prioritizes widows/widowers - evidence of local/private info (Sumarto et al. 2025) [Details ⮱]

Phone-based targeting prioritizes socially isolated households (based on peer ranking data, not CDR)

What features are predictive? ⮱

CBT does not prioritize low consumption households

PMT prioritizes low consumption neighborhoods

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

N=1,252
The regressors were chosen by whatever demographic data we had available, and seemed potentially relevant.
�
Results of regressions for which types of households are selected by each targeting method. The dependent variable of each regression an indicator for whether a household was targeted by the method in question. Regressions are run jointly with all explanatory variables in the first column. All explanatory variables are standardized. Connectedness (under neighborhood characteristics) represents the average self-reported knowledge that households have of other households in their community, elicited during the peer rankings exercise in our household survey. Connectedness (in) under household characteristics represents the average knowledge that other households had of the household in question during the peer ranking exercise; connectedness (out) represents the average knowledge that the household in question had of other households in their community. Regressions are run using data from a single train-test split. Standard errors are in parentheses.
In the peer rankings module, each household was asked, for eight randomly selected households in their neighborhood, how well they know the household on a scale of 1-4. Connectedness at the neighborhood level is defined as the average knowledge ranking for all households in the neighborhood. A household's ``connectedness'' is defined as the average knowledge ranking others assign to that household.

19 of 73

Do some methods work better on subpopulations?

For all types of neighborhood, PMT > phone-based > CBT

19

Example 2:

The PMT performs better in poorer neighborhoods (green bar > red bar)

Example 1:

Phone-based performs better than CBT for neighborhoods of all sizes (each colored PBT bar is higher than the corresponding CBT bar)

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

20 of 73

Do some methods work better on subpopulations?

For almost all types of household, PMT > phone-based > CBT

20

One exception:

CBT may perform “better” than phone-based for large households

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

21 of 73

Combining multiple approaches

In principle, different approaches use complementary information

Adding phone or CBT predictors does not meaningfully improve PMT performance
Adding CBT predictors slightly improves phone-based targeting

21

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

22 of 73

Combining multiple approaches

In principle, different approaches use complementary information

If we combine rankings, rather than predictors, we still can’t beat a pure PMT

22

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

23 of 73

Other considerations

Community perceptions

Higher satisfaction with CBT; perceived as more fair

Targeting the poorest within a neighborhood

Nothing changes

How do we handle households with multiple phones?

Nothing changes

23

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

24 of 73

Comparison to Togo: similar, but different

In Togo, PMT > phone-based, but differences are not as stark

Togo: AUC is 18% higher
Bangladesh: AUC is 34% higher

24

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

25 of 73

How does our targeting compare to literature?

Targeting accuracy is low, but not unusually low (for both CBT and PMT)

Targeting accuracy increases as fraction targeted increases
In our setting (Cox’s bazaar), small geography -> less variation in poverty

In general, targeting is worse in more homogenous locations [details ⮱]

25

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

26 of 73

Outline of Talk

Motivation and Background
Research setting, Data, and Methods
Results: Accuracy
Results: Welfare
Discussion and Conclusion

26

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

27 of 73

Accuracy vs. cost-effectiveness

Summary thus far:

PMT is most accurate; CBT is least accurate

However, PMT/CBT costs are higher [details ⮱]

Next focus: Which approach is most “cost-effective”? (Hanna & Olken 2018)

Simple framework: Each dollar spent on targeting is not given to beneficiaries

With fixed budget, this induces a tradeoff between accuracy and cost

Requires that we articulate a policy objective

E.g., we might care more about excluding those far below the poverty line
We seek a policy that maximizes “welfare”, using CRRA utility [details ⮱]

27

	Fixed	Variable
Phone	$46,600	$0
PMT	$52,900	$1.25
CBT	$19,300	$2.33

Targeting costs

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

28 of 73

Results: cost-effectiveness and welfare

We analyze welfare gains relative to “perfect” targeting

We aggregate population utility from all transfers [details ⮱] and compare to “oracle” benchmark
Bangladesh: $5M budget; 100k HH; $1.25 PMT

Despite higher PMT screening costs, net welfare is still highest for PMT

Togo: $5M budget; 207k HH; $4.00 PMT

Phone-based preferred (higher PMT cost, more hh screened, more accurate phone-based)

28

Phone-based

	Bangladesh	Togo
Budget	$5 Million	$5 Million
HH screened	100,000	207,000
% targeted	Poorest 21%	Poorest 29%
Recall (PMT / Phone)	32% / 50%	50% / 63%
PMT cost per HH	$1.25	$4.00

[ Details⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

29 of 73

Generalization: When is phone-based > PMT?

Phone-based targeting is more effective when the total program budget is small relative to the number of households that must be screened

29

Welfare ratio

(Phone/PMT)

Phones better

PMT better

GiveDirectly Bangladesh: $5M budget, 100k beneficiaries, $1.48 PMT cost

Novissi Togo: $5M budget, 200k beneficiaries, $4.00 PMT cost

Decision threshold

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

For example, government cash transfer programs in Bangladesh typically have budgets of \$10-300 million, and a mandate to screen all 41 million households in the country for eligibility.

Visual comparison of the performance of phone-based targeting and PMTs for a range of hypothetical programs, varying both the total program budget (horizontally) and the number of households screened (vertically).

Figure fixes the accuracy of the targeting technology, varies budget and # of households screened

When transfer programs have a large budget relative to the size of the population screened, like GiveDirectly's programs in Bangladesh and Togo, then PMT-based targeting results in larger welfare gains. However, for programs with small budgets relative to the number of households screened, which characterizes many real-world government-run social protection programs in Bangladesh, phone-based targeting is preferred. This is mainly because the marginal cost of screening additional beneficiaries using mobile phone meta-data -- once an algorithm is already developed -- is essentially zero.

30 of 73

Global perspective

Using the World Bank Aspire Database: PMT preferred in 66 countries; phone-based preferred in 10 countries ; 19 others depend on PMT costs

30

Phone-based preferred (N=10)

PMT preferred (N=66)

Depends on cost of PMT (N=19)

[ Show Details⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

31 of 73

Key factor: Budget per household screened

In Bangladesh: Phone-based targeting preferred if program budget is less than $4 (or $15) per household screened, given a $1.25 (or $4) PMT

31

Our PMT ($1.25)

Typical PMT ($4.00)

Phone-based

CBT

GiveDirectly program

$4/household

$15/household

[ Show Togo⮱ ]

[ Show Details⮱ ]

Govt. of Bangladesh S.A. budget

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

Thresholds at which different targeting methods provide the largest increase in utility, as a function of the program budget per household screened
In Bangladesh (Panel A), phone-based targeting (red line) is preferred to the PMT for programs with budgets under \$4 per household screened if the PMT costs \$1.25 (solid green line);
However, if the PMT costs the industry-standard $4.00 (dashed green line), phone-based targeting is preferred for budgets up to \$15 per household screened.

In constructing Figure 7, we ignore fixed costs. For medium- to large-scale programs,these will be a tiny fraction of total costs; for example, fixed costs make up 4% of screening costs for a PMT-targeted program screening 1 million households, but only 0.4% of screeningcosts for a PMT-targeted program screening 10 million households

32 of 73

Robustness / Sensitivity

Other factors that affect targeting method selection

Accuracy of phone-based targeting (relative to PMT/CBT)

BD accuracy: $15 threshold; TG accuracy: $40 threshold

PMT costs

As PMT costs increase, relative benefit decreases

CRRA risk aversion

Less curvature: accuracy matters less, phone-based preferred

Share of population targeted

Results qualitatively similar

32

Sensitivity to accuracy

[ Show Details⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

33 of 73

Discussion / Limitations

We focus only on financial costs

PMT and CBT impose time cost to respondents and community
Suggestive evidence of a (strong) community preference for CBT [Details ⮱]

One-shot targeting

Real-world programs do not need to screen every person every year
Less frequent updates: ↓ costs, but also ↓ accuracy

We do not capture dynamics of poverty

In principle, phone-based and CBT might better capture this? [Details ⮱]

Data privacy considerations [Details ⮱]
Politics of implementation [Details ⮱]

33

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

CBT in each of the 180 neighborhoods included in our study, following the protocol used by BRAC.
First, in neighborhoods of more than 100 households, enumerators split neighborhoods into contiguous segments of 50-100 households
Enumerators worked with senior community members to identify 12-25 households to join the meeting, inviting households from all walks of life and ensuring participation from women, students, farmers, businessmen, and laborers.
Each meeting began with a ``social mapping'' exercise in which a community map was drawn with each household identified by name and occupation.
Meeting attendees then worked together to rank the wealth of all households in the community by placing index cards representing each household on a string in the order of wealth.
To make the CBT exercises consequential, participants were informed at the start of the meeting that the 20\% poorest-ranked households would receive a one-time cash transfer of 1,000 Taka (\$31.88 USD PPP) following the meeting.
The normalized wealth ranking within each village is used to identify the poorest households for our community-based targeting method. The implicit assumption of this approach is that poverty distributions across villages are comparable.
Households that are not ranked in the community-based targeting approach (0.4\% of households) are considered to be targeted last for benefits by the CBT approach.

34 of 73

Summary / Conclusion

Accuracy of CBT < phone targeting << PMT

Consistent with past literature showing that CBT < PMT

Alatas et al. (2012), Premand and Schnitzer (2021), Trachtman et al. (2022)

“Best” approach depend on program parameters

Phone targeting best with small budgets over large populations

Thanks!

jblumenstock@berkeley.edu

34

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

35 of 73

BACKUP SLIDES

35

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

36 of 73

Summary statistics: Household Survey

36

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

37 of 73

CBT Details: Based on BRAC

Community selection: drawn from village census

Separately conducted in each of 180 neighborhoods
Neighborhoods with > 100 households split into contiguous segments of 50-100 households

Recruitment: Senior community members identify 12-25 households to join meeting

Households from all walks of life, ensuring participation from women, students, farmers, businessmen, and laborers

Social Mapping: Community uses map to identify each household by name and occupation
Ranking: Attendees worked together to rank the wealth of all households in community

By placing index cards representing each household on a string in the order of wealth
Households that are not ranked in the community-based targeting approach (0.4\% of households) are considered to be targeted last for benefits by the CBT approach.

Stakes: Participants informed at the start of the meeting that the 20% poorest-ranked households would receive a one-time cash transfer of 1,000 Taka ($32 USD PPP) following the meeting

37

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

38 of 73

Feature Engineering: Overview

How to convert raw data into “features”?

Key point: We need to do this algorithmically, not intuitively

38

Number of calls per day

Total expenditures on airtime

Average duration of weekend calls

Number of international contacts

Entropy of SMS volume

Unique cell towers visited

Caller-ID	Receiver-ID	Date-Time	Caller-Tower	…
0mDzqJpd	DO3L2Bdg	2015-11-01 11:27:25	4122010103	…

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

39 of 73

Feature Engineering: Details

Deterministic Finite Automata (DFA)

Advantages: Intuitive, creates “interpretable” features

39

In-sample

Feature	R²
Weighted average of all first-degree neighbor’s “Day of week (DoW) entropy” of outgoing SMS volume	0.376
Avg. of alter’s “Hour of Day (HoD) entropy”, outgoing SMS	0.371
Avg. of alter’s DoW entropy for incoming SMS	0.364
Avg. of alter’s HoD entropy for incoming SMS	0.356
Median of alter’s incoming call length, weekends only	0.354
Median of alter’s incoming call length, calls over 60s	0.351
Median of alter’s incoming call length, weekend evenings	0.338

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

40 of 73

Phone-Based Targeting: Generalization

Note: A model trained in one country does not work in another

40

Blumenstock (2018 AER: P&P)

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

41 of 73

Targeting accuracy: Additional metrics

Accuracy results are not sensitive to how accuracy is evaluated

If anything, CBT is slightly more likely to miss extreme poor [Details ⮱]

41

[ Details⮱ ]

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

42 of 73

Targeting accuracy: Additional metrics

42

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

43 of 73

43

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

44 of 73

What features are predictive?

PMT: Large households, many children, geography, assets, disabilities

Phone-based targeting: Recharge activity, geography, data use (proxy for phone type?)

44

[ Back ⮱ ]

44

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

45 of 73

Sumarto et al. (2025)

45

[ Back ⮱ ]

What do communities target on in Indonesia?

Baseline characteristics from Alatas et al (2012)

Communities choose households:

That spend a greater share of their consumption on food
Who may be more vulnerable on metrics beyond consumption.

Widows
Experienced illness

No discrimination against outsiders

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

46 of 73

Consumption distribution of targeted HH

Which method is including the poorest households?

Broadly, PMT does better than phone-based, which does better than CBT
Exclusion errors are “worst” for CBT

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

47 of 73

Targeting within Neighborhoods

CBT could be better at targeting within neighborhoods, because communities were asked to rank poverty within their own neighborhood

If we redefine the objective as targeting the 21% poorest of each village, qualitative patterns are unchanged

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

48 of 73

How to deal with multi-phone households?

Performance of phone-based targeting is not sensitive to how we determine phone-based scores for households with multiple phones

Our main results use the phone of the most senior household member

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

49 of 73

Targeting works better in heterogeneous areas

Performance is lower in simulations that restrict to poorest households

(x-axis varies the % of households included, selecting poorest first)

49

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

50 of 73

50

Targeting accuracy by geographic extent

[ Back ⮱ ]

50

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

51 of 73

Cost data

PMT

$53,517 fixed cost (consumption survey of 5,000 households)
$1.25 variable cost per household screened -- very low!
Census took ~15 minutes (similar to a typical PMT scorecard)

Phone-based targeting

$44,528 fixed cost (consumption survey of 5,000 households)

(Assume no costs to developing ML algorithm)

Marginal cost per household approximately zero

CBT exercises

CBT: $19,335 fixed cost (training), $2.32 variable cost per household screened
(Assume no costs to community of participating in meetings)

51

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

52 of 73

Cost data: Comparison to literature

Our PMT costs ($1.25) are much lower than “industry standard”

52

[ Back ⮱ ]

Source: Aiken, Bedoya, Blumenstock, Coville (2023 JDE)

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

53 of 73

Cost data: Comparison to literature

Our CBT costs ($2.30) were similar to literature

53

[ Back ⮱ ]

Source: Aiken, Bedoya, Blumenstock, Coville (2023 JDE)

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

54 of 73

How to maximize “social welfare”?

54

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

55 of 73

Utility calculation

55

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

56 of 73

Welfare for GD program - more options

56

[ Back ⮱ ]

56

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

57 of 73

Welfare gains from different approaches

57

[ Back ⮱ ]

Utility gain (relative to pre-program utility)

Our PMT ($1.25)

Typical PMT ($4.00)

Phone-based

Oracle

Random

CBT

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

58 of 73

Togo: Budget per household screened

In Togo: Phone-based targeting preferred if program budget is less than $31 (or $51) per household screened, given a $1.25 (or $4) PMT

58

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

Thresholds at which different targeting methods provide the largest increase in utility, as a function of the program budget per household screened
In Bangladesh (Panel A), phone-based targeting (red line) is preferred to the PMT for programs with budgets under \$4 per household screened if the PMT costs \$1.25 (solid green line);
However, if the PMT costs the industry-standard $4.00 (dashed green line), phone-based targeting is preferred for budgets up to \$15 per household screened.

In constructing Figure 7, we ignore fixed costs. For medium- to large-scale programs,these will be a tiny fraction of total costs; for example, fixed costs make up 4% of screeningcosts for a PMT-targeted program screening 1 million households, but only 0.4% of screeningcosts for a PMT-targeted program screening 10 million households

59 of 73

ASPIRE details

59

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

60 of 73

ASPIRE details

60

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

61 of 73

What if phone-based targeting were more accurate?

61

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

Thresholds at which different targeting methods provide the largest increase in utility, as a function of the program budget per household screened
In Bangladesh (Panel A), phone-based targeting (red line) is preferred to the PMT for programs with budgets under \$4 per household screened if the PMT costs \$1.25 (solid green line);
However, if the PMT costs the industry-standard $4.00 (dashed green line), phone-based targeting is preferred for budgets up to \$15 per household screened.

In constructing Figure 7, we ignore fixed costs. For medium- to large-scale programs,these will be a tiny fraction of total costs; for example, fixed costs make up 4% of screeningcosts for a PMT-targeted program screening 1 million households, but only 0.4% of screeningcosts for a PMT-targeted program screening 10 million households

62 of 73

What if CBT were more accurate?

62

[ Back ⮱ ]

CBT would have to be substantially more accurate to beat PMT

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

Thresholds at which different targeting methods provide the largest increase in utility, as a function of the program budget per household screened
In Bangladesh (Panel A), phone-based targeting (red line) is preferred to the PMT for programs with budgets under \$4 per household screened if the PMT costs \$1.25 (solid green line);
However, if the PMT costs the industry-standard $4.00 (dashed green line), phone-based targeting is preferred for budgets up to \$15 per household screened.

In constructing Figure 7, we ignore fixed costs. For medium- to large-scale programs,these will be a tiny fraction of total costs; for example, fixed costs make up 4% of screeningcosts for a PMT-targeted program screening 1 million households, but only 0.4% of screeningcosts for a PMT-targeted program screening 10 million households

63 of 73

Sensitivity

63

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

64 of 73

Community perceptions

64

[ Back ⮱ ]

Which targeting method do beneficiaries prefer?

CBT may increase satisfaction (Alatas et al. 2012)
But PMT perceived as more legitimate (Premand and Schnitzer 2022)

Our setting is not ideal for answering this question

We did not randomize assignment of targeting (unlike Alatas, Premand/Schnitzer)

CBT occurred in villages that had already received PBT

Recency: Survey was 2 months after CBT transfers; 10 months after phone-based
CBT transfer ($9) was much smaller than phone-based transfer ($300)
CBT process included daylong meetings with explanations of process

Qualitative evidence

Most people remembered the programs (90% CBT, 80% phone-based)
Very poor understanding of how programs worked – especially phone-based

Examples ⮱

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

65 of 73

Community perceptions

65

[ Back ⮱ ]

Caveats notwithstanding, we observe:

Significantly higher satisfaction with CBT (29 pp.)

“How satisfied were you with how the approach determined who was eligible to receive cash aid?”

CBT more likely to be perceived as fair (35 pp.)

“In your opinion, was the selection process to receive cash aid in the program fair?”

Much higher ratings from beneficiaries of their own program [details ⮱]

	Satisfied with process	Selection was fair
CBT (vs phone)	0.294*** (-0.019)	0.351*** (-0.021)
Participated in CBT meeting	0.019 (-0.079)	0.084* (-0.065)
Beneficiary of phone-based	0.339*** (-0.025)	0.296** (-0.025)
Beneficiary of CBT	0.174*** (0.022)	0.053** (-0.024)
Aware of phone-based	0.105* (-0.085)	0.246** (-0.096)
Aware of CBT	0.273 (-0.21)	0.365* (-0.19)
N	2,026	2,026

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

66 of 73

66

	Share satisfied or very satisfied			Share viewing process as fair
	(1)	(2)	(3)	(4)	(5)	(6)
	0.294***	0.294***	0.211***	0.351***	0.351***	0.256***
	-0.019	-0.019	-0.062	-0.021	-0.02	-0.074
CBT		0.170***	-0.017		0.049**	-0.023
		(0.021)	(0.027)		-0.023	-0.029
Beneficiary of CBT			0.376***			0.144***
			-0.038			-0.041
CBT * Beneficiary of CBT		0.335***	0.613***		0.292**	0.562***
		-0.022	-0.028		-0.023	-0.03
Beneficiary of Phone			-0.556***			-0.540***
			-0.04			-0.042
CBT * Beneficiary of Phone		0.056	0.023		0.056*	0.040
		-0.022	-0.028		-0.024	-0.031
Participated in CBT meeting			0.067*			0.032
			-0.04			-0.044
CBT * Participated in CBT meeting		0.173***	0.055		0.192*	-0.021
		-0.033	-0.043		-0.038	-0.05
Aware of CBT program			0.234***			0.426***
			-0.061			-0.07
CBT * Aware of CBT program		0.064*	0.155***		0.160**	0.271***
		-0.025	-0.033		-0.03	-0.039
Aware of Phone program			-0.181***			-0.221***
			-0.046			-0.056
CBT * Aware of phone program	2,026	2,026	2,026	2,026	2,026	2,026

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

67 of 73

Example Responses

67

[ Back ⮱ ]

#	How was eligibility determined for phone-based targeting?	…for community-based targeting?
1	I have no idea about this program. So I don't know how the eligibility criteria were determined.	Most of the rich got aid, so don't think it's appropriate.
2	I don't know how they decided because no one told me about it.	I know that the real poor families have been determined through meetings or getting together in the area.
3	Don't know how eligibility is determined under this program.	I heard they had organized a meeting but I am not familiar with the process they used for checking the eligibility for cash assistance.
4	I don't know anything about this programme. I don't know how the eligibility for the aid programme was determined.	She doesn't know the selection process, but they heard it helped 20% of poor people.
5	I don't know about this method, so I'm not sure how the selection was made.	I really liked how households were selected for cash assistance in this project because it was based on everyone's opinions and the households were surveyed accordingly.
6	I don't know how they determined it, so I have no idea about the process.	Through a lottery in a neighborhood meeting.
7	I do not know how eligibility selection program of providing aid has been set.	The poor were selected in a community gathering.
8	I don't know how eligibility for cash assistance was determined in this program; I've never heard of it.	It was decided through the meeting with everyone's opinion because I was at the meeting so I know.
9	I am not aware of this process. But I heard that they provided money through mobile.	I think that the rich and the poor in the area, all together chose who is the richest, who is the poorest, and the poorest were given the money.
10	I don't know how the eligibility was determined but many others and I received monetary aid through the process.	She heard that assistance will be provided to 20% of people living in poverty, but no further details have been shared.

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

68 of 73

Decay of PMT accuracy over time

As data go out of date, policies and interventions are less effective

Accuracy of poverty registries decreases by roughly 9 percentage points per year

Aiken, Ohlenberg, & Blumenstock (2024). “Moving targets: How often should a PMT be recalibrated?” Working Paper

Accuracy of PMT (R²)

Error rate of targeted policy

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

69 of 73

Data privacy and ethics

Data privacy

Informed consent from each respondent before requesting CDR from operator

Data minimization

Protocols to ensure that each actor (researcher, GiveDirectly, government, MNO) had access to as little data as possible

Only the phone company had access to transaction-level CDR
The phone company did not have access to survey data

Data security

All analysis was performed on a pseudonymized version of the dataset that was housed on a secure server at a2i

69

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

70 of 73

Implementing CDR at Scale

Key actors:

Mobile Network Operators (CDR Data and Implementation)
Economists (“Who should be Targeted?” | Design of CCT | Survey Data)
Machine Learning & CS Modelers (“How do we identify beneficiaries?”)
Central Government (Regulatory Permission for Data Sharing)
Mobile money or Local Government (“How do we distribute funds?”)

Sensitivities

Each operator’s competitive positioning in the MNO market
Politics – Politicians would prefer to create their own list and distribute
Civil Society – Are women, vulnerable groups protected?
Citizens – Can we really trust this system?

70

[ Back ⮱ ]

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

71 of 73

Dumpster slides

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

72 of 73

Data: Bangladesh (1)

Household Census: 106,000 households in 201 villages

Household characteristics, asset ownership (to compute PPI), phone numbers
98% of households had a phone; other 2% excluded from GiveDirectly’s program

Household consumption survey: 5006 randomly selected households

Consumption expenditures, demographics, assets (PMT), peer rankings, consent
PPP-adjusted PCE: Our “ground truth” measure of poverty
Proxy Means Test (PMT): Learning from household survey

LASSO optimized out of sample (selected 45 covariates)
(Also tried step-wise forward selection, random forest)

Peer rankings: each household ranks 8 randomly selected households in their neighborhood, and states how “well off” the household is (1-5)

72

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

73 of 73

Data: Bangladesh (2)

Mobile phone data (call detail records)

Complete mobile phone metadata from all consenting survey respondents from all 4 mobile network operators active in Cox’s bazar in 2023

Community-based targeting in 180 neighborhoods

Assembled 12-25 community members from all walks of life
Collectively identified the poorest 20% for a 1100 Taka transfer
Used BRAC’s standard CBT protocol for social safety net programs

73

Aiken, Ashraf, Blumenstock , Gutieras, Mobarak

CBT in each of the 180 neighborhoods included in our study, following the protocol used by BRAC.
First, in neighborhoods of more than 100 households, enumerators split neighborhoods into contiguous segments of 50-100 households
Enumerators worked with senior community members to identify 12-25 households to join the meeting, inviting households from all walks of life and ensuring participation from women, students, farmers, businessmen, and laborers.
Each meeting began with a ``social mapping'' exercise in which a community map was drawn with each household identified by name and occupation.
Meeting attendees then worked together to rank the wealth of all households in the community by placing index cards representing each household on a string in the order of wealth.
To make the CBT exercises consequential, participants were informed at the start of the meeting that the 20\% poorest-ranked households would receive a one-time cash transfer of 1,000 Taka (\$31.88 USD PPP) following the meeting.
The normalized wealth ranking within each village is used to identify the poorest households for our community-based targeting method. The implicit assumption of this approach is that poverty distributions across villages are comparable.
Households that are not ranked in the community-based targeting approach (0.4\% of households) are considered to be targeted last for benefits by the CBT approach.