Table 2.

Sources of exclusion from rural Novissi benefits

Exclusion source	Proportion included	Data and calculations
Voter ID possession	83–98%	According to administrative data, 3,633,898 individuals are registered to vote in Togo. The electoral commission of Togo reports that this corresponds to 86.6% of eligible adults44. The total adult population in Togo is not certain (the last census was in 2011), but Togo’s national statistical agency (https://inseed.tg/) estimates that there are 3,715,318 adults in Togo; the United Nations estimates 4.4 million adults45. These imply a voter ID penetration rate of either 82.6% or 97.8%, respectively.
SIM card and mobile phone access	50–85%	65% of individuals interviewed in the 2018–2019 field survey (n = 6,171) reported owning a phone; 85% of individuals were in a household with one or more phones. Rural penetration is lower (50% of individuals and 77% of households), as is penetration among women (53% for women vs 79% for men; in rural areas, it is 33% for women and 71% for men) (Supplementary Fig. 3). Phone penetration in Togo probably increased between the field survey (2018–2019) and the Novissi expansion (October 2020); the Togolese government estimates 82% SIM card penetration44.
Past mobile phone use	72–97%	Poverty estimates were constructed only for subscribers who placed at least one outgoing transaction between March and September 2020. In a typical month, 2.5% of all phone numbers are newly registered (Supplementary Fig. 6), so with a one-month gap between poverty inference and programme registration we would expect 95–97% of registrations to be associated with a poverty score. However, 27% of all Novissi registrations (November–December 2020) did not match to CDR, probably owing to new SIM purchases or registration on infrequently used SIMs (Methods, ‘Programme exclusions’).
Programme awareness	35–46%	245,454 unique subscribers attempted to register for the rural Novissi programme. The total voting population of eligible areas is 528,562, implying a maximum registration rate of 46.44%. However, not all 245,454 registration attempts were made by people living in eligible areas; examining administrative data on home location from successful registrations, we estimate that 87% of registration attempts came from eligible areas, implying an attempted registration rate of 40.40%. An alternative way to estimate attempted registration rates involves comparing the number of registration attempts made by phones below the poverty threshold (69,753) with our estimate of the number of voters in eligible cantons below the poverty threshold based on inferred home locations from mobile phone data (174,425; see Supplementary Methods section 4 for details), which implies an attempted registration rate of 34.79% after scaling by 87% (to account for registrations that came from outside of eligible areas).
Registration challenges	72%	Registration for the Novissi programme requires entering basic information into a USSD (phone-based) platform. According to programme administrative data, of the 245,454 subscribers who attempted registration, 176,517 (71.95%) eventually succeeded. The average registration required four attempts.
Targeting errors	47%	Based on the estimates from our targeting simulations using the 2020 phone survey (Table 1), the exclusion error rate of the phone-based targeting algorithm is 53%.

We use multiple sources of administrative data, survey data and government sources to estimate the extent to which different elements of the design of the Novissi programme may have led to errors of exclusion. Eligibility requirements for Novissi included: a valid voter ID (as a unique identifier and for home location), access to a mobile phone (to fill the register using the unstructured supplementary service data (USSD) platform), past mobile network transactions (to estimate poverty from mobile network behaviour), programme awareness (to know that the programme exists and to attempt to register), ability to register via the USSD platform (which requires basic digital literacy), as well as targeting errors from the phone-based machine-learning algorithm. This table calculates sources of exclusion as though they were all independent; Extended Data Table 5 uses survey data to calculate overlaps in exclusions.