Estimating Food Consumption and Poverty Indices with Mobile Phone Data

Introduction

• Mobile phone data are proposed as timely alternatives or complements to costly, infrequent surveys for development and humanitarian decision-making.Relevant signals include remote sensing, social media, remittances, and anonymized mobile phone records.Source: Introduction · p. 1
• Prior studies linked mobile phone activity to mobility, epidemics, social communities, and socioeconomic conditions, but often faced coarse or outdated poverty ground truth.The cited Côté d’Ivoire study compared phone features with poverty data that had coarse spatial resolution and were several years older.Source: Introduction · pp. 1–2
• Social diversity and mobility-based measures have been examined as proxies for poverty and socioeconomic status at population, area, and individual levels.Social diversity measures how equally communication time is shared across contacts, while other work used mobility and transaction histories.Source: Introduction · p. 2
• The study assesses CDR and airtime-credit metrics against contemporaneous, wide-scale household-survey data for food security and non-monetary poverty in a low-income country.The research addresses the limited prior evidence relating airtime purchases to countrywide food-security ground truth.Source: Introduction · p. 2

CDR data and Airtime credit purchases

• The mobile dataset contains seven months of CDR activity and six months of airtime-credit purchases from the same users and one large carrier.CDRs include caller, callee, initiating tower, date, and time; top-ups include user, amount, date, and time.Source: CDR data and Airtime credit purchases · p. 2

Food consumption survey

• The representative 2012 food-consumption survey covered 7,500 households and recorded detailed food-access, consumption, demographic, shelter, and income information.The questionnaire contained 486 questions, including weekly consumption frequencies and seasonal food-stock information.Source: Food consumption survey · p. 3
• Survey-derived food measures include the Food Consumption Score and Coping Strategy Index, capturing weighted food consumption and coping responses to restricted food access.The FCS classifies households as poor, borderline, or acceptably food secure, while the CSI measures coping-strategy severity and frequency.Source: Food consumption survey · p. 3

Poverty indices

• Poverty comparisons use area-level percentages poor, deprivation intensity, and a Multidimensional Poverty Index combining education, health, and living standards.The MPI is computed as the product of the poverty and deprivation-intensity variables.Source: Poverty indices · p. 3

Methodology

• Mobile-phone and survey measures are compared at the sector level, covering roughly 10,000–50,000 inhabitants depending on whether areas are rural or urban.The survey indicators are aggregated by taking household means within each sector.Source: Methodology · p. 3, Food consumption survey data processing · p. 4

Mobile phone data processing

• Users are assigned home locations using their most frequent calling tower between 6pm and 8am, which performed better than all-day assignment.The analysis was repeated with all-day calls, but night-time calls produced better correspondence with survey variables.Source: Mobile phone data processing · p. 3
• User features summarize top-up amounts and social diversity, while sector features include means, medians, standard deviations, and coefficients of variation.Top-up measures include average, six-month sum, minimum, and maximum values; social diversity follows.Source: Mobile phone data processing · p. 3

Food consumption survey data processing

• Survey variables cover household characteristics, food consumption, wealth, food expenditure, income, and total expenses, aggregated by sector.Food measures include access, consumption frequency, and composite indicators such as FCS.Source: Food consumption survey data processing · p. 4

Comparing food consumption and mobile phone data

• The study computes Pearson correlations for 13 mobile variables against 232 food-consumption and poverty indicators, with p-values and confidence intervals.Shuffling sector order produced maximum correlations of only 0.3, and the highly correlated variables differed across random trials.Source: Comparing food consumption and mobile… · p. 4
• Linear and symbolic regression are used to fit mobile-phone variables to selected food indicators such as Food Consumption Scores and total food expenditure.These models complement the pairwise correlation analysis for indicators of particular interest.Source: Comparing food consumption and mobile… · p. 4

Results

• Correlations above 0.7 were found for several food-security indicators, with sum of expenses and maximum top-up generally performing best.Coefficients of variation for summed and average top-ups did not yield high correlations.Source: Results · pp. 4–5
• 0.89 was the highest correlation, from a quadratic model using sum of expenses and average top-up to shadow food expenditure.A linear model correlated 0.69 with the Food Consumption Score, whereas the Coping Strategy Index correlated only 0.25.Source: Results · p. 4
• Vitamin-rich vegetables, rice, bread, sugar, and fresh meat each correlated above 0.7 with airtime purchases, whereas cassava and beans showed no relation.The authors relate this contrast to market-purchased versus personally cultivated foods.Source: Results · p. 5
• Airtime expenses had the only significant negative correlation with white sweet potato consumption, and correlated −0.64 with the share of food from household production.The authors suggest that households able to afford more may reduce consumption of this cheap, broadly cultivated item.Source: Results · p. 5
• Mobile-phone expenses correlated 0.72 with education, school-fee, and uniform expenses, while reported communication expenses correlated 0.69 with top-up expenses.The survey’s communication category was broader than mobile expenses alone.Source: Results · p. 5
• Top-up information produced a sector-level proxy for multidimensional poverty with correlation above 0.8.The denomination distribution was stable and dominated by common fixed purchase amounts.Source: Results · p. 5

• Top-up metrics significantly correlate with composed food security indicators and consumption of vitamin-rich vegetables, rice, bread, sugar, and fresh meat.These results support using mobile phone expenditure as a proxy for food expenditure in market-dependent households.Source: Discussion · p. 5
• In this study, households spend about 5-10 % of the money spent on food on mobile services.Source: Discussion · p. 6
• Mobile phone activity can infer poverty levels at a granular level, with top-up metrics achieving strong correlations and narrower confidence intervals than activity metrics.The comparison also reports poor and opposite correlations for social diversity, potentially reflecting differing indicators, social structures, and network coverage.Source: Discussion · p. 6
• Dynamic monitoring remains constrained because the study lacked household surveys at multiple time points needed to validate temporal changes in mobile indicators.The exploratory analysis found similar trends between top-up activity and food-stock levels, but the authors present early-warning implications as hypotheses.Source: Discussion · p. 6
• Privacy risks remain pressing because uniquely identifiable phone activity may be recoverable even from de-identified CDRs.Spatial aggregation can reduce this risk but may also reduce operational detail; the study reports that aggregated data can still provide indicators at equivalent spatial resolution.Source: Discussion · p. 6
• Simple top-up statistics could provide fine-grained, near-real-time information between infrequent household surveys rather than replacing them.The authors envision aggregated carrier-held measures supporting early warning and policy responses while sensitive personal data remains within mobile companies.Source: Discussion · pp. 6–7

• Figure 1 organizes absolute Pearson correlations by household, food-consumption, and expense variables against aggregated top-up and social-diversity measures.Source: Figure 1 · p. 9
• 0.89 is the correlation for a quadratic combination of CDR variables against food expenses.Source: Figure 2 · p. 9
• 0.69 is the correlation for a linear combination of CDR variables against the Food Consumption Score.Source: Figure 3 · p. 10
• 0.89 is the correlation for a quadratic combination of CDR variables against carrots and orange-vegetable consumption.Source: Figure 4 · p. 10
• 0.8 is the correlation for a linear combination of CDR variables against MPI.Source: Figure 5 · p. 11
• Figure 6 places seasonal food-stock percentages above sector-level average mobile expenses over 30-day rolling windows on a logarithmic scale.Source: Figure 6 · p. 11

• Table 1 reports correlations between different food types and the sum of airtime expenses, with 95% confidence intervals and p-values below 10^-15.Source: Table 1 · p. 12, Tables · p. 13
• White sweet potato is the only food group identified as significantly negatively correlated with the sum of mobile-phone expenses.Source: Tables · p. 12
• Table 2 selects household-survey variables with high correlation with the sum of mobile-consumption expenses.Source: Table 2 · p. 13