📊 statistics

Context-Aware Displacement Estimation from Mobile Phone Data: A Methodological Framework

This paper presents a methodological framework that leverages mobile phone data to generate context-aware, near real-time population displacement estimates by distinguishing regular commuters from disaster-affected individuals, thereby reducing overestimation errors and providing humanitarian actors with actionable metrics and uncertainty bounds for decision support.

Original authors: Rajius Idzalika, Muhammad Rheza Muztahid, Radityo Eko Prasojo

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rajius Idzalika, Muhammad Rheza Muztahid, Radityo Eko Prasojo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to count how many people have fled their homes after a massive storm. In the past, aid workers had to drive to villages, knock on doors, and ask, "Did you leave?" This is slow, expensive, and often impossible when roads are washed out.

This paper proposes a faster way: using mobile phone data as a "digital footprint" to see where people are moving in near real-time. However, the authors found a major problem with how this was usually done, and they built a new "smart filter" to fix it.

Here is the breakdown of their method, using simple analogies.

The Problem: The "Commuter Confusion"

Think of a typical day for a person who lives in Town A but works in Town B.

Monday to Friday: They wake up in Town A, go to work in Town B, and sleep in Town A.
The Old Method: If a storm hits, the old computer program looks at the phone data. On a Tuesday, it sees the person in Town B. Since their "home" is Town A, the computer screams, "This person has been displaced! They left their home!"

This is a false alarm. The person isn't displaced; they are just at work. If you count every commuter as a refugee, you end up with a massive overestimate, leading aid groups to send too many supplies to the wrong places.

The Solution: The "Context-Aware" Filter

The authors created a new framework that acts like a smart bouncer at a club. Instead of just checking if someone is at a different location, the bouncer checks who they are and what day it is.

They do this in three steps:

1. Building a "Personality Profile" (Mobility Classification)

Before the storm hits, the system watches everyone for a few weeks to learn their habits. It sorts people into groups:

The Local: Always stays in the same town.
The Daily Commuter: Lives in Town A, works in Town B, every weekday.
The Weekend Traveler: Only moves on Saturdays and Sundays.

2. The "Smart Bouncer" Rules (Context-Aware Detection)

When the storm hits, the system applies different rules based on the profile:

If you are a Local: If you are not in your home town, you are Displaced.
If you are a Daily Commuter:
- On a Tuesday: If you are in your work town (Town B), you are NOT Displaced. (The bouncer lets you in).
- On a Saturday: If you are in your work town (Town B) instead of your home (Town A), you ARE Displaced. (The bouncer stops you, because you should be home).

This simple change prevents the system from panicking every time a worker goes to their office.

3. The "Safety Margin" (Uncertainty Bounds)

The authors know this isn't perfect. Sometimes phones die, or towers break during a storm. So, they don't just give a single number (e.g., "5,000 people displaced"). Instead, they give a range, like a weather forecast.

Example: "We estimate 5,000 people are displaced, but the real number could be between 4,300 and 5,700."
They calculate this range by looking at how much people's movements usually wiggle around on normal days, then adding a "disaster factor" to account for the chaos of a storm.

The Real-World Test: Typhoon Nando

The team tested this in the Philippines during Super Typhoon Nando (2025) in a town called Aparri.

The Result: The old method said about 9% of people were displaced on weekdays. The new "smart" method said only 6.5%.
The Difference: The old method was counting about 2.5% of the population (the commuters) as refugees when they were actually just at work.
The Insight: The new method also showed where people went. It found that many people moved to the regional capital (Tuguegarao) for safety, not just to the next village over.

What This Paper Does NOT Do

To be clear about the limits:

It doesn't count people who stayed home. If a storm hits and you stay in your house, your phone might still be connected to the local tower. The system thinks you are "home," so it doesn't count you as displaced, even if your house is flooded.
It doesn't know why you moved. It sees you moved, but it doesn't know if you left because of the storm or to visit a friend. It assumes the storm caused the move because it happened right after the storm.
It's not a perfect census. It relies on people having phones and keeping them on. If you are poor, elderly, or in a rural area with no signal, you might be invisible to this system.

The Bottom Line

This paper offers a faster, smarter way to guess where people are going after a disaster. By teaching the computer to understand the difference between "going to work" and "fleeing a storm," aid organizations can stop wasting resources on false alarms and focus on the people who actually need help. It turns a blurry, noisy signal into a clearer picture of who is safe and who is in trouble.

Technical Summary: Context-Aware Displacement Estimation from Mobile Phone Data

Problem Statement
Timely and accurate population displacement estimates are critical for humanitarian response during disasters. Traditional methods, such as household surveys and field assessments, are resource-intensive, slow to deploy, and often restricted by physical access to disaster zones. While mobile phone data offers a mechanism for near real-time tracking, existing approaches typically apply uniform displacement definitions that fail to distinguish between forced displacement and regular mobility patterns. Specifically, standard methods often misclassify regular commuters as displaced persons when they are observed in locations different from their home, leading to systematic overestimation of displacement figures.

Methodological Framework
The paper presents a methodological framework designed to separate potential displacement from regular mobility patterns using mobile phone data. The framework operates through a six-stage pipeline:

Daily Location Signal: The system utilizes a single vendor-supplied location per user per day (external mode), distinguishing weekend locations (residential-like) from weekday locations (daytime-activity-like).
Residential Baseline Establishment: A stable "true home" is determined from a pre-disaster baseline period (minimum 6 weeks). The methodology prioritizes weekend observations (requiring $\ge$ 2 days) to identify the residential location, falling back to the modal weekday location (requiring $\ge$ 5 days) if weekend data is insufficient.
Mobility Profile Classification: Users are categorized based on the relationship between their residential baseline and their modal weekday location:
- Local Resident: Same location on weekdays and weekends.
- Intra-City Commuter: Different location within the same city.
- Inter-City Commuter: Different location in a different city.
- Weekend-Only User: No weekday data available.
Context-Aware Displacement Detection: Displacement rules are applied dynamically based on user type and day of the week.
- Local Residents/Intra-City Commuters: Displaced if observed outside their home city.
- Inter-City Commuters: Displaced only if observed outside both their home city and their work city on weekdays. On weekends, they are displaced if outside their home city.
- Weekend-Only Users: Displaced if outside their home city.
Metric Calculation: The framework computes three metrics: displacement rates, origin-destination (O-D) flows, and return dynamics.
Population Scaling and Uncertainty: Subscriber counts are scaled to population levels using a city-specific ratio. Uncertainty bounds are derived from the baseline coefficient of variation (CV) of daily user counts, adjusted by a disaster multiplier (default 2.0) to account for crisis-induced behavioral variability.

Key Contributions
The paper identifies three primary methodological innovations:

Mobility Profile Classification: Explicitly categorizing users by commuting patterns (local, intra-city, inter-city) to differentiate regular mobility from displacement.
Context-Aware Detection Logic: Applying displacement rules that vary by user type and day of the week, specifically creating an "exception" for inter-city commuters on weekdays to prevent false positives.
Operational Uncertainty Quantification: Utilizing a CV-based approach with a disaster adjustment factor to provide decision-support bounds rather than formal statistical inference intervals. This approach acknowledges the partial identification problem where "missing" users (unobservable due to network issues or lack of activity) cannot be classified as displaced or non-displaced.

Case Study Results
The framework was applied to data from Aparri, Cagayan, following Super Typhoon Nando in September 2025, using daily location data from Globe Telecom.

Reduction in Overestimation: The context-aware method reduced estimated between-municipality displacement by 1.6 to 2.7 percentage points on weekdays compared to naive methods. This difference is attributed to the exclusion of inter-city commuters who were incorrectly flagged as displaced by uniform rules.
Weekend Convergence: On weekends, the context-aware and naive methods produced nearly identical results (difference of 0.1 pp), confirming that the commuter exception logic functions as intended only for weekdays.
Displacement Patterns: The study identified that displaced persons moved to both immediately neighboring municipalities and the regional capital, Tuguegarao City (approx. 80 km away).
Return Dynamics: A retrospective analysis showed a cumulative return rate rising from 0% to 87.6% over 15 days, with 50% of displaced subscribers returning by day 9.
Comparison with Official Data: The framework's estimate of 6.5% between-municipality displacement was lower than the official DSWD count of 15.5%. The authors note this is expected, as the official count includes internal evacuees (within-municipality), which the mobile data framework does not detect.

Significance and Limitations
The paper positions this framework as a tool for operational decision support rather than formal statistical inference. Its primary significance lies in providing rapid, municipality-level estimates that account for regular commuting patterns, thereby reducing the risk of misallocating humanitarian resources based on inflated displacement numbers.

The authors explicitly acknowledge several limitations:

Scope: The method detects only between-municipality displacement; within-municipality evacuation is outside its scope.
Validation: The commuter exception logic has not been independently validated against GPS ground truth or self-reported data; the results are presented as a proof of concept.
Data Bias: Estimates rely on mobile subscribers, potentially underrepresenting elderly, rural, and low-income populations. Single-operator data (Globe Telecom) may introduce market-share biases.
Generalizability: The framework is designed for sudden-onset disasters. Application to slow-onset disasters or conflict zones requires methodological modifications.
Missing Data: Users unobservable during the post-disaster period cannot be classified, creating a "partial identification" scenario where the true displacement rate lies between the observed rate and the observed rate plus the missing rate.

The paper concludes that while external validity across multiple events remains future work, the framework offers a viable, privacy-preserving approach for generating timely displacement intelligence to complement traditional field assessments.