Wildfire Evacuation Statistics: Alerts, People and Mobile Data

What does “people evacuated” mean in a wildfire report?

Start by finding the source’s counting rule. A population inside an evacuation boundary might describe the intended scope of an order. A traffic observation might count vehicles passing one location. A follow-up interview might document a household’s reported departure. These can inform the same event while remaining separate evidence objects.

NIST’s Camp Fire NETTRA publication, dated 18 July 2023, describes integrating more than 2,600 observations and data points with the fire-progression timeline. It examines notifications, traffic, refuge areas and other life-safety evidence. That number describes mixed evidence records, not 2,600 independently counted evacuees. This guide uses the agency’s publication abstract for that scope statement. [1]

Ask for both the unit and the boundary: who, where, when, and under what inclusion rule? A figure for an evacuation zone should not be silently relabeled as an observed movement count. Likewise, vehicle counts cannot become people without an occupancy estimate and a way to address repeat passages.

Keep an alert’s identity, updates and timestamps

OASIS CAP 1.2 distinguishes a message’s identifier, sender and origination time. Its message types include initial alerts, updates and cancellations, with references linking earlier messages. The status field also distinguishes actual messages from tests, exercises and drafts. The standard describes message records; it does not certify household departure. [2]

Practical implication: one warning carried over two channels is not automatically two unique households warned. An update is not necessarily a new affected population. Preserve the message history, recipient identifiers and channel records, then define exactly which unique objects the final total counts.

Keep the original timestamp and timezone alongside the normalized analysis time. An alert’s origination, its effective time and the expected start of the subject event are separate fields in CAP. Do not replace one with another simply because all three resemble dates. [2]

The synthetic demonstration below is an analysis table, not a generated CAP message or a certification of CAP conformance. Its delivery flags are invented system records. They should not be read as evidence that a human saw or understood a real emergency alert.

One invented dataset, six different statistics

Entirely synthetic, not observations from a wildfire. The example targets 400 unique households. It records 600 notification attempts to 300 unique household identifiers; 250 identifiers have at least one invented delivery-success record. A separate survey has 200 respondents: 150 report departure and 50 report not departing at the same cutoff.

The last row does not imply that the remaining 62.5% stayed. Half the target supplied no survey response. Assuming accurate reports at a common cutoff, departures could range from 150 to 350 households, or 37.5–87.5% of the target. This is a logical range for the invented records, not a statistical confidence interval.

The survey respondents are deliberately selected from different parts of the target in this demonstration. Only 100 respondents intersect the 250-identifier delivery-success group. That intersection is specified in the synthetic example; the marginal totals of 200 and 250 alone cannot reconstruct it.

Consequently, multiplying the delivery percentage by the survey departure percentage would combine different groups and invent a causal chain. A matched household analysis would need linked observations, explicit coverage and appropriate handling of missing evidence. Even a correct join would not, by itself, prove that the notification caused departure.

What does a mobile-data “evacuation rate” measure?

Jiang and colleagues’ 2026 repository version studies Palisades and Eaton using Facebook population and movement data. Their methods compare baseline and crisis counts to infer evacuation-related changes. Observations are aggregated into approximately 2.4 × 2.4 km tiles and eight-hour intervals. Privacy filtering can remove rows or null a count; some missing observations are imputed. The authors state that they cannot share the underlying data. [3]

Tecnobosque interpretation: retain the inference rule and observation unit with any reported rate. An aggregated count difference is a different evidence object from an individually verified departure. An eight-hour observation window also cannot establish a person’s departure minute. These are limits on the questions a product can answer, not a reason to discard it.

For a deliberately invented tile, a baseline count of 80 and a later count of 28 give (80 − 28) ÷ 80 = 65.0%. That arithmetic describes a relative count deficit under the specified baseline. Calling it a verified 65% departure rate for every resident would require additional evidence about membership, coverage and movement.

Keep a missing count missing. A null crisis value is not a measured zero; replacing it with zero would create a 100% deficit against any positive baseline. A zero baseline also makes the ordinary relative-change denominator undefined. Neither case should quietly become a dramatic evacuation finding.

This review examined the repository’s version 1, submitted 3 January 2026. Its record lists a journal version, but this guide does not claim a version-by-version comparison or independent reprocessing of the original mobile data. The examples supplied here are newly generated synthetic records.

Repeated population deficits are not unique departure events

A stock describes a quantity present during an observation. A flow describes movement over an interval. Keeping those concepts separate matters when a population stays below baseline across several observations.

Unchanged counts between the last two snapshots do not prove that nobody moved; entries could balance exits. They do show why the two aggregate values alone cannot identify individual movements. Summing deficits can produce a separately defined index, but it must retain that definition rather than become a count of unique departures.

A stronger movement estimate distinguishes entries, exits, repeat trips and continued absence. Publish whether a result concerns net population change, outgoing movement records, unique tracked users or a measure accumulated across time. Those labels are part of the result, not optional presentation details.

Which clock starts an evacuation delay?

Consider an invented linked household record: an order is issued at 11:00 UTC, a system delivery record appears at 11:02, the household reports receiving the message at 11:20, and reports departure at 11:25. The example keeps system evidence separate from human recollection.

The 25-minute interval is not automatically 25 minutes of decision-making after receipt. Nor does the five-minute interval measure road clearance or travel to a destination. State the start event, end event, timing source and precision before comparing delays.

If a person departed before an order, a negative signed difference can describe that chronology. Automatically clipping it to zero erases the earlier departure. Whether an analysis uses signed differences, truncation or a different event rule should be stated and justified.

A compact framework for auditing evacuation figures

1. Define the target. Record the geography, population unit, cutoff and how eligibility was established.
2. Retain identifiers. Deduplicate the objects the statistic claims to count; keep transmission attempts separately.
3. Preserve message versions. Updates, cancellations and test status can change how a record belongs in an analysis.
4. Preserve missingness. Distinguish unavailable observations, privacy suppression, imputation and actual zeros.
5. Match the clocks. Keep source timestamps, timezone, normalization and uncertainty.
6. Publish the numerator and denominator. Readers should be able to reconstruct the percentage without guessing.

This is a proposed analytical workflow, not an official performance rating or a substitute for an incident investigation. Its purpose is to make comparisons reviewable. A historical dataset can support useful conclusions while leaving some households, movement sequences or causal questions unresolved.

Primary sources and reproducible examples

1. NIST: A Case Study of the Camp Fire — Notification, Evacuation, Traffic, and Temporary Refuge Areas. Technical Note 2252; agency publication page dated 18 July 2023. Abstract consulted for the study’s scope; the original event database was not reanalyzed.
2. OASIS: Common Alerting Protocol Version 1.2. Standard dated 1 July 2010. Sections 3.2.1–3.2.2: identity, status, message relationships and time fields. The proposed audit table is Tecnobosque’s explanation, not a copied standard or a CAP implementation.
3. Jiang and colleagues: Wildfire Evacuation Analysis Using Facebook Data: Evidence from Palisades and Eaton Fires. Repository version 1, 3 January 2026; methods 3.1 and 4.2, limitations 6.2, and data-availability statement. Version and bibliographic record. Repository text is licensed CC BY 4.0; its original authorship is retained.

Sources checked 13 October 2026. The household records, notification attempts, survey responses, mobile snapshots, timestamps, arithmetic and chart are Tecnobosque’s synthetic demonstration. The calculations above use explicitly labeled synthetic inputs. They contain no real household identifiers or incident microdata.

Cite: Tecnobosque, Wildfire Evacuation Statistics: Alerts, People and Mobile Data, TB-METHOD-EVACUATION-STATISTICS, version 1.0, reviewed 13 October 2026.