Applied research / Floods & information seeking

GeoDemand.

When floods happen,
when do people search?

Connecting reported flood episodes with search-interest signals—and tracing how the evidence changes with the data behind it.

A bounded descriptive study · GeoDemand 1.2 Explore the findings ↓
22 eligible exploratory flood episodes
+3 days median individual peak after reported onset
19/19 shared episodes with the same peak date across archive versions

01 / The signal

Attention has a timeline.

In the selected flood archive, search interest varies around reported onset. Explore the average response, then compare two archived versions on exactly the same 19 episodes.

Mean standardized interest · baseline standard deviations

Mean standardized flood-search interest across 22 exploratory episodes Daily means from 28 days before to 28 days after reported onset. Each episode is standardized against its own pre-event baseline. This is a descriptive curve, not a causal estimate. 0 1 2 3 4 5 Reported onset -28 -21 -14 -7 +0 +7 +14 +21 +28 Days from reported onset
Partial-preserving archive
Equal-weight daily means, aligned to reported onset; shading marks the baseline. The median individual peak lag (+3 days) is a different statistic from the peak of this average curve. Peak metrics use the full event-end-anchored window, which can extend beyond this chart.

Across the 19 shared episodes, mean absolute peak-lift difference is 2.31 baseline SD. Identical peak dates do not imply identical magnitudes or independent replication.

02 / The evidence

Two evidence levels.
One transparent account.

These samples answer different questions. Their counts are kept separate and must not be added together.

Provenance-checked subset

5 episodes · weather context

Five episode/concept units pass the frozen admission and numerical checks. Only weather-context signals are eligible; this subset does not establish a flood-specific response.

“Verified” means file hashes, lineage and values pass the pipeline checks. It does not independently authenticate the event or acquisition service.

Exploratory archive

22 episodes · flood searches

Twenty-two of 40 archived episodes pass unchanged numerical rules in the partial-preserving version. Median peak lift is 4.56 baseline SD; peak lags range from −7 to +24 days.

Acquisition provenance remains incomplete. Nineteen episodes are eligible in both versions, forming the comparison above.

03 / How it works

Start with the event.
Keep the history of the data.

  1. Link

    Connect reported episodes, geography and search concepts. Preserve request identity, raw-file hashes and repeat lineage where available.

  2. Screen

    Check daily coverage, sparse signals and baseline variability before computing a normalized response. Keep exclusion reasons visible.

  3. Compare

    Describe the size and timing of peaks. Compare archive versions on shared eligible episodes without treating them as independent repeats.

What does “peak lift” measure?

The maximum search index from onset −7 through event end +28, minus the baseline mean, divided by the baseline population standard deviation. It is neither a percentage increase nor a search count. Earliest tied peaks determine timing. Longer events have longer response windows and more opportunities for a maximum.

Which numerical rules stay fixed?

The baseline spans onset −28 through −8 and needs at least 14 valid days. Each response phase needs 80% valid daily coverage. Zero fraction ≥0.5 or baseline population SD ≤0.001 excludes the normalized response. No threshold relaxation, denominator epsilon or outcome imputation is used.

What can this study conclude?

It describes search variation in selected episodes. There are zero eligible treated/control contrasts, so it cannot attribute peaks to floods. Prediction remains not estimable under the original independent-event and concept-coverage gates. The 22 exploratory episodes do not override those gates. Sparse-series selection and unresolved archive history constrain generalization.

Why compare archived versions?

Version choice affects both eligibility and magnitudes: 22 episodes pass in the partial-preserving archive and 19 in the alternative. Retrieval dates and attempt identities are unresolved, and the alternative omits partial flags. Agreement is within-archive sensitivity, not independent repeat reliability.

04 / Open the work

Useful signals. Explicit limits.

The descriptive study is complete. Read the evidence, inspect the project source, or discuss a related research question.

Read the research summary ↗ Project source on GitHub ↗ Discuss the research

This page presents aggregate findings. Raw acquisitions and local analysis packages are not distributed here.