Disaster Linking and Causal Chains

Project snapshot

DaedalMap keeps a shared disaster-links layer for formal event-to-event relationships across hazards. The goal is precision, not recall. The public graph answers questions such as what this event triggered, which chains are strongest, and how one hazard cascades into another without quietly mixing in weak time-place coincidence.

That means the linking system is intentionally narrower than the aggregate system. Aggregates ask what happened in a county or over a time window. Links ask whether one specific event plausibly caused or directly generated another specific event. The standard is stricter.

Public contract

The shared links table is thin on purpose. It stores event identities and relationship metadata, not copied event summaries.

• Identity fields. Parent and child event identity using canonical loc_id plus additive event_id helpers.
• Relationship type. The current public link_type stays narrow as triggered.
• Source and confidence. Each link carries its origin plus a confidence score so downstream users can judge the assumptions.
• No copied impact fields. Time, magnitude, area, and hazard-specific interpretation stay in the canonical event tables.

This separation matters. It keeps the graph stable while hazard-specific ranking, search, and interpretation can evolve in runtime without rewriting the stored relationship layer every time methodology improves.

The same page also works as a compact reference for the live method. The public-facing assumptions here track the active hardening work on the shared disaster graph, but the copy stays focused on what the published system does and how to evaluate it.

Methodology

There is no single universal industry standard for multi-hazard event linking. In practice, disaster systems use a hierarchy of evidence. The preferred order in DaedalMap is:

• Explicit source-native linkage. Foreign keys or agency-provided causal references are the strongest case.
• Hazard-specific physical rules. When no foreign key exists, use the established mechanism for that hazard pair such as storm track overlap, burn-scar overlap, or tight quake-landslide timing.
• Bounded spatiotemporal heuristics. Time overlap and distance only matter when they match a real process model, not as a generic co-occurrence rule.
• No automatic link when causality is ambiguous. Weak or speculative relationships stay out of the canonical graph.

That standard is auditable. If an outside researcher disagrees with a pairing rule, they should be able to inspect the assumptions and decide whether to accept, ignore, or rebuild that layer.

Confidence scoring

The confidence value is a public methodology signal, not just an internal convenience field.

• 1.0. Explicit source-native or agency-confirmed causal relationship.
• 0.8. Strong deterministic inference using hazard logic plus tight time and spatial agreement.
• 0.6. Plausible but more interpretive inference. Good enough for research exploration, not strong enough for the canonical shared table.
• Below 0.6. Excluded from public links.parquet.

The published shared graph includes only links scoring 0.8 or above. That keeps the graph smaller and easier to trust, while lower-confidence possibilities stay in research-oriented interpretation layers rather than the canonical public table.

Current and planned link families

Today the strongest shared cross-hazard links are still the original source-supported disaster pairs: earthquake to tsunami and volcano to earthquake or tsunami. The next planned expansions are chosen because their causal mechanisms are stronger and more operationally defensible.

• Earthquake -> aftershock. This stays in the earthquake-native contract rather than the shared cross-hazard graph. The current sequence logic follows the standard declustering tradition built around magnitude-scaled time and distance windows, rather than a simple fixed-radius cutoff.
• Volcano -> earthquake. Use source-supported event references first. More generally, volcano-tectonic earthquake swarms are a standard monitoring signal in volcanology, but they are not the same as every nearby tectonic earthquake.
• Hurricane -> tornado. Use storm timing plus track proximity. This is one of the best next auto-link candidates.
• Hurricane -> flood. Use storm timing, track or storm corridor overlap, and ideally cause agreement for rainfall, coastal flooding, or surge-related flooding.
• Wildfire -> flood. Use ordered timing and tight burn-scar or downstream overlap. This represents post-fire runoff or debris-flow style chains, not any later flood in the same region.
• Earthquake -> landslide. Use tight time, distance, and magnitude thresholds once the landslide pack is production-ready.
• Thunderstorm -> wildfire. A strong future pair, but only once storm or lightning-source data is detailed enough to link ignition timing and location cleanly.

Some pairs remain intentionally narrow or curated only. Earthquake to flood and volcano to wildfire are real in specific cases, but the default heuristics are much weaker. Those are not auto-generated broadly until the evidence model is stronger.

Assumptions and limits

The main assumption is that the canonical public graph should maximize precision over coverage. That means some true chains will be missing if the source did not make the relationship explicit or if the physical rule cannot be defended tightly enough with current fields. This is a known tradeoff, not a hidden gap.

The second assumption is that causal linking stays hazard-pair specific. A generic rule such as "two events happened nearby in the same week" does not scale honestly across hurricanes, floods, earthquakes, wildfires, and future hazards. The method must follow the process model for the specific pair.

Finally, the link table is not the whole system. Same-pack sequences such as earthquake aftershocks stay inside their native pack contracts, while the shared graph focuses on cross-hazard links and the search/ranking layer above them.

The calculations are simple on purpose: event identity, bounded time and space rules, source attribution when available, and a confidence score that determines whether a link is published. Outside users can inspect those assumptions directly instead of treating the graph as an opaque truth layer.

Reference basis

The linking rules are anchored in outside methodology when a strong reference exists. DaedalMap follows three levels of support:

• Agency-backed operational products. Best for earthquake-triggered ground failure and other hazards with active public response systems.
• Peer-reviewed hazard literature. Best for hazard pairs where the science is established but no single operational foreign key exists.
• DaedalMap heuristics. Used only when the pair is operationally useful but the public methodology is less standardized. In those cases the rule stays narrow and the confidence threshold stays strict.

The strongest external references in the current method stack are:

• Earthquake -> landslide. U.S. Geological Survey Ground Failure product and background materials. USGS treats earthquake-triggered landslides and liquefaction as a distinct secondary hazard class and models them from shaking plus susceptibility layers.
• Earthquake -> aftershock. The sequence logic follows the standard seismology declustering tradition associated with Gardner-Knopoff style magnitude-scaled windows, combined with ordinary aftershock behavior such as Omori-style decay.
• Volcano -> earthquake. Volcano-tectonic earthquake swarm literature and USGS volcano monitoring practice. The key point is that magma movement commonly produces earthquake swarms, but those swarms are not treated as generic tectonic aftershock sequences.
• Hurricane -> tornado. NOAA and severe-storm literature on tropical-cyclone tornadoes, including track-relative tornado occurrence and the strong role of storm timing and storm sector.
• Hurricane -> flood. NOAA/NHC/WPC rainfall and inland-flooding literature. Tropical-cyclone flooding is a standard hazard pathway and is typically associated through storm timing, track-relative rainfall footprint, and local amplification.
• Wildfire -> flood. USGS post-wildfire debris-flow and burn-scar hazard work. This provides the cleanest outside basis for linking a fire to later runoff, debris-flow, or related flood impacts.
• Thunderstorm -> wildfire. Operationally useful, but not published broadly until the storm-side source data is strong enough to support ignition attribution rather than rough co-occurrence.

Where no world-leading public method exists for a pair, the preferred fallback is to use the narrowest defensible causal rule and publish only higher-confidence results. That is why some real but weaker pairings such as earthquake to flood or volcano to wildfire stay curated-only for now.