A Guangxi snakebite death fits five defensible slots in the disaster loss record

Thirty-nine becoming 159 in six weeks is a factor of about 4.1 ^[1], roughly 20 further deaths entering the official record every week ^[2]. Someone who queried that toll on day three and someone who queried it on day forty were both reading the best available number for the same storm, and they would have drawn very different conclusions from it.

Revision does not only push upward. Srikanth and Petley ^[1] open with a sequence that runs 20, then 27, then 23, because some people first listed as missing are confirmed alive while some listed as injured later die ^[4]. Both corrections are legitimate even though they point in opposite directions, which means a count can fall on its way to being right.

Under the timing problem sits the harder one. Mortality registries look for the underlying cause of death, the injury or circumstance that started the chain, while disaster databases attach deaths to a particular hazard event and admit indirect consequences alongside direct ones ^[9]. The Guangxi snakebite has a legitimate home in each system under a different name: a snakebite death, a flood-related death, a consequence of the reservoir failure, a typhoon death, or an indirect disaster fatality, with none of those labels necessarily wrong ^[8]. The same fatality can therefore occupy several positions in one causal chain and pass through several interpretive layers before it becomes a row ^[10]. Reported delays in reaching medical care ^[7] are a further thread that no single label carries.

The event boundary compounds it. Databases want one event as the unit of analysis, and a cyclone that generates rainfall, flooding, landslides and infrastructure failures offers several defensible cut points for a death that arrived through more than one of them ^[11]. WMO is working towards a globally accepted standard for the temporal or spatial extent of a hazard event ^[13], which is itself a statement that no accepted one exists yet. The same review counted consultation with national hydrological and meteorological services in about 18 of 91 post-disaster needs assessments ^[3], so most of those attributions were made without the agencies best placed to say where the hazard started and stopped.

These examples establish a mechanism, and they leave the magnitude open. The authors point to published work on missing data in EM-DAT, on loss-data fallacies and on flood loss databases to show that definitions, thresholds and reporting practices shape what gets recorded ^[3], and none of the cases here puts a size or a sign on how far a given entry sits from the truth. The mechanism does say something about the shape of the error. A classification convention applied consistently across thousands of entries does not average out across them the way sampling noise does; it carries the whole series in whichever direction the convention leans. That is an argument from process, and it is the kind of argument that a tail fitted to per-event losses cannot detect from the losses alone.

Two analysts working carefully from the same database can arrive at different tolls for the same typhoon. Until each row carries its attribution rule and the date it was last revised, neither of them can show the other is wrong.