Volume is the number that bounds everything else in this analysis. Spread evenly across 2016 through 2022, 24,598 posts works out to about 3,514 a year [2], or roughly ten a day for the whole of California [3]. That is why the analysis rests on county-level covariates for income, housing, transportation and tree canopy [6] rather than on anything finer. The county is the unit doing the statistical work; the posts are what gets counted inside it.
The housing result, reported in Environmental Research Communications [3], is the one that should interest anyone building an early-warning workflow, and it cuts against the intuition. Counties with high levels of RV, van and boat households had 22 per cent lower odds of posting about coping strategies, and were 15 per cent less likely to post positively about heat [9]. The team calls that category sensitive housing and notes those dwellings are often poorly insulated and can become dangerously hot [10]. So the population flagged as most exposed indoors is producing the least talk about hydration, air conditioning and pools [5]. That volume looks like quiet, but the authors read it as a gap in who heat-safety messaging reaches [14].
The label layer deserves scrutiny. Categories were assigned by a large language model and checked against a hand-sorted set, with accuracy between 74 and 90 per cent across heat impacts, coping strategies and other [8]. At the low end, that is roughly one post in four in the wrong bucket [4]. That accuracy is fine for grouping thousands of posts into themes, but it sits less comfortably underneath effect sizes of 15 and 22 per cent [9], since misclassification washes out only if it is unrelated to the county variables being tested, and nothing in the design as described guarantees that.
The commuter finding has the cleaner mechanism. Counties with more people walking, cycling or taking transit to work showed substantially greater odds of negative posts and posts about heat's effects on daily life [11], and active commuters take their exposure directly, while travelling [12]. The general pattern held the same way: housing tracked coping talk, active commuting predicted life-impact talk [7]. Counties with larger Hispanic populations showed a slightly greater likelihood of life-impact posts [13]. All of these are associations between county aggregates, and the distinction matters: the data show that counties with more RV, van and boat households produce less coping talk, not that any individual living in a van posts less about coping [9].
This design leaves timing unresolved, whether any of it arrives early enough to matter. The comparison here is post content against county indicators [6], not post content against dated illness records, so lead time is untested in this design. What the data do carry is coping behaviour, which temperature readings and heat-illness counts do not record [15]. That makes this a reasonable instrument for deciding which counties to go survey, and a poor one for deciding where to send a cooling van, until someone shows the categories moving before emergency-department visits in the same weeks.
