Same numbers as the individual pages. This page loads the same data files and
runs the same functions — the county-day matrix, the least-squares fit and the
Gaussian smoother are shared code, not a reimplementation. A cell here and the map on
the hazard page at the same hazard, threshold, period and
smoothing agree to the last digit. That is checked by the test suite rather than
assumed.
What a cell shows. The large number is the climatology — mean hazard
days per year (derecho is in events per year). Beneath it is the same figure as a
return period, 1÷rate: the long-run average spacing between events,
not a schedule, and not a promise about any particular year. The bottom line is the
least-squares trend per decade, given twice: in the hazard's own units, and
beneath that as a percentage of the climatology in the same cell. The
percentage is what makes cells comparable — 0.05 days/yr per decade
is a rounding error on a hazard that happens three times a year and a large
change on one that happens once a decade.
The arrow is the trend, not a significance test. It points up for
rising and down for falling, with its length scaled to the magnitude. Where
the trend is smaller than 10% of that cell's climatology — the percentage
printed in the cell — it lies flat
(↔) instead: a change that small is not worth a direction whichever way
it points. Both sides of that comparison are the smoothed values the cell displays.
p-values are deliberately absent — “could this slope have come
from noise?” is a different question from “is it changing enough to
matter?”, and only the second one is this page's business.
The trend's colour is that same percentage. Red for rising, blue for
falling, deepening in four steps at 10%, 20%, 35% and 60% of the climatology per
decade, and grey below 10%. It is relative rather than absolute because no
single absolute scale can serve a hail cell at 4.8 days/yr and a derecho
cell at 0.02 events/yr at once.
Colour is an absolute scale. Nine classes in events per year, the same in
every cell, every hazard and every county, so two cells can be compared by eye. The
breaks step by roughly half an order of magnitude because the rates span four;
even spacing would put nearly every cell in one class. A climatology of exactly
zero is grey rather than the pale end of the scale. The palette stops at a strong
red-orange rather than running to near-black, so one text colour stays legible on
every class.
Always smoothed, sigma 2°, area-weighted and normalised — the same
setting the individual pages open on. That means a cell is a neighbourhood
average, not the county's own count: half the weight comes from within
230 km and nothing beyond 444 km contributes. A single county's reports
are far too sparse to carry a trend on their own, which is why there is no raw
option here; the hazard page has one.
One fixed period, 2000 onward, shown beside each hazard and clipped to that
hazard's own record. There is no period control on purpose: pre-2000 windows
measure the growth of reporting as much as the weather (see below), and this page
is meant to be read without knowing that. Use
the hazard page to choose a window.
Severity runs left to right and the thresholds are cumulative, so a 2″
hailstone is also counted as a 1″ one. Hazards have different numbers of
levels, so rows are ragged: tornado has five, derecho three.
Sources. Hail, tornado and thunderstorm wind are NOAA/NCEI Storm Events.
Derecho combines the swath archive of
Squitieri, Wade and Jirak (2026) with the Storm Events reports
falling inside those swaths; that archive is NEXRAD-era only, so the derecho record
begins in 1996.
Freezing rain and peak wind are not here yet. They are station data — twelve
Indiana sites, not a national county field — so there is no county value to
show. They remain on the hazard page.
June and July 1993 are missing from Storm Events entirely, so 1993 is dropped
from any window that contains it rather than counted as a quiet year.
Storm Events is a report database, not a fixed observing network. National
counts rose roughly thirtyfold between 1955 and the late 1990s as population,
spotter networks and radar changed, so trends spanning that era partly measure the
growth of reporting. That is why this page starts at 2000. Even within it, the
higher severities are much less sensitive to reporting practice than
“any report”.