Five NEON products, normally explored apart, share field sites. This app lines them up by year into one bottom-up cascade :
climate sets the water and warmth → green-up timing (the hinge) → producers (plants) → consumers (small mammals, birds).
Each site has only a handful of years. This tool states what the literature predicts and checks whether the data agrees. It does not hunt for whichever correlation looks biggest.
Browse all 46 sites
The annual signals available at this site, by trophic layer.
The cascade, year by year
Each strip is one trophic layer; each line a signal,
standardised
(z-scored) so layers are comparable on one scale. Shared x-axis = year.
Watch whether pulses
march down the ladder
: a wet year, then a green flush, then more plants, then more rodents. The chips on the right show whether each link matches its
literature-predicted direction
.
Short series: read the shapes, not a single number. Verdicts are gated by how many years overlap.
Reading the ladder
Each strip is one trophic layer; each line a signal, standardised (z-scored) so layers are comparable on one scale. Shared x-axis = year.
Watch whether pulses march down the ladder : a wet year, then a green flush, then more plants, then more rodents. The chips on the right show whether each link matches its literature-predicted direction .
Short series: read the shapes, not a single number. Verdicts are gated by how many years overlap.
Standardised annual signals, stacked by trophic layer. The eye does the judging; the chips keep it honest.
Each signal is rescaled so 0 = its own average year and +1 = one standard deviation above. Signals in different units can then share one axis, so you compare the TIMING of the bumps, not their heights.
A delay. A 1-year lag means this year's driver shows up in NEXT year's response: rain grows a seed crop that feeds the rodents the following year. In this app, lag is evaluated on annual site-year summaries; month-level seasonality is represented by separate winter/monsoon and spring-temperature signals.
Each arrow is a predicted driver→response link with its expected sign and lag. The colour says whether the data agrees, and is honest about how few years stand behind it.
In bimodal deserts, rain comes in two ecologically independent seasons: winter (Oct–Mar, which grows the spring forbs) and the summer monsoon (Jul–Sep, which grows the grass seeds desert rodents eat).
They can even move in opposite directions from year to year, so a single annual rainfall total blends them into noise. Splitting them back out is what recovers the desert cascade.
For the response you pick (left), every candidate driver the literature proposes, with its expected
sign
and
lag
, checked against this site's data.
At
n < 6 years
no verdict is given (the series is too short); links are shown as
exploratory
. Expected lag/sign are literature priors on annual site-year summaries (not selected by monthly correlation in this table). At n≥6 the bootstrap interval sets the per-site direction verdict; the permutation p is reported but cannot reach significance at this n (that is the cross-site pooled test's job). Never read these as causation.
Driver Lab
For the response you pick (left), every candidate driver the literature proposes, with its expected sign and lag , checked against this site's data.
At n < 6 years no verdict is given (the series is too short); links are shown as exploratory . Expected lag/sign are literature priors on annual site-year summaries (not selected by monthly correlation in this table). At n≥6 the bootstrap interval sets the per-site direction verdict; the permutation p is reported but cannot reach significance at this n (that is the cross-site pooled test's job). Never read these as causation.
Which drivers explain this response here, measured against what ecology predicts, not against whichever lag looks best.
The driver (x) against the response (y) at the predicted lag, one point per overlapping year. Few points = wide uncertainty; that's the honest picture.
Permutation p uses an autocorrelation-preserving circular-shift null (response years are rotated, not freely shuffled). The honest test is the cross-site pooling on Across NEON.
95% CI is a bootstrap interval (wide at this n; indicative, not a precision claim).
Lag & season explorer: see why we lock the prior, not chase the best fit
Re-examine a stated prior by sliding the lag and toggling annual vs seasonal climate. Watch how easily a better-looking r appears when you search, and why that p is not a real one. The gold diamond is the prior lag, fixed before looking; it is the only honest reading here.
r (correlation) how tightly the two move together, from -1 (they move opposite) through 0 (no link) to +1 (in lockstep). It never means one causes the other.
n how many years of overlap the number rests on. Fewer years means shakier, and under 6 we give no verdict at all.
adjusted p the chance a correlation this strong could turn up by luck, after counting every lag and season you tried. Near 1 means easily luck; near 0 means hard to explain by chance. We adjust for the search because trying many combinations finds a winner even in pure noise.
un-adjusted p the same chance for one lag on its own. It looks more impressive than it should once you have scanned several, so we show it small.
bootstrap 95% range the span the true correlation could plausibly sit in, given so few years. A wide range, or one that crosses 0, means we cannot pin it down.
the bar how small the p would have to be to count as real after that many tries. Sliding rarely beats it.
on the prior vs EXPLORED the gold lag is the literature value, fixed before looking, and the one honest reading. Slide off it and you are exploring, not testing, so that p shows struck through.
You just looked at up to K lag-by-season combinations. With about 6 years, the best-looking one reaches p around 0.05 roughly 1 in 3 by chance. A p you found by sliding is not a p=.03 result. The honest tests are the prior lag (gold) and the cross-site pooling on Across NEON.
Does the cascade hold across NEON?
One site's handful of years can't settle anything. But each predicted link can be
pooled across the sites where it's expected
(one vote per site), which is the statistically honest way past the short-series problem.
The grid shows every link's verdict at every site.
Green
columns marching down a biome block mean the predicted direction holds cleanly at those sites (the per-site verdict is direction, not significance, a short series can't be significant on its own); the
pooled
test above the grid is where significance lives.
Grey
means too few years to tell, and we show that, rather than hide it.
The scoreboard
One site's handful of years can't settle anything. But each predicted link can be pooled across the sites where it's expected (one vote per site), which is the statistically honest way past the short-series problem.
The grid shows every link's verdict at every site. Green columns marching down a biome block mean the predicted direction holds cleanly at those sites (the per-site verdict is direction, not significance, a short series can't be significant on its own); the pooled test above the grid is where significance lives. Grey means too few years to tell, and we show that, rather than hide it.
Per-site series are short; pooling across sites gives a stronger test of direction. Each cell is one link's verdict at one site. Click any site to open it.
Rows are sites grouped by biome; columns are the predicted links. Cell colour = the verdict; faded cells aren't the mechanism expected for that biome.
Search the network
Two ways to query all
46
sites at once, off the bundled index (instant, no download):
Find a link
: pick one driver→response prior and see every site where it was tested, sorted by how well the data agrees.
Cascade strength
: rank sites by how many of their biome-expected links the data agrees with.
Per-site tests rest on a handful of years and are
underpowered
. A site missing from a result usually means too few years, not a real absence. The honest cross-site test is the pooled result on
Across NEON
.
Searching the cascade
Two ways to query all 46 sites at once, off the bundled index (instant, no download):
Find a link : pick one driver→response prior and see every site where it was tested, sorted by how well the data agrees.
Cascade strength : rank sites by how many of their biome-expected links the data agrees with.
Per-site tests rest on a handful of years and are underpowered . A site missing from a result usually means too few years, not a real absence. The honest cross-site test is the pooled result on Across NEON .
Query every NEON site in the cascade at once. Pick a single link to see where it holds, or rank sites by how much of their expected cascade the data agrees with.
Each row is this link tested at one site: r is within-site, n is overlapping years, p is an autocorrelation-preserving circular-shift permutation p. “Expected” marks the biome where the mechanism is predicted; faded sites still ran it as out-of-biome corroboration. A single site's short series can't settle the link; the pooled test on Across NEON can.
“Agree” = the data points the literature-predicted direction on a biome-expected link with at least 6 overlapping years. It does NOT require statistical significance, since at this n almost nothing reaches it. Read this as k of K expected links agree, a direction tally, not an absolute ranking of ecosystems. Magnitudes (the within-site indices) are not comparable across sites.
Every number here is a within-site, short-series screen. It points you to sites worth opening, not to settled results. The statistically honest, cross-site answer is the pooled binomial on Across NEON.
Data-quality review for this site
The suite gold-standard QC panel: ranked
“verify, not wrong”
flags for the selected site, worst-first.
Every flag is a value to
look at
before over-reading it, never a bug. The cascade's QC rules (the ≥5-individual green-up gate, the within-site temperature outlier filter, the CI-spans-zero guard) are correct; this just shows where they bit. Tap a flag for the exact rows; export the whole review as CSV.
Verify, not wrong
The suite gold-standard QC panel: ranked “verify, not wrong” flags for the selected site, worst-first.
Every flag is a value to look at before over-reading it, never a bug. The cascade's QC rules (the ≥5-individual green-up gate, the within-site temperature outlier filter, the CI-spans-zero guard) are correct; this just shows where they bit. Tap a flag for the exact rows; export the whole review as CSV.
A clean site shows a single green all-clear. Pick another site on the Overview to review it.