What presence-only data cannot tell you
Revised 10 August 2026
One absence, in Uppland, in 2005
Artportalen, run by SLU Artdatabanken, is one of the largest biodiversity databases in Europe. The version of it published to GBIF held 121,526,685 occurrence records when I queried the API in July 2026. Of those, 707,249 — roughly six in every thousand — carry an occurrence status of absent: a record that someone looked and the species was not there.
Narrow it to what a mushroom picker cares about and the ratio gets worse. Kingdom Fungi in Sweden: 6,826,703 GBIF records, 528,389 of them absences — but three quarters of those absences come from two Swedish National Forest Inventory presence-absence vegetation datasets, not from anybody reporting voluntarily. Go to the species and it collapses entirely. Cantharellus cibarius, kantarell: 15,705 Swedish records, and exactly one absence, logged in Uppland on 5 August 2005. Boletus edulis, karljohan: 12,987 records, no absences. Craterellus tubaeformis, trattkantarell: 10,007 records, no absences.
That is the whole argument in four numbers. We have an enormous record of where mushrooms were found and essentially no record of where they were sought.
Why the pile grows this way
Nothing dishonest produced this. A find is an event — it has a name, a date, a coordinate, and a photograph. An empty afternoon is a non-event, and non-events do not naturally become rows in a database. Artportalen actually does have a mechanism for the other half: report a checklist as complete and you are documenting arters frånvaro, the absence of the species you did not see. It is used, mostly by lichen and vascular-plant recorders revisiting known sites. It is barely used at all for edible fungi.
So the pile grows lopsided, and it grows lopsided in space as well as in kind. Mair and Ruete (2016) took Swedish species-observation data across thirteen taxonomic groups — fungi among them — and asked what predicts how hard a given square has been looked at. Road density and log population density came out consistently important. That is the finding stated plainly: the Swedish record is, in substantial part, a map of where Swedes can conveniently park. Add the temporal spikes — a county inventory, a mycological society excursion, one prolific reporter's twenty active years — and a model trained naively on these points learns human behaviour at least as well as it learns biology.
The missing response variable
This is not a data-cleaning problem. It is a structural one. To fit a model that says “the probability a mushroom is here is p” you need cases where it was and cases where it was not, drawn from the same sampling process. Presence-only data gives you only one of those, so the standard workaround is to manufacture the other: drawbackground or pseudo-absence points from the landscape and treat them as the contrast class. Barbet-Massin et al. (2012) is the practical reference for how many to draw and where — and the fact that the question needed a paper at all should tell you how much the answer matters.
Phillips et al. (2009) showed the sharp edge of it. If your presences cluster along roads and your background is drawn uniformly at random, the model does not learn what habitat the species likes; it learns what roadsides look like. Their proposed fix — the “target-group background”, sampling background only from cells where somebody recorded something — helped enough that they concluded the choice of background data mattered as much as the choice of modelling method. That is a startling sentence. The contrast you invent matters as much as the algorithm you invent it for.
And it does not buy you a probability. Elith et al. (2011) are explicit that MaxEnt’s raw output is a relative occurrence rate, not a probability of presence; prevalence is simply not in the data. Hastie and Fithian (2013) went further in Ecography, arguing that claims to recover absolute occurrence probability from presence-only records rest on unjustified parametric assumptions — that the rigidity of the model manufactures information the data does not contain. Guillera-Arroita et al. (2015) put the practical version of it best: presence records suffice for some applications, and the ones that need occurrence probability are “unattainable without reliable absence records”.
Detection is not occurrence
There is a second gap underneath the first, and foragers already know it in their legs. A mushroom can be present and not seen. MacKenzie et al. (2002) is the canonical statement: non-detection does not imply absence unless the probability of detection is one, and it never is. Their occupancy framework separates the two by using repeat visits to the same site — you estimate how often you miss a thing that is there, then correct the occupancy estimate for it. Repeat visits are the price of admission, and presence-only data has none.
For fungi the correction is large. Straatsma, Ayer and Egli (2001) surveyed a 1,500 m² plot at La Chanéaz in Switzerland weekly for 21 years and encountered more than 400 species, many of them transient, appearing in a single year and never again — and concluded that the species count would keep climbing if the survey continued. Halme and Kotiaho (2012), working on wood-inhabiting fungi, found variation between surveys high enough that single-visit diversity estimates are not trustworthy. If two decades of weekly visits to one small plot does not exhaust it, one walk through a hillside establishes very little indeed.
What an empty trip is actually worth
Which is why the record that costs you the most to make is worth the most. A presence pin is the ten-thousandth of its kind and is confounded with road access. A record that saysI searched this stand for ninety minutes on 12 September and found nothing is rare, and it carries three things at once: a location, a negative outcome, and the effort that produced it.
That third element is what makes the negative interpretable. Isaac et al. (2014) is the standard treatment of how effort proxies — list length, visit duration — let you separate real change from changes in how hard people looked. Johnston et al. (2018) showed that modelling observer expertise as a covariate improved species distributions estimated from citizen-science data. Effort recorded at the time is not recoverable afterwards. Nobody can reconstruct, in 2035, how long you looked in 2026.
None of this makes presence-only data worthless. It is excellent for describing where the habitat is right — the environmental envelope a species occupies, the phenological window it fruits in, the tree associations it keeps. That is real and useful knowledge. It is simply a different claim from “there will be mushrooms here on Saturday”, and the honest distance between the two is exactly the distance between presence data and absence data. A blank notebook page from a real walk is not a failed trip. On the evidence, it is the more valuable half of the record — and the half Sweden has almost none of.