Endangered, and studied
Does being endangered get a bird studied? For 9,113 birds, yes — a Critically Endangered species has about 2.65 times the papers of a Least Concern one, once family, size, range and description date are held fixed. Which direction that runs is not settled here: a bird cannot be listed without evidence, and evidence is research. The published models, on other taxa and older indexes, mostly found the opposite.
Everyone knows that pandas get more attention than frogs. The expected correction is that endangered species get more than their share, because danger draws study. Across the vertebrates it does not work that way. Amphibians are 41% threatened, the highest rate of any fully assessed class, and hold 25% of the threatened vertebrate species; they get 6% of the papers that name their class, 9 per threatened species. Birds hold 11% of the threatened species and get 27% of the papers, 102 per threatened species. Reptiles sit with the amphibians. That is the gap at the level of whole classes, and it is where most treatments stop.
The question this page asks is narrower, and the answer came out the other way. Take one class where every term can be measured for every species, and ask whether a species' own Red List category predicts how much it is studied, once the things that make a species easy or appealing to study are held fixed. For 9,113 birds it does, and upward: a Critically Endangered bird has 2.65 times the papers of a Least Concern one. Adjusting does not shrink the gap — it widens it.
That is not what the literature says. Ten species-level models across mammals, birds, reptiles, carnivores, turtles and plants have asked this question since 2014; of the 6 whose full text could be read for this page, 5 report threat status as null or negative. The disagreement is worth stating before anything is made of it. Those models used Web of Science and Zoological Record, over series ending in 2008 to 2010, across several classes. This page uses OpenAlex title-and-abstract counts for 2015-2024, on birds alone. Fifteen years and a different index is a sufficient explanation for a reversal, and it is the more likely one: a single result that contradicts a literature is more likely to be wrong than the literature. What follows is what one instrument measured on one taxon over one window.
Every number on this page is generated
Six fetch scripts, one build and one figure script. The archive holds all of them and every input whose license allows it. It does not hold the Red List column: the Red List's terms forbid redistributing it, so the build fetches it with the reader's own free token, and the page says so below.
Download code and data
The five vertebrate classes: the share of all threatened vertebrate species each holds, against the share of papers of 2015-2024 whose title or abstract names the class. Assessment coverage is printed beside each class because it is the denominator: the Red List has assessed 78 to 100% of each vertebrate class, and 0.1 to 21% of the invertebrate, plant and fungal groups, which is why no other group is drawn. Threatened counts from Red List 2026-1, Table 1a1; papers from OpenAlex, biology and environment topics only.
What it finds
The class-level picture is the one a reader expects, and it is consistent with two different stories: that attention follows appeal, or that attention follows the practical business of being able to study a species at all. Birds are large, diurnal, widely distributed and watched by millions of amateurs; amphibians are small, nocturnal, often tropical and often described in the last thirty years. To separate the stories you need species, not classes.
Birds are the one class where every term has a free, species-level source. AVONET gives body mass2, range size and family for 11,009 species. The iratebirds project3 gives a rated attractiveness score, from several hundred thousand photograph ratings, for 9,386 of them. The GBIF backbone gives the year4 each species was described. English Wikipedia gives a public-attention measure5 for the 10,645 species with an article. OpenAlex gives the number of papers6 of 2015 to 2024 whose title or abstract names the species. The Red List gives the category. Joined, with the Data Deficient, Extinct and Extinct in the Wild species set aside, that is 9,113 birds.
Papers per bird species by Red List category: as counted, adjusted, and adjusted with public attention held fixed as well. "Typical" is the back-transformed mean of log(1 + papers), which is closer to a median than to a mean for counts this skewed. Each adjusted series predicts every one of the 9,113 species under each category in turn, with its own family, mass, range and age held as they are, and averages. Whiskers are 95% intervals from 300 bootstrap resamples over species. The dotted drop is the share of the advantage that runs through attention; its direction is assumed, for the reason given under Limits.
As counted, a Least Concern bird has 18.9 papers and a Critically Endangered one 30.5. After adjustment the figures are 17.6 and 48.1, a 95% interval of 43.3 to 54.7 for the Critically Endangered species. The coefficient is +0.97 on the log scale — 2.65 times — with a robust p-value under 0.01. The adjustment is doing the work the question asks for: it is what remains of "endangered" once "small, restricted, recently named and in an obscure family" has been accounted for. What remains is larger than what it started with, because the convenience terms were working against threat, not with it: threatened birds tend to be the small, narrowly ranged, recently described ones, and holding that fixed removes a handicap rather than an advantage.
Holding public attention fixed as well takes the coefficient to +0.64, 1.90 times. The difference between the two is the share of the advantage that travels with attention: 16% at Near Threatened, 21% at Vulnerable, 21% at Endangered and 34% at Critically Endangered. It rises with severity. Both numbers are on the page because they answer different questions — the first is what being endangered is worth in total, the second is what it is worth among birds people look up equally often — and neither is the "true" one.
What the full model's fit depends on: the fall in R² when each term is removed and the model refitted, with 95% bootstrap intervals. The full model adds rated attractiveness and Wikipedia views to the adjusted model above; R² rises from 0.48 to 0.59.
Ranked by what the fit loses without them, family comes first (0.141) and wikipedia views second (0.104). Red List category is 3rd of seven at 0.021, above range size and years since description.
The two measures of appeal do not behave alike, and this page cannot use the word undivided. Rated attractiveness — scored from photographs by several thousand people — contributes 0.001, below body mass and indistinguishable from zero. Wikipedia views contribute 0.104, 135 times as much. The thing this project was named for explains nothing here; the thing that does is how often people look a bird up. Weigh that second term with care: views are partly a consequence of research as well as a cause of it, and the model cannot tell the two apart.
The published record says the same, taxon by taxon. Ten species-level models of research effort since 2014 have included threat status beside size, range, description date or phylogeny. Where the full text could be read, the threat term is null in 3, negative in 2, and positive only for being assessed at all rather than for being threatened. Three are marked as read from the abstract alone because the full text sits behind a paywall or a bot wall, and one could not be read at all; their verdicts are left as unread rather than guessed.
| Model | Taxon, species | Threat term | Read from |
|---|---|---|---|
| Brooke et al. 20147 | Carnivores, 286 | null | full text |
| Ducatez & Lefebvre 2014 | Birds, 10,064 | negative | full text |
| McKenzie & Robertson 20158 | British breeding birds, 225 | null | full text |
| Fleming & Bateman 20169 | Australian terrestrial mammals, 331 | unread | abstract |
| dos Santos et al. 202010 | Terrestrial mammals, 4,108 | weak | abstract |
| Tam et al. 202211 | Mammals, 5,497 | negative | full text |
| Mammola et al. 202312 | 29 phyla, 3,019 | assessed, positive | full text |
| Guedes et al. 202313 | Reptiles, 10,531 | unread | abstract |
| Adamo et al. 202114 | Alpine plants, 113 | unread | none |
| Ducatez & DeVore 202615 | Turtles, 370 | null | full text |
"Assessed, positive" means the paper found that species on the Red List at all are more studied than unassessed ones, with no further effect of the category; the authors of that study note the circularity, since an assessment needs data to exist first. "Negative, unadjusted" means a bivariate comparison rather than a model term. One entry, on British birds, found the international Red List category null while the national action-plan listing was a significant positive predictor, which is the one place in this record where a conservation listing does buy attention: a national one, with money and a monitoring obligation behind it.
Funding behaves the way the papers do. In the first study of United States endangered-species spending16, over half of the identifiable money went to ten species, and body length and being a mammal or bird outweighed the agency's own endangerment rank. Birds and mammals take 75% of the European Union's LIFE species-project budget17. In a global database of conservation projects18 over twenty-five years, about 6% of threatened species were ever funded and 29% of the money went to species of Least Concern. Those are looked up, not recomputed here.
Is it just the paperwork?
A listing generates its own literature. Status reviews, Red List updates and assessment documents all name the species, so a Critically Endangered bird could look well studied because being assessed produces writing rather than because anyone studies the animal. That would make the whole finding an artefact of the instrument.
It is testable, and the test was fixed in writing before the data to run
it existed. Every OpenAlex work carries a primary topic, so the model can
be refitted inside one field at a time. The rule — which field is the
primary test, how many species each category needs before a field may be
read at all, and what result would count as failure — was committed to
PRESPEC_topics_2026-09-09.md before a single topic count was
fetched, because choosing the field after seeing the coefficients would
leave no trace: every number correctly computed and the conclusion still
unearned.
The Critically Endangered coefficient refitted inside each topic field, against the same terms as the adjusted model. Brown marks the fields that carry conservation writing; blue marks those that cannot. The pre-registered primary field is Biochemistry, Genetics and Molecular Biology, the floor was set at +0.20, and a field is read only if every Red List category has at least 30 species with a paper in it. 3 fields fail that floor and are named without a value rather than shown at a smaller sample.
The effect survives. In molecular biology — where a status review cannot be published — the Critically Endangered coefficient is +0.721, 2.06 times Least Concern, against a floor set in advance at +0.20. Earth and Planetary Sciences, the other field clearing the cell rule, gives +0.358. Environmental Science, the positive control, gives the largest value on the page, +0.983: the artefact is real and measurable, and it is not what the effect is made of.
Its size can be put a number on. Refitting on all papers except the two conservation-carrying fields gives +0.888 against +0.966 on the whole corpus — about 8% of the headline coefficient is assessment writing. That comparison was predicted in the pre-specification to understate the effect, because deleting whole fields also deletes genuine conservation biology, which is research about the species; it is reported because it was predicted, not because it came out well.
Method
The response is log(1 + papers), where a paper is an OpenAlex work of 2015 to 2024 whose title or abstract contains the species' scientific name as a phrase. That definition is assumed: it misses papers that use only a common name and catches a few that mention a species in passing, and it treats a note and a monograph alike. Two ordinary least-squares fits, each with a fixed effect for every family and heteroskedasticity-robust standard errors. The adjusted model has log body mass, log range size, years since description to 2024, and the Red List category with Least Concern as the reference. The full model adds rated attractiveness and log(1 + mean monthly Wikipedia views). Family fixed effects standing in for a phylogeny is assumed; the published bird model used a tree19 and attributed 74% of the variance to it.
The category means are computed rather than read off coefficients: every species is predicted under each of the five categories with its other terms as they are, and the predictions are averaged and back-transformed. The share of the fit each term carries is the fall in R² when that term alone is dropped and the model refitted. Intervals are percentile bootstrap over species, 300 resamples, seed 20260905. Species without an English Wikipedia article are given zero views rather than dropped, which is assumed and keeps the obscure species in; 260 species assessed as Data Deficient, Extinct or Extinct in the Wild are left out because the question is about the categories that rank risk.
The Red List column on this page came from the Red List API, version 2026-1, matched by scientific name to the HBW-BirdLife v5 names AVONET uses; 10,669 of 11,009 names matched. The class-level figure uses the Red List's own summary table and a class-name search of OpenAlex, restricted to works whose primary topic is in the biological or environmental sciences so that "bird" does not count ornithopter patents. That proxy is defensible for vertebrate classes, whose conservation papers name the class, and not beyond them, where a paper on rice never says "angiosperm"; the feasibility work behind this page tried the wider version and found the numbers it produced were artefacts of vocabulary.
The exact form of the species query is pinned, and the reason is the
clearest illustration on this site of a number being wrong while nothing
goes wrong. OpenAlex offers two ways to ask for works mentioning a phrase.
Until February 2026 the short one20, search=, meant the title
and the abstract; it now means the full text, and
filter=title_and_abstract.search: means what the short one
used to. On 5 September 2026 a species with a large literature,
the lion, returned 11,759 works through the short
parameter and 3,459 through the pinned one: a factor of
3.4 in the quantity this whole page rests on. Both queries are
valid, both are answered, and a build that reached for the shorter one
would have finished without an error and published counts 3.4
times too large. So the query is fixed in one place and a test asserts the
URL the fetcher constructs, which is the only way that particular mistake
can be caught: it is a change in someone else's API, not a fault in this
one's code.
Limits
Association, not cause, and here that is not a formality. A bird cannot be listed Critically Endangered without evidence: Least Concern is the cheap default, while the higher categories need population size, trend and range data, and that data is research. So the arrow may run from research to listing rather than from listing to research, and the Red List category may be partly a record of how well studied a species already was. The gradient this page reports — more threatened, more studied — is exactly what rising evidence requirements would produce, from a mechanism that has nothing to do with attention or with threat.
The cheapest check on that story is the category that means nobody has enough data. 34 birds are assessed Data Deficient, and they are the least-studied group in the table: a median of 9 papers against 15 for Least Concern, the lowest at every quartile. That is 34 species, far too few to fit anything, so it is reported as a description and given no coefficient and no point on any figure. Those raw counts are on a wider frame than the model: all 10,669 birds with a Red List category and a paper count, where every fitted number on this page uses the 9,113 that also carry an attractiveness rating. But it points the way the evidence story predicts. This page reports an association between threat category and research volume and does not establish its direction. Two mechanisms produce that association — threat attracting research, and research enabling the listing — and the stratum test above separates only a third one, the paperwork.
The raw counts are not monotone, and the fitted values are. Among the raw medians, Near Threatened (13) sits below Least Concern (15), and Critically Endangered (21) below Endangered (22). Only the adjusted series rises without exception, because the raw counts are confounded by exactly what the model controls for — family, body mass, range size and years since description — and threatened birds are on the wrong side of all four. Extinct in the Wild, which the model leaves out, is the most-studied group of all at a median of 28 papers across 5 species: birds with captive populations generating literature, which is the evidence mechanism at its most visible.
Rated attractiveness is itself correlated with size, colour and familiarity, and public attention with research; what the model can say is how the fit partitions among named terms, and what it cannot say is that any one of them makes scientists choose. "Beauty determines funding" is a sentence this page does not write. Nor is the ordering of threat, attention and research a finding: a cross-section with attention measured at one moment cannot tell threat → attention → research from research → a better article → views. The 34% is what the model attributes, and its direction is assumed.
The model needs a rated attractiveness score, which 1,472 of the 10,585 otherwise-complete birds do not have, and those species are marginally less studied and more often Critically Endangered. That could bias the coefficient, so it was measured rather than argued: fitting the same terms on both samples moves the Critically Endangered coefficient from +0.864 to +0.966, with every interval overlapping. The selection is real and its effect on this page's claim is not.
The attention measure is the weakest of the inputs and is measured badly in two known ways. 364 species have no English Wikipedia article and are given zero views, which is a floor rather than a measurement. And the bird taxonomy this page uses splits species that Wikipedia still covers in a single article, so 273 species share 129 articles between them and each is credited with the whole of an article's traffic: the four pheasant pigeons are one page. Refitting without the 12 such species that reach the model moves the Red List term from 0.021 to 0.021, and refitting on only the species that have an article of their own moves it to 0.013. The finding does not rest on either.
The denominator is the bias. The Red List has assessed between 78 and 100% of each vertebrate class and between 0.1 and 21% of everything else, so the groups that carry most of the extinction burden are invisible in the very statistic used to count it. Of the roughly two million named species, one estimate puts21 7.5-13% extinct since 1500 against 0.04% on the Red List; 64% of Data Deficient mammals22 and 85% of Data Deficient amphibians23 are predicted to be threatened once modeled. This page's class figure stops at vertebrates for that reason, and its species model stops at birds because birds are the one class with a rated-attractiveness dataset.
Some stories about neglected extinction do not survive checking, and they are left out. The ridge of Centinela in Ecuador, long cited as a place where dozens of plant species vanished before anyone had studied them, was resurveyed in 202424 and almost all of its supposed endemics have since been collected elsewhere. The Christmas Island pipistrelle, cited as a bat lost for want of attention, was monitored for fifteen years25 before it disappeared; the failure was a decision delayed, not a species unstudied. Both are better told as they happened, and neither is evidence for this page's claim.
What is absent, and why. There is no zoo figure: captive-collection data at species level are held in a members-only database, and the published finding that zoo holdings favor large26, attractive and non-threatened species is cited27 rather than drawn. There is no funding computation: the per-species expenditure reports exist only as scanned tables. There is no forecast of what the gap will cost; that is an argument, and this page carries findings. Mammals, with a public h-index dataset of their own28, are the obvious next class and are not built.
The code
Six fetch scripts, one per source, each resumable; the build, which
joins, fits and bootstraps and writes the payload; the figure script; and
the page renderer. The archive is thinner than the site's others by one
column. The IUCN Red List's terms of use prohibit redistributing Red List
data in any form, so no file in it carries a category beside a species
name: fetch_iucn.py gets the column from a free personal
token, and the build labels its output with the Red List version it ran
against. Every other input travels: the AVONET and attractiveness tables
under CC BY 4.0, the Wikipedia and OpenAlex counts under CC0, the Red
List summary table as a cited aggregate.
What travels in place of the withheld table is a fingerprint of it. The
archive's manifest.json carries the SHA-256 of the
11,185 species and categories this page was built from, sorted and
canonicalised, 1f1d4b6f1ff2c81d; after fetching the
categories yourself, build_beauty.py --verify-categories
tells you whether you have the same table, a Red List reassessed since,
or a name-matching problem. Sixty-four characters cannot be turned back
into eleven thousand categories, so the digest goes where the table may
not, and the one input a reader has to supply is still the one input they
can check.
OpenAlex has metered its API since February 2026. The per-species counts shipped in the archive cost $11.01 to fetch across 11,009 species, on a free account's allowance of a dollar a day; a reader who trusts the shipped counts pays nothing, and one who re-fetches them needs a free key and 12 days. The build runs without a Red List token and says so in its output, taking the 2021 categories that travel with the attractiveness deposit; the page that ships was built from the API.
Download source · 1,488 KB All code
Sources
- IUCN, The IUCN Red List of Threatened Species, version 2026-1, summary statistics Table 1a, 2026, https://www.iucnredlist.org.Described, evaluated and threatened species by major group, and the category per bird species through the Red List API.
- Tobias et al., Ecology Letters 25(3):581-597, 2022, doi:10.1111/ele.13898.Body mass, range size, family and order for 11,009 bird species on the HBW-BirdLife v5 taxonomy; CC BY 4.0.
- Santangeli et al., npj Biodiversity 2:20, 2023, doi:10.1038/s44185-023-00026-2.Rated attractiveness per species and sex from over 400,000 photograph ratings by 6,212 raters; the deposit is CC BY 4.0.
- GBIF Secretariat, GBIF Backbone Taxonomy, checklist dataset, doi:10.15468/39omei.The full scientific name with authorship, from which the year of description is read.
- Mittermeier et al., Conservation Biology 35(2):412-423, 2021, doi:10.1111/cobi.13702.Wikipedia pageviews as a measure of public interest in species, the method this page borrows.
- Priem, Piwowar & Orr, arXiv 2205.01833, 2022.The open index of scholarly works whose title-and-abstract search gives the paper counts; data CC0.
- Brooke, Bielby, Nambiar & Carbone, PLoS ONE 9(4):e93195, 2014, doi:10.1371/journal.pone.0093195.286 carnivores; IUCN status estimate -0.047, z = -0.82, p = 0.41 beside body mass, range and diet breadth.
- McKenzie & Robertson, PLoS ONE 10(7):e0131004, 2015, doi:10.1371/journal.pone.0131004.225 British breeding birds; Red List status null (p = 0.34), national action-plan status positive (t = 3.14).
- Fleming & Bateman, Mammal Review 46(4):241-254, 2016, doi:10.1111/mam.12066.331 Australian mammals; the threat-term result is in the paywalled full text and was not read for this page.
- dos Santos et al., Animal Conservation 23(6):679-688, 2020, doi:10.1111/acv.12586.4,108 terrestrial mammals, hurdle model; the abstract reports threat status as weakly associated with research presence and volume.
- Tam, Lagisz, Cornwell & Nakagawa, GigaScience 11:giac074, 2022, doi:10.1093/gigascience/giac074.5,497 mammals; IUCN status linear -16.4 [-19.3, -13.6] with a quadratic term; Google Trends the strongest fixed effect.
- Mammola et al., eLife 12:RP88251, 2023, doi:10.7554/eLife.88251.3,019 species across 29 phyla; being on the Red List raises interest whether the species is endangered or of Least Concern.
- Guedes, Moura & Diniz-Filho, Ecography 2023:e06491, 2023, doi:10.1111/ecog.06491.10,531 reptiles; body size, description year and proximity to institutions dominate. Full text not reached for this page.
- Adamo et al., Nature Plants 7:574-578, 2021, doi:10.1038/s41477-021-00912-2.113 Alpine plants; color, conspicuousness and range predict research attention. Paywalled; not read for this page.
- Ducatez & DeVore, PLoS ONE 21:e0347198, 2026, doi:10.1371/journal.pone.0347198.370 turtles; extinction risk -0.10, t = -0.46, p = 0.64; phylogeny 66% of the variance; assessed species more studied than unassessed.
- Metrick & Weitzman, Land Economics 72(1):1-16, 1996, doi:10.2307/3147153.Endangered Species Act spending FY1989-91: over half to ten species; body length and taxon outweigh endangerment rank.
- Mammides, Biodiversity and Conservation 28(5):1291-1296, 2019, doi:10.1007/s10531-019-01725-8.Over 800 EU LIFE species projects since 1992; birds and mammals take 69% of projects and 75% of the species budget.
- Guénard et al., PNAS 122:e2412479122, 2025, doi:10.1073/pnas.2412479122.About 14,600 projects over 25 years: 6% of threatened species ever funded, 29% of funds to Least Concern species.
- Ducatez & Lefebvre, PLoS ONE 9(2):e89955, 2014, doi:10.1371/journal.pone.0089955.10,064 birds; phylogeny takes 74% of the variance in research effort; Least Concern species have twice the papers of threatened ones.
- OpenAlex, “New features and usage-based pricing”, 24 February 2026, https://blog.openalex.org/openalex-api-new-features-and-usage-based-pricing/.The date the bare search parameter was redefined as full-text search and the API became metered, at $0.0001 for a filtered list call.
- Cowie, Bouchet & Fontaine, Biological Reviews 97(2):640-663, 2022, doi:10.1111/brv.12816.7.5-13% of the roughly two million known species estimated extinct since 1500, against 0.04% recorded on the Red List.
- Bland et al., Conservation Biology 29(1):250-259, 2015, doi:10.1111/cobi.12372.64% of Data Deficient mammals predicted to be threatened.
- Borgelt et al., Communications Biology 5:679, 2022, doi:10.1038/s42003-022-03638-9.85% of Data Deficient amphibians predicted to be threatened.
- White et al., Nature Plants 10:1627-1634, 2024, doi:10.1038/s41477-024-01832-7.99% of Centinela's supposed microendemic plants have since been collected elsewhere.
- Martin et al., Conservation Letters 5(4):274-280, 2012, doi:10.1111/j.1755-263X.2012.00239.x.The Christmas Island pipistrelle was monitored from 1994 and lost in 2009 while a captive-breeding decision was delayed.
- Frynta et al., PLoS ONE 8(5):e63110, 2013, doi:10.1371/journal.pone.0063110.Rated beauty and body size predict which mammal families zoos hold and in what numbers; Red List status does not.
- Conde et al., PLoS ONE 8(12):e80311, 2013, doi:10.1371/journal.pone.0080311.3,955 species in 837 zoos, 23% threatened; only 2 of 59 orders hold more threatened species than random would.
- Tam et al., GigaDB dataset 102237, 2022, doi:10.5524/102237.Per-species h-index, publication count, Red List category and Google Trends for 7,521 mammals; CC0.