Your captures
Result
Every group
best model by AIC, with its intervalChao, with its intervalanimals seen
Estimates
N is the whole number with the highest likelihood; the interval is every N the data cannot rule out at the chosen level. "Unbounded" means the data cannot rule out a very large population. AIC weights compare M0, Mt and Mb only.
Where each estimate sits
Captures per occasion
The example: how many chorus frogs bred at Lily Pond?
Boreal chorus frogs at Lily Pond in Colorado were caught, marked and released on three to five days of each breeding season from 1986 to 2014 (Muths 2020, a U.S. Geological Survey data release). Within one season the breeding population is close to closed, so each year is its own estimate: put year in the group box. Three things stand out.
- The count you see is not the population. In 2010, the best-recaptured year (57 of 270 frogs caught more than once), model Mt puts the breeding population at 539, 95% interval 455 to 656: about two frogs for every one seen.
- The model can matter more than the data. In 1999, 120 frogs were caught on three days, with falling catches (68, 39, 23). Model Mb, which reads that fall as most frogs already being marked, wins on AIC with 95% of the weight and says 135; model Mt, which reads it as falling activity, says 493. Effort records, not AIC, should settle that.
- Few recaptures, no answer. In 2007 only one of 88 frogs was caught twice, and Mt's interval runs from 474 to about 33,300. The page says so rather than pretending to a number.
The bundled file is a reshaping of the USGS file, one row per frog per year it was caught (eval/build-example.py); the original study fitted open-population models across years, which this page does not attempt. Most frogs in the file are calling males, so these are estimates of the frogs that came to be caught, not of every frog nearby.
Method and citation
From the capture histories of the r animals seen over k occasions the page counts, for each occasion, the animals caught (nj), first caught (uj) and already marked (Mj), and how many animals were caught exactly once (f1), twice (f2) and so on. The full likelihood of the histories, a multinomial over the N animals including the N − r never caught, is maximised for three models of Otis, Burnham, White and Anderson (1978): M0, one capture probability; Mt, one per occasion (Darroch 1958); and Mb, one for first capture and another for recapture (a trap response; with no recaptures it is the removal method). The capture probabilities are profiled out in closed form, the best whole-number N is found by a scan and golden-section search, and the interval is the profile-likelihood interval. Models are ranked by AIC (Akaike 1974). Chao's estimator (1987), r + f1²/(2f2), is a lower bound when catchability varies between animals; its interval is log-normal on the number never caught. The jackknife of Burnham and Overton (1978) is shown at orders 1 and 2, as point estimates. With two occasions the page adds Lincoln-Petersen and Chapman (1951) with Seber's (1982) variance.
Closed population. Nothing born, dying, arriving or leaving between the first and last occasion. Estimate separate periods separately.
Marks are kept and read correctly, and animals are caught independently of one another.
Not here (yet): heterogeneity models by likelihood (Mh mixtures, Mth, Mbh), individual covariates (Huggins), open populations (Cormack-Jolly-Seber, Jolly-Seber, the robust design) and spatial capture-recapture.
Checked against: an independent Python reference that maximises the full multinomial over all 2k histories numerically with scipy 1.18.1 on six simulated studies (same whole-number N everywhere, limits to 1e-6); the cottontail numbers printed in Chao (1987); and a coverage simulation of 1,000 studies under each of five models (README, "Verification").
@article{otis1978closed,
author = {Otis, David L. and Burnham, Kenneth P. and White, Gary C. and Anderson, David R.},
title = {Statistical inference from capture data on closed animal populations},
journal = {Wildlife Monographs}, volume = {62}, pages = {3--135}, year = {1978}}
@article{darroch1958multiple,
author = {Darroch, J. N.},
title = {The multiple-recapture census. {I}: Estimation of a closed population},
journal = {Biometrika}, volume = {45}, number = {3/4}, pages = {343--359}, year = {1958},
doi = {10.2307/2333183}}
@article{chapman1951some,
author = {Chapman, Douglas G.},
title = {Some properties of the hypergeometric distribution with applications to zoological sample censuses},
journal = {University of California Publications in Statistics}, volume = {1}, pages = {131--160}, year = {1951}}
@book{seber1982estimation,
author = {Seber, G. A. F.},
title = {The Estimation of Animal Abundance and Related Parameters}, edition = {2nd},
publisher = {Griffin}, address = {London}, year = {1982}}
@article{chao1987unequal,
author = {Chao, Anne},
title = {Estimating the population size for capture-recapture data with unequal catchability},
journal = {Biometrics}, volume = {43}, number = {4}, pages = {783--791}, year = {1987},
doi = {10.2307/2531532}}
@article{burnham1978jackknife,
author = {Burnham, K. P. and Overton, W. S.},
title = {Estimation of the size of a closed population when capture probabilities vary among animals},
journal = {Biometrika}, volume = {65}, number = {3}, pages = {625--633}, year = {1978},
doi = {10.1093/biomet/65.3.625}}
@article{akaike1974new,
author = {Akaike, Hirotugu},
title = {A new look at the statistical model identification},
journal = {IEEE Transactions on Automatic Control}, volume = {19}, number = {6}, pages = {716--723}, year = {1974},
doi = {10.1109/TAC.1974.1100705}}
@misc{muths2020chorus,
author = {Muths, E. L.},
title = {Demographic data from two chorus frog populations in {Colorado}},
howpublished = {U.S. Geological Survey data release}, year = {2020},
doi = {10.5066/P9BZAPLB}}
Questions
What is a closed population?
One with no births, deaths or movement in or out between the first and the last occasion, so the same N animals were there to be caught every time. Keep the occasions close together (a few nights, one breeding season) and estimate each period separately with a group column. If animals arrive or leave between occasions you need an open-population model (Cormack-Jolly-Seber or a robust design), which this page does not fit.
Which estimate should I report?
The page ranks M0, Mt and Mb by AIC and shows each model's weight, but AIC only compares the models it was given. When animals differ in how catchable they are (heterogeneity, almost always true in the field), every one of those three underestimates N, and Chao's estimator, which is a lower bound under heterogeneity, is the better guide. When the models disagree by a lot, report the disagreement: it is real uncertainty about how animals were caught, not noise.
Why do Mb and Mt give such different answers?
Both explain a fall in the number of new animals caught on later occasions, in opposite ways. Mt reads it as lower catch effort or activity, so many animals are still out there; Mb reads it as most animals already being marked and avoiding recapture less or more, so few are left. With falling catches the two can differ several-fold, and the data alone cannot always separate them. Record effort per occasion, and if it changed, trust Mt over Mb.
What is the difference between Lincoln-Petersen and Chapman?
Lincoln-Petersen is n1 times n2 divided by m, the number marked in the second sample. It is biased upwards when m is small and undefined when m is 0. Chapman's version, (n1+1)(n2+1)/(m+1) minus 1, is nearly unbiased when n1 plus n2 is at least N, and the page gives it with Seber's variance. With two occasions model Mt is the same model, fitted by maximum likelihood.
Can I use this for things that are not animals?
Yes, wherever several independent searches find overlapping sets: defects found by several reviewers or test suites, cases on several registries (multiple-list or multiple-systems estimation), duplicate records across sources. Each item is a row and each search is an occasion. The independence assumption is the one to worry about: two lists that draw on the same source overlap more than chance and make the estimate too small.
How are the intervals computed?
For M0, Mt and Mb, by profile likelihood: every N whose log-likelihood is within half the chi-square critical value of the best. They are asymmetric and never go below the number of animals seen, and an upper limit can be unbounded when the data cannot rule out a very large population. Chao's interval is log-transformed on the number never caught (Chao 1987), and Chapman's is the usual normal interval with Seber's variance.
How well do the intervals work?
In 1,000 simulated studies for each model, each model's own 95% interval held the true N 94.5% to 95.5% of the time when the data were made by that model. Under heterogeneity (capture probabilities drawn from a beta distribution) M0, Mt and the AIC-chosen model held it only 2% to 13% of the time, underestimating by about 19%, while Chao's interval held it 85.5% of the time. Chao's estimator in turn overestimated by 14% to 64% when every animal was equally catchable over few occasions. The full table is in eval/coverage-output.txt.
Is my data uploaded anywhere?
No. Every number is computed by JavaScript in this browser tab. There are no accounts, cookies or analytics vendors.