= Negative binomial information in capture-recapture sampling
{title2=$\widehat p=\frac r{r+\bar x},\quad I_n(p)=\frac{nr}{p^2(1-p)}$}
Suppose a population has $N$ individuals, $k$ tagged, and captures are made independently with immediate replacement. Stop each run at the $r$th tagged capture, with known positive <integer> $r$, and let $X$ count untagged captures before stopping. Then $X$ has a failure-count <negative binomial distribution> with tagged-capture <probability> $p=k/N$ and mass $\binom{x+r-1}{x}p^r(1-p)^x$.
For $n$ <independent> runs and $s=\sum_i x_i$, the log-<likelihood function> apart from constants is $nr\log p+s\log(1-p)$. Treating $p$ as continuous in $(0,1)$ gives
$$
\widehat p=\frac{nr}{nr+s}=\frac r{r+\overline x},\qquad
I_n(p)=\frac{nr}{p^2(1-p)}.
$$
Indeed, $\mathbb EX=r(1-p)/p$ yields the second-derivative <Fisher information>. For an interior true $p$, the usual continuous-parameter <asymptotic variance> is $p^2(1-p)/(nr)$. If $s=0$, the likelihood is maximized at $p=1$ on the closed <parameter space>.
For actual unknown population size $N$, the constraint $p=k/N$ restricts $p$ to a discrete set. The relaxed estimate $\widehat N=k/\widehat p$ need not be an <integer>; evaluate the likelihood at the admissible <integers> on either side of it rather than claiming the continuous optimum is an attained integer-population <maximum-likelihood estimator>. A <confidence interval> $(L,U)$ for $p$, with $0<L<U$, transforms to $(k/U,k/L)$ and then is intersected with the allowed <integer> population sizes. Immediate replacement is essential; sampling without replacement gives a different likelihood.
Back to article page